The Capco Institute
Journal of Financial Transformation #32 08.2011
JournalThe Capco Institute Journal of Financial Transformation
#32Applied Finance
CApCo.Com
AmsterdamAntwerp
BangaloreChicago
FrankfurtGenevaLondon
New Yorkparis
San FranciscoToronto
Washington, D.C.Zurich
Recipient of the Apex Awards for Publication Excellence 2002-2011
08.2011
With an unmatched breadth of products and depth of knowledge, FIS is uniquely capable of helping you build and maintain a distinctive position in your market.
We are the trusted partner that offers you the greatest range of choices – not only in what technologies and services to deploy, but also how. This can include customizing software functions, private brandingand tailoring customer interfaces or leveraging our deep suite of outsourcing and managed services. With 30,000 dedicated people in over 100 countries and the recent addition of the Capco organization, wecan apply unprecedented power to your unique market strategy. Look closer – and discover the power andvalue of FIS. Go to fisglobal.com
One company has the precision and focusto help you redefine value in a competitive market.
fisglobal.com
Visit our new Web site
Learn about our industry-leadingproducts, services and solutions.
Go to fisglobal.com
© 2011 Fidelity National Information Services, Inc. and its subsidiaries.
FINANCIAL SOLUTIONS • PAYMENT SOLUTIONS • BUSINESS CONSULTINg • TECHNOLOgY SERVICES
FS244_gem Ad.indd 1 2/16/11 4:44 PM
JournalEditorShahin Shojai, Global Head of Strategic Research, Capco
Advisory EditorsCornel Bender, Partner, CapcoChristopher Hamilton, Partner, CapcoNick Jackson, Partner, Capco
Editorial BoardFranklin Allen, Nippon Life Professor of Finance, The Wharton School, University of PennsylvaniaJoe Anastasio, Partner, CapcoPhilippe d’Arvisenet, Group Chief Economist, BNP ParibasRudi Bogni, former Chief Executive Officer, UBS Private BankingBruno Bonati, Strategic Consultant, Bruno Bonati ConsultingDavid Clark, NED on the board of financial institutions and a former senior advisor to the FSAGéry Daeninck, former CEO, RobecoStephen C. Daffron, Global Head, Operations, Institutional Trading & Investment Banking, Morgan StanleyDouglas W. Diamond, Merton H. Miller Distinguished Service Professor of Finance, Graduate School of Business, University of ChicagoElroy Dimson, BGI Professor of Investment Management, London Business SchoolNicholas Economides, Professor of Economics, Leonard N. Stern School of Business, New York UniversityMichael Enthoven, Former Chief Executive Officer, NIBC Bank N.V.José Luis Escrivá, Group Chief Economist, Grupo BBVAGeorge Feiger, Executive Vice President and Head of Wealth Management, Zions BancorporationGregorio de Felice, Group Chief Economist, Banca IntesaHans Geiger, Professor of Banking, Swiss Banking Institute, University of ZurichPeter Gomber, Full Professor, Chair of e-Finance, Goethe University FrankfurtWilfried Hauck, Chief Executive Officer, Allianz Dresdner Asset Management International GmbHMichael D. Hayford, Corporate Executive Vice President, Chief Financial Officer, FISPierre Hillion, de Picciotto Chaired Professor of Alternative Investments and Shell Professor of Finance, INSEADThomas Kloet, Chief Executive Officer, TMX Group Inc.Mitchel Lenson, former Group Head of IT and Operations, Deutsche Bank GroupDonald A. Marchand, Professor of Strategy and Information Management, IMD and Chairman and President of enterpriseIQ®
Colin Mayer, Peter Moores Dean, Saïd Business School, Oxford UniversityJohn Owen, Chief Operating Officer, Matrix GroupSteve Perry, Executive Vice President, Visa EuropeDerek Sach, Managing Director, Specialized Lending Services, The Royal Bank of ScotlandManMohan S. Sodhi, Professor in Operations & Supply Chain Management, Cass Business School, City University LondonJohn Taysom, Founder & Joint CEO, The Reuters Greenhouse FundGraham Vickery, Head of Information Economy Unit, OECDNorbert Walter, Managing Director, Walter & Daughters Consult
Part 19 Measuring Financial Supervision
Architectures and the Role of Central BanksDonato Masciandaro, Marc Quintyn
15 What Reforms for the Credit Rating Industry? A European PerspectiveKarel Lannoo
23 Money Market Funds and Financial Stability: Comparing Sweden to the U.S. and IcelandGudrun Gunnarsdottir, Maria Strömqvist
35 Interest Rates After the Credit Crunch: Markets and Models EvolutionMarco Bianchetti, Mattia Carlicchi
49 Fat Tails and (Un)happy Endings: Correlation Bias, Systemic Risk and Prudential RegulationJorge A. Chan-Lau
59 Empirical Analysis, Trading Strategies, and Risk Models for Defaulted Debt SecuritiesMichael Jacobs, Jr.
75 Systemic Risk, an Empirical ApproachGonzalo de Cadenas-Santiago, Lara de Mesa, Alicia Sanchís
89 Price of Risk – Recent Evidence from Large FinancialsManmohan Singh, Karim Youssef
Part 299 Simulation and Performance Evaluation of
Liability Driven Investment (LDI)Katharina Schwaiger, Gautam Mitra
107 Behavioral Finance and Technical AnalysisKosrow Dehnad
113 The Failure of Financial Econometrics: Assessing the Cointegration “Revolution”Imad Moosa
123 A General Structural Approach For Credit Modeling Under Stochastic VolatilityMarcos Escobar, Tim Friederich, Luis Seco, Rudi Zagst
133 A Stochastic Programming Model to Minimize Volume Liquidity Risk in Commodity TradingEmmanuel Fragnière, Iliya Markov
143 The Organization of Lending and the Use of Credit Scoring Techniques in Italian BanksGiorgio Albareto, Michele Benvenuti, Sauro Mocetti, Marcello Pagnini, Paola Rossi
159 Measuring the Economic Gains of Mergers and Acquisitions: Is it Time for a Change?Antonios Antoniou, Philippe Arbour, Huainan Zhao
169 Mobile Payments Go Viral: M-PESA in KenyaIgnacio Mas, Dan Radcliffe
Applied Finance
Over the last ten years we have been
proud to share with you the finance
perspectives of both academics and
practitioners, many of which have been
provocative and heavily debated in light
of the stress scenario we experienced in
the 2008 financial crisis. Even now the
markets remain challenged by the weight
of sovereign debt and the potential for
default.
Applied finance is more than just impres-
sive models. In recent years many had
started to believe that finance was rela-
tively simple, even though the underpin-
ning methodologies and mathematical
models had become increasingly more
complicated. However, the recent crisis
proved that finance is significantly more
complex than the models that support it.
Many of our Journals have focused on
the complex models and finance theo-
ries that have become the tools of our
trade. More recently we have been high-
lighting the operational and technology
complexities that make these theories
and models more vulnerable. These ele-
ments simply compound the complex-
ity of the systemic risk within our world
economies. The simple click of a button,
to trade a single share, sets in motion an
extremely complicated process in which
there are still far too many human inter-
ventions. As the complexity of the instru-
ments increase, so do the levels of risk
and the complexity of the processes that
follow the initiation of the trade. Conse-
quently, for us at Capco, the complexi-
ties of finance are as much about the
processes that are involved in managing
the industry and its many different par-
ticipants as about simply price and man-
aging complex instruments.
That is why, in this issue of the Journal
and the next, the focus is on the com-
plexity in finance from a more practical,
and more importantly, managerial per-
spective.
As always we welcome your candor and
healthy debate. By so doing we continue
to strive with you to form the future of
finance.
Rob Heyvaert,
Founder and CEO, Capco
Dear Reader,
The finance discipline is one of those
rare branches of economics that is
deemed to be practical by its very
nature. It deals with how corporations
use banks or financial markets to raise
capital for investments that lead to
economic development, how individu-
als or their representatives participate
in this economic process by investing
in shares and bonds, and how com-
panies can put these funds to the best
possible use.
The finance discipline not only advises
recipients of capital on how to best
apply it. It also provides guidelines on
how the markets should be structured
to best serve and protect borrowers
and investors, and how these investors
should make investment decisions.
For over 40 years it was accepted wis-
dom that the finance discipline was
very effective in ensuring that markets
operated in ways that best met the
needs of the participants. Of course,
with the introduction of more quanti-
tative tools, the subject became even
more “applied”. This led to the devel-
opment of more complex financing
tools and investment vehicles.
The recent crisis has put the discipline
under a very critical spotlight. Many
have started to question whether 40
years of academic research have really
helped create markets that are indeed
efficient. They also question the valid-
ity of the tools developed to price com-
plex instruments and risk itself – the
foundation of the entire discipline.
We at the Journal of Financial Transfor-
mation firmly believe that there is still
quite a large gap between the research
undertaken at leading finance acad-
emies, some of which has permeated
into the real world, and what investors
and borrowers need and apply. We
believe that future research in finance
should focus on what actually happens
in the markets rather than what should.
That is why this issue of the Journal
is dedicated to applied finance. The
articles that have passed our rigorous
review process not only make a genu-
ine contribution to the field, they are
also practical in their focus.
We hope that you enjoy the practical
nature of the papers in this edition of
the Journal and that you will continue
to support us by submitting your best
ideas to us.
On behalf of the board of editors
Finance Applied
Part 1Measuring Financial Supervision Architectures and the Role of Central Banks
What Reforms for the Credit Rating Industry? A European Perspective
Money Market Funds and Financial Stability: Comparing Sweden to the U.S. and Iceland
Interest Rates After the Credit Crunch: Markets and Models Evolution
Fat Tails and (Un)happy Endings: Correlation Bias, Systemic Risk and Prudential Regulation
Empirical Analysis, Trading Strategies, and Risk Models for Defaulted Debt Securities
Systemic Risk, an Empirical Approach
Price of Risk – Recent Evidence from Large Financials
9
PART 1
Measuring Financial Supervision Architectures and the Role of Central Banks
AbstractToday, policymakers in all countries, shocked by the financial
crisis of 2007-2008, are reconsidering carefully the features
of their supervisory regimes. This paper reviews the chang-
ing face of the financial supervisory regimes before and after
the crisis, introducing new indicators to measure the level of
consolidation in supervision and the degree of the involve-
ment of the central banks.
Donato Masciandaro — Professor of Economics, Chair in Economics of Financial Regulation, Director, Paolo Baffi Centre on Central Banking and Financial Regulation, Bocconi University
Marc Quintyn — Division chief, IMF Institute, International Monetary Fund
10
Until around 15 years ago, the issue of the shape of the financial super-
visory architecture was considered irrelevant. The fact that only banking
systems were subject to robust and systematic supervision kept several
of the current issues in the sphere of irrelevance. Since then, financial
market development, resulting in the growing importance of insurance,
securities, and pension fund sectors, has made supervision of a grow-
ing number of non-bank financial intermediaries, as well as the investor
protection dimension of supervision, highly relevant.
In June 1998, most of the responsibility for banking supervision in the
U.K. was transferred from the Bank of England to the newly established
Financial Services Authority (FSA), which was charged with supervising
all segments of the financial system. For the first time a large industrial-
ized country – as well as one of the main international financial centers
– had decided to assign the main task of supervising the entire financial
system to a single authority, other than the central bank (the U.K. regime
was labeled the tripartite system, stressing the need for coordination in
pursuing financial stability between the FSA, the Bank of England, and
the Treasury). The Scandinavian countries – Norway (1986), Iceland and
Denmark (1988) and Sweden (1991) – had preceded the U.K. in the after-
math of domestic financial crises. But after that symbolic date of 1998,
the number of unified supervisory agencies started to grow rapidly.
Europe has been the center of gravity regarding the trend towards super-
visory consolidation (or unification). In addition to the U.K., three “old”
European Union member states – Austria (2002), Belgium (2004), and
Germany (2002) – also placed financial supervision under a non-central
bank single authority. In Ireland (2003) and the Czech and Slovak Re-
publics (2006), the supervisory responsibilities were concentrated in the
hands of the central bank. Five countries that are part of the E.U. en-
largement process – Estonia (1999), Latvia (1998), Malta (2002), Hungary
(2000), and Poland (2006) – also concentrated all supervisory powers in
the hands of a single authority. Outside Europe, unified agencies have
been established in, among others, Colombia, Kazakhstan, Korea, Ja-
pan, Nicaragua, and Rwanda.
Then came the crisis. Most accounts of the contributing factors point
to macroeconomic failures and imbalances as well as regulatory failures
[see for instance, Allen and Carletti (2009), Brunnermeier et al. (2009),
Buiter (2008)]. While some of these studies also mention, in passing, cer-
tain supervisory failures, some other studies analyze in more depth the
contribution of supervisory failures to the crisis [for a survey see Mascian-
daro et al. 2011]. Some of them explicitly mention flaws in the supervisory
architecture as a contributing factor in some countries [Buiter (2008) for
the U.S. and the U.K. and Leijonhufvud (2009), for the U.S.].
Thus, in response to the 2007-2008 financial crisis, different countries
and regions – the European Union, Belgium, Germany, Ireland, the U.K.,
and the U.S., among others – are either implementing or evaluating the
possibility of introducing reforms aimed at reshaping the supervisory ar-
chitecture and the role of the central bank. In addition, Finland also es-
tablished a unified supervisor in 2009.
In July 2010, U.S. President Barack Obama signed into law the Dodd-
Frank Act, which is considered the most important U.S. financial regu-
lation overhaul since the Great Depression. A rethink of the roles and
responsibilities of the Fed has been part of the broad financial legislation
restyling. Despite the fact that during the discussions of the bill U.S.
lawmakers debated the possibility of restricting some of the Fed’s regu-
latory responsibilities (supervision of small banks, emergency lending
powers), as well as to increase the political control on the central bank
with changes in its governance (congressional audits of monetary policy
decisions, presidential nomination of the New York Fed Presidents), the
Dodd-Frank law ended up increasing the powers of the Fed as a bank-
ing supervisor.
In Europe, policymakers are moving to finalize reforms concerning the
extent of central banks’ involvement in supervision, both at international
and at the national levels. In 2009, the European Commission enacted
a proposal for the establishment of a European Systemic Risk Council
(ESRC) for macro-prudential supervision which should be dominated by
the ECB. With regards to the individual E.U. members, in 2008 the Ger-
man grand-coalition government expressed its willingness to dismantle
the unique financial supervisor (BAFIN) in favor of the Bundesbank. Bel-
gium also expressed an interest to follow suit. In June 2010, the U.K.
government unveiled a reform of the bank supervisory system aimed
at consolidating powers within the Bank of England. The key functions
of the Financial Services Authority would be moved inside the Bank of
England, which will become the Prudential Regulatory Authority. Finally,
in summer 2010 the Irish Financial Services Regulatory Authority was
legally merged with the central bank.
These episodes show that the financial supervisory architecture remains
in a state of flux, and the latest round seems to provide signals of a sort of
“Great Reversal,” given that it has been shown that before the crisis the
direction of the changes in the supervisory structure was characterized
by central bank specialization in pursuing monetary policy as a unique
mandate [Masciandaro and Quintyn (2009), Orphanides (2010)].
This paper reviews the changing face of the financial supervisory regimes
by introducing new indicators to measure the level of consolidation in su-
pervision and the degree of the involvement of the central banks. We de-
fine supervision as the activity that implements and enforces regulation.
While regulation refers to the rules that govern the conduct of the inter-
mediaries, supervision is the monitoring practice that one or more public
authorities undertake in order to ensure compliance with the regulatory
11
framework [Barth et al. (2006)]. However, we use the term “regulatory”
and “supervisory” authorities interchangeably, as is done in most of the
literature.
In general, the focus of this paper is on micro-prudential supervision
and consumer protection, where macro-prudential supervision is usually
carried out by the central bank and competition policy is in the hands
of a specialized authority [Borio (2003), Kremers et al. (2003), Čihák and
Podpiera (2007), Herring and Carmassi (2008)]. Micro-prudential super-
vision is the general activity of safeguarding financial soundness at the
level of individual financial firms, while macro-supervision is focused on
monitoring the threats to financial stability that can arise from macroeco-
nomic developments and from developments within the financial system
as a whole [Commission of the European Communities (2009)]. In par-
ticular we shed light on the reforms of the supervisory architecture. We
classify as reforms those institutional changes implemented in a country
which involved either the establishment of a new supervisory authority
and/or the changing the powers of at least one of the already existing
agencies.
Cross country comparisons of financial regulation architectures: the existing indicatorsThe literature on the economics of financial supervision architectures has
zoomed in on the following phenomenon: before the crisis, an increasing
number of countries had moved towards a certain degree of consolida-
tion of powers, which in several cases has resulted in the establishment
of unified regulators, different from the national central banks. Various
studies [Barth et al. (2002), Arnone and Gambini (2007), Čihák and Pod-
piera (2007)] claim that the key issues for supervision are (i) whether there
should be one or multiple supervisory authorities and (ii) whether and
how the central bank should be involved in supervision. More important-
ly, these two crucial features of a supervisory regime seem to be related.
The literature tried to undertake a thorough analysis of the supervisory
reforms, measuring these key institutional variables [Masciandaro (2004,
2006, 2007 and 2008), Masciandaro and Quintyn (2009)], i.e. the degree
of consolidation in the actual supervisory regimes, as well as the central
bank involvement in supervision itself.
How can the degree of consolidation of financial regulation be measured?
This is where the financial supervision unification index (FSU Index) be-
comes useful [Masciandaro (2004, 2006, 2007, and 2008)]. This index
was created through an analysis of which, and how many, authorities in
each of the examined countries are empowered to supervise the three
traditional sectors of financial activity: banking, securities markets, and
insurance.
To transform the qualitative information into quantitative indicators, a nu-
merical value has been assigned to each regime, to highlight the number
of the agencies involved. The rationale by which the values are assigned
simply considers the concept of unification (consolidation) of supervisory
powers: the greater the unification, the higher the index value.
The index was built on the following scale: 7 = single authority for all
three sectors (total number of supervisors = 1); 5 = single authority for
the banking sector and securities markets (total number of supervisors
= 2); 3 = single authority for the insurance sector and the securities mar-
kets, or for the insurance sector and the banking sector (total number of
supervisors = 2); 1 = specialized authority for each sector (total number
of supervisors = 3).
A value of 5 was assigned to the single supervisor for the banking sec-
tor and securities markets because of the predominant importance of
banking intermediation and securities markets over insurance in every
national financial industry. It is also interesting to note that, in the group
of integrated supervisory agency countries, there seems to be a higher
degree of integration between banking and securities supervision than
between banking and insurance supervision. Consequently, the degree
of concentration of powers, ceteris paribus, is greater.
This approach does not, however, take into account another qualita-
tive characteristic, namely that there are countries in which one sector
is supervised by more than one authority. It is likely that the degree of
concentration rises when there are two authorities in a given sector, one
of which has other powers in a second sector. On the other hand, the
degree of concentration falls when there are two authorities in a given
sector, neither of which has other powers in a second sector. It would,
therefore, seem advisable to include these aspects in evaluating the vari-
ous national supervisory structures by modifying the index in the follow-
ing manner: adding 1 if there is at least one sector in the country with two
fields of authority, and one of these is also for at least one other sector;
subtracting 1 if there is at least one sector in the country with two au-
thorities assigned to its supervision, but neither of these authorities has
responsibility for another sector; 0 elsewhere.
Now we can consider what role the central bank plays in the various na-
tional supervisory regimes. Here, the index of the central bank’s involve-
ment in financial supervision has been proposed [Masciandaro (2004,
2006, 2007 and 2008), Masciandaro and Quintyn (2009)]: the Central
Bank as the Financial Authority index (CBFA).
For each country, and given the three traditional financial sectors (bank-
ing, securities, and insurance), the CBFA index is equal to: 1 if the central
bank is not assigned the main responsibility for banking supervision; 2 if
the central bank has the main (or sole) responsibility for banking supervi-
sion; 3 if the central bank has responsibility in any two sectors; 4 if the
central bank has responsibility in all three sectors. In evaluating the role
The Capco Institute Journal of Financial TransformationMeasuring Financial Supervision Architectures and the Role of Central Banks
12
of the central bank in banking supervision, we considered the fact that,
whatever the supervision regime, the monetary authority has responsi-
bility in pursuing macro-financial stability. Consequently, we chose the
relative role of the central bank as a rule of thumb: we assign a greater
value (2 instead of 1) if the central bank is the sole or the main authority
responsible for banking supervision.
Measuring consolidation and central bank involvement: new indicatorsWhat are the shortcomings of the previous indices? They has been de-
signed to be consistent with the aim of measuring the degree of consoli-
dation of the supervisory powers using subjective weights in differentiat-
ing some cases, for example, in giving more relevance to the supervision
on both banking and securities industries, or in evaluating other situa-
tions, for example, the degree of consolidation when there are at least
two supervisors in one sector, or when a supervisor is in charge of more
than one sector. Consequently, one type of improvement could be to
reduce the role of the subjective weights.
Starting with the supervisory architectures, we introduce two indicators
to evaluate the two main characteristics highlighted in the literature: the
degree of supervisory consolidation and central bank involvement in su-
pervision.
We propose the Financial Supervision Herfindahl Hirschman (FSHH) in-
dex. The FSHH is a measure of the level of consolidation of the super-
visory powers that we derive by applying to this novel field the classical
index proposed by Herfindahl and Hirschman [Hirschman (1964)]. We
use the FSHH index to calculate the degree of supervisory consolidation.
The robustness of the application of the FSHH to analyze the degree of
concentration of power in financial supervision depends on the following
three crucial hypotheses [Masciandaro et al. (2011)].
First of all, it must be possible to define both the geographical and institu-
tional dimension of each supervisory market so that it would be possible
to define the different sectors to be supervised (institutional dimension)
in each country (geographical dimension). In other words, in every coun-
try each financial market forms a distinct market for supervision. It is still
possible to identify both the geographical dimension, i.e., the existence
of separate nations, and the institutional dimension, i.e., the existence
of separate markets – notwithstanding the fact that the blurring of the
traditional boundaries between banking, securities, and insurance activi-
ties and the formation of large conglomerates diluted the definition of the
intermediaries. Then, in each sector we can define the distribution of the
supervisory powers among different authorities – if more than one agency
is present – and consequently their shares without ambiguity. For each
sector, as the degree of supervision consolidation falls, the greater the
number of authorities involved in monitoring activity.
Secondly, we consider the supervision power as a whole. Given differ-
ent kinds of supervisory activity (banking supervision, securities markets
supervision, insurance supervision) there is perfect substitutability among
them in terms of supervisory power and/or supervisory skills. The supervi-
sory power is a feature of each authority as agency, irrespective of where
this supervisory power is exercised (agency dimension). Consequently, in
each country and for each authority, we can sum the share of the super-
visory power it enjoys in one sector with the share it owns in another one
(if any). For each authority, as the degree of supervisory power increases,
the greater the number of sectors over which that agency exercises moni-
toring responsibility. All three dimensions – geographical, institutional and
agency – have both legal foundations and economic meaning.
Finally, we prefer to adopt the HH index rather than the classic Gini in-
dex in order to emphasize the fact that the overall number of authorities
matters. In general, the use of the HH index rather than other indices of
concentration – such as the entropy index – gives more weight to the
influence of the unified authorities, which is, as we stressed above, the
main feature of the recent evolution in the shape of the supervisory re-
gimes. We calculate the FSHH index by summing up the squares of the
supervisory shares of all the regulators of a country. For each country, the
FSHH index is equal to:
H = ∑ni=1
s2i
where si is the share of supervisory power of the authority i and N is the
total number of authorities. For each authority i, we consider that in each
country there are three sectors to supervise (each sector has the same
importance) and that in each sector we can have more than one authority
(each authority has the same importance). We use the following formula
si = ∑mj=1
sj ; and sj = 1/m · 1/qj
where m is the number of sectors where the authority i is present as su-
pervisor and q is the number of authorities involved in supervision in each
sector j. In other words, if in one sector there is more than one authority,
the supervisory power is equally divided among the supervisors.
We can use the FSHH index to provide a quantitative perspective on
the state of the art of the supervisory regimes. We analyze the situation
before and after the recent crisis, according to country – income and re-
gional adherence. Figure 1 provides this perspective. First, it shows that
before the crisis – 2007, blue bars – the degree of consolidation was on
average greater in the European Union than in the industrial countries as
a whole or the overall European region. These three groups score higher
than the overall country sample. Second, the consolidation process has
progressed in the above three groups of countries during the crisis –
2009, red bars – while the overall sample shows a slight reduction.
13
Our data, therefore, show that also during the crisis the supervisory re-
forms were driven by a general tendency to reduce the number of agen-
cies to reach the unified model, unknown before 1986, or the so-called
peak model, which characterized the two decades between 1986-2006
[Masciandaro and Quintyn (2009)].
Now, the new methodology can also be used to construct the index of
the central bank involvement in supervision: the Central Bank as Finan-
cial Supervisor (CBFS) index. The intuition is quite simple: the greater the
share of the central bank, the higher the odds that the central bank will be
involved into the overall regulatory organization. In other words, central
bank involvement in supervision is likely to be at its maximum where the
central banker is the unified supervisor in charge, while the involvement
is likely to be low, the smaller the number of sectors where the central
bank has supervisory responsibilities. To construct the CBFS index we
just need to take the share of the central bank in each country, which can
go from 0 to 1.
Again we can use the CBSS index to offer a numerical description of the
degree of central bank involvement in supervision, before and after the
crisis. Two facts emerge. Figure 2 shows that before the crisis – 2007,
yellow bars – the industrial countries have on average a lower level of
involvement and that the European countries as well as the E.U. member
states experience even less central bank involvement in supervision than
the overall sample.
However, in response to the crisis we note a sort of “Great Reversal”: the
2009 data – green bars – show that in the industrialized, European, and
E.U. countries central bank involvement increased. The new trend can be
explained using at least two different reasons.
First of all, in some countries the central bank can be more involved in
supervision because the monetary responsibilities are not completely in
their hands. Among the central banks which do not have full responsibil-
ity for monetary policy, such as those of the countries belonging to the
European Monetary Union, some countries chose the route of the central
bank specialization in supervision, with the following being examples of
those countries in which this is strongest: Czech Republic, Ireland, Neth-
erlands, and the Slovak Republic. In general, it has been noted [(Herring
and Carmassi (2008)] that the central banks of members of the EMU have
become financial stability agencies.
Second, the experience of recent years has stressed the importance of
overseeing systemic risks in the system. In other words, it is crucial to
monitor and assess the threats to financial stability that can arise from
macroeconomic as well as macro-financial developments (the so-called
macro-supervision). The increasing emphasis on macro-supervision
motivates the policymakers to identify specific bodies responsible for
macro-supervision.
To carry out macro-prudential tasks information on the economic and
financial system as a whole is required. The current turmoil has stressed
the role of the central banks in the prevention, management, and resolu-
tion of financial crises. Consequently, this view is gaining momentum that
the central banks are in the best position to collect and analyze this kind
of information, given their role in managing in normal times the monetary
policy and in exceptional times the lender of last resort function.
From the policymakers’ point of view, therefore, the involvement of the
central bank in the macro-supervision area means potential benefits in
terms of information gathering. At the same time, they can postulate that
The Capco Institute Journal of Financial TransformationMeasuring Financial Supervision Architectures and the Role of Central Banks
0
10
20
30
40
50
60
70
80
ALL OECD EUROPE E.U.
FSHH index
Countries
Figure 1 – Financial supervision unification
0
5
10
15
20
25
30
35
ALL OECD EUROPE E.U.
CBBSS
Countries
Figure 2 – Central bank involvement in supervision
14
the potential costs of the involvement are smaller with respect to the case
of micro-supervision [moral hazard risk, conflict of interest risk, powerful
bureaucracy risk; see Masciandaro (2008)]. In other words, the separa-
tion between micro- and macro-supervision can be used to reduce the
arguments against the central bank involvement.
ConclusionThe wave of reforms in supervisory architectures across the globe that
we have witnessed since the end of the 1990s leaves the interested
bystander with a great number of questions regarding the key features
of the emerging structure, their true determinants, and their effects on
the performance of banking and financial industries. This paper reviews
the changing face of the financial supervisory regimes, introducing new
indicators to measure the level of consolidation in supervision and the
degree of the involvement of the central banks. We show that the new Fi-
nancial Supervision Herfindahl Hirschman indexes are (i) consistent with
the previous one, (ii) but at the same time more precise, (iii) more robust,
given that they exclude subjective weights and (iv) more easy to use and
interpret, given that they apply a well-known methodology for measure-
ment. The new indices can be used in empirical studies on the determi-
nants of the reform process and on the impact of the new architectures
of financial supervision.
ReferencesAllen, F. and Carletti, E., 2009, “The global financial crisis: causes and consequences,” Mimeo•
Arnone, M. and A. Gambini, 2007, “Architecture of supervisory authorities and banking •
supervision,” in: Masciandaro D. and M. Quintyn (eds.), Designing financial supervision
institutions: independence, accountability and governance, Edward Elgar
Barth, J. R., G. Caprio, and R. Levine, 2006, Rethinking bank regulation. Till angels govern, •
Cambridge University Press
Barth, J. R., D. E. Nolle, T. Phumiwasana, and G. Yago, 2002, “A cross country analysis of •
the bank supervisory framework and bank performance,” Financial Markets, Institutions &
Instruments, 12:2, 67-120
Borio C., 2003. “Towards a macroprudential framework for financial regulation and •
supervision?” BIS Working Papers, no.128, Bank of International Settlements, Geneva
Brown, E. F, 2005, “E Pluribus Unum – out of many, one: why the United States need a single •
financial services agency,” American Law and Economics Association Annual Meeting
Brunnermeier, M. et al., 2009, “The fundamental principles of financial regulation”, Geneva •
Reports on the, World Economy number 11
Buiter, W., 2008, “Lessons from the North Atlantic financial Crisis,” paper presented at the •
conference “The role of money markets,” Columbia Business School and Federal Reserve of
New York, May 29-30
CEPS, 2008, “Concrete steps towards more integrated financial oversight,” CEPS Task Force •
Report, December
Č• ihák, M. and R. Podpiera, 2007, Experience with integrated supervisors: governance
and quality of supervision, in: Masciandaro, D. and M. Quintyn (eds.), Designing financial
supervision institutions: independence, accountability and governance, Edward Elgar
Coffee, J., 1995. Competition versus consolidation: the significance of organizational structure •
in financial and securities regulation,” Business Lawyer, vol 50
Commission of the European Communities, 2009, “Communication from the Commission. •
European financial supervision,” COM (2009) 252 final
Council of the European Union, 2009, “Council conclusions on strengthening EU financial •
supervision,” 2948th Economic and Financial Affairs, Luxembourg, June 9
De Larosière Group, 2009, Report of the high level group on supervision•
Department of the Treasury, 2008, Blueprint for a modernized financial regulatory structure, •
Washington, D.C.
Department of the Treasury, 2009, Financial regulatory reform: a new foundation, Washington, •
D.C.
Financial Services Authority, 2009, The Turner review, March, London•
Kremers J., D. Schoenmaker, and P. Wierts, 2003, “Cross – sector supervision: which model?” •
in Herring, R., and R. Litan (eds), Brookings-Wharton Papers on Financial Service
Goodhart, C. A. E., 2007, Introduction, in Masciandaro D. and M. Quintyn (eds.), Designing •
financial supervision institutions: independence, accountability and governance, Edward Elgar
Hardy, D., 2009, “A European mandate for financial sector supervisors in the EU,” IMF Working •
Paper WP/09/5
Herring R. J and J. Carmassi, 2008, “The structure of cross–sector financial supervision,” •
Financial Markets, Institutions and Instruments, 17:1, 51-76
Hirschman, A. O., 1964, “The paternity of an index,” American Economic Review, 54:5, 761-762•
House of the Lords, 2009, “The future of EU of financial regulation and supervision,” European •
Union Committee, 14th report of session 2008-2009, HL paper 106-I, Authority of the House of
the Lords
Leijonhufvud, A., 2009, “Curbing instability: policy and regulation,” CEPR Policy Insight, no. 36, •
July
Masciandaro, D., 2004, “Unification in financial sector supervision: the trade off between central •
bank and single authority,” Journal of Financial Regulation and Compliance, 12:2, 151-169
Masciandaro, D., 2006, “E Pluribus Unum? Authorities design in financial supervision: trends •
and determinants,” Open Economies Review, 17:1, 73-102
Masciandaro, D., 2007, “Divide et Impera: financial supervision unification and the central bank •
fragmentation effect,” European Journal of Political Economy, 23:2, 285-315
Masciandaro, D, 2008, “Politicians and financial supervision unification outside the central •
bank: why do they do it?” Journal of Financial Stability, 5:2, 124-147
Masciandaro, D. and M. Quintyn, 2009, Reforming financial supervision and the role of the •
central banks: a review of global trends, causes and effects (1998-2008), CEPR Policy Insight,
n.30, 1-11
Masciandaro, D., R. V. Pansini, and M. Quintyn, 2011, “Economic crisis: do financial supervision •
matter?” Paper presented at 29th SUERF Colloquium, Brussels.
Orphanides A., 2010, “Monetary policy lessons from the crisis,” Discussion Paper Series, •
CEPR, no. 7891
Scott, K., 1977, “The dual banking system: a model of competition in regulation,” Stanford Law •
Review, 30:1, pp. 1–49
World Bank and IMF, 2005, Financial sector assessment: a handbook, World Bank and IMF, •
Washington D.C.
15
PART 1
What Reforms for the Credit Rating Industry? A European Perspective
AbstractCredit rating agencies were the first victim of the crisis, with
a regulation adopted in a period of six months – a record by
E.U. standards. The regulation subjects E.U.-based CRAs to
a mandatory license and strict conduct of business rules,
whereas, unlike in the U.S., no rules had been in place before.
This article discusses the role of credit ratings agents today,
the regulatory role of ratings, the scope of the E.U. regula-
tion, and the regulatory approach of the business models
of the large ratings agents. It concludes that the regulation
should have impacted the business model of ratings agents
more fundamentally.
Karel Lannoo — Chief Executive Officer, CEPS, and Director, ECMI1
This paper was initially prepared for the Competition Committee of the 1
OECD, meeting in Paris, 16 June 2010. Comments from participants at that
meeting are gratefully acknowledged, as well as from Piero Cinquegrana,
Chris Lake, Barbara Matthews, and Diego Valiante.
16
Credit rating agencies (CRAs) continue to find themselves in the eye of
the storm. Despite having singled out the industry early on in the financial
crisis as needing more regulation, policymakers seem not to be reas-
sured by the measures that have been adopted in the meantime, and
want to go further. Faced with a rapid downgrading in ratings in the con-
text of the sovereign debt crisis, European Commissioner Michel Barnier
raised the possibility last May of creating a new E.U.-level agency that
would specialize in sovereign debt and regulating CRAs.
The debate on the role of rating agents considerably pre-dates this crisis.
As early as the 1997 South-East Asia crisis, the delayed reaction of rat-
ing agents to the public finance situation of these countries was strongly
criticized. The same criticism of CRAs was leveled in the dot.com bubble
in 2001. Many reports were written on their role in that episode, but it was
not until mid-2008 that a consensus emerged in the E.U. that the industry
was in need of statutory legislation. In the meantime, the U.S. had ad-
opted the Credit Rating Agency Reform Act in 2006. At the global level, in
2003, the International Organisation of Securities Commissions (IOSCO)
adopted a Statement of Principles on the role of credit rating agencies,
but apparently the initiative has not been successful.
Rating agents pose a multitude of regulatory problems, none of which
can be solved easily. Some of these are specific to the profession and the
current market structure, whereas others are of a more generic nature.
Some are related to basic principles of conduct in the financial services
sector, while others are part of horizontal market regulation. The financial
crisis also demonstrated the important role of rating agents in financial
stability, which involves macro-prudential authorities.
The credit ratings industry todayThe credit ratings industry is a global business, controlled by a handful of
players, two of which are of U.S. parentage. Moody’s and Standard & Poor’s
alone possess more than four-fifths of the market. With Fitch, the three lead-
ing players dominate over 94 percent of the global market [European Com-
mission (2008)]. See brief portraits of these three companies in Box 1.
As shown in Table 1, the three groups have suffered serious declines in rev-
enue since 2007, especially Fitch. Its revenues have declined by 26 percent
since 2007, and its net income by 70 percent. This may confirm the finding
discussed below that more competition does not necessarily improve the
quality, but that newcomers, in this case Fitch, attempt to attract market
share with a short-term strategy. Firms may also have abandoned ratings,
which cost between €45,000 and €90,000 per annum, plus 0.05 percent
of total value of a bond emission. Table 1 further indicates that the market
share of the three firms has been fairly constant over the period 2006-09.
That the credit rating business is essentially of American parentage
should be no surprise, as it is an intrinsic part of the market-driven system
pioneered by the U.S. Unlike the bank-driven model, which is common
in Europe, a market-driven system relies upon a multi-layered system to
make it work [Black (2001)]. Reputational intermediaries, such as invest-
ment banks, institutional investors, law firms, and rating agents, and self-
regulatory organizations, such as professional federations and standard-
setters, play an important role to make the system, in between issuers
and supervisors, work. In effect, financial markets are constantly affected
by adverse selection mechanisms, and investors need third-party tools
such as credit ratings in order to reduce asymmetric information and in-
crease their ability to understand the real risk of financial instruments.
Since there had not been much of a capital market in Europe until re-
cently, banks have essentially performed the credit-risk analysis function,
and continue to do so. But the credit-risk analysis function of European
banks declined, possibly as a result of the reputational strength of the
U.S. capital market model. The introduction of the euro and a set of E.U.
regulatory measures led to the rapid development of European capital
markets, and demand for ratings. Moreover, European authorities cre-
ated a captive market for an essentially U.S.-based industry.
Moody’s investor services was incorporated in 1914 as a bond rating
and investment analysis company. Today, the listed company Moody’s
Corporation is the parent company of Moody’s Investors Service, which
provides credit ratings and research covering debt instruments and
securities, and Moody’s Analytics, which encompasses non-ratings
businesses, including risk management software for financial institutions,
quantitative credit analysis tools, economic research and data services,
data and analytical tools for the structured finance market, and training
and other professional services. Combined, they employ about 4,000
people.
Standard & Poor’s was incorporated in 1941, following the merger of
two firms active in credit risk analysis. Both firms originated in similar cir-
cumstances as Moody’s, in the context of the huge industrial expansion
of the U.S. in the second half of the 19th and early 20th centuries. S&P
was taken over by McGraw Hill in 1966, the listed media concern, and
it forms the most important part of the group in terms of revenues, and
even more so in profits (about 73 percent), although these have seriously
declined since 2007. S&P financial services, which includes the ratings
service, employs about 7,500 people.
Fitch Ratings – by far the smaller “European” player in the sector with
headquarters in New York and London – is part of the Fitch Group. The
Fitch Group also includes Fitch Solutions, a distribution channel for Fitch
Ratings products, and Algorithmics, a leading provider of enterprise risk
management solutions. The Fitch Group has been a majority-owned
subsidiary (60 percent) of Fimalac S.A. since 1997, which has headquar-
ters in Paris, and is listed on Euronext, but with a very low free float. Fitch
grew through acquisitions of several smaller ratings agents, including
IBCA and Duff & Phelps. Fitch employs 2,266 people.
Box 1 – The “big three”
17
A captive market for CRAs in the E.U.Two forms of “regulation” have given the CRAs a captive market in the
E.U.: Basel II,2 implemented in Europe as the Capital Requirements Direc-
tive (CRD), and the liquidity-providing operations of the European Central
Bank (ECB). Both explicitly use the rating structure of the CRAs to deter-
mine risk weighting for capital requirement purposes – thresholds in the
former case and “haircuts”3 for the ECB’s liquidity-providing operations.
The U.S. does not use either of these practices, as it has not implemented
Basel II (largely because because the Federal Reserve did not want to have
the vast majority of U.S. banks relying on CRAs for setting regulatory risk
weights), and the discount window of the Fed is not based upon ratings.
The Dodd-Frank Wall Street Reform and Consumer Protection Act of July
20104 goes even further, requiring regulators to remove any references
from their rules to “investment grade” and “credit ratings” of securities.5
The Basel II proposals were finalized in November 2005 after lengthy
discussions, among other things, because of its pro-cyclical impact and
the use of private sector rating agents. In its “standardized approach,” to
be used by less sophisticated banks, it bases risk weightings on rating
agents’ assessments. The capital requirements increase with the decline
in the rating, from 0 percent for AA-rated (and higher) government bonds,
or a minimum of 20 percent for banks and corporates up to 150 percent for
CCC or below. However, in the E.U.’s CRD, the risk weighting is 0 percent
across the board for all sovereigns in the European Economic Area (EEA)
funded in domestic currency. A zero-risk weighting means that a bank does
not have to set any capital aside for these assets. No indication has been
given so far that the reliance on rating agents for the risk weightings will be
changed in the Basel III proposals, published on 12 September 2010.
Since CRAs were not subject to E.U. regulation at the time the CRD was
adopted, the Committee of European Banking Supervisors (CEBS) issued
“Guidelines on the recognition of external credit assessment institutions”
in January 2006. These guidelines set criteria for “determining” external
credit assessments on the basis of the CRD risk weights. The use of a
rating agent for the purposes of the CRD is thus the prerogative of the
national supervisory authorities. For comparison, the Japanese FSA has
designated five rating firms as qualified to calculate risk weights for the
standardized approach: the “big three” and two smaller Japanese firms.
The use of rating agents is possibly even more prevalent in the assess-
ment of marketable assets used as collateral in the ECB’s liquidity-pro-
viding operations. The credit assessment for eligible collateral is pre-
dominantly based on a public rating, issued by an eligible External Credit
Assessment Institution (ECAI). In the ECB’s definition, an ECAI is an in-
stitution whose credit assessments may be used by credit institutions
for determining the risk weight of exposures according to the CRD.6 The
minimum credit quality threshold is defined in terms of a “single A” credit
assessment,7 which was temporarily relaxed during the financial crisis to
BBB-. If multiple and possibly conflicting ECAI assessments exist for the
same issuer/debtor or guarantor, the first-best rule (i.e., the best available
ECAI credit assessment) is applied.8
The liquidity categories for marketable assets are subdivided into five
categories, based on issuer classification and asset type, with an in-
creasing level of valuation haircuts, depending on the residual maturity.9
An important group of assets in the context of the financial crisis, classi-
fied as “category V”, are the asset-backed securities (ABS), or securitiza-
tion instruments. The extent to which banks used ABS collateral in liquid-
ity operations rose dramatically after mid-2007, from 4 percent in 2004
to 18 percent in 2007 and 28 percent in 2008 [Fitch (2010)]. Within ABS,
residential mortgage-backed securities (RMBS) form the most important
element, exceeding 50 percent. These securitization instruments, and in
particular the residential mortgage-backed securities segment, were an
extremely important market for CRAs. Moody’s, for example, assigned
the AAA rating to 42,625 RMBS from 2000 to 2007 (9,029 mortgage-
backed securities in 2006 alone), and later had to downgrade the assets.
In 2007, 89 percent of those originally rated as investment grade were
reduced to junk status.10
The Capco Institute Journal of Financial TransformationWhat Reforms for the Credit Rating Industry? A European Perspective
The second set of the recommendations issued in June 2004 by the Basel Committee on 2
Banking Supervision, Basel II creates an international standard that banking regulators may
use in establishing regulations governing how much capital banks must set aside to counter
the financial and operational risks they face.
A deduction in the market value of securities being held by brokerage and investment 3
banking firms as part of net worth for calculating their net capital.
Public Law 111 – 203 – Dodd-Frank Wall Street Reform and Consumer Protection Act 4
(http://www.gpo.gov/fdsys/pkg/PLAW-111publ203/pdf/PLAW-111publ203.pdf)
Clifford Chance (2010), p. 73.5
ECB (2006), The implementation of monetary policy in the Euro Area,” General documen-6
tation on Eurosystem monetary policy instruments and procedures, September, p. 43.
“Single A” means a minimum long-term rating of “A-” by Fitch or Standard & Poor’s, or a 7
“A3” rating by Moody’s [ECB (2006)].
ECB (2008), “The implementation of monetary policy in the Euro Area,” General documentation 8
on Eurosystem monetary policy instruments and procedures, November, p. 42.
The liquidity categories were changed in September 2008 and the valuation haircuts 9
increased in July 2010. See latest changes to risk control measures in Eurosystem credit
operations, European Central Bank, Press notices, 4 September, 2008 and 28 July, 2010.
According to Phil Angelides, Chairman of the ten-member Financial Crisis Inquiry 10
Commission appointed by the U.S. government to investigate the causes of the financial
crisis, quoted in Bloomberg, 2 June 2010.
2006 2007 2008 2009 Δ 07-09
Moody’s Turnover 2037.1 2259 1775.4 1797.2 -20.4
Net income 753.9 701 461.6 407.1 -41.9
S&P’s Turnover 2750 3046.2 2653.3 2610 -14.3
Net income n.a. 440.16 327.8 307.4 -30.2
Fitch Turnover 655.6 827.4 731.2 613.5 -25.9
Net income n.a. 120.2 44 35.8 -70.2
Sources: 10-K filings to the U.S. SEC by Moody’s and McGraw-Hill, other filings by Fimalac
and Hoover, S&P’s and Fitch’s website.
Table 1 – Turnover and net income of the “big three” ratings businesses, 2006-09 ($ millions)
18
The E.U. rating agencies regulationAs the financial crisis erupted, the developments recounted above and
others rapidly led to a policy consensus that rating agents should be
regulated at E.U. level. The proposal for a regulation was published in No-
vember 2008, and adopted in April 2009, a minimum interval in E.U. de-
cision-making.11 The regulation was the first new E.U. legislative measure
triggered by the financial crisis. It is also one of the first financial services
measures to be issued as a regulation, meaning it is directly applicable,
rather than a directive, which has to be implemented in national law.
The E.U. was not starting from scratch. Back in 2004, further to an own
initiative report of the European Parliament (Katifioris report), the Europe-
an Commission asked the Committee of European Securities Regulators
(CESR) for technical advice regarding market practice and competitive
problems in the CRAs. In a Communication published in December 2005,
it decided that no legislation was needed for three reasons: 1) three E.U.
directives already cover ratings agents indirectly: the market abuse Direc-
tive, the CRD, and MiFID; 2) the 2004 Code of Conduct Fundamentals for
Credit Rating Agencies,12 published by the IOSCO; and 3) self-regulation
by the sector, following the IOSCO Code.13
In 2006, in a report for the Commission, the CESR concluded that the
rating agents largely complied with the IOSCO Code.14 But concerns re-
mained regarding the oligopoly in the sector, the treatment of confidential
information, the role of ancillary services, and unsolicited ratings. In a fol-
low-up report published in May 2008, focusing especially on structured
finance, the CESR strongly recommended following the international
market-driven approach by improving the IOSCO Code. Tighter regula-
tion would not have prevented the problems emerging from the loans to
the U.S. subprime housing market, according to the CESR.
Notwithstanding the CESR’s advice, the Commission went ahead and
issued a proposal in November 2008, after two consultations in July and
September 2008. The E.U. regulation requires CRAs to be registered and
subjects them to ongoing supervision; defines the business of the issuing
of credit ratings; sets tight governance (board structure and outsourc-
ing), operational (employee independence and rotation, compensation,
prohibition of insider trading, record keeping) and conduct of business
(prohibition of conflicts of interest in the exercise of ratings or through the
provision of ancillary services to the rated entity) rules for CRAs; requires
CRAs to disclose potential conflicts of interest and its largest client base;
and requires CRAs to disclose their methodologies, models, and rating
assumptions. CESR is mandated to set standards for methodologies and
establish a central repository with the historical performance data.
The regulation came into force 20 days after its publication in the Of-
ficial Journal, on 7 December, 2009. But guidance had to be provided by
CESR before the regulation could take effect, by 7 June, 2010, regarding
registration, supervision, the endorsement regime, and supervisory re-
porting; and by 7 September 2010, regarding enforcement practices, rat-
ing methodologies, and certification. CESR has to report annually on the
application.
The novelty in the regulation is the central role of CESR in providing ad-
vice regarding the requirement for registration by a CRA in an E.U. mem-
ber state, and in informing all the other member states. The home and
host member states to the CRA are required to establish a college and
are required to cooperate in the examination of the application and in
day-to-day supervision. Host member states are not only those where
a CRA has a branch, they are also those where the use of credit ratings
is widespread or has a significant impact. In these circumstances, the
host country authority may at any time request to become a member
of the college (Art. 29.3). Host countries can also act against an agency
deemed to be in breach of its obligations (Art. 25). CESR has the author-
ity to mediate between the competent authorities (Art. 31), which had the
effect of pre-empting its transformation into a securities market authority
under the proposals discussed as further to the de Larosière report.15
As the industry is essentially of U.S. parentage, a focal point in the dis-
cussions was the third country regime. The regulation states that CRAs
established in a third country may apply for certification, provided that
they are registered and subject to supervision in their home country, and
that the Commission has adopted an equivalence decision. However,
credit ratings issued in a third country can only be used if they are not
of systemic importance to the E.U.’s financial stability (Art. 5.1), meaning
that all large CRAs need to be fully registered in the E.U. system. In ad-
dition, credit ratings produced outside the E.U. have to be endorsed by
the CRA registered in the E.U., subject to a series of conditions (Art. 4.3).
It has been argued that this regime will unnecessarily fragment global
capital markets. Foreign companies will be less inclined to raise capital
in the E.U. as they need a local endorsement of their rating. E.U. financial
institutions will invest less abroad, as the ratings on third country invest-
ments may be seen to be of insufficient quality, unless they are endorsed
in the E.U., or their rating agents are equivalent. The regime could also be
qualified as anti-competitive, as smaller CRA without an E.U. presence,
such as the two largest CRAs in Asia, may stop rating E.U. sovereigns
and issuers. Establishing a local presence in the E.U. could be too costly,
and the client base for these ratings would as a result diminish, since they
can no longer be used by European banks [St. Charles (2010)].
Regulation 1060/2009 of 16 September 2009, OJ 17.11.2009.11
See http://www.iosco.org/library/pubdocs/pdf/IOSCOPD180.pdf.12
Communication from the Commission on Credit Rating Agencies (2006/C 59/02), OJ C 59/2 13
of 11.03.2006.
CESR’s Report to the European Commission on the compliance of rredit rating agencies 14
with the IOSCO Code, CERS, 06-545.
Report of the High-Level Group on Financial Supervision in the E.U., chaired by Jacques de 15
Larosière, 25 February, 2009, Brussels.
19
hazard, it gives a regulatory “blessing” and will further reduce the in-
centives for banks to conduct proper risk assessments. It creates the
illusion that the industry will live up to the new rules, and that these will
adequately supervised.
For Pagano and Volpin (2009), the preferred policy is more drastic: 1) rat-
ings should be paid for by investors, and 2) investors and ratings agen-
cies should be given free and complete access to all information about
the portfolios underlying structured debt securities. The investor-pays
principle was the rule in the U.S. until the 1970s, but because of increas-
ingly complex securities in need of large resources and the fear of declin-
ing revenues resulting from the dissemination of private ratings through
new information technologies, the issuer-pays principle was introduced.
Pagano and Volpin do not discuss, however, how to deal with free riding.
But moving back to the investor-pays principle may also require further
regulation to prohibit the sale of ancillary services by CRAs to issuers.
The E.U. regulation goes in the direction of requiring more disclosure (see
Annex I, Section E of the regulation), but it is questionable whether inves-
tors will read this. On the contrary, given that a supervisory fiat has been
given, investors may be even less inclined to read all the information, as
was demonstrated during the financial crisis.
Making investors pay would bring the ratings agents closer to the profes-
sion of analysts and investment advisors, which is regulated under the
E.U.’s Market in Financial Instruments Directive (2004/39). MiFID requires
investment advisors to be licensed, to act in the best interests of their cli-
ents, and to identify, disclose, and avoid conflicts of interest. MiFID also
states that firewalls must be constructed between analysts and sales de-
partments in banks.
Ponce (2009) discusses an interesting alternative to the issuer-pays and
investor-pays models: the platform-pays model. He demonstrates on the
basis of large datasets that the transition from the investor-pays to the
issuer-pays model had a negative impact on the quality of the ratings.
Under the issuer-pays model, a rating agency may choose a quality stan-
dard below the socially efficient level. In this case, Ponce argues, a rating
agency does not internalize the losses that investors bear from investing
in low-quality securities. A rating agent may give ratings to low-quality
securities in order to increase its revenues. To avoid this, Ponce proposes
the “platform-pays” model, which takes the form of a clearing house for
ratings, complemented by prudential oversight of ratings’ quality to con-
trol for bribery. The platform assigns the agent (based on performance
and experience) and the issuer pays up front. This would at the same
time overcome the oligopoly problem. The problem with this model, how-
ever, is that its governance wills need to be completely watertight.
The amendments tabled by the Commission on 2 June, 2010 modify the
regulation to accommodate the imminent creation of the European Se-
curities Market Authority (ESMA), and to further centralize the supervi-
sion of CRAs.16 ESMA would become the sole supervisor, for the sake
of efficiency and consistency, doing away with the complex system de-
scribed above. National supervisors will remain responsible, however, for
the supervision of the use of credit ratings by financial institutions, and
can request ESMA to withdraw a license. ESMA can ask the European
Commission to impose fines for non-respect of provisions of the regula-
tions (see Annex III of the proposal). ESMA may also delegate specific su-
pervisory tasks to national authorities. The proposal does not, however,
propose any involvement of the European Systemic Risk Board (ESRB),
which could have been useful in the control of the methodologies and
the macroeconomic models used by CRAs. The draft regulation finally
requires issuers of structured finance instruments to disclose the same
information which they have given to the CRA, as is the case under the
U.S. SEC’s Rule 17g-5. This change was welcomed by the markets as it
would make both regimes convergent.
The regulatory debateThe E.U.’s regulation does not alter the fundamental problem that CRAs
pose from a public policy perspective: 1) the oligopolistic nature of the
industry, 2) the potential conflict of interest through the issuer-pays prin-
ciple and, 3) the public good of the private rating. The E.U. approach
seems to be a second-best solution. A more fundamental review is need-
ed of the business model of the CRAs, and for which other industry sec-
tors could provide useful alternative models.
On the structure of the industry, the E.U. increases the barriers to entry by
introducing a license and setting tight regulation, rather than taking the
oligopolistic nature as one of the fundamental reasons for the abuses. In
addition, since statutory supervision of the industry may increase moral
The Capco Institute Journal of Financial TransformationWhat Reforms for the Credit Rating Industry? A European Perspective
The new E.U. regime for CRAs is very closely aligned with the new
U.S. regime, as introduced by the Dodd-Frank Bill. Whereas the U.S.
had already legislated the sector in 2006 with the Credit Rating Agency
Reform Act, this was a light regime that required CRAs to register with
the Securities and Exchange Commission (SEC) in Washington, D.C.,
as a Nationally Recognized Statistical Rating Organization (NRSRO).
The Dodd-Frank Bill fundamentally alters this regime by requiring tight
operational (internal controls, conflicts of interest, qualification standards
for credit rating analysts) and governance requirements, and detailed
disclosure requirements (including disclosure of the methodologies used).
The SEC is required to create an Office of Credit Ratings to implement
the measures of the Bill, to issue penalties and to conduct annual
examinations and reports.
Source: Clifford Chance (2010)
Box 2 – The Dodd-Frank Bill and CRAs
Proposal for a regulation of the European Parliament and of the Council on amending 16
regulation (EC) No 1060/2009 on credit rating agencies, COM(2010) 289/3.
20
Other research fi nds that more competition would not necessarily im-
prove standards, however. New entrants do not necessarily improve the
quality of ratings. On the contrary, they attract business by friendly and
infl ated ratings. As competition reduces future rents, it increases the risk
of the short-term gains by cheating. In an analysis of the corporate bond
markets, Becker and Milbourn (2009) fi nd a signifi cant positive correla-
tion between the degree of competition and the level of the credit ratings
(Figure 1). Concretely, they fi nd a positive correlation between Fitch’s en-
trance in the market and ratings levels, without exception.
Considering that incentives and reputational mechanisms are key, Larry
Harris (2010) proposes an entirely different approach. He takes his in-
spiration from the bonus debate in the banking sector, and proposes to
defer a part of the payment based upon results. Given that credit ratings
are about the future, the performance of the securities rated would be the
indicator for the fees rating agents can charge. An important part of the
fees would be put into a fund, against which the ratings agencies could
borrow to fi nance their operations. Disclosure of these deferred contin-
gent compensation schemes would be required, so that investors can
decide for themselves which schemes provide adequate incentives.
Another possibility for creating the right incentives is to move to a part-
nership structure in the rating business, as is common in the audit sector.
The audit sector has several similarities with rating agencies: in the type
of work, the importance of reputation and global presence, the network
economies and the oligopolistic structure, and the confl icts of interest.
The audit sector is regulated by an E.U. Directive (2006/43/EC) that
brought the sector under statutory supervision. It sets tight rules on gov-
ernance and quality control, and limits the degree of non-audit services
that audit fi rms can perform for an audit client. This directive also has an
important third country equivalence regime. It is interesting to note in this
context that at least two audit fi rms have recently expressed an interest
in starting a rating business. The downside of the partnership model is
the liability problem, however, which may deter many from being active
in that business.
During the sovereign debt crisis, European and national policymakers
have repeatedly raised the possibility of “creating” local CRAs, eventually
even government-sponsored entities. A state-controlled CRA would lack
independence, and hence credibility, and, as demonstrated above, it is
not necessarily more competition that will solve the problem.
ConclusionConsidering the policy alternatives outlined above, the E.U. and the U.S.
should probably have considered the specifi cities of the sector more
carefully before embarking upon legislation. The legislation that was ad-
opted does not alter the business model of the industry and gives rise to
side effects, the most important of which is the supervisory seal. Given
the depth of the fi nancial crisis and the central role played by ratings
agents, certainly in the E.U., a more profound change would be useful,
towards the “platform-pays” model or a long-term incentive structure, as
discussed above.
The E.U. regulation, as adopted, consolidates the regulatory role of CRAs
in the E.U. system, but the price is high. It fragments global capital mar-
kets, as it introduces a heavy equivalence process, and requires a local
presence of CRAs and endorsement of systemically important ratings.
It is at the same time protectionist. Under the new set up, CESR and its
successor, ESMA, are given a central role in the supervision of CRAs, but
the question is whether they will be able to cope. The supervisor needs
to check compliance with the basic requirements to decide on a license
and to verify adherence to the governance, operational, methodological,
and disclosure requirements imposed upon CRAs. This is a heavy work-
load, especially considering that no supervision had been in place until a
few months ago. Given the present debate on the role of CRAs in fi nan-
cial stability and the need for technical expertise, the European Systemic
Risk Board could have been involved, but this seems not to have been
considered, at least for now.
On the other hand, the advantage of having a regulatory framework in
place is that the Commission’s competition directorate can start scruti-
nizing the sector from its perspective. To our knowledge, the competition
policy dimensions of the CRA industry in Europe have not been closely
investigated so far, as no commonly agreed defi nitions and tools were
available at E.U. level, and since the sector is essentially of U.S. parent-
age. E.U. registration for the large CRAs will allow the authorities to check
their compliance with E.U. Treaty rules on concerted practices and abuse
of dominant position. This may raise some feathers.
12%
10%
8%
6%
4%
4%
2%
0%C CC CCC- CCC CCC+ B- B B+ BB- BB BB+ BBB- BBB BBB+ A- A A+ AA- AA AA+ AAA
Low competition
High competition
Source: Becker & Milbourn (2009).
Figure 1 – Firm credit ratings distribution: high and low competition in the industry
21
The Capco Institute Journal of Financial TransformationWhat Reforms for the Credit Rating Industry? A European Perspective
ReferencesArbak, E. and P. Cinquegrana, 2008, “Report of the CEPS-ECMI joint workshop on the reform •
of credit rating agencies,” November (available at www.eurocapitalmarkets.org)
Becker, Bo. and T. Milbourn, 2009, “Reputation and competition: Evidence from the credit •
rating industry,” Harvard Business School, Working Paper No. 09-051, Cambridge, MA
Black, B. S., 2001, “The legal and institutional preconditions for strong securities markets,” •
UCLA Law Review, 48, 781-855
Cinquegrana, P., 2009, “The reform of the credit rating agencies: a comparative perspective,” •
ECMI Policy Brief, February (www.ceps.eu and www.eurocapitalmarkets.org)
Clifford Chance, 2010, “Dodd-Frank Wall Street Reform and Consumer Protection Act,” Client •
Briefing Report, July
Coffee, J., 2010, “Ratings reform: a policy primer on proposed, pending and possible credit •
rating reforms,” paper presented at a meeting of the OECD Competition Committee, 16 June,
Paris
ECB, 2006, “The implementation of monetary policy in the Euro Area,” General documentation •
on Eurosystem monetary policy instruments and procedures, September, Frankfurt
ECB, 2008, “The implementation of monetary policy in the Euro Area,” General documentation •
on Eurosystem monetary policy instruments and procedures, November, Frankfurt
European Commission, 2008, “Commission staff working document accompanying the •
proposal for a regulation of the European Parliament and of the Council on Credit Rating
Agencies – impact assessment,” SEC/2008/2746 final, November
European Commission, 2010, Proposal for a regulation of the European Parliament and the •
Council on amending Regulation (EC) No 1060/2009 on credit rating agencies,” June
Fitch, 2010, “The role of the ECB –temporary prop or structural underpinning?” Special Report, •
11 May
Harris, L., 2010, “Pay the rating agencies according to results,” Financial Times, 4 June•
Pagano, M. and P. Volpin, 2009, “Credit ratings failures: causes and policy options”, in •
Dewatripont, M., X. Freixas, and R. Portes, (eds.), Macroeconomic stability and financial
regulation: key issues for the G20, VoxEU, 2 March (http://www.voxeu.org/index.
php?q=node/3167)
Ponce, J., 2009, “The quality of credit ratings: a two-sided market perspective,” August •
(http://www.bcu.gub.uy/autoriza/peiees/jor/2009/iees03j3471009.pdf)
Richardson, M. and L. J. White, 2009, “The rating agencies: is regulation the answer?”, in •
Acharya, V. V. and M. Richardson, (eds.), Restoring financial stability, New York: Wiley
St. Charles, C., 2010, “Regulatory imperialism: the worldwide export of European regulatory •
principles on credit rating agencies,” Minnesota Journal of International Law, 19:1, 177-200.
23
PART 1
Money Market Funds and Financial Stability: Comparing Sweden to the U.S. and Iceland
AbstractThe financial crisis, in particular the collapse of Lehman
Brothers, has revealed that money market funds are more
risky than had previously been believed. We discuss the im-
portance of money market funds for financial stability and
whether situations similar to those during the recent crisis
in the U.S. and Icelandic markets could arise in Sweden. We
find that there are similarities between the Swedish and Ice-
landic funds, but few similarities with the U.S. funds. In Swe-
den, as was the case in Iceland, the assets under manage-
ment are concentrated in a few funds and the connection to
the major banks is strong. However, given the relatively small
size of the money market funds in Sweden, we do not find
that they, in isolation, are of major systemic importance as a
source of funding for the Swedish banks. The funds are more
likely to have a systemic impact through spill-over effects
on the banking system, especially in a market already char-
acterized by high uncertainty and risk aversion. The money
market funds are thus more important to certain parts of the
financial market, such as the market for corporate commer-
cial paper and covered bonds.
Gudrun Gunnarsdottir — Financial Economist, Financial Stability Department, Sveriges Riksbank
Maria Strömqvist — Quantitative Analyst, Brummer Multi-Strategy1
This paper was written when Maria Strömqvist was a senior economist at 1
the Riksbank. We are grateful for useful comments and help with data from
Elias Bengtsson, Anders Bjällskog, Heidi Elmér, Johanna Fager Wettergren,
David Forsman, Johannes Holmberg, Kerstin Mitlid, Kjell Nordin, Fredrik
Pettersson, Anders Rydén, Kristian Tegbring, and Staffan Viotti. The views
expressed in this paper are the authors’ and not necessarily those of the
Riksbank. We are responsible for all errors.
24
As the recent financial crisis has shown, in certain situations money market
funds can be considered important for financial stability. These situations
are characterized by extensive uncertainty and instability in the markets.
The money market funds in both the U.S. and Iceland were severely af-
fected by the recent financial crisis. This paper discusses the importance
of money market funds for financial stability and whether situations similar
to those during the recent crisis in the U.S. and Icelandic markets could
arise in Sweden. Do the Swedish money market funds have significant
similarities to the U.S. and Icelandic money market funds?
Factors that influence the importance of money market funds, apart from
the market situation, include the risk of spill-over effects to the banking
system, investor sentiment, and whether the funds are an important source
of funding for banks and mortgage institutions. This paper examines the
importance of these factors for the Swedish money market funds.
Data has been collected from several different sources for 2003 to the
third quarter of 2009. Data at the aggregate level has been collected
from the Swedish Investment Fund Association (Fondbolagen), Statistics
Sweden, and Morningstar. To analyze specific holdings, we examined
the portfolios of seven large money market funds. The data on individual
funds was collected from the Swedish Financial Supervisory Authority
(Finansinspektionen), the funds’ annual and semi-annual reports, and
from the fund companies themselves. The paper focuses on money mar-
ket funds mainly investing in securities issued in domestic currency. For
the Swedish and Icelandic markets, money market funds are defined as
short-term bond funds with an average maturity of less than one year.
The corresponding definition for U.S. money market funds is 90 days.
Differences in definitions will be discussed later in the paper.
What happened to the money market funds in the U.S. and Iceland, and why were these funds considered important for financial stability?A run on U.S. money market fundsThe U.S. money market mutual fund market is the largest of its kind in the
world (about U.S.$4 trillion). The funds invest in short-term assets and
the weighted average maturity of the portfolios of money market funds is
restricted to 90 days [Baba et al. (2009)].2 The U.S. money market funds
are structured to maintain a stable net asset value of U.S.$1 per share;
this is called the Buck system.3 This simplicity is important because there
are a lot of transactions in the funds as they are often used as a cash
management tool. As the money market funds do not have capital buf-
fers, they instead rely on discretionary financial support from their spon-
sors whenever the value of a share threatens to fall below U.S.$1. The
Buck system had only been broken once in 30 years, until the fall of
Lehman Brothers in September 2008.4 On 16 September 2008, the day
after Lehman Brothers’ fall, the Reserve Primary Fund announced a share
price for its flagship fund of 97 cents. This was the first money market
fund open to the general public to ever break a buck. The U.S.$64.8 bil-
lion fund held U.S.$785 million in Lehman Brothers’ commercial paper
[Waggoner (2009)]. In the end, it was the combination of the holdings, the
large redemptions, and the lack of resources from the sponsor (the fund
company Reserve) to back the fund that led the fund’s net asset value to
drop to 97 cents [McCabe and Palumbo (2009)].
This event triggered a run on U.S. money market funds, especially funds
that invested in non-government securities. Investors moved their money
to funds that, for example, only invested in government securities and
bank deposits. Institutional investors liquidated much more than retail
investors. As an example, institutional investors liquidated 16 percent of
their holdings in a couple of days, while individuals liquidated 3 percent
at the same time [Baba et al. (2009)]. This had severe financial stabil-
ity implications, including freezing the commercial paper market. U.S.
money market funds held nearly 40 percent of the outstanding volume of
U.S. commercial paper in the first half of 2008 [Baba et. al. (2009)]. The
U.S. government stepped in and guaranteed U.S. money market mutual
funds on 18 September, 2008. In the press release from the U.S. Treasury
(2008), the justification for the action was to protect and restore investor
confidence and the stability of the global financial system. The money
market funds were considered to be of systemic importance, as they
have an important role as a savings and investment vehicle, as well as
a source of financing for the capital markets.5 The U.S. government be-
lieved that the concerns about the net asset value of money market funds
falling below U.S.$1 had exacerbated global financial market turmoil,
causing a spike in certain short-term interest rates and increased volatil-
ity in foreign exchange markets. The event also provoked severe liquid-
ity strains in world markets. European banks, which have experienced a
large growth in U.S. dollar assets, were affected when the opportunities
for dollar funding were reduced, partly due to the problems with the U.S.
money market funds [Baba and Packer (2009)]. In its press release, the
U.S. Treasury concluded that actions from the authorities were necessary
to reduce the risk of further heightened global instability.6
The weighted average maturity of Swedish and Icelandic funds is around one year.2
The U.S. money market funds are categorized by their investment objectives and the type 3
of investors in the fund. For example, they can be divided into prime funds that invest in
non-government securities, which can be divided further into institutional or retail prime
funds depending on the investors. The Buck system provides convenience and simplicity to
investors in terms of tax, accounting, and record keeping. Returns on investments are paid
out as dividends with no capital gains or losses to track.
A small institutional money market fund, Community Bankers Money Fund, broke the buck 4
in 1994. However, this had no effect on financial stability.
The U.S. money market funds were also important for the asset-backed commercial paper 5
market (ABCP) and thus, there was a connection between the funds and the real economy.
However, the money market funds did not experience large redemptions during the ABCP
crisis that started in mid-2007 (sponsor support played a large role there).
New rules from the SEC for the U.S. money market funds were adopted in January 2010. 6
These new rules include new liquidity requirements, tighter constraints on credit quality,
new disclosure requirements, and new procedures for the orderly shutdown of money
market funds that break the buck [McCabe and Palumbo (2009)].
25
The Capco Institute Journal of Financial TransformationMoney Market Funds and Financial Stability: Comparing Sweden to the U.S. and Iceland
The Icelandic money market funds crashed with the financial systemIn Iceland, the money market funds were not the source of the crisis, but
they were severely affected and thus aggravated the crisis. The Icelandic
case is interesting for two reasons: firstly, it shows that money market
funds can be risky and make large losses; secondly, it points to the risks
of excessive connections between the mutual funds and the domestic
banks. The largest money market funds in Iceland, in terms of assets
under management in Icelandic kronor, were owned by Landsbanki,
Kaupthing, and Glitnir’s fund companies.7 These banks were taken over
by the Icelandic government in October 2008. Around the time of their
collapse, the money market funds were closed.
When the financial system in Iceland collapsed, it affected all major is-
suers of fixed income securities in Iceland, financial institutions as well
as corporations. New emergency legislation was implemented in Iceland
on 6 October, 2008,8 in which bank deposits were given priority before all
other claims. Before this new law was implemented, bonds and certifi-
cates had the same right to claims as deposits. Thus, the new legislation
had a negative impact on the money market funds’ recovery rate from se-
curities. To protect the investors in money market funds, the government
decided that the banks, now government-owned, would themselves re-
solve the issue, with the interests of their investors as their primary goal.9
The banks then bought back securities from the money market funds for
a total of about €552 million (ISK 83 billion) before they paid their inves-
tors. According to Morgunbladid, 60 percent of that amount has now
been written off [Juliusson (2009)].10 The money market funds returned
between 69 and 85 percent of their value to their investors after the se-
curities had been bought from the funds [Juliusson (2009)]. Securities is-
sued by financial institutions and corporations accounted for the majority
of losses [Sigfusdottir (2008)], despite the fact that securities seem to
have been bought back at higher prices than they were ultimately worth,
given the write-offs.
Two out of three money market funds did not report negative returns be-
fore they were closed and, subsequently, reported losses of between 15
to 31 percent. The exception was Glitnir, where the money market fund’s
returns decreased when the bank received emergency funding from the
government a few days before the system collapse. The fund had to be
closed for three days due to large outflows following negative media at-
tention and problems encountered by corporations linked to Glitnir. The
fund then opened again for a short period until the bank was fully taken
over by the government. The other money market funds also experienced
outflows in 2008, although the amounts of these varied between funds.
Outflows were especially large around the time that Glitnir received emer-
gency funding from the government [Sigfusdottir (2008)].
The Icelandic money market funds were poorly diversified with substantial linkages to parent banksThe Icelandic money market funds had mainly invested in domestic se-
curities issued by financial institutions and corporations. For example,
the money market fund owned by Landsbanki had 60 percent of its in-
vested capital with exposure to financial institutions11 and 40 percent
invested in securities issued by corporations at its closing [Sigfusdottir
(2008)]. In addition, all the funds had invested a large proportion of their
assets in securities linked to the Icelandic banking system, either directly
in securities issued by financial institutions, by corporations with owner-
ship stakes in the Icelandic banks, or even the bank’s major debtors. At
the time of closing, the money market fund of Landsbanki had 41 percent
and Kaupthing 21 percent in securities connected to their own parent
banks [Gunnarsdottir (2008)]. Glitnir’s fund, Sjodur 9, had 46 percent in
securities connected to Glitnir, according to its last semi-annual report in
2008. In addition, the money market funds also had some deposits with
their own bank. For example, in Kaupthing’s case, deposits amounted
to 66 percent of the fund at its closing, a large part of which was held
with Kaupthing [SIC (2010)].12 It is, therefore, evident that the Icelandic
money market funds formed a source of funding for their parent banks to
a certain extent. Nevertheless, the direct access to funding in the money
market funds was helpful when foreign funding markets closed for the
Icelandic banks in 2008. The three large money market funds amounted
to €4.4 billion (ISK 400 billion) at their peak at the end of 2007, an amount
equivalent to approximately 30 percent of Iceland’s GDP in 2007. At the
same time, household deposits with the Icelandic banks amounted to €6
billion (ISK 550 billion).13 In fact, many households used money market
funds as a substitute for normal bank deposits.
Money market funds are often considered to involve very low risk because
their returns are stable. However, in Iceland, the money market funds were
exposed to large systematic risk because of the small size of the bond
market and the concentration of market securities.14 The funds mainly
The three parent banks accounted for 90 percent of the total financial system in Iceland 7
according to Sedlabanki’s 2008 Stability Report.
Icelandic law number 125/2008: http://www.althingi.is/lagas/nuna/2008125.html.8
A great deal of media attention focused on the money market funds around the time of their 9
closure, when it was evident that losses would be made. Many household investors had not
fully understood the possible risks involved in investing in those funds, as they had been
marketed “almost” as a substitute to bank accounts with similar risk [SIC (2010)].
The buyback was based on the expected recovery rate of securities, although there was 10
great uncertainty at the time.
In Landsbanki’s data the exposure to financial institutions included both issued securities 11
and deposits. A graph of the development of the fund indicates that deposits seem to have
been around half of the exposure to financial institutions [Sigfusdottir (2008)].
If the emergency law making deposits priority claims had not been implemented, the losses 12
of Kaupthing’s money market fund investors (for example) would have been a lot larger, as
a large part of its portfolio was in deposits.
According to data from the Central Bank of Iceland.13
35 percent could be invested in one counterparty, and then other counterparties had to 14
count for less than 20 percent [Kaupthing (2008)].
26
invested in securities issued in Icelandic kronor. Investment in government
securities was minimal and the reason given for this was the small supply
of government bonds available in the market. In normal times, funds like
these should be diversified enough to be able to handle losses stemming
from one issuer. However, when the whole financial system collapses, the
situation is very different. In such a situation, not even a high degree of di-
versification will help. Even though the money market funds had invested
in securities with the highest credit rating available on the Icelandic mar-
ket, they made large losses. Despite that fact, it can be argued that the di-
versification in the funds was not satisfactory and that substantial linkage
to the money market funds’ parent banks created large risks.15
How are money market funds important for financial stability?The U.S. and Icelandic crises concerning money market funds point to
some explicit ways in which money market funds are important for finan-
cial stability. These are specified in more detail in this section and then in-
vestigated further in terms of the Swedish market in the following section.
Spill-over effectsIn the U.S. situation, the authorities explicitly stated that one of the main
reasons for guaranteeing the value of the money market funds after the
collapse of Lehman Brothers was spill-over effects resulting in further
heightened global instability. In Iceland, it is likely that the government
owned banks’ purchases of assets were performed in order to avoid fur-
ther financial instability and decrease the losses of investors. In order to
minimize panic in the market, at the time of the system failure, the gov-
ernment of Iceland emphasized that all deposits would be guaranteed.
Spill-over effects from problems in money market funds to the banking
system are more likely if there is a high concentration in a small number
of funds and if the fund market is dominated by the major banks’ mutual
fund companies, which was the case in Iceland.
Investors’ degree of sophistication affects flowsInvestors’ expectations can have an effect on fund flows in a financial cri-
sis. Investors include both retail investors (households) and institutional
investors. If investors believe that money market funds are liquid and
have low risk, their reactions may be stronger in the event of problems
than would otherwise be the case. According to Henriques (2008), retail
investors in the U.S. considered money market funds to be as safe as
bank savings accounts. This also appeared to be the case in Iceland. The
main outflow from Icelandic money market funds occurred after nega-
tive media attention focused on Glitnir’s money market fund. Sirri and
Tufano (1998) find that U.S. mutual fund flows are directly related to the
current media attention received by the funds. Klibanoff et al. (1998) also
find that investors react more strongly to headlines in the newspapers.
Consequently, extensive media coverage of problems in mutual funds
could have a major impact on fund flows. Investor sentiment can also
be important. According to Davis and Stein (2001), households are more
likely to have more diverse views than groups of institutional investors.
This is supported by the fact that the largest outflows from the U.S. mon-
ey market funds came from institutional investors.
Potentially large market impact from fire salesThe liquidity of financial securities (for example covered bonds) and cor-
porate securities decreased during the crisis, especially around the col-
lapse of Lehman Brothers. This lower liquidity was manifested in higher
bid-ask spreads, lower prices and turnover. It was, in fact, even difficult
to get a price for financial and corporate securities that, in normal times,
had been liquid in the market.16 Investors’ risk aversion increased sharply.
The liquidity problems for money market funds were evident both in Ice-
land and the U.S. If funds are forced to liquidate securities in such a
market, unless the managers are willing to realize losses, it is likely that all
funds would have to sell the most liquid and (in relative terms) most fairly
priced securities, such as, for example, government bonds.
Source of funding for systemically important financial institutions and marketsMoney market funds can be a source of funding for banks. This was
particularly evident in Iceland, where the money market funds had in-
vested a large share of their capital in securities issued by the domestic
banks, and in particular, securities linked to the bank that owned the
fund. If money market funds buy a large share of the short-term bonds
and commercial paper issued by banks and other systemically important
institutions, they become an important source of funding for these insti-
tutions. The continuous and consistent availability of short-term funding
for these institutions is essential to financial stability. If the money market
funds are the dominating investors in a specific submarket of the finan-
cial market, their problems may have negative consequences that may
spread through the financial system as well as to the real economy in the
long run. This is illustrated by the U.S. money market funds, which were
essential for the commercial paper market. When the U.S. money market
funds incurred losses from defaults in connection with the Lehman Broth-
ers collapse and were not able to continue investing to the same degree
Althingi’s Special Investigation Commission’s report about the collapse of the Icelandic 15
banks includes a chapter on the Icelandic money market funds. The main conclusions
of the Committee are that the funds were excessively linked to their parent companies
in terms of investment selection and that the separation between the fund company and
parent bank was unsatisfactory. The money market funds grew very fast and became
too large for the Icelandic securities market since the supply of solid and liquid securities
was limited. The interests of the parent bank seem to have been prioritized ahead of the
interests of the investors. In some cases, the investors were not provided with reliable
information about the standing of their investments, which were frequently worse than the
returns of the funds implied. Outflows from the funds in 2008 were also investigated and
the Commission has reason to believe that some investors (individuals and corporations
linked to the banks) had better information than others. This issue has been sent to the
Public Prosecutor in Iceland and the Financial Supervisory Authority for further investigation
[SIC (2010)].
From conversations with fund managers of the largest Swedish money market funds.16
27
as before because of large redemptions, the commercial paper market
froze. If this had continued for a more extensive period, it could have had
a negative effect on the financial and corporate sector and, in the end,
the real economy. In addition, there were wide-ranging spill-over effects
from the U.S. market to the global financial markets.
Characteristics of Swedish money market funds that may affect their influence on financial stabilityCertain factors increase the risk of spill-over effects from the money market
funds to the rest of the financial system (and especially the major banks).
Here we look at three factors applying to the Swedish money market funds
– firstly, whether the money market funds have substantial amounts of as-
sets under management; secondly, whether the assets under management
are concentrated in a few large funds; and, thirdly, whether there is any
significant connection with the systemically important banks.
The size of the Swedish market for money market funds is relatively smallTable 1 presents summary statistics for the Swedish market, both for
money market funds registered in Sweden and abroad. According to
Morningstar, there are 45 mutual funds classified as money market funds
investing in securities issued in SEK. The total assets under management
by these funds were about €20 billion (SEK 204 billion) at the end of the
third quarter of 2009. According to Fondbolagen’s data, the money mar-
ket funds constitute about 14 percent of the total Swedish mutual fund
market, while long-term bond funds constitute about 11 percent.17
The Swedish market is highly concentrated and dominated by the major banksThe average fund has €0.4 billion (SEK 4.5 billion) in assets under man-
agement, but the median is only €0.2 billion (SEK 1.6 billion). This im-
plies that there are several large funds in the sample, as illustrated by the
maximum levels shown in Table 1. The largest fund has €2.8 billion (SEK
28 billion) in assets under management. The smallest fund has only €12
million (SEK 100 million) under management. Of the 45 funds, 19 funds
are managed by the four largest Swedish banks’ mutual fund compa-
nies (Svenska Handelsbanken (SHB), SEB, Nordea and Swedbank).
Even though these funds account for 42 percent of the total number of
funds, the market share of assets under management is equivalent to
81 percent. Nordea has the largest market share (30 percent) followed
by Swedbank (24 percent). SEB and SHB have market shares of 17 and
11 percent, respectively.
Table 2 presents the assets under management and the percentage of
total market assets under management for seven large money market
funds investing in Swedish securities at the end of the third quarter of
2009. The largest fund is Nordea Sekura, with a market share of 13 per-
cent, followed by Swedbank Robur Svensk Likviditetsfond (market share
of 9 percent). The seven funds have a total market share of 55 percent.
The Swedish market for money market funds is thus highly concentrated,
with seven of the largest funds having more than half of the assets under
management. That implies that it could be enough for one or a small
number of money market funds to run into trouble for there to be larger
implications for the financial system. Financial stability could be affected,
especially if the general market situation at the time were to be charac-
terized by large uncertainty, as was shown by the Icelandic example.
The fact that the market is dominated by the four largest banks in terms
of assets under management also has implications for financial stability.
These implications are mainly spill-over effects from problems with the
money market funds to the banking sector. For example, there is a risk of
a negative reputation effect with lower trust in the banks as a result. As
an example, according to Svenska Dagbladet (2009), Swedbank Robur
compensated mutual fund investors in Estonia to protect the reputation
of the mutual fund company as well as the parent bank. If investors lose
To compare, the European market for money market funds has about €1,000 billion in 17
assets under management, which is 21 percent of the total European fund market [EFAMA
(2010)]. The numbers are for UCITS funds.
The Capco Institute Journal of Financial TransformationMoney Market Funds and Financial Stability: Comparing Sweden to the U.S. and Iceland
All SHB Nordea SEB Swedbank Others
Number of funds 45 2 7 5 5 26
AUM (€million) 20037 2107 5928 3322 4776 3903
Average AUM 445 1053 847 664 955 150
Median AUM 158 1053 287 917 969 73
Max AUM 2781 1512 2781 1224 1634 814
Min AUM 12 595 27 35 177 12
Market share 11% 30% 17% 24% 19%
The table shows summary statistics for Swedish money market funds investing in SEK.
The first column is for all funds in the sample, the next four columns represent funds
owned by the four largest banks’ fund companies and the last column all other mutual fund
companies. The data was collected from Morningstar on 30 September 2009.
Table 1 – Summary statistics
(30 September 2009)
Fund name AUM (€million) % market AUM
Nordea Sekura 2676 13%
Swedbank Robur Svensk Likviditetsfond 1714 9%
Swedbank Robur Penningmarknadsfond 1601 8%
Handelsbanken Lux Korträntefond 1418 7%
Nordea Institutionell Penningmarknadsfond 1393 7%
Nordea Likviditetsinvest 1198 6%
SEB Penningmarknadsfond SEK 1019 5%
Total 11019 55%
The data has been collected from the Swedish Financial Supervisory Authority, the funds’
annual and semi-annual reports and directly from the fund companies.
Table 2 – Seven large Swedish money market funds
28
their trust in a certain bank’s funds, the result could be that they withdraw
their money, not only from the funds, but also from their deposits in the
bank. Bank deposits are an important source of funding for the Swedish
banks [Sveriges Riksbank (2009)].
Retail investors are the largest investor groupKnowledge of the investors in mutual funds is of interest for financial sta-
bility given that this will provide information on who would be affected by a
decrease in value of mutual funds. In addition, different types of investors
may induce financial instability through their behavior. Households are
the largest investor group in Swedish money market funds, with a share
of 68 percent18 on average between 2007 and 2009 (Figure 1). Swedish
corporations have kept their proportion to around 23 percent. The fact
that household investors constitute a major component of the investors in
these mutual funds indicates that the average investor is relatively unso-
phisticated. As stated before, this has positive and negative implications
from a financial stability perspective. Retail investors are more sensitive to
media attention about the funds, but they do not tend to react to market
events as strongly, as quickly, or with as much coordination as institution-
al investors. In discussions, various Swedish fund managers stated that,
during the crisis, it was mainly institutional investors that asked questions
about the implications of the market events for the money market funds.
This could be due to limited media attention concerning the issue in Swe-
den, with less information thus reaching the retail investors. However, the
situation could be different in a future crisis.
Securities issued by financial institutions is the largest component in the portfoliosFrom a financial stability perspective, it is interesting to know what secu-
rities these funds invest in, what they would potentially have to sell in the
event of major redemptions, and which submarkets might be affected.
Data at the aggregate level is available from Statistics Sweden.19 Taking
a look at the components of the total investment of money market funds
in Figure 2, bonds issued by financial institutions are the largest compo-
nent, followed by securities issued by corporations. The share of bonds
issued by financial institutions has increased from 46 percent in the first
quarter of 2007 to 59 percent in the third quarter of 2009. This increase is
likely to depend on three factors: a limited supply of available securities,
the relatively higher yield they give compared to T-bills, and the reduc-
tion of risk due to the government guarantee programs that came into
effect late in 2008. However, this increase may have consequences on
the systematic risk taken by funds due to lower diversification between
asset classes.
Investment in foreign securities (issued in Swedish kronor) has decreased
in every period, from a level of about 12 percent in the first quarter of 2007
to 5 percent in the second quarter of 2009, indicating a stronger home
bias. It is likely that mutual funds decreased their investments in foreign-
issued securities in Swedish kronor because of higher uncertainty about
their issuers and poor performance by these securities during the crisis.
In discussions, fund managers stated that the holding of geographically-
diversified portfolios had a negative effect on funds’ ability to manage
the crisis.
A closer investigation of the holdings of seven large money market fundsTo get a better understanding of the investments undertaken by Swedish
money market funds, we look at the holdings of seven large funds at three
points in time: December 2007, December 2008, and June 2009. Table 3
shows the average exposure to asset classes (using nominal values) for the
seven funds for the three periods in time. The assets are divided into cov-
ered bonds, bonds that are guaranteed by the government, government
bonds, general bonds and notes, subordinated bonds, T-bills, commercial
paper, and cash.20 The fact that Swedish money market funds can invest in
This includes direct investments by households, individual pension saving (IPS), premium 18
pension savings (PPM), and unit linked investments (fondförsäkring).
The data only includes funds registered in Sweden.19
The asset classes are categorized by Bloomberg with some revisions. Under general 20
bonds and notes, we put Bloomberg categories bonds, notes, senior unsecured, senior
notes, unsecured, unsubordinated, and company guarantee. Covered bonds, government
guaranteed bonds, and general bonds and notes can all be FRNs.
67% 66% 66% 67% 68% 67% 69% 70% 70% 70% 71%
6% 6% 5% 6% 6% 5% 5% 5% 4% 5% 5%
23% 24% 24% 22% 22% 25% 23% 23% 24% 23% 21%
3% 4% 6% 4% 4% 3% 3% 2% 2% 3% 3%
0%
50%
100%
Q1 2007 Q2 2007 Q3 2007 Q4 2007 Q1 2008 Q2 2008 Q3 2008 Q4 2008 Q1 2009 Q2 2009 Q3 2009
Other Swedish corporations Non pro�t inst. for households Swedish households
Figure 1 – Investor groups (aggregate data)
24% 22% 19% 18% 17% 13% 19%
27% 18% 18%
13%
46% 45% 45% 50% 54%
55% 51%
50%
54% 53% 59%
18% 22% 25% 23% 20% 26% 25%
18% 23% 25% 23%
12% 12% 11% 9% 9% 6% 5% 5% 4% 5% 5%
0%
50%
100%
Q1 2007 Q2 2007 Q3 2007 Q4 2007 Q1 2008 Q2 2008 Q3 2008 Q4 2008 Q1 2009 Q2 2009 Q3 2009
Rest of the world Other domestic Financial institutions Central government
Figure 2 – Investments (aggregate data)
29
securities with longer maturity (unlike U.S. funds) comes from the fact that
these can have a weighted average maturity of up to one year.
The share of covered bonds increased during the crisisCovered bonds and general bonds and notes are the largest asset
classes in the portfolios. At the end of 2007, they constituted, on aver-
age, 40 percent of the portfolio each. In June 2009, the share invested
in bonds and notes had decreased, while the share invested in covered
bonds had increased to 45 percent. This could be interpreted as the re-
sult of managers decreasing the risk in their portfolios during the financial
crisis by increasing the proportion of covered bonds to regular bonds. It
could also partly be a result of the issue, by the banks, of large quanti-
ties of covered bonds during the period. Government bonds increased
from almost nothing to 3 percent, and bonds guaranteed by the govern-
ment increased from 1 to 7 percent in the same period, again indicating
increased risk aversion. In addition, commercial paper increased from
12 percent in 2007 to 20 percent in 2008. On the other hand, according to
fund managers, the commercial paper market closed for a period of time
after the collapse of Lehman Brothers. It became difficult to sell com-
mercial paper in the secondary market, which meant that these papers
became less liquid.21 This, in turn, made it more difficult for commercial
paper issuers to get buyers for their issuance.
A large share of the securities are floating rate notesFloating rate notes (FRNs)22 constitute a large part of the bonds in the
portfolios. For example, for the largest fund, Nordea Sekura, around
74 percent of the total assets under management were invested in FRNs
in 2009. The interest rate on FRNs changes with short intervals (for ex-
ample every three months). Consequently, the interest rate risk is still
low but, since there is a credit risk premium, the return is higher than
for securities with shorter maturities. The longer maturity length of FRNs
does not affect the portfolio’s sensitivity to interest rate changes to any
great extent. However, having a portfolio dominated by FRNs may have
implications for the liquidity risk of the portfolio, especially if the credit
risk increases considerably in a financial crisis.
No large bias towards investing in the parent bank’s securitiesTable 4 sorts the holdings of the seven largest funds into exposure to
issuers. The issuers are divided into financial institutions (including mort-
gage institutions and real estate companies), government, corporations,
and cash. Confirming the findings from aggregate data, financial institu-
tions issue the vast majority of the securities in the portfolios, between
74 and 79 percent over time. The share of securities from mortgage in-
stitutions (including commercial paper, covered and regular bonds, and
notes) has increased during the period.23 At the end of 2007, 47 percent
of the portfolio was invested in securities from mortgage institutions. In
June 2009, the corresponding figure was 55 percent. Table A1 in the
Appendix displays the five largest exposures by issuer in each of the
seven funds at three points in time. In general, the majority of the largest
exposures are to mortgage institutions, which is consistent with previous
findings. All funds but one (SHB Korträntefond) have one of their top five
exposures, at all three points in time, to the mortgage institution Stads-
hypotek, a subsidiary of Svenska Handelsbanken. This is due to the fact
that Stadshypotek is one of the largest issuers on the Swedish covered
bond market. The fact that Stadshypotek remains the largest exposure
from 2007 to 2009 indicates that it is considered a good investment even
in turbulent times, given the flight to quality that normally occurs during a
financial crisis. The largest exposure ranges from 8 to 24 percent, indicat-
ing some differences in counterparty diversification by the funds. Also,
there is no large bias towards investing in the parent bank’s securities,
as in the Icelandic funds. The average exposure to securities issued by
the parent bank is 11 percent for the Swedish funds (the corresponding
figure for the Icelandic banks is approximately 36 percent).24
The Capco Institute Journal of Financial TransformationMoney Market Funds and Financial Stability: Comparing Sweden to the U.S. and Iceland
Dec-07 Dec-08 June-09
Covered bonds 40% 35% 45%
Government guarantee 1% 1% 7%
Government bonds 0% 2% 3%
Bonds/notes 40% 33% 25%
Subordinated bonds 2% 2% 2%
T-bills 4% 4% 2%
Commercial paper 12% 20% 13%
Cash 2% 3% 2%
Nominal values are used. Data is collected from Bloomberg
Table 3 – Average exposures to asset classes (7 funds)
Dec-07 Dec-08 June-09
Financial (including mortgage) 76% 74% 79%
Government 4% 6% 4%
Corporations 18% 17% 15%
Cash 2% 3% 2%
(Mortgage) 47% 47% 55%
Nominal values are used. Data is collected from Bloomberg
Table 4 – Average exposures to type of issuer (7 funds)
During this period all securities with credit risk became more difficult to trade. The 21
commercial paper market is one example (especially issuance by financial institutions
and corporations). Bonds issued by financial institutions and corporations and structured
products are other examples.
FRNs are issued by all types of institutions and can be covered bonds, government 22
guaranteed bonds, general bonds, and notes, etc.
The mortgage institutions are the Swedish hypotek.23
This is without deposits, which would increase the exposure substantially for the Icelandic 24
funds, given large deposit position by some.
30
Cross-investments lower in the crisisTable A2 in the Appendix shows, in detail, the cross-investments between
funds. That is, the proportion of a fund’s assets invested in exactly the
same securities as another fund. If the funds largely invest in the same
securities, this increases the systematic risks and the potential market im-
pact in the event that the funds should need to liquidate securities. Some
cross-investments are to be expected, given the relatively small size of
the Swedish market and the limited selection of securities. This limitation
arises from the fact that the funds invest in securities issued in Swedish
kronor. As Nordea Sekura is the largest fund, it could be expected that
the other funds would have the highest percentages of cross-investment
with this fund. However, the cross-investments with Sekura were much
higher in 2007 (73 percent on average) than in 2008 (37 percent) and 2009
(47 percent). We observe a similar trend between the other funds.
Swedish funds do not hold a large share of the outstanding amount of a securityAnother point to consider is how much the funds own of each security
compared to the outstanding amount. For example, if they hold a large
share of the outstanding amount of one security, this will affect liquidity,
making this security harder to sell in the market, should the need arise.
Making a rough estimate of the funds’ ownership in bonds, compared to
their outstanding amounts in June 2009 according to Bloomberg, we see
that the weighted average of ownership in each bond is around 15 per-
cent.25 This indicates that the Swedish funds, in general, do not own large
amounts in single bonds, as was the case in Iceland, where, in some
cases, the funds even held the full amount, which severely affected the
liquidity of the bonds [SIC (2010)].
Greater home bias could indicate higher risks in the futureAlthough lower cross-investments between the funds are positive for fi-
nancial stability and the potential market impact of money market funds,
the decrease of diversification in the portfolios during the financial crisis
can have negative effects in the long run. The share of covered bonds has
increased, making the funds more dependent on this market. Also, there
is less diversification among issuers and the home bias has increased.
Normally, it is always better to have a more diversified portfolio. However,
in the special case of the recent financial crisis, higher diversification was
negative from the fund managers’ perspective. In a financial crisis, the
correlation between assets increases, which decreases the positive ef-
fects of diversification. Also, Adrian and Shin (2008) put forward that,
when there is a high degree of diversification in the financial system, a
small shock can be amplified through market prices. The increased home
bias in the managers’ portfolios is a natural development, as the funds
that diversified their portfolio with foreign securities were punished in the
crisis. However, from a financial stability perspective, a strong home bias
could indicate a higher risk for the Swedish financial system if problems
with the domestic markets occur in the future.
Swedish money market funds are not an important source of funding for major banks but more important for certain marketsFor the funds to be an important source of funding for the Swedish banks,
a substantial part of the banks’ outstanding bonds should be held by the
money market funds. However, since that information cannot easily be
obtained from the data, we have made a rough estimate. According to
Statistics Sweden, Swedish financial institutions (banks and mortgage
institutions, etc.) issued securities in Swedish kronor in the amount of
€138 billion (SEK 1,509 billion) in the second quarter of 2009 (both long-
term and short-term).26 This constituted 24 percent of the banks’ total
interest bearing funding [Blomberg (2009)]. Given that the money market
funds at that time had total assets under management of €20 billion and,
on average, 79 percent was invested in securities issued by financial in-
stitutions, only a small part of the institutions’ securities funding would
potentially come from money market funds.
Additionally, according to Statistics Sweden’s financial market statis-
tics, deposits and borrowings from Swedish non-financial institutions
amounted to €164 billion (SEK 1,792 billion) in the second quarter of
2009, accounting for 29 percent of the banks’ total interest bearing fund-
ing [Blomberg (2009)]. Out of that €164 billion, households’ deposits
amounted to 43 percent of this funding. Consequently, if household in-
vestors lose their trust in a certain bank’s funds and withdraw their money,
not only from the funds but also from their deposits in the bank, this may
have an effect on the bank in question. However, if we look at the cov-
ered bond market in particular, we find that, on average, the largest funds
invest 45 percent of their portfolios in covered bonds. Assuming that, on
average, the share is the same for all Swedish money market funds, the
funds would have around 8 percent of all outstanding covered bonds
denominated in SEK.27 The Riksbank estimates the Swedish market for
corporate commercial paper to be worth about €10 billion. According to
the same assumption, the money market funds would thus have about
12 percent of the outstanding commercial paper issued by corporations.
Although this is not a huge figure, it is not entirely insignificant.
It is not only the size of the invested capital that matters but also the mo-
bility of the capital. The experience of the Swedish covered bond market
in 2008 shows that quick withdrawals of capital can have a substantial
effect on the stability of the market. After Lehman Brothers’ collapse,
Note that this is only a rough estimate as information on certain securities could not be 25
found in Bloomberg (although the majority could be found). In addition, commercial paper
is not included, as Bloomberg does not provide information on those (commercial paper
counted for 13% of the portfolios in June 2009).
This data does not include subsidiaries.26
The estimate of the size of the covered bond market comes from the Association of 27
Swedish Covered Bonds’ homepage.
31
many foreign investors wanted to sell their holdings of Swedish covered
bonds quickly. The market makers had problems handling the large vol-
umes of selling orders, which then disrupted the market. The problems
in the covered bond market were reduced in mid-September 2008, when
the Swedish National Debt Office started to issue large extra volumes of
T-bills to meet heightened demand for these short-term securities. The
money from these extra auctions was placed in reverse repos with cov-
ered bonds as collateral. In 2008, foreign investors decreased their hold-
ings of Swedish covered bonds by around €11 billion compared to 2007.
Although that figure is only about 7 percent of the covered bond market,
the outflow had a substantial impact due to its rapid pace.
How similar is the Swedish situation to the U.S. and Icelandic situations?Few similarities with the U.S. marketU.S. money market funds are very different from their Swedish peers. For
example, as previously mentioned, the weighted average maturity of the
portfolios of U.S. money market funds is restricted to 90 days. In Swe-
den, there is no set rule concerning the weighted average maturity of the
portfolio. Statistics Sweden’s definition of money market funds is that the
weighted average maturity is 397 days or less.28 Consequently, it is not
easy to compare Swedish and U.S. funds directly. Constraining the matu-
rity of the money market funds should have a positive effect on financial
stability. Given the fact that Swedish money market funds can invest in
both commercial paper and long-term bonds (like covered bonds), they
can potentially affect both markets if compelled to sell securities. The
problems in the U.S. funds mainly affected the money markets.
Concerning the funds’ potential market impact, money market funds ac-
count for 30 percent of the U.S. fund market, compared to around 15 per-
cent in Sweden. Institutional owners play a large role in the U.S. money
market funds.29 The Investment Company Institute in the U.S. estimates
that around 80 percent of U.S. companies use money market funds for
their cash management. There are no corresponding figures for Sweden,
but only about 23 per cent of the assets of Swedish money market funds
are held by corporations.
The money market funds in the U.S. were also important for the asset-
backed commercial paper market, and, thus, problems with the money
market funds had direct implications for the real economy. In Sweden,
the money market funds have not invested in structured products, but
the covered bonds are linked to the Swedish housing market.
Because the reporting of earnings in Swedish mutual funds is differ-
ent from the Buck system, it is likely that, when funds show a negative
performance, this will have more severe consequences in the U.S. than
in Sweden. Fund sponsors in the U.S. have provided financial support
when the market value of a share threatened to fall substantially below
one dollar, although there is no official guarantee that the fund shares
should always be above the dollar.30 The Swedish money market funds,
on the other hand, can both increase and decrease in value, a fact known
to most investors, even though, in normal times, the funds have shown
stable positive returns. Swedish funds are more sensitive to changes in
interest rates than U.S. funds, given that they can hold securities with
longer maturities.
Several similarities with Icelandic fundsIf, on one hand, there are few similarities between Swedish and U.S.
funds, there are, on the other, several similarities between Swedish and
Icelandic funds. The Icelandic money market funds were similar to the
Swedish money market funds in terms of investments, returns, and the
purpose they serve for investors.
As in Sweden, there is currently no clear definition of money market funds
in Iceland. It is up to the funds to define their average maturity and in both
countries money market funds are usually defined as having an average
maturity of about a year or less. Although there is no exact data on the
investors in the Icelandic money market funds, large proportions were
households, as in Sweden. Also, the Icelandic funds were not a major
cash management tool for corporations, unlike the funds in the U.S. In
Sweden and Iceland, the supply of government bonds (i.e., bonds is-
sued in domestic currency) was small, so the funds consisted mostly of
securities issued by financial institutions and corporations, although cash
increased in the Icelandic funds in the period before the system collapse.
On the other hand, the Swedish bond market, of which covered bonds
form a large part, is larger than the Icelandic bond market. Consequently,
the diversification opportunities are better in the Swedish market, al-
though both markets are still small compared to the U.S. market.
The investments made by the Swedish money market funds are also
closely linked by ownership to the major banks in Sweden, a situation
similar to that in Iceland. However, Swedish funds do not have the same
strong bias towards investing in securities issued by the parent bank as
the Icelandic funds did. Another important factor to consider is that Swe-
den is a larger country than Iceland and its banking sector is not as big
in terms of GDP as the Icelandic banking sector was before the collapse.
Also, largely due to the careful regulation of financial markets (including
money market funds) and the experience of the domestic banking crisis
in the 1990s, Sweden was better prepared for the crisis than Iceland.
The Capco Institute Journal of Financial TransformationMoney Market Funds and Financial Stability: Comparing Sweden to the U.S. and Iceland
According to email correspondence with Johannes Holmberg, Statistics Sweden.28
According to email correspondence with Fredrik Pettersson, Fondbolagen.29
Given the short maturity of U.S. money market funds (90 days), the volatility in the funds is 30
low and thus, in most times the support does not involve much risk for the sponsors.
32
How did the recent financial crisis affect Swedish money market funds?No run on Swedish money market fundsFigure 3 shows the monthly net capital flows to money market funds from
2007 to 2009. The figure also plots the repo rate (monthly averages) as
an indication of the general level of interest rates. The largest inflow into
money market funds was in August 2007, amounting to €918 million (SEK
7.4 billion). In that month, there was substantial outflow in equity funds.
This is directly linked to the beginning of the subprime crisis, a liquidity
crisis that turned into a long period of recession. In a financial crisis,
investors want more liquid and less risky investment portfolios, thus turn-
ing to money market funds. However, money market funds turned out to
be more risky than previously thought. The Lehman Brothers’ collapse
completely changed the risk tendencies in the market. Swedish money
market funds did not experience runs in the period after Lehman Broth-
ers’ collapse. However, sales and redemptions in the funds increased
rapidly, even though net capital flows show inflows. This caused a lot of
stress on the funds, especially on those (few) that had securities issued
by Lehman Brothers or other U.S. financial institutions. These securities
were probably held by some managers because they had a relatively high
credit rating as well as yield before the failure and because the managers
found it unlikely that the authorities would let the investment bank fail.
However, this assumption turned out to be incorrect.
Lower liquidity in the Swedish bond marketIn addition, liquidity disappeared in the Swedish bond market for a few
days after Lehman Brothers’ collapse. Consistent with this is the extreme
increase in bid-ask spreads during those days for covered bonds in the
Swedish market (illustrated in Figure 4 by data for one large issuance of
a benchmark covered bond). Liquidity is crucial for money market funds,
given that they have to be able to pay out redemptions on the same day.
However, according to the statistics collected by the Riksbank, there was
still some turnover in the Swedish covered bond market, indicating that
there were investors willing (or forced) to trade during these days of acute
stress (Figure 5).
In a situation with large redemptions and low liquidity in the markets,
money market funds can use the repo market to access cash. The se-
curities they invest in are commonly used in the repo market. It was,
therefore, important that the repo market in Sweden continued to func-
tion throughout the financial crisis, although it was more difficult to use
collateral with higher risk in repo agreements. Fortunately, the Swedish
mutual money market funds were able to handle the redemptions dur-
ing the most acute phase of the crisis. Although there was some media
attention31 regarding the situation after Lehman Brothers’ collapse, the
outflows stayed at moderate levels.
See, for example, e24 (2008).31
0,0
0,5
1,0
1,5
2,0
2,5
3,0
3,5
4,0
4,5
5,0
1.000
800
600
400
200
0
200
400
600
800
1.000
1.200
2007 2008 2009
New �ows Repo rate
MEUR
%
Source: Fondbolagen and the Riksbank.
Figure 3 – Monthly net flows
0
2
4
6
8
10
12
14
jan-0
8
feb-0
8
mar
-08
apr-0
8
maj-
08
jun-0
8
jul-0
8
aug-
08
sep-0
8
okt-0
8
nov-
08
dec-0
8
jan-0
9
feb-0
9
mar
-09
apr-0
9
maj-
09
jun-0
9
jul-0
9
aug-
09
sep-0
9
okt-0
9
Stadshypotek 6 12-16-2009 (6 yr)
Source: Bloomberg
Figure 4 – Bid-ask spread for a Swedish covered bond
-
2.000
4.000
6.000
8.000
10.000
12.000
14.000
16.000
Jul/0
8
Jul/0
8
Jul/0
8
Jul/0
8
Jul/0
8
Aug/0
8
Aug/0
8
Aug/0
8
Aug/0
8
Sep/0
8
Sep/0
8
Sep/0
8
Sep/0
8
Oct/0
8
Oct/0
8
Oct/0
8
Oct/0
8
Oct/0
8
Nov/0
8
Nov/0
8
Nov/0
8
Nov/0
8
Dec/0
8
Dec/0
8
Dec/0
8
Dec/0
8
Dec/0
8
Jan/
09
Jan/
09
Jan/
09
Jan/
09
2008 2009
Source: Riksbank. MEUR.
Figure 5 – Turnover covered bonds
33
The Capco Institute Journal of Financial TransformationMoney Market Funds and Financial Stability: Comparing Sweden to the U.S. and Iceland
ConclusionWe find that there are similarities between the Swedish and Icelandic
money market funds, but few similarities with the U.S. funds.
Lack of diversification creates risksIn Sweden, money market funds invest, on average, almost 60 percent of
their capital in securities issued by financial institutions (the correspond-
ing number for the seven large funds is 79 percent). This lack of diversi-
fication may have consequences on the systematic risk taken by funds,
as shown by the Icelandic experience. Although lower cross-investments
between the funds are positive from the standpoint of financial stability
and the potential market impact of money market funds, the decrease
of diversification in the portfolios during the financial crisis could have
a negative effect in the long run. The share of covered bonds has in-
creased, making the funds more dependent on this market. Additionally,
the home bias has increased. Although lower diversification might have
had a positive effect in this crisis, given that there were fewer problems
on the Swedish market, compared to, for example, the U.S., lower di-
versification implies more systematic risk in the Swedish money market
funds. Consequently, if a problem were to arise in Sweden, this would
have a greater impact on the funds.
Swedish funds were able to handle the effects from Lehman Brothers’ collapseLiquidity disappeared in the bond and money markets after Lehman
Brothers’ collapse. Liquidity is crucial for money market funds, given that
they have to be able to pay out redemptions on the same day. It was,
therefore, important that the repo market in Sweden continued to func-
tion throughout the financial crisis. Fortunately, the Swedish mutual mon-
ey market funds were able to handle the redemptions during the most
acute phase of the crisis. Although there was some media attention in
Sweden regarding the situation following Lehman Brothers’ collapse, the
outflows stayed at moderate levels.
Funds can have a systemic impact through spill-over effectsInvestigating the risks associated with the Swedish money market funds,
we do not find that the funds, in isolation, constitute a large systemic
risk. However, the funds are large enough and connected enough to the
financial system to be able to aggravate an already vulnerable situation.
This was the case in both Iceland and the U.S. Given the relative size
of the money market funds in Sweden, we find it unlikely that they are
of systemic importance as a source of funding for the Swedish banks.
The funds are more likely to have a systemic impact through spill-over
effects on the banking system, especially in a market already character-
ized by high uncertainty and risk aversion. The money market funds are
then more important for some parts of the financial markets, such as the
markets for corporate commercial paper and covered bonds.
Future regulations may have an impactIn the future, the new proposed regulations for banks in Europe may af-
fect the Swedish money market funds. If banks are required to extend
their funding, focusing on issuing securities with longer maturities, this
could mean that the market for money market securities will decline. This
would have an adverse effect on the investment opportunities for money
market funds.
ReferencesAdrian, T. and Shin, H. S., 2008, “Liquidity and financial contagion,” Banque de France, •
Financial Stability Review – Special issue on liquidity, no. 11, February
Baba, N., R. McCauley, and S. Ramaswamy, 2009, “U.S. dollar money market funds and non-•
U.S. banks,” BIS Quarterly Review, March
Blomberg, G., 2009, “Funding of the banks during the financial crisis,” Sveriges Riksbank’s •
Economic Commentaries, No. 11.
Davis, E. P. and B. Stein, 2001, “U.S. mutual funds and financial stability,” Institutional Investor•
E24, 2008, “Lehman-kollapsen drabbar “säkra” investeringar” 16 September•
EFAMA, 2010, ”Trends in the European investment fund industry in the fourth quarter of 2009 •
and results for the full year 2009,” Quarterly Statistics Release, no.40, March
Gunnarsdottir, G., 2008, “Fjarfestu i tengdum felogum,” Morgunbladid, 17 December•
Henriques, D. B., 2008, “Treasury to guarantee money market funds,” New York Times, 20 •
September
Juliusson, T., 2009, “Reidufe fyrir skuldabref,” Morgunbladid, 8 October•
Kaupthing, 2008, “Uppgjor Kaupthings peningamarkadssjods,” Kaupthing, December•
Klibanoff, P., O. Lamont, and T. A. Wizman, 1998, “Investor reaction to salient news in closed-•
end country funds,” Journal of Finance, 53:2, 673-699
McCabe, P. and M. Palumbo, 2009, “A case study of financial stability analysis and policy •
reform: The run on U.S. money market funds in September 2008,” presentation at the Federal
Reserve Board, January 29
Sigfusdottir, E., 2008, “Opid bref til hlutdeildarskirteinishafa,” Landsbanki, 10 December•
Sirri, E. R. and P. Tufano, 1998 “Costly search and mutual fund flows”, Journal of Finance, 53:5, •
1589-1622
Special Investigation Commission (SIC), 2010, “Factors preceding and causing the collapse of •
the Icelandic banks in 2008 and related events,” 12 April
Svenska Dagbladet, 2009, “Ny kalldusch för Swedbank,” 16 October•
Sveriges Riksbank, 2009, “Financial Stability Report,” 2009:2•
U.S. Department of the Treasury, 2008, “Treasury announces guarantee program for money •
market funds,” press release, 18 September
Waggoner, J., 2009, “Money market mutual funds remain at risk of “breaking the buck’,” USA •
Today, 17 September
34
Appendix
31 December 2007
SEB Penningm SHB Kortränte Nordea Sekura Nordea Likviditet Nordea Instit Penn Swedbank Sv Likvid Swedbank PenningmSEB Penn. 63% 70% 49% 65% 13% 12%SHB Kortränte 16% 37% 18% 20% 29% 27%Nordea Sekura 68% 75% 91% 51% 86% 82%Nordea Likviditet. 41% 31% 84% 34% 17% 17%Nordea Instit. Penn. 61% 68% 67% 59% 25% 25%Swedbank Sv. Likvid. 20% 38% 98% 17% 15% 86%Swedbank Penn. 18% 38% 83% 17% 15% 82%
31 December 2008
SEB Penningm SHB Kortränte Nordea Sekura Nordea Likviditet Nordea Instit Penn Swedbank Sv Likvid Swedbank PenningmSEB Penn. 16% 38% 36% 53% 34% 38%SHB Kortränte 10% 11% 10% 5% 16% 21%Nordea Sekura 31% 10% 67% 32% 24% 28%Nordea Likviditet. 25% 11% 84% 33% 20% 26%Nordea Instit. Penn. 18% 7% 22% 19% 16% 19%Swedbank Sv. Likvid. 32% 11% 34% 34% 29% 83%Swedbank Penn. 29% 9% 32% 32% 31% 75%
30 June 2009
SEB Penningm SHB Kortränte Nordea Sekura Nordea Likviditet Nordea Instit Penn Swedbank Sv Likvid Swedbank PenningmSEB Penn. 26% 54% 52% 50% 38% 41%SHB Kortränte 15% 9% 9% 9% 4% 4%Nordea Sekura 20% 10% 68% 49% 28% 30%Nordea Likviditet. 23% 8% 83% 48% 27% 27%Nordea Instit. Penn. 36% 16% 70% 60% 41% 41%Swedbank Sv. Likvid. 26% 10% 37% 35% 35% 98%Swedbank Penn. 23% 7% 31% 30% 29% 86%
Table A2 – Cross-investments
December 2007 December 2008 June 2009
Fund Issuer Share Issuer Share Issuer ShareNordea Sekura Stadshypotek 13% Stadshypotek 9% Stadshypotek 11%
Spintab 10% Landshypotek 8% SBAB 9%Landshypotek 7% Swedbank 6% Landshypotek 8%Nordea Hypotek 6% Nordea 6% Swedbank 7%Nordea 5% DnB Nor 6% Nordea Hypotek 6%
Swedbank Nordea Hypotek 23% Stadshypotek 23% Stadshypotek 24% Likviditetsfond Stadshypotek 12% Swedish Government 15% Nordea Hypotek 15%
Swedbank Hypotek 9% SEB 8% Swedbank Hypotek 15%Swedbank 6% Nordea Hypotek 7% SBAB 9%SBAB 6% Swedbank Hypotek 6% Landshypotek 7%
SHB Lux. Swedish Government 16% Swedish Government 13% Länsforsäk.Hypotek 8% Korträntefond Nordbanken Hypotek 6% Stadshypotek 10% Swedish Covered 5%
Sandvik 4% Cash 6% Handelsbanken 4%Vasakronan 3% Swedish Covered 6% Landshypotek 4%General Electric 3% Länsförsäkringar 5% Volvo 4%
Nordea Inst. Stadshypotek 15% Nordea Hypotek 22% Stadshypotek 24% Penn. Nordea Hypotek 14% Stadshypotek 17% Nordea Hypotek 15%
SEB Bolån 9% Spintab 11% Swedish Government 9%Spintab 8% SBAB 9% Länsforsäk. Hypotek 8%Swedish Government 6% SEB Bolån 5% SBAB 7%
Swedbank Nordea Hypotek 24% Stadshypotek 24% Stadshypotek 23% Penn. Stadshypotek 14% Swedbank hypotek 13% Swedbank hypotek 15%
Swedbank Hypotek 9% SEB 8% Nordea hypotek 14%Swedbank 6% Landshypotek 8% Landshypotek 7%Swedish Covered 6% Swedish Government 7% SBAB 7%
Nordea Likv. Stadshypotek 10% SBAB 10% Stadshypotek 14%Volvo 8% Stadshypotek 10% SBAB 12%Landshypotek 7% Swedbank 8% Swedbank 9%SBAB 7% SEB 8% Landshypotek 7%Spintab 6% Landshypotek 7% Länsförsäk. Hypotek 6%
SEB Penn. Stadshypotek 21% Stadshypotek 22% Stadshypotek 23%SEB 20% Swedbank Hypotek 15% Swedbank Hypotek 12%Swedbank Hypotek 8% Nordea Hypotek 8% Nordea Hypotek 8%Nordea Hypotek 8% SBAB 5% SBAB 7%Swedish Covered 5% Swedish Government 3% Cash 5%
Table A1 – Largest exposures by issuer
35
PART 1
Interest Rates After the Credit Crunch: Markets and Models Evolution
AbstractWe present a quantitative study of the evolution of markets
and models during the recent crisis. In particular, we focus
on the fixed income market and we analyze the most relevant
empirical evidence regarding the divergence between Libor
and OIS rates, the explosion of basis swaps spreads, and
the diffusion of collateral agreements and CSA-discounting,
in terms of credit and liquidity effects. We also review the
new modern pricing approach prevailing among practitio-
ners, based on multiple yield curves reflecting the different
credit and liquidity risk of Libor rates with different tenors
and the overnight discounting of cash flows originated by
derivative transactions under collateral with daily margin-
ation. We report the classical and modern no-arbitrage pric-
ing formulas for plain vanilla interest rate derivatives, and the
multiple-curve generalization of the market standard SABR
model with stochastic volatility. We then report the results
of an empirical analysis on recent market data comparing
pre- and post-credit crunch pricing methodologies and
showing the transition of the market practice from the classi-
cal to the modern framework. In particular, we prove that the
market of interest rate swaps has abandoned, since March
2010, the classical single-curve pricing approach, typical of
the pre-credit crunch interest rate world, and has adopted
the modern multiple-curve CSA approach, thus incorporat-
ing credit and liquidity effects into market prices. The same
analysis is applied to European caps/floors, finding that the
full transition to the modern multiple-curve CSA approach
has been deferred until August 2010. Finally, we show the ro-
bustness of the SABR model to calibrate the market volatility
smile coherently with the new market evidences.
Marco Bianchetti — Market Risk Management, Intesa San Paolo
Mattia Carlicchi — Market Risk Management, Intesa San Paolo1
The authors gratefully acknowledge fruitful interactions with A. Battauz, 1
A. Castagna, C. C. Duminuco, F. Mercurio, M. Morini, M. Trapletti, and
colleagues at Market Risk Management and Fixed Income trading desks.
36
The financial crisis that began in the second half of 2007 has triggered,
among its many other implications, a deep evolution phase of the clas-
sical framework adopted for trading derivatives. In particular, credit and
liquidity issues were found to have macroscopical impacts on the prices
of financial instruments, both plain vanillas and exotics. These are clearly
visible in the market quotes of plain vanilla interest rate derivatives, such
as deposits, forward rate agreements (FRA), swaps (IRS), and options
(caps, floors, and swaptions). Since August 2007 the primary interest
rates of the interbank market, such as Libor, Euribor, Eonia, and Federal
Funds rate2, display large basis spreads that have reached up to 200 ba-
sis points. Similar divergences are also found between FRA rates and the
forward rates implied by two consecutive deposits, and similarly, among
swap rates with different floating leg tenors. Recently, the market has
also included the effect of collateral agreements widely diffused among
derivatives counterparties in the interbank market.
After the market evolution the standard no-arbitrage framework adapted
to price derivatives, developed over forty years following the Copernican
revolution of Black and Scholes (1973) and Merton (1973), became obso-
lete. Familiar relations described in standard textbooks [see, for example,
Brigo and Mercurio (2006), Hull (2010)], such as the basic definition of
forward interest rates, or the swap pricing formula, had to be abandoned.
Also the fundamental idea of the construction of a single risk free yield
curve, reflecting at the same time the present cost of funding of future
cash flows and the level of forward rates, has been ruled out. The finan-
cial community has thus been forced to start the development of a new
theoretical framework, including a larger set of relevant risk factors, and
to review from scratch the no-arbitrage models used on the market for
derivatives pricing and risk analysis. We refer to such old and new frame-
works as “classical” and “modern,” respectively, to remark the shift of
paradigm induced by the crisis.
The topics discussed in this paper sit at the heart of the present deriva-
tive’s market, with many consequences for trading, financial control, risk
management, and IT, and are attracting a growing attention in the finan-
cial literature. To our knowledge, they have been approached by Kijima et
al. (2008), Chibane and Sheldon (2009), Ametrano and Bianchetti (2009),
Ametrano (2011), Fujii et al. (2009a, 2010a, 2011) in terms of multiple-
curves; by Henrard (2007, 2009) and Fries (2010) using a first-principles
approach; by Bianchetti (2010) using a foreign currency approach; by Fujii
et al. (2009b), Mercurio (2009, 2010a, 2010b) and Amin (2010) within the
Libor Market Model; by Pallavicini and Tarenghi (2010) and Moreni and
Pallavicini (2010) within the HJM model; by Kenyon (2010) using a short
rate model; by Morini (2009) in terms of counterparty risk; by Burghard
and Kjaer (2010), Piterbarg (2010a, 2010b), Fujii et al. (2010b), Morini and
Prampolini (2010) in terms of cost of funding.
Market evolutionIn this section we discuss the most important market data showing the
main consequences of the credit crunch, which started in August 2007.
We will focus, in particular, on euro interest rates, since they show rather
peculiar and persistent effects that have strong impacts on pricing meth-
odologies. The same results hold for other currencies, USDLibor and
Federal Funds rates in particular [Mercurio (2009, 2010b)].
Euribor – OIS basisFigure 1 reports the historical series of the Euribor deposit six-month
(6M) rate, of the Eonia overnight indexed swap3 (OIS) six-month (6M) rate,
and the iTtraxx financial senior CDS 5Y index value for the European
market over the time interval Jan. 06 – Dec. 10. Before August 2007, the
two interbank rates (Euribor and Eonia) display strictly overlapping trends
differing by no more than 6 bps. In August 2007, we observe a sudden
increase of the Euribor rate and a simultaneous decrease of the OIS rate
that lead to the explosion of the corresponding basis spread, touching
the peak of 222 bps in October 2008, when Lehman Brothers filed for
bankruptcy protection. Successively the basis has sensibly reduced and
stabilized between 40 bps and 60 bps. Notice that the pre-crisis level
has never been recovered. The same effect is observed for other similar
couples, such as Euribor 3M versus OIS 3M. The abrupt divergence be-
tween the Euribor and OIS rates can be explained by considering both
the monetary policy decisions adopted by international authorities in re-
sponse to the financial turmoil and the impact of the credit crunch on the
credit and liquidity risk perception of the market, as well as the different
financial meaning and dynamics of these rates.
The Euribor rate is the reference rate for over-the-counter (OTC) ■■
transactions in the Euro Area. It is defined as “the rate at which euro
interbank deposits are being offered within the EMU zone by one
prime bank to another at 11:00 a.m. Brussels time.” The rate fixings
for a strip of 15 maturities, ranging from one day to one year, are
constructed as the trimmed average of the rates submitted (excluding
the highest and lowest 15 % tails) by a panel of banks. The contribu-
tion panel is composed, as of September 2010, of 42 banks, selected
among the E.U. banks with the highest volume of business in the
Eurozone money markets, plus some large international banks from
non-E.U. countries with important Eurozone operations. Thus, Euribor
Libor, sponsored by the British Banking Association (BBA), is quoted in all the major 2
currencies and is the reference rate for international over-the-counter (OTC) transactions
(see www.bbalibor.com). Euribor and Eonia, sponsored by the European Banking
Federation (EBF), are the reference rates for OTC transactions in the euro market (see
www.euribor.org). The Federal Funds rate is a primary rate of the USD market and is set by
the Federal Open Market Committee (FOMC) according to the monetary policy decisions of
the Federal Reserve (FED).
The overnight index swap (OIS) is a swap with a fixed leg versus a floating leg indexed to 3
the overnight rate. The euro market quotes a standard OIS strip indexed to Eonia rate (daily
compounded) up to 30 years maturity.
37
The Capco Institute Journal of Financial TransformationInterest Rates After the Credit Crunch: Markets and Models Evolution
rates reflect the average cost of funding of banks in the interbank mar-
ket at each given maturity. During the crisis, the solvency and solidity
of the entire financial sector was brought into question and the credit
and liquidity risk and premia associated with interbank counterpar-
ties increased sharply. The Euribor rates immediately reflected these
dynamics and raise to their highest values for more than a decade. As
seen in Figure 1, the Euribor 6M rate suddenly increased on August
2007 and reached 5.49 percent on 10th October 2008.
The Eonia rate is the reference rate for overnight OTC transactions in ■■
the Euro Area. It is constructed as the average rate of the overnight
transactions (one day maturity deposits) executed during a given
business day by a panel of banks on the interbank money market,
weighted with the corresponding transaction volumes. The Eonia
contribution panel coincides with the Euribor contribution panel. Thus
the Eonia rate includes information about the short-term (overnight)
liquidity expectations of banks in the euro money markets. It is also
used by the European Central Bank (ECB) as a method of effecting
and observing the transmission of its monetary policy actions. During
the crisis, the central banks were mainly concerned with stabilizing
the level of liquidity in the market, thus they reduced the level of the
official rates: the “deposit facility rate” and the “marginal lending facil-
ity rate.” Empirical evidence suggests that the Eonia rate is always
higher than the deposit facility rate and lower than the marginal lend-
ing facility rate, defining the so-called “rates corridor.” Furthermore,
the daily tenor of the Eonia rate makes the credit and liquidity risks
reflected in it negligible. For this reason the OIS rates are considered
the best proxies available in the market for the risk-free rate.
Thus the Euribor-OIS basis explosion of August 2007 plotted in Figure 1
is essentially a consequence of the different credit and liquidity risk re-
flected by Euribor and Eonia rates. We stress that such a divergence is
not a consequence of the counterparty risk carried by the financial con-
tracts, deposits and OISs, exchanged in the interbank market by risky
counterparties, but depends on the different fixing levels of the underly-
ing Euribor and Eonia rates. The different influences of credit risk on Libor
and overnight rates can also be appreciated in Figure 1 by comparing the
historical series for the Euribor-OIS spread with the iTraxx financial senior
CDS 5Y index value for the European market, which represents the level
of the CDS spread related to a generic European primary bank. We ob-
serve that the Euribor-OIS basis explosion of August 2007 matches the
CDS explosion, corresponding to the generalized increase of the default
risk seen in the interbank market.
The liquidity risk component in Euribor and Eonia interbank rates is dis-
tinct but strongly correlated to the credit risk component. According to
Acerbi and Scandolo (2007), liquidity risk may appear in at least three
circumstances: when there is a lack of liquidity to cover short term-debt
obligations (funding liquidity risk), when it is difficult to liquidate assets on
the market due to excessive bid-offer spreads (market liquidity risk), and
when it is difficult to borrow funds on the market due to excessive fund-
ing cost (systemic liquidity risk). According to Morini (2009), these three
elements are, in principle, not a problem so long as they do not appear
together, because a bank with, for instance, problem 1 and 2 (or 3) will
be able to finance itself by borrowing funds (or liquidating assets) on the
market. During the crisis these three scenarios manifested themselves
jointly at the same time, thus generating a systemic lack of liquidity [Mi-
chaud and Upper (2008)].
Clearly, it is difficult to disentangle liquidity and credit risk components
in the Euribor and Eonia rates, because, in particular, they do not refer
to the default risk of one counterparty in a single derivative deal but to a
money market with bilateral credit risk [Morini (2009)].
Finally, we stress that, as seen in Figure 1, the Libor-OIS basis is still per-
sistent today at a non-negligible level, despite the lower rate and higher
liquidity regime reached after the most acute phase of the crisis and the
strong interventions of central banks and governments. Clearly the mar-
ket has learnt the lesson of the crisis and has not forgotten that these
interest rates are driven by different credit and liquidity dynamics. From
an historical point of view, we can compare this effect to the appearance
of the volatility smile on the option markets after the 1987 crash [Derman
and Kani (1994)]. It is still there.
FRA rates versus forward ratesThe considerations above, with reference to spot Euribor and Eonia rates
underlying deposit and OIS contracts, also apply to forward rates and FRA
rates. The FRA 3x6 rate is the equilibrium (fair) rate of a FRA contract start-
ing at the spot date (today plus two working days in the euro market), ma-
turing in six months, with a floating leg indexed to the forward interest rate
between three and six months, versus a fixed interest rate leg. The paths
-50
0
50
100
150
200
250
0%
1%
2%
3%
4%
5%
6%
7%
2/01/06
2/07/06
2/01/07
2/07/07
2/01/08
2/07/08
2/01/09
2/07/09
2/01/10
2/07/10
Spread (bps) Rate (%)
Euribor deposit 6M
Eonia OIS 6M
iTraxx Fin. CDS 5Y Ind.
Euribor deposit 6M - Eonia OIS 6M spread
The spread between Euribor deposit 6M and Eonia OIS 6M is shown on the right axis
(Jan. 06 – Dec. 10 window, source: Bloomberg)
Figure 1 – Historical series of Euribor deposit 6M rate, Eonia OIS 6M rate, and iTraxx senior financial CDS 5Y index value for the European market.
38
of market FRA rates and of the corresponding forward rates implied in two
consecutive Eonia OIS deposits are similar to those observed in Figure 1 for
the Euribor deposit and Eonia OIS respectively. In particular, a sudden diver-
gence between the quoted FRA rates and the implied forward rates arose
in August 2007, regardless of the maturity, and reached its peak in October
2008 with the Lehman crash. Mercurio (2009) has proven that the effects
above may be explained within a simple credit model with a default-free
zero coupon bond and a risky zero coupon bond issued by a defaultable
counterparty with recovery rate R. The associated risk free and risky Libor
rates are the underlyings of the corresponding risk free and risky FRAs.
Basis swapsA third evidence of the regime change after the credit crunch is the ex-
plosion of the basis swaps spreads. Basis swaps are quoted on the euro
interbank market in terms of the difference between the fixed equilibrium
swap rates of two swaps. For example, the quoted Euribor 3M versus
Euribor 6M basis swap rate is the difference between the fixed rates of
a first standard swap with a Euribor 3M floating leg (quarterly frequency)
versus a fixed leg (annual frequency), and of a second swap with a Eu-
ribor 6M floating leg (semi-annual frequency) versus a fixed leg (annual
frequency). The frequency of the floating legs is the “tenor” of the cor-
responding Euribor rates. The Eonia rate has the shortest tenor (one day).
The basis swap spreads were negligible (or even not quoted) before the
crisis. They suddenly diverged in August 2007 and peaked in October
2008 with the Lehman crash.
The basis swap involves a sequence of spot and forward rates carrying
the credit and liquidity risk discussed above. Hence, the basis spread ex-
plosion can be interpreted, in principle, in terms of the different credit and
liquidity risks carried by the underlying Libor rates with different tenor.
From the findings described above we understand that, after the crisis,
market players have a preference for receiving floating payments with
higher frequency (i.e., 3M) indexed to lower tenor Euribor rates (i.e., Euri-
bor 3M), with respect to floating payments with lower frequency (i.e., 6M)
indexed to higher tenor Euribor rates (i.e., Euribor 6M), and are keen to
pay a premium for the difference. Hence in a basis swap (i.e., 3M versus
6M), the floating leg indexed to the higher rate tenor (6M) must include
a risk premium higher than that included in the floating leg indexed to
the shorter rate tenor (3M, both with the same maturity). Thus, a positive
spread emerges between the two corresponding equilibrium rates (or, in
other words, a positive spread must be added to the 3M floating leg to
equate the value of the 6M floating leg).
According to Morini (2009), a basis swap between two interbank coun-
terparties under collateral agreement can be described as the difference
between two investment strategies. Fixing, for instance, a basis swap
Euribor 3M versus Euribor 6M with 6M maturity, scheduled on three dates
T0, T1=T0+3M, T2=T0+6M, we have the following two strategies:
1 6M floating leg – at T0 choose a counterparty C1 with a high credit
standing (that is, belonging to the Euribor contribution panel) with
collateral agreement in place, and lend the notional for six months
at the Euribor 6M rate prevailing at T0 (Euribor 6M flat because C1 is
an Euribor counterparty). At maturity T2 recover notional plus interest
from C1. Notice that if counterparty C1 defaults within six months we
gain full recovery thanks to the collateral agreement.
2 3M+3M floating leg – at T0 choose a counterparty C1 with a high
credit standing (belonging to the Euribor contribution panel) with col-
lateral agreement in place, and lend the notional for three months at
the Euribor 3M rate (flat) prevailing at T0. At T1 recover notional plus
interest and check the credit standing of C1: if C1 has maintained its
credit standing (it still belongs to the Euribor contribution panel), then
lend the money again to C1 for three months at the Euribor 3M rate
(flat) prevailing at T1, otherwise choose another counterparty C2 be-
longing to the Euribor panel with collateral agreement in place, and
lend the money to C2 at the same interest rate. At maturity T2 recover
notional plus interest from C1 or C2. Again, if counterparties C1 or C2
default within six months we gain full recovery thanks to the collateral
agreements.
Clearly, the 3M+3M leg implicitly embeds a bias towards the group of
banks with the best credit standing, typically those belonging to the Euri-
bor contribution panel. Hence the counterparty risk carried by the 3M+3M
leg must be lower than that carried by the 6M leg. In other words, the
expectation of the survival probability of the borrower of the 3M leg in the
second 3M-6M period is higher than the survival probability of the borrow-
er of the 6M leg in the same period. This lower risk is embedded into lower
Euribor 3M rates with respect to Euribor 6M rates. But with collateralization
the two legs have both null counterparty risk. Thus a positive spread must
be added to the 3M+3M leg to reach equilibrium. The same discussion can
be repeated, mutatis mutandis, in terms of liquidity risk. We stress that the
credit and liquidity risk involved here are those carried by the risky Libor
rates underlying the basis swap, reflecting the average default and liquidity
risk of the interbank money market (of the Libor panel banks), not those
associated to the specific counterparties involved in the financial contract.
We stress also that such effects were already present before the credit
crunch, as discussed in Tuckman and Porfirio (2004), and well known to
market players, but not effective due to negligible basis spreads.
Collateralization and OIS-discountingAnother effect of the credit crunch has been the great diffusion of collat-
eral agreements to reduce the counterparty risk of OTC derivatives posi-
tions. Nowadays most of the counterparties on the interbank market have
mutual collateral agreements in place. In 2010, more than 70 percent of
all OTC derivatives transactions were collateralized [ISDA (2010)].
Typical financial transactions generate streams of future cash flows, whose
39
total net present value (NPV = algebraic sum of all discounted expected
cash flows) implies a credit exposure between the two counterparties.
If, for counterparty A, NPV(A)>0, then counterparty A expects to receive,
on average, future cash flows from counterparty B (in other words, A
has a credit with B). On the other hand, if counterparty B has NPV(B)<0,
then it expects to pay, on average, future cash flows to counterparty A
(in other words, B has a debt with A). The reverse holds if NPV(A)<0 and
NPV(B)>0. Such credit exposure can be mitigated through a guarantee,
called “collateral agreement,” or “credit support annex” (CSA), follow-
ing the International Swaps and Derivatives Association (ISDA) standards
widely used to regulate OTC transactions. The main feature of the CSA
is a margination mechanism similar to those adopted by central clearing
houses for standard instruments exchanges (i.e., futures). In a nutshell,
at every margination date the two counterparties check the value of the
portfolio of mutual OTC transactions and regulate the margin, adding
to or subtracting from the collateral account the corresponding mark
to market variation with respect to the preceding margination date. The
margination can be regulated with cash or with (primary) assets of cor-
responding value. In any case the collateral account holds, at each date,
the total NPV of the portfolio, which is positive for the creditor counter-
party and negative for the debtor counterparty. The collateral amount is
available to the creditor. On the other side, the debtor receives an inter-
est on the collateral amount, called “collateral rate.” Hence, we can see
the collateral mechanism as a funding mechanism, transferring liquidity
from the debtor to the creditor. The main difference with traditional fund-
ing through deposit contracts are that, using derivatives, we have longer
maturities and stochastic lending/borrowing side and amount. We can
also look at CSA as a hedging mechanism, where the collateral amount
hedges the creditor against the event of default of the debtor. The most
diffused CSA provides a daily margination mechanism and an overnight
collateral rate [ISDA (2010)]. Actual CSAs provide many other detailed
features that are beyond the scope of this paper.
Thus, a first important consequence of the diffusion of collateral agree-
ments among interbank counterparties is that we can consider the de-
rivatives prices quoted on the interbank market as counterparty riskfree
OTC transactions. A second important consequence is that, by no-arbi-
trage, the CSA margination rate and the discounting rate of future cash
flows must match. Hence the name of “CSA discounting.” In particular,
the most diffused overnight CSA implies overnight-based discounting
and the construction of a discounting yield curve that must reflect, for
each maturity, the funding level in an overnight collateralized interbank
market. Thus overnight indexed swaps (OIS) are the natural instruments
for discounting curve construction. Hence the alternative name of “OIS
discounting” or “OIS (yield) curve.” Such discounting curve is also the
best available proxy of a riskfree yield curve.
In the case of absence of CSA, using the same no-arbitrage principle
between the funding and the discounting rate, we conclude that a bank
should discount future cash flows (positive or negative) using its own
“traditional” cost of funding term structure. This has important (and rather
involved) consequences, such that, according to Morini and Prampolini
(2009), each counterparty assigns a different present value to the same
future cash flow, breaking the fair value symmetry; that a worsening of its
credit standing allows the bank to sell derivatives (options in particular) at
more competitive prices (the lower the rate, the higher the discount, the
lower the price); as well as the problem of double counting the debt value
adjustment (DVA) to the fair value.
Presently, the market is in the middle of a transition phase from the classical
Libor-based discounting methodology to the modern CSA-based method-
ology. OTC transactions executed on the interbank market normally use
CSA discounting. In particular, plain vanilla interest rate derivatives, such
as FRA, swaps, basis swaps, caps/floor/swaptions are quoted by the main
brokers using CSA discounting [ICAP (2010)]. However, presently just a
few banks have declared full adoption of CSA discounting also for balance
sheet revaluation and collateral margination [Bianchetti (2011)].
Finally, we stress that before the crisis the old-style standard Libor curve
was representative of the average funding level on the interbank market
[Hull (2010). Such curve, even if considered a good proxy for a riskfree
curve, thanks to the perceived low counterparty risk of primary banks
(belonging to the Libor contribution panel), was not strictly riskfree be-
cause of the absence of collateralization.
Modeling evolutionAccording to Bianchetti and Morini (2010), the market “frictions” dis-
cussed above have created a sort of “segmentation” of the interest rate
market into sub-areas, mainly corresponding to instruments with 1M, 3M,
6M, 12M underlying rate tenors. These are characterized, in principle, by
different internal dynamics, liquidity, and credit risk premia, reflecting the
different views and interests of the market players. In response to the
crisis, the classical pricing framework, based on a single yield curve used
to calculate forward rates and discount factors, has been abandoned,
and a new modern pricing approach has prevailed among practitioners.
The new methodology takes into account the market segmentation as an
empirical evidence and incorporates the new interest rate dynamics into
a multiple curve framework as follows:
■■ Discounting curves – these are the yield curves used to discount
futures cash flows. As discussed above, the curve must be con-
structed and selected such that it reflects the cost of funding of the
bank in connection with the actual nature of the specific contract that
generates the cash flows. In particular, an OIS-based curve is used
to discount cash flow generated by a contract under CSA with daily
margination and overnight collateral rate; a funding curve is used in
The Capco Institute Journal of Financial TransformationInterest Rates After the Credit Crunch: Markets and Models Evolution
40
case of contracts without CSA; and in the case of non-standard CSA
(i.e., different margination frequency, rate, threshold, etc.), appropri-
ate curves should be, in principle, selected (we will not discuss this
topic here since it applies to a minority of deals and it would be
beyond of the scope of the present paper). We stress that the funding
curve for no-CSA contracts is specific to each counterparty, that will
have its specific funding curve. This modern discounting methodology
is called CSA-discounting.
Forwarding curves –■■ these are the yield curves used to compute
forward rates. The curve must be constructed and selected according
to the tenor and typology of the rate underlying the actual contract
to be priced. For example, a swap floating leg indexed to Euribor
6M requires a Euribor 6M forwarding curve constructed from quoted
instruments with Euribor 6M underlying rate.
Following Bianchetti (2010), we report in Table 1 the comparison between
the classical and the modern frameworks, called single-curve and multi-
ple-curve approach, respectively. The adoption of the multiple-curve ap-
proach has led to the revision of no-arbitrage pricing formulas. According
to Mercurio (2009, 2010a, 2010b), we compare in Table 2 the classical
and modern pricing formulas for plain vanilla interest rate derivatives. We
stress that the fundamental quantity of the modern pricing framework
is the FRA rate F̃. Indeed, following Mercurio (2009, 2010a, 2010b), the
correct probability measure to be used in expectations is that associated
with the discounting curve Cd, under which the forward rate is no longer
a martingale. Instead, the FRA rate, by definition, is a martingale under
such measure.
Empirical pricing analysisIn the following sections we present the results of an empirical analysis
comparing the results of the three pricing frameworks described above
against market quotations of plain vanilla interest rate derivatives at two
different valuation dates. The aim of this analysis is to highlight the time
evolution of the market pricing approach as a consequence of the finan-
cial crisis.
Market data, volatility surfaces and yield curvesThe reference market quotes that we considered are, in particular, euro
forward start interest rate swap contracts (FSIRS): market swap rates
based on Euribor 6M, published by Reuters, and euro cap/floor European
options: market premia and Black implied volatility surfaces based on
Euribor 6M, published by Reuters.
Classical methodology (single-curve) Modern methodology (multiple-curve)
Yield curves
construction
Select a single finite set of the most convenient (i.e., liquid) vanilla interest
rate market instruments and build a single yield curve C using the preferred
bootstrapping procedure. For example, a common choice in the European
market is a combination of short-term EUR deposits, medium-term Futures/FRAs
on Euribor 3M, and medium-long-term swaps on Euribor 6M.
Build one discounting curve Cd using the preferred selection of vanilla interest
rate market instruments and bootstrapping procedure.
Build multiple distinct forwarding curves Cx using the preferred selections of
distinct sets of vanilla interest rate market instruments, each homogeneous in
the underlying rate tenor (typically x = 1M, 3M, 6M, 12M) and bootstrapping
procedures. For example, for the construction of the forwarding curve C6M only
market instruments with six-month tenor are considered.
Computation of
expected cash flows
For each interest rate coupon compute the relevant forward rates using the given
yield curve C and applying the standard formula, with t≤Tk-1≤Tk:
where τk is the year fraction related to the time interval [Tk-1,Tk].
Compute cash flows as expectations at time t of the corresponding coupon
payoffs with respect to the Tk-forward measure QTk , associated to the numeraire
P(t,Tk) from the same yield curve C:
For each interest rate coupon compute the relevant FRA rate F̃ x,k (t) with tenor x
using the corresponding forwarding curve Cx and applying the following formula,
with t≤Tk-1≤Tk:
where τx,k is the year fraction associated to the time interval [Tk-1,Tk].
Compute cash flows as expectations at time t of the corresponding coupon
payoffs with respect to the discounting Tk-forward measure QdTk, associated to
the numeraire Pd (t,Tk) from the discounting curve Cd:
Computation of
discount factors
Compute the relevant discount factors P(t,Tk) from the unique yield curve C
defined in step 1.
Compute the relevant discount factors Pd (t,Tk) from the discounting curve Cd of
step 1.
Computation of the
derivative’s price
Compute the derivative’s price at time t as the sum of the discounted expected
future cash flows:
Compute the derivative’s price at time t as the sum of the discounted expected
future cash flows:
We refer to a general single-currency interest rate derivative under CSA characterized by m future coupons with payoffs π = {π1, …, πm}, generating m cash flows c = {c1, …, cm} at future dates T =
{T1, …, Tm} with t <T1 < ···<Tm.
Table 1 – Comparison table between the classical single-curve methodology and the modern multiple-curve methodology
41
All the market data were archived at close of business (17:30 CET) on 31st
March and 31st August 2010, and refer to instruments traded among col-
lateralized counterparties in the European interbank market. Moreover,
for the pricing analysis we defined the following yield curves:
Euribor standard –■■ the classical yield curve bootstrapped from short-
term euro deposits (from 1D to the first Futures), mid-term futures on
Euribor 3M (below 3Y) and mid-long term swaps on Euribor 6M (after
3Y) (Table 1, left column).
Euribor 6M Standard –■■ the “pure” Euribor 6M forwarding yield curve
is bootstrapped from the euro deposit 6M, mid-term FRAs on Euribor
6M (below 3Y), and mid-long term swaps on Euribor 6M (after 3Y). The
discounting curve used in the bootstrapping procedure is the Euribor
Standard. This curve differs from the one above on the short- to mid-
term, while after the three-year node it tends to coincide with the
Euribor standard since the bootstrapping instruments are the same.
Eonia OIS –■■ the modern discounting yield curve bootstrapped from
quoted Eonia OIS. This yield curve usually presents lower rates and,
hence, greater discount factors than those obtained with the Euribor
standard above.
Euribor 6M CSA –■■ the modern Euribor 6M forwarding yield curve
bootstrapped from the same market instruments as the Euribor 6M
standard curve, but using the Eonia OIS as discounting curve. This
curve usually presents small, but not negligible, differences (±2 bps)
with respect to the Euribor 6M Standard.
Pricing methodologiesWe have tested three different pricing methodologies as described below.
Single-curve approach –■■ we use the Euribor standard yield curve
to calculate both the discount factors P(t, Tk) and the forward rates
F needed for pricing any interest rate derivatives. This is the classical
single-curve methodology adopted by the market before the credit
crunch, without collateral, credit, and liquidity effects.
Multiple-curve no-CSA approach –■■ we calculate discount factors
Pd(t, Tk) on the Euribor standard curve and FRA rates F̃x,k (t) on the
Euribor 6M standard curve. This is the “quick and dirty” methodology
adopted by the market in response to the credit crunch after August
2007, which distinguishes between discounting and forwarding
curves. It is defined “no-CSA” because it does not include the effect
of collateral. Indeed, the Euribor standard discounting curve reflects
the average cost of (uncollateralized) funding of a generic European
interbank counterparty (belonging to the Euribor panel). Also the
Euribor 6M standard forwarding curve construction does not take into
account collateralization, but it does include the tenor-specific credit
and liquidity risk of the underlying Euribor 6M rate.
The Capco Institute Journal of Financial TransformationInterest Rates After the Credit Crunch: Markets and Models Evolution
Classical approach (single-curve) Modern approach (multiple-curve)
FRA
with with
Sw
ap with with
Bas
is s
wap
with with
Cap
/flo
or
with with
Table 2 – Comparison table of classical and modern formulas for pricing plain vanilla derivatives. In red we emphasize the most relevant peculiarities of the multiple-curve method
42
Multiple-curve CSA approach –■■ we calculate discount factors Pd(t,
Tk) on the Eonia OIS curve and FRA rates F̃x,k (t) on the Euribor 6M
CSA curve. This is the “state of the art” modern methodology, fully
coherent with the CSA nature of the interest rate derivatives consid-
ered and with the credit and liquidity risk of the underlying Euribor
6M rate.
The arbitrage-free formulas used in the analysis are those reported in
Table 2: the single-curve approach is in the left column and the two mul-
tiple-curve approaches are on the right column. The two following sec-
tions report the findings of the analysis.
Forward start interest rate swapsThe forward start interest rate swap contracts considered here are char-
acterized by a floating leg on Euribor 6M with six-month frequency versus
a fixed leg with annual frequency, a forward start date and maturity dates
ranging from 1 to 25 years. We selected forward start instead of spot start
swaps because the former are more sensible to the choice of the pric-
ing methodology. For each methodology and each valuation date (31st
March and 31st August 2010) we computed the theoretical equilibrium
FSIRS rates and we compared them with the market quotes. In Figure 2
below (left side graphs) we report the result for the valuation date 31st
March 2010, while in Table 3 we compare the most important numbers:
the range of minimum and maximum discrepancies and the standard de-
viation. The spikes observed in the graphs for start dates below 3Y may
be explained in terms of differences in the short-term yield curve con-
struction, where there is a significant degree of freedom in choosing the
bootstrapping instruments (deposits, FRAs, and futures). Smaller spikes
are also present for short tenor FSIRS with maturities below 3Y because
these swaps depend on a few forwards and discounts and, thus, are
more sensitive to minor differences in the yield curves. Hence, in Table
3 we present the result including and excluding the two “stripes” below
2 years start/maturity date and considering both the valuation dates (31st
March and 31st August 2010).
From Figure 2 we observe that the first methodology has the worst per-
formance, producing, on average, overestimated FSIRS rates. The sec-
ond methodology introduces small improvements, at least below three
years. This is expected, because the two curves used are very similar
after three years, both using standard euro swaps on Euribor 6M. The
third methodology is by far the best in reproducing the market data. The
remaining differences of around 1 basis points may be explained with
minor differences with respect to the market yield curves. The same ob-
servations apply to the results of 31st August 2010.
We conclude that the market of interest rate swaps since March 2010
has abandoned the classical single-curve pricing methodology, typical
of the pre-credit crunch interest rate world, and has adopted the modern
multiple-curve CSA approach, thus incorporating into market prices the
credit and liquidity effects described above.
Cap/floor optionsThe European cap/floor options considered here are characterized by
floating payments with 6M frequency indexed to Euribor 6M, spot start
date, maturity dates ranging from 3 to 30 years, and strikes ranging from
1 percent to 10 percent. The first caplet/floorlet, already known at spot
date, is not included in the cap/floor premium. The market quotes floor
premia for strikes below the at-the-money (ATM) and cap premia for
strikes above ATM. For each methodology and each valuation date (31st
March and 31st August 2010) we computed the theoretical cap/floor pre-
mia and we compared them with the market premia.
The computation of the theoretical European cap/floor premia using
the standard Black’s formula in Table 2 requires two inputs: the pair of
discounting and forwarding curves and the Black implied term volatility
[Mercurio (2009, 2010b)]. Even if the market-driven quantity is the pre-
mium of the traded option, it is standard market convention to quote the
option in terms of its Black implied term volatility. Clearly, once the pre-
mium is fixed by the market supply and demand, the value of this volatil-
ity depends on the curves used for discounting and forwarding. Thus, a
change in the market yield curves implies a corresponding change in the
Black implied volatilities. Actually, in August 2010 the market began to
quote two distinct volatility surfaces: one implied using the classical Euri-
bor discounting curve and one implied using the modern Eonia discount-
ing curve [ICAP (2010)]. The Eonia implied volatility is generally lower
than the Euribor implied volatility, since the effect of lower Eonia rates
Forward start interest rate swaps differences
31st March 2010 31st August 2010
Range Standard deviation Range Standard deviation
Single-curve [-18.4;+20.8] [-3.2;+2.7] 2.84 1.89 [-16.3;+24.4] [-3.9;+1.9] 2.58 1.15
Multiple-curve no-CSA [-2.9;+3.1] [-2.9;+2.6] 1.77 1.86 [-5.7;+2.9] [-3.7;+1.7] 1.11 1.09
Multiple-curve CSA [-2.9;+2.3] [-1.0;+1.5] 0.53 0.37 [-4.1;+2.4] [-1.4;+1.0] 0.47 0.26
Table 3 – FSIRS differences (in basis points) on 31st March and 31st August 2010
43
The Capco Institute Journal of Financial TransformationInterest Rates After the Credit Crunch: Markets and Models Evolution
FSIRS rates differences: Single-curve versus market (31 Mar 2010) Cap/floor premia differences: single-curve versus market (31 Aug 2010)
1Y
5Y
9Y
13Y
25Y -6
-4
-2
0
2
4
1Y
3Y
5Y
7Y
9Y
11Y
13Y
15Y
25Y Forward start
Diff
eren
ce (b
ps)
Maturity
1Y
5Y
9Y
13Y
25Y -20
-15
-10
-5
0
5
10
15
20
25
1Y
3Y
5Y
7Y
9Y
11Y
13Y
15Y
25Y Forward start
Diff
eren
ce (b
ps)
Maturity
1Y
5Y
9Y
13Y
25Y -6
-4
-2
0
2
4
1Y
3Y
5Y
7Y
9Y
11Y
13Y
15Y
25Y Forward start
Diff
eren
ce (b
ps)
Maturity
1,0%
2,0%
2,5%
3,5%
4,5%
6,0%
10,0%
-10
-5
0
5
10
15
20
3Y
5Y 7Y
9Y 12Y
20Y
30Y Strike
Diff
eren
ce (b
ps)
Maturity
1,0%
2,0%
2,5%
3,5%
4,5%
6,0%
10,0%
-10
-5
0
5
10
15
20
3Y
5Y 7Y
9Y 12Y
20Y
30Y Strike
Diff
eren
ce (b
ps)
Maturity
1,0%
2,0%
2,5%
3,5%
4,5%
6,0%
10,0%
-10
-5
0
5
10
15
20
3Y
5Y 7Y
9Y 12Y
20Y
30Y Strike
Diff
eren
ce (b
ps)
Maturity
FSIRS rates differences: multiple-curve no-CSA versus market (31 Mar 2010) Cap/Floor premia differences: multiple-curve no-CSA versus market (31 Aug 2010)
1Y
5Y
9Y
13Y
25Y -6
-4
-2
0
2
4
1Y
3Y
5Y
7Y
9Y
11Y
13Y
15Y
25Y Forward start
Diff
eren
ce (b
ps)
Maturity
1Y
5Y
9Y
13Y
25Y -20
-15
-10
-5
0
5
10
15
20
25
1Y
3Y
5Y
7Y
9Y
11Y
13Y
15Y
25Y Forward start
Diff
eren
ce (b
ps)
Maturity
1Y
5Y
9Y
13Y
25Y -6
-4
-2
0
2
4 1Y
3Y
5Y
7Y
9Y
11Y
13Y
15Y
25Y Forward start
Diff
eren
ce (b
ps)
Maturity
1,0%
2,0%
2,5%
3,5%
4,5%
6,0%
10,0%
-10
-5
0
5
10
15
20
3Y
5Y 7Y
9Y 12Y
20Y
30Y Strike
Diff
eren
ce (b
ps)
Maturity
1,0%
2,0%
2,5%
3,5%
4,5%
6,0%
10,0%
-10
-5
0
5
10
15
20
3Y
5Y 7Y
9Y 12Y
20Y
30Y Strike
Diff
eren
ce (b
ps)
Maturity
1,0%
2,0%
2,5%
3,5%
4,5%
6,0%
10,0%
-10
-5
0
5
10
15
20
3Y
5Y 7Y
9Y 12Y
20Y
30Y Strike
Diff
eren
ce (b
ps)
Maturity
FSIRS rates differences: multiple-curve CSA versus market (31 Mar 2010) Cap/floor premia differences: multiple-curve CSA versus market (31 Aug 2010)
1Y
5Y
9Y
13Y
25Y -6
-4
-2
0
2
4
1Y
3Y
5Y
7Y
9Y
11Y
13Y
15Y
25Y Forward start
Diff
eren
ce (b
ps)
Maturity
1Y
5Y
9Y
13Y
25Y -20
-15
-10
-5
0
5
10
15
20
25
1Y
3Y
5Y
7Y
9Y
11Y
13Y
15Y
25Y Forward start
Diff
eren
ce (b
ps)
Maturity
1Y
5Y
9Y
13Y
25Y -6
-4
-2
0
2
4
1Y
3Y
5Y
7Y
9Y
11Y
13Y
15Y
25Y Forward start
Diff
eren
ce (b
ps)
Maturity
1,0%
2,0%
2,5%
3,5%
4,5%
6,0%
10,0%
-10
-5
0
5
10
15
20
3Y
5Y 7Y
9Y 12Y
20Y
30Y Strike
Diff
eren
ce (b
ps)
Maturity
1,0%
2,0%
2,5%
3,5%
4,5%
6,0%
10,0%
-10
-5
0
5
10
15
20
3Y
5Y 7Y
9Y 12Y
20Y
30Y Strike
Diff
eren
ce (b
ps)
Maturity
1,0%
2,0%
2,5%
3,5%
4,5%
6,0%
10,0%
-10
-5
0
5
10
15
20
3Y
5Y 7Y
9Y 12Y
20Y
30Y Strike
Diff
eren
ce (b
ps)
Maturity
Panels on the left: differences between theoretical FSIRS equilibrium rates and market quotes. Panels on the right: cap/floor options premia differences (light colors: floors, dark colors: caps). Upper
panels: they report the results obtained through the single-curve approach. Middle panels: they report the results obtained through the multiple-curve no-CSA approach. Lower panels: they report
the results obtained through the multiple-curve CSA approach. Valuation dates: 31st March 2010 for the FSIRS (left side graphs) and 31st August 2010 for cap/floor options (right side graphs). Note
that the y-axis scales of the middle and lower graphs on the left hand side have been magnified to better highlight lower price differences than the ones of the upper left graph (source: Reuters).
Figure 2 – FSIRS and cap/floor differences (in basis point) from market data
44
(higher Eonia discount factors) must be compensated with lower values
of implied volatility. In conjunction with the market change, we used the
Euribor volatility to compute multiple-curve no-CSA prices and the Eonia
volatility to compute multiple-curve CSA prices at 31st August 2010. The
results for the valuation date 31st August 2010 are shown in Figure 2
and Table 4. We do not graphically report the results on 31st March 2010
because they are not necessary for the purpose of the analysis; all the
relevant numbers are contained in Table 4.
Overall, we notice again that, on both dates, the single-curve methodology
(upper panels) has a very bad performance. The multiple-curve no-CSA
methodology (middle panels) has a good performance on both dates, with
an absolute average difference of 1.4/1.6 bps over a total of 169 options
and a standard deviation of 2.06/2.28 bps. Finally the multiple-curve CSA
methodology (lower panels) shows a bad performance on the first date
(standard deviation 15.82 bps) and a performance as good as that of the
multiple-curve CSA methodology on the second date, with absolute aver-
age difference of 1.7 bps and standard deviation of 2.43 bps.
We conclude that the results discussed above are coherent with the inter-
est rate market evolution after the credit crunch and, in particular, with the
market changes announced in August 2010 [ICAP (2010)] and with our
findings for forward start IRS discussed above. First of all, the market, at
least since March 2010, has abandoned the classical single-curve pricing
methodology, typical of the pre-credit crunch interest rate world, and has
adopted the modern multiple-curve approach. Second, the transition to
the CSA-discounting methodology for options has happened just in Au-
gust 2010, thus incorporating into market prices the credit and liquidity
effects described above. In the latter case, contrary to FSIRS, both the
two modern multiple-curve methodologies (if correctly applied) lead to
good repricing of the market premia, because, at constant market pre-
mia, the change in the yield curves (switching from Euribor discounting
to Eonia discounting) are compensated by the corresponding changes in
the Black implied volatilities.
SABR model calibrationThe SABR (stochastic alpha beta rho) model developed by Hagan et al.
(2002) is one of the simplest generalizations of the well-known Black’s
model with stochastic volatility, preserving Black-like closed formulas for
caps, floors, and swaptions, leading to a market coherent description
of the dynamics of the volatility and allowing calibration to the interest
rate smile. Thanks to its mathematical simplicity and transparent financial
interpretation, it imposed itself as the market standard for pricing and
hedging plain vanilla interest rate options and to calibrate the market
volatility, often called “SABR volatility surface” (for caps/floors) or “cube”
(for swaptions).
Similarly to the Black model, the modern version of the SABR model
is obtained from the corresponding classical SABR version of Hagan et
al. (2002) just by replacing the classical forward rate with the modern
FRA rate and the TK-forward Libor measure associated with the classi-
cal single-curve numeraire P(t, Tk) with the modern TK-forward measure
Cap/floor premia differences
31st March 2010 31st August 2010
Range Standard
deviation
Range Standard
deviation
Single-curve [-5.8;+14.1] 6.3 [+0.2;+20.0] 9.7
Multiple-curve no-CSA [-7.0;+5.8] 2.1 [-6.3;+7.4] 2.3
Multiple-curve CSA [-8.9;+77.7] 15.8 [-6.8;+9.6] 2.4
Table 4 – Differences (in basis points) from Figure 2
Classical SABR (single-curve) Modern SABR (multiple-curve)
SABR dynamics
SABR volatility
Table 5 – Classical (top left column) versus modern (top right column) SABR model dynamics and volatility expression consistent with the multiple-curve approach (bottom) [Hagan et al. (2002)]
45
associated with the discounting numeraire Pd(t, Tk). The SABR volatility
formula remains unchanged, but takes the FRA rate as input. Cap/floor
options are priced as in Table 2 using the standard Black’s formula and
input SABR volatility. In Table 5 we show the classical and the modern
SABR equations.
Using different multiple-curve pricing methodologies, based on different
choices of discounting and forwarding yield curves, leads to the defini-
tion of two distinct implied volatility surfaces referring to the same col-
lateralized market premia, as discussed above: the Euribor implied term
volatility, which is consistent with the multiple-curve no-CSA approach;
and the Eonia implied term volatility, which is consistent with the multi-
ple-curve CSA approach.
Notice that the SABR model refers to forward (not term) volatilities implied
in caplets/floorlets (not caps/floors). We denote with σ̃x(t; Tk-1, Tk, K) the
implied forward volatility seen at time t of an European caplet/floorlet on the
spot Euribor rate Lx(Tk-1, Tk) and strike K, with x = {Euribor 6M standard,
Euribor 6M CSA}. Thus, we stripped the two forward volatility surfaces im-
plied in the cap/floor premia published by Reuters on the 31st March and
on 31st August 2010, using the two multiple-curve methodologies above.
The stripping procedure requires many technicalities that we do not report
here, we refer you to section 3.6 in Brigo and Mercurio (2006).
The SABR calibration procedure is applied to each smile section, corre-
sponding to the strip of caplets/floorlets with the same maturity date Tk,
underlying FRA rate F̃x,k,(t) and different market strikes Kj, j = {1,…, 14}.4
Thus the calibration returns the values of the model’s parameters α, β, ν,
ρ that minimize the distance between the market implied forward volatili-
ties σ̃xMkt (t;Tk-1,Tk,Kj ) and the corresponding theoretical SABR volatili-
ties σ̃xSABR (t;Tk-1,Tk,Kj ) obtained through the closed analytic formula in
The Capco Institute Journal of Financial TransformationInterest Rates After the Credit Crunch: Markets and Models Evolution
Short term (2Y) – Euribor – 31 Mar 2010 Short term (2Y) – Eonia – 31 Mar 2010
0,000
0,002
0,004
0,006
0,008
0,010
0,012
0,014
10%
20%
30%
40%
50%
60%
0% 2% 4% 6% 8% 10%
Vega Vola
tility
Strike
Euribor Impl. Fwd. Vol. Vega-Wght. SABR Vol. Standard SABR Vol. Vega
0,000
0,002
0,004
0,006
0,008
0,010
0,012
0,014
10%
20%
30%
40%
50%
60%
0% 2% 4% 6% 8% 10%
Vega Vola
tility
Strike
Eonia Impl. Fwd. Vol. Vega-Wght. SABR Vol. Standard SABR Vol. Vega
Medium term (10Y) – Euribor – 31 Mar 2010 Medium term (10Y) – Eonia – 31 Mar 2010
0,000
0,002
0,004
0,006
0,008
0,010
0,012
0,014
10%
20%
30%
40%
50%
60%
0% 2% 4% 6% 8% 10%
Vega
Vola
tility
Strike
Euribor Impl. Fwd. Vol. Vega-Wght. SABR Vol. Standard SABR Vol. Vega
0,000
0,002
0,004
0,006
0,008
0,010
0,012
0,014
10%
20%
30%
40%
50%
60%
0% 2% 4% 6% 8% 10%
Vega Vola
tility
Strike
Eonia Impl. Fwd. Vol. Vega-Wght. SABR Vol. Standard SABR Vol. Vega
The blue dots represents the market implied forward volatility, the red line refers to the standard calibration, the green line refers to the vega-weighted calibration, and the purple line (right y-axis)
reports the values of the vega. The graphs on the left are related to the market Euribor implied forward volatility. The graphs on the right are associated with the market Eonia implied forward
volatility. Upper panels: smile section with maturity date 2-year. Lower panels: smile section with maturity date 10-year. Valuation date: 31st March 2010 (source: Reuters).
Figure 3 – SABR model calibration results
14 is the number of strikes quoted in the market.4
46
Table 5. Thus we obtain a set of SABR parameters for each smile sec-
tion. For the two dates (31st March and on 31st August 2010) and the
two pricing methodologies (multiple-curve no-CSA and multiple-curve
CSA) associated with the two corresponding forward volatility surfaces
(Euribor, Eonia), we performed two minimizations using two distinct error
functions:
a standard error function defined as the square root of the sum of ■■
the square differences between the SABR and the market forward
volatilities:
Errorstd(Tk)={∑14j=1[σ̃x
Mkt(t;Tk-1,Tk,Kj) - σ̃xSABR (t;Tk-1,Tk,Kj )]2}½ (1)
a vega-weighted error function: ■■
Errorvw(Tk)={∑14j=1[(σ̃x
Mkt(t;Tk-1,Tk,Kj) - σ̃xSABR (t;Tk-1,Tk,Kj ))Wj,x]2}½
(2)
where Wj,x= υ(Tk,Kj)
∑14j=1
υ(Tk,Kj)
and υ(Tk,Kj) is the Black’s vega sensitivity of the caplet/floorlet option
with strike Kj, FRA rate F̃x,k(t), and maturity Tk. Weighting the errors by
the sensitivity of the options to shifts of the volatility allows us, during the
calibration procedure, to give more importance to the near-ATM areas of
the volatility surface, with high vega sensitivities and market liquidity, and
less importance to OTM areas, with lower vega and liquidity.
The initial values of α, ρ, and ν were respectively 4.5 percent, -10 percent,
and 20 percent. Different initializations gave no appreciable differences in
the calibration results. According to Hagan et al. (2002) and West (2005),
in the calibration of the model we decided to fix the value of the redun-
dant parameter β to 0.5. The minimization was performed using the built-
in Matlab’s function “patternsearch.” A snapshot of the SABR calibration
results for the 31st March 2010 is shown in Figure 3 where we report two
smile sections at short-term (two-year maturity) and mid-term (ten-year
maturity). We do not include here the graphs of the calibration results for
longer term and for the 31st August 2010 because they would lead to
the same considerations. However, in Table 6 we compare the two cali-
bration approaches on both the two valuation dates reporting the most
important numbers: the range of minimum and maximum errors and the
standard deviation.
Overall, the SABR model performs very well at both dates with both pric-
ing methodologies. In particular, we notice that in the short term (two-
year, upper panels in Figure 3) the standard SABR calibration (red line)
seems, at first sight, closer to the market volatility (blue dots) and to bet-
ter replicate the trend in the OTM regions. However, a closer look reveals
that there are significant differences in the ATM area, where even small
calibration errors can produce sensible price variations. Instead, the ve-
ga-weighted SABR calibration (green line) gives a better fit of the market
volatility smile in the ATM region, in correspondence to the maximum
vega sensitivity, and allows larger differences in the OTM regions where
the vega sensitivity is close to zero. Thus the vega-weighted calibration
provides a more efficient fit in the volatility surface regions that are critical
for option pricing. The effects are less visible for the long term (middle
panels in Figure 3) because of the higher vega sensitivity in the OTM
regions.
Both the standard and the vega-weighted approaches lead to similar re-
sults in terms of range of minimum and maximum errors and standard
deviation (Table 6). In particular, the standard deviation measures of the
errors over the 30-year term structure are almost the same. This is due
to the fact that the two calibrations differ only in the short term (up to
four years) and using a vega-weighted minimization can ensure a better
fitting of the market data, as shown in the upper panels of Figure 3. We
conclude that the SABR model is quite robust under generalization to the
modern pricing framework and can be applied to properly fit the new dy-
namics of the market volatility smile and to price off-the-market options
coherently with the new market practice.
SABR calibration errors
31st March 2010 31st August 2010
Implied volatility Euribor Implied volatility Eonia Implied volatility Euribor Implied volatility Eonia
Range Standard
deviation
Range Standard
deviation
Range Standard
deviation
Range Standard
deviation
Standard calibration [-0.2%;+0.1%] 0.0003 [-0.1%;+0.1%] 0.0003 [-0.3%;+0,2%] 0.0004 [-0.3%;+0.2%] 0.0004
Vega-weighted calibration [-0.2%;+0.1%] 0.0003 [-0.1%;+0.1] 0.0002 [-0.1%;+0.1] 0.0004 [-0.1%;+0.1%] 0.0004
For each calibration procedure (standard and vega-weighted) and for each valuation date (31st March and 31st August 2010), we report the range of minimum and maximum calibration errors and
the standard deviation of the errors (equally-weighted for standard calibration and vega-weighted for vega-weighted calibration).
Table 6 – SABR model calibration errors over all the market volatility smile
47
ConclusionIn this work we have presented a quantitative study of the markets and
models evolution across the credit crunch crisis. In particular, we have
focused on the fixed income market and we have analyzed the most
relevant literature regarding the divergences between Libor versus OIS
rates, between FRA versus forward rates, the explosion of basis swaps
spreads, and the diffusion of collateral agreements and CSA-discount-
ing, in terms of credit and liquidity effects. These market frictions have
induced a segmentation of the interest rate market into sub-areas, cor-
responding to instruments with risky underlying Libor rates distinct by
tenors, and risk free overnight rates, and characterized, in principle, by
different internal dynamics, liquidity, and credit risk premia reflecting the
different views and preferences of the market players. In response to the
crisis, the classical pricing framework, based on a single yield curve used
to calculate forward rates and discount factors, has been abandoned, and
a new modern pricing approach has prevailed among practitioners, tak-
ing into account the market segmentation as an empirical evidence and
incorporating the new interest rate dynamics into a multiple curve frame-
work. The latter has required a deep revision of classical no-arbitrage
pricing formulas for plain vanilla interest rate derivatives, now funded on
the risk neutral measure associated to the risk free bank account and on
the martingale property of the FRA rate under such measure. In particular,
we have reported the multiple-curve generalization of the SABR model,
the simplest extension of the well-known Black’s model with stochas-
tic volatility, routinely used by market practitioners to fit the interest rate
volatility smile and to price vanilla caps/floors and swaptions.
We have reported the results of an empirical analysis on recent market
data comparing three different pre- and post-credit crunch pricing meth-
odologies and demonstrated the transition of the market practice from
the classical to the modern pricing framework. In particular, we have
proved that in the case of interest rate swaps the markets have, since
March 2010, abandoned the classical single-curve pricing methodology,
typical of the pre-credit crunch interest rate world, and instead adopted
the modern multiple-curve CSA approach, thus incorporating into market
prices the credit and liquidity effects. The same happened with European
caps/floors, with the full transition to the CSA-discounting methodology
deferred until August 2010. Finally, we have proved that the SABR model
is quite robust under generalization to the modern pricing framework and
can be applied to properly fit the new dynamics of the market volatility
smile and to price non-quoted off-the-market options coherently with the
new market practice.
The work presented here is a short step in the long-run theoretical refun-
dation of the interest rate modeling framework in a post-crisis financial
world, with Libor rates incorporating credit and liquidity risks. We believe
that such risks and the corresponding market segmentation expressed
by large basis swap spreads will not return as negligibly as they were in
the pre-crisis world, and will be there also in the future, exactly as the
volatility smile has been there since the 1987 market crash. Expected
future developments will also consider the extension of pre-crisis pric-
ing models to the multiple-curve world with stochastic basis, and the
pricing of non-collateralized OTC derivatives including consistently the
bilateral credit risk of the counterparties in the form of credit value ad-
justment (CVA) and debt value adjustment (DVA), and the liquidity risk of
the lender in the form of liquidity value adjustment (LVA) [Bianchetti and
Morini (2010)].
ReferencesAcerbi, C. and G. Scandolo, 2007, “Liquidity risk theory and coherent measures of risk,” SSRN •
working paper, http://ssrn.com/abstract=1048322
Ametrano, F. M., 2011, “Rates curves for forward Euribor estimation and CSA-discounting,” •
QuantLib Forum, London, 18 January
Ametrano, F. M. and M. Bianchetti, 2009, “Bootstrapping the illiquidity: multiple curve •
construction for coherent forward rates estimation,” in Mercurio, F., ed., Modelling interest
rates: latest advances for derivatives pricing, Risk Books, London
Amin, A., 2010, “Calibration, simulation and hedging in a Heston Libor market model with •
stochastic basis,” SSRN working paper, http://ssrn.com/abstract=1704415
Bianchetti, M., 2010, “Two curves, one price,” Risk Magazine, 23:8, 66-72•
Bianchetti, M., 2011, “Switching financial institutions to marking to market under the CSA •
regime,” presentation at Global Derivatives, April
Bianchetti, M. and M. Morini, 2010, Interest rate after the credit crunch, WBS Course, 25-26 •
Oct. Milan
Black, F. and M. Scholes, 1973, “The pricing of options and corporate liabilities,” The Journal of •
Political Economy, 81:3, 637-654
Brigo, D. and F. Mercurio, 2006, Interest rate models: theory and practice, 2nd edition, Springer, •
Heidelberg
Burgard, C. and M. Kjaer, 2010, “PDE representation of options with bilateral counterparty risk •
and funding cost,” SSRN working paper, http://ssrn.com/abstract=1605307
Chibane, M. and G. Sheldon, 2009, “Building curves on a good basis,” SSRN working paper, •
http://ssrn.com/abstract=1394267
Derman, E. and I. Kani, 1994, “The volatility smile and its implied tree,” Risk Magazine, 7:2, •
139-145
Fries, C. P., 2010, “Discounting revisited: valuation under funding, counterparty risk and •
collateralization,” SSRN working paper, http://ssrn.com/abstract=1609587
Fujii, M. and A. Takahashi, 2011, “Choice of collateral currency,” Risk Magazine, 24:1, 120-125•
Fujii, M., Y. Shimada, and A. Takahashi, 2009a, “A survey on modelling and analysis of basis •
spread,” CARF working paper, CARF-F-195, available from http://ssrn.com/abstract=1520619
Fujii, M., Y. Shimada, and A. Takahashi, 2009b, “A market model of interest rates with dynamic •
basis spreads in the presence of collateral and multiple currencies,” CARF working paper,
CARF-F-196, available from http://ssrn.com/abstract=1520618
Fujii, M., Y. Shimada, and A. Takahashi, 2010a, “A note on construction of multiple swap curves •
with and without collateral,” CARF working paper, CARF-F-154, available from http://ssrn.com/
abstract=1440633
Fujii, M., Y. Shimada, and A. Takahashi, 2010b, “Collateral posting and choice of collateral •
currency – implications for derivative pricing and risk management,” CARF working paper,
CARF-F-216, available from http://ssrn.com/abstract=1601866
Giura, R. 2010, Banca IMI swap trading, private communication•
Hagan, P. S., D. Kumar, A. S. Lesniewski, and D. E. Woodward, 2002, “Managing smile risk,” •
Wilmott Magazine, July, 84-108
Henrard, M., 2007, “The irony in the derivative discounting,” SSRN working paper, http://ssrn.•
com/abstract=970509
Henrard, M., 2009, “The irony in the derivative discounting part II: the crisis,” SSRN working •
paper, http://ssrn.com/abstract=1433022
Hull, J., 2010, Options, futures and other derivatives, 7th edition, Prentice Hall•
ICAP, 2010, communications to clients, 11 Aug. 2010 and 15 Sep. 2010•
ISDA, 2010, “ISDA margin survey 2010”, available form http://www.isda.org•
Kenyon, C., 2010, “Post-shock short-rate pricing,” Risk Magazine, 23:11, 83-87•
Kijima, M., K. Tanaka, and T. Wong, 2008, “A multi-quality model of interest rates,” Quantitative •
Finance, 9:2, 133-145
Mercurio, F., 2009, “Interest rates and the credit crunch: new formulas and market models,” •
SSRN working paper, http://ssrn.com/abstract=1332205
48
Mercurio, F., 2010a, “LIBOR market model with stochastic basis,” SSRN working paper, http://•
ssrn.com/abstract=1583081
Mercurio, F., 2010b, “Modern Libor market models: using different curves for projecting rates •
and for discounting,” International Journal of Theoretical and Applied Finance, 13:1, 113-137
Mercurio, F. and M. Morini, 2009, “Joining the SABR and Libor models together,” Risk •
Magazine, 22,:3, 80-85
Merton, R. C., 1973, “Theory of rational option pricing,” The Bell Journal of Economics and •
Management Science, 4:1, 141-183
Michaud, F. and C. Upper, 2008, “What drives interbank rates? Evidence from the Libor panel,” •
BIS Quarterly Review, March
Moreni, N. and A. Pallavicini, 2010, “Parsimonious HJM modelling for multiple yield-curve •
dynamics,” SSRN working paper, http://ssrn.com/abstract=1699300
Morini, M., 2009, “Solving the puzzle in the interest rate market (part 1 and part 2),” SSRN •
working paper, http://ssrn.com/abstract=1506046
Morini, M. and A. Prampolini, 2010, “Risky funding: a unified framework for counterparty and •
liquidity charges,” SSRN working paper, http://ssrn.com/abstract=1669930
Pallavicini, A. and M. Tarenghi, 2010, “Interest-rate modelling with multiple yield curves,” SSRN •
working paper, http://ssrn.com/abstract=1629688
Piterbarg, V., 2010a, “Funding beyond discounting: collateral agreements and derivative •
pricing,” Risk Magazine, 23:2, 97-102
Piterbarg, V., 2010b, “Effects of funding and collateral,” Global Derivatives & Risk Management •
Conference, May 2010, Paris
Tuckman, B. and P. Porfirio, 2003, “Interest rate parity, money market basis swap, and •
cross-currency basis swap,” Lehman Brothers Fixed Income Liquid Markets Research – LMR
Quarterly, 2004, Q2
West, G., 2005, “Calibration of the SABR model in illiquid markets,” Applies Mathematical •
Finance, 12:4, 371-385
49
PART 1
Fat Tails and (Un)happy Endings: Correlation Bias, Systemic Risk and Prudential Regulation
AbstractThe correlation bias refers to the fact that claim subordination
in the capital structure of the firm influences claim holders’
preferred degree of asset correlation in portfolios held by the
firm. Using the copula capital structure model, it is shown
that the correlation bias shifts shareholder preferences to-
wards highly correlated assets. For financial institutions, the
correlation bias makes them more prone to fail and raises
the level of systemic risk given their interconnectedness. The
implications for systemic risk and prudential regulation are
assessed under the prism of Basel III, and potential solutions
involving changes to the prudential framework and corpo-
rate governance are suggested.
Jorge A. Chan-Lau — International Monetary Fund and The Fletcher School, Tufts University1
The views presented in this paper are those of the author and do not 1
necessarily reflect those of the IMF or IMF policy. The paper benefits from
comments by Stijn Claessens, Dora Iakova, Robert Rennhack, and Marcos
Souto. The author is responsible for any errors or omissions.
50
A firm or financial institution that holds a portfolio of diverse projects and/or
assets is subject to correlation risk, or the risk arising from the correlation
of cash flows accrued to the different projects/assets in the portfolio. Most
firms and financial institutions finance their portfolios using a mix of claims,
such as equity and debt, where each claim is differentiated by its seniority
in the capital structure; for example, equity is subordinated to debt.
The correlation bias is the preference of different claim holders on the
firm for different levels of projects/asset correlation in the firm’s portfolio.
In particular, junior claim holders prefer portfolios where assets are highly
correlated while senior claim holders would prefer a more uncorrelated
portfolio. Since the control of the firm is usually exercised by managers
that tend to act on behalf of the most junior claim holders, shareholders,
the choice of portfolio assets would tend to be biased towards highly
correlated assets. This bias leads to portfolio outcomes characterized by
fat tails that increase the likelihood of observing scenarios with extreme
upside and downside risks. In particular, the lack of diversity in the firm’s
portfolio increases the likelihood that it may fail since all the assets in the
portfolio will be equally affected by a negative shock.
The implications of the correlation bias are not circumscribed to indi-
vidual institutions though. In a financial system where the correlation bias
of junior claim holders is dominant and herd behavior prevalent, it would
not be rare to observe “black swan” events often following on the heels
of extended periods of tranquility [Taleb (2009)]. The stronger the bias
of junior claim holders the more likely the financial system will oscillate
between extreme periods of tranquility and financial disruption, contrib-
uting to increased procyclicality in the event of negative shocks. This is
not just a mere theoretical implication: fat tail “tales” and their unhappy
endings were dramatically illustrated by the recent global financial crisis
originated by problems in the U.S. subprime mortgage market.
This paper argues that a copula approach to the capital structure, build-
ing on the copula pricing model first developed to analyze structured
credit products, provides the right framework for understanding the cor-
relation bias arising from the capital structure of the firm. Furthermore,
the copula capital structure model is a natural generalization of the con-
tingent claim approach to the capital structure of the firm first proposed
by Black and Scholes (1973) and Merton (1974). Insights on the correla-
tion bias derived from the copula pricing model are useful to understand
how the bias interacts with systemic risk and whether recent financial
regulatory reform could address these interactions effectively.
Understanding the correlation biasThe basic contingent claim modelBlack and Scholes (1973) and Merton (1974), by noting that the payoffs
to equity and debt were equivalent to options on the asset value of the
firm, established the foundations of the contingent claim approach for
analyzing the capital structure of the firm. The approach sheds light on
the potential conflicts of interest between shareholders and debt hold-
ers. Specifically, the payoff of equity, as illustrated in Figure 1, resembles
the payoff of a call option where the underlying is the asset value of the
firm and the strike price is determined by what is owed to debt holders.
Ceteris paribus, shareholders, who hold a long asset volatility position,
benefit from projects and/or portfolios that increase the volatility of the
firm’s asset value, as increased volatility increases the value of the call
option. Shareholders, hence, exhibit a high volatility bias while the op-
posite is true for debt holders.
While useful for recognizing conflicts of interest between different claim
holders, the Black-Scholes-Merton capital structure model does not
account for the fact that the asset value of the firm is determined by
the portfolio of projects and assets held by the firm. The volatility of the
profit/loss of the firm, its equity returns, and its asset value are deter-
mined not only by the volatility of individual projects but on their mutual
dependence (or correlation). Accounting for the correlation in the firm’s
project/asset portfolio requires an understanding of how correlation af-
fects different claims on the firm, which requires setting up an appropri-
ate analytical framework.
Tail risk and correlationNevertheless, it can be shown that the basic intuition of the Black-
Scholes-Merton capital structure model, that shareholders benefit from
increased volatility of the asset value of the firm, extends to the case of a
firm undertaking a portfolio of several projects. More importantly, share-
holders can increase the volatility of the firm by increasing the correlation
among the different projects in the firm’s portfolio.
To illustrate this point, assume a firm that can choose between the fol-
lowing two-project portfolios, each one requiring an initial investment of
100 per project. The first portfolio contains uncorrelated projects A and
B. The probability of success and failure of each project is 0.5. For simpli-
fication, it is further assumed a zero rate of return. In case of success, the
project returns the original investment of 100, and in the case of failure
the original investment is lost. The second portfolio contains projects C
Payoff
Equity
Asset value of the firm Asset value of the firm
Debt
Figure 1 – Payoff schedules of equity and debt
51
and D, with similar characteristics as projects A and B in the first port-
folio. The only difference is that projects C and D are uncorrelated. The
three potential profit/loss scenarios are that the project portfolio loses
100, breaks even at 0, or gains 100. Figure 2 shows the probability distri-
bution of each scenario for both portfolios.
The capital structure financing the portfolio and who controls the firm
influence the degree of correlation of the assets in the portfolio. The first
and second portfolios both have an expected value of 100. If the project
portfolio is financed only with equity, the shareholders would prefer the
first portfolio, as its standard deviation is smaller and the probability of
losing money is only 25 percent. Once the project portfolio is financed
partly with debt, as the share of debt increases the stronger the incen-
tives for the shareholders to choose the second portfolio, as they would
accrue positive returns only in the scenario where both projects are
succesful. For instance, for a 50-50 mix of debt and equity, portfolio 1
yields shareholders an expected profit/loss of 25 while portfolio 2 deliv-
ers twice that amount, 50. Shareholders, hence, would prefer portfolio
2 even though the odds that the debtholders would suffer a total loss is
50 percent compared to 25 percent in portfolio 1. In contrast, bondhold-
ers would prefer portfolio 1 to portfolio 2.
The copula capital structure model and the correlation biasThis section will exploit the analogy between the capital structure of
the firm and structured and securitized products to explain how struc-
tured credit pricing models can serve as an analytical and conceptual
framework to evaluate issues related to portfolio correlation, systemic
risk, and the correlation bias. To observe this, note that as in the case
of a securitization or structured product, the total asset value of the firm
depends on cash flow from its different projects, or in the case of a bank,
its bank and trading portfolios. In turn, the value of each corresponding
claim depends on its priority over the cash flow determined by its senior-
ity in the capital structure. Figure 3 shows the analogy between the firm
and a structured product. As in the case of structured products, senior
debt is protected by the buffer consisting of, in a first instance, equity,
and in a second instance subordinated debt, which in that order absorb
any loss incurred by the firm or financial institution.
The observation above suggests that the copula pricing model, as origi-
nally proposed by Vacisek (1977) and specialized to the Gaussian copula
by Li (2000), could be the natural framework for analyzing conflicts of
interest between shareholders and debt holders. The analysis of the capi-
tal structure based on the copula pricing model, henceforth, would be
referred to as the copula capital structure model. In the copula approach,
the value of the projects/assets included in the firm’s portfolio is deter-
mined by a common factor and a project/asset idiosyncratic factor. The
correlation of the assets is determined by the correlation of each asset
with the common factor. This framework can be easily generated to mul-
tifactor copula models.2
Furthermore, it can be shown that the copula capital structure model is a
natural extension of the contingent claim approach of Black and Scholes
(1973) and Merton (1974) since the value of equity is determined by a
portfolio of options in several projects, or that each tranche comprises
long and short options on the value of the portfolio [Rajan et al. (2007)].
The natural advantage vis-à-vis the contingent claim approach is that the
copula approach naturally accommodates several types of claims and the
fact that firms hold portfolios comprising multiple projects and assets.
The main insight from the copula approach to the capital structure is
that it makes explicit the correlation bias, or differences in claim holders’
preferences for the degree of correlation among the different projects/as-
sets held by the firm. The particular instance of the correlation bias that
leads to higher idiosyncratic and systemic risk is that associated with the
correlation bias of shareholders since they prefer portfolios with highly
correlated assets.
The copula capital structure model explains how different factors affect
The Capco Institute Journal of Financial TransformationFat Tails and (Un)happy Endings: Correlation Bias, Systemic Risk and Prudential Regulation
For a concise exposition, see Andersen et al. (2003), Gibson (2004), Hull and White (2004), 2
and Chapter 9 in Lando (2004). Coval et al. (2009) offer an accessible introduction. Rajan
et al. (2007) is textbook treatment by market practitioners. The reader should bear in
mind, though, that the insights from the copula benchmark model are not restricted to the
particular distributional assumptions of the model but extend to other copula models, such
as Student-t, semi-parametric, and non-parametric copulas.
Probability Probability
0,50 0,50
0,25 0,25
-100 0 100Pro�t/loss
-100 0 100Pro�t/loss
Portfolio 1 Portfolio 2
Figure 2 – Profit/loss of hypothetical two-project portfolios
Tranched equity product
Firm
Senior tranche
Mezzanine tranche
Equity tranche
Senior debt
Subordinated debt
Equity
Figure 3 – The analogy between the capital structure of a tranched structured product and the capital structure of the firm
52
the value of the different claims in the capital structure of the firm. One
such factor is the riskiness of the individual projects as proxied by their
probability of success. An improvement on the odds of success benefits
all claims as long as the payoff of a successful project is independent of
the riskiness of the project (Figure 4). If the payoff of the project/assets in-
creases with the level of risk, the standard result from the Black-Scholes-
Merton contingent claim model applies: equity shareholders benefit when
the firm undertakes high risk-high return projects. The reason for this
result, again, is due to the convexity of the shareholders’ payoff which
gains the most from upside risks. In contrast, ceteris paribus, an increase
in the correlation of the cash flows of the different projects (or assets)
and the corresponding increase in the risk correlation of the projects (or
assets) benefit shareholders but not senior debt holders.
The intuition underlying this result is as follows. Increased risk correlation
leads to outcomes where either the majority of projects in the portfolio
succeed or fail. In the extreme case of perfectly correlated projects or
assets, the outcome is binary, either the portfolio succeeds or it fails.
The downside to shareholders from outcomes close to the binary case
is limited to their equity stake in the firm, which only finances part of the
total portfolio or assets of the firm. Shareholder, thus, are indifferent to all
scenarios where losses exceed the equity of the firm.
The downside scenarios are accompanied by upside scenarios where,
due to the high correlation among different projects or assets, the portfo-
lio bears minimal or no losses at all. As in the case of the Black-Scholes-
Merton contingent claim model, the upside scenarios benefit sharehold-
ers more than debtholders due to the convexity of the payoff structure
of the former: they accrue all the gains in excess of what is owed to
debtholders. The nature of the payoffs associated with the downside and
upside scenarios provide incentives to shareholders to bias the firm’s
portfolio towards highly correlated projects and/or assets. In contrast,
senior debt holders have the opposite bias and would rather prefer that
the assets/projects held by the firm exhibit low correlation.
Another important result, which also bears on the design of prudential
regulation of financial institutions, relates to how subordinated debt re-
acts to changes in correlation. The sensitivity of subordinated debt to
project/asset correlation is non-monotonic: at low levels of correlation,
the value of subordinated debt declines but after a certain threshold is
reached, its value starts increasing. Why does subordinated debt exhibit
a non-monotonic relationship with correlation? The answer is as follows:
subordinated debt becomes the loss absorbing buffer once the equity
buffer is exhausted in case of large portfolio losses. For very low levels of
projects/asset correlation, losses are small and are fully absorbed by eq-
uity. As correlation increases beyond certain threshold, the losses could
potentially exceed the equity buffer forcing partial losses in the value of
subordinated debt. Below the threshold, it is in the best interest of sub-
ordinated debt holders to keep project/asset correlation as low as pos-
sible to maximize the value of their stake in the firm since low correlation
minimizes their partial losses. Once the correlation threshold is crossed,
however, the loss scenarios would likely imply full losses to subordinated
debt holders and would bias their preference towards highly correlated
projects/assets since they will maximize the likelihood of a sufficiently
low number of failed projects.
The bias of junior claim holders towards highly correlated assets and
projects is determined by the relative size of equity and subordinated
debt in the capital structure of the firm.3 The smaller the size of these
subordinated claims (or tranches in structured credit parlance), the stron-
ger the incentive for subordinated claim holders to opt for high correla-
tion on the asset side of the balance sheet. In the case of subordinated
debt, the smaller the buffer provided by equity, the lower the threshold
correlation would be.
Correlation bias, systemic risk, and prudential regulationThe correlation bias, in combination with corporate control, could poten-
tially induce high correlation in the asset and trading portfolios of indi-
vidual institutions, raising the likelihood of their failure. Since excess port-
folio correlation at the individual firm level could translate into systemic
risk at the aggregate level it becomes important to assess whether the
problems induced by the correlation bias could be addressed through
corporate governance and prudential regulation.
Systemic riskThe design of a prudential regulatory framework should account for the
correlation bias exhibited by the different claim holders. On the one hand,
if the correlation bias of junior stakeholders dominates the choice of proj-
ects and assets included in the firm portfolio, the likelihood that the firm
will fail increases. On the other hand, the correlation bias induces both
Value ValueSenior debt Senior debt
Subordinated debt
Subordinated debt Equity
Equity
Project riskiness Correlation
Figure 4 – Sensitivity of corporate claims value to the riskiness of a single project and to portfolio correlation
For a rigorous but accessible derivation of this a result, see Gibson (2004) and Coval et al. 3
(2009).
53
upside and downside fat-tail risk in the portfolio, which could lead to
wild swings from periods of tranquility to periods of turmoil. Furthermore,
the excess portfolio correlation is compounded by the well documented
procyclical behavior of asset correlation which tends to increase during
periods of turmoil.4
Systemic risk would not be an issue if the failure of a financial institution
were an isolated event. The interconnected nature of the financial system,
however, raises the risk that an individual failure could have a domino ef-
fect on other institutions and prompt serious disruptions across different
markets. The interconnectedness nature of the financial system arises
from direct linkages between the different market participants, such as
the cross-institutions claims, and from indirect linkages such as exposure
to common risk factors and feedback effects increased volatility and de-
clining prices from the use of similar accounting and risk management
practices. Furthermore, the failure of one bank could lead to a bank panic
and a run on the banking system even in the absence of direct or indirect
linkages.
Besides interconnectedness, the impact of the correlation bias on sys-
temic risk can be compounded by herding behavior among financial in-
stitutions. Herding behavior, which is rather common in financial markets,
increases the chances of facing a too-many-to-fail (TMTF) problem since
a single adverse shock could prompt the failure of several institutions
holding similar portfolios biased towards highly correlated assets [Acha-
rya and Yorulmazer (2007)]. Herding behavior could be prompted by sev-
eral factors, such as reputational issues that cause institutions to mimic
each other’s investment behavior.5 Similarly, the trading strategies of dif-
ferent institutions are likely to converge to specific trades, overcrowding
them, and raising the possibility of generalized losses in specific sectors
of the financial system. The losses incurred by credit correlation traders
in 2005 and quantitative alternative asset managers in 2007 are two re-
cent examples of this problem.
The discussion above suggests that reducing systemic risk from cor-
relation bias requires corporate governance structures and regulatory
requirements aimed at reducing correlation in the asset portfolio of fi-
nancial institutions.
Corporate governanceCorrelation bias risk is present when corporate control is exercised by
shareholders (or management acting on behalf of shareholders in the ab-
sence of informational asymmetries).6 In the particular case of banks,
the share of equity in the capital structure is relatively small since, in the
presence of asymmetric information, high leverage is the market solution
for dealing with governance problems, as noted by Kashyap et al. (2008).
But as noted above, high leverage would bias the bank even more to-
wards highly correlated assets in its portfolio.7
One potential solution for reducing correlation bias risk involves changing
the corporate governance of financial institutions to shift some corporate
control to claim holders other than shareholders. The simpler way to en-
able debt holders partial control on the firm would be to require them to
hold also some equity in the firm. In the real world it is not uncommon to
find that type of structure in bank-based financial systems in Japan and
continental Europe, where banks are both lenders to and equity owners
in other corporations.8 Whether banks’ portfolios in these countries are
less correlated than those based in market-based, arm’s-length financial
systems, such as as those in the U.K. and the U.S., is an open empirical
question.
Absent substantial changes in corporate governance structures and cor-
porate control, the monitoring role of debt holders must be overtaken
by prudential regulation. As argued in Dewatripont et al. (2010), the debt
holder base of financial institutions could be too diversified and lack the
required expertise to monitor or exercise control efficiently. Moreover, for
institutions deemed systemic, there are explicit or implicit guarantees like
deposit insurance that work against close monitoring of the financial in-
stitution. Under such circumstances, prudential regulation needs to fill
the monitoring role of debt holders.
Prudential regulationThe 2008-9 global financial crisis has prompted a number of reform ini-
tiatives to the prudential regulation of banks and financial institutions,
including the use of contingent capital, increases in minimum capital re-
quirements, the imposition of systemic risk capital charges, and requiring
originators of structured and securitized products to have some “skin in
the game” to align their incentives with those of investors in securitized
notes. The relative merits of these initiatives, from the perspective of the
correlation bias, are explored in detail.
The Capco Institute Journal of Financial TransformationFat Tails and (Un)happy Endings: Correlation Bias, Systemic Risk and Prudential Regulation
See, for example, Hartmann et al. (2004). The increase in asset price correlation during 4
periods of turmoil is prompted by several factors, including liquidity shortages due to fire sales
prompted by mark-to-market and risk management practices [Brunnermeier and Pedersen
(2008), Shin (2008)]. In general, the financial system is characterized by procyclicality (Borio et
al. (2001)], which suggest the need for counter-cyclical regulation [Brunnermeier et al. (2009)].
See Scharfstein and Stein (1990) for an early reference to herding behavior among mutual 5
fund managers and Bikhchandani and Sharma (2001) for a more recent survey. On herding
behavior and fat-tails, see Nirei (2006).
Agency problems between managers and shareholders can be accommodated into the 6
copula approach by noting that managers would be the most junior claimants on the cash
flows of the firm’s portfolio, and would be even more biased than shareholders towards
highly correlated projects and/or assets.
In emerging market countries, domestic banks are usually controlled by families or financial 7
conglomerate holdings that own a substantial majority of shares. The low leverage of these
banks works against correlation bias risk but family or financial conglomerates-controlled
firms may raise issues related to the protection of minority shareholders’ rights.
See Allen and Gale (2000). An added benefit from shared corporate control is that, 8
besides reducing the correlation bias risk, it also helps reduce the agency cost of debt,
or the underinvestment problem [Chan-Lau (2001)]. It should be borne in mind that bank-
based financial systems, while effective in reducing the correlation bias problems and
underinvestment, may generate other problems, such as overinvestment and inefficient
liquidation [Allen and Gale (2000)].
54
Contingent capital and hybrid securitiesAcademics, policy makers, and market practitioners have argued that the
use of contingent capital and/or hybrid securities could help reduce too-
big-to-fail (TBTF) risk [Corrigan (2009), Dudley (2009), Strongin et al. (2009),
Tarullo (2009), Tucker (2009), French et al. (2010), and BCBS (2010a)]. Un-
der this proposal, the capital structure of systemic financial institutions
should include two tranches of subordinated claims. The first and most ju-
nior tranche would be common equity, and the second most junior tranche
would be subordinated debt which will convert to equity once a pre-es-
tablished, pre-insolvency threshold is crossed, i.e., a decline in the ratio of
common equity or regulatory capital to risk-weighted assets ratio.
There are at least two strong arguments for the use of contingent capital.
The first is that during distress periods, contingent capital facilitates an
orderly recapitalization of a bank and/or financial institution, especially
under circumstances when accessing capital markets and/or obtaining
equity capital injections are difficult. The second is that the risk of dilution
during periods of distress would provide shareholders with incentives to
avoid excessive risk taking. At the same time, by removing ambiguity
about a potential bailout of subordinated creditors in case of an insti-
tution failure, holders of convertible subordinated debt will have strong
incentives to price risks correctly. In turn, more reliable prices would have
a signalling and disciplinary effect on financial institutions.
While contingent capital could be useful for ensuring that banks will be
able to comply with minimum regulatory capital requirements under se-
vere circumstances, it cannot address the correlation bias problem suc-
cessfully. From a functional perspective, contingent capital and hybrid
securities can be classified as subordinated debt. Moreover, the convert-
ibility feature forces contingent capital and hybrid securities to resemble
equity more closely and induces a stronger bias towards a highy cor-
related portfolio than in the case of plain subordinated debt. From the
perspective of shareholders, the incentives from the equity dilution effect
from exercising the convertibility option are offset by the fact that the
option increases the subordination of equity in the capital structure. The
more subordinated equity is, the stronger the incentives to gamble on
increased volatility, including increasing the asset correlation in the bank-
ing and trading books.
Minimum capital requirementsIn September 2010, the Basel Committee on Banking Supervision
(BCBS) announced higher global capital minimum standards for com-
mercial banks following a number of recommendations aimed at revising
and strengthening the prudential framework [BCBS (2009, 2010b) and
Caruana (2010)]. The specific recommendations are to enhance the capi-
tal quality by imposing a stricter definition of common equity, for example
“core capital,” and requiring more capital, by raising the minimum com-
mon equity requirements to 4½ percent of risk-weighted assets (RWA)
from 2 percent, and adding a conservation buffer of 2½ percent on top
of it, raising total common equity ratio to 7 percent of RWA. National
authorities could also impose a countercylical macroprudential common
equity overlay of up to 2½ percent of RWA. Tier-1 capital requirements
are increased to 6 percent, and the minimum total capital requirement
remains equal to 8 percent of RWA (Table 1).
In principle, increasing the share of common equity in the capital struc-
ture of a bank helps reduce the bias towards highly correlated assets. But
the increase in common equity is relative to risk-weighted assets rather
than to the total assets of the firm which creates a loophole for banks
wishing to exploit the correlation bias. Banks, by concentrating their port-
folio on highly correlated low risk-weighted assets could satisfy both the
minimum and required capital requirements while reducing common eq-
uity and increasing leverage.9
The introduction of a non-risk based leverage ratio requiring Tier-1 assets
of no less than 3 percent of non-weighted assets plus off-balance sheet
structures could contribute to limiting the build up of leverage. Neverthe-
less, since Tier-1 assets comprise assets other than common equity, the
leverage ratio may not be enough to address the correlation bias risk
posed by shareholders’ correlation preferences.
Systemic risk capital chargesAnother reform proposal contemplates imposing systemic capital charg-
es, or capital charges, proportional to the contribution of each bank’s
contribution to systemic risk.10 While systemic capital charges do not
address the correlation bias directly, if the charges reflect how the failure
of one bank would spillover to other banks, they could provide incentives
for the bank to reduce its default risk, which should be reflected in a rela-
tively diversified banking and trading book.
There are several difficulties, however. One is related to the measure-
ment of how much each bank contributes to systemic risk. The risk mea-
sures are based on market measures including but not limited to mark-
to-market firm value and profit/loss statements, measures based on the
prices of equity and bonds or credit default swap spreads, or measures
based on a combination of balance sheet data and market prices such
as distance-to-default, Moody’s KMV expected default frequencies, and
Altman Z-scores [Adrian and Brunnermeier (2009), Chan-Lau (2009), and
For a simple account of the problems created by the use of risk-weights, seeTriana (2010), 9
The Economist (2010).
On systemic risk charges, one of the pioneering papers is Acharya (2001). The recent 10
global financial crisis has spurred work in this area, including, among others, Adrian and
Brunnermeier (2009), Chan-Lau (2010), Gauthier et al. (2010), Lester et al. (2008), Tarashev
et al. (2010). See also Chapter 13 in Acharya and Richardson (2009), and Brunnermeier et
al. (2009).
55
The recent performance of structured vehicles during the 2008-9 crisis
suggests that the previous argument rings true. Standard market practice
in tranched securitizations and structured products is for the originating
bank to retain the equity tranche to assure investors in the more senior
tranches that the structured vehicle is safe. Additional safeguards to se-
nior claims include a principal and interest payment waterfall structure
which directs first the cash flows from the collateral assets to the pay-
ment of principal and interests to senior claims, effectively reducing the
maturity of the senior claims [Rajan et al. (2007)].
None of these safeguards, however, addressed the correlation bias in-
centives of the originating bank. From the perspective of the copula pric-
ing model framework it is not surprising that, despite these safeguards,
the correlation among the collateral assets was way higher than the one
used to price and manage the risk of the structured vehicles. While model
risk and lack of historical data could also be singled out as responsible
for the low correlation estimates used to model the prices of structured
vehicles, the recent indictment of an investment bank by the U.S. Securi-
ties and Exchange Commission suggests that gaming model parameters
such as correlation to benefit a certain group of claim holders is not at
all unusual.12
What may work against correlation bias riskThe evaluation of the regulatory initiatives, especially Basel III, suggests
that they would be rather ineffective for reducing the correlation bias of
shareholders in financial institutions. The finding should not be surprising
since until now, the correlation bias has not been identified as a potential
source of idiosyncratic or systemic risk. The natural question for regula-
tors is what measures could work towards reducing the correlation bias
of shareholders.
On incentives associated with executive compensation see Acharya and Richardson (2009), 11
Part III, and French et al. (2010), Chapter 6.
U.S. SEC (2010). The investment bank settled the charge by paying a $550 million fine.12
The Capco Institute Journal of Financial TransformationFat Tails and (Un)happy Endings: Correlation Bias, Systemic Risk and Prudential Regulation
Lester et al. (2008)]. For the systemic risk capital charges to work, the
market-based measures need to capture potential spillovers reliably, but
it is often the case that markets fail to price risk correctly.
Moreover, as described above, the correlation bias of shareholders also
induces upside fat-tail risk that could be reflected in long periods of tran-
quility. If the data available for calculating the systemic risk charge spans
such a period, and periods of turmoil have yet to be realized, the sys-
temic risk charge could understimate spillover risks. The global nature
of systemic financial institutions requires the harmonization of systemic
capital charges across different jurisdictions, a feat that could be difficult
to accomplish. Finally, the ever evolving nature of investment and trad-
ing strategies suggest that the nature of spillovers, and in turn, a bank’s
contribution to systemic risk is constantly changing, therefore banks that
are currently deemed relatively safe from a systemic risk perspective may
not be so going forward.
“Skin-in-the-game” measuresThe dramatic implosion of the structured credit market in 2008-9, es-
pecially for tranched residential mortgage-backed securities, prompted
initiatives to require banks originating structured vehicles to have more
“skin-in-the-game,” or in other words, to hold a relatively risky claim on
the vehicles they originate. By forcing banks to have more of their capital
at stake when structuring a vehicle, both investors and regulators expect
that banks would have incentives to perform due diligence on the quality
of the collateral assets. In addition, to avoid regulatory arbitrage, the total
capital held against all claims or tranches in a securitized product should
not be less than the capital that would be held against the collateral as-
sets [BCBS (2010)].
Abstracting from issues related to executive compensation and the short-
horizon of managers, the copula capital structure model suggests that
more “skin-in-the-game” measures concentrated in junior claims on struc-
tured products are bound to fail in offseting the correlation bias.11 The
copula pricing model suggests that a bank holding the most junior, subor-
dinated claim in a structured product has a strong incentive to include as
many highly correlated assets as possible, as shown in Figure 4.
In percent of risk-
weighted assets
Capital requirements Additional macroprudential
overlay
Common equity Tier 1 capital Total capital Countercyclical buffer
Minimum Conservation
buffer
Required Minimum Required Minimum Required Required
Basel II 2.0 n.a. n.a. 4.0 n.a. 8.0 n.a. n.a.
Basel III 4.5 2.5 7.0 6.0 8.5 8.0 10.5 0 – 2.5
Source: Caruana (2010)
Table 1 – Basel II and Basel III – capital requirements
56
Reduce leverageThe simpler solution, probably, is to require financial institutions to hold
higher levels of common equity relative to unweighted assets rather than
risk-weighted assets. The current Basel III requirements specify a maxi-
mum leverage ratio of 33 for Tier-1 capital, so that the leverage ratio of
common equity could exceed that number under certain circumstances.
Even if Tier-1 were to comprise only common equity, that recommend-
ed leverage ratio implies a very thin equity layer supporting the capital
structure. Calibrating the maximum common equity leverage ratio would
first require setting the acceptable level of risk for a financial institution
and evaluating precisely the impact of the correlation bias on the risk of
the institution. Alternatively, a rough calibration exercise under simulated
conditions could yield rough estimates of common equity leverage ratios
deemed safe based on historical experience.
Enforce the Volcker Rule and portfolio diversification requirementsAnother solution is to require financial institutions to hold diversified
banking and trading portfolios. Since trading portfolios can be relatively
complex and the reality of trading implies frequent portfolio changes,
monitoring the portfolio’s degree of asset correlation is extremely diffi-
cult, both for supervisory agencies and the institution itself. Probably, the
cleanest way to avoid correlation bias risk is to adopt the Volcker rule and
ban systemic financial institutions from proprietary trading activities.13
In the case of the banking portfolio, existing requirements such as con-
centration and large position limits work towards increasing portfolio di-
versification. These requirements, however, are only rough guidelines to
ensure portfolio diversification since seemingly unrelated sectors may be
actually correlated due to their exposure to common risk factors. One
way to ensure diversification would be to establish quantitative limits
guided by a risk factor analysis performed by the regulatory agency or by
the bank using its own internal models. As indicated before, model risk,
and in the case of internal models, incentives to game the system could
work against this solution.
Force originators to hold “skin, flesh, and bones” in securitized productsThe correlation bias shows that investors in structured and securitized
products would prefer that the collateral assets exhibit different levels of
correlation depending on the subordination of their claims. In particular,
senior tranche holders would prefer low correlation while equity holders
would prefer high correlation. To align incentives, originators usually hold
the equity tranche and waterfall structures that reduce the maturity of se-
nior tranches, but these measures cannot address the correlation bias of
the equity holders. Rather than holding the riskiest tranche, from a default
risk perspective, or “skin-in-the-game,” originators should be required
to hold stakes in every single tranche of the structured and securitized
product, or in other words they have to put “skin, flesh, and bones” in
the game.
Enhance corporate control by debt holdersMoving beyond potential solutions related with the regulatory framework,
reducing correlation bias risk may require introducing changes in the cor-
porate governance structure of financial institutions to yield more corpo-
rate control to debt holders. The incentives for debt holders to participate
actively in the corporate control of a bank are relatively weak.
Banks’ debt-like instruments are mainly of a short-term nature, held by
a diversified investor base, and in the case of deposits, they are guar-
anteed by the government. To shift incentives, banks could be required
to increase the share of long-term debt in their liabilities, and long-term
creditors should be represented in the board of directors. An added ben-
efit of increasing the share of long-term debt is the enhancement of the
bank’s asset-liability management. Going into this direction would require
a careful design of the the balance of power between directors represent-
ing shareholders and those representing debt holders. The copula capital
structure model could provide guidance towards this end.
ConclusionsViewing the firm as a portfolio of projects and assets funded by claims
differentiated by their seniority in the capital structure helps us to extend
the insights of the contingent claim model of the capital structure, based
on the Black-Scholes-Merton model, to a copula capital structure model
based on the copula structured credit pricing model.
More importantly, by shifting to a portfolio perspective, it becomes
straightforward to identify the correlation bias, which influences the pref-
erences of different claim holders for the degree of asset correlation in
the firm’s portfolios. The more junior the claim is the higher the preference
for higher asset correlation.
The existence of the correlation bias, combined with the fact that cor-
porate control is exercised by the most junior claimants, suggest that
financial institutions including systemic banks may bias their banking
and trading portfolios towards highly correlated assets. The portfolio bias
could increase the risk that the institution fails, and owing to the high de-
gree of interconnectedness in the financial system, systemic risk would
increase since its failure would cause severe market disruptions and raise
the likelihood of the subsequent failures.
The Volcker rule, initially proposed by former U.S. Federal Reserve Chairman Paul Volcker, 13
was incorporated as section 619 in the Dodd-Frank Act approved by the U.S. Congress on
June 27, 2010.
57
Because the systemic risk implications of the correlation bias have yet
to be recognized explicitly, current prudential regulatory reform initiatives
including the use of contingent capital and hybrid securities, systemic
risk charges, and market practices like waterfall structures in securitized
products cannot reduce correlation bias risk and its impact on systemic
risk. Measures that could potentially reduce correlation bias risk include
increasing substantially the share of common equity in the capital struc-
ture of financial institutions, enforcing diversification requirements, ban-
ning proprietary trading in systemic institutions, requiring securitization
originators to hold stakes in all tranches of the capital structure of securi-
tized products, and giving debt holders some control over the firm.
ReferencesAcharya, V. V., 2001, “A theory of systemic risk and design of prudential bank regulation,” •
working paper, New York University
Acharya, V. V. and T. Yorulmazer, 2007, “Too many to fail – an analysis of time-inconsistency in •
bank closure policies,” Journal of Financial Intermediation, 16:1, 1–31
Acharya, V. V. and M. Richardson, editors, 2009, Restoring financial stability: how to repair a •
failed system, John Wiley and Sons
Adrian, T. and M. Brunnermeier, 2009, “CoVaR,” working paper, Federal Reserve Bank of New •
York
Allen, F. and D. Gale, 2000, Comparing financial systems, MIT Press•
Andersen, L., J. Sidenius, and S. Basu, 2003, “All your hedges in one basket,” Risk, November, •
67–72
Basel Committee on Banking Supervision, 2009, “Strengthening the resilience of the banking •
sector,” Consultative Document, Bank for International Settlements
Basel Committee on Banking Supervision, 2010a, “Proposal to ensure the loss absorbency of •
regulatory capital at the point of non-viability,” Consultative Document, Bank for International
Settlements
Basel Committee on Banking Supervision, 2010b, “Group of governors and heads of •
supervision announces minimum global standards,” press release, September 12
Bikhchandani, S. and S. Sharma, 2001, “Herd behavior in financial markets,” IMF Staff Papers, •
47, 279–310
Black, F. and M. Scholes, 1974, “The pricing of options and corporate liabilities,” Journal of •
Political Economy, 81:3, 637–54
Borio, C., C. Furfine, and P. Lowe, 2001, “Procyclicality of the financial system and financial •
stability: issues and policy options,” BIS Paper No. 1, Bank for International Settlements
Brunnermeier, K. and L. Pedersen, 2008, “Market liquidity and funding liquidity,” Review of •
Financial Studies, 22:6, 2201-2238
Brunnermeier, K., A. Crockett, C. Goodhart, A. D. Persaud, and H-S. Shin, 2009, “The •
fundamental principles of financial regulation,” Geneva reports on the World Economy, Vol. 11,
International Center for Monetary and Banking Studies
Caruana, J., 2010, “Basel III: towards a safer financial system,” speech delivered at the Third •
Santander International Banking Conference, Madrid, September
Chan-Lau, J.A., 2001, “The impact of corporate governance structures on the agency cost of •
debt,” available at http://ssrn.com/author=96623
Chan-Lau, J.A., 2009, “Default risk codependence in the global financial system: was the Bear •
Stearns bailout justified?” in Gregoriou, G., (ed.), The banking crisis handbook, CRC Press
Chan-Lau, J.A., 2010, “Regulatory capital for too-connected-to-fail institutions: a practical •
proposal,” forthcoming in Financial Markets, Institutions, and Instruments, 19:5, 355–379
Cont, R. and J. P. Bouchaud, 1998, “Herd behavior and aggregation fluctuations in financial •
markets,” working paper
Corrigan, G., 2009, “Containing too big to fail,” The Charles F. Dolan Lecture Series at Fairfield •
University November
Coval, J., J. W. Jurek, and E. Stafford, 2009, “The economics of structured finance,” Journal of •
Economic Perspectives, 23:1, 3–25
Dewatripont, M., J.-C. Rochet, and J. Tirole, 2010, Balancing the banks, Princeton University •
Press
Dudley, W., 2009, “Some lessons learned from the crisis,” speech delivered at the Institute of •
International Bankers Membership Luncheon, New York City, October
The Economist, 2010, “Third time’s the charm?” September 13•
Forbes, K. and R. Rigobon, 2002, “No contagion, only interdependence: measuring stock •
market comovements,” Journal of Finance, 57:5, 2223–2261
French, K. R., M. N. Baily, J. Y. Campbell, J. H. Cochrane, D. W. Diamond, D. Duffie, A. K. •
Kashyap, F. S. Mishkin, R. G. Rajan, D. S. Scharfstein, R. J. Shiller, H. S. Shin, M. J. Slaughter,
J. C. Stein, and R. M. Stulz 2010, The Squam Lake report: fixing the financial system, Princeton
University Press
Gauthier, C., A. Lehar, and M. Souissi, 2010, “Macroprudential regulation and systemic capital •
requirements,” Working Paper No. 2010-4, Bank of Canada
Gibson, M., 2004, “Understanding the risk of synthetic CDOs,” Finance and Economics •
Discussion Series 2004-36, Board of Governors of the Federal Reserve
Hartmann, P., S. Straetmans, and C. de Vries, 2004, “Asset market linkages in crisis periods,” •
The Review of Economics and Statistics, 86:1, 313–326
Hull, J. and A. White, 2004, “Valuation of a CDO and n-th to default CDS without Monte Carlo •
simulation,” Journal of Derivatives, 12:2, 8-23
Kashyap, A. K., R. G. Rajan, and J. C. Stein, 2008, “Rethinking capital regulation,” Proceedings •
of the 2008 Jackson Hole Symposium, Federal Reserve Bank of Kansas City
Lando, D., 2004, Credit risk modeling, Princeton University Press•
Lester, A., L. H. Pedersen, and T. Philippon, 2008, “Systemic risk and macroeconomic capital,” •
working paper, New York University
Li, D. X., 2000, “On default correlation: a copula function approach,” Journal of Fixed Income, •
9:4, 43–54
Merton, R. C., 1974, “On the pricing of corporate debt: the risk structure of interest rates,” •
Journal of Finance, 29:2, 449–470
Nirei, M., 2006, “Herd behavior and fat tails in financial markets,” working paper, Carleton •
University
Rajan, A., G. McDermott, and R. Roy, 2007, The structured credit handbook, Wiley•
Shin, H.-S., 2008, Risk and liquidity, Clarendon lectures in finance, Oxford University Press•
Scharfstein, D. and J. Stein, 1990, “Herd behavior and investment,” American Economic •
Review, 80, 465–479
Strongin, S., A. Hindlian, and S. Lawson, 2009, “Effective regulation: part 5 – ending too big to •
fail,” Global Markets Institute, Goldman Sachs, December
Tarashev, N., C. Borio, and C. Tsatsaronis, 2010, “Attributing systemic risk to individual •
institutions,” Working Paper
Tarullo, D., 2009, “Confronting too big to fail,” speech at The Exchequer Club, Washington, •
D.C., October
Taleb, N., 2009, The black swan, Basic Books•
Triana, P., 2010, “Basel III contains seeds of more chaos,” Financial Times, September 19•
Tucker, P., 2009, “The crisis management menu,” speech at the SUERF, CEPS, and Belgian •
Financial Forum Conference, Crisis management at the cross-roads, Brussels, November
U.S. Securities and Exchange Commision, 2010, “SEC charges Goldman Sachs with fraud in •
structuring and marketing of CDO tied to subprimer mortgages,” press release, April 16
Vacisek, O., 1977, “Probability of loss on loan portfolio,” Moody’s KMV, San Francisco•
The Capco Institute Journal of Financial TransformationFat Tails and (Un)happy Endings: Correlation Bias, Systemic Risk and Prudential Regulation
59
PART 1
Empirical Analysis, Trading Strategies, and Risk Models for Defaulted Debt Securities
AbstractThis study empirically analyzes the historical performance of
defaulted debt from Moody’s Ultimate Recovery Database
(1987-2010). Motivated by a stylized structural model of
credit risk with systematic recovery risk, we argue and find
evidence that returns on defaulted debt covary with determi-
nants of the market risk premium, firm specific and structural
factors. Defaulted debt returns in our sample are observed
to be increasing in collateral quality or debt cushion of the
issue. Returns are also increasing for issuers having supe-
rior ratings at origination, more leverage at default, higher
cumulative abnormal returns on equity prior to default, or
greater market implied loss severity at default. Considering
systematic factors, returns on defaulted debt are positively
related to equity market indices and industry default rates.
On the other hand, defaulted debt returns decrease with
short-term interest rates. In a rolling out-of-time and out-
of-sample resampling experiment we show that our leading
model exhibits superior performance. We also document
the economic significance of these results through excess
abnormal returns, implementing a hypothetical trading strat-
egy, of around 5-6 percent (2-3 percent) assuming zero (1bp
per month) round-trip transaction costs. These results are
of practical relevance to investors and risk managers in this
segment of the fixed income market.
Michael Jacobs, Jr. — Senior Financial Economist, Credit Risk Analysis Division, Office of the Comptroller of the Currency1
The views expressed herein are those of the author and do not necessarily 1
represent a position taken by the Office of the Comptroller of the Currency
or the U.S. Department of the Treasury.
60
There exists an economic argument that to the extent there may be op-
portunity costs associated with holding defaulted debt, and that the per-
formance of such debt may vary systematically, the required return on
the defaulted instruments should include an appropriate risk premium.
Thus far, most research studying systematic variation in defaulted debt
recoveries has focused on the influence of either macroeconomic fac-
tors [Frye (2000 a,b,c; 2003), Hu and Perraudin (2002) Cary and Gordy
(2007), Jacobs (2011)], supply/demand conditions in the defaulted debt
markets [Altman et al. (2003)], or some combination thereof [Jacobs and
Karagozoglu, (2011)]. Probably the reason for this focus is the conven-
tional wisdom that determinants of recoveries (i.e., collateral values) are
thought covary with such systematic macroeconomic measures. How-
ever, the results concerning systematic variation in recoveries have been
mixed. We believe that this is due to the unmeasured factors influencing
the market risk premium for defaulted debt. Adequately controlling for
other determinants of defaulted debt performance, potentially imper-
fectly correlated with standard macroeconomic indicators, is critical to
understanding this.
We propose to extend this literature in several ways. First, we quantify
the systematic variation in defaulted debt returns with respect to factors
which influence the market risk premium for defaulted debt, which are
related to investors’ risk aversion or investment opportunity sets; in the
process, we specify a simple stylized model of credit risk in structural
framework [Merton (1974)], having testable implications that are inves-
tigated herein. Second, we are able to analyze defaulted debt perfor-
mance in segments homogenous with respect to recovery risk, through
controlling for both firm and instrument specific covariates, and examine
whether these are associated with recoveries on defaulted debt securi-
ties. Third, departing from most of the prior literature on recoveries, hav-
ing predominantly focused on measures around the time of default or
at settlement, we will be studying the relationship amongst these in the
form of returns. We believe that such focus is most relevant to market
participants – both for traders and buy-and-hold investors (i.e., vulture
funds, or financial institutions managing defaulted portfolios) – since this
is an accepted measure of economic gain or loss. Finally, we are able to
build parsimonious and robust econometric models, in the generalized
linear model (GLM) class, that are capable of explaining and predicting
defaulted debt returns, and we use these to construct trading strategies
demonstrating their economic significance.
In this study, we quantify the performance of defaulted debt relative to
the previously and newly proposed determinants of corporate debt re-
coveries, through a comprehensive analysis of the returns on this asset
class. The dataset that we utilize, Moody’s Ultimate Recovery Database™
(MURD™), contains the market prices of defaulted bonds and loans near
the time of default, and the prices of these instruments (or market value
of the bundle of instruments) received in settlement (or at the resolution)
of default. We have such data for 550 obligors and 1368 bonds and loans
in the period 1987-2010. We examine the distributional properties of the
individual annualized rates of return on defaulted debt across different
segmentations in the dataset (i.e., default type, facility type, time period,
seniority, collateral, original rating, industry), build econometric models to
explain observed returns, and quantify potential trading gains to deploy-
ing such models.
Our principle results are as follows. We find returns to be in line with
(albeit to the upper end of the range of results) what has been found in
the previous literature, a mean of 28.6 percent.3 We find returns on de-
faulted debt to vary significantly according to contractual, obligor, equity/
debt markets, and economic factors. At the facility structure level, there
is some evidence that returns are elevated for defaulted debt having bet-
ter collateral quality rank or better protected tranches within the capital
structure. At the obligor or firm level, returns are elevated for obligors
rated higher at origination, more financially levered at default, or hav-
ing higher cumulative abnormal returns (CARs) on equity prior to default.
However, we also find returns to be increasing in the market implied loss
severity at default. We also find evidence that while defaulted debt returns
vary counter to the credit cycle, as they increase with industry default
rates, they also increase with aggregate equity market returns. Further,
we observe that short-term interest rates are inversely related to returns
on defaulted debt. Finally, we document the economic significance of
these results through excess abnormal returns, in a debt-equity arbitrage
trading experiment, of around 5-6 percent (2-3 percent) assuming zero
(1bp per month) round-trip transaction costs.
In addition to the relevance of this research for resolving questions in the
finance of distressed debt investing, and aiding practitioners in this space,
our results have implications for recently implemented supervisory Basel II
capital standards for financial institutions [BCBS (2004)]. Our results indi-
cate that time variation in the market risk premium for defaulted debt may
be an important systematic factor influencing recoveries on such instru-
ments (and by implication, their loss-given-default – LGD), which is likely
to not be perfectly correlated with the business cycle. Hence, any financial
institution, in making the decision about how much capital to hold as a
safeguard against losses on corporate debt securities, should take into ac-
count factors such as the systematic variation in investor risk aversion and
Standard portfolio separation theory implies that, all else equal, during episodes of 2
augmented investor risk aversion, a greater proportion of wealth is allocated to risk-
free assets [Tobin (1958), Merton (1971)], implying lessened demand, lower price, and
augmented expected returns across all risky assets.
The probable reason why we are closer to the higher end of estimates, such as Keenan et 3
al (2000), is that we have included several downturn periods, such as the early 1990s and
recently.
61
investment opportunity sets.4 Indeed, Basel II requires that banks quantify
“downturn effects” in LGD estimation [BCBS (2005, 2006)], and for the
relevant kind of portfolio (i.e., large corporate borrowers having marketable
debt), and our research provides some guidance in this regard.
Review of the related literatureAltman (1989) develops a methodology – at the time new to finance – for
the measurement of risk due to default, suggesting a means of ranking
fixed-income performance over a range of credit-quality segments. This
technique measures the expected mortality of bonds, and associated
loss rates, similarly to actuarial tabulations that assess human mortality
risk. Results demonstrate outperformance by risky bonds relative to risk-
less Treasuries over a ten-year horizon and that, despite relatively high
mortality rates, B-rated and CCC-rated securities outperform all other
rating categories in the first four years after issuance, with BB-rated se-
curities outperforming all others thereafter.
Gilson (1995) surveys the market practices of so-called “vulture investors,”
noting that as the risks of such an investment style exposes one to a high
level of idiosyncratic and non-diversifiable risk, those who succeed in this
space must have a mastery of legal rules and institutional setting that gov-
ern corporate bankruptcy. The author further argues that such mastery can
result in very high returns. Hotchkiss and Mooradian (1997) study the func-
tion of this investor class in the governance and reorganization of defaulted
firms using a sample of 288 public debt defaults. They attribute better rela-
tive operating performance after default to vulture investors gaining control
of the target firm in either a senior executive or an ownership role. They
also find positive abnormal returns for the defaulted firm’s equity or debt
in the two days surrounding the public revelation of a vulture purchase of
such instruments. The authors conclude that vulture investors add value
by disciplining managers of distressed firms.
The historical performance of the Moody’s Corporate Bond index [Keenan
et al. (2000)] shows an annualized return of 17.4 percent in the period
1982-2000. However, this return has been extremely volatile, as most of
this gain (147 percent) occurred in the period 1992-1996. Keenan et al.
(2000) and Altman and Jha (2003) both arrive at estimates of a correlation
to the market on this defaulted loan index of about 20 percent, implying a
market risk premium of 216 bps. Davydenko and Strebuleav (2002) report
similar results for non-defaulted high-yield corporate bonds (BB rated) in
the period 1994-1999.
From the perspective of viewing defaulted debt as an asset class, Guha
(2003) documents a convergence in market value as a proportion of par
with respect to bonds of equal priority in bankruptcy approaching de-
fault. This holds regardless of contractual features, such as contractual
rate or remaining time-to-maturity. The implication is that while prior to
default bonds are valued under uncertain timing of and recovery in the
event of default, that varies across issues according to both borrower
and instrument characteristics. Upon default such expectations become
one and the same for issues of the same ranking. There is cross-sectional
variation in yields is due to varied perceived default risk as well as instru-
ment structures, but as default approaches the claim on the debt col-
lapses to a common claim on the expected share of emergence value
of the firm’s assets due to the creditor class. Consequently, the contract
rate on the debt pre-default is no longer the relevant valuation metric with
respect to restructured assets. This was predicted by the Merton (1974)
theoretical framework that credit spreads on a firm’s debt approach the
expected rate of return on the firm’s assets, as leverage increases to the
point when the creditors become the owners of the firm. Schuermann
(2003) echoed the implications of this argument by claiming that cash
flows post-default represent a new asset.
Altman and Jha (2003), regressing the Altman/Solomon Center defaulted
bond index on the S&P 500 returns for the period 1986-2002, come up
with an 11.1 percent required return (based upon a 20.3 percent correla-
tion estimate.) Altman et al. (2003) examine the determinants of recover-
ies on defaulted bonds, in a setting of systematic variation in aggregate
recovery risk, based on market values of defaulted debt securities shortly
following default. The authors find that the aggregate supply of defaulted
debt securities, which tends to increase in downturn periods, is a key de-
terminant of aggregate as well as instrument level recovery rates. The au-
thors’ results suggest that while systematic macroeconomic performance
may be associated with elevated LGD, the principle mechanism by which
this operates is through supply and demand conditions in the distressed
debt markets. More recently, Altman (2010) reports that the Altman-NYU
Salomon Center Index of defaulted bonds (bank loans) returned 12.6 per-
cent (3.4 percent) over the period 1986-2009 (1989-2009).
Machlachlan (2004), in the context of proposing an appropriate discount
rate for workout recoveries for regulatory purposes in estimating economic
LGD [BCBS (2005)], outlines a framework that is motivated by a single
factor CAPM model and obtains similar results in two empirical exercises.
First, regressing Altman-NYU Salomon Center Index of Defaulted Public
Bonds in the period 1987-2002 on the S&P 500 equity index, he obtains a
20 percent correlation, implying a market risk premium (MRP) of 216 bps.
Second, he looks at monthly secondary market bid quotes for the period
April 2002-August 2003, obtaining a beta estimate of 0.37, which accord-
ing to the Frye (2000c) extension of the Basel single factor framework,
implies a recovery value correlation of 0.21 and an MRP of 224 bps.
The Capco Institute Journal of Financial TransformationEmpirical Analysis, Trading Strategies, and Risk Models for Defaulted Debt Securities
Our research also has a bearing on the related and timely issue of the debate about the 4
so-called “pro-cyclicality” of the Basel capital framework [Gordy ( 2003)], an especially
relevant topic in the wake of the recent financial crisis, where a critique of the regulation is
such that banks wind up setting aside more capital just at the time that they should be using
capital to provide more credit to businesses or to increase their own liquidity positions, in
order to help avoid further financial dislocations and help revitalize the economy.
62
Finally, considering studies of recovery rates (orLGDs), Acharya et al.
(2007) examine the empirical determinants of ultimate LGD at the instru-
ment level, and find that the relationship between the aggregate supply
of defaulted debt securities and recoveries does not hold after controlling
for industry level distress. They argue for a “fire-sale effect” that results
when most firms in a troubled industry may be selling collateral at the
same time. These authors’ results imply that systematic macroeconomic
performance may not be a sole or critical determinant of recovery rates
on defaulted corporate debt. Carey and Gordy (2007) examine whether
there is systematic variation in ultimate recoveries at the obligor (firm-
level default incidence) level, and find only weak evidence of system-
atic variation in recoveries. Recently, building upon these two studies,
Jacobs and Karagozoglu (2011) empirically investigate the determinants
of LGD and build alternative predictive econometric models for LGD on
bonds and loans using an extensive sample of most major U.S. defaults
in the period 1985–2008. They build a simultaneous equation model in
the beta-link generalized linear model (BLGLM) class, identifying several
that perform well in terms of the quality of estimated parameters as well
as overall model performance metrics. This extends prior work by model-
ing LGD both at the firm and the instrument levels. In a departure from
the extant literature, the authors find the economic and statistical sig-
nificance of firm-specific, debt, and equity-market variables; in particular,
that information from either the equity or the debt markets at around the
time of default (measures of either distress debt prices or cumulative
equity returns, respectively) have predictive powers with respect to the
ultimate LGD, which is in line with recent recovery and asset pricing re-
search. They also document a new finding, that larger firms (loans) have
significantly lower (higher) LGDs.
Theoretical frameworkIn this section we lay out the theoretical basis for returns on post-default
recoveries, denoted rsD, where s denotes a recovery segment (i.e., senior-
ity classes, collateral types, etc.). Following an intertemporal version of
the structural modelling framework for credit risk [(Merton (1971), Vasicek
(1987, 2002)5], we may write the stochastic process describing the instan-
taneous evolution of the ith firm’s5 asset returns at time t as: dVi,t/Vi,t = µidt
+ σiWi,t (1), where Vi,t is the asset value, σi is the return volatility, µi is the
drift (which can be taken to be the risk-free rate r under risk-neutral mea-
sure), and Wi,t is a standard Weiner process that decomposes as (this is
also known as a standardized asset return): dWi,t = ρi,XdXt + (1- ρ2i,X)1/2
dZi,t (2), where the processes (also standard Weiners) Xt and Zi,t are the
systematic risk factors (or standardized asset returns) and the idiosyncratic
(or firm-specific) risk factors, respectively; and the factor loading ρi,X is
constant across all firms in a PD segment homogenous with respect to
default risk (or across time for the representative firm).7 It follows that the
instantaneous asset-value correlation amongst firms (or segments) i and j
is given by: 1/dt · CorVi,j [dVi,t/Vi,t , dVj,t/Vj,t] = ρi,Xρj,X (3).
Defining the recovery rate on the ith defaulted asset8 at time t as Ri,t, we
may similarly write the stochastic process describing its evolution as:
dRi,t/Ri,t = µiRdt + σi
RdWRi,t (4), where µi
R is the drift (which can be taken to
be the expected instantaneous return on collateral under physical mea-
sure, or the risk-free rate under risk-neutral measure), σiR is the volatility
of the collateral return, and WRi,t is a standard Weiner process that for re-
covery segment SR decomposes as: dWRi,t = ρi,sR dXt
R + (1 – ρ2i,sR)1/2 dZR
i,t
(5), where the two-systematic factors are bivariate standard normal, each
standard normal, but with correlation r between each other:
( ) 0 1, ~ ,
0 1TR
t t
rdX dX N
r
⎛ ⎞⎛ ⎞ ⎛ ⎞⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠⎝ ⎠ (6)
This set-up follows various extensions of the structural model framework
for systematic recovery risk. What they have in common is that they allow
the recovery process to depend upon a second systematic factor, which
may be correlated with the macro (or market) factor Xt [Frye (2000 a, b,
c), Pyktin (2003), Dullman and Trapp (2004), Giese (2005), Rosch and
Scheule (2005), Hillebrand (2006), Barco (2007), Jacobs (2011)]. In this
general and more realistic framework, returns on defaulted debt may be
governed by a stochastic process distinct from that of the firm. This is
the case where the asset is secured by cash, third party guarantees, or
assets not used in production. In this setting, it is possible that there are
two salient notions of asset value correlation, one driving the correlation
amongst defaults, and another driving the correlation between collateral
values and the returns on defaulted assets in equilibrium. This reasoning
implies that it is entirely conceivable that, especially in complex bank-
ing facilities, cash flows associated with different sources of repayment
should be discounted differentially according to their level of systematic
risk. In not distinguishing how betas may differ between defaulted instru-
ments secured differently, it is quite probable that investors in distressed
debt may misprice such assets.
It is common to assume that the factor loading in (5) is constant amongst
debt instruments within specified recovery segments, so that the recov-
ery-value correlation for segment SR is given by ρ2i,sR ≡ RsR.9 If we take
the further step of identifying this correlation with the correlation to a
market portfolio – arguably a reasonable interpretation in the asymptotic
Note that this is also the approach underlying the regulatory capital formulae [BCBS (2003)], 5
as developed by Gordy (2003).
This could also be interpreted as the ith PD segment or an obligor rating.6
Vasicek (2002) demonstrates that under the assumption of a single systematic factor, an 7
infinitely granular credit portfolio, and LGD that does not vary systematically, a closed-form
solution for capital exists that is invariant to portfolio composition.
We can interpret this as an LGD segment (or rating) or debt seniority class.8
Indeed, for many asset classes the Basel II framework mandates constant correlation 9
parameters equally across all banks, regardless of particular portfolio exposure to industry
or geography. However, for certain exposures, such as wholesale non-high volatility
commercial real estate, this is allowed to depend upon the PD for the segment or rating
[BCBS (2004]).
63
The Capco Institute Journal of Financial TransformationEmpirical Analysis, Trading Strategies, and Risk Models for Defaulted Debt Securities
single risk-factor (ASRF) framework [(Vasicek (1987), Gordy (2003)] – then
we can write R2sR = ρ2
sR,M
. It then follows from the standard capital asset
pricing model (CAPM) that the relationship between the defaulted debt
instrument and market rates of return is given by the beta coefficient:
, ,,
, ,
,,
,
,R
R
R
i t M tRs Mii t M t s
s MMM t
MM t
dR dVCov
RR V
dVVar
V
σβ
σ
⎡ ⎤⎢ ⎥⎣ ⎦ = =⎡ ⎤⎢ ⎥⎣ ⎦ (7)
Where σM is volatility of the market return. We may now conclude that in
this setting the return on defaulted debt on the sth exposure (or segment)
rsD is equal to the expected return on the collateral, which is given by the
sum of the risk-free rate rrf and a debt-specific risk-premium δs:
( ) ,
R
R R
Ri sD
s rf M rf rf rfs M sM
Rr r r r r MRP r
σβ δ
σ= + − = + = +
(8)
Where the market risk premium is given by MRP ≡ rM – rrf (also assumed
to be constant through time) and the debt-specific risk premium is given
by δsR = βsR,M
MRP. This approach identifies the systematic factor with the
standardized return on a market portfolio rM, from which it follows that
the asset correlation to the former can be interpreted as a normalized
“beta” in a single factor CAPM (or just a correlation between the de-
faulted debt’s and the market’s return), which is given by ρi,sR ≡ (RsR)1/2.
In subsequent sections, we pursue alternative estimations ρ̂i,sR, through
regressing actual defaulted debt returns on some kind of market factor
or other measure of systematic risk (i.e., aggregate default rates),10 while
controlling for firm or instrument specific covariates.
Empirical methodologyWe adopt a simple measure, motivated in part by the availability of a rich
dataset of defaulted bonds and loans available to us, which analyzes the
observable market prices of debt at two points in time: the default event
(i.e., bankruptcy or other financial distress qualifying as a default) and the
resolution of the default event (i.e., emergence from bankruptcy under
Chapter 11 or liquidation under Chapter 7). We calculate the annualized
rate of return on the ith defaulted debt instrument in segment j as:
, ,,
,
1
, ,,
, ,
1E Di j i jE
i j
Di j
E t ti j tD
i j D
i j t
Pr
P
−⎛ ⎞⎜ ⎟= −⎜ ⎟⎝ ⎠ (9)
where
,, , Di j
D
i j tP (
,, , Ei j
E
i j tP ) are the prices of debt at time of default t i,j
D (emer-
gence t i,jE ). An estimate for the return, the jth segment (seniority class of
collateral type), can then be formed as the arithmetic average across the
loans in that segment:
, ,,
,
1
, ,
1 , ,
1 1E DDi j i jj E
i j
Di j
E t tNi j tD
j D Dij i j t
Pr
N P
−
=
⎡ ⎤⎛ ⎞⎢ ⎥⎜ ⎟= −⎢ ⎥⎜ ⎟⎢ ⎥⎝ ⎠⎣ ⎦
∑ (10)
Where NjD is the number of defaulted loans in the recovery group j. A
measure of the recovery uncertainty in recovery class s is given by the
sample standard error of the mean annualized return:
, ,,
,
21
, ,
1 , ,
1ö 11
E DDi j i jj E
i j
Dj
Di j
E t tNi j t D
jD Drij i j t
Ps r
N P
−
=
⎡ ⎤⎧ ⎫⎛ ⎞⎢ ⎥⎪ ⎪⎜ ⎟= − −⎢ ⎥⎨ ⎬
− ⎜ ⎟⎢ ⎥⎪ ⎪⎝ ⎠⎢ ⎥⎩ ⎭⎣ ⎦
∑
(11)
Empirical results: summary statistics of returns on defaulted debt by segmentIn this section and the following, we document our empirical results.
These are based upon our analysis of defaulted bonds and loans in the
Moody’s Ultimate Recovery Database™ (MURD™) release as of August,
2010. This contains the market values of defaulted instruments at or near
the time of default,11 as well as the values of such pre-petition instruments
(or of instruments received in settlement) at the time of default resolution.
This database is largely representative of the U.S. large-corporate loss
experience, from the late 1980s to the present, including most of the
major corporate bankruptcies occurring in this period.
Table A1, in the Appendix, summarizes basic characteristics of simple
annualized return on defaulted debt (RDD) in (10) by default event type
(bankruptcy under Chapter 11 versus out-of-court settlement) and instru-
ment type (loans – broken down by term and revolving versus bonds).
Here we also show the means and standard deviations of two other
key quantities: the time-to-resolution (i.e., time from default to time of
resolution) and the outstanding-at-default, for both the RDD sample as
well as for the entire MURD™ database (i.e., including instruments not
having trading prices at default). We conclude from this that our sample
is for the most part representative of the broader database. Across all
instruments, average time-to-resolution is 1.6 (1.4) years and average
outstanding at default is U.S.$216.4M (U.S.$151.7M) for the analysis
(broader) samples.
Alternatively, we can estimate the vector of parameters (10 µ, µiR, ρi,s, ρi,sR, r)T by full-
information maximum likelihood (FIML), given a time series of default rates and realized
recovery rates. The resulting estimate ρ̂i,sR can be used in equation (8) – in conjunction with
estimates of the market volatility σM, debt-specific volatility σiR, the MRP (rM – rf), and the
risk-free rate rf – in order to derive the theoretical return on defaulted debt within this model
[Machlachlan (2004)]. Also see Jacobs [2011] for how these quantities can be estimated
from prices of defaulted debt at default and at emergence of different seniority instruments.
Experts at Moody compute an average of trading prices from 30 to 45 days following the 11
default event, where each daily observation is the mean price polled from a set of dealers
with the minimum/maximum quote thrown out.
64
The bottom panel of Table A1 represents the entire Moody’s database,
whereas the top panel summarizes the subset for which we can calcu-
late RDD measures. The version of MURD™ that we use contains 4,050
defaulted instruments, 3,500 (or 86.4 percent) of which are bankruptcies,
and the remaining 550 are distressed restructurings. On the other hand,
in the RDD subset, the vast majority (94.6 percent or 1,322) of the total
(1,398) are Chapter 11. One reason for this is that the times-to-resolution
of the out-of-court settlements are so short (about 2 months on average)
that it is more likely that post-default trading prices at 30-45 days from
default are not available. Second, many of these were extreme values
of RDD, and were heavily represented in the outliers that we chose to
exclude from the analysis (30 of 35 statistical outliers).12
The overall average of the 1,398 annualized RDDs is 28.6 percent, with a
standard error of the mean of 3.1 percent, and ranging widely from -100
percent to 893.8 percent. This suggests that there were some very high
returns – as the 95th percentile of the RDD distributions is 191 percent,
or that in well over 70 cases investors would have more than doubled
their money holding defaulted debt. We can observe this in Figure 1,
the distribution of RDD, which has an extremely long tail to the right. We
observe that the distribution of RDD is somewhat different in the case
of out-of-court settlements as compared to bankruptcies, with respec-
tive mean RDDs of 37.3 percent for the former, and 28.1 percent in the
latter. The standard errors of mean RDDs are also much higher in the
non-bankruptcy population, 15.3 percent for out-of-court versus 3.2 per-
cent for bankruptcies. The data is well-represented by bank loans, 36.8
percent (38.1 percent) of the RDD total MURD™ sample, or 514 (1543)
out of 1398 (4050) instruments. Loans appear to behave somewhat dif-
ferently than bonds, having slightly higher mean and standard error of
mean RDDs, 32.1 percent and 26.4 percent, respectively.
Table A2 summarizes the distributional properties of RDD by senior-
ity rankings (bank loans; senior secured, unsecured and subordinated
bonds; and junior subordinated bonds) and collateral types.13 Generally,
since this does not hold monotonically across collateral classes or is con-
sistent across recovery risk measures, better secured or higher ranked
instruments exhibit superior post-default return performance. However,
while the standard error of mean RDD (which we can argue reflects re-
covery uncertainty) tends to be lower for more senior instruments, it
tends to be higher for those which are better secured. Average RDD is
significantly higher for secured as compared to unsecured facilities, 34.5
percent versus 23.6 percent respectively. Focusing on bank loans, we
see a wider split of 33.0 percent versus 19.8 percent for secured and
unsecured, respectively. However, by broad measures of seniority rank-
ing, mean RDD exhibits a non-monotonic increasing pattern in seniority,
while the standard error of RDD is decreasing in seniority. Average RDD
is 32.3 percent and 36.6 percent for loans and senior secured bonds, as
compared to 23.7 percent and 33.2 percent for senior secured and senior
subordinated bonds, decreasing to 15.6 percent for junior subordinated
instruments. However, while unsecured loans have lower post-default
Based upon extensive data analysis in the Robust Statistics package of the S-Plus 12
statistical computing application, we determined 35 observations to be statistical outliers.
The optimal cutoff was determined to be about 1,000%, above which we removed the
observation from subsequent calculations. There was a clear separation in the distributions,
as the minimum RDD in the outlier subset is about 17,000%, more than double the
maximum in the non-outlier subset.
We have two sets of collateral types: the 19 lowest level labels appearing in MURD™ 13
(Guarantees, Oil and Gas Properties, Inventory and Accounts Receivable, Accounts
Receivable, Cash, Inventory, Most Assets, Equipment, All Assets, Real Estate, All Non-
current Assets, Capital Stock, PP&E, Second Lien, Other, Unsecured, Third Lien, Intellectual
Property and Intercompany Debt), and a six level high level grouping of that we constructed
from the (Cash, Accounts Receivables & Guarantees; Inventory, Most Assets & Equipment;
All Assets & Real Estate; Non-Current Assets & Capital Stock; PP&E & Second Lien; and
Unsecured & Other Illiquid Collateral). The latter high-level groupings were developed with
in consultation with recovery analysis experts at Moody’s Investors Services.
Quintiles of time from default to resolution date
1 2 3 4 5 Total
Average
(%)
Std error
of the
mean (%)
Average
(%)
Std error
of the
mean (%)
Average
(%)
Std error
of the
mean (%)
Average
(%)
Std error
of the
mean (%)
Average
(%)
Std error
of the
mean (%)
Average
(%)
Std error
of the
mean (%)
Quintiles
of time
from last
cash pay
to default
date
1 64.19 24.57 25.75 26.11 38.32 13.54 29.75 14.08 -4.99 9.21 35.04 9.66
2 22.10 15.41 38.34 17.09 28.24 19.03 26.69 9.21 8.23 6.82 25.93 6.78
3 20.81 12.16 30.55 18.16 10.04 8.12 27.19 11.21 8.90 5.06 19.28 5.26
4 91.53 31.75 41.38 19.92 19.79 9.16 23.55 6.26 8.96 3.91 28.67 5.51
5 92.08 34.68 57.99 20.85 8.82 8.21 34.22 20.16 -2.97 8.57 38.32 9.23
Total 58.90 11.57 39.71 8.89 20.02 5.71 27.32 4.99 6.03 2.61 28.56 3.11
Table 1 – Returns on defaulted debt (RDD)1 of defaulted instruments by quintiles of time-to-resolution (TTR)2 and time-in-distress (TID)3 from last cash pay to default date (Moody’s Ultimate Recovery Database 1987-2010)1 – Annualized “return on defaulted debt” from just after the time of default (first trading date of debt) until the time of ultimate resolution.
2 – TTR: Duration in years from the date of default (bankruptcy filing or other default) to the date of resolution (emergence from bankruptcy or other settlement).
3 – TID: Duration in years from the date of the last interest payment to the date of default (bankruptcy filing or other default).
65
The Capco Institute Journal of Financial TransformationEmpirical Analysis, Trading Strategies, and Risk Models for Defaulted Debt Securities
returns than secured loans, within the secured loan class we find that
returns exhibit a humped pattern as collateral quality goes down in rank,
an increase in RDD from 22.6 percent for cash, to 46.2 percent for “All
assets and real estate,” to 29.0 percent for “PP&E and second lien.”
Table 1 summarizes RDDs by two duration measures: the “time-in-dis-
tress” (TID), defined as the time (in years) from the last cash pay date to
the default state, and the “time-to-resolution” (TTR), the duration from
the date of default to the resolution or settlement date. Analysis of these
measures helps us to understand the term-structure of the defaulted
debt returns. We examine features of RDD by quintiles of the TTR and
TID distributions, where the first refers to the bottom fifth of durations in
length, and the fifth quintile the top longest. The patterns we observe are
that RDD is decreasing (albeit non-monotonically) in TTR, while it exhibits
a U-shape in TID.
Table 2 summarizes RDD by the earliest available Moody’s senior un-
secured credit rating for the obligor. This provides some evidence that
returns on defaulted debt are augmented for defaulted obligors that had,
at origination (or time of first public rating), better credit ratings or higher
credit quality. Mean RDD generally declines as credit ratings worsen, al-
beit unevenly. While the average is 22.9 percent for the AA-A category, it
goes up to 45.1 percent for BBB, then down to 17.9 percent for BB, but
up again to 31.6 percent for B, and finally down to 21.99 percent for the
lowest category CC-CCC.
Table 3 summarizes RDD by measures of the relative debt cushion of the
defaulted instrument. MURDTM provides the proportion of debt either
above (degree of subordination) or below (debt cushion) any defaulted in-
strument, according to the seniority rank of the class to which the instru-
ment belongs. It has been shown that the greater the level of debt below,
or the less debt above, the better the ultimate recovery on the defaulted
debt [Keisman et al. (2000)]. We can also think of this position in the
capital structure in terms of “tranche safety” – the less debt above, more
debt below, then the more likely it is that there will be some recovery.
While this is not the entire story, this measure has been demonstrated to
be an important determinant of ultimate recovery, so we suspect that it
will have some bearing on the performance of defaulted debt. Here, we
offer evidence that returns on defaulted debt are increasing in the degree
of tranche safety, or relative debt cushion, as measured by the difference
between debt below and debt above. To this end, we define the tranche
safety index (TSI) as:
TSI ≡ ½[%debt below – %debt above + 1] (12)
This ranges between zero and 1. When it is near zero the difference be-
tween the debt above and below is greatest (i.e., the thinnest tranche or
the most subordinated), and closest to unity when debt below is maximized
and the debt above is nil (i.e., the thickest tranche or the greatest debt
cushion). In Table 3, we examine the quintiles of the TSI, where the bottom
20th percentile of the TSI distribution represents the least protected instru-
ments, and the top 20th percentile the most protected. Additionally, we
define several dummy variables in order to capture this phenomenon, as
in Brady et al. (2006). “No debt above and some debt below” (NDA/SDB)
represents a group that should be the best protected, while “Some debt
above and some debt below” (SDA/SDB) and “No debt above and no debt
below” (NDA/NDB) represent intermediate groups, and “No debt below
Count Average of
RDD
Standard error
of mean RDD
Rating
groups
AA-A 146 22.94% 5.04%
BBB 586 45.09% 13.25%
BB 285 17.92% 5.24%
B 65 31.57% 5.66%
CC-CCC 125 21.99% 8.29%
Investment grade
(BBB-A)
211 29.77% 8.43%
Junk grade
(CC-BB)
996 23.71% 5.94%
Total 1398 28.56% 3.34%
Table 2 – Returns on defaulted debt1 of defaulted instruments by credit rating at origination (Moody’s Ultimate Recovery Database 1987-2010)1 – Annualized “Return on defaulted debt” (RDD) from just after the time of default
(first trading date of debt) until the time of ultimate resolution.
Count Average
of RDD
Standard error
of mean RDD
Debt
tranche
groups
1st quintile TSI 172 35.06% 12.88%
2nd quintile TSI 373 10.98% 4.85%
3rd quintile TSI 413 25.77% 5.40%
4th quintile TSI 342 42.41% 6.33%
5th quintile TSI 98 47.48% 9.57%
NDA / SDB1 449 42.77% 4.89%
SDA / SDB2 259 24.06% 7.65%
NDA / NDB3 164 25.23% 9.44%
NDB / SDA4 526 19.67% 5.25%
Total 1398 28.56% 3.11%
1 – No debt above and some debt below.
2 – Some debt above and some debt below.
3 – No debt above and no debt below.
4 – No debt below and some debt above.
Table 3 – Returns on defaulted debt5 of defaulted instruments by Tranche Safety Index6 (TSI) quintiles and categories (Moody’s Ultimate Recovery Database 1987-2010)5 – Annualized “return on defaulted debt” (RDD) from just after the time of default
(first trading date of debt) until the time of ultimate resolution.
6 – An index of the tranche safety calculated as TTS = (% debt below – % debt above + 1)/2.
66
total assets. Results show generally a negative correlation between cash
flow ratios and RDD, notably a strong negative correlation for FAR of 9.0
percent. The intuition here may be considered strained, as it is natural to
think that the ability to throw off cash may signal a firm with an underlying
business model that is viable, which is conducive to a successful emer-
gence from default and well performing debt; however, this may also be
taken to mean an “excess” of cash with not good investments to apply it
to and a basically poor economic position.
Finally for the financials, we have a set of variables that measure some no-
tion of accounting profitability: net income/book value of total assets, net
income/market value of total assets, retained earnings/book value of total
assets, return on assets, and return on equity. These have generally a mod-
est inverse relation to RDD. As with other dimensions of risk considered
here, we resort to a “backward story,” relative to the expectation that least-
bad profitability mitigates credit or default risk: that is, if already in default,
then better accounting profitability may be a harbinger of deeper woes for
the firm, as reflected in the better returns on the debt from default to reso-
lution of default. However, none of these enter the multiple regressions.
Equity price performance metrics were extracted from CRSP at the
date nearest to the first default date of the obligor, but no nearer than
on month to default. These are shown in the second from top panel of
Table A3. The 1-month equity price volatility, the standard deviation of
daily equity returns in the month prior to default, exhibits a small modest
positive correlation of 2.5 percent to RDD. This sign is explainable by an
option theoretic view of recoveries, since the value of a call-option on the
residual cash flows of the firms to creditors are expected to increase in
asset value volatility, which is reflected to some degree in equity volatility.
On the other hand, the one-year expected equity return, defined as the
average return on the obligor’s stock in excess of the risk-free rate the
year prior to default, exhibits a modest degree of negative correlation
(-6.4 percent) which we find to be somewhat puzzling. The cumulative
abnormal returns on equity, the returns in excess of a market model in
the 90 days prior to default, have the strongest positive relationship to
RDD of the group, 10.3 percent. This is understandable, as the equity
markets may have a reasonable forecast of the firm’s ability to become
rehabilitated in the emergence from default, as reflected in “less poor”
stock price performance relative to the market. Note this is one of two
variables in this group that enters the candidate regression models, and
is also the basis of our trading analysis. Market capitalization of the firm
relative to the market as a whole, defined as the logarithm of the scaled
market capitalization,14 also has a significant negative univariate corre-
lation to the RDD of -8.6 percent, and enters all of the regressions, as
does CAR. We have no clear a priori expectation for this variable, as
The scale factor is defined as the market capitalization of the stock exchange where the 14
obligor trades times 10,000.
and some debt above” (NDB/SDA) should be the least protected group.
Table 3 shows that there is there is U-shape overall in average RDD with
respect to quintiles of TSI: starting at 35.1 percent at the bottom quintile,
having a minimum in the second of 11.0 percent, and increasing thereafter
to 25.8 percent, 42.3 percent and 47.5 percent at the top. With regards
to the dummy variables, we observe a general decrease in average RDD,
from the most to the least protected categories: 42.8 percent, 24.1 per-
cent, 25.2 percent, and 19.7 percent from NDA/SDB to NDB/SDA.
Summary statistics and distributional properties of covariatesIn this section we first analyze the independent variables available to
us and calculated from MURD™, as well as data attached to this from
Compustat and CRSP, and then discuss a multivariate regression model
to explain RDD. Table A3 in the Appendix summarizes the distributional
properties of key covariates in our database and their univariate corre-
lations to RDD. We have grouped these into the following categories:
financial statement and market valuation, equity price performance, capi-
tal structure, credit quality/credit market, instrument/contractual, macro/
cyclical, and durations/vintage.
The financial variables, alone or in conjunction with equity market met-
rics, are extracted from Compustat or CRSP. The Compustat variables
are taken from the date nearest to the first instrument default of the ob-
ligor, but no nearer than one month, and no further than one year, to
default. These are shown in the top panel of Table A3 in the Appendix.
First, we see some evidence that leverage is positively related to RDD,
suggesting that firms that were nearer to their “default points” prior to
the event had defaulted debt that performed better over the resolution
period, all else equal. This is according to an accounting measure, book
value of total liabilities/book value of total assets, which has a substantial
positive correlation of 17.2 percent.
Next, we consider a set of variables measuring the degree of market valu-
ation relative to stated value, or alternatively the degree of intangibility in
assets: Tobin’s Q, market value of total assets/book value of total assets
(MVTA/BVTA), book value of intangibles/book value of total assets, and the
price/earnings ratio. In this group, there is evidence of a positive relation-
ship to the RDD, which is strongest by far for MVTA/BVTA, having a cor-
relation of 18.5 percent. This enters into some of our candidate regression
models significantly, but not the final model chosen. We speculate that the
intuition here is akin to a “growth stock effect” – such types of firms may
have available a greater range of investment options, that when come to
fruition result in better performance of the defaulted debt on average.
We display 3 covariates in Table A3 that measure the cash-flow gen-
erating ability of the entity: free asset ratio (FAR), free cash flow/book
value of total assets, and the cash flow from operations/book value of
67
The Capco Institute Journal of Financial TransformationEmpirical Analysis, Trading Strategies, and Risk Models for Defaulted Debt Securities
perhaps we would expect larger companies to have the “resiliency” to
better navigate financial distress, counter to what we are measuring. The
stock price relative to the market, which is the percentile ranking or the
absolute level of the stock price in the market, has a moderate negative
correlation to RDD of -4.4 percent. As this variable is intended to capture
the delisting effect when a stock price goes very low, we might expect the
opposite sign on this correlation. Finally, the stock price trading range,
defined as the stock price minus its three-year low divided by the differ-
ence between its three-year high and three-year low, is showing only a
small negative correlation to RDD of -2.9 percent. This is another coun-
terintuitive result, as one might expect that when a stock is doing better
than its recent range that it should be a higher quality firm whose debt
might have a better performance in default, but the data is not showing
that, or much less of any kind of relationship here.
Capital structure metrics, extracted from the MURD™ data at the default
date of the obligor, are shown in the third from top panel of Table A3. The
two measures of capital structure complexity, number of instruments (NI)
and number of creditor classes (NCC), show an inverse relationship to
defaulted debt performance. NI (NCC) has a modest negative correlation
to RDD of -4.0 percent (-3.0 percent). We might expect a simpler capital
structure to be conducive to favorable defaulted debt performance ac-
cording to a coordination story. Note that neither of these variables en-
ters the final regression models. While most companies in our database
have relatively simple capital structures, with NI and NCC having medi-
ans of 6 and 2, respectively, there are some rather complex structures
(the respective maxima are 80 and 7).
We have three variables in this group that measure the nature of debt compo-
sition: percent secured debt (PSCD), percent bank debt (PBD) and percent
subordinated debt (PSBD). The typical firm in our database has approxi-
mately 40 percent to 50 percent of its debt either secured, subordinated,
or bank funded. All of these exhibit moderate positive correlation to RDD
of 8.8 percent, 9.4 percent, and 8.7 percent for PSCD, PBD, and PSBD,
respectively. The result on PBD may be attributed to either a monitoring on
the one hand, or alternatively an “optimal foreclosure boundary choice,”
kind of story [Carey and Gordy (2007), Jacobs (2011)]. However, as with the
complexity variables, none of these appear in the regression model.
The credit quality/credit market metrics were extracted from the MURD™
database and Compustat just before the default date of the obligor. These
are shown in the fourth from top panel of Table A3. Two of the variables
in this group have, what may seem to be at first glance, counterintuitive
relationships to RDD. First, the Altman Z-Score, which is available in Com-
pustat, has a relatively large negative correlation of -8.8 percent (note that
higher values of the Z-score indicate lower bankruptcy risk). Second, the
LGD implied by the trading price at default, which forms the basis for the
RDD calculation, exhibits a moderate positive correlation to RDD of 11.3
percent. As this variable has been shown to have predictive power for ul-
timate LGD [Emery et al. (2007), Jacobs and Karagozoglu (2011)], at first
glance this relationship may seem difficult to understand. But note that
the same research demonstrates that LGD at default is also an upwardly
biased estimate of ultimate LGD in some sense. Consequently, we might
just as well expect the opposite relationship to hold, as intuitively it may be
that otherwise high quality debt may perform better on average if it is (per-
haps unjustifiably) “beaten down.” Indeed, LGD enters all of our regression
models with this sign, and as a more influential variable than suggested by
this correlation, but the Z-score does not make it to any of our regression
models. The remaining variables in this group are reflective of the Moody’s
ratings at the first point that the debt is rated. These are the Moody’s Origi-
nal Credit Rating Investment Grade Dummy (MOCR-IG), Moody’s Original
Credit Rating – Major Code (MOCR-MJC; i.e., numerical codes for whole
rating classes), Moody’s Original Credit Rating – Minor Code (MOCR-MNC;
i.e., numerical codes for notched rating classes) and Moody’s Long Run De-
fault Rate – Minor Code (MLRDR-MNC; i.e., empirical default rates associ-
ated with notched rating classes). The only meaningful univariate result here
is the small positive correlation of 2.4 percent in the case of MOCR-IG. This
variable enters significantly into all of our candidate regression models.
Next we consider instrument/contractual metrics, extracted from the
MURD™ database at the default date of the obligor. These are shown in
the third from bottom panel of Table A3. Consistent with the analysis of the
previous section, the correlations with RDD in this group reflect the extent
to which instruments which are more senior, better secured, or in a safer
tranches experience better performance of defaulted debt. The seniority
rank (SR) and collateral rank (CR) codes both have negative and reasonably
sized correlation coefficients with RDD, -9.6 percent and -10.0 percent for
SR and CR, respectively. Percent debt below and percent debt above are
positively (negatively) correlated to RDD, coefficients of 9.4 percent (-5.2
percent). And the TSI, constructed from the latter two variables as detailed
in the previous section, has a significant positive correlation with RDD of 9.7
percent. TSI enters into two of our three candidate regression models.
In this section we consider macroeconomic/cyclical metrics measured
near the default date of the obligor. These are shown in the second from
bottom panel of Table A3. These correlations are evidence that defaulted
debt returns vary counter-cyclically with respect to the credit cycle, or
that debt defaulting in downturn periods tends to perform better. We
have measures of the aggregate default rate, extracted from Moody’s
Default Rate Service (DRS™) database. These are lagging 12-month de-
fault rates, with cohorts formed on an overlapping quarterly basis.15 The
For example, the default rate for the fourth quarter of 2008 would represent the fraction 15
of Moody’s rated issuers in the beginning of 4Q07 that defaulted over the subsequent
year. We follow the practice of adjusting for withdrawn ratings by subtracting one-half the
number of withdrawn obligors from the number of available-to-default (or the denominator
of the default rate.)
68
four versions of this are for the all-corporate and speculative grade seg-
ments, both in aggregate and by industry. All of these have a mild, albeit
significant, positive linear correlations with RDD. The Moody’s All-Corpo-
rate Quarterly Default Rate (MACQDR), having a 6.7 percent correlation
with RDD, is one of the systematic risk variables to enter the candidate
regression models.
The next set of variables represent measures of aggregate equity and
money market performance, the Fama and French (FF) portfolio returns
commonly used in the finance literature, measured on a monthly basis in
the month prior to instrument default.16 These are excess return on the
market (FF-ERM), relative return on small stocks17 (FF-ERSS), and the rel-
ative return on value stocks18 (FF-ERVS). We see that RDD is somewhat
positively associated with aggregate return on the market factor FF-ERM,
having a modest correlation of 7.2 percent.19 Similarly, RDD is positively
but weakly related to FF-RRSS, a correlation of only 2.8 percent. On the
other hand, RDD seems to have a small negative correlation to FF-RRVS
of -4.3 percent. We have one more aggregate equity market variable,
two-year stock market volatility, defined as the standard deviation of the
S&P 500 return in the two years prior to default, which shows a modest
positive linear correlation to RDD of 5.7 percent. Note that FF-ERM is
the only of these aggregate equity market variables to enter significantly
in the multiple regression models. Another set of systematic variables
are aggregate interest rates, the one-month treasury bill yield and the
ten-year treasury bond yield, which exhibit moderate negative correlation
to RDD of -10.2 percent and -7.0 percent, respectively. However, only
the one-month treasury bill yield appears in the final regressions. The
intuition here may be that defaulted debt performs better in low interest
rate environments, which is associated with lower aggregate economic
activity, as well as a higher marginal utility of consumption on the part of
investors.20
The final set of variables that we consider in this section are duration/vin-
tage metrics, based on calculations from extracted dates in the MURD™
database. These are shown in the bottom panel of Table A3. We can
conclude from this section that the duration/vintage measures that would
be in one’s information set at the time of instrument default are largely
Variables Model 1 Model 2 Model 3
Partial effect P-value Partial effect P-value Partial effect P-value
Intercept 0.3094 1.42E-03 0.5101 9.35E-04 0.4342 6.87E-03
Moody’s 12-month lagging speculative grade default rate by industry 2.0501 1.22E-02 2.2538 6.94E-03 2.1828 1.36E-02
Fama-French excess return on market factor 1.3814 8.73E-03 1.5085 6.35E-03 1.5468 9.35E-03
Collateral rank secured 0.2554 7.21E-03 0.2330 1.25E-02 0.2704 9.36E-04
Tranche safety index 0.4548 3.03E-02 0.4339 3.75E-02
Loss given default 0.3273 1.44E-02 0.2751 3.88E-02
Cumulative abnormal returns on equity prior to default 0.3669 1.51E-03 0.3843 1.00E-03 0.4010 9.39E-04
Total liabilities to total assets 0.2653 5.22E-08
Moody’s original rating investment grade 0.2118 2.80E-02 0.2422 6.84E-03 0.1561 6.25E-02
One-month treasury yield -0.4298 3.04E-02 -0.3659 1.01E-02 -0.4901 3.36E-02
Size relative to the market 0.0366 4.76E-02 0.0648 3.41E-03
Market value to book value 0.1925 2.64E-05 0.1422 5.63E-03
Free-asset ratio -0.2429 2.25E-02
Degrees of freedom 959 958 783
Log-likelihood -592.30 -594.71 -503.99
McFadden pseudo R-squared (in-sample) 32.48% 38.80% 41.73%
McFadden pseudo R-squared (out-of-sample) – bootstrap mean 21.23% 12.11% 17.77%
McFadden pseudo R-squared (out-of-sample) – bootstrap standard error 2.28% 1.16% 1.70%
Table 4 – Beta-link generalized linear model for annualized returns on defaulted debt1 (Moody’s Ultimate Recovery Database 1987-2008)1 – Annualized “return on defaulted debt” (RDD) from just after the time of default (first trading date of debt) until the time of ultimate resolution.
These can be downloaded from Kenneth French’s website: http://mba.tuck.dartmouth.edu/16
pages/faculty/ken.french/data_library.html
This is more commonly termed the “small minus large” (SML) portfolio [Fama and French 17
(1992)].
This is more commonly termed the “high minus low” (HML) portfolio, meaning high versus 18
low book-to-market ratio [Fama and French (1992)].
Results for the S&P 500 return, not shown, are very similar.19
The term spread, or the difference in a long and short term treasury yield, was neither 20
significant on a univariate basis nor in the regressions. This held across several different
choices of term structures. Consequently, we do not show these results.
69
uninformative regarding the performance of defaulted debt. The variables
that we have chosen to display include time from origination to default,
time from first rating to default, time from last cash pay date to default,
time from default to emergence, and time from origination to maturity.
Multivariate regression analysis of defaulted debt returnsIn this section, we discuss the construction and results of multiple regres-
sion models for RDD. In order to cope with the highly non-normal nature
of the RDD distribution, we turn to the various techniques that have been
employed in the finance and economics literature to classify data in mod-
els with constrained dependent variables, either qualitative or bounded
in some region. However, much of the credit risk related literature has
focused on qualitative dependent variables, which the case of probabil-
ity-of -default (PD) estimation naturally falls into. Maddala (1991, 1983)
introduces, discusses, and formally compares the different generalized
linear models (GLMs). Here we consider the case most relevant for RDD
estimation, and the least pursued in the GLM literature. In this context,
since we are dealing with a random variable in a bounded region, this is
most conveniently modeled through employing a beta distribution. Con-
sequently, we follow Mallick and Gelfand (1994), in which the GLM link
function21 is taken as a mixture of cumulative beta distributions, which we
term the beta-link GLM (BLGLM) [see Jacobs and Karagozoglu (2011) for
an application of the GLM model to estimating the ultimate LGD].
The coefficient estimates and diagnostic statistics for our “leading” three
models are shown in Table 4. These are determined through a combina-
tion of automated statistical procedures22 and expert judgment, where
we try to balance sometimes competing considerations of statistical
quality of the estimates with the sensibility of the models. Essentially, the
three models shown in Table 4 had the best fit to the sample data, while
spanning what we thought was the best set of risk factors, based upon
prior expectations as well as the univariate analysis. Note that there is
much overlap between the models, as Model 2 differs from Model 1 by
two variables (it has MV/BV instead of TL/TA, and has RSIZ), and Model 3
from Model 2 by two variables (FAR in lieu of TSI and LGD).
Across the three candidate models, we observe that all coefficients esti-
mates attain a high degree of statistical significance, in almost all cases
at better than the 5 percent level,23 and in many cases at much better
than the 1 percent level. The number of observations for which we had
all of these explanatory variables is the same for Models 1 and 2 (968),
but there is a sizable drop-off for Model 3 to only 792 observations. In all
cases, the likelihood functions converged to a stable global maximum.24
Model 3 achieves the best in-sample fit by McFadden pseudo r-squared
of 41.7 percent, followed by Model 2 (38.8 percent), and Model 1 (32.5
percent). In terms of maximized log-likelihood, Model 3 is far better than
the others (-504.0), and Model 1 is only slightly better than Model 2
(-592.3 versus -594.7) in spite of having one less explanatory variable.
However, as these models are not nested this may not be so meaningful
a comparison. Overall, we deem these to signify good fit, given the non-
linearity of the problem, the relatively high dimension, as well as the high
level of noise in the RDD variable.
We now turn to the signs and individual economic significance of the vari-
ables, note that we report partial effects (PEs), which are akin to straight
coefficient estimates in an ordinary least squares regression. Roughly
speaking, this represents a change in the dependent variable for a unit
change in a covariate, holding other variables fixed at their average sam-
ple values.25
First, we consider the systematic risk variables. In the case of the Moody’s
speculative default rate by industry, appearing in all models, we see PEs
ranging in 2.05-2.25. This implies that a percentage point elevation in
aggregate default rates adds about 2 percent in return on defaulted debt
on average, all else equal, which can be considered highly significant in
an economic sense. For example, the near quadrupling in default rates
between 1996 and 2001 would imply an increase in expected RDD of
about 12 percent. On the other hand, the PEs on the one-month treasury
yield are in the range of -0.49 to -0.37, so that debt defaulting when
short-term rates are about 2 percent higher will experience close to 1
percent deterioration in performance, ceteris paribus. Second, across all
three regression models, RDD has a significant (at the 5 percent level)
and positive loading on the FF-ERM, with PEs ranging from 1.38 to 1.55,
implying that a 5 percent increase in the aggregate equity market return
augments defaulted debt returns by about 6 percent.
Next, we consider the contractual variables. The dummy variable for se-
cured collateral has PEs ranging in 0.23-0.27 across models, suggesting
that the presence of any kind of security can be expected to augment
expected RDD by about 25 percent, which is an economically significant
result. The TSI, appearing only in Models 1 and 2, has a PE ranging in
0.43-0.45, suggesting that going up a single decile in this measure can
increase RDD by anywhere between 4 percent to 5 percent.
The Capco Institute Journal of Financial TransformationEmpirical Analysis, Trading Strategies, and Risk Models for Defaulted Debt Securities
In the terminology of GLMs, the link function connects the expectation of some function 21
of the data (usually the random variable weighted by density, in the case of the expected
value) to a linear function of explanatory variables.
To this end, we employ an alternating direction stepwise model selection algorithm in the 22
mass( ) library of R statistical software. There were five candidate leading models that were
tied as best, we eliminated two of them that we judged to have economically unreasonable
features.
Moody’s investment grade rating in Model 3 is on the borderline, having a p-value of 0.06, 23
just shy of significance at the 5 percent level.
The estimation was performed in S+ 8.0 using built-in optimization routines.24
See Maddala (1981) for a discussion of this concept in the context of probit and logit 25
regressions.
70
Turning to the credit quality/market variables, for LGD at default, only in
Models 1 and 2, PEs are about 0.28-0.33, implying that a 10 percent low-
er expected recovery rate by the market at default can lead to about a 3
percent higher expected RDD. The dummy variable for a Moody’s invest-
ment grade rating at origination, appearing in all models, has PEs ranging
from 0.16 in Model 3 to 0.24 in Model 2. This tells us that “fallen angels”
are expected to have about 15-25 percent better return on their defaulted
debt. On the other hand, the single relative stock price performance vari-
able CAR, in all three models, has PEs ranging in 0.37-0.40. This says
that, for example, a firm with 10 percent better price performance relative
to the market in the 90 days prior to default will experience about 4 per-
cent better return on its defaulted debt.
In the case of the financial ratios, TL/TA appears only in Model 1, hav-
ing a PE of 0.27. This means that the debt of a defaulted firm having 10
percent higher leverage at default will have about 3 percent greater return
on its debt. MV/BV appears in Models 2 and 3, with respective PEs of
0.19 and 0.14, so that a 10 percent higher market valuation translate on
average into nearly a 2 percent better return on defaulted debt. Finally in
this group, the cashflow measure FAR only appears in Model 3, with a PE
of -0.24. This implies that if a defaulted firm has 10 percent greater cash
generating ability by this measure, then holding other factors constant its
RDD should return about 2.5 percent less.
Finally, the size of the firm relative to the market appears in only Models 2
and 3, with PEs of about 0.06 to 0.04. As this is in logarithmic terms, we
interpret this as if a defaulted firm doubles in relative market capitaliza-
tion, we should expect its RDD to be augmented by around 5 percent, all
other factors being held constant.
In order to settle upon a “favored” or “leading” model, we perform an
out-of-sample and out-of-time analysis. We reestimate the models for
different subsamples of the available data, starting from the middle of
the dataset in year 1996. We then evaluate how the model predicts the
realized RDD a year ahead. We employ a resampling procedure (a “non-
parametric bootstrap”), sampling randomly with replacement from the
development dataset (i.e., the period 1987-1996), and in each iteration
reestimating the model. Then from the year ahead, we resample with
replacement (i.e., the 1997 cohort), and evaluate the goodness-of-fit for
the model. This is performed 1000 times, then a year is added, and this is
repeated until the sample is exhausted. At the end of the procedure, we
collect the r-squareds, and study their distribution for each of the three
models. The results of this show that the mean out-of-sample r-squared
in Model 1 is highest, at 21.2 percent, followed by Model 3 (17.8 percent),
and Model 2 (12.1 percent). On the basis of the numerical standard er-
rors (on the order of 1-2 percent), we deem these to be significantly dis-
tinct. Given the best performance on this basis, in conjunction with other
considerations, we decide that Model 1 is the best. The other reasons
for choosing Model 1 are its parsimony relative to Model 2, and that it
contains a credit market variable (LGD), the latter we believe makes for
a more compelling story. Note that this procedure is robust to structural
breaks, as the model is redeveloped over an economic cycle, in each
iteration the same variables are chosen, and the models display the same
relative performance over time.
Finally, in Table 5, we evaluate the economic significance of these results.
We formulate a trading strategy as follows. At the time of default, if fore-
casted returns according to the model over the expected time-to-resolution
exceed cumulative excess of returns equity in the three months prior to de-
fault, then we form a long position in the debt, else we form a short position
on the defaulted instrument. Abnormal excess returns are then measured
relative to a market model (3-factor Fama-French) from the time of default
to resolution. The results show excess abnormal returns, in this defaulted
debt trading experiment, of around 5-6 percent (2-3 percent) assuming
zero (1bp per month) round-trip transaction costs. These are statistically
significant, and understandably lower and having higher p-values when we
factor in transaction costs. Also, results are not highly differentiated across
models, with Model 3 performing about 1 percent better assuming no
transaction costs, and Model 1 having a similar margin of outperformance
relative to the other models assuming transaction costs. Given that the lat-
ter is arguably a more realistic scenario, we still favor Model 1 because it
generates superior excess returns in this trading strategy.
Model 1 Model 2 Model 3
Mean P-value Mean P-value Mean P-value
Zero transaction costs 0.0051 3.65E-03 0.0049 2.79E-04 0.0062 1.98E-03
1 bp per month round trip transaction costs 0.0032 7.76E-02 0.0019 7.90E-03 0.0025 6.45E-02
Table 5 – Excess abnormal trading returns1 of beta-link generalized linear model for annualized returns on defaulted debt2 (Moody’s Ultimate Recovery Database 1987-2008)1 – We formulate a trading strategy as follows. At the time of default, if forecasted returns according to the model over the expected time-to-resolution in excess of returns of returns on equity in
the three months prior to default are positive (negative), then we form a long (short) position in the debt. Abnormal excess returns are then measured relative to a market model (3-factor Fama-
French) from the time of default to resolution.
2 – Annualized “return on defaulted debt” (RDD) from just after the time of default (first trading date of debt) until the time of ultimate resolution.
71
ConclusionIn this paper, we have empirically studied the market performance of a
long history of defaulted debt. We examined the distributional properties
of the return on defaulted debt (RDD) measure across different segmen-
tations in the dataset (i.e., default type, facility type, time period, senior-
ity, industry), and developed multiple regression models for RDD in the
generalized linear model (GLM) class.
We found that defaulted debt returns vary significantly according to cer-
tain different factors. There is some evidence that RDD is elevated for
debt having better collateral quality rank or better protected tranches
within the capital structure; and for obligors rated higher at origination,
larger in market capitalization relative to the market, more financially le-
vered, or having higher cumulative abnormal returns on equity (CARs)
at default. However, RDD is increasing in market implied loss severity
at default (loss given default – LGD). We also find evidence that returns
vary countercyclically, as they are positively correlated with industry de-
fault rates. Furthermore, they are inversely related to short-term interest
rates, and positively related to returns on the equity market. We identify
a leading econometric model of RDD that performs well out-of-time and
out-of sample. Finally, we document the economic significance of these
results through excess abnormal returns, in a debt-equity arbitrage trad-
ing experiment, of around 5-6 percent (2-3 percent) assuming zero (1bp
per month) round-trip transaction costs.
ReferencesAcharya, V., S. T. Bharath, and A. Srinivasan, 2007, “Does industry-wide distress affect •
defaulted firms? – evidence from creditor recoveries,” Journal of Political Economy, 85, 787-821
Altman, E. I., 1989, “Measuring corporate bond mortality and performance,” Journal of Finance, •
44, 909-922
Altman, E. I. and S. Jha, 2003, “Market size and investment performance of defaulted bonds •
and loans: 1987-2002,” Working paper, New York University Stern School of Business
Altman, E. I. and B. Karlin, 2010, “Special report on defaults and returns in the high-yield bond •
and distressed debt market: The year 2009 in review and outlook,” NYU Solomon Center
Report, February
Altman, E. I., A. Resti, and A. Sironi, 2003, “Default recovery rates in credit risk modelling: a •
review of the literature and empirical evidence,” Working paper, New York University Solomon
Center, February
Barco, M., 2007, “Going downturn,” Risk, September, 38-44•
Basel Committee on Banking Supervision, 2003, “International convergence on capital •
measurement and capital standards,” Bank for International Settlements (BIS), June
Basel Committee on Banking Supervision, 2005, “Guidance on paragraph 468 of the framework •
document,” BIS, July
Basel Committee on Banking Supervision, 2006, “International convergence on capital •
measurement and capital standards: a revised framework,” BIS, June
Carey, M. and M. Gordy, 2007, “The bank as grim reaper: debt decomposition and recoveries •
on defaulted debt,” Working paper, Federal Reserve Board
Davydenko, S. and I. Strebuleav, 2002, “Strategic behavior, capital structure and credit spreads: •
an empirical investigation,” Working paper, London Business School
Dullman, K. and M. Trapp, 2004, “Systematic risk in LGD – an empirical analysis,” Working •
paper, University of Mannheim
Emery, K., R. Cantor, D. Keisman, D. and S. Ou, 2007, “Moody’s Ultimate Recovery Database: •
special comment,” Moody’s Investor Service, April
Fama, E. F. and K. R. French, 1992, “The cross-section of expected stock returns,” Journal of •
Finance, 47, 42-480
Frye, J., 2000a, “Collateral damage,” Risk, April, 91-94•
Frye, J., 2000b, “Collateral damage detected,” Federal Reserve Bank of Chicago, Emerging •
Issues Series, October, 1-14
Frye, J., 2000c, “Depressing recoveries,” Risk, 13:11, 108-111•
Frye, J., 2003, “A false sense of security”, Risk, August, 63-67•
Giese, G, 2005, “The impact of PD/LGD correlations on credit risk capital,” Risk, April, 79-84•
Gilson, S., 1955, “Investing in distressed situations: a market survey,” Financial Analysts •
Journal, November-December, 8-27
Gordy, M., 2003, “A risk-factor model foundation for ratings-based bank capital rules,” Journal •
of Financial Intermediation, 12, 199-232
Guha, R., 2003, “Recovery of face value at default: empirical evidence and implications for •
credit risk pricing,” Working paper, London Business School
Hillebrand, M., 2006, “Modelling and estimating dependent loss given default,” Risk, •
September, 120-125
Hotchkiss, E. S. and R. M. Mooradian, 1997, “Vulture investors and the market for control of •
distressed firms,” Journal of Financial Economics, 43, 401-432
Hu, Y. T. and W. Perraudin, 2002, “The dependence of recovery rates and defaults,” working •
paper, Bank of England
Hotchkiss, E. and R. Mooradian, 1997, “Vulture investors and the market for control of •
distressed firms,” Journal of Financial Economics, 43, 401-432
Jacobs, Jr., M., 2011, “Empirical implementation of a 2-factor structural model for loss-given-•
default,” Journal of Financial Transformation, 31, 31-43
Jacobs, Jr., M. and A.K. Karagozoglu, 2011, “Modeling ultimate loss-given-default on corporate •
debt,” Journal of Fixed Income (Forthcoming)
Keenan S. C., D. T. Hamilton, and A. Berthault, 2000, “Historical default rates of corporate bond •
issuers: 1920-1999,” Moody’s Investors Services, January
Keisman, D. and K. van de Castle, 2000, “Suddenly structure mattered: insights into recoveries •
of defaulted debt, Corporate ratings,” commentary, Standard and Poors
Maclachlan, I., 2004, “Choosing the discount factor for estimating economic LGD,” Working •
paper, Australia and New Zealand Banking Group, Ltd
Maddala, G. S., 1983, Limited dependent and qualitative variables in finance, 3rd ed., •
Cambridge University Press
Maddala, G. S., 1991, “The perspective on the use of limited-dependent and qualitative •
variables models in accounting research,” The Accounting Review, 66, 788-807
Mallick, B. K. and A. E. Gelfand, 1994, “Generalized linear models with unknown link functions,” •
Biometrika, 81, 237-245
Merton, R. C., 1971, “Optimum consumption and portfolio rules in a continuous-time model,” •
Journal of Economic Theory, 3, 373-413
Merton, R. C., 1974, “On the pricing of corporate debt: the risk structure of interest rates,” •
Journal of Finance, 29, 449–470
Pykhtin, M., 2003, “Unexpected recovery risk”, Risk, August, 74-79•
Rosch, D. and H. Scheule, 2005, “A multifactor approach for systematic default and recovery •
risk”, Risk, September, 62-75
Schuermann, T., 2003, “What do we know about loss given default?” in Shimko, D., ed. Credit •
risk models and management, 2nd Edition, Risk Books
Vasicek, O., 1987, “Probability of loss on a loan portfolio,” Working paper, KMV Corporation•
Vasicek, O., 2002, “Loan portfolio value,” Risk, December, 160-162•
The Capco Institute Journal of Financial TransformationEmpirical Analysis, Trading Strategies, and Risk Models for Defaulted Debt Securities
72
Bankruptcy Out-of-court Total
Cnt Average
Std Err of
the mean Minimum Maximum Cnt Average
Std Err of
the mean Minimum Maximum Cnt Average
Std Err of
the mean Minimum Maximum
Sub
-po
pul
atio
n o
f M
oo
dy’
s re
cove
ries
dat
abas
e ha
ving
tra
din
g p
rice
of
deb
t at
def
ault
Bonds
and term
loans
RDD
1072
28.32% 3.47% -100.00% 893.76%
59
45.11% 19.57% -91.87% 846.73%
1131
29.19% 3.44% -100.00% 893.76%
Time-to-
resolution21.7263 0.0433 0.0027 9.0548 6.65% 3.33% 0.27% 144.38% 1.6398 0.0425 0.0000 9.0548
Principal at
default3207,581 9,043 163 4,600,000 416,751 65,675 6,330 2,250,000 218,493 9,323 0 4,600,000
Bonds RDD
837
25.44% 3.75% -100.00% 893.76%
47
44.22% 21.90% -91.87% 846.73%
884
26.44% 3.74% -100.00% 893.76%
Time-to-
resolution21.4089 0.0436 0.0548 9.0548 0.2044 0.0786 0.0027 1.4438 1.3194 0.0427 0.0027 9.0548
Principal at
default3205,028 10,590 0 4,000,000 432,061 72,727 6,330 2,250,000 207,647 10,325 0 4,000,000
Revolvers RDD
250
26.93% 7.74% -100.00% 893.76%
17
10.32% 4.61% -0.04% 61.18%
267
25.88% 7.26% -100.00% 893.76%
Time-to-
resolution21.4089 0.0798 0.0548 9.0548 0.0027 0.0000 0.0027 0.0027 1.3194 0.0776 0.0027 9.0548
Principal at
default3205,028 19,378 0 4,000,000 246,163 78,208 32,000 1,250,000 207,647 18,786 0 4,000,000
Loans RDD
485
32.57% 5.71% -100.00% 893.76%
29
26.161% 18.872% -91.87% 532.76%
514
32.21% 5.49% -100.00% 893.76%
Time-to-
resolution21.4089 0.0548 0.0027 9.0548 18.12% 9.96% 0.0027 2.8959 1.2458 0.0743 0.0027 9.0548
Principal at
default3193,647 11,336 0 4,000,000 291,939 78,628 24,853 1,750,000 199,192 16,088 0 4,000,000
Total RDD
1322
28.05% 3.17% -100.00% 893.76%
76
37.33% 15.29% -91.87% 846.73%
1398
28.56% 3.11% -100.00% 893.76%
Time-to-
resolution21.6663 0.0384 0.0027 9.0548 0.0522 0.0260 0.0000 1.4438 1.5786 0.0376 0.0000 9.0548
Principal at
default3207,099 8,194 0 4,600,000 378,593 54,302 0 2,250,000 216,422 8,351 0 4,600,000
Ent
ire
po
pul
atio
n o
f M
oo
dy’
s re
cove
ries
dat
abas
e
Bonds
and term
loans
Time-to-
resolution2
2798
1.6982 0.0253 0.0027 9.3151
433
0.2084 0.0261 0.0027 3.8767
3231
1.4986 0.0239 0.0027 9.3151
Principal at
default3149,623 4,585 0 4,600,000 204,750 16,469 0 3,000,000 157,011 4,553 0 4,600,000
Bonds Discounted
LGD3
2162
48.57% 0.83% -69.78% 100.00%
345
14.50% 1.37% -27.66% 100.00%
2507
43.83% 0.78% -69.78% 100.00%
Time-to-
resolution21.7786 0.0290 0.0027 9.3151 0.2084 0.0292 0.0027 3.8767 1.5620 0.0275 0.0027 9.3151
Principal at
default3157,488 5,608 0 4,600,000 204,750 18,450 0 3,000,000 166,781 5,551 0 4,600,000
Revolvers Discounted
LGD3
702
39.47% 1.47% -69.78% 100.00%
117
18.00% 2.76% -3.58% 100.00%
819
36.40% 1.35% -69.78% 100.00%
Time-to-
resolution41.3944 0.1062 0.0027 9.0548 0.1490 0.0476 0.0027 2.8959 1.2165 0.0407 0.0027 9.0548
Principal at
default5131,843 21,396 0 4,000,000 124,199 17,836 347 1,250,000 130,751 #### 0 4,000,000
Loans Discounted
LGD3
1338
40.03% 1.08% -69.78% 100.00%
205
17.20% 2.03% -27.66% 100.00%
1543
37.00% 0.99% -69.78% 100.00%
Time-to-
resolution41.4089 0.0330 0.0027 9.0548 0.1812 0.0375 0.0027 2.8959 1.2458 0.0309 0.0027 9.0548
Principal at
default5127,586 5,521 0 4,000,000 124,671 14,739 347 1,750,000 127,199 5,171 0 4,000,000
Total Discounted
LGD3
3500
45.31% 0.66% -69.78% 100.00%
550
15.25% 1.13% -27.66% 100.00%
4050
41.23% 0.61% -69.78% 100.00%
Time-to-
resolution41.6373 0.0221 0.0027 9.3151 0.1958 0.0026 0.0027 3.8767 1.4415 0.0208 0.0027 9.3151
Principal at
default5146,057 4,064 0 4,600,000 187,615 13,576 0 3,000,000 151,701 3,972 0 4,600,000
1 – “Return on defaulted debt”: annualized simple rate of return on defaulted debt from just after the time of default (first trading date of debt) until the time of ultimate resolution.
2 – Total instrument outstanding at default.
3 – The time in years from the instrument default date to the time of ultimate recovery.
Table A1 – Characteristics of return on defaulted debt (RDD)1 observations by default and instrument type (Moody’s Ultimate Recovery Database 1987-2010)
73
The Capco Institute Journal of Financial TransformationEmpirical Analysis, Trading Strategies, and Risk Models for Defaulted Debt Securities
Collateral Type
Revolving credit/term
loan
Senior secured
bonds
Subordinated
bonds
Senior unsecured
bonds
Senior subordinated
bonds
Total
instrument
Cnt
Avg
(%)
Std Err
(%) Cnt
Avg
(%) Std Err Cnt
Avg
(%)
Std Err
(%) Cnt
Avg
(%) Std Err Cnt
Avg
(%) Std Err Cnt
Avg
(%) Std Err
Min
or
colla
tera
l cat
ego
ry
Guarantees 2 -96.0 4.0 0 N/A N/A 0 N/A N/A 0 N/A N/A 0 N/A N/A 2 -96.0 4.0
Oil and gas properties 2 77.5 68.3 0 N/A N/A 0 N/A N/A 0 N/A N/A 0 N/A N/A 2 77.5 68.3
Inventory and
accounts receivable
28 20.2 23.4 0 N/A N/A 0 N/A N/A 0 N/A N/A 0 N/A N/A 28 20.2 23.4
Accounts receivable 5 24.5 28.5 2 23.9 40.6 0 N/A N/A 0 N/A N/A 0 N/A N/A 7 24.4 21.6
Cash 2 114.8 17.0 0 N/A N/A 0 N/A N/A 0 N/A N/A 0 N/A N/A 2 114.8 17.0
Inventory 1 -100.0 N/A 1 29.3 N/A 0 N/A N/A 0 N/A N/A 0 N/A N/A 2 -35.3 64.7
Most assets 6 25.6 30.4 1 161.5 N/A 0 N/A N/A 0 N/A N/A 1 72.1 N/A 8 48.4 28.1
Equipment 1 -100.0 N/A 17 41.9 8.9 0 N/A N/A 0 N/A N/A 0 N/A N/A 18 34.0 11.5
All assets 363 32.4 6.9 36 33.4 23.8 1 86.5 N/A 0 N/A N/A 0 N/A N/A 400 32.6 6.6
Real estate 4 132.0 73.6 2 63.8 110.9 0 N/A N/A 1 57.4 N/A 0 N/A N/A 7 101.8 48.3
All non-current assets 2 -41.7 58.3 3 -60.7 35.1 0 N/A N/A 0 N/A N/A 0 N/A N/A 5 -53.1 27.0
Capital stock 36 40.0 15.4 38 65.2 19.1 0 N/A N/A 0 N/A N/A 0 N/A N/A 74 52.9 12.3
PP&E 8 106.0 70.7 17 6.9 17.4 0 N/A N/A 0 N/A N/A 0 N/A N/A 25 38.6 26.3
Second lien 21 23.2 26.8 17 24.1 18.4 1 119.6 N/A 1 -46.0 N/A 0 N/A N/A 40 24.3 16.2
Other 0 N/A N/A 1 -24.7 N/A 0 N/A N/A 0 N/A N/A 0 N/A N/A 1 -24.7 N/A
Unsecured 32 19.8 7.9 3 -27.7 36.6 452 23.7 4.9 158 31.0 10.2 117 15.7 11.1 762 23.6 4.0
Third lien 1 106.1 N/A 1 4.9 N/A 7 3.5 22.3 1 439.4 N/A 2 -21.8 2.5 12 44.3 39.2
Intellectual property 0 N/A N/A 2 28.6 43.9 0 N/A N/A 0 N/A N/A 0 N/A N/A 2 28.6 43.9
Intercompany debt 0 N/A N/A 1 143.1 N/A 0 N/A N/A 0 N/A N/A 0 N/A N/A 1 143.1 N/A
Maj
or
colla
tera
l cat
ego
ry
Cash, accounts
receivables and
guarantees
39 22.6 18.2 2 23.9 40.6 0 N/A N/A 0 N/A N/A 0 N/A N/A 41 22.6 17.4
Inventory, most assets
and equipment
8 -5.8 30.3 19 47.5 10.2 0 N/A N/A 0 N/A N/A 1 72.1 N/A 28 33.2 11.7
All assets and real
estate
367 33.5 6.9 38 35.0 22.9 1 86.5 N/A 1 57.4 N/A 0 N/A N/A 407 33.8 6.6
Non-current assets
and capital stock
38 35.7 15.0 41 56.0 18.6 0 N/A N/A 0 N/A N/A 0 N/A N/A 79 46.2 12.0
PPE and second lien 29 46.1 27.6 35 14.3 12.3 1 119.6 N/A 1 -46.0 N/A 0 N/A N/A 66 29.0 13.9
Unsecured & Other
Illiquid Collateral
33 22.4 8.1 7 17.4 28.5 459 23.4 4.8 159 33.6 10.4 119 15.1 10.9 777 24.1 3.9
Total unsecured 32 19.8 7.9 3 27.7 36.6 452 23.7 4.9 158 31.0 10.2 117 15.7 11.1 762 23.6 4.0
Total secured 482 33.0 5.8 139 38.0 9.1 9 25.6 22.6 3 150.3 147.6 3 9.5 31.4 636 34.5 4.9
Total collateral 514 32.2 5.5 161 36.6 8.4 461 23.7 4.8 161 33.2 10.3 120 15.6 10.8 1398 28.6 3.1
Table A2 – Return on defaulted debt (RDD1) by seniority ranks and collateral types (Moody’s Ultimate Recovery Database 1987-2010)1 – Annualized “return on defaulted debt” from the time of default until the time of ultimate resolution.
74
Category Variable Count Minimum Median Mean Maximum Std err of
the mean
Correlation
with RDD
P-value of
correlation
Fina
ncia
l sta
tem
ent
and
mar
ket
valu
atio
n
Book value total liabilities/book value total assets 1106 38.00% 115.00% 137.42% 392.00% 2.33% 17.20% 4.32E-04
Market-to-book (market value assets/book value
assets)
1106 44.00% 123.00% 152.61% 673.00% 2.71% 18.50% 6.34E-05
Intangibles ratio (book value intangibles/book value
assets)
773 0.00% 18.34% 21.02% 87.85% 0.75% 11.91% 4.20E-04
Free asset ratio 941 -95.51% 9.24% 5.89% 95.86% 1.18% -8.97% 2.34E-03
Free cash flow/book value of total assets 1006 -107.64% -1.41% -10.77% 34.61% 0.70% -2.42% 6.11E-04
Cash flow from operations/book value of total assets 1014 (669.12) (0.48) 57.09 7,778.00 30.65 -3.32% 8.33E-04
Retained earnings/book value of total assets 1031 -757.97% -25.80% -61.21% 56.32% 3.01% -5.91% 1.47E-03
Return on assets 1031 -159.12% -8.52% -22.18% 36.35% 0.92% -6.50% 1.62E-03
Return on equity 1031 -2950.79% 3.10% 23.11% 6492.67% 17.19% -4.31% 1.07E-03
Eq
uity
pri
ce
per
form
ance
One-year expected return on equity 1106 -132.00% -80.00% -72.40% 161.00% 1.26% -6.42% 1.07E-03
One-month equity price volatility 1106 13.00% 209.00% 259.49% 6116.00% 11.18% 2.48% 1.14E-04
Relative size (market cap of firm to the market) 1106 -17.3400 -12.7200% -13.0487 -6.9300 0.0599 8.60% 1.31E-03
Relative stock price (percentile ranking to market) 1106 0.47% 11.00% 13.76% 81.00% 0.42% -4.36% 1.05E-03
Stock price trading range (ratio of current to 3 Yr
high/low)
1106 0.00% 0.71% 2.95% 88.00% 0.22% -2.92% 7.02E-04
Cumulative abnormal returns (90 days to default) 1171 -127.70% 0.00% -4.87% 147.14% 0.84% 10.30% 2.42E-03
Cap
ital
stru
ctur
e
Number of instruments 4050 0.0000 6.0000 9.9511 80.0000 0.1938 -4.04% 5.07E-04
Number of creditor classes 4050 0.0000 2.0000 2.4669 7.0000 0.0188 -2.98% 3.74E-04
Percent secured debt 4050 0.00% 47.79% 47.13% 100.00% 0.56% 8.76% 1.10E-03
Percent bank debt 4050 0.00% 44.53% 45.23% 100.00% 0.54% 9.44% 1.19E-03
Percent subordinated debt 4050 0.00% 41.67% 43.26% 100.00% 0.53% 8.68% 1.09E-03
Cre
dit
qua
lity/
cred
it
mar
ket
Altman Z-score 793 -8.5422 0.3625 -0.3258 4.6276 0.0804 -8.75% 2.49E-03
LGD at default 1433 -8.50% 59.00% 55.05% 99.87% 0.83% -11.28% 2.38E-04
Moody’s original credit rating investment grade
dummy
3297 0.0000 0.0000 0.2014 1.0000 0.0070 12.40% 2.37E-04
Moody’s original credit rating (minor code) 3342 3.0000 14.0000 12.4054 20.0000 0.0588 3.63% 5.01E-04
Inst
rum
ent/
cont
ract
ual
Seniority rank 4050 1.0000 1.5000 1.7262 7.0000 0.0142 -9.60% 2.28E-04
Collateral rank 4050 1.0000 6.0000 4.5879 6.0000 0.0254 -10.00% 5.29E-04
Percent debt below 4050 0.00% 9.92% 25.89% 100.00% 0.48% 9.36% 1.18E-03
Percent debt above 4050 0.00% 0.00% 21.41% 100.00% 0.45% -5.16% 6.48E-04
Tranche safety index 4050 0.00% 50.00% 52.24% 100.00% 0.40% 9.70% 1.04E-03
Mac
ro/c
yclic
al
Moody’s all-corporate quarterly default rate 1322 0.00% 7.05% 7.14% 13.26% 0.09% 6.68% 1.47E-03
Moody’s speculative quarterly default rate 1322 1.31% 7.05% 7.16% 13.26% 0.09% 6.40% 1.41E-03
Fama-French excess return on market factor 4050 -1076.00% 77.00% 31.06% 1030.00% 7.20% 7.22% 3.02E-04
Fama-French relative return on small stocks factor 4050 -2218.00% 31.00% 20.15% 843.00% 6.00% 2.81% 3.52E-04
Fama-French excess return on value stock factor 4050 -912.00% 54.00% 79.35% 1380.00% 5.75% -4.27% 5.35E-04
Short-term interest rates (1-month treasury yields) 1322 6.00% 32.00% 31.75% 79.00% 0.46% -10.22% 2.26E-03
Long-term interest rates (10-month treasury yields) 1106 332.00% 535.00% 538.42% 904.00% 3.61% -7.00% 1.69E-03
Stock-market volatility (2-year IDX) 1106 4.00% 9.00% 10.03% 19.00% 0.12% 5.70% 1.37E-03
Dur
atio
ns/
vint
age
Time from origination to default 3521 0.2500 2.9096 4.0286 29.9534 0.0631 0.57% 7.68E-05
Time from last cash-pay date to default 4050 0.0000 0.2384 0.3840 4.3808 0.0075 4.49% 5.63E-04
Time from default to resolution 4050 0.0027 1.1534 1.4415 9.3151 0.0208 -13.41% 1.70E-03
Time from origination to maturity date 3521 0.1000 7.5890 8.9335 50.0329 0.1111 -0.85% 1.14E-04
Table A3 – Summary statistics on selected variables and correlations with RDD1 (Moody’s Ultimate Recovery Database 1987-2010)1 – Annualized “return on defaulted debt” (RDD) from just after the time of default (first trading date of debt) until the time of ultimate resolution.
75
PART 1
Systemic Risk, an Empirical Approach
AbstractWe have developed a quantitative analysis to verify the ex-
tent to which the sources of systemic risk identified in the
academic and regulatory literature actually contribute to it.
This analysis shows that all institutions contribute to system-
ic risk albeit to a different degree depending on various risk
factors such as size, interconnection, unsubstitutability, bal-
ance sheet, and risk quality. From the analysis we conclude
that using a single variable or a limited series of variables as
a proxy for systemic risk generates considerable errors when
identifying and measuring the systemic risk of each institu-
tion. When designing systemic risk mitigation measures, all
contributing factors should be taken into account. Likewise,
classifying institutions as systemic/non-systemic would
mean giving similar treatment to institutions that could bear
very different degrees of systemic risk, while treating differ-
ently institutions that may have very similar systemic risk in-
side. Consequently, we advocate that some continuous ap-
proach to systemic risk in which all institutions are deemed
systemic, but to varying degrees, would be preferable. We
acknowledge that this analysis may prove somehow lim-
ited given that it is not founded on a predefined conceptual
approach, does not fully consider other very relevant quali-
tative factors,2 and accounts only for some of the relevant
sources of systemic risk in the banking system.3 These limits
are currently set due to data availability and the current state
of empirical research, but we believe that these should not
hinder our work in identifying the true sources of systemic
risk and our aim to help avoid any partial and thus limited
prudential policy approach.
Gonzalo de Cadenas-Santiago — Economic Research & Public Policy Department, Banco Santander
Lara de Mesa — Economic Research & Public Policy Department, Banco Santander
Alicia Sanchís — Economic Research & Public Policy Department, Banco Santander1
We would like to thank our colleagues at the Economic Research 1
Department, the Financial Supervision and Investors Relations Division
and our colleagues at the general intervention division for their helpful
comments. Special thanks to Alejandra Kindelán and Antonio Cortina for
their invaluable comments and support. The authors would also like to
thank the Economic Capital Models and Risk Metrics Department for their
invaluable help providing the validation analysis of the model. We thank
Jordi Gual and Jose Peydrós for their helpful comments in the discussion
of this paper at the XXVII Moneda y Crédito Conference held in Madrid
on November 3rd, 2010. We would also like to thank Mauricio Pinto’s
collaboration during his stay at the Santander premises as well as Nyla
Karsan for her supervision and help.
Such as the quality of supervision, the regulatory framework, and the crisis 2
management framework of each jurisdiction.
That is the case of interconnectedness, where only theoretical work has 3
been put in place and scarce variables to approximate it are available.
76
Motivation and main objectivesAcademic and regulatory literature defines systemic risk as the risk of
a major dysfunction or instability of the financial system caused by one
or more parties, resulting in externalities whose potential effects pose a
severe threat to both the financial system and the real sector [Acharya
(2009)4, FSB/IMF/BIS (2009)]. International regulatory bodies have identi-
fied the need to control such risk as a priority in order to guarantee the
financial stability of the system and avoid distortions that could prevent
the efficient assignment of savings, investment, consumption, and the
correct payment process within the economy [Vesala (2009)]. All those
functions are nowadays carried out by the financial system.
The negative externalities associated with the potential risk stemming
from financial activity justified the introduction of a specific prudential su-
pervisory framework. However, this framework is insufficient to address
the effects of systemic dimensions such as capital quality and quantity,
liquidity management, risk inherent to trading activities, etc. Basel III is a
new framework designed to fill in the gaps and mitigate the weaknesses
identified to date and the Financial Stability Board (FSB) and the Ba-
sel Committee on Banking Supervision (BCBS) are considering ways of
modulating regulatory pressure on systemic risk at both the system and
institutional level. This would result in heavier penalties for those deemed
to be more systemic.
The report by the FSB/IMF/BIS5 shows the enormous variety of ap-
proaches to systemic risk that are currently used by the different do-
mestic supervisors, which ranges from sophisticated models to simple
synthetic indicators. The report concludes that the nature of systemic
risk is both time varying and multi-factor led. This renders its analysis
the more challenging. In this context, considering qualitative aspects of
banks that could mitigate factors of risk are very appealing to both regu-
lators and academics. The novelty and complexity of the topic suggest
adding some type of ex-ante expert’s view in order to better supplement
the identification process of our analysis.
The disparity of approaches, together with the complexity of the prob-
lem, and the data availability shortages, may encourage regulators to
take excessively simplistic solutions to tackle the analysis, such as limit-
ing the number of variables to define systemic risk or to define closed
lists of institutions characterized as systemic. Giving too much weight
to a limited number of variables (like using solely “size” as proxy for risk)
simply because they are easier to gather and measure and ignoring other
equally important sources of risk could not only result in a very partial
view of the problem, but also have very severe consequences for the
financial system itself. Doing this could disregard the benefits of diversi-
fication and impede the financial sector from reaping the full benefits of
economies of scale and scope. The unintended consequence could be a
loss of strength and stability of the financial system and the atomization
and increased cost of the banking business with questionable benefits
for the stability of the financial system.
This article aims to contribute to the debate by building a “proxy” for
systemic risk for each institution and for the system itself and discussing
its potential uses and shortcomings. It will be based on criteria gathered
from academic and regulatory literature (FSB, BIS, IMF and ECB) that
relates to the following factors: size, interconnectedness, unsubstitut-
ability, and other equally important sources of systemic risk such as the
strength of institutions (measured in terms of balance sheet quality and
composition), the liquidity and financing position, and the quality of risk
management. Our goal is to extract the information about systemic risk
contained in variables identified by regulators and academics as those
most contributing to systemic risk. We will address this by creating a
synthesis indicator to use as a proxy for the latent systemic risk nature of
each financial institution within the sample. This will enable us to assess
the contribution of each risk variable to global systemic risk, evaluate the
contribution of each institution to risk, assess the degree of systemicity
of each institution through time, and to give a wide measure of systemic
risk for the entire system through the cycle.
The aim of this analysis is two-fold: on the one hand it is to explain how
a simplistic approach to systemic risk may fail to provide a correct clas-
sification of banks according to their systemic nature, and on the other to
give an insight into the alternative areas prudential policy could take care
of when addressing systemic risk (leverage, interconnectedness, etc.).
Following this line, we find that though far from perfect, the conclusions
that yield from our analysis are robust against any possible methodologi-
cal flaws since we emphasize the value of the approach from a policy
perspective, not methodological.
We deem it very important to contribute to a theoretical foundation for
the approach to this matter but also believe that it will take years if not
decades to consolidate a proven and unified theoretical approach. We
are aware that our analysis may fail because it does not rely explicitly on
a defined conceptual framework, however we have tried to incorporate
all the possible approaches found in the literature by means of including
the variables that academics and regulators use to set their conceptually
sustained views of systemic risk. Given that our indictor is equivalent
to a weighted average of other variables that the literature finds related
to systemic risk, we believe that our indicator could be interpreted as a
summary measure of these variables, too. Since each of these variables
are related to systemic risk within an specific conceptual framework, we
Acharya (2009) offers an overview of the regulatory and positive aspects of systemic risk, 4
under which he identifies it with the risk of the failure of the system as a whole arising from
the correlation of the elements (assets) on the balance sheets of institutions.
Guidance to assess the systemic importance of financial institutions in BIS-FMI-FSB (2009).5
77
more general definition of a systemic event as that generated by one
or more financial institutions and that has severe consequences on the
financial and real sectors. These documents presented a wide range
of approaches for the analysis that still coexist today with the different
strands of the literature, such as: network analysis, risk-based portfolio
models, stress testing, and scenario analysis. A vast amount of literature
attributes systemic risk to one of the sources mentioned above (size,
interconnectedness, unsubstitutability, and balance sheet quality8) and
uses one of the alternative approaches detailed below.
Bartman et al. (2009) use three methods to quantify the probability of
the systemic failure of the world banking system: correlations between
returns on assets, the Merton Structural Method (1974), and the infer-
ence of PDs from options prices. Their findings state that the probability
of a bank’s crisis leading to a systemic crisis is small. Tarashev et al.
(2009) use the Shapley’s Value to attempt to identify the impact of differ-
ent determining factors (balance sheet characteristics of an institution)
in the probability of the latter causing a crisis of similar characteristics.
Brunnemeier (2009) uses market data and proposes a measurement
(CoVar) for calculating the incremental risk produced by an institution,
measured as the difference between the VaR of the system, conditional
on the fact that an institution is in stress, less the VaR of the system in
normal conditions. Huang et al. (2009) measure the systemic risk of lead-
ing financial institutions by using an approximation of such risk called
“distress insurance premium,” which is equivalent to the premium that
the banking sector would pay to cover itself against a potential loss ex-
ceeding a specific share of the liabilities of the system (as an insurance
franchise). The premium is calculated with PDs and using the correla-
tion data among balance sheet assets. Acharya et al. (2009) use market
data to identify systemic risk with the expected contribution from each
institution to a systemic event (systemic expected shortfall, SES). And
they measure it as the cost to the banking sector when it becomes “infra-
capitalized,” weighted by the contribution made by the institution to such
event. The marginal contribution of each bank is linked to a rate requiring
it to internalize the negative externality of its participation in the system.
Their document has a prudential focus. Leaven and Valencia (2008) find
common patterns in the crisis that they identify as systemic; significant
signs of financial stress such as bank runs, bank liquidations, or close of
wholesale market financing accompanied by large-scale public sector
intervention in the system. However, the number of cases considered
systemic is scarce. Even in the latter cases, their frequency, intensity,
believe that by summarizing them into one it helps in gauging some parts
of those frameworks used.
We have decided not to take a unique and explicit conceptual approach
because we think that every framework is conditioned by its specific non-
closed and non-observable representation of systemic risk and will thus
remain subject to constant change and revision over the years.
Overview of existing research into the identification and measurement of systemic riskThe analysis of risk as a problem caused by an institution capable of caus-
ing damage to the general system has led to a widespread academic and
regulatory debate. The macro-prudential identification, assigning, mea-
surement, and management of the systemic nature of institutions either
focus on the implications of being too big, interconnected or important
to fail; measurement methodologies (CoVar, Systemic Risk shortfall, ex-
posure analysis, etc.); or variety in the use of market data versus bal-
ance sheet data. This has resulted in different ways of identifying risk, its
causes, and the policies needed to address it. It bears mentioning that a
systemic event is by nature non-observable and as such requires the use
of proxies to identify it. As stated in the memorandum of the Cross Border
Financial Stability Group6 [CBFSG (2008)], the identification of a systemic
event is changeable over time and cannot just be based on quantitative
criteria. There is no straightforward way of identifying a factor as being
systemic, but it is always conditional on a specific economic, regulatory,
or financial endowment that jointly with such factor may result in the ex-
ternality produced leading to a general system failure. Moreover, systemic
risks are not immutable; they depend on time and context and as such
may appear or disappear in a given institution. This makes it very difficult
to identify a systemic crisis and its connection to specific characteristics
of the banking system at a given moment. That is the reason why our
analysis does not attempt to seek causal relations between regressors
and a particular independent variable seemingly accounting for risk.
Although an ad-hoc quantitative criterion linking characteristics to sys-
temic events cannot be established, the expert’s view in the literature
may help to confine what we understand as systemic risk in its different
formats (collected in different variables). That such risk has systemic con-
sequences will not only be difficult to measure for the reasons mentioned
above, but will always be conditional on the definition we have given to a
“systemic event” and the proxy we have chosen to identify it with.
Seminally, the Bank of England’s Financial Stability Review (2001) pub-
lished some criteria for selecting a large and complex institution as a can-
didate for being systemic,7 although it offered no definition of systemic
risk as such. Jointly, the BIS/IMF/FSB in their paper “Guidance to assess
the systemic importance of financial institutions” (2009) and the Euro-
pean Central Bank (ECB)’s Financial Stability Review (2006) provided a
The Capco Institute Journal of Financial TransformationSystemic Risk, an Empirical Approach
Group made up of supervisors, central banks, and ministers of finance of the European 6
Union.
A large and complex financial institution (LCFI) is systemic if it is within the top ten in one or 7
more of the following categories: equity book runners, bond bookrunners, syndicated loan
bookrunners, interest rate derivatives outstanding, and holders of custody assets.
The Squam Lake Working Group on Financial Regulation (2009) also makes reference to 8
variables such as solvency, balance sheet composition, and liquidity of institutions.
78
and duration changes over the time, is heterogeneous and difficult to link
with the characteristics of the institutions making up the banking system
of that period.
This paper presents a pragmatic approach to systemic risk through the
use of quantitative analysis. Owing to the aforementioned difficulties in
isolating factors that explicitly provide information on systemic risk, the
methodology includes all the variables identified as significant based on
our approach, with the intention of obtaining the most robust, efficient,
and least biased results possible. To that end, the variables commonly
identified as sources of systemic risk have been used. Owing to the dif-
ficulty in specifying variables, sample, and method, the most general ap-
proach possible has been undertaken in the selection of variables and
the sample of banks used, allowing for the processing of the informa-
tion as objectively as possible. The methodological criterion followed to
obtain the indicator is the maximization of the information contained in
the sample and that such information remains structurally invariant over-
time.
The main weakness of our approach is the inherent lack of an unbiased
and robust indicator of systemic risk. However, our analysis is based on
a theoretical and conceptual approach whose validity is commonly ac-
cepted by academics and policymakers as conventional wisdom. The
methodology we have chosen particularly addresses the limits of CoVar,
which relies on market information, fails to account for procyclical effects
on institutions, and implicitly incorporates the probability of banks being
bailed out. Other approaches, such as network analysis, are data inten-
sive, computationally costly, and usually imply the introduction of strong
assumptions about system structure.
Variables accounting for systemically important sources of riskThere are different dimensions throughout the systemic nature of risks
assumed by an institution that may be analyzed. Table A1, in the Ap-
pendix, lists those variables classified by dimension that have been used
in creating our indicator. They depict the generally accepted sources of
systemic risk.
Size is the main criterion found in the literature. Many believe that size
is a direct channel for systemic risk to pass on to the real sector and the
rest of the financial system. Crisis within a large financial institution will
have a relatively bigger impact for the entire system. This proportion-
ally bigger impact comes from the greater difficulty in replacing the vital
functions provided by these institutions and to find a solution for orderly
wind down (i.e., more difficult to find a buyer) as well as from its greater
interconnections with the rest of the financial system. Size is also associ-
ated with riskier balance sheet structures as they face lower funding cost
just because creditors consider them as systemic and therefore more
likely to be bailed out. Typically, size variables focus on the size of the
balance sheet either as a whole or each of its parts (assets and liabilities).
Frequently, size indicators are based on assets, deposits, and/or credit.
In this analysis we will use variables of the three making them relative to
world GDP. We do this to render each measure comparable with each
other. Since we identify size as a source of systemic risk, these indicators
are expected to have a positive weighting on the total risk indicator.
The quality of the balance sheet – as was made clear by the Squam Lake
Working Group (2009) – plays a very important role, too, when defining
systemic risk. A weak and unstable balance sheet makes an institution
vulnerable to market, credit, and liquidity risks. Barrel et al. (2009) found
evidence of an increased probability of a systemic banking crisis when-
ever solvency, leverage, and liquidity are below the desired levels. In this
sense, it sounds sensible to direct policy to guarantee enough loss ab-
sorption capacity, sufficient liquidity, a correct risk administration, and a
sound ability to organically generate capital.
The literature divides balance sheet information into variable groups that
have gained/lost prominence over the years, in line with the nature of the
various emerging distresses during the lifecycle of the recent crisis. The
crisis began as a liquidity issue and then turned into a problem of adverse
selection when nobody knew the true quality of the assets in the balance
sheets. In the latter parts of the crisis, financial distress evolved into a
solvency issue as such, for which it was not only necessary to rely on
existing capital buffers (static solvency) but also on the ability to generate
capital organically9 (dynamic solvency).
Solvency indicatorsAn increase in these indicators augments balance sheet stability and
should result in lowering the risk of the institution. In this synthesis indi-
cator we would expect risk elasticity to solvency levels to be negative,
that is, we would expect our indicator to drop whenever our solvency
variables increase. The literature identifies the following ratios as espe-
cially significant:
Core tier 1 –■■ resembles the amount of pure capital held by an insti-
tution and therefore gives information about its maximum reaction
capacity in a going concern situation. It is generally made up of equity
and retained income.
Tier 1 –■■ takes into consideration the additional capital with which an
institution may deal with its solvency before going into liquidation. The
additional capital consists of preferred shares whose nature, incen-
tives, and consequences are less able to absorb losses.
In this sense, the strength of its results (what we refer to here as performance indicators) 9
gain true significance.
79
Leverage Ratio (and corrected leverage ratio in respect of off-■■
balance sheet assets) – both ratios refer to level of leverage an
institution has to deal with.
Off-balance sheet assets –■■ the aforementioned variables do not
give sufficient consideration to the contribution made to solvency by
off-balance sheet assets to the institution’s financial circumstances.
It is, therefore, necessary to include this variable, since it reflects the
greater level of flexibility that it provides an institution. Off balance
sheet items are guarantees, committed credit lines, and other contin-
gent liabilities. Other types of off balance sheet items, which are the
normally referred to as riskier off balance sheet items, such as struc-
tured products, are never available under balance sheet data.
Liquidity indicatorsThe outbreak of the crisis was identified on several occasions with a tem-
porary lack of market liquidity, either due to the interbank failure (in gen-
eral for the entire banking sector at the beginning of 2009), or because of
the absence of a wholesale market in which to liquidate structured assets
(i.e., Lehman), or because of bank runs in the retail sector (i.e., North-
ern Rock). The following are found among the indicators most frequently
used in available literature:
Deposits to assets –■■ in principle, deposits are a relatively stable
source of financing in normal times, although such stability may vary
in terms of the type of deposit and the depositor, as well as the level
of coverage. However, in times of crisis, deposits may become more
volatile.
Money market and short-term funding to assets –■■ financing in
wholesale markets should in principle pose a greater systemic risk,
inasmuch as they are a source of interconnections in the financial
system. On the other hand, the ability to finance itself in these mar-
kets gives a bank greater diversification in its sources of financing
and therefore contributes to its financial strength. Since the effect of
greater interconnection is taken up by the variable for financing in the
interbank market, it cannot be ruled out that the contribution of this
variable to systemic risk will be negative, as it picks up the additional
(risk reducing) effect mentioned.10
Asset mix indicatorsAs a measure of balance sheet quality, the available literature identifies
credit risk and market risk using two indicators:
Total gross lending to assets –■■ this indicator reflects the weight
of the traditional businesses on a bank’s balance sheet. We would
expect a negative contribution of this variable to systemic risk.
Share of on-balance sheet securities –■■ this reflects the weight that
market activities can have on the bank’s balance sheet. These activi-
ties, as well as being more volatile, present a greater degree of inter-
connection with the rest of the financial system. We would, therefore,
expect a positive contribution to systemic risk.
Risk management quality indicatorsThese work as proxies of the business concern for a correct risk admin-
istration.
Short-term debt to total debt –■■ this variable accounts for the risk of
refinancing, which could be acute in a general funding crisis, as was
experienced during the recent crisis. The (negative) sign of this vari-
able could be conditioned by the fact that during the sample period
it becomes more difficult to get access to long term market funding.
This suggests that factors limiting the issuance of debt have a greater
impact on short term debt that cannot be refinanced than medium
and long term debt. Accordingly, the ratio of short term versus total
debt also decreases. This effect should be more pronounced for those
institutions that have greater refinancing difficulties. Consequently, a
lower ratio could be positively correlated with systemic risk in crisis
times when market discipline is exaggerated, as opposed to what we
would expect in normal times when maturity mismatches could result
in the buildup of huge liquidity risks.
Net interest margin –■■ a culture that focuses on maximizing earnings
is deemed to help mitigate risk. The elasticity of the system to this
factor is interpreted as negative. Net interest margin is not net of loan
loss provisions (LLP). We acknowledge that this could pose problems
for comparability, but believe that the effect of LLP is already consid-
ered in the ROE item of the model. On the other hand, detracting LLPs
from the margin could bring some degree of distortion to our analysis
because LLPs are not homogenously accounted for in every bank.
The unsubstitutability of certain strategic services may increase the
negative externality produced by its disappearance. Unsubstitutability
depends on an institution’s monopoly position in a certain market and
the complexity of the service it offers. A service that is difficult to replace
is that of clearing and custody. This paper documents unsubstitutability
as the degree to which an institution contributes infrastructure to the fi-
nancial system. In this case we address the implications related to clear-
ing and settlement infrastructure. We approximate this variable with the
proportion of assets safeguarded by custodian banks over the total vol-
ume of assets under custody. We understand that this indicator should
contribute to the total risk of the system and therefore have a positive
weighting. We advise caution with data of this type, as the source from
which they are extracted is not the balance sheet but voluntary contribu-
tions to the Global Custody Net. This source of data does not take into
We have tested this by estimating the indicator in absence of the interbank financing 10
variable and the result was as expected.
The Capco Institute Journal of Financial TransformationSystemic Risk, an Empirical Approach
80
recommend caution in this regard, we also consider that other variables
also indirectly contribute additional information on interconnection and
therefore filter the possible distortion introduced by the infra-specifica-
tion of the model. The custodian nature of a bank also provides informa-
tion about the degree to which it is interconnected with other institutions.
This is in addition to the interconnections via the value of the assets of
other banks that are held in its custody.
Performance indicatorsPerformance indicators take into consideration the contribution of an
institution to the capacity of the system to organically generate liquid-
ity and resources to remain solvent. Returns and efficiency indicators
provide insights into the stream of flows that could be retained in order
to increase capital. They are also important signals that the market tracks
when banks seek to raise capital from them.
Return on average equity and return on average assets –■■ both
indicators should reduce the systemicity of a banking institution rep-
resented in the market, as they are indicators of strength. We expect
the elasticity of risk to these components to be negative.
Cost-to-income/efficiency –■■ it is understood that the greater the
efficiency of an institution, the more its capacity to retain profit and
generate capital organically. Efficiency weights negatively on total
risk. Hence, cost-to-income should contribute positively.
In the current literature we find other variables identified as sources of
risk. In considering size, the ECB finds it important to use the number
of affiliates, whilst the BIS/IMI/FSB take into consideration the market
share of each institution (in loans and deposits). These variables have
not been used in this analysis as they are only found at the aggregate
level. The ECB finds sources of risk on the side of liquidity in contin-
gent liabilities and additional factors arising from the asset mix related
to cross-border and interbank assets. With respect to risk management
quality, several approaches in academic literature find risk mitigation
factors in the degree of diversification, and incremental effects in the
degree of intermediation with financial derivatives (trade income) and
the foreign exchange market (FX income). These variables have not been
used in the analysis because they are not available in the 11 databases
used (Bankscope or any other publicly available database with data for
the sample used). Figure A1 in the Appendix, also shows a list of market
Nicholo and Kwast (2002) identify the degree of interconnection as a source of systemic 11
risk. IMF GFSR introduces the “too interconnected to fail” concept. CBFSG also identifies
the degree of interconnection as one of the sources of risk. Chan-Lau (2009 a,b) uses
network analysis on balance sheet data to evaluate interconnection risk when one or
more institutions suffer a credit or financing shock, and draws up an interconnection risk
measurement that is equal to the average capital loss suffered by the banking system due
to one or more types of shocks to such the system.
consideration the nested positions of accounts under the custody of one
unit within another, therefore the data contributed by some institutions
will contain a certain level of bias.
The variables relating to the system of payments also offer information
on the degree of unsubstitutability of a specific institution. From the per-
spective of the ECB, banks, in terms of the unsubstitutability of their pay-
ments function in the system, are either considered systemically impor-
tant (SIRP) or not (PIRP). However, the ECB does not provide any metric
that could be used within the scope of this analysis.
As stated before some indicators of the degree of concentration of mar-
ket services may also have a reading for the degree of substitutability
that these have. In this vein, although we have grouped some indicators
under the size category, the information that they carry also refers to the
degree of substitutability though only in a very rough and approximate
way since only size and not concentration factors have been taken into
account.
The current literature also distinguishes between the factors that attri-
bute risk indirectly,11 that is to say, factors that are transmitters of the
externality produced when the risk becomes systemic. These variables
are gathered together as indicators of the interconnectedness among the
items making up the financial system. In practice, market indicators are
used, such as those shown on the attached Table A1, or alternatively,
balance sheet data. We could only approximate this variable by means of
the level of interbank exposure of each institution and acknowledge that
this approximation may be relatively partial. The importance of intercon-
nectedness may be underestimated in our analysis for this reason but it is
the furthest we can get using this type of data and approach. Most of the
exercises in this vein are inevitably theoretical or simulation-based due
to the absence of data and require large computational workload, thus
falling beyond possible scope of this analysis.
We have chosen to introduce an interconnectedness variable based on
the degree of exposure of each bank to the interbank market. We under-
stand that this variable gives information on the amplifying effect of its
own risk to the rest of the market, and we therefore believe that it should
have a positive weighting in our risk indicator. Furthermore, this variable
can be read from a liquidity standpoint since it may be a source of in-
stability given the dependence of the institution on the interbank market
and not the real sector of the business as such. The recent crisis during
which financing was almost impossible in the interbank market, is a good
example of this risk. Once again, the elasticity of systemic risk to the
banking interconnection indicator of an institution should therefore be
positive. The shortage of variables (available as balance sheet data) that
could provide information on interconnection risk could pose a specifica-
tion problem for the estimation of the indicator. However, although we
81
creating a global indicator. This means that our analysis underestimates
the systemic risk of local banks.
A recognized problem of any approach dealing with bank data is that
it fails to gauge other financial but not banking-based sources of risk
such as those stemming from the insurance sector and other financial
intermediaries. We would like to extend the scope of this analysis in the
near future.
Certain variables that are important for defining the indicator (ROA, in-
terest margin, etc.) depend on local conditions which make comparison
among banks located in different geographical locations difficult (i.e., in-
terest rates or margins resulting from the tightness of the market in which
each institution is located), but this does not prevent such factors from
being a significant risk factor for each institution, and should therefore be
considered as part of their vulnerabilities (or strength).
Furthermore, we acknowledge that our dataset could also be partially
influenced by public aid programs put in place, which have an impact on
the balance sheets of a number of banks. However, we believe that these
data are the best one can obtain and are significantly less distorted than
market data. We, therefore, choose to use them.
An additional issue arising from the use of data is that balance sheet data
are unable to highlight latent imbalances within the banking system that
ultimately result in crises. This issue is wholly uncontested, since even
the market data used under the ECB indicator for systemic risk did not
anticipate the looming crisis. Nevertheless, that distortion is considered
in our conclusions.
MethodologyOur approach adheres to the position taken by the Bank of Spain (2010)
and strives to come close to the approaches found at the ECB (2006) and
the IMF (2010). These define a series of balance sheet variables to estab-
lish a systemic order score to individual banks. To this end, and with the
aim of allowing the systemic relevance of each variable to the total aggre-
gate indicator to be assigned objectively, we have employed a “principal
component analysis” approach (PCA). We have chosen this methodology
over a CoVar or network based approach, even though we consider them
very useful approaches for theoretical and monitoring purposes, since they
fail to provide insight on the effect of time on systemic risk sources that we
attempt to address. Besides, we aim to anticipate the affordable alterna-
tives that the regulators will consider in their assessment of the systemic
nature of banks. The latter will very likely be a compound measure of the
Many of these banks (24) make up the list of banks requiring cross-border supervision, 12
according to the FSB.
driven variables that tend to be used as interconnectedness proxies.
These variables have not been used because of the problems mentioned
and due to their availability.
DatabaseThe analysis below has been undertaken using balance sheet data. The
current literature documents the distortion produced on market data when
the financial markets are distressed. The volatility of returns and CDSs, for
example, was strongly altered by public aid programs implemented across
the board during the recent crisis, programs that actually prevented policy
makers from gauging the true picture of each bank’s real risk state. On the
other hand, there is a consensus that in times of instability financial mar-
kets tend to become shortsighted and irrational. An example of this is the
fact that for a long time the market attributed the same risk (CDS spread)
to the bond of a Spanish multinational company (with 83 percent of its
market revenues coming from outside the country) as it did to a Spanish
Treasury bond. In view of these problems we consider the second alterna-
tive the most reliable way of approximating sources of risk.
We recall that the variables considered are all those available indica-
tors stated in academic and regulatory literature. With the same spirit
of generality, we have chosen a sample of banks with total assets ex-
ceeding one hundred billion euros.12 Institutions for which Bankscope
and Bloomberg did not have the balance sheet and earnings statement
data needed have been excluded from this analysis. As a result, we have
a sample of 100 of the largest banks. The estimation was carried out in a
panel data fashion for the 2006 -2009 sampling period. We have focused
on the results of a panel data to avoid inconsistencies arising from the
size of the sample and the variability of the sources of risk from year to
year. With the panel data analysis, we obtain robust estimations of fac-
tors over time, not determined by the nature of the emerging risk during
a given year.
A cautious approachWe recognize that certain information related to risk mitigating factors has
not been considered in the selection of variables. This was due to avail-
ability reasons as well as the diversity of the data. The corporate structure
of the different institutions also makes it impossible to compare many di-
versification data among the different banks. Other risk mitigation sources
are the degree of diversification and the type of legal structure of each
bank, elements that have proved to be fundamental pillars of the institu-
tions that have best survived the crisis. This means that our analysis over-
estimates the systemic risk of highly diversified and organized banks.
Variables of size have been taken relative to world GDP. We understand
that this decision may limit to a certain degree the ability of the indica-
tor to perceive the systemic risk of an institution at the local level, but
this decision is justified for reasons of comparability and consistency in
The Capco Institute Journal of Financial TransformationSystemic Risk, an Empirical Approach
82
various systemic risk sources identified. Performing conditional PCAs was
an appealing alternative but it would have implied identifying, a priori, some
closed form of systemic risk. This could introduce some bias to our analy-
sis. As stated before, we are trying to implement the general and over-
arching view found in the literature. This method allows us to reduce the
information gathered in our sample of indicators to a much lower number
of synthetic indicators that provide segments of independent information.
Obtaining such indicators requires estimating the factor loads on which
the original variables will rely so that they form the final indicator. Once the
information has been reduced and segmented into such indicators, we
resort to an expert opinion to decide which of them could be interpreted
as an indicator of systemic risk. By construction, once it has been decided
that a specific component contributes specific information, such as that of
systemic risk, the possibility that such information could be found in anoth-
er of the discarded indicators is excluded. Once the indicator is obtained,
its loads are used to weigh up the characteristics gathered from each bank
and its systemic score is calculated. The estimation process is detailed in
the methodological appendix.
ResultsCharacterizing the sources of riskBy means of the PCA, we extract the first p-factors that explain about
two-thirds of the information contained in the dataset. These factors are
weighted sums of the systemic risk variables. From these factors, we
select our systemic risk indicator candidate on the basis of the proportion
of information from the original data that it summarizes and on the basis
of its consistency with the preconceived idea of the characteristics that
such indicator should have, both internally and externally.
We are focusing the analysis on the results extracted under a panel data
approach estimated through (part of) the cycle that goes along the crisis
years 2006 to 2009. Examining the loads from the aforementioned indica-
tors (Table 1), we again identify a distinctive pattern in the first indicator.
We observe a unique attribute in that this indicator is a linear combination
(a weighted sum) of variables whose value increases when: the ability to
organically generate resources deteriorates, the activity devoted to retail
business is reduced, leverage is increased, size is increased, the volume
of financial assets (securities) on the balance sheet is increased, access
to market funding falls, the ability to retain resources deteriorates due to
efficiency costs, interbank market funding dependency is high, exposure
to the interbank market increases, the number of asset under custody
increases, solvency is reduced.
Consequently, all the characteristics included in the current literature on
the typification of a synthetic systemic risk indicator are met. Also, the
indicator preserves the desirable characteristics mentioned above con-
cerning information explained and stability. We will call this indicator the
synthesis systemic risk indicator (SSRI).
Direct and total impact of risk sources on the SSRIOn examination of the indicator loads, it is possible to assess the direct
effects of the various sources of risk on the SSRI. We use the weighted
elasticity of SSRI on each of the risk variables to derive these direct ef-
fects.14 By doing so, one would find that the risk arising from size weights
are approximately 19 percent of the indicator, while solvency and liquidity
strike are 26 percent. The capacity to organically generate capital in a
recurrent manner contributes 21 percent, while the sound risk manage-
ment policy is 14 percent. Asset mix weights are 12 percent and the
risk arising from interconnectedness and unsubstitutability of services
would seemingly add 8 percent to the weighting. It is important to bear
in mind, however, that risk sources interact with each other throughout
Variables13 / Factors γ1SR Weight
of γ1SR
γ2 γ3 γ4 γ5
Assets 0.32 8.0% -0.26 -0.27 0.27 -0.02
Credit 0.21 5.3% -0.41 -0.16 0.35 -0.12
Size Deposits 0.24 6.0% -0.42 -0.29 -0.04 -0.04
Core Tier 1 -0.15 3.9% 0.11 -0.30 -0.12 -0.23
Tier 1 -0.05 1.3% 0.24 -0.41 -0.20 -0.17
Off Bal. -0.13 3.3% 0.03 -0.09 0.30 0.23
Leverage 0.34 8.5% 0.10 0.17 -0.03 0.37
Solvency Real LR 0.13 3.3% 0.13 0.04 0.12 0.19
Depot/Assets 0.02 0.5% -0.36 -0.12 -0.52 0.07
Liquidity ST. Funding -0.19 4.8% 0.25 -0.10 0.45 0.01
Gross Loans -0.20 5.0% -0.30 0.34 0.17 -0.12
Asset Mix Securities 0.28 7.1% 0.16 -0.36 0.03 0.15
Short T. Debt -0.24 6.1% 0.03 -0.32 -0.12 -0.03
RMQ NIIM -0.34 8.6% -0.03 -0.17 0.14 0.03
Custody 0.08 2.1% 0.13 -0.27 0.25 -0.02
Unsust. &
Interc.
Interbank Exp. 0.22 5.5% 0.17 -0.02 -0.06 0.46
ROAA -0.36 8.9% -0.14 -0.19 -0.02 0.31
ROAE -0.24 6.1% -0.17 -0.12 -0.08 0.53
Performance Cost-to-Inc. 0.23 7.7% 0.29 -0.02 -0.16 -0.19
Est.Quality λ 4.26 – 2.43 2.3 1.4 1.3
% of variance 22.4% – 12.5% 12.2% 7.6% 6.9%
Determinant 4.26 – 10.4 24 34.5 46
Table 1 – Sources of risk extracted principal components and candidates
for systemic risk indicator (λSR)
Size variables are expressed relative to world GDP. Leverage Ratios are expressed in 13
times. Liquidity and asset mix ratios are expressed as share of Assets. Short-term debt is
expressed relative to total debt. Custody assets are rendered to total custody assets in the
system. RMQ stands for Risk Management Quality.
We take the effect by normalizing the elasticity of the SSRI on each of the identified 14
sources of risk.
83
innumerable channels. This is easily observable by having a look at the
correlations among the risk variables selected for this analysis (Table A2
in the Appendix) where we find that size, solvency, balance sheet quality,
and performance are significantly correlated. This means that risk factors
operate not only in a direct manner but also throughout other indirect
channels that reduce or amplify their direct effect according to the cor-
relation shared with other risk variables.15 We use these correlations to
assess the direct and indirect impact of each of the risk variables and
find that the total impact of each variable group, taking into consideration
such cross-effects, would yield: 31 percent for size, 18 percent for sol-
vency and liquidity, 17 percent for risk management quality, 15 percent
to the capacity to organically generate capital in a recurrent manner, 12
percent for balance sheet composition, and 7 percent for interconnect-
edness and unsubstitutability as a whole.
From a policy point of view, the latter brings us to consider the cross ef-
fects among risk factors, as these may amplify or mitigate the originally
envisaged major sources of risk. A policy rule that considers risk fac-
tors should strive to be as diversified as possible to remain neutral on
the innumerable cross effects among variables. Consequently, taking a
wide set of variables in our measure would be advisable. Likewise, we
consider that from the prudential viewpoint, the systemic risk mitigation
measures imposed on an institution should depend on the factors that
cause such risk. Thus, certain solutions may be presented:
1. Through an increase of the levels of solvency (Core tier 1, leverage
ratio, etc.) and liquidity, an option already addressed by Basel III,
2. Treatment of the rest of the sources of systemic risk that have also
proved to be significant.
As stated before, any policy rule should bear in mind not only the ben-
efits but also the costs involved in such measures. These measures are
not substitute but complementary measures. Enhancing balance sheet
composition, risk management quality, and the ability to generate capital
organically in a recurrent manner are measures as effective as others and
do not involve direct costs that could restrict credit in quantity and price.
On the other hand, additional capital charges could have a nil marginal
effect on the level of total risk and their cost could exceed the additional
stability they provide.
Systemic risk through the cycle 2006 -2009Owing to the fact that the objective of the indicator is a prudential frame-
work, we should bear in mind that creating an indicator based on these
factors in a specific year to apply systemic measures in future periods is
not forward-looking and could result in penalizing the institutions for char-
acteristics they no longer have, or (or vice versa) reducing their systemic
load, although the risk information deduced from the ratios in a specific
year had worsened. Table 2 shows average band errors committed one
and/or two periods forward if the load estimated in the year (t) to add to
the score of each institution according to the characteristics found in
year t+k16 was used. We observe that the risk scores are systematically
undershot. However, the error made, if we use the average of the loads,
would be much less, and that is the reason for choosing to use a central
risk measurement. A panel estimate offers that central risk measure, con-
sistent with the scores obtained year by year, which preserves the ability
to correctly delimit the systemic institutions.
Table 3 shows the weighted impact of each risk indicator selected every
year between 2006 and 2009. As can be seen, the interpretation of the
loads remains consistent throughout time, only varying in the intensity
of the factors, but not in their sign. In the same way, we can also see
that the indicator estimated over the years collects similar proportions of
information, as reflected by trace criterion of the indicators. We see that
although the significance of certain factors has changed over time, this
change has been consistent with the changing nature of the crisis. For
example, we observe that the weight of exposure to the interbank market
in the indicator was twice as much in 2007/2008 than in 2009, when this
market began to reactivate itself. And we also see that certain factors
have lost importance as a result of the absence of problems arising from
them, such as size, custody, and the existence of off-balance sheet as-
sets. The changes in the composition of the indicator have only involved
intensity (weight of the factors) and not sign, showing the stability of the
estimation. However, the variability of such composition should be borne
in mind on designing the indicator for macro-prudential purposes. The
risk factors collected in 2007 (adverse selection versus balance sheet
composition) were different from the 2008 risk factors (solvency crisis) or
2009 (liquidity crisis), and will therefore have a different weighting in the
risk score obtained each year.
A measure of system-wide systemic riskOn having an estimated indicator by means of panel data, we can consis-
tently apply the loads of the SSRI indicator to the variables of the banks
The Capco Institute Journal of Financial TransformationSystemic Risk, an Empirical Approach
One step Two steps Average
Upper quartile -1% 16% 6%
Median -28% 14% -13%
Lower quartile -72% -87% -40%
Table 2 – Average forward scoring error – one and two steps ahead
Asset volume, for instance, represents a risk in itself and through other size and balance 15
sheet factors that add to the weighting. However, since the same variable is negatively
correlated to performance and solvency variables, these detract from the final effect, too.
The error is the difference between the score obtained in the current year, obtained with 16
loads estimated on the basis of balance sheet information for that year, less the same
score obtained with the characteristics of the current year but with loads estimated on the
basis of information existing one or two years previously.
84
each year. This does not just make it possible to establish intra-group
bank comparisons each year, but also to check an institution’s degree of
systemicity over the time. Additionally, given that institutions would be
comparable with themselves and with the rest of institutions over time,
we are able to obtain a measure of systemicity of the entire system over
the period analyzed.
In Figure 1, a measure of system-wide systemic risk has been plotted.
This measure is obtained by showing the median risk score found in the
system along the sample period. One may distinguish how (according
to our measure) banking systemic distress was rather contained in 2006
and how systemic risk increased to full-blown levels during 2008. It was
only in 2009 when after massive international policy action (monetary and
fiscal stimulus, bail out, new regulatory regimes, accounting changes) the
risk was leveled down somehow, although the looming threat to the sys-
tem is far higher than normal times (2006). This indicator offers the same
reading as the systemic risk indicator plotted by the ECB in its Financial
Stability Report since 2009.
Distribution and classification of banks according to systemic scoringWhen we compare how the 100 banking institutions are distributed ac-
cording to the chosen indicators for systemic risk (figures that are avail-
able from the authors on request), we find that the distribution of the
institutions is very asymmetrical in all the variables used. The asymmetry
in the separate distributions of each variable may lead to a biased selec-
tion of banks which are considered most likely to be systemic. This is
even more likely when fewer criteria are used to delimit the systemic risk
Variables/factors 2009 2008 2007 2006 Average Consistent
Size Assets 0.32 0.29 0.31 0.36 0.31 yes
Credit 0.21 0.19 0.2 0.26 0.21 yes
Deposits 0.24 0.21 0.21 0.29 0.23 yes
Solvency Core -0.15 -0.25 -0.17 -0.02 -0.14 yes
Tier 1 -0.05 -0.03 -0.16 -0.08 -0.03 yes
Off balance sheet -0.13 -0.12 -0.15 -0.19 -0.11 yes
Leverage 0.34 0.33 0.35 0.33 0.32 yes
Liquidity Real LR 0.13 0.23 0 0.13 0.12 yes
Deposits/assets 0.02 -0.03 -0.01 0.08 0.01 yes
Asset mix ST. funding -0.19 -0.14 -0.2 -0.18 -0.19 yes
Gross loans -0.2 -0.19 -0.2 -0.21 -0.2 yes
RMQ Securities 0.28 0.3 0.23 0.27 0.28 yes
Short-term debt -0.24 -0.25 -0.23 -0.16 -0.24 yes
NIIM -0.34 -0.31 -0.35 -0.34 -0.33 yes
Unsubstitutability and
inter connected ness
Custody 0.08 0.07 0.08 0.08 0.1 yes
Interbank exposure 0.22 0.2 0.25 0.24 0.2 yes
Performance ROAA -0.36 -0.33 -0.38 -0.37 -0.35 yes
ROAE -0.24 -0.31 -0.22 -0.15 -0.24 yes
Cost-to-income 0.23 0.2 0.25 0.17 0.23 yes
Estimated quality 2009 2008 2007 2006 Average
Cumulative variance λ1 22.4 26.4 24.8 21.9 23.9
λ2 35.2 39.7 37.6 37.1 37.4
λ3 47.4 51.0 50.2 48.8 49.4
λ4 55 58.9 58.5 57.2 57.4
λ5 63 66.5 66.0 64.9 65.1
Cumulative determinant λ1 4.6 5.0 4.7 4.2 4.4
λ2 10.4 12.7 11.5 12.0 11.8
λ3 24.0 27.2 27.4 26.8 27.5
λ4 34.5 40.9 43.5 42.5 44.0
λ5 45.5 58.8 62.2 62.6 63.8
Table 3 – Risk weightings evolution along 2006-2009 for (λSR)
85
perimeter. When we examine what would happen if we use the criterion
of size on classifying the 100 institutions as systemic when they breach
median threshold we find that 16 percent of banks are classified as sys-
temic when they are not (type I error), whilst at the same time one would
be ignoring 14 percent of the sample institutions that (according to SSRI)
are systemic (type II error). The expected stability cost of such a misclas-
sification justifies the selection of a synthetic indicator that uses a richer
set of risk variables to gauge all possible systemic risk sources embed-
ded in each bank. An indicator combining all the variables, such as the
SSRI, renders the distribution of banks the closest to normality (accord-
ing to JacqeBera) vis-a-vis the systemic variables taken independently.
This allows a scaling of the resulting banking systemic scores closer to a
continuum; thus, making the setting of thresholds for qualifying systemic
less discrete and less partial.
These results have several policy implications. A prudential policy based
on a single variable would give rise to a partial and possibly wrong clas-
sification, of both systemic and not systemic banks (type 1 and 2 errors),
introducing a high expected cost in the event of crisis. Similarly, focusing
on a single dimension of one variable would bias the choice towards a
single source of risk, when the criterion followed indicates that the sourc-
es of such risk are multiple (i.e., in the case of size: assets, liabilities,
credit risk, market risk, etc.).
ConclusionThe main conclusions of this analysis are:
All financial institutions contribute to systemic risk, albeit to a different ■■
degree. Their contribution is driven by various risk factors embedded in
multiple groups of variables such as size, interconnection, unsubstitut-
ability balance sheet quality, and the quality of risk management.
Using a single variable, or a limited series of variables, as a proxy for ■■
systemic risk fails to provide a full explanation and results in numer-
ous errors when attempting to identify and measure the systemic risk
of each institution.
Based on this analysis, we advocate the following from the policy point
of view:
When designing systemic risk mitigation measures, all contributing ■■
factors need to be taken into account in the policy cost function. A pru-
dential policy based on a single variable and on a single dimension of
a variable would give rise to an incorrect classification both of systemic
institutions (type 1 error) and institutions not considered systemic (type
2 error), introducing a high expected cost in the event of crisis.
Likewise, classifying institutions as systemic/non-systemic would ■■
mean giving similar treatment to institutions that might actually have
very different degrees of systemic risk whilst treating very differently
institutions that may present similar levels of systemic risk.
A continuous measure of risk is preferable to a simple systemic/non-■■
systemic classification; under this approach all institutions would be
systemic but to varying degrees.
Systemic risk may be addressed through different measures: reinforc-■■
ing solvency and liquidity standards as already envisaged by Basel III
(increasing capital levels and adjusting capital charges to riskier activ-
ities like those held in the trading book); addressing the other relevant
sources of systemic risk linked to each institution’s risk profile, that
is improving risk management, corporate governance, and adopting
a business model capable of organically generating recurrent capital;
and addressing an institution’s complexity, market services unsubsti-
tutability, and interconnectedness within the system.
We consider that a purely quantitative approach has serious limitations
as it does not take into account some relevant unquantifiable factors that
could mitigate or exacerbate systemic risk. This analysis attempts to be
as broad and objective as possible and also considers the qualitative
dimension by using other quantitative indicators as a proxy for it.17 How-
ever, due to the unavailability and diversity of relevant data,18 the analysis
fails to consider risk mitigation sources such as the quality of supervision,
the regulatory framework, and the crisis management framework of each
jurisdiction, or the degree of diversification and the legal structure of a
bank. All of which are elements that have been proven to be fundamental
pillars of the financial systems that have best survived the crisis. That
said, we believe that this reinforces the conservative approach of our
analysis, and hence its conclusions for those systems where these vari-
ables are of a high quality.
The Capco Institute Journal of Financial TransformationSystemic Risk, an Empirical Approach
-
0,3
0,5
0,8
1,0
2006 2007 2008 2009
SSRI system-wide = median of systemic charge
Figure 1 – System-wide level of systemic risk
That would be the case for example of risk management quality variables used here.17
The corporate structure of the different institutions also makes it impossible to include 18
diversification as a specific variable for the analysis.
86
A macro-prudential tool should not focus solely on quantitative indica-
tors but consider qualitative factors as well. These factors prove crucial
in defining the risk profile of an entity. The complexity, volatility of bank
operations, transparency, its cross-border legal structure, and degree of
exposure to financial markets and the existence of living wills are all good
examples of the aforementioned qualitative dimension that could help to
mitigate or exacerbate the degree of systemic risk in the system. Particu-
larly relevant to assess systemic risk is the information provided by the
recovery and resolution plans as they provide information on a crisis as-
sumption that complements the business as usual information provided
by accounting statements.
Although we recognize the relevance of having sound capital and liquidity
standards, we also deem very important having good internal risk man-
agement and corporate governance practices. All the latter framed within
a set of close micro-prudential supervision policy scheme. In order to
avoid excessive complexity at the bank level and to facilitate substitut-
ability of vital services, we find crucial the development of recovery and
resolution plans (living wills). The latter, jointly with a strengthening of
international coordination for crisis management, will reduce the risk of
contagion. Equally important is that macro-supervision is reinforced to
take into full consideration the risks derived from interconnectedness.
Finally, a prudential solution should consider not only the benefits but
also the costs involved in such measures. The above-mentioned factors
are equally effective as other measures, but crucially they do not involve
direct costs that could restrict the viability of institutions or the availability
of credit, in quality and price. Having said that, additional capital charges
could have a nil marginal effect on the level of total risk and their cost
could exceed the additional stability they provide.
ReferencesAcharya,V., 2009, “A theory of systemic risk and design of prudential bank regulation,” CEPR •
Discussion Papers, 7164
Acharya, V., H. L. Pedersen, T. Philippon, and M. Richardson, 2009, “Measuring systemic risk,” •
Federal Reserve Bank of Cleveland Working Paper, 1002
Aikman, D., P. Alessandri, B. Eklund, P. Gai, S. Kapadia, E. Martin, N. Mora, and G. Sterne, •
2009, “Funding liquidity risks in a quantitative model of systemic stability,”Bank of England
working papers, 372
Bank of England, 2001, “Financial Stability Review,” 11•
Barrell, R., E. P. Davis, T. Fic, D. Holland, S. Kirby, and I. Liadze, 2009, “Optimal regulation of •
bank capital and liquidity: how to calibrate new international standards,” Financial Services
Authority, Occasional Papers, 38
Basel Committee on Banking Supervision, 2010, “Report and recommendations of the cross-•
border bank resolution group,” March
BdE - Associate directorate general banking supervision, 2010, “The treatment of systemic •
institutions”
Beville, M., 2009, “Financial pollution: systemic risk and market stability,” Florida State •
University Law Review, 36, 245-247
Blanchard, O. J., C. Cottarelli, and J. Viñals 2009, “Exiting from crisis intervention policies,” •
Fiscal Affairs, Monetary and Capital Markets, and Research Departments, Policy Papers, IMF
Brunnermeier,A., 2009, “CoVaR,” Federal Reserve Bank of New York - Staff Reports, 348•
Chan-Lau, J., 2009a, “Regulatory capital charges for too connected to fail institutions: a •
practical proposal,” IMF Working Papers, 10/98
Chan-Lau, J., 2009b, “Balance sheet network analysis of too-connected -to-fail in global and •
domestic banking systems,” IMF Working Papers, 10/107
Claessens, S., G. Dell’Ariccia, D. Igan, and L. Laeven, 2010, “Lessons and policy implications •
from global financial crisis,” IMF Working Papers, 10/44
Cross Border Financial Stability Group, 2008, Memorandum of understanding on cooperation •
between financial supervisory authorities, central banks and finance ministries of the EU on
cross border financial stability
Cumming, C. and R. Eisenbeis, 2010, “Resolving troubled systematically important cross-•
border financial institutions: is a new corporate organizational form required?” Federal Reserve
Bank of New York, Staff Report, 457
De Bandt, O. and P. Hartmann, 2000, “Systemic risk: a survey,” ECB working paper, 35•
De Bandt, O., P. Hartmann, and J. L. Peydrós, 2009, “Systemic risk in banking“, Oxford •
Handbook of Banking, 145-160
Detragiache E. and D. Ho, 2010, “Responding to banking crisis: Lessons from cross-country •
evidence,” IMF Working Papers, 10/18
Demirgu-Kunt, A. and E. Detragiache, 2010, “Basel core principles and bank risk: does •
compliance matter?” IMF Working Papers, 10/81
Dungey, M., R. Fry, B. Gonzalez-Hermosillo, V. Martin, and C. Tang, 2010, “Are financial crisis •
alike?” IMF Working Papers, 10/14
ECB, 2006, “Financial Stability Review,” December and June•
Espinosa-Vega, M. and J. Solé, 2010, “Cross-border financial surveillance: a network •
perspective,” IMF Working Papers, 10/105
Fontene, W., W. Bossu, L. Cortavarria, A. Giustiniani, and D. Hardy, 2010, “Crisis management •
and resolution for a European banking system,” IMF Working Papers, 10/70
FSB-IMF-BIS,2009, “Guidance to assess the systemic importance of financial institutions, •
markets and instruments: initial considerations,”
Gauthier, C., A. Lehar, and M. Souissi, 2010, “Macro-prudential regulation and systemic capital •
requirements,” Bank of Canada Working Paper/Document de travail, 2010-4
Huang, X., H. Zhou, and H. Zhu, 2009, “A framework for assessing the systemic risk of major •
financial institutions,” BIS Working Papers, 281
Hund, J., S. M. Bartram, and G. W. Brown, 2009, “Estimating systemic risk of the financial •
system,” Journal of Financial Economics, 86:3, 835-869
IMF, 2010, “Systemic and the redesign of financial regulation,” Financial Stability Report, April•
Kick, T., T. Koetter, and T. Poghosyan, 2010, “Recovery Determinants of distressed banks: •
regulators, market discipline or the environment?“ Deutsche Bundesbank, Research Centre –
Discussion Paper Series 2: Banking and Financial Studies, 2010/02
Leaven, L. and F. Valencia, 2008, “Systemic banking crises: anew database,” IMF Working •
Papers, 08/224
Leaven, L. and F. Valencia, 2010, “Resolution of banking crisis: the Good, the bad, and the •
ugly,” IMF Working Papers, 10/146
Nicholo, G. and N. Kwast, 2002, “Systemic risks and the macroeconomy,” IMF Working Papers, •
10/29
Nicholo, G. and M. Lucchetta, 2010, “Systemic risks and the macroeconomy,” IMF Working •
Papers no. 10/29
Osborne, M., 2009, “Bank regulation, capital and credit supply: measuring the impact of •
prudential standards,” Bank of England Occasional Paper Series, 36
Popov, A. and G. F. Udell, 2010, “Cross-border banking and the international transmission of •
financial distress during the crisis of 2007-2008,” European Central Bank in its series Working
Paper Series, 1203
Sarkar, A. and J. Shrader, 2010, “Financial amplification mechanisms and the Federal Reserve’s •
supply and liquidity during the crisis,” Federal Reserve Bank of New York -Staff Reports, 431
Schoenmaker, D., 2009, “The quest for financial stability in Europe,” Chapters in SUERF •
Studies, 61-4, 25
Squam Lake Group, 2009, “Reforming capital requirements for financial institutions,” Squam •
Lake Group working papers, April
Tarashev, N. and C.Borio, 2009, “Attributing systemic risk to individual institutions,” BIS •
Working Papers, 308
Tarashev, N., C.Borio, and K. Tsatsaronis, 2009, “The systemic importance of financial •
institutions,” BIS Quarterly Review, 2009
Vesala, J., 2009, “How to bring systemic risk considerations into financial regulation and •
supervision,” SUERF Studies, 2010/4
87
The Capco Institute Journal of Financial TransformationSystemic Risk, an Empirical Approach
Possible Risk Indicators Indicator Information Selected Reference
Size Assets to world GDP (AtGDP) Total assets as a percentage of domestic GDP √ EU MoU CBFS
Deposits to world GDP (DtGDP) Market share in banking deposits and lending √ EU MoU CBFS
Share of credit to world credit (%CtGDP) Market share of credit √
Market share in payment transactions EU MoU CBFS
Degree of concentration of various markets (CONC) EU MoU CBFS
Number of transactions (NoT) Volume of transactions engaged PAPER
Assets under custody/total assets in custody (AuC) Volume of assets warehoused or managed √ ECB
Number of subsidiaries (NS) ECB
Balance Structure Balance quality
Leverage ratio (equity/total assets) Leverage position √ PAPER
Leverage ratio (including off balance) Leverage position real √ PAPER
Tier 1 (T1) Risk-based indicators √ PAPER
Core Tier 1 (CT1) Risk-based indicators √ PAPER
Off balance assets/total assets (OBAtA) Shadow risk proxy √ PAPER
Liquidity
Maturity mismatch PAPER
Short term liquidity indicators PAPER
Long term liquidity indicators PAPER
Deposits/assets (DtA) Capturing the funding decision of the bank √ ECB
Money market and short term funding/assets (MM&STFtA) Capturing the vulnerability to rate volatility, especially
relevant during liquidity dry up periods
√ PAPER
Contingent liabilities (CL) ECB
Asset mix
Gross loans/assets (LtA) Exposure to real cycle √ PAPER
Securities/assets (StA) Exposure to financial cycle √ ECB
Mortgages ECB
Cross-border assets (CBA) Exposure to foreign cycle ECB
Interbank assets (IA) ECB
Other assets (OA) ECB
Risk management culture
Short-term debt/total debt (STDtTD) Incentive alignment and risk franchising √ PAPER
Diversification degree PAPER
Non-interest income to total income (NIIR) Measure of income diversification (stable if ok) ECB
Net interest margin (%) Measure of the size of retail operations over overall
operations
√ ECB
Trade income (TI) Measure of Investment-type business ECB
Performance
Return on average assets (ROAA) (%) Capability to generate organically resources √ PAPER
Return on average equity (ROAE) (%) Capability to generate organically resources √ PAPER
Cost to income ratio (%) Measure of efficiency √ PAPER
Interconnectivity Eonia overnight lending contribution (EOLC) Interbank market weight ECB
Market capitalization (MC) Exposure of others to bank ECB
Subordinated debt issuance (SDI) Exposure of others to bank ECB
Interbank exposure/Tier 1 capital Exposure of others to bank √ EU MoU CBFS
Intra-group exposures EU MoU CBFS
Ranking in markets in which the institution is a significant player EU MoU CBFS
Share in transactions volume in markets in which the institution
participates
EU MoU CBFS
Credit spreads, bond spreads, and price to book value (level
and correlation)
EU MoU CBFS
Substitutability Worldwide custody (and clearing and compensation) (AuC) Infrastructure provider √ BoE
Sectoral breakdown of deposits and lending EU MoU CBFS
Volume interbank activity EU MoU CBFS
Volume of corresponding banking EU MoU CBFS
Hirschman Herfindahl Index Concentration levels in services BoE
ECB – The reference was found in ECB’s Financial Stability Report (2009)
BoE – The reference was found in BoE’s Financial Stability Review (2001)
PAPER – The reference was found in any of the papers used so far (as from SUERF, Squam Lake, etc.)
Table A1 – Candidate variables for systemic risk – all held in literature1
1 – Variables not selected (√) were due to availability reasons
88
Size Balance quality Asset quality
Risk management
culture Performance
Correlation matrix/Sig. (1-tailed) Ass
ets
to w
orl
d G
DP
(%)
Sha
re o
f cr
edit
to
wo
rld
cre
dit
(%)
Dep
osi
ts t
o G
DP
(%)
Ass
ets
und
er c
usto
dy
/to
tal a
sset
s in
cust
od
y (%
)
Tie
r1 (%
)
Co
re T
ier1
(%)
Off
bal
ance
ass
ets/
tota
l ass
ets
Leve
rag
e ra
tio
(sb
voln
ego
cio
) (%
)
Leve
rag
e ra
tio
(%)
Gro
ss lo
ans/
asse
ts (%
)
Sec
urit
ies/
asse
ts (%
)
Sho
rt t
erm
deb
t /t
ota
l deb
t (%
)
Net
inte
rest
mar
gin
(%)
Inte
rban
k ex
po
sure
(int
erb
ank
exp
)/ti
e
Ret
urn
on
aver
age
asse
ts (R
OA
A) (
%)
Ret
urn
on
aver
age
equi
ty (R
OA
E) (
%)
Co
st t
o in
com
e ra
tio
(%)
Siz
e
Assets to world GDP (%)100 61 48 33 -6 -36 -10 -28 -26 -24 34 -36 -40 3 -51 -26 24
0 1 3 5 3 2 0
Share of credit to world credit (%)61 100 63 -5 -40 -69 -40 -11 -26 48 -17 9 -28 29 48 -51 3
0 0 2 0 2 1 1 0
Deposits to GDP (%)48 63 100 -4 -18 -45 -28 -11 -23 23 8 7 -31 15 -35 -29 -2
1 0 1 4
Assets under custody/total assets in
custody (%)
33 -5 -4 100 7 34 43 -2 18 -38 15 1 4 -10 -14 -9 -4
4 1 3
Bal
ance
qua
lity
Tier 1 (%)-6 -40 -18 7 100 37 54 -27 2 -63 56 -27 8 -10 24 39 31
2 3 0 0 0 3
Core Tier 1 (%)-38 -69 -45 34 37 100 38 2 11 -45 14 -13 9 -7 22 34 0
3 0 1 4 3 3 1 5
Off balance assets/total assets-10 -40 -28 43 54 38 100 -7 45 -53 20 23 38 -18 31 27 -15
2 1 0 3 1 0 3
Leverage ratio (%)-28 -11 -11 -2 -27 2 -7 100 82 43 -76 13 54 -1 41 18 -41
0 1 0 0 2 2
Leverage ratio (%)-26 -26 -23 18 2 11 45 82 100 10 -57 31 74 -13 53 26 -44
1 0 0 0 0 1
Ass
et q
ualit
y
Gross loans/assets (%)-24 48 23 -38 -63 -45 -53 43 10 100 -69 27 9 28 -3 -32 -34
1 3 0 1 0 1 5
Sec/assets (%)34 -17 8 15 56 14 20 -76 -57 -69 100 -38 -34 -18 -15 15 47
5 0 0 0 0 3 5 1
Ris
k m
anag
emen
t
cult
ure
Short-term debt/total debt (%)-36 9 7 1 -27 -13 23 13 31 27 -38 100 41 -10 26 -14 -32
3 3 2
Net interest margin (%)-40 -28 -31 4 8 9 38 54 74 9 -34 41 100 -8 72 39 -49
2 3 0 0 5 2 0 2 1
Interbank exposure/tie3 29 15 -10 -10 -7 -18 -1 -13 28 -18 -10 -8 100 -39 -46 -18
2 1
Per
form
ance
Return on average assets (ROAA) (%)-51 -48 -35 -14 24 22 31 41 53 -3 -15 26 72 -39 100 83 -54
0 1 4 2 0 0 2 0 0
Return on average equity (ROAE) (%)-26 -51 -29 -9 39 34 27 18 26 -32 15 -14 39 -46 83 100 -37
0 3 5 2 1 0 3
Cost to income ratio (%)24 3 -2 -4 31 0 -15 -41 -44 -34 47 -32 -49 -18 -54 -37 100
2 1 5 1 1 0 3
Table A2 – Correlation matrix of systemic risk variablesBlue marked fields mean significant correlation between variables.
Black – white figures mean correlation degrees – significance levels
89
PART 1
Price of Risk – Recent Evidence from Large Financials1
AbstractProbability of default (PD) measures have been widely used
in estimating potential losses of, and contagion among, large
financial institutions. In a period of financial stress however,
the existing methods to compute PDs and generate loss es-
timates have generated results that vary significantly. This
paper discusses three issues that should be taken into ac-
count in using PD-based methodologies for loss or conta-
gion analyses: (i) the use of “risk-neutral probabilities” versus
“real-world probabilities;” (ii) the divergence between move-
ments in credit and equity markets during periods of finan-
cial stress; and (iii) the assumption of stochastic versus fixed
recovery for the assets of financial institutions. All three ele-
ments have non-trivial implications for providing an accurate
estimate of default probabilities and associated losses as
inputs for setting policies related to large banks in distress.
Manmohan Singh — Senior Economist, Monetary and Capital Markets Department, IMF
Karim Youssef — Economist, Strategy, Policy, and Review Department, IMF
The authors wish to thank Darrel Duffie, Kenneth Singleton, Inci Ötker-1
Robe, Andre Santos, Laura Kodres, Mohamed Norat, and Jiri Podpiera
for their helpful comments. The views expressed are those of the authors
and do not reflect those of the IMF. This paper has been authorized for
distribution by Martin Mühleisen.
90
Measures for the probability of default (PD) of financial institutions have
been widely used in estimating potential losses of, and contagion among,
large financial institutions.2 However, different methodologies used to ar-
rive at such estimates have not necessarily produced uniform results. Dur-
ing the recent financial crisis, two types of PDs (based on CDS spreads
and Moody’s KMV, respectively) have differed markedly for large banks,
and the resulting loss estimates have also varied significantly. In order to
properly identify policies with respect to large banks in distress, a closer
review of the key differences arising from the various methods to extract
PDs is necessary. Indeed, the difficulties in harmonizing the results of the
methodologies discussed need to be spelled out, as they could potentially
have an impact on authorities’ reactions and subsequent policy advice.
These differences start with the underlying market signals used to calcu-
late the PDs. Credit default swap (CDS) spreads providing signals from
debt and/or credit markets – given an assumed level of recovery – have
been used to arrive at a PD measure. By design, it is risk neutral because
it does not take into account investors’ varying degrees of risk aversion.
Risk neutrality allows us to bypass the need to calibrate a real world mea-
sure of investors’ utility by assuming that all investors are risk neutral.
That is to say, risk neutral methods assign greater probabilities to worse
outcomes. PDs derived via the risk neutrality assumption are widely ac-
cepted when pricing credit instruments, or assessing the impact of de-
fault risk on a portfolio of assets with similarly priced components.
The Moody’s KMV methodology, which accounts for investors’ risk aver-
sion by extracting signals from equity markets to arrive at a “real world”
measure of risk have also been used to extract PDs. In contrast to risk
neutral PDs, which use only market prices as inputs, risk measures based
on the real world approach also use balance sheet inputs. It is generally
accepted that real world measures provide for a better approximation
of investors’ risk aversion and are as such better suited to carrying out
scenario analysis to calculate potential future losses caused by defaults
[Hull (2009)]. Nevertheless, the nature of the inputs used for real world
measures also provide for the potential of missing important market sig-
nals (especially during distress).
The resulting implication is that losses computed from risk neutral PDs
may need to be adjusted downward to arrive at the real world probabili-
ties, while during periods of market stress, the assumptions underlying
some of the models yielding real world PDs may become tenuous. The
difficulties associated with the transformation of risk neutral PDs to real
world PDs are discussed below, along with issues that need to be con-
sidered and explored further. In particular, in adjusting the risk neutral
probabilities with a conversion factor (the price of risk), we explore the
importance of: (i) deviation between credit and equity prices during pe-
riods of financial market stress and (ii) the role of the assumption of sto-
chastic versus fixed recovery for financial institution assets.
Adjusting probabilities: the price of riskThe price of risk can be defined as the ratio needed to convert risk-neu-
tral probabilities (associated with CDS spreads) to real world probabili-
ties. The recent literature on this topic converges on the methodology of
Amato (2005), which proxies the conversion factor as follows: price of
risk = CDS spread/equity market signal. An example of an equity mar-
ket signal would be taking the Moody’s KMV expected default frequency
(EDF) as a real world measure. An enhancement to this would be to proxy
the conversion factor by also accounting for the recovery (R) expected
at default (40 percent is a common assumptions for R), that is to say,
adjusted price of risk = CDS spread/EDF(1-R).
The BIS Quarterly Report (March 2009) uses this approximation to show
that the price of risk during the 2007–08 period had fluctuated from an
average of about 4 to 12. In other words, risk neutral probabilities derived
from CDS spreads would need to be adjusted by a large and significant
factor to determine real-world probabilities. For example, if CDS spreads
were implying a PD of 0.9 percent, and the associated price of risk con-
version factor for a given corporate entity was 10, then the relevant ad-
justed PD would be 0.09 percent. The price of risk for large global banks
has indeed been sizable and varies across institutions. Our results sug-
gest that at the time of Lehman bankruptcy in September 2008, the price
of risk for many large banks was about 5, and for European banks, in
particular, higher than 10 in some cases (Figure 1).
There have been efforts to use Moody’s model to adjust real world prob-
abilities into risk neutral measures. This is mainly done via the use of the
Sharpe ratio and a correlation coefficient between individual returns and
market returns. However, it should be highlighted that this framework
assumes that investors treat financial and non-financial firms in a similar
fashion (even during the recent crisis). Additionally, in this framework, the
Sharpe ratio is updated only once a year, which presents an inconsis-
tency with most asset allocation models, especially during the distress
periods of 2008. The price of risk approach, which avoids these compli-
cations, may better reflect the transformation from risk neutral to the real
world probability of default.
Divergence between credit and equity markets during the recent crisisIn transforming risk neutral probabilities to real life probabilities, the im-
plicit assumption is the co-movement of equity and bond market signals
that drive EDF and CDS spreads, respectively. However, in the case of
large banks, the equity markets have been far more volatile than the bond
markets since 2008. In most cases, CDS spreads for the large banks have
Probability of default or distress is used here in a broader context, to include conditional 2
probabilities of default, joint probability of default, distance to distress, and joint default
dependence (i.e., via the off-diagonal elements of the distress dependence matrix).
91
The Capco Institute Journal of Financial TransformationPrice of Risk – Recent Evidence from Large Financials
Price of risk: American banks Price of risk: American banks
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Goldman Sachs JPMorgan Morgan Stanley
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Bank of America Citigroup Wells Fargo
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Barclays RBS Svenska
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
BBVA BNP Paribas Santander SocGen
0
1
2
Jul-08 Oct-08 Jan-09 Apr-09
Commerzbank Deutsche Bank UBS
Credit Suisse Danske Raiffeisen
0
5
10
15
20
Jul-08 Oct-08 Jan-09 Apr-09
Intesa KBC
Standard Chartered Erste
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Goldman Sachs JPMorgan Morgan Stanley
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Bank of America Citigroup Wells Fargo
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Barclays RBS Svenska
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
BBVA BNP Paribas Santander SocGen
0
1
2
Jul-08 Oct-08 Jan-09 Apr-09
Commerzbank Deutsche Bank UBS
Credit Suisse Danske Raiffeisen
0
5
10
15
20
Jul-08 Oct-08 Jan-09 Apr-09
Intesa KBC
Standard Chartered Erste
Price of risk: European banks Price of risk: European banks
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Goldman Sachs JPMorgan Morgan Stanley
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Bank of America Citigroup Wells Fargo
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Barclays RBS Svenska
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
BBVA BNP Paribas Santander SocGen
0
1
2
Jul-08 Oct-08 Jan-09 Apr-09
Commerzbank Deutsche Bank UBS
Credit Suisse Danske Raiffeisen
0
5
10
15
20
Jul-08 Oct-08 Jan-09 Apr-09
Intesa KBC
Standard Chartered Erste
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Goldman Sachs JPMorgan Morgan Stanley
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Bank of America Citigroup Wells Fargo
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Barclays RBS Svenska
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
BBVA BNP Paribas Santander SocGen
0
1
2
Jul-08 Oct-08 Jan-09 Apr-09
Commerzbank Deutsche Bank UBS
Credit Suisse Danske Raiffeisen
0
5
10
15
20
Jul-08 Oct-08 Jan-09 Apr-09
Intesa KBC
Standard Chartered Erste
Price of risk: European banks Price of risk: European banks
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Goldman Sachs JPMorgan Morgan Stanley
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Bank of America Citigroup Wells Fargo
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Barclays RBS Svenska
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
BBVA BNP Paribas Santander SocGen
0
1
2
Jul-08 Oct-08 Jan-09 Apr-09
Commerzbank Deutsche Bank UBS
Credit Suisse Danske Raiffeisen
0
5
10
15
20
Jul-08 Oct-08 Jan-09 Apr-09
Intesa KBC
Standard Chartered Erste
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Goldman Sachs JPMorgan Morgan Stanley
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Bank of America Citigroup Wells Fargo
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
Barclays RBS Svenska
0
1
2
3
4
5
Jul-08 Oct-08 Jan-09 Apr-09
BBVA BNP Paribas Santander SocGen
0
1
2
Jul-08 Oct-08 Jan-09 Apr-09
Commerzbank Deutsche Bank UBS
Credit Suisse Danske Raiffeisen
0
5
10
15
20
Jul-08 Oct-08 Jan-09 Apr-09
Intesa KBC
Standard Chartered Erste
Figure 1 – Price of risk
92
remained subdued given the perception of “too-large-to-fail.” Compared
with non-financial firms, such as GM or Chrysler, where bondholders
have recently taken a haircut and losses, bondholders of large complex
financial institutions have so far been kept whole. As a result, the varia-
tions in prospective returns have been reflected more immediately in the
equity markets, relative to the volatility of their bond prices (see Figure 2,
which is derived from Bloomberg’s OVCR function).3
The asymmetric signals from debt and equity markets, in turn, have
implications for estimating losses or guarantees from the implied bal-
ance sheet components. Disentangling the implications of this differen-
tial in volatility needs to be considered in probability models. Moreover,
the asymmetry in signals from credit and equity markets is important
to consider in models using distance-to-distress where debt and equity
market volatilities are important variables in determining the final results
(see Appendix II). Consequently, from a policy perspective, the estimates
of losses and guarantees need to be interpreted with caution when the
models do not account for dynamics of these relationships.
Assumption between fixed and variable recoveryModels estimating PDs have commonly assumed fixed recovery values.
However, stochastic recovery value assumptions may be necessary dur-
ing distress episodes. Unlike sovereigns or corporates, financial institu-
tions have few tangible assets, and recovery during the credit crisis was
very different from the 40 percent assumption (Lehman and Landsbanki
were roughly 8 cents and 1 cent on the dollar respectively, while Fannie
Mae and Freddie Mac were both above 90 cents on the dollar). Hence,
the use of a time-varying or stochastic recovery rate is all the more impor-
tant in the case of distressed financial institutions. In cases where cash
bonds trade below par, the cheapest-to-deliver (CTD) bond is a good
proxy for stochastic recovery [Duffie (1999); Singh (2003, 2004); Singh
and Spackman (2009)], as it reflects a more realistic inference of the value
of debt obligations than the fixed recovery assumption. Moreover, the
use of CTD is also in line with the physical settlement covenants of the
ISDA contracts.4 In the case of Iceland’s Landsbanki Bank, for example,
probabilities stemming from using a fixed recovery rate (green line) ver-
sus a stochastic recovery (blue line) proxied by cheapest-to-deliver bond
are markedly different (Figure 3).5
The use of PDs with a fixed recovery assumption has implications also
for assessing institutional interconnectedness through joint probability
of distress (JPoD). Inaccurate estimates of conditional probabilities may
result if the independent probabilities are biased. For example, if we are
to estimate the probability of default for Goldman conditional on Citi’s
default, incorrect estimation of the probability for Citi would contaminate
the conditional probabilities for Goldman.6
Bloomberg’s OVCR function (equity volatility and credit risk) converts equity prices, 3
leverage, and implied volatility to a CDS spread. This “theoretical” equity implied CDS
spread can be compared to actual CDS spread. The OVCR function is described in
Appendix 1.
In most models, including those using CDS and Moody’s EDF data, the general assumption 4
has been to hold recovery value constant (in the range of 20–40). The probability of default
(i.e., the hazard rate) and the recovery value more or less offset each other when bonds
trade near par. Such approximation works poorly when bonds trade at high spreads.
To further augment the use of stochastic recovery, the cheapest priced Citi and Goldman 5
bonds illustrate that their bond prices have traded well below par in the recent crisis,
despite the implicit forbearance offered to bondholders of large financial institutions, unlike
the bondholders of GM, Chrysler, or even Fannie Mae and Freddie Mac (Figure 4).
See IMF Working Paper No. 08/258 (page 14, second paragraph), which states: “using CDS 6
data after Lehman’s default will require the use of variable recovery value assumption, or in
its absence, CTD bonds.” There may be other factors such as funding costs during crisis
that can contribute to probability estimates.
Citibank U.S. Deutsche Bank
0
200
400
600
800
1000
1200
Aug
-08
Sep
-08
Oct
-08
Nov
-08
Dec
-08
Jan-
09
Feb
-09
Mar
-09
Ap
r-09
May
-09
Jun-
09
Model SD
Actual SD
0
100
200
300
400
500
Aug-08 Sep-08 Oct-08 Nov-08 Dec-08 Jan-09 Feb-09 Mar-09 Apr-09 May-09 Jun-09
Model SD
Actual SD
0
20
40
60
80
100
120
1/1/
2008
3/26
/200
8
7/24
/200
8
11/2
0/20
08
CTD bond price
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Probability of default (PoD)
CTD bond price (left scale)
PoD R=Stochastic (right scale)
PoD R=40% (right scale)
We use a distressed bond price “cut off” of below 95% of par.
Default
40
50
60
70
80
90
100
Jan-08 Apr-08 Jul-08 Oct-08 Jan-09 Apr-09 Jul-09
frac
tion
of p
ar
Citigroup
Goldman Sachs
Sources: Bloomberg and Moody’s KMV
Figure 2 – Volatility divergence in equity and credit markets for LCFIs
93
Conclusion and policy implicationsThis paper has argued that during periods of stress, measures for the
probability of default of financial institutions should be adjusted to reflect
the price of risk and address potential divergence of credit/equity mar-
ket relationships and the stochastic nature of asset recovery. Not taking
these elements into consideration may result in different, and perhaps
misguided, results and policy recommendations.
From a policy angle, until we have a more precise idea of the magnitude
of the biases and how best to revise the existing models, loss estimates
based on distance-to-distress models should be interpreted with caution
for large banks. Also, modeling the degree of interconnectedness of large
banks based on joint probabilities of distress should incorporate the low
recovery rates observed in the context of the recent credit events that
involved large financials, so as to avoid over- or under-estimation of the
degree of connectedness.
ReferencesAmato, J., 2005, “Risk aversion and risk premia in the CDS market,” Bank for International •
Settlements, Quarterly Review, December, 55–68
Amato, J., 2009, “Overview: investors ponder depth and duration of global downturn,” Bank for •
International Settlements, Quarterly Review, March, 1–28
Bloomberg, “Inferring default probabilities from capital structure information,” Version 1.0, •
equity valuation and credit risk function (OVCR).
Duffie, D., 1999, “Credit swap valuation,” Financial Analyst Journal, 55:1, 73–87•
Hull, J., 2009, “The credit crunch of 2007: what went wrong? Why? What lessons can be •
learned?” Journal of Credit Risk, 5:2, 3-18
Moody’s, 2007, “EDF™ 8.0 model enhancements.”•
Segoviano, B., A. Miguel, and M. Singh, 2008, “Counterparty risk in the over-the-counter •
derivatives market,” IMF Working Paper No. 08/258 (Washington: International Monetary Fund)
Singh, M., 2003, “Are credit default swap spreads high in emerging markets? An alternative •
methodology for proxying recovery value,” IMF Working Paper No. 03/242 (Washington:
International Monetary Fund)
Singh, M., 2004, “A new road to recovery,” RISK, September•
Singh, M. and C. Spackman, 2009, “The use (and abuse) of CDS spreads during distress,” IMF •
Working Paper No. 09/62 (Washington: International Monetary Fund)
The Capco Institute Journal of Financial TransformationPrice of Risk – Recent Evidence from Large Financials
0
200
400
600
800
1000
1200
Aug
-08
Sep
-08
Oct
-08
Nov
-08
Dec
-08
Jan-
09
Feb
-09
Mar
-09
Ap
r-09
May
-09
Jun-
09
Model SD
Actual SD
0
100
200
300
400
500
Aug-08 Sep-08 Oct-08 Nov-08 Dec-08 Jan-09 Feb-09 Mar-09 Apr-09 May-09 Jun-09
Model SD
Actual SD
0
20
40
60
80
100
120
1/1/
2008
3/26
/200
8
7/24
/200
8
11/2
0/20
08
CTD bond price
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Probability of default (PoD)
CTD bond price (left scale)
PoD R=Stochastic (right scale)
PoD R=40% (right scale)
We use a distressed bond price “cut off” of below 95% of par.
Default
40
50
60
70
80
90
100
Jan-08 Apr-08 Jul-08 Oct-08 Jan-09 Apr-09 Jul-09fr
actio
n of
par
Citigroup
Goldman Sachs
Sources: Bloomberg and Fund Staff estimates, and Singh and Spackman (2009)
Figure 3 – Landsbanki, Iceland – PDs from stochastic and fixed recovery assumptions
0
200
400
600
800
1000
1200
Aug
-08
Sep
-08
Oct
-08
Nov
-08
Dec
-08
Jan-
09
Feb
-09
Mar
-09
Ap
r-09
May
-09
Jun-
09
Model SD
Actual SD
0
100
200
300
400
500
Aug-08 Sep-08 Oct-08 Nov-08 Dec-08 Jan-09 Feb-09 Mar-09 Apr-09 May-09 Jun-09
Model SD
Actual SD
0
20
40
60
80
100
120
1/1/
2008
3/26
/200
8
7/24
/200
8
11/2
0/20
08
CTD bond price
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Probability of default (PoD)
CTD bond price (left scale)
PoD R=Stochastic (right scale)
PoD R=40% (right scale)
We use a distressed bond price “cut off” of below 95% of par.
Default
40
50
60
70
80
90
100
Jan-08 Apr-08 Jul-08 Oct-08 Jan-09 Apr-09 Jul-09
frac
tion
of p
ar
Citigroup
Goldman Sachs
Source: Bloomberg
Figure 4 – U.S. banks – Citi and Goldman’s bond prices
94
Appendix I – Bloomberg’s equity volatility and credit risk (OVCR) function7
Merton (1974) assumes that the value of a firm, A, follows a geometric
Brownian motion. Under constant interest rate r assumption, there exists
a risk-neutral measure Q under which we can write the dynamics of the
firm value as, dAt/At = rdt + σAdW (1)
Merton further assumes that the firm’s debt is in the form of a zero-cou-
pon debt with face value, D, and a single fixed expiry date T. At time T, the
firm pays off the debt if its firm value is higher than the face value of the
debt and claims bankruptcy if its firm value is below the face value of the
debt, A<D. The equity holder claims the remainder of the firm value.
Under these assumptions, the equity of the firm can be regarded as a
call option on the firm value with the strike equal the face value of the
debt and the maturity being the maturity of the debt. Since the firm value
follows a geometric Brownian motion, we can value the equity of the firm
based on the option pricing formula developed by Black and Scholes
(1973) and Merton (1973). The time-0 value of the firm’s equity, E0, can
be written as, E0 = A0N(d1) – De-rTN(d2) (2)
d1 = [lnA0/D + rT + 1/2σA2T] ÷ σA√T,
d2 = [lnA0/D + rT – 1/2σA2T] ÷ σA√T (3)
In particular, N(d2) represents the risk-neutral probability that the call op-
tion will finish in the money and hence the firm will not default. Therefore,
1-N(d2) denotes the risk-neutral probability of default.
In principal, we can use this pricing relation to estimate the default prob-
ability, if we are willing to accept the simplifying assumptions of Merton
(1974). In reality, there are several more assumptions and choices that we
need to make. The equity value can be observed from the stock market
if we regard the stock price as the equity value per share and ignore the
added value of warrants. The stock return volatility, σE, can also be esti-
mated either from the time series data or stock options. Furthermore, the
face value of the debt can be obtained from the balance sheet informa-
tion if we are willing to make further simplifying assumptions regarding
the debt structure and their maturities. Nevertheless, the firm value A0
and the firm volatility σA are normally regarded as not observable. These
two quantities can be solved from the following two equation,
E0 = A0N(d1) – De-rTN(d2), σE = N(d1) σAA0/E0 (4)
Where the second equation is a result of Ito’s lemma. These two equa-
tions contain the two unknowns (A0, σA) if we can obtain estimates on
(D, E0, σE, r, T). Therefore, one can solve for the two unknown quantities
using standard numerical nonlinear least square procedures.
Appendix II – The Moody’s KMV model8
The concept of using an option theoretic framework combined with stock
prices to estimate default risk was controversial, when first developed.
Consider a holding company that owns stock in another company and
that the market value of these holdings is V. Further, the company has a
debt payment of D due at a fixed point in time, T. Owning the equity of
such a holding company is equivalent to owning an option to buy the
stock at a price of D with an expiration date of T. Owning the debt is
equivalent to owning a risk-free bond that pays D at time T and being
short a put option on the stock with an exercise price of D and an expira-
tion date of T. In this example, the firm defaults if the value of assets, V, is
below D at the expiration date T. One can use the simple Black-Scholes
option formula to determine the value of equity. The four inputs to this
equation are the debt payment, D which we refer to as the default point,
the market value of the firm’s assets, V, the volatility of assets, A σ, and
the risk-free interest rate, r. The probability that the obligations will not be
met is a function of the firm’s DD, which represents the number of stan-
dard deviations that the firm is away from the default point. DD can also
be viewed as a volatility-adjusted market-based measure of leverage. As
the VK model is a barrier model, the model relates the asset value, the
default point, and volatility to the default probability via a first passage
through time formula. Vasicek has noted that the probability of default for
a first passage through time model is approximately equal to: 2Φ(−DD),
where DD is the so-called distance-to-default and Φ is the cumulative
normal distribution. Distance-to-default can be defined as:
DD(V, XT, σA,T, µ, δ) = [log(V/(XT + aT)) + (µ – 1/2σA2T] ÷ σA√T
Where V is the value of a firm’s assets, T X is the default point to the
horizon, µ is the drift term, A σ is the volatility of assets, T is the horizon,
and a represents cash leakages per unit time due to interest payments,
coupons, and dividends. The value of the firm’s assets and volatility is
computed as described above. The default point is computed as cur-
rent liabilities plus a portion of long-term debt. For longer horizons, a
larger portion of long-term debt is included in the default point to reflect
that long-term debt becomes more important at longer horizons. Note
that the DD varies considerably with the horizon under consideration. At
longer horizons, the weight on volatility increases relative to the default
point.
Empirically, there is a strong relationship between DD and the observed
default rates – firms with a larger DD are less likely to default. Never-
theless, the actual default rate found in the data differs from the literal
See Bloomberg’s “Inferring Default Probabilities from Capital Structure Information,” 7
version 1.0.
See Moody’s EDF™ 8.0 Model Enhancements.8
95
predictions of the model. Taken literally, the Brownian motion assumption
on asset value implies a Gaussian relationship between DD and the EDF
credit measure. Specifically, for a DD greater than 4, a Gaussian relation-
ship predicts that defaults will occur 6 out of 100,000 times. This would
lead to one half of actual firms being essentially risk-free. This implication
is not found in the data. Consequently, when implementing the model,
we depart from the Gaussian assumption by implementing an empirical
mapping.
The Capco Institute Journal of Financial TransformationPrice of Risk – Recent Evidence from Large Financials
Part 2Simulation and Performance Evaluation of Liability Driven Investment (LDI)
Behavioral Finance and Technical Analysis
The Failure of Financial Econometrics: Assessing the Cointegration “Revolution”
A General Structural Approach For Credit Modeling Under Stochastic Volatility
A Stochastic Programming Model to Minimize Volume Liquidity Risk in Commodity Trading
The Organization of Lending and the Use of Credit Scoring Techniques in Italian Banks
Measuring the Economic Gains of Mergers and Acquisitions: Is it Time for a Change?
Mobile Payments Go Viral: M-PESA in Kenya
99
PART 2
Simulation and Performance Evaluation of Liability Driven Investment (LDI)
AbstractIn contrast to decision models which use in-sample scenar-
ios we use out-of-sample scenarios to conduct simulation
and decision evaluation including backtesting. We propose
two simulation methodologies and describe six decision
evaluation techniques that can be applied to test the per-
formance of liability-driven investment (LDI) models. A good
practice of evaluating the performance of funds is to apply
risk-adjusted performance measures; we have chosen two
widely applied ratios: the Sortino ratio and the funding ratio.
We perform statistical tests and establish the quality of the
portfolios. We report our empirical findings by testing an as-
set and liability management model for pension funds where
we propose one deterministic linear programing and three
stochastic programming models.
Katharina Schwaiger — Financial Engineer, Markit
Gautam Mitra — Distinguished Professor and Director of CARISMA, Brunel University, and Chairman, OptiRisk Systems
100
The decisions suggested by the generic models proposed in Schwaiger
et al. (2010) are examined and evaluated in detail through simulation
studies. Two simulation methodologies and six evaluation techniques are
introduced and statistical methods and performance measures are used
to interpret these solutions. We start with a brief literature review of simu-
lation and evaluation methods applied to asset and liability management
(ALM) so far. The simulation and decision evaluation framework proposed
is generic and can also be applied to areas other than ALM.
We review briefly simulation methodologies used in portfolio planning
and ALM applied to pension funds, banks, and insurance companies.
A bank ALM model is introduced by Kusy and Ziemba (1986), where the
aim is to find the optimal trade off between risk, return, and liquidity of
the bank. A multiperiod stochastic linear programming model is intro-
duced and then compared with the Bradley-Crane stochastic decision
tree (SDT) model. The simulation study determines which model of the
two gives better first-period results by testing two hypotheses: whether
or not the ALM initial period profit is superior to the SDT and whether
the mean profit of the ALM is superior to the SDT. Oguzsoy and Güven
(1997) tested their bank ALM model in different ways. Firstly, they relaxed
their management policy constraints and compared its results with the
results of the fully included management policy constraints to show their
importance in avoiding risk. This also allows the determination of the
most important policy constraints. Secondly, they conducted sensitiv-
ity analysis on the rate of loans becoming irregular, different inflation ef-
fects, and on the rate of short term loans returning before maturity. And
finally they analyzed the occurrence probabilities of possible outstanding
deposit values. Zenios et al. (1998) developed a multi-stage stochastic
programing model with recourse for fixed-income portfolio management
under uncertainty. They show that their stochastic programing models
perform well using both market data (historical backtesting) and simulat-
ed data (using Monte Carlo simulation procedures) compared to portfolio
immunization models and single period models. They conclude that that
these results show robustness with respect to changes in the input data.
Mulvey et al. (2000) use stress testing in their pension plan and insurance
ALM model for Towers Perrin-Tillinghast with generated “devastating”
cases such as dropping equity markets and falling interest rates simul-
taneously. Once the desired scenario characteristic is set, the CAP:Link
system (scenario generator) filters the desired scenarios. Their stochastic
planning approach was compared to the current equity/bond mix of the
company, an equity hedged mix, and a long-bond strategy. Kouwenberg
(2001) developed a multi-stage stochastic programing ALM model for a
Dutch pension fund. The model is tested using rolling horizon simula-
tions and compared to a fixed mix model. The objective of the SP model
is to minimize the average contribution rate and penalizing any deficits
in the funding ratio. The rolling horizon simulations are conducted for
a period of five years, where the model is resolved using the optimal
decisions after each year. Then the average contribution rate and the
information about underfunding is saved after each year and compared
to the solution of the fixed mix model. The multistage simulation method
in our present paper is comparable to this method and explained in more
detail below. Their findings are that the SP model dominates the fixed mix
model and that the trade off between risk and costs are better in the SP
model than in the fixed mix model. Fleten et al. (2002) compare the per-
formance of a multistage stochastic programming insurance ALM model
and a static fixed mix insurance ALM model through in-sample and out-
of-sample testing. The stochastic programing model dominates the fixed
mix model, but the degree of domination decreases in the out-of-sample
testing. Only the first stage solutions are used to test the quality of the
models. The models are solved, the first stage decisions are fixed, and at
the next time period the models are solved again and the new first stage
decisions are used. Our multistage simulation methodology also draws
upon the algorithm described in their paper. More papers on decision
models and simulations for asset liability management can be found in
Mitra and Schwaiger (2011).
SimulationWe want to see how decision models perform and compare them in the
setting of a rolling forward simulation. An individual model can be tested
to exhibit its performance, but also a set of models can be compared
with each other for benchmarking. We then compare the models at each
time period by looking at the wealth (and/or funding status) distribution.
There are two decision making frameworks we can consider: (a) fixing the
first stage decisions and then only looking at the recourse actions, which
means in our models we decide on the set of bonds and the initial inject-
ed cash today and then only look at the reinvested spare cash and the
money borrowed from the bank in the future; or (b) we roll the model at
fixed time intervals of one year, which means in our models we decide on
the set of bonds and the initial injected cash today and then in the future
we might not only want to reinvest spare cash or borrow cash, but we
might also want to sell some bonds and/or buy new available bonds to
match the liabilities even better. In later time periods some bonds might
have even defaulted and we will not receive any cash flow payments from
it anymore. The two decision making frameworks are called two-stage
simulation and multistage simulation from now on.
The in-sample scenarios over which the decision models are initial-
ized over will be called optimization scenarios, while the out-of-sample
scenarios over which the decision models will be evaluated over will be
called simulation scenarios [introduced by Fleten et al. (2002)]. Fleten et
al. (2002) argue that a potential error source is if the scenarios used in
the simulation differ from the original scenarios of the optimization prob-
lem. The stochastic programing models adapt to the information in the
scenario tree during in-sample testing. If the out-of-sample scenarios dif-
fer too much from the in-sample scenarios, then this gained information
might lead to too bad or too good decisions. In our computational study
101
The Capco Institute Journal of Financial TransformationSimulation and Performance Evaluation of Liability Driven Investment (LDI)
the out-of-sample scenarios are generated using random sampling from
the original scenarios of the optimization problem. The distribution of the
out-of-sample scenarios will be the same as the distribution of the in-
sample scenarios.
The procedures for conducting a two-stage simulation and multistage
simulation is explained in the following two subsections. The main dif-
ference between the two is that in the two-stage simulation framework
the decision models using the optimization scenarios is solved only once
and the decisions gained from it is used within the simulation scenarios,
while during the multistage simulation the decision models are rerun at
subsequent times.
Methodology (1): two-stage simulationThe aim of the two-stage simulation is to solve the models at the initial
time period and see how they perform in the future under different (new)
scenarios. The steps of the two-stage simulation (Methodology (1)) are
as follows: (1) Generate the optimization scenarios; (2) solve the models
and fix the first stage decisions; (3) generate the simulation scenarios; (4)
use the fixed first stage decisions and run the model with the simulation
scenarios; and (5) compute the wealth (surplus and deficit) and mismatch
information at each time period
Figure 1 shows graphically the two-stage simulation, where the green
nodes and connecting arcs are the optimization scenarios and the red
nodes and arcs are the simulation scenarios. The models are solved at
the initial time point and their performance is measured under the new
simulation scenarios.
Methodology (2): multistage simulationThe second methodology is used to solve the models at the initial time
period but also at possible later time periods under different scenarios.
The steps of the multistage simulation [Methodology (2)] are as follows
[see also Fleten et al. (2002) and Kouwenberg (2001)]: (1) Generate the
optimization scenarios around the market expectations (i.e., the interest
rate scenarios (or mortality rates or salary curves) for periods 1 to T) and
solve the stochastic models and obtain the initial decisions at time t=1;
(2) generate a high number of sampled simulation scenarios around the
optimization scenarios for periods 1 to T; (3) for each simulation path
node, generate a conditional scenario tree; (4) if t<T, solve the models
at each simulation node; and (5) compute the current wealth (surplus or
deficit) and the new set of bonds, then go back to step 4.
In Figure 2, again the green lines represent the optimization scenarios
and the red lines represent the simulation scenarios. The generated opti-
mization scenarios after time period one are conditional on the simulation
node: the scenario generator takes all information until the current simu-
lation node into consideration. The multistage simulation applied to our
model is in a telescoping manner. The liabilities time period is fixed and
reduced after each year. The first portfolio optimization has a time horizon
from 1 to T, the year after the time horizon is 1 to T-1.
Decision evaluation and performance measurementThere are different types of decision evaluation techniques, where we use
six types: stress testing, backtesting, in-sample testing, out-of-sample
testing, scenario analysis, and what-if analysis. Stress testing is used to
test the models by using “extreme” data. Three cases can be considered:
how well will the models perform, if (a) there are low interest rates in the
future, i.e., which will mean the discounted present values of liabilities
are relatively larger and if (b) some bonds will default in the future, mak-
ing it impossible to match all the liabilities. Another stress case is (c) –
a possible decrease in mortality in the future. Pension funds are using
Figure 1 – Two-stage simulation tree
Figure 2 – Multistage simulation tree
102
current mortality tables, whereby they know that the mortality rates will
decrease in the future, making their liability stream longer and distributed
differently. Both simulation methods can be used for this type of decision
evaluation. For backtesting the models are run with past data and the
solutions show what happens if it had been decided to implement them
in the past. Using data from 2000 onwards is especially important, since
in that period pension funds were experiencing higher deficits. Figure 3
shows historical data of the LIBOR from 1997 till 2007.
Good results for that period will show the quality of the models. Only the
multistage simulation method is appropriate for backtesting, where the
model is rerun every year with the new knowledge of the world. Fleten et
al. (2002) showed that the discrepancy between the dynamic stochastic
approach and the fixed mix approach solutions is less in the out-of-sam-
ple testing than in the in-sample testing, this is due to the fact that the
stochastic model adapts to the information available in the scenario tree
in in-sample more than out-of-sample. By scenarios analysis we analyze
possible future events by considering alternative scenarios. Only the two-
stage simulation methodology is suitable for this analysis. The models
are solved for each scenario, the first stage decision are fixed and then
solved again with the other scenarios. In what-if analysis we project what
will happen under each scenario. Only the two-stage simulation method
can be used, whereby the optimization scenarios are the same as the
simulation scenarios.
The performance of any fund can be measured using different risk mea-
surements. The simplest one being the standard deviation. Fund manag-
ers look at the standard deviation of excess return over the risk-free rate
or some benchmark. The standard deviation of the difference between
the funds return and a benchmark return is also called the tracking error.
The higher the standard deviation, the higher the expected return for an
investment and the more risk bearing the investor. We will be using the
Sortino ratio and the funding ratio. The Sortino ratio roots from the Sharpe
ratio [Sharpe (1994)], but it is widely used across the industry since it only
penalizes a portfolios underperformance via downside deviation.
These two ratios are used substantially, we provide the formulae in sum-
mary form:
The Sortino ratio is calculated by: S ≡ RP – RI / σd (1)
where
1))(1(1
1=
−+≡ ∏ NPi
N
iP RR
2
1=,0))((1IPi
N
id RRmin
N−≡ ∑σ
The term σd is also called the downside deviation or target semidevia-
tion.
The funding ratio is the ratio of a pension funds assets to its liabilities; a
funding ratio of greater than one suggests that the pension fund is able
to cover all its obligations to its members. A funding ratio of less than
one indicates that the pension fund is currently unable to meet all its
obligations.
Ft = Assetst/Liabilitiest (2)
Computational investigationComputational results are given that highlight the methods of pension
fund performance evaluation: (a) backtesting using data from the last de-
cade, (b) stress testing using low interest rate scenarios, (c) stress testing
using low interest rates, and (d) stress testing including the possibility of
bonds defaulting. The other portfolio evaluation tests include the Sortino
ratio and the funding ratio, which are both widely understood as risk-
adjusted performance measurements in the industry by fund managers.
Table 1 shows the number of bonds available at each year within our con-
strained rating classes. Each year some bonds migrate to another rating
class, if they fall out of our given rating range, they cannot be bought
Figure 3 – Historical LIBOR data (12 month LIBOR)
Rating t=0 t=1 t=2 t=3 t=4 t=5 t=6 t=7 t=8 t=9
AAA 156 148 140 133 126 120 114 108 102 97
AA 88 89 90 90 90 89 89 88 87 85
A 72 75 79 82 85 88 90 92 94 96
BBB 60 61 61 62 63 63 64 65 65 66
Total 376 373 369 367 364 360 356 352 348 344
Table 1 – Number of bonds available after migration after t years
103
The Capco Institute Journal of Financial TransformationSimulation and Performance Evaluation of Liability Driven Investment (LDI)
a) Two-stage simulation results of the LP model b) Two-stage simulation results of the SP model
c) Two-stage simulation results of the CCP model d) Two-stage simulation results of the ICCP model
Figure 4 – What-if performance of the SP, CCP, and ICCP model from 1997-2006
into the portfolio anymore, i.e., if their rating is below BBB. The last row
shows the total number of bonds available for our portfolio. The data is
based on using average transition matrices from 1995-2005, where each
matrix gives the probability of staying in or changing the rating class.
The computational study is tested on the decision models proposed in
Schwaiger et al. (2010), where four models are proposed for a pension
fund ALM problem: a linear deterministic programming model minimiz-
ing PV01 deviations between assets and liabilities and minimizing initial
injected cash (from the members and the sponsoring company) and
three stochastic programing models incorporating uncertainty around
the interest rates that affect both assets and liabilities. The stochastic
programing models minimize present value deviations between assets
and liabilities and initial injected cash. The two-stage stochastic pro-
graming model is extended to include chance constraints and integrated
chance constraints. The models have a 45 year time horizon, 376 bonds
to choose from, and the stochastic models are initialized with 150 interest
rate scenarios.
104
The first set of results show the what-if analysis using the two-stage
simulation methodology: Figures 4a to 4d show the performance of the
liabilities and asset present values of the LP, SP, CCP, and ICCP model,
respectively. The present values of the assets and liabilities are stochas-
tic since they are discounted using the stochastic interest rates. Figure
4a shows the expected performance of the assets and liabilities of the
LP model, while Figures 4b, 4c, and 4d show the asset and liability per-
formance within each scenario. The LP model performs the worst, since
it is only useful for small shifts in the yield curve, while the tests allow
for large shifts in the yield curve. The SP model performs well on aver-
age, however it does not match the liabilities perfectly, although this is
one of the objective functions of the model. Looking at the performances
of the assets and liabilities within each scenario and plotting the differ-
ence a positive performance (surplus) can be seen in most time periods,
but also deficits in other time periods. In our tests, the SP model has
outperformed its liabilities, but there is no guarantee that this might not
turn into underperformance in some cases, since we are minimizing both
underdeviations and overdeviations at a same level. The seemingly bad
performance of the CCP model can be explained as follows: the chance
constraints are only included in the first three time periods, and it can be
seen that the assets match the liabilities perfectly until time period five.
The conclusion is to either set the chance constraints for more time pe-
riods or to set them for low time periods and rerun the model in between
again. The performance of the ICCP perfectly matches its liabilities with
the assets. Especially looking at Figure 4d, it can be seen that no mat-
ter what scenario occurs, the pension fund has a perfect asset/liability
match (suggested by the converging straight line at zero of A-L).
The next set of results are stress testing results using low interest rate
scenarios:
The surplus/deficit frequency distribution at four different time periods:
t=1 t=10
0
20
40
60
80
100
120
140 0
1714
14
3428
29
5142
44
6856
58
8570
73
1028
48
1199
90
1371
31
1542
73
1714
14
1885
56
2056
97
2228
39
2399
80
2571
22
2742
63
2914
04
3085
46
3256
87
3428
29
t=40
0
20
40
60
80
100
120
0
1300
18
2600
37
3900
55
5200
74
6500
92
7801
113
9101
29
1040
14
1170
16
1300
18
1430
20
1560
22
1690
24
1820
25
1950
27
2080
29
2210
31
2340
33
2470
35
2600
37
t=20
0
10
20
30
40
50
60
70
80
90
100
0
1180
81
2361
62
3542
44
4723
25
5904
07
7084
88
8265
70
9446
51
1062
73
1180
81
1298
89
1416
97
1535
05
1653
14
1771
22
1889
30
2007
38
2125
46
2243
54
2361
62
t=10
0
20
40
60
80
100
120
0
1304
04
2608
08
3912
12
5216
16
6520
20
7824
24
9128
29
1043
23
1173
63
1304
04
1434
44
1564
84
1695
25
1825
65
1956
06
2086
46
2216
87
2347
27
2477
67
2608
08
t=2
t=20 t=40
0
20
40
60
80
100
120
140 0
1714
14
3428
29
5142
44
6856
58
8570
73
1028
48
1199
90
1371
31
1542
73
1714
14
1885
56
2056
97
2228
39
2399
80
2571
22
2742
63
2914
04
3085
46
3256
87
3428
29
t=40
0
20
40
60
80
100
120
0
1300
18
2600
37
3900
55
5200
74
6500
92
7801
113
9101
29
1040
14
1170
16
1300
18
1430
20
1560
22
1690
24
1820
25
1950
27
2080
29
2210
31
2340
33
2470
35
2600
37
t=20
0
10
20
30
40
50
60
70
80
90
100
0
1180
81
2361
62
3542
44
4723
25
5904
07
7084
88
8265
70
9446
51
1062
73
1180
81
1298
89
1416
97
1535
05
1653
14
1771
22
1889
30
2007
38
2125
46
2243
54
2361
62
t=10
0
20
40
60
80
100
120
0
1304
04
2608
08
3912
12
5216
16
6520
20
7824
24
9128
29
1043
23
1173
63
1304
04
1434
44
1564
84
1695
25
1825
65
1956
06
2086
46
2216
87
2347
27
2477
67
2608
08
t=2
Figure 5 – SP model stress testing low interest rate surplus/deficit distribution
105
t=2, t=10, t=20 and t=40 and for all four models is calculated. The re-
sults in Figure 5 show the surplus/deficit frequency distribution of the SP
model using low interest rate scenarios. The SP model performs relatively
well (in matching terms) compared to the other models and in both cases
of high and low interest rate scenarios.
The next results of the surplus/deficit frequency distribution of the CCP
model using low interest rate scenarios show that the CCP model gener-
ates a good asset/liability matching outcome, with surplus/deficit being
around zero over the four time periods. It can match the asset and liability
present value deviations no matter what happens to the interest rates,
this is due to the restriction of underfunding events which was incorpo-
rated into the model.
In low interest rate times the ICCP model guarantees fairly well a perfect
matching. In low interest rate events, the discounted liabilities are higher
than in high interest events; when liabilities are higher valued the ICCP
model protects the pension fund from interest rate risk and generating a
deficit.
As expected the LP model underperforms far more than all the other
models in low interest rate scenarios since it is only a static decision
model and does not take into account any future uncertainty.
Considering the performance of the funds low interest rate scenarios
confirm our previous expectations: the SP model even performs well dur-
ing low interest rates generating only low deficits at a few time periods.
Due to a high reliability level the CCP only generates a surplus, which is
relatively high. The ICCP model does generate a deficit at some time pe-
riods, but it is restricted and it best matches the assets with the liabilities.
The LP model generates large deficits in the long term. The same study
has been conducted using extreme events of high interest rate scenarios
and the outcomes in summary are: the ICCP model never generates a
deficit and outperforms the other models in terms of matching. The LP
model performs far better during high interest rate times, but still does
not outperform any of the stochastic programming models.
Performance measurementAgain we wish to look at the funding ratio and at the Sortino ratio of the
SP, CCP, and ICCP models during low interest rate scenarios. This is plot-
ted in Figure 6. As expected the ICCP model has a funding ratio of one,
which suggests a perfect matching of assets and liabilities at all time pe-
riods. The SP model has a desired funding ratio of one in the low interest
rate scenarios. In the CCP model the assets outperform the liabilities dur-
ing low interest rates. Looking now at the Sortino ratio, the CCP model
has an upward sloping curve for low interest rate scenarios. The SP has
a Sortino ratio close to zero during low interest rates, which means again
a close PV match of the assets and liabilities. The ICCP model is close to
zero during low interest rates.
ConclusionsWe introduced a framework for two simulation methodologies: two-stage
simulation and multistage simulation. We have also applied six differ-
ent decision evaluation techniques to test the results of four pension
fund ALM decision models. We report the results of the empirical study
which test the models proposed in Schwaiger et al. (2010). The inter-
est rate scenarios are generated using the CIR model and the liability
stream is generated using pension mathematics practice, U.K. mortality
tables, and inflation rate scenarios. Computational results which are re-
ported highlight the methods of pension fund performance evaluation: (a)
t=20 t=40
-10000000
0
10000000
20000000
30000000
40000000
50000000
60000000
70000000
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43
CCP
ICCP
SP
-0,6
-0,4
-0,2
0
0,2
0,4
0,6
0,8
1
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43
CCP
ICCP
SP
-10000000
0
10000000
20000000
30000000
40000000
50000000
60000000
70000000
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43
CCP
ICCP
SP
-0,6
-0,4
-0,2
0
0,2
0,4
0,6
0,8
1
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43
CCP
ICCP
SP
Figure 6 – Funding and Sortino ratios for stress testing (low interest rates)
The Capco Institute Journal of Financial TransformationSimulation and Performance Evaluation of Liability Driven Investment (LDI)
106
backtesting using data from the last decade, (b) stress testing using low
interest rate scenarios, (c) stress testing using high interest rates, and (d)
stress tests including the possibility of bonds defaulting. The other port-
folio evaluation tests include the use of the Sortino ratio and the funding
ratio, which are both widely understood as risk-adjusted performance
measurements in the industry by fund managers.
Based on the empirical study we can make the following summary obser-
vations: (i) the stochastic programing models dominate the deterministic
programing models when interest rate risk is considered, (ii) the chance
constrained programing model and integrated chance constraint pro-
graming model need a higher initial cash injection to limit possible future
deficit events, (iii) the integrated chance constrained programing model
does not lead to a deficit during backtesting and stress testing while at
the same time it does not lead the pension fund to severe overfunding,
and (iv) from a computational performance perspective the ICCP model
has similar CPU solving times to the SP model, while the CCP model
leads to expensive CPU times.
Pension fund managers should adopt stochastic programming models to
hedge against changes in the present values of the liabilities. Although,
the SP models are more costly to implement, they guarantee the model
to be less exposed to a deficit. From a computational perspective fund
managers can easily run the LP, SP, and ICCP models with a large asset
universe and time horizon within a reasonable time.
ReferencesFleten, S. E., K. Hoyland, and S. W. Wallace, 2002, “The performance of stochastic dynamic •
and fixed mix portfolio models,” European Journal of Operational Research, 140:1, 37-49
Kouwenberg, R., 2001, “Scenario generation and stochastic programming models for asset •
liability management,” European Journal of Operational Research, 134:2, 279-292
Kusy, M. I. and W. T. Ziemba, 1986, “A bank asset and liability management model,” Operations •
Research, 34:3, 356-376
Mitra, G. and K. J. Schwaiger, (eds.), 2011, Asset and liability management handbook, Palgrave•
Mulvey, J. M., G. Gould, and C. Morgan, 2000, “An asset and liability management system for •
Towers Perrin-Tillinghast,” Interfaces, 30:1, 96-114
Oguzsoy, C. B. and S. Güven, 1997, “Bank asset and liability management under uncertainty,” •
European Journal of Operational Research, 102:3, 575-600
Schwaiger, K. J., 2009, Asset and liability management under uncertainty: models for decision •
making and evaluation, PhD thesis, CARISMA, Brunel Univeristy
Schwaiger, K. J., C. Lucas, and G. Mitra, 2010, “Alternative decision models for liability •
determined investment,” Journal of Asset Management: special issue on ALM
Zenios, S. A., M. R. Holmer, R. McKendall, and C. Vassiadou-Zeniou, 1998, “Dynamic models •
for fixed-income portfolio management under uncertainty,” Journal of Economic Dynamics and
Control, 22:10, 1517-1541
107
PART 2
Behavioral Finance and Technical Analysis
AbstractBehavioural finance has challenged many claims of efficient
market hypothesis (EMH). Unfortunately many of these chal-
lenges are in the form of anecdotal evidence and lack quan-
tification. This article uses market data together with some
simple statistics to show that in practice certain assertions
of EMH and mathematical finance can be rejected with a
high degree of confidence. The working of the FX market is
used to demonstrate certain shortcomings of elegant results
in mathematical finance that render them irrelevant in prac-
tice. An approach based on Markov chains is developed to
model certain heuristic notions such as “fast market,” “sup-
port,” and “resistance,” that are widely used by “technical
analysts” and practitioners. Using market observation, it is
shown that this model better fits historical data than that im-
plied by the assumption that daily returns are independent
and normally distributed.
Kosrow Dehnad — IEOR Department, Columbia University, Quantitative Trading, Samba Financial Group
108
Behavioral finance started as an irritating challenge to advocates of ef-
ficient market hypothesis (EMH) and was initially ignored by them. It ex-
posed instances of investors’ behavior that contradicted the main tenets
of EMH which assumes that investors are rational decision makers and
that at each instant the prices of securities reflect all available informa-
tion about them. This implies that all the efforts by security analysts and
traders to beat the market are futile and that investors should simply con-
struct a portfolio consisting of a combination of risk-free assets and the
“market” portfolio which best suits their risk appetite and lay back and
enjoy the fruits of their investments. The EMH found an ally in the math-
ematical finance community where the above assumptions implied that
prices were a martingale process, thus enabling the researchers in that
field to use the techniques of Brownian motion to prove many elegant
theorems. Further, by blending some supposedly “minor” simplifying as-
sumptions with their work, such as absence of bid-ask spread or taxes,
or homogeneity of time, or the ability of investors to lend and borrow
at the same rate and as much as they desire, theoreticians opened the
floodgates of dynamic hedging and option replication that has become
an industry in itself.
There has, however, been growing evidence contradicting EMH such
as: performance of investors such as Warren Buffet, George Soros, and
Peter Lynch who have beaten the market year after year; the mini crash
of 1987; the great recession of 2008, and the flash crash of 2010. More-
over, studies in behavioral finance have shown that certain hardwired
biases and habits of humans make them a far cry from the rational and
calculating decision makers who pounce on any opportunity to maxi-
mize their profit. As such anecdotal evidence have accumulated, ex-
ponents of EMH had to reluctantly modify their stances by introducing
different flavors of efficiency, such as strong form efficiency, weak form
efficiency, etc., and in all cases the statement of “all available informa-
tion about a stock” is left as a nebulous concept. It should be pointed
out that certain market practitioners, the so-called technical analyst or
chartists, never paid any attention to the results of mathematical finance
and EMH that are against the very grain of their work. These practitio-
ners believe that certain price patterns repeat themselves and provide
profit opportunities. Consequently, they pore over historical data and
draw charts with acronyms such as “support,” “resistance,” “channel,”
“head-and-shoulder,” and “momentum,” which according to EMH have
no informational value whatsoever, in order to gain insight into market
sentiment that hopefully will give them a trading edge. Justification of
chartists for their approach to market is very intuitive and suffers from
a lack of quantification though they use certain statistical terms such
as moving averages or ratios. To justify some of these approaches psy-
chologists have joined the foray and tried to provide an explanation
for the way market behaves using psychology. Figure 1 is a cognitive
psychology explanation of oscillation of prices that fall into a rising or
falling band called a “channel.”
Most market observers are familiar with statements such as, “the market
showed a number of abrupt rises interrupted by sideways movement in
the congestion area, previous buyers selling to take home profits, and
the new buyers taking advantage of an opportunity to get in. Sellers, also
participating all the way up, each time noted that the market reached new
higher peaks and that they should have stayed firm. Small drops were
therefore used to come back in, and each increase provoked new buy in-
terests.” Such descriptions that often lack hard data and statistical analy-
sis to support them are in part responsible for the fact that the EMH camp
dismiss some of the findings of behavioral finance by statements such
as, “are these the tip of the iceberg or the whole iceberg?” Because in
the absence of any statistical analysis, such descriptions of the causes of
market behavior remain an interesting read at best. This article attempts
to use statistical methods and market data to demonstrate that continu-
ous time models based on Brownian motion disregard some of the basic
characteristics of some markets and the behavior of their participants.
This inattention has major practical implications and renders some of the
results in mathematical finance a theoretical construct at best.
Homogeneity of timeOne of the assumptions of mathematical finance is homogeneity of time.
Namely, the evolution of asset prices in one hour is assumed to be in-
dependent of whether this one hour is from 5 to 6 pm New York time
when all major financial markets are closed, or from 8 to 9 of the first
Friday of the month when important U.S. economic data are released
that can considerably move the markets. Practitioners pay special atten-
tion to this difference and market data also support this distinction. For-
mally, independent increment assumption implies that price movements
in the interval [t, t + Δ] is a function of Δ only and do not depend on “t.”
This premise makes Brownian motion tools and techniques applicable in
0
5
10
15
20
25
30
1 2 3 4 5 6 7 8 9 10 11 12 13 14
Price
Time
Lower band
Higher band
Price
Figure 1 – Channel formation
109
The Capco Institute Journal of Financial TransformationBehavioral Finance and Technical Analysis
modeling price movements. We propose the following statistics for test-
ing this hypothesis. Let Q[t, t + Δ] be a certain observed quantity during
time interval [t, t + Δ], for example the volume of trades on a certain stock,
or its realized volatility, etc. For a given day i and two different periods [ti,
ti + Δ] and [t’i, t’i + Δ] let Qi = Q[t, t + Δ] and Q’i = Q[t’i, t’i + Δ].
If time is homogeneous, then Pr[Qi ≥ Q’i] = Pr[Qi ≤ Q’i] = ½. For example, if
Q represents trade volume, the above hypothesis implies that it is equally
likely for the volume from 9:30 am when the market opens till 12:30 to be
more or less the same as that from 1 till to 4 pm when the market closes.
Let us define Xi = 1 if [Qi ≥ Q’i]; Xi = 0 otherwise.
Under the assumption of homogeneity of time, Xi’s are iid observations
with Pr[Xi = 1] =1/2. And for a sufficiently large number of observation “n”
the statistics S = (2∑ Xi – n)/ √n will have a standard normal distribution.
This result can be used to test the hypothesis of homogeneity of time.
For illustration, consider the volume of trades in AT&T (T) shares between
9:30 am until 12:30, Qi=Volume[9:30,12:30] and from 1 till 4:00, Q’i = Vol-
ume[1, 4], for business days from Monday, October 4th, 2010 until Friday,
April 29th, 2011. During this time less than 12 percent of time the volume
of the shares traded in the morning was larger than that in the afternoon.
The statistics S= 9.03 that implies the null hypothesis of homogeneity of
time can be rejected with p value well below 0.001.
Time varying nature of volatility and its practical implicationsUnlike equity markets, foreign exchange (FX) markets are theoretically
open 24 hours a day, 7 days a week and there is no central exchange for
them. They, however, exhibit some of the same characteristics of equity
markets with regards to volume, liquidity, and bid/ask spread. For illustra-
tion consider the Japanese yen market (JPY). The general course of daily
trading in JPY is along the following lines: there is very little liquidity and
trading activity until the Tokyo market opens and price action begins. The
liquidity and price action markedly increase with the opening of London
and reach their peak when New York opens and its traders also join the
action. Liquidity starts to taper off with the closing of Tokyo, followed
by London and New York. With the closing of New York the price action
becomes minimal, particularly if it is the last working day of the week or
the start of a major holiday such as New Year’s day. And for all practical
purposes liquidity and volatility die down until Tokyo opens again and the
cycle repeats itself. All practitioners are acutely aware of the above facts
and adjust their trading accordingly. In other words, bid-ask spreads can
be quite wide when the market lacks liquidity and it drops to one “tic”
when the market is most active. This implies that trading time must not
be treated as homogeneous and market behavior and volatility during the
interval [t, t + Δt] do indeed depend on “t.”
Of course, the importance of bid spread, liquidity, and volatility in trading
depends on the investment horizon and trading style. In the case of the
so-called “global macro” strategies, these factors are of secondary im-
portance since such strategies try to capture the long-term market trends
and ignore day to day price variations. On the other hand, in daily trading
and market making the important issues could be the levels of “stop loss”
(SL) and “take profit” (TP) and where such trigger levels should be placed
relative to those of other market participants and whether these levels are
bunched together creating “support” or “resistance” levels. The practical
implication of these points is that a wide and supposedly safe stop loss
might be triggered over night because of low liquidity and high bid-ask
spread. The trigger could also be caused by sudden spikes and drops
in price because of temporary imbalances in supply and demand due to
say, arrival of unexpected news or large orders. Continuous time finance
assumes all price movements, over a short period of time, to be infini-
tesimal and uses an annual volatility number to describe the distribution
of these movements. In practice, arrival of a large order or a new piece
of information could cause the market to spike and “gap” even for assets
with a deep and liquid market such as JPY/USD. In particular, a sudden
spike could trigger stop losses that in turn trigger additional stop loss
orders that temporarily result in a destabilizing positive feedback system.
For example, the continuous time finance is incapable of explaining the
sudden drop of 1000 point in the Dow index in a few minutes and its
subsequent recovery in a few hours on May 6th 2010 because of some
computer orders by “fat fingers.”
These sudden movements when translated into an annual volatility can
result in unrealistically high numbers such as 20000 percent! To model
such phenomenon or the fact the volatility changes overtime a number
of techniques and models have been proposed from jump processes to
ARCH and GARCH models. These models, which are mathematically el-
egant and sophisticated, require estimation of a number of parameters
and pose the daunting task of selecting among a number of competing
models that all fit the data “reasonably” well. Some argue that the vola-
tilities used in option pricing models should be viewed as an “estimate”
of the “average” “long run” variability of the underlying parameters. In
practice, however, market participants must adjust their behavior and ac-
tions based on what they perceive to be those of other participants and
balance the elegance of purely mathematical theories against relevance
of market realities such as risk limits.
Let us not forget that after “dropping a lot of money” and getting fired for
it, there is rarely a chance to prove to the management that, in the “long
run” and “on average,” our losing trades would have been “profitable” if
only it had been given enough time for the market to turn!
Assumption of normal returnsAnother simplifying assumption of continuous time models that mani-
festly does not hold in practice is that returns are normally distributed
110
even on infinitesimal scale, i.e., price movements during a day. Assets
with sufficient liquidity move in small and quantum increments, say one
tic. For example, in the case of USD/JPY at each instant the price can
either go up or down by that quantum amount of one tic irrespective of
whether JPY is trading at 90.50 or 95.50. Consequently, on infinitesimal
scale the returns cannot be normally distributed since the size of each
move, i.e., one tic, is independent of the price level. On the other hand,
the assumption of normal returns on even micro levels that renders such
processes infinitely divisible is essential in order to use powerful tools and
techniques of Brownian motion and derive elegant mathematical results.
Further, there is always a bid-ask spread associated with each trade and
this fact is also conveniently glossed over in some mathematical results.
The bid-ask spread and minimum size of a move have major practical
implications and practitioners are not oblivious to this fact. For example,
in dynamic hedging of an option position if rebalancing is followed blindly
the way that it is prescribed by theory, it will take the option trader to the
poor house and will render dynamic hedging impractical. Practitioners
are not oblivious to this fact and refrain from rebalancing their hedges
for every minor move of the markets. In practice, after placing the ini-
tial hedges on – the so-called delta hedging the position – subsequent
rebalancing is done less frequently and on a portfolio basis in order to
benefit from portfolio effect and save the bid-ask spread. Assuming zero
bid-ask spread is like driving under the assumption that a car can stop
immediately when the brake is pressed. It might be a good approximation
but in never holds in practice. This assumption also implies that liquidity
providers do this as a public service or an act of charity!
Another implication of the assumption of normality of return is the pos-
sibility of, though highly improbable, very large daily price moves. In
FX markets, daily exchange rates have some rather hard lower bounds
based on the size of economies and purchasing power parity which make
such large movements impossible. For example, if JPY/USD exchange
rate suddenly falls to 20 from 85 it implies that in a short period Japanese
economy becomes multiple of that of the U.S. Though this is a theoretical
possibility under the assumption of normal returns of assets, in practice it
should be treated as a fantasy.
Markov chain modelAn approach that better reflects the daily behavior of FX markets is to
divide the trading day into intervals with different intensity of arrival of
new trades. This provides the flexibility to model “fast markets” or “slow
markets,” etc. This division of time can be based on historical data and
can accommodate trader’s view of market liquidity by assigning a bid-
ask spread to each segment. For example, from close of New York to 7
am Tokyo time the bid-ask spread can be eight tics with low intensity of
new trades while from 9:00 to 9:15 when economic data is released and
trading is very heavy the bid ask spread is only one tic. In this model if pi
is the price at instant I, then with probability half the next trade is either at
pi + 0.5*(bid-ask) or pi - 0.5*(bid-ask). The support and resistance levels
can be modeled as follows. We assume a certain concentration of buy
orders, say 20, at a certain level, say 90.50. Note the buy order refers to
buying USD, the base currency, and selling JPY, the term currency. This
concentration of orders creates a support level for USD at 95.50. This
means that if the next trade is a sell order, the price rather than falling to
95.49 will stay at 95.50. However, buy concentration is reduced by 1 unit
to 19. If the next trade is again a sell order the concentration is further
reduced to 18 and this will continue until all buy orders are taken out and
in that case the next sell order will move the market down to 90.49. By
adjusting the probability of up or down after the support is breached one
can model the sharp drop in prices that usually follow at this time and the
next support is where the next buy orders are bunched together. Most of
the technical analysis jargon and concepts can be modeled by changing
the transition probability and resistance of support and concentration of
buy or sell orders, i.e., stop loss or take profit orders.
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,001 0,101 0,201 0,301 0,401 0,501 0,601 0,701 0,801 0,901
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,001 0,101 0,201 0,301 0,401 0,501 0,601 0,701 0,801 0,901
Figure 2 – Q-Q plot of U[0,1] versus actual daily JPY (close-low)/(high-low)
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,001 0,101 0,201 0,301 0,401 0,501 0,601 0,701 0,801 0,901
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,001 0,101 0,201 0,301 0,401 0,501 0,601 0,701 0,801 0,901
Figure 3 – Q-Q Plot of U[0,1] versus simulated values (close-low)/(high-low)
111
The Capco Institute Journal of Financial TransformationBehavioral Finance and Technical Analysis
Market observationsIn this section, we use a simple test based on market observation to
determine which of the two models, i.e., Brownian motion or the Markov
chain model described in the previous section is a better representation
of daily USD/JPY market. Let us consider the daily close of JPY relative
to high and low of that day. If we let low = 0 and high = 1, the ratio (close-
low)/(high-low) represents the relative position of closing price in the in-
terval [0,1]. The Markov chain model of price movements implies that the
above number should have uniform distribution U[0,1]. This follows from
the observation that all states of the Markov chains are accessible and
recurrent, hence the asymptotic distribution of the states, i.e., that the
closing price is uniform. This ratio has been calculated for the past ten
years. Figure 2 is the Q-Q plot of the empirical distribution of these ratios
against that of U[0,1] – implied by Markov chain model. Figure 3 shows
a similar plot where simulation is used to generate the data assuming a
lognormal distribution of prices. It is clear that U[0,1] is a better fit for the
distribution of the above ratios, hence, the Markov chain model seems to
be a better model of price movement than continuous time models based
on normal distribution of returns.
ConclusionIn daily trading of very liquid assets, such as JPY, one has to be cognizant
of notions such as “market sentiments,” “support,” “resistance,” etc.
Many of the mathematical models applied in finance ignore these notions
and dismiss them despite the fact that all practitioners are acutely aware
of them. Continuous time finance also makes certain simplifying assump-
tions that greatly reduce the applicability of its results to daily trading.
This article demonstrates some of these issues and proposes a modified
version of the Markov chain that will enable practitioners to model, in a
consistent manner, concepts that are used in “technical analysis.” It also
remedies one of the major shortcomings of technical analysis which is
its inability to assign probabilities to various outcomes – a fundamental
ingredient of successful trading.
ReferencesChung, K. L., 1967, Markov chains, Springer Verlag•
Freedman, D., 1971, Markov chains, Holden Day•
Kahnerman, D., P. Slovic, and A. Tversky (eds.), 1982, Judgement under uncertainty: heuristics •
and biases, Cambridge University Press Thaler, R. H., 1993, Advances in behavioural finance,
Russell Sage Foundation Publications
Thaler, R. H., 1994, The winner’s curse: paradoxes and anomalies of economic life, Princeton •
University Press
113
PART 2
The Failure of Financial Econometrics: Assessing the Cointegration “Revolution”
AbstractOne aspect of the failure of financial econometrics is the use
of cointegration analysis for financial decision making and
policy analysis. This paper demonstrates that the results
obtained by using different cointegration tests vary consid-
erably and that they are not robust with respect to model
specification. It is also demonstrated that, contrary to what
is claimed, cointegration analysis does not allow distinction
between spurious relations and genuine ones. Some of the
pillars of cointegration analysis are not supported by the
results presented in this study. Specifically it is shown that
cointegration does not necessarily imply, or is implied by,
a valid error correction representation and that causality is
not necessarily present in at least one direction. More im-
portantly, however, cointegration analysis does not lead to
sound financial decisions, and a better job can be done by
using simple correlation analysis.
Imad Moosa — Professor of Finance, School of Economics, Finance and Marketing, RMIT1
I am grateful to the editor of this journal for encouraging me to write this 1
paper and to Kelly Burns for useful comments.
114
In the second half of the 1980s, specifically following the publication of
the “seminal” paper of Engle and Granger (1987), the world of academic
finance and economics experienced a “revolution” similar to that expe-
rienced by the worlds of music and dancing as a result of the introduc-
tion of rock and roll and the twist. Engle and Granger formalized their
work on cointegration and error correction and subsequently adapted
the causality test of Granger (1969) to take into account the possibility of
cointegration. The introduction of these techniques has created a thriving
industry with a rapidly growing output of papers written by academics
testing theories in economics and finance that were previously tested
using straightforward regression analysis.
Tens of thousands of papers and PhDs later, it is about time to ask wheth-
er or not the cointegration “revolution” has changed our lives and led to
discoveries that enhance our understanding of the working of the econo-
my and financial markets, which is presumably the objective of scientific
research. One would tend to imagine that, since this work was awarded
the Nobel Prize, it must be valued the same way as the discovery of
penicillin, which was awarded the same prize. However, it seems to me
that while cointegration analysis has provided the means for finance and
economics academics to get their promotion and students to obtain their
PhDs, the technique has contributed almost nothing to the advancement
of knowledge.
The objective of this paper is to demonstrate that cointegration analysis,
error correction modeling, and causality testing are misleading, confusing,
and provide a tool for proving preconceived ideas and beliefs. More impor-
tant, however, is the hazardous practice of using the results of cointegra-
tion analysis to guide policy and financial operations, including investment,
financing, and hedging. With the help of examples on stock market integra-
tion and international parity conditions it will be demonstrated that cointe-
gration analysis produces results that tell us nothing and that for practical
purposes these results are useless at best and dangerous at worst.
Cointegration, error correction, and causalityCointegration implies that a linear combination of two (or more) variables
is stationary although the variables themselves are nonstationary in the
sense that they tend to wander around over time. When variables are
cointegrated, they are said to be tied up by long-run equilibrium relations.
This means that while it is possible to deviate from the long-run condition
in the short run, these deviations tend to disappear with the passage of
time as a result of the tendency to move back to the equilibrium condition
(the phenomenon of mean reversion).
A simple two-variable cointegrating regression (normally includes a con-
stant term) may be written as yt = α + βxt + εt (1). For xt and yt to be
cointegrated, the necessary condition is that xt ~ I(1) and yt ~ I(1), where-
as the sufficient condition is that εt ~ I(0). What distinguishes a pair of
cointegrated variables from another pair that are spuriously related is that
a linear combination of I(1) series produces another I(1) series except when
their underlying long-run movements affect each other so that the residu-
als obtained from the linear combination is I(0). Conversely, variables that
are spuriously related (perhaps due to a dominant time trend) but without
the same underlying long-run movement would not produce I(0) residuals.
It will be demonstrated later that this proposition is questionable.
The basic test for cointegration between xt and yt utilizes the DF statistic,
which is the t ratio of Φ in the Dickey-Fuller regression as applied to the
residuals of the cointegrating regression (1). The regression used to con-
duct the residual-based test is specified as εt = Φεt-1 + ut (2). Engle and
Granger (1991) assert that the t statistic of Φ no longer has the Dickey-
Fuller distribution – this is because when the cointegrating parameter (β)
is estimated the residuals εt appear slightly more stationary than if the
true value of the parameter is used. The distribution of the test statistic
is known as the Engle-Granger distribution. To reject the null hypothesis
of nonstationarity of the residuals (and therefore non-cointegration), the
value of the test statistic must be significantly negative (since a value
of zero implies random walk). The critical values of this test statistic are
tabulated in Engle and Granger (1987) and Engle and Yoo (1987, 1991),
and more precise values are found in MacKinnon (1991). The Dickey-
Fuller regression may be modified to produce the augmented version,
which is written as Δεt = Φεt-1 + ∑mi=1
δi Δεt-i + ut (3).
The residual-based approach, as proposed by Engle and Granger, has
been criticized on the following grounds. First, conflicting results are likely
to be obtained from the DF and ADF tests (depending on the lag length),
which may be attributed to the low power of these tests in small samples.
Second, extending the method to the multivariate case produces weak
and biased results [Gonzalo (1994)], and there is no way to tell whether
this linear combination is an independent vector or a linear combination
of independent vectors. Third, the results are not invariant or robust with
respect to the direction of normalization, that is, the choice of the variable
on the left-hand side of the cointegrating regression. Dickey et al. (1991)
argue that while the test is asymptotically invariant to the direction of
normalization, the test results may be very sensitive to it in finite samples.
Finally, there is a substantial finite sample bias [for example, Banerjee et
al. (1986)], and there is also the problem of implicit common factor re-
striction [for example, Kremers et al. (1992)]. Apart from that, two serious
shortcomings of the residual-based test are (i) the Dickey-Fuller test is
based on a simple AR(1) representation, which means that the underlying
model is misspecified in the sense that it should contain a moving aver-
age component; and (ii) the test is rather weak in distinguishing between
unit root and near-unit root processes.
In the late 1980s, the Johansen (1988) test for cointegration took the
world of academic economics and finance by a storm. This test quickly
115
The Capco Institute Journal of Financial TransformationThe Failure of Financial Econometrics: Assessing the Cointegration “Revolution”
became a “crowd pleaser” since it allowed anyone to prove anything
they wanted to prove: all it takes to obtain the results that you want is
a simple modification to the specification of the underlying model. The
test is based on a vector autoregressive representation that allows the
estimation of the cointegration matrix. Subsequently, two test statistics
can be calculated to determine the number of significant cointegrating
vectors: the maximal eigenvalue test (Max) and the trace test (Trace).
The claim to fame of the Johansen test is that, unlike the Engle-Granger
test, it produces results that are invariant with respect to the direction of
normalization. This is because all variables are explicitly endogenous,
so that there is no need to pick the left-hand side variable in an arbitrary
manner. Another perceived advantage of this test is that it provides esti-
mates of all of the cointegrating vectors that exist within a set of variables
and offers test statistics for their number. It has also been put forward
that (i) the Johansen test fully captures the underlying time series proper-
ties of the data; (ii) it is more discerning in its ability to reject a false null
hypothesis; (iii) it is based on a fully specified statistical model; and (iv) it
has the important advantage of using all of the information in the dataset,
thereby increasing estimation efficiency.
This seems to be a superb list of credentials for the Johansen test but
what about its shortcomings? I have often argued that this test is the
biggest scandal in modern econometrics because it can, at the touch
of a button, be used to prove the researcher’s underlying beliefs. This is
convenient because the majority of empirical research in economics and
finance is directed at proving preconceived ideas and producing “good”
results, rather than to go on a quest for the truth. In this sense, the test is
also dangerous, because it can be used to support faulty policy actions
and financial decisions. Imagine that you want to prove that privatization
is good under any circumstances to please a policymaker who believes
that for ideological reasons or that you fancy the proposition that interna-
tional diversification pays off. Not a problem, you will get the results you
want, thanks to the Johansen test.
The Johansen test suffers from major problems. One important short-
coming is that it does not allow the identification of separate functional
relations in a structural simultaneous equation model [for example Moosa
(1994)]. If, by applying the method to a set of variables, two cointegrat-
ing vectors are obtained, these vectors cannot be identified as specific
structural equations. As a matter of fact, no one knows what the cointe-
grating vectors are: structural equations, reduced-form equations, or a
combination thereof [Wickens (1996)]. Moreover, Reimers (1991) asserts
that the test over-rejects the null hypothesis of non-cointegration when
it is true, hence providing the ammunition for those wanting to prove
preconceived ideas.2 And if that does not work, then all it takes to obtain
the “desired” results is to change the specification of the underlying VAR,
for example, by changing the lag structure. Last, but not least, the test
invariably produces implausible point estimates of the coefficients of the
cointegrating vectors, hence you may get 178.6 for the estimate of a
demand elasticity that is supposed to be around unity.
Once any of the tests shows that cointegration is present, the correspond-
ing dynamic relation should be represented by an error correction model,
which combines short-term dynamics (as represented by first differenc-
es) and deviations from the long-run equilibrium relation (as represented
by the error correction term). The error correction model corresponding
to (1) is Δyt = ∑ki=1
αi Δyt-i + ∑ki=0
βi Δxt-i + γεt-i + ut (4), where the coefficient
on the error correction term measures the speed of adjustment towards
the long-run relation or the rate at which the deviation is eliminated. For
a valid error correction model, the coefficient on the error correction term
(γ) must be significantly negative.
Granger’s representation theorem [Engle and Granger (1987)] states that
cointegration implies and is implied by a valid error correction model.
With respect to equations (1) and (3), if εt ~ I(0), then γ should be sig-
nificantly negative and vice versa. This means that it is possible to test
for cointegration via an error correction model, in which case the null
of no cointegration is H0 : γ = 0 against the alternative H1 : γ < 0. More
generally, Granger’s representation theorem stipulates that cointegration
implies the existence of a valid error correction model in a regression of
Δy on Δx (as in equation 4), or vice versa.
Causality testing was popularized by Granger (1969). While the test was
initially based on a straight first difference model, the advent of cointe-
gration analysis led to a rethink of causality testing. If the variables are
cointegrated then the test should be based on an error correction model
because the first difference model would be misspecified. If this is the
case, causality should be detected in at least one direction, from x to
y or vice versa. The model used to test for causality in the presence of
cointegration is Δyt = ∑ki=1
αi Δyt-i + ∑ki=1
βi Δxt-i + γεt-i + ut (5), which is
the same as equation (4) except that the contemporaneous term Δxt is
deleted. This is because causality in economics and finance is not really
causality (as it is in physics). It is effectively temporal ordering – some-
thing causes something else because the first something occurs before
the other. Consequently, the results of causality testing mean nothing.
Furthermore, for x to be judged to affect y, x must be exogenous, which is
hardly the case in most application (for example, purchasing power par-
ity). Yet another problem is that the test results are sensitive to the choice
of the lag structure (the value of k), which induces scope for manipulating
the model to obtain the desired results.
Hjalmarsson and Österholm (2007) use Monte Carlo simulations to show that “in a system 2
with near-integrated variables, the probability of reaching an erroneous conclusion
regarding the cointegrating rank of the system is generally substantially higher than the
nominal size,” which means that “the risk of concluding that completely unrelated series are
cointegrated is therefore non-negligible.”
116
Applications in economics and financeCointegration analysis has been used extensively in economics and
finance. One of the applications in finance is an investment strategy
known as “pairs trading.” The underlying idea is that the spread between
(or ratio of) two cointegrated stock prices may widen temporarily, which
provides the opportunity to go short on one of the stocks and long on
the other, then exiting the two positions when the spread is back where
it should be, in other words, when the spread has gone through mean
reversion. Hence what matters is cointegration, not correlation, because
cointegration implies mean reversion. Schmidt (2008) uses the Johansen
test to detect stock pairs for this strategy as applied to some stocks
listed on the ASX200 (Australian Stock Exchange). Based on the results,
it is stated that “two cointegrated stocks can be combined in a certain
linear combination so that the dynamics of the resulting portfolio are gov-
erned by a stationary process.” Without experimenting with a trading rule
and based on plots of the residual series (showing a high rate of “zero
crossings” and large deviations around the mean), she concludes that
“this strategy would likely be profitable.” I would imagine that it is typical
to find that those suggesting these investment strategies are not willing
to bet their own money on the predictions of cointegration-based tests.
Alexander (1999) explains how cointegration can be used for the purpose
of hedging, arguing that “hedging methodologies based on cointegrated
financial assets may be more effective in the long term” and that “in-
vestment management strategies that are based only on volatility and
correlation of returns cannot guarantee long term performance.” She
suggests that “since high correlation alone is not sufficient to ensure the
long term performance of hedges, there is a need to augment standard
risk-return modeling methodologies to take account of common long-
term trends in prices,” which is “exactly what cointegration provides.”
But it has been demonstrated that accounting for cointegration makes no
difference whatsoever for the estimation of the hedge ratio and hedging
effectiveness. For example, it has been suggested by Lien (1996) that if
the price of the unhedged position and that of the hedging instrument
are cointegrated the position will be under-hedged if the hedge ratio is
estimated from a first difference model. This follows from Granger’s rep-
resentation theorem, which implies that if the prices are cointegrated the
first difference model will be misspecified because it ignores cointegra-
tion. However, it has been found that using an error correction model to
estimate the hedge ratio does not make the hedge more effective [Moosa
(2011a)]. In other words, the theoretically sound results of Lien (1996)
have negligible empirical significance or ramifications.
Cointegration and causality have been used to investigate market in-
terdependence and integration, inter alia, by Taylor and Tonks (1989),
Mathur and Subrahmanyam (1990), Eun and Shin (1989), and Malliaris
and Urrutia (1992). For example, Taylor and Tonks (1989) used cointe-
gration analysis to examine the effect of the 1979 abolition of the U.K.
exchange controls on the degree of integration between the British mar-
ket and other markets (Germany, the Netherlands, Japan, and the U.S.).
The results show that U.K. stock prices became cointegrated with prices
in other markets in the post-1979 period, which reduced the scope for
diversification. Mathur and Subrahamanyan (1990) used causality testing
to find out if the Nordic markets (of Denmark, Norway, Finland, and Swe-
den) are integrated with that of the U.S. The results reveal that the U.S.
market affects the Danish market only and that the Norwegian, Danish,
and Finnish markets do not affect any of the other markets (naturally, no
explanation is suggested for differences in the results).3
Conversely, there was much more uniformity in the results obtained by
Eun and Shin (1989), who estimated a nine-market VAR. They detected
a considerable amount of multilateral interaction and significant effect of
U.S. innovations on other markets. However, they also found that no sin-
gle foreign market could adequately explain the U.S. market movements.
Finally, Malliaris and Urrutia (1992) investigated the lead-lag relations
among six markets in diverse time zones for the period before, during,
and after the October 1987 crash. They concluded that the crash was
probably an international crisis and that it might have begun simultane-
ously in all national stock markets. The implication of these results is that
international diversification will not work if long positions are taken on a
group of integrated markets, as shown by the results of cointegration.
In international finance, one of the most popular applications of cointe-
gration has been the testing of international parity conditions, starting
with the testing of purchasing power parity (PPP), which was initiated by
Taylor (1988) and Enders (1988). The production of papers testing PPP
by cointegration is yet to come to an end, but we are not better off with
respect to our understanding of PPP. This hypothesis works well over
very long periods of time and under hyperinflation. The use of cointegra-
tion analysis to test PPP, particularly with the Johansen test, gives a false
indication that PPP works over a short period of time. But to any observ-
er, exchange rates are too volatile to be explained by the smooth price
movements over time. Cointegration has also been used to test covered
and uncovered interest parity, CIP and UIP, respectively. In particular, I
have been puzzled by attempts to use cointegration to test CIP, because
this condition must hold by definition, as an arbitrage or a hedging con-
dition [Moosa (2004)]. Strangely, cointegration tests may tell us that CIP
does not work, as we are going to see.
Cointegration analysis is also used for policy making. For example, Dreh-
mann et al. (2010) use the Johnansen test to derive results to design a
It is invariably the case that when several countries, currencies, or whatever are examined 3
using cointegration analysis, the results turn out to be all over the place. Typically, the
results would show that A and B are cointegrated but A and C are not, and no one knows
why because no one presents an explanation why.
117
The Capco Institute Journal of Financial TransformationThe Failure of Financial Econometrics: Assessing the Cointegration “Revolution”
program for countercyclical capital buffers, which has been introduced
as part of the so-called Basel III provisions [Moosa (2011b)]. The Johan-
sen test is also used by Ericsson et al. (1998) to estimate an econometric
model for the U.K. money, output, prices, and interest rates. This notori-
ous test is used to judge exchange rate misalignment by estimating a
cointegrating vector relating the exchange rate to its determining factors.
It is this kind of work that has led to the conclusion that the Chinese yuan
is undervalued against the dollar, which is the pretext for a potential full-
scale trade war between the U.S. and China [Moosa (2011c, 2011d)]. We
are talking about some serious business here, too serious to be sorted
out on the basis of the fragile and misleading cointegration tests.
Correlation and cointegrationApplications of cointegration in economics and finance are based on two
propositions pertaining to the difference between cointegration and cor-
relation. The first proposition is that cointegration tells you a different
story from that told by correlation, which is the basis of pairs trading and
the hedging argument. The second is that cointegration enables us to
distinguish a spurious relation between two highly correlated variables
from a genuine one. Hence it is suggested that only the variables that are
genuinely related exhibit cointegration. The problem is that no one has
shown how any cointegration test can do this sort of forensic investiga-
tion. The empirical evidence does not support this contention either.
The proposition on the distinction between correlation and cointegration
is not very well thought of. For example, Chan (2006) distinguishes be-
tween the two concepts by using the co-movements of two (theoreti-
cally constructed) stock prices. He argues that the two prices are cor-
related if they rise and fall “in synchrony,” whereas they are cointegrated
if they do not “wander off in opposite directions for very long without
coming back to a mean distance eventually.” By “synchrony,” it is meant
that prices rise and fall together on a daily, weekly, or monthly basis. It
follows, therefore, that the spread between two cointegrated prices is
mean reverting while the spread between two perfectly correlated prices
is constant. But this is misleading because “synchrony” implies perfect
correlation, which can never be the case in practice. This also means that
perfectly or highly correlated prices are necessarily cointegrated because
a constant spread is by definition stationary. Alexander (1999) argues that
“high correlation does not necessarily imply high cointegration in prices.”
She shows a graph of the German mark and Dutch guilder over the pe-
riod 1975-85, arguing that “they appear to be cointegrated.” Then she
adds a very small daily incremental return to the guilder, suggesting that
the series are still highly correlated but not cointegrated.” However, this
addition of incremental return affects both cointegration and correlation,
to the extent that the more the two variables drift apart the less correlated
they become. Does this also mean that variables that are negatively cor-
related cannot be cointegrated as Chan (2006) argued explicitly (“wander
off in opposite directions”)? This is an issue that we will explore later.
0
10
20
30
40
50
60
70
80
0 10 20 30 40 50 60
Highly correlated variables
x1 x2
-50
0
50
100
150
200
0 10 20 30 40 50 60
Moderately correlated variables
x3 x4
Figure 1 – Cointegration and correlation
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
1945 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
x1 x2 x3 x4 x5
Source: Federal Reserve
Figure 2 – Correlation and cointegration of measures of U.S. indebtedness
I will argue here that cointegration analysis is not capable of distinguish-
ing between a spurious relation and a genuine (causal) one. For this pur-
pose, I present two examples, one is based on artificially generated time
series (hence they can only be spuriously related, if at all) and another
that is based on actual data on U.S. debt. Starting with the artificial data,
four time series variables are generated over 51 points in time ranging be-
tween t=0 and t=50. The four time series (x1, x2, x3, and x4) are generated
118
as follows: x1t = 10 + 1.2t + ε1t; x2t = 10 + 0.8t + ε2t; x3t = 10 + 4t + ε3t;
x4t = 10 + ζtt + ε4t (6), where ε1, ε2, ε3, ε4 and ζ are random variables
falling in the ranges (-1,1), (-2,2), (-25,25), (-10,10), and (0,2) respectively.
By design, x1 and x2 are highly correlated (correlation = 0.99) whereas x3
and x4 are not (correlation is 0.51). Figure 1 shows how the four variables
move over time. Since they are not genuinely related by design we should
obtain a finding of no cointegration between x1 and x2 and between x3
and x4. However, that turned out not to be the case: x1 and x2 are cointe-
grated (ADF = -5.01) but x3 and x4 are not (ADF = -2.50).
The same argument can be illustrated with the use of actual data on five
variables representing measures of U.S. indebtedness over the period
1946-2010 as displayed in Figure 2: x1 is the liabilities of non-financial
business, x2 is the liabilities of financial institutions, x3 is the U.S. lia-
bilities to the rest of the world, x4 is total foreign assets, and x5 is GDP.
Table 1 reports correlations and the ADF statistics of the residuals of
the cointegrating regressions for each possible pair of the five variables.
While there is no exact correspondence between correlation and the find-
ing of cointegration, the latter is found in three cases where correlation
is at least 0.97. But why would the U.S. liabilities to the rest of the world
have a genuine relation with GDP but not the liabilities of financial institu-
tions and total foreign assets? In general, what makes the relations (x2,
x3), (x3, x4), and (x3, x5) genuine whereas the other relations are not?
Where is the theory or the intuition that explains these findings? A genu-
ine relation must be formally or intuitively explainable, not in this example
though. Consequently, a finding of cointegration does not necessarily
imply a genuine relation in the sense that the genuineness of the relation
can be justified theoretically or intuitively.
The distinction between cointegration and correlation as described by
Chan (2006) and Alexander (1999) seems to suggest that two nega-
tively correlated variables cannot be cointegrated because they do not
move “in synchrony,” they drift apart from each other, and they move
in opposite directions. Consider two time series variables, x’5t and x’1t,
which are generated as follows from the variable x1t as in equation (6):
x’5t = -0.5x1t + ε5t; x’5t = 50 + x5t; x’1t = 50 + x1t (8); where ε5t is a ran-
dom variable falling in the range (-2,2). x’5t and x’1t are negatively cor-
related by construction, which is obvious from Figure 3 (the correlation
coefficient is -0.99). The two variables drift apart, hence they cannot be
cointegrated. Yet, by running a cointegrating regression of x’1t on x’5t
and applying the Dickey-Fuller test to the residuals of the regression, the
results show that ADF=-5.59, which is statistically significant. Hence, x’5t
and x’1t, two variables that drift apart from each other, are cointegrated.
It is interesting, however, that the logs of these variables are not cointe-
grated (ADF=-2.81).4
An illustration: stock market integrationThe idea behind testing for cointegration between the stock prices of
two countries is simple. If cointegration is established, then the stock
markets of the two countries are said to be integrated, which means that
international diversification involving these two markets is not effective in
the sense that it does not reduce risk. Two flaws can be detected in this
line of reasoning. First, the underlying assumption is that stock prices
are positively correlated, which means that taking long positions on the
two markets will not produce risk reduction. However, even if this is the
case diversification can be implemented by taking a short position on
one market and a long position on the other. The second flaw is that the
effectiveness of hedging depends on correlation, not on cointegration,
which is contrary to the argument put forward by Alexander (1999).
This illustration is based on quarterly stock price data from a diverse
set of countries over the period 2001-2010.5 These countries are the
U.S., Japan (JP), U.K., Canada (CA), Australia (AU), New Zealand (NZ),
Switzerland (SW), Singapore (SN), Korea (KO), India (IN), South Africa
(SA), Brazil (BR), Turkey (TR), and Kuwait (KW). We will test the relation
between each one of these markets and that of the U.S. by subjecting
the stock price time series to the following tests: the residual-based test
(ADF), the Johansen test with two different lag lengths, the test based
on the error correction term, and causality testing from the U.S. market
to another market, and vice versa. Following normal practice logarithmic
prices are used. The results are reported in Table 2.
Judged by the ADF statistic and the Johansen test (with four lags), none
of the markets is integrated with that of the U.S., a result that may be dis-
appointing for those who want to show that markets should be integrated
The use of logs in cointegration analysis is rather arbitrary. A typical justification for the use 4
of logs is that the first log difference is more likely to be stationary than the first absolute
difference. Hence logs are used for convenience, not for a theoretical or empirical reason.
The problem is that the use or otherwise of logs may introduce significant qualitative
differences in the results, as we are going to see later.
The source of the data is the IMF’s International Financial Statistics.5
x1 x2 x3 x4 x5
x1 1.00
x2 0.89
-1.81 1.00
x3 0.81 0.97 1.00
-2.23 -4.06*
x4 0.92 0.99 0.96 1.00
-2.33 -4.06* -3.22
x5 0.81 0.97 0.98 0.95 1.00
-1.54 -2.28 -4.39* -0.62
* Significant at the 5% level
Table 1 – Correlation coefficients (top) and ADF statistics (bottom) for measures of U.S. indebtedness
119
in the age of globalization. But this is not a problem. Just change the
model specification in the magical Johansen procedure to 12, then all of
the markets become integrated. The test based on the error correction
term shows that only four markets are integrated with that of the U.S.:
U.K., New Zealand, South Africa, and Kuwait. So, what are we supposed
to believe? The tendency would be to report the results that support a
preconceived idea. If I accepted the proposition that cointegration has
implications for the benefits or otherwise of international diversification,
and if I wanted to prove a preconceived idea, I would do the following. I
would report the results of the ADF test if I thought that there was scope
for international diversification, but I would report the results of the Jo-
hansen test with lag 12 if I thought that international diversification was
not effective. If I held the view of sometimes/sometimes not and perhaps/
perhaps not, I would report the results based on the error correction test.
But then how would I explain the finding that the markets of Kuwait,
South Africa and New Zealand are integrated with the U.S. but those of
Japan, Canada, and Singapore are not?6 It is simply embarrassing and
even hazardous to derive inference from any set of these results.
Checking the validity of some predictions of the theory of cointegration
produces interesting results. Specifically, we want to examine the valid-
ity of the propositions that (i) cointegration implies and is implied by the
validity of the corresponding error correction model; and (ii) cointegra-
tion implies that causality should run in at least one direction. For this
purpose we will use the ADF test results, since the error correction terms
are taken to be the lagged residuals of the bivariate cointegration regres-
sions. The results tell us that there is no cointegration in any case, but the
coefficient on the error correction term is significant in four cases, when
Granger’s representation theorem stipulates that it should be significant
in none. The cointegrating regressions were run both ways, but that did
not change the results. As far as causality is concerned, the results show
that the Singapore stock prices, which are not cointegrated with U.S.
stock prices, are causally related to them – which is fine, because cointe-
gration does not necessarily preclude causation. The problem is how to
explain why, out of all of these markets, only the Singapore market has an
effect on the U.S. market. The only other case of causality is that the U.S.
market has an effect on that of South Africa but not on any of the others
(again, why South Africa?). Out of the four cases that show cointegration,
the only case that produces unidirectional causality is that of South Af-
rica. So much for the implications of cointegration on causality.
Now we examine the practical significance of these results in terms of
the benefits of international diversification. For this purpose we construct
portfolios consisting of a position on the U.S. market and an opposite
position (of the same size) on one of the other markets. If the hedge is
effective then the variance of the rate of return on U.S. stocks should be
significantly higher than the variance of the rate of return on the portfolio.
For this purpose we conduct a variance ratio test, which requires the
calculation of the variance ratio: VR = σ2(RUS)/ σ2(RP) (8), where σ2(RUS)
is the variance of the rate of return on U.S. stocks (the first log difference
of the stock price index) and σ2(RP) is the variance of the rate of return
The Capco Institute Journal of Financial TransformationThe Failure of Financial Econometrics: Assessing the Cointegration “Revolution”
It is interesting to note that rising (and falling) oil prices should lead the Kuwait and U.S. 6
markets to drift away from each other.
0
20
40
60
80
100
120
0 10 20 30 40 50
x'1 x'5
Source: Federal Reserve
Figure 3 – Cointegration of negatively-correlated variables
Test statistic JP UK AU CA NZ SW SN KO IN SA BR TR KW
ADF -1.36 -2.72 -1.47 -2.11 -2.69 -2.93 -0.85 -0.49 -0.82 -2.12 -0.91 -0.93 -2.86
Johansen (4 lags) max 13.60 10.6 7.14 9.63 10.38 12.28 8.04 4.94 6.85 13.64 7.58 8.62 11.63
Johansen (4 lags) trace 14.68 14.93 8.21 11.01 14.79 15.27 8.15 5.72 7.03 16.57 8.07 8.87 18.05*
Johansen (12 lags) max 73.25* 57.01* 104.20* 117.22* 136.41* 161.67* 133.59* 97.75* 80.63* 58.26* 91.58* 74.04* 44.16*
Johansen (12 lags) trace 79.15* 15.44* 157.62* 120.36* 141.20* 172* 141.40* 97.57* 130.88* 66.61* 92.10* 75.15* 44.17*
Error correction t(γ) -0.25 -2.08* -1.09 -0.36 -2.15* -1.14 0.39 -1.42 -0.56 -6.36* -0.64 -0.04 -3.17*
US→ XX [χ2(4)] 6.49 1.51 3.21 2.20 7.93 5.49 5.23 3.42 0.27 17.55* 1.81 0.86 4.02
XX→ US [χ2(4)] 3.69 1.61 5.18 6.63 0.89 3.89 22.98* 4.91 3.02 1.60 4.51 0.97 7.30
*Significant at the 5% level. The 5% critical values are: ADF (-3.52), max (14.88), trace (17.86), error correction (-2), χ2(4) (9.48).
Table 2 – Testing the hypothesis of market integration
120
on the portfolio. For simplicity, we assume an equally-weighted portfolio,
hence: RP = 0.5(RUS – RXX) = 0.5(ΔpUS – ΔpXX) (9), where RXX is the rate
of return on the stocks of the other country.7 Hedging is effective (that is,
international diversification is useful) if VR > F(n-1, n-1) (10), where n is
the sample size. The VR test can be supplemented by a measure of vari-
ance reduction (VD), which is calculated as: VD = 1 – 1/VR (11).
The results are presented in Table 3, which reports the variance ratio,
variance reduction, and the correlation between U.S. returns and returns
on the stock market in the other country. We can see that hedging is ef-
fective (international diversification is useful) in two cases only: the U.K.
and Canada, one showing cointegration while the other is not, according
to the error correction test. Why these two markets? Because they exhib-
it the highest correlation with returns on the U.S. market. Consequently,
a simple concept like correlation leads to the right inference but the “so-
phisticated” tests of cointegration produce messy results that may lead
to faulty financial decisions. The claim that what matters for hedging is
cointegration, not correlation, is unfounded.
Another illustration: international parity conditionsIn this section we illustrate how misleading cointegration analysis can be
by testing two international parity conditions: purchasing power parity
and covered interest parity. For this purpose, we employ the residual-
based test and the error correction test, but not the Johansen test be-
cause we have established that it is unreliable (actually, highly reliable if
you want to produce “desirable results’).
PPP is tested using two specifications: the restricted specification in
which the exchange rate is a function of the price ratio, and the unre-
stricted specification in which the exchange rate is a function of the price
indices in the two countries. The two specifications are written in loga-
rithmic form respectively as: st = α0 + α1(pxx, t – pUS, t) + εt (12) and st =
β0 + β1pxx, t + β2pUS, t) + ζt (13), where s is the (log of) exchange rate,
pUS is the (log of) price level in the U.S. and pxx is the (log of) price level
in the other country.8 The results of testing PPP for 13 countries against
the U.S., using quarterly data covering the period 2001Q1-2010Q3, are
reported in Table 4.
The results show that when PPP is tested by applying the Dickey-Fuller
test to the residuals of the cointegrating regression, evidence for cointe-
gration is apparent between the U.S. and South Africa when the un-
restricted specification of the model is used. What is so special about
South Africa to be the only country for which PPP is valid against the
U.S.? When the error correction test is used, most cases exhibit cointe-
gration, which casts doubt on the validity of Granger’s representation
theorem. But then in some cases cointegration is present when the re-
stricted specification is used but not when the unrestricted specification
is used, which is a contradiction. If PPP is valid, as shown by testing the
restricted specification, then the restriction (resulting from the imposition
of the condition of symmetry) must be valid. If this is so, then the cor-
responding unrestricted version must also be valid, but this is not so in
all cases.
Moosa and Al-Deehani (2009) show how to calculate the portfolio weights that minimize the 7
variance of the rate of return on the portfolio.
The use of log-linear specification is appropriate in this case because the “raw” 8
specification of PPP is that the exchange rate is a function of the price ratio. Hence the
specification is justified by theoretical reasoning.
Market VR VD Correlation
JP 1.11 10.15 0.74
UK# 6.30* 84.13 0.92
CA 3.28* 69.53 0.89
AU 1.52 34.39 0.73
NZ# 2.45 59.22 0.77
SW 1.33 24.54 0.76
SN 0.74 -34.58 0.76
KO 1.00 0.05 0.78
IN 0.38 -160.28 0.66
SA# 0.42 -137.65 0.49
BR 0.69 -44.58 0.82
TR 0.47 -111.73 0.75
KW# 0.36 -177.87 0.56
# Markets cointegrated with the U.S.
* Significant at the 5% level (critical value is 2.68).
Table 3 – Results of the variance ratio test
Restricted
(ADF)
Unrestricted
(ADF)
Restricted
(EC)
Unrestricted
(EC)
JP -1.43 -1.90 -1.36 -1.54
UK -2.45 -2.81 -2.36* -2.39*
CA -2.23 -2.17 -1.90 -1.38
AU -2.53 -2.78 -2.25* -2.07*
NZ -2.06 -2.58 -2.11* -2.06*
SW -2.71 -2.99 -2.61* -2.95*
SN -0.10 -3.64 -0.67 0.00
KO -1.76 -2.13 -2.07* -0.97
IN -1.83 -2.09 -2.89* -1.93
SA -2.32 -4.61* -2.92* -3.93*
BR -1.94 -2.89 -5.04* -3.42*
TR -2.59 -2.53 -3.77* -3.51*
KW -2.74 -2.81 -2.79* -2.57*
* Significant at the 5% level (critical value of the ADF is -3.52 for the restricted specification
and -4.01 for the unrestricted specification). The critical value for the EC test is -2.
Table 4 – Testing PPP
121
Covered interest parity is a rather peculiar case. One way to test CIP is to
find out if the actual and forward exchange rates are cointegrated. How-
ever, CIP must hold as a truism because a bank will never quote a forward
rate that is different from that compatible with CIP, the so-called interest
parity forward rate. This is because this is the only rate that precludes the
possibility of profitable risk-free arbitrage (hence CIP is a no-arbitrage
condition) and it is the only rate that enables the quoting bank to hedge
its exposure perfectly (hence CIP as a hedging condition).9 The interest
parity forward rate is calculated as Ft = St [(1 + iXX)/(1 + iUS)] (14), where
F is the forward rate, S is the spot rate, iXX is the interest rate in the other
country and iUS is the interest rate in the U.S. Although CIP must hold by
definition and design, cointegration tests may reveal that it does not hold.
Consider the cointegrating regression: ƒt = α + βst + εt (15), where lower-
case letters imply logarithms. Cointegration between ƒt and st requires
that εt ~ I(0). The results, presented in Table 5, show that cointegration is
present in four cases only when the error correction test is used and in no
case when the ADF test is used. A plot of the spot versus forward rates
shows similar close relations irrespective of whether or not cointegration
is present. The spot and forward rates do not drift apart from each other,
which makes one wonder why they are not cointegrated.
But irrespective of the results of cointegration testing, CIP must always
hold in the sense that a bank can only quote the forward rate implied by
equation (14). There cannot be any deviations from this condition be-
cause deviation implies the availability of riskless arbitrage profit. Cointe-
gration results have no meaning whatsoever in the case of CIP, and the
hundreds of tests that have been done to find out whether or not CIP
holds are a total waste of time.
The Capco Institute Journal of Financial TransformationThe Failure of Financial Econometrics: Assessing the Cointegration “Revolution”
ConclusionIt seems that the cointegration “revolution” was not a revolution at all. It
was another econometric “trick” that does not help our understanding of
how the economy and financial system work. On the contrary, cointegra-
tion analysis may produce results that are highly misleading (for example,
CIP does not hold, hence free money is available for no risk). One can
only wonder why this gimmick is considered as important for our lives as
the discovery of Penicillin (at least in the eyes of the Nobel Prize Com-
mittee).
The problems associated with cointegration analysis are plentiful. To start
with, the results obtained by using different tests typically vary consid-
erably and they are not robust with respect to model specification (for
example, linear versus log linear, restricted versus unrestricted, changing
the lag structure, changing the direction of normalization, and the addi-
tion or deletion of a time trend). Hence, the technique offers the opportu-
nity for anyone to prove whatever they like. This can be very dangerous if
the results are used for policy formulation or for financial decisions. The
claim to fame of cointegration analysis that it can distinguish spurious
relations from genuine ones is false. We can only do that by common
sense, theory, and/or intuition. Furthermore, some of the pillars of cointe-
gration analysis are not supported by the results presented in this study:
cointegration does not necessarily imply and is implied by a valid error
correction representation, and it does not necessarily imply that causality
must be present in at least one direction. In some cases simple correla-
tion analysis does a better job than cointegration testing.
Cointegration analysis, like many of the techniques of financial econo-
metrics, is not worthy of the brain power spent on its development. While
it has provided the means for finance and economics academics to pub-
lish papers and for students in the two disciplines to obtain their PhDs,
the technique has not provided any useful insights. On the contrary, it
typically provides faulty, inconsistent, and robustness-lacking results that
may be hazardous to use. I would certainly advocate the use of a warning
phrase such as “handle with care” to describe results based on cointe-
gration testing. It is yet another manifestation of the failure of financial
econometrics.
See Moosa (2004) for a distinction between CIP as an arbitrage and a hedging condition.9
ADF EC
JP -1.96 -3.15*
UK -2.34 -1.76
CA -2.47 -2.58*
AU -2.14 -1.65
NZ -2.86 -2.09*
SW -2.38 -1.15
SN -2.02 -0.58
KO -3.27 -1.25
IN -2.80 -1.69
SA -3.17 -1.60
BR -1.48 -1.88
TR -3.24 -2.51*
KW -2.07 -0.10
* Significant at the 5% level (critical value is -1.96)
Table 5 – Testing CIP
122
ReferencesAlexander, C., 1999, “Optimal hedging using cointegration,” Philosophical Transactions of the •
Royal Society, Series A 357, 2039-2085
Banerjee, A., J. J. Dolado, D. F. Hendry, and G. W. Smith, 1986, “Exploring equilibrium •
relationships in econometrics through static models: some Monte Carlo evidence,” Oxford
Bulletin of Economics and Statistics, 48, 253-77
Chan, E., 2006, “Cointegration is not the same as correlation,” TradingMarkets.Com, 13 •
November
Dickey, D. A., D. W. Jansen, and D. L. Thornton, 1991, A primer on cointegration with an •
application to money and income,” Federal Reserve Bank of St. Louis Economic Review,
March/April, 58-78
Drehmann, M., C. Borio, L. Gambacorta, G. Jimenez, and C. Trucharte, 2010, “Countercyclical •
capital buffers: exploring options,” BIS Working Papers, No 317
Enders, W., 1988, ARIMA and cointegration tests of PPP under fixed and flexible exchange rate •
regimes,” Review of Economics and Statistics, 70, 504-508
Engle, R. F. and C. W. J. Granger, 1987, Cointegration and error correction: representation, •
estimation and testing, Econometrica, 55, 251-76
Engle, R. F. and C. W. J. Granger (eds.), 1991, Long-run economic relationships: readings in •
cointegration, Oxford University Press
Engle, R. F. and B. S. Yoo, 1987, “Forecasting and testing in cointegrated systems,” Journal of •
Econometrics, 35, 143-59
Engle, R.F. and B. S. Yoo, 1991, “Cointegrated economic time series: an overview with new •
results,” in Engle, R. F. and C. W. J. Granger (eds.), 1991, Long-run economic relationships:
readings in cointegration, Oxford University Press
Ericsson, N. R., D. F. Hendry, and G. E. Mizon, 1998, “Exogeniety, cointegration and economic •
policy analysis,” Federal Reserve System, International Finance Discussion Papers, No 616,
June
Eun, C. S. and S. Shin, 1989, “International transmission of stock market movements,” Journal •
of Financial and Quantitative Analysis, 24, 41-56
Gonzalo, J., 1994, “Comparison of five alternative methods of estimating long-run equilibrium •
relationships,” Journal of Econometrics, 60, 203-33
Granger, C. W. J., 1969, “Investigating causal relations by econometric models and cross-•
spectral methods,” Econometrica, 37, 424-438
Hjalmarsson, E. and P. Österholm, 2007, “Testing for cointegration using the Johansen •
methodology when variables are near-integrated,” IMF Working Papers, June
Johansen, S., 1988, “Statistical analysis of cointegrating vectors,” Journal of Economic •
Dynamics and Control, 12, 231-54
Kremers, J. J. M., N. R. Ericsson, and J. J. Dolado, 1992, “The power of cointegration tests,” •
Oxford Bulletin of Economics and Statistics, 54, 325-48
Lien, D., 1996, “The effect of cointegration relationship on futures hedging: a note,” Journal of •
Futures Markets, 16:7, 773-780
MacKinnon, J. G., 1991, “Critical values for cointegration tests,” in Engle, R. F., and C. W. •
J. Granger (eds.), 1991, Long-run economic relationships: readings in cointegration, Oxford
University Press Malliaris, A. G., and J. L. Urrutia, 1992, “The international crash of October
1987: causality tests,” Journal of Financial and Quantitative Analysis, 27, 353-364
Mathur, T., and V. Subrahmanyam, 1990, “Interdependencies among the Nordic and US Stock •
Markets,” Scandinavian Journal of Economics, 92, 587-597
Moosa, I. A., 1994, “The monetary model of exchange rates revisited,” Applied Financial •
Economics, 4, 279-87
Moosa, I. A., 2004, “Is covered interest parity an arbitrage or a hedging condition?” Economia •
Internazionale, 57, 189-194
Moosa, I. A., 2011a, “The failure of financial econometrics: estimation of the hedge ratio as an •
illustration, Journal of Financial Transformation, 31, 67-72
Moosa, I. A., 2011b, “Basel II to Basel III: a great leap forward?” in La Brosse, J. R., R. Olivares-•
Caminal, and D. Singh (eds), Managing risk in the financial system, Edward Elgar
Moosa, I. A., 2011c, “Undermining the case for a trade war between the U.S. and China,” •
Economia Internazionale, forthcoming
Moosa, I. A., 2011d, “On the U.S.-Chinese trade dispute,” Journal of Post Keynesian •
Economics, forthcoming
Moosa, I. A. and T. Al-Deehani, 2009, “The myth of international diversification,” Economia •
Internazionale, 62, 83-406
Reimers, H. E., 1991, “Comparison of tests for multivariate cointegration,” Christian-Alberchts •
University, Discussion Paper No. 58
Schmidt, A. D., 2008, “Pairs trading: a cointegration approach,” Honours Thesis, University of •
Sydney
Taylor, M. P., 1988, “An empirical examination of long-run purchasing power parity using •
cointegration techniques,” Applied Economics, 20, 1369-1381
Taylor, M. P. and I. Tonks, 1989, “The internationalisation of stock markets and the abolition of •
UK exchange control,” Review of Economics and Statistics, 71, 332-336
Wickens, M. R., 1996, “Interpreting cointegrating vectors and common stochastic trends,” •
Journal of Econometrics, 74, 255-271
123
PART 2
A General Structural Approach For Credit Modeling Under Stochastic Volatility
AbstractThis paper assumes a structural credit model with underlying
stochastic volatility combining the Black/Cox approach with
the Heston model. We model the equity of a company as a
barrier call option on its assets. The assets are assumed to
follow a stochastic volatility process; this implies an equity
model with most documented stylized facts incorporated.
We derive the price of this option under a general framework
where the barrier and strike are different from each other,
allowing for richer financial applications. The expression for
the probability of default under this framework is also pro-
vided. As the calibration of this model gets much more com-
plex, we present an iterative fitting algorithm with which we
are able to nicely estimate the parameters of the model, and
we show via simulation the consistency of the estimator. We
also study the sensitivity of the model parameters to the dif-
ference between the barrier and strike price.
Marcos Escobar — Associate Professor, Ryerson University
Tim Friederich — PhD candidate, HVB Institute for Mathematical Finance, Technische Universität München, and a Financial Engineer, Risklab GmbH
Luis Seco — Director, RiskLab, and Professor, University of Toronto
Rudi Zagst — Director, HVB Institute for Mathematical Finance, and Professor, Technische Universität München
124
When we aim to characterize the performance of a stock, the stock price
is mainly the result of the behavior of the company’s assets and its liabili-
ties. Yet, the evolution of assets and liabilities is usually not reported on a
daily basis. For this reason, we modeled a company’s equity as a barrier
call option on the assets, with the liabilities as barrier and strike price. The
value of the equity is therefore given by the price of the down-and-out
call option [Escobar et al. (2010)]. In this manner, we can calculate the
asset price time series from the given equity price time series by inverting
the option pricing formula, in particular, the asset price volatility becomes
the volatility implied from the option price. This model is mainly based
on the foundations of structural credit models laid by Merton (1973) and
Black and Cox (1976). We combined this interpretation of the equity price
as a call option with the Heston model [see Heston (1993)] by modeling
the asset of the company as a stochastic volatility process. We also de-
rived estimators for the calibration of the model, inspired by the work of
Genon-Catalot et al. (2000).
In Escobar et al. (2010), we assumed the knock-out barrier and strike
price to be equal and found the option pricing formula to be of a very
simple form with a straightforward inverse formula, which allows us to
calculate the asset price if the value of the option is known. In this paper,
we do not make this assumption anymore and derive the price of the
equity in the form of a general barrier call option. A strike price D+K with
K>0 can be economically interpreted in various ways. For example, the
additional costs of fulfilling the option contract at maturity. These costs
are not involved when selling the option before maturity. Such costs are
likely for over-the-counter (OTC) options and common for customized
options. As a practical example, for OTC options on commodities or
other physical goods, the transportation or storage costs incurred when
the option is exercised might affect the price someone is willing to pay
for such an option.
An alternative explanation of D and K is to assume that total liabilities
consist of the debt D and an additional debt K granted to the cash ac-
count. The credit limit on the cash account does not require any backing
by assets but is granted for the liquidity of daily business cash flows on
the company’s cash account. However, the company only defaults if the
assets fall below D. In other words, the debt on the cash account is not
considered for determining the case of a default, but only the true debt
D. Another interpretation comes from taking D+K1 as the actual debt with
K1 < K so that K1 and T are the maximum amount and time that equity
holders (EH) and debt holders (DH) are willing to wait to avoid bankrupt-
cy. Default would mean a total loss for EH as they are the first to be left
out, so allowing the asset to wander in [D, D+K1] is beneficial to them. On
the other hand, DH are usually willing to accept this risk for some extra
return K - K1 on the money lent.
The model presented for the equity not only has more stylized facts de-
scribed in the literature, see Cont (2001), like for example stochastic vola-
tility and correlation between the volatility and the equity, but also allows
for estimation which is a widely underestimated problem in the literature.
A key challenge with calibrating the model parameters is that the param-
eters are needed two steps: for the calculation of the inverse of the option
price, and for fitting the parameters to the asset process. This is a very
common estimation problem in finance, as it appears, for example, when
estimating the parameters of the intensity in a reduced form model using
credit prices or the parameters of the instantaneous interest rate from
bond prices, or simply the implied volatility from option prices. In this pa-
per, we present an iterative fitting method, which dramatically decreases
the computation time and makes the more complex setting in this paper
manageable. We furthermore conduct tests to examine numerically the
quality of the fitting, and with a case study examine the sensitivities of
the model parameters to the difference between the barrier and the strike
price for which we allow in this paper.
The stochastic volatility modelAs asset returns are, in general, not normally distributed, applying a
Black/Scholes model assuming geometric Brownian motions for the per-
formance of asset prices is a significant simplification. In particular, in
falling stock markets, Engle (1982) and Heston (1993) find that volatil-
ity increases. To account for this so-called heteroscedasticity, we use a
model incorporating stochastic volatility.
Let (Ω, F, F, Q) be a filtered probability space with filtration F = {Ft}t>0.
The underlying asset process A and the variance v of that process are
expressed through the following SDEs:
dA(t) = µA(t)dt + A(t)dZ(t) (1);
dv(t) = κv[v∞ - v(t)]dt + εv dZvt (2)
where A is the underlying asset process, v is the variance of asset pro-
cess with volatility σA = , µ is drift of the assets, Z, Zv are two inde-
pendent Wiener processes in the probability space (Ω, F, Q), v∞ is the
long-term value of volatility, εv is the volatility of the variance process, and
κv is the mean-reversion speed.
The parameters v∞, κv, and εv have to fulfill the following condition: κv ·
v∞ > ½εv2. This condition, which is also called the Feller condition [Feller
(1951)], guarantees that the variance process is always greater than zero.
Furthermore, claiming κv > 0 ensures strict stationarity and α-mixing [see
Genon-Catalot et al. (2000)]. The debt is assumed to be exponentially
growing with the risk-free rate r: D(t) = D(0) · = D(T) · . The
model defined by (1) and (2) is a so-called stochastic volatility model and
is the same model setup we used in Escobar et al. (2010). For a discus-
sion of this model, we refer the interested reader to Heston (1993). Due to
the fact that it was introduced by this author, this model is also referred
to as the “Heston model.”
125
The Capco Institute Journal of Financial TransformationA General Structural Approach For Credit Modeling Under Stochastic Volatility
Pricing of barrier optionsIn the structural credit model, the value of the company is interpreted as
a call option on its assets with its liabilities as strike price. As the com-
pany can default over time, a barrier is introduced. Thus, if the value of
the assets falls below the barrier at any point in time, the option expires
worthless. As opposed to Escobar et al. (2010), where both the barrier
and strike price of the option were assumed to be equal, in the more
general setting which we examine in this study, they can assume differ-
ent values.
Figure 1 displays three possible paths with the following parameteriza-
tion: A(0) = 100, r = 0 (for the purpose of simplicity), v(0) = v∞ = 0.01, κv =
0.75, εv = 0.01. As an example, consider a down-and-out call option with
barrier D = 80 (constant due to r = 0) and strike D + K = 90. The underly-
ing path 1 never falls below neither the strike nor the barrier and has a
terminal value of 125. Thus, the payoff at maturity T = 5 is 125 – 90 = 35.
Path 2 falls below the strike price but not below the barrier. This does not
cause the option to expire worthless. As the terminal value lies above the
strike price again, the payoff at maturity amounts to 104 – 90 = 14. Path
3 breaks the barrier after two and a half years. Consequently, the option
is knocked out from then on, and has a zero-payoff, even though the
underlying recovers and has a terminal value above the strike.
The aim is to price a barrier option with strike D(T) + K on the assets A:
C(t, A) = EQ[ · max{A(T) – D(T) – K, 0} · 1{τ>T} | Ft] (3), where C(t,
A) is option price of the barrier option with underlying A at time t and τ
is the time of default of the option, i.e., the first time the asset process A
crosses (or reaches) the barrier D. It is modeled as a stopping time on the
interval (t, T]: τ = inf{t’∈(t,T] : A(t’) ≤ D(t’)} (4).
The symbol EQ denotes the expected value under the arbitrage-free
measure Q. Equation (3) describes a down-and-out call option with un-
derlying A, strike D(T) + K and knock-out barrier D.1
Whereas, for K=0 the value of the option can be shown to fulfill the
straightforward formula C(t,A) = [A(t) – D(t)]1{τ>t} using optional sampling
theory, for the general case K≥0, the evaluation of the option pricing for-
mula (3) is done inspired by the approach of Sepp (2006):
Proposition 1: The price of the down-and-out call option (3) is C(t, A) =
C(t, A, v) = e½x+½a · G(T - t, x - a, v) · D(t) (5), where a = ln [(D(T) + K)/
(D(T))], x = ln [A(t)/D(t)] and G(T-t, x-a, v) can be evaluated via G(T-t, x-a,
v) = e½(x-a) - e-½(x+a) – 1/π ∫0∞[eα(T-t, k) + β(T-t, k)v [cos((x-a)k) – cos((x+a)k)]/
k2 + ¼] dk, with
β(T-t, k) = -[k2 + 1/4] · (1 – e-(T-t)ζ)/(ψ- + ψ+ e-(T-t)ζ)
α(T-t, k) = -κvv∞/ε2v · [(T-t) ψ+ + 2ln[(ψ- + ψ+ e-(T-t)ζ)/2ζ]]
ψ± = ±κv + ζ
ζ = [κ2v + ε2
v · (k2 + 1/4)]1/2
Note that solving (3) for K = 0 as a special case also results in C(t, A) =
[A(t) – D(t)]1{τ>t}.
The model implied for the equity (5), which is an explicit function not only
of the assets but also the volatility at time t, is quite rich. As a direct result
from Ito’s lemma the presence of stochastic volatility on the equity can be
shown. Moreover, due to (5) being a function of v(t), a correlation between
the equity and the equity’s volatility can be observed. This correlation
could even be stochastic. For example, if we denote the log-equity as R
= logC then the variance of R, the stochastic term in the variance of R,
denoted as vR, and the correlation between R and vR, denoted ρR,vR, are:
dR = (…)dt + ∂R/∂lnA · dZv + ∂R/∂v · εv dZv = (…)dt + dZR.
dvR = (…)dt + [g + v(∂g/∂v)] εv dZv + v(∂g/∂lnA) dZ
ρR,vR = [v2(∂g/∂lnA)(∂g/∂lnA) + (g + v · ∂g/∂v) ε2
vv · ∂R/∂v] ÷
[ (6)
respectively, where ZR is a Wiener process in (Ω, F, F, Q) and g(lnA, v) =
(∂R/∂lnA)2 + ε2v(∂R/∂v)2
The stochastic part of vR depends not only on the Brownian driving the
volatility of the assets but also on the Brownian driving the asset itself
due to the fact that g is dependent on both v and lnA. Figure 2 shows a
plot of the correlation ρR,vR as a function of the log-asset and the volatility
of the asset for the following set of parameters: µ = 0.075, v∞ = 0.01 = κv
= 0.75, εv = 0.1, D(0) = 4 (where the asset is assumed to be in the range
[4, 12] at t=0), K = 0.2, r = 0.03, T = 1, v(0) = [0.002, 0.02].
Sepp (2006) prices an option on the equity (which he denotes as S) which itself is modeled 1
as an option on the assets A. This so-called compound option has similar characteristics
as (3).
70
80
90
100
110
120
130
0 1 2 3 4 5
asset price
time (years) path 1 path 2 path 3 strike barrier
Figure 1 – Simulated paths in the stochastic volatility model.
126
The correlation in Figure 2 is negative which further support the useful-
ness of the model. Recall negative correlation is called leverage effect
and accounts for smile or the skew structure of the volatility implied by
options prices.
Probability of defaultIn this framework the default of a company for a fi xed period of time (t, T)
could be triggered by both the behavior at maturity T as well as the path
of the assets before maturity. Consequently, the probability of default,
denoted P(t, T) can be represented by the following expectation: P(t, T) =
1 – EQ[1{A(T)-D(T)-K, 0} · 1{τ>T} | Ft] (7).
A formula for the risk-neutral probability of default can be derived similar
to the price of a barrier option under deterministic interest rates. As (3)
has an additional discount factor under the expectation in con-
trast to (7), we need an additional factor to arrive at the result for
(7). The result is provided next.
Proposition 2: The risk neutral probability of default is:
P(t, T) = P(t, T, v) = 1 - [e1/2x+1/2a · Gp(T – t, x –a, v) · D(t)] (5), with
a, x as in Proposition 1 and where Gp(T – t, x –a, v) can be evaluated via
[D(T) + K]Gp(T – t, x –a, v) = [[1/π∫0∞
[eα(T-t, k) + β(T-t, k)v (sin((x-a)k)+sin((x+a)
k)] ÷ k2 +1/4]kdk + [1/2π∫0∞
[eα(T-t, k) + β(T-t, k)v (cos((x-a)k)+cos((x+a)k)] ÷
k2 +1/4]dk
with β(T - t, k), α(T - t, k) as defi ned in Proposition 1.
Calibrating the modelEquity markets allow us to observe the value of company’s stocks with
high frequency. However, we only have scarce and inaccurate informa-
tion on the daily value of the assets of a company. Thus, calibrating the
asset process (1) of a company has the diffi culty that this process itself is
not observable. Fitting the parameters involves two steps. First, inverting
the option pricing formula (5) to solve for A subject to the parameters v∞,
κv, and εv and the actual fi tting of the parameters to the asset process.
Other than in the setting with k = 0, the parameters which are to be fi tted
are not only used to describe the asset process, but also for estimating
the asset process from the equity process via the option pricing formula.
Solving this problem by standard optimization algorithms would be too
time consuming, because every step of fi tting the asset process would
require a recalculation (subject to the parameters of the current fi tting
step).2 Consequently, we propose a recursive algorithm which is evenly
stable and robust, yet much more effi cient.
Recursive fi tting algorithmThe fi tting method to estimate the parameters Θ = (µ, v∞, κv, and εv) is a
recursive methodology involving as many steps as we have data points
for the time series. The time series considered are growing with each
step. For step i, the analyzed equity time series consists of i data points.
In order not to have too short a time series, we wait for the fi rst 100 data
points to start the estimation procedure.
The algorithm is divided into several parts assuring both accuracy and
effi ciency at the same time.
(A) Every single step, the following is done: we calculate the new asset
data point applying the option pricing formula (5) and the parameters
of step i-1 and with the new asset time series, estimate the new pa-
rameters Θ̂ along the lines of Escobar et al. (2010).
(B) Every n(B) steps, the asset time series is updated completely, with
the just fi tted parameter vector Θ̂ (i) the entire asset time series is
updated, i.e., all values are calculated according to the option pricing
formula (5) subject to Θ̂ (i).
(C) Every n(C) steps, Θ̂ (i) is refi ned according to a grid, from every single
point in the grid, Θ̃ (i) is fi tted. That set of parameters Θ̃ (i) is chosen
as new Θ̂ (i), which minimizes the error. That way, we prevent the al-
gorithm from terminating in local minima, which could also be due to
numerical issues.
The reason why (B) and (C) are not done every single step (although that
would be desirable) is simply a matter of computing time. As already
Actually, for a time series of 10,000 data points we expect to wait a man’s lifetime to get 2
the fi nal parameters.
6,53
8,63
10,74
-1,0000
-0,8000
-0,6000
-0,4000
-0,2000
0,0000
0,001 0,001
0,006
0,010
0,015
Correlation between log-equity and its volatility
Asset value at time t. Volati4,42 Asset value at time t. Volati4,42 0,001
Asset value at time t. Volati0,001
lity of asset at time t
Figure 2 – Correlation of log-equity and equity-volatility versus log-assets and assets-volatility.
127
The Capco Institute Journal of Financial TransformationA General Structural Approach For Credit Modeling Under Stochastic Volatility
mentioned, the error function is very sensible (and thus smooth) in µ and
ν∞, yet very rough (due to numerical reasons) in κv and εv. In order to pre-
vent the algorithm from being stuck in a local minimum resulting from this
“unsmoothness,” the grid overcomes this risk. In our analysis, we choose
n(B) = 50 and n(C) = 50, which then requires approximately 10 hours for
one time series of 10,000 data points.
Estimating the parametersFor estimating the parameters we use the estimators we derived in Esco-
bar et al. (2010) based on the work by Genon-Catalot et al. (2000).
In the following, a recovery test is applied in order to numerically validate
the above-described method for fitting the parameters. The general idea
behind it is to simulate several time series given a known set of param-
eters Θ = (µ, v∞, κv, and εv) as well as the information on the debt, D and
K. From the simulated asset time series, the according equity time series
are calculated applying the option pricing formula (3). Given the equity
time series as well as the information on the debt, the parameters of the
underlying (yet unobservable) asset process are estimated. The so recov-
ered parameters should then be equal to the initial parameters with which
the time series had been simulated.
Recovering the parameters with kappa knownWe first assume the parameter κv to be known. To start with, one single
time series is analyzed in detail. An asset time series has been simulated
with the following parameters: µ = 0.075, v∞ = 0.01 = κv = 0.75, εv = 0.1,
D(0) = 4 (where the equity is assumed to have a value of 1 at t = 0)3, K =
0.2, r = 0.03, T = 5, Δ = 1/252 (representing daily data) n = 10,000 (number
of data points, i.e., approximately 40 years of data)).
Figure 3 visualizes the evolution of the fitted parameters in every step
from 100 to 10,000. The test aims to recover the parameters Θ = (µ, v∞,
κv, εv), assuming κv to be known. The first 100 steps are not included in
the graphs as the estimation starts after 100 data points.
The resulting final parameters are the following: µ̂ = 0.05755, v̂∞ =
0.01107, κ̂v = 0.75, ε̂v = 0.1010
As this is only one single time series, which is examined here, it does not
come as a surprise that the true parameters were not matched perfectly.
Yet, the fitted parameters are very close to the true parameters. Having
a look at the evolution of µ̂ (i), v̂∞(i), and ε̂v(i) in Figure 3, one can see that
the parameters are nicely converging as they scatter less, yet at the same
time accounting for the new information added by the new data points. In
order to emphasize this statement, Figure 4 shows that the standard de-
viation of the fitted parameters µ̂ (i) up to step i is decreasing. Of course,
one single time series does not allow for drawing any conclusions on the
quality of the method. Consequently, we simulated 250-asset time series
of 10,000 data points each with the same parameters as above. In one
case, we calculated the equity from the simulated assets with K = 0.2,
and in the other case with K = 0.05. Table 1 summarizes the outcomes
of this test. The final parameters (i.e., in step 10,000) match the initial
parameters very well.4
From top to bottom, the table gives information about the true parameters
for Θ = (µ, v∞, κv, εv), the means of the fitted parameters, the standard
0,0
0,2
0,4
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
mu
0,005
0,010
0,015
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
v_inf
0,0
0,5
1,0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
kappa
0,0
0,1
0,2
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
eps
Figure 3 – Estimated model parameters over time (assuming κv known).
0,00
0,02
0,04
0,06
0,08
0,10
0,12
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
standard deviation
Figure 4 – Standard deviation of estimated µ̂ (i) up to step i.
An equity value of one together with a debt of four implies an approximate value of five for 3
the asset. This scenario is a particular case of Figure 2.
Note that for calculating the statistics, we did not consider those time series of which one 4
of the parameters lies in the top or bottom 2 percent of the range of this parameter in order
to exclude outliers.
128
deviation, the 5 percent and 95 percent confidence interval (derived from
that standard deviation assuming normally distributed parameters), and
the median. As expected, the difference between K = 0.05 and K = 0.2
is negligible, indicating that the value of K does not have an influence on
the quality of the fitting method.
Recovering the parameters with kappa unknownOf course, in reality, we do not know the true value for κv. The difficulty is
that two parameters, κv and εv, are describing the variance process of the
assets. The variance process is, of course, not observable, and even the
asset process itself which it describes is not observable (only the equity
time series is). Thus, it is a challenge to accurately capture both of these
“twice-unobservable” parameters.
Figure 5 shows the evolution of the fitted parameters Θ̂ over the same
time series as in Figure 3, yet now with κv unknown. The two parameters
µ̂ and v̂∞ evolve almost exactly the same as with κv known with their final
values deviating less than 0.1 percent in relative terms. Yet, the picture
looks different for κ̂v and ε̂v. The final values for those two parameters
are 0.492 and 0.0818, around which they scatter mostly, yet with a high
noise (as compared with µ̂ and v̂∞). The higher noise is due to the fitting
being much more sensitive to the parameters κv and εv than to the other
two for the reason just mentioned above. Other than that, κ̂v often hits its
boundary which we set to the value 10 (relaxing that boundary to higher
values made the κ̂v(i) hit that higher boundary).
In order to overcome the shortcoming of the parameters κ̂v(i) and ε̂v(i)
suddenly hitting the boundaries and scattering widely, we impose a pen-
alty term to the error function, which is to be minimized (namely Σ, the
K = 0.05
µ̂ v̂∞ κ̂v ε̂v
True parameter 0.0750 0.0100 0.7500 0.1000
Mean 0.0760 0.0098 0.7500 0.0915
Standard deviation 0.0138 0.0018 0.0000 0.0135
5% conf. 0.0533 0.0068 0.7500 0.0693
Median 0.0766 0.0096 0.7500 0.0910
95% conf. 0.0987 0.0127 0.7500 0.1138
K = 0.2
True parameter 0.0750 0.0100 0.7500 0.1000
Mean 0.0761 0.0097 0.7500 0.0915
Standard deviation 0.0137 0.0018 0.0000 0.0135
5% conf. 0.0535 0.0067 0.7500 0.0692
Median 0.0768 0.0096 0.7500 0.0912
95% conf. 0.0987 0.0128 0.7500 0.1138
Table 1 – Recovering the parameters with kappa known
0,0
0,2
0,4
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
mu
0,005
0,010
0,015
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
v_inf
0
5
10
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
kappa
0,0
0,2
0,4
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
eps
Figure 5 – Estimated model parameters over time (including κv).
0,0
0,2
0,4
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
mu
0,005
0,010
0,015
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
v_inf
0,0
1,0
2,0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
kappa
without penalty term with penalty term
0,0
0,1
0,2
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
eps
without penalty term with penalty term
Figure 6 – Estimated model parameters over time (including κv, with and without penalty).
The graphs represent the estimations for µ, v∞, κv, and εv respectively. The last two
graphs (for κv and εv) show the estimation with penalty (smooth curve) and without penalty
(random-like curve). Both with and without penalty are the same in the first two graphs.
sum of squared deviations between theoretical and empirical estimators),
preventing large jumps in the evolution of the single parameters. Σ’(i) =
Σ(i) + ψκ · [κ̂v(i) - κ̂v(i -1)]2 + ψε · [εv(i) – εv(i-1)]2.
We examined various combinations for ψκ and ψε. Of course, the higher
the values that are chosen the more “phlegmatic” the parameters behave,
and, conversely, the lower the values the faster they adjust. We found ψκ =
0.05 and ψε = 0.5 to be reasonable values. Figure 6 compares the original
evolution of the parameters Θ̂ (from Figure 5) with the evolution imposing
129
the penalty. The differences can be observed only in the last two graphs,
where the smooth curve represents the estimation with penalty and the
random-like curve shows the estimation without penalty. Different scales
are used in the last two graphs of Figures 5 and 6 in order to clearly show
the smooth estimators. The figures show that the parameters µ̂ and v̂∞ do
not deviate from the original parameter evolution – irrespective of which
penalty is chosen for κv and εv (note that no penalty is set for the two pa-
rameters µ and v∞). For κv and εv we observe that the evolution is smooth-
er and these parameters do not exhibit jumps any more.
Table 2 summarizes the test for the 250 time series for which the pa-
rameters have been recovered, including κv.5 Comparing the values for
µ and v∞ with those of Table 1, one can barely tell the difference. These
two parameters could be recovered just as if κv is known. However, now
that κv and εv are targeted and both describe the unobservable variance
process of the again unobservable asset process, ε̂v scatters with a big-
ger standard deviation than before, yet still very well around the true pa-
rameter. For κ̂v we can assess a skewed parametric distribution with a
median of 0.84, very close to the true 0.75. Thus, we can conclude that
the proposed fitting algorithm proved its quality in the presented complex
problem, with its biggest advantage being that it makes problem compu-
tationally manageable.
Case study: Merrill LynchAs an application of the model, we want to examine the downfall of the
Wall Street investment bank Merrill Lynch. Founded by Charles Merrill in
1914, Merrill Lynch had always been an icon on Wall Street with its stock
performing strongly over decades. Yet, the investment bank could not
survive the financial crisis in which it suffered billions of losses from its
activities in the MBS and subprime markets. Merrill Lynch was taken over
by the Bank of America in September 2008 and has been part of Bank of
America since January 2009. Figure 7 displays the stock price of Merrill
Lynch stocks quoted on the New York Stock Exchange (NYSE) over the
five years from 2003 until 2007, i.e., prior to the takeover.
Using the method of calibration as described in Escobar et al. (2010)
to estimate the parameters for the Merrill Lynch time series between
2003 and 2007 for the presented model yields the following parameters
– assuming K = 0: µ̂ = 0.04135, v̂∞ = 0.0006383, κ̂v = 0.6002, and ε̂v =
0.02433.
To obtain the parameters above, we need the level of debt. This is ob-
tained by averaging the reported ratio between assets and equity for the
selected period. Scaling the equity time series for ease of comparison to
start with the value 1 in 2003, we get D(0) = 12.6. Furthermore, the matu-
rity T of the modeled option is required to correspond to the maturity of
the liabilities. As the information is not provided in the annual reports, we
assumed T = 5, but found that the parameters are hardly sensitive to the
choice of maturity. In accordance with this assumption, we used a risk-
free rate of 3.94 percent, which is the average of the five-year treasury
rate over the years 2003 until 2007.
We next examine the sensitivity of the equity price and the probability of
default towards K. These analyses are performed using the estimators
for Θ = (µ, v∞, κv, εv) as well as the values for [T, D(0), r] and equity pro-
vided above. We resort to two scenarios, first constant total debt D(0) +
K and varying K is assumed. A second scenario assumes constant debt
D(0) and varying K. In the first scenario K would represent the additional
debt limit on the cash account, while the total debt is assumed constant
and the assets remain the same, i.e., the parameters describing the as-
set process do not change subject to K. Figure 8 plots the equity price
subject to K. We can observe a monotonously increasing equity value as
K increases. However, the slope of the curve is decreasing, i.e., granting
The Capco Institute Journal of Financial TransformationA General Structural Approach For Credit Modeling Under Stochastic Volatility
K = 0.2
µ̂ v̂∞ κ̂v ε̂v
True parameter 0.0750 0.0100 0.7500 0.1000
Mean 0.0762 0.0096 1.4831 0.1148
Standard deviation 0.0140 0.0018 1.4479 0.0590
5% confidence 0.0531 0.0067 0.0000 0.0177
Median 0.0767 0.0094 0.8410 0.0983
95% confidence 0.0992 0.0126 3.8647 0.2119
Table 2 – Recovering the parameters with penalty function
30
40
50
60
70
80
90
100
2003 2004 2005 2006 2007 2008
Figure 7 – Stock price of Merrill Lynch on NYSE (2003-2007).
Again, for calculating the statistics, we did not consider those time series of which one of 5
the parameters lies in the top or bottom 2 percent of the range of this parameter. Prior to
that, we excluded those time series where the boundary was hit (which was the case for
18 of the 250 cases). In those 18 cases, v did not jump to the boundary, but monotonously
increased until it reached the boundary.
130
a higher cash limit is more effective for the first dollar than the second.
If the company holds, for example, 1 percent of its market capitaliza-
tion as a cash liability (which would not cause an immediate default as
it is the case for the “regular” liabilities), the equity value increases from
1 to 1.0031 if the assets of the company stay the same. If a cash limit
of 10 percent is granted (and used instead of taking “regular” debt) the
equity value increases to 1.0289. And if (consider that as an illustrative,
not realistic case) this debt would be of the same size as the equity value
(in the case K = 0) itself, the equity value would increase to 1.1990. These
findings can also be explained theoretically: granting a higher limit on the
company’s cash account simply lowers the default barrier while leaving
the total debt untouched. This corresponds to a lower barrier line in the
illustration given in Figure 1. Lowering the barrier simply decreases the
number of paths being knocked-out and expiring worthless. Thus, the
lower the barrier the higher the value of the option price.
Figure 9 plots the equity price as a function of K under the second sce-
nario, that of constant debt. The figure shows a steep decrease of the
equity price when additional debt is taken in the form of a cash account.
This can also be observed from Equation (3), an increase in K implies
lower chances for the assets to be above the new total debt at maturity
[D(T) + K], therefore decreasing the expected value. Note the indicator
term in Equation (3) does not change because of the assumption of con-
stant debt.
The analysis of the probabilities of default is performed first under the
0,95
1,00
1,05
1,10
1,15
1,20
1,25
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
relative equity price
value for K
Figure 8 – Relative change in equity price subject to K. Assuming D+K constant.
0,22
0,23
0,24
0,25
0,26
0,27
0,28
0,29
0,30
0,31
0,32
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
probabilitiy of default
value for K
Figure 10 – Change in probability of default subject to K. Assuming D+K constant.
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
relative equity price
value for K
Figure 9 – Relative change in equity price subject to K. Assuming D constant.
0,30
0,35
0,40
0,45
0,50
0,55
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0
probabilitiy of default
value for K
Figure 11 – Change in probability of default subject to K. Assuming D constant.
131
The Capco Institute Journal of Financial TransformationA General Structural Approach For Credit Modeling Under Stochastic Volatility
first scenario. In this scenario, a decrease in the probability of default is
observed when plotting this probability versus K, see Figure 10.
This decreasing behavior of the probability of default could also be im-
plied from Equation (7). In this scenario the total debt remains constant
therefore the probability of defaulting at maturity remains the same for
all K. On the other hand, an increase in K leads to a decrease in “regular
debt” D(t) hence lowering the probability of default prior to reaching ma-
turity. This insight favors the use of K as an alternative to taking regular
debt.
The second scenario is explored in Figure 11. Here K is taken as addi-
tional debt so the total debt D(0) + K increases. This implies an increase
in the probability of default. This can be observed from Equation (7), an
increase in the total debt means an unchanged probability of default be-
fore maturity but a higher probability at maturity. This shows the risk in
taking additional debt in the form of a cash account instead of transform-
ing regular debt using a cash account.
A question, which might be even more interesting, is how the estimated
parameters change if we still observe the same equity time series and
assume different values for K while total debt [D(0) + K] is kept constant.
This answer is provided in Table 3 giving the estimated parameters µ̂ , v̂∞,
κ̂v, ε̂v, subject to K. From this table, we can make the following obser-
vation: due to the fact that an increase in K would lead to higher equity
values (Figure 8), the estimated drift of the asset process should mo-
notonously decrease with K. Intuitively, if we now fit the parameters to
the same equity time series (but increased K), the assets do not have to
perform that strongly as before to result in the same equity values. This
decrease in drift gives room for an increase in volatility. Note that the
long-term mean v∞ of the variance of the asset process is monotonous-
ly increasing for higher values of K. All other parameters are calibrated
around these relationships.
Examining the parameters κ̂v and ε̂v, we cannot observe a monotonous
trend as we did for the other two parameters µ̂ and v̂∞. However, we
have to bear in mind that these parameters are much harder to capture
because they are describing the variance process, and that we only have
a comparatively small dataset consisting of five years of daily data to cali-
brate the parameters. Having said that, a general decrease in the param-
eter κ̂v can be assessed, whereas we cannot observe a clear trend for ε̂v.
The trend in κ̂v can also be explained by the nature of the model. A higher
κv means that the variance is returning faster to its long-term mean v∞.
Thus, for low values of κv, the likelihood of an extremely high variance
increases and so does the likelihood that the option matures worthless.
This is perfectly in line with the argument raised for µ̂ and v̂∞.
We have studied the implications of assuming K>0. In general, a positive
K could be beneficial if it replaces regular debt as it reduces the probabil-
ity of default (Figure 10) and increases the equity value (Figure 8). On the
other hand, taking a positive K at the expense of increasing the overall
debt leads to the opposite scenario hence a higher probability of default
(Figure 11) and a lower equity value (Figure 9). Unfortunately, K cannot
be properly estimated from a time series of equity and debt values due
to a problem of identifiability, and should thus be obtained from addi-
tional information from the company. In fact, if we examine the annual
reports of Merrill Lynch, we learn that the short-term borrowings (divided
by the total stockholder’s equity) increase in upwards-sloping markets
and decrease in recessions. Over the calibration period 2003-2007, the
short-term borrowings average to a value of approximately 0.4. Let us
assume, for an expository sensitivity analysis, that K can be represented
by the short-term borrowings and apply the parameters in Table 3 for K
= 0.4. If we furthermore assume that the total liabilities remain constant,
but K is eaten up totally by a market crash, Proposition 2 tells us that the
probability of default increases from 25.6 percent to 32.96 percent and
the equity value drops by 8.87 percent.
ConclusionThis paper continues the work of Escobar et al. (2010) where we com-
bined the Heston model with the Black/Cox framework providing a model
for the company’s asset process if only the equity process is observable.
The assets follow a Heston model, and the company’s equity value is
K µ̂ v̂∞ κ̂v ε̂v
0 0.04135 0.0006383 0.6002 0.02433
0.01 0.04134 0.0006393 0.5869 0.02417
0.02 0.04132 0.0006411 0.5774 0.02403
0.03 0.04131 0.0006429 0.5872 0.02424
0.04 0.04130 0.0006446 0.5795 0.02405
0.05 0.04128 0.0006463 0.5758 0.02409
0.06 0.04127 0.0006480 0.5607 0.02379
0.07 0.04125 0.0006496 0.5610 0.02383
0.08 0.04124 0.0006512 0.5602 0.02386
0.09 0.04122 0.0006523 0.5558 0.02374
0.1 0.04121 0.0006544 0.5511 0.02371
0.2 0.04104 0.0006681 0.5348 0.02360
0.3 0.04086 0.0006792 0.5160 0.02341
0.4 0.04067 0.0006883 0.5000 0.02322
0.5 0.04047 0.0006955 0.5023 0.02341
0.6 0.04026 0.0007016 0.5022 0.02350
0.7 0.04005 0.0007066 0.5152 0.02398
0.8 0.03983 0.0007101 0.5044 0.02383
0.9 0.03962 0.0007148 0.5022 0.02370
1 0.03940 0.0007183 0.4940 0.02365
Table 3 – Estimated parameters for different values of K
132
modeled as a barrier call option on the assets. In contrast to our previous
work, we now allow the barrier and strike price of the option to be differ-
ent from each other and thus for richer financial applications. We present
a closed-form solution of the option price, the equity, in this model, which
allows for most stylized facts on the equity process.
This much more complex option pricing formula, including an improper
integral, demonstrates the difficulties one often comes across. The com-
putational effort to optimize a set of parameters is beyond available com-
putational power, because in this case it requires the evaluation of the
formula (and thus also the integral) in every iteration of the optimization
for every point of the time series, which is calculated as the inverse of this
option pricing formula. We present a method to overcome this problem,
which drastically limits the number of times the option pricing formula
has to be evaluated. This method could also be transferred to various
financial applications, for example: in the reduced form credit framework
where the unobservable intensity is the one modeled and calibrated
based on observable credit derivative prices; for fixed income products
where bond prices are observed but the unobservable instantaneous in-
terest rate is the one being modeled and therefore calibrated; or in cases
where the implied volatility is targeted and the observable values come
from option prices.
Besides numerical validation of the convergence of the proposed fitting
algorithm, we provide a possible application of the model with Merrill
Lynch as an example. We show the sensitivities of the model parameters
subject to the relationship between the barrier and the strike price and
give a theoretical interpretation of this behavior.
ReferencesBlack, F. and J. Cox, 1976, “Valuing corporate securities: some effects of bond indenture •
provisions,” Journal of Finance, 31:2, 351-367
Cont, R., 2001, “Empirical properties of asset return stylized facts and statistical issues,” •
Quantitative Finance, 1, 223–236
Cox, J., J. Ingersoll, and S. Ross, 1985, “A theory of the term structure of interest rates,” •
Econometrica, 53:2, 385-408
Dewynne, J., S. Howison, and P. Wilmott, 1993, Option pricing, Oxford Financial Press•
Engle, R., 1982, “Autoregressive contidional heteroscedasticity with estimates of the variance of •
United Kingdom inflation,” Econometrica, 50:4, 987-1008
Escobar, M., T. Friederich, M. Krayzler, L. Seco, and R. Zagst, 2010, “Structural credit modeling •
under stochastic volatility,” submitted to the Journal of Financial and Quantitative Analysis
Feller, W., 1951, “Two singular diffusion problems,” The Annals of Mathematics, 54:1, 173-182•
Genon-Catalot, V., T. Jeantheau, and C. Laredo, 2000, “Stochastic volatility models as hidden •
Markov models and statistical applications,” Bernoulli, 6:6, 1051-1079
Heston, S., 1993, “A closed-form solution for options with stochastic volatility with applications •
to bond and currency options,” Review of Financial Studies, 6:2, 327-343
Lipton, A., 2004, Mathematical methods for foreign exchange. A financial engineer’s approach, •
World Scientific Publications
Merton, R., 1973, “Theory of rational option pricing,” Bell Journal of Economics and •
Management Science, 4, 141-183
Rosenblatt, M., 1956, “A central limit theorem and a strong mixing condition,” Proceedings of •
the National Academy of Sciences, 42, 43-47
Seber, G. A. F. and C. J. Wild, 2005, Nonlinear regression, John Wiley and Sons•
Sepp, A., 2006, “Extended credit grades model with stochastic volatility and jumps,” Wilmott •
Magazine, September, 50-62
133
PART 2
A Stochastic Programming Model to Minimize Volume Liquidity Risk in Commodity Trading
AbstractThe goal of this paper is to study a very important risk metric
in commodity trading: volume liquidity risk. It begins by ex-
amining the statistical properties of volume and settlement
price change of futures contracts of different maturities. The
results are used in the construction of a model for the mini-
mization of volume liquidity risk – the inability to cover an un-
profitable position due to lack of trading volume. The model
is embedded in a stochastic program designed to construct
a portfolio of futures contracts of different maturities with
the aim of minimizing price and volume liquidity risk. The
results of the case study (grain market) show that the model
predicts the best spread trade accurately in 75 percent of
cases. In the remaining cases the inaccuracy is due to the
market shock present in the year 2008. A tool has been cod-
ed in Excel VBA to make the model available to traders and
risk managers. This contribution directly relates to Energy
ETF recent issues (i.e., roll-over).
Emmanuel Fragnière — University of Bath School of Management, and Haute Ecole de Gestion de Genève
Iliya Markov — School of Mathematics, University of Edinburgh
134
Liquidity risk is already an important factor in the management of finan-
cial risk. The current global economic crisis emphasized the need to in-
clude liquidity risk in risk management models. The financial meltdown of
2007-2008 saw the freezing of the markets for commercial paper, asset-
backed securities, and collateralized debt obligations among many oth-
ers. Various risk models were proposed that try to deal with liquidity dry-
ups in the financial markets in different ways – estimation of probability,
pricing of liquidity risk, etc. [Pedersen (2008)]. Liquidity in the commodity
sector, on the other hand, has not been given a proper academic treat-
ment. The main trading products in the commodity sector are futures
contracts. Traders usually hedge risk by entering into contracts of dif-
ferent maturity and opposing direction. Distant maturities, however, are
associated with very low trading volume, which means that an already
established position may not be closed or changed due to lack of trading
volume. Thus, if an already established spread trade is found to be un-
profitable, a trader may not be able to cover it because of what this paper
refers to as volume liquidity risk, or simply volume risk.
This paper sets out to explore the patterns and relationships associated
with volume and settlement price change and to build a stochastic pro-
gram whose purpose is the construction of a portfolio where the risk of
a volume liquidity trap is minimized or avoided. The use of stochastic
programming is grounded on the uncertainty of the prices of distant ma-
turities on the forward curve. The ultimate purpose of the paper is that
the stochastic model should be applicable to real life situations. Conse-
quently, it was developed in VBA in Excel, which is the most widely used
software in the commodity sector. All data used in this paper is from the
Kansas City Board of Trade [KCBT (2010)], the world’s largest exchange
for hard red winter wheat.
Literature reviewThe purpose of the literature review below is to position the paper in the
context of the three main academic areas with which it is concerned.
First and foremost, it discusses the lack of academic literature dealing
with the concept of volume risk. Second, it gives a brief overview of sto-
chastic programming and the reasons why it was chosen as the portfo-
lio optimization technique. Third, it explains what past statistical results
concerning futures related time series this paper corroborates and what
contributions it makes.
Volume risk To the best of our knowledge, currently there is no academic literature
that deals with the problem of volume liquidity risk as outlined above.
Table 1 is a demonstration of the volume liquidity of certain commodity
and financial futures traded at various boards throughout the world.
Table 1 lists the futures according to their liquidity. The more liquid the
contracts, the easier it is to buy and sell. Interestingly, the hard red winter
wheat futures traded at the KCBT are last, which means that they are
relatively illiquid. Consequently, the results obtained in subsequent sec-
tions should be especially useful to traders at the KCBT. The analysis and
conclusions, however, are applicable to any commodity trading board.
An important contribution of this paper is the fact that the analysis is fo-
cused on commodity instead of financial futures. The good performance
of the financial markets until a few years ago led to a lateral treatment of
the commodity sector in academic literature. Partly due to the financial
meltdown, however, recent years have seen the increased importance of
the precious metals, oil, and grain markets. The increased participation
of hedge funds in the commodity markets is another sign of their growing
importance.
There are numerous papers that study positions optimization and the vol-
umes associated with different maturities on the forward curve. The two
concepts, however, are never integrated and volume risk is never taken
into account when establishing futures positions. Boyd and Kaastra
(1995), for example, devise a model to forecast futures trading volumes
of different maturities at the Winnipeg Commodity Exchange. Their mod-
el is an application of a neural network, an artificial intelligence method
that acts like a human brain in finding patterns. The neural network uses
a gradient descent algorithm and is able to provide predictions of trading
volumes up to nine months into the future. Its predictive power is better
than both the discussed naïve model and the autoregressive integrated
moving average model. De Roon and Veld-Merkoulova (2003) discuss
long-term hedging strategies when only the first few contract maturities
of a given commodity are actively traded. Instead of focusing on volume
risk, however, they construct a hedging strategy using futures conve-
nience yields that minimizes spot-price risk and rollover risk by using
futures contracts of different maturities. This paper’s main contribution,
therefore, is the introduction, explanation, and modeling of volume risk in
Commodity futures Board Relative contract liquidity
Eurodollar interest rate CME ••••••••••••••••••••••••••••••••••••••>>
three-month Euribor interest
rate
LIFFE ••••••••••••••••••••••••••••••••••••••>>
Short sterling LIFFE ••••••••••••••••••••
30-day Federal funds CBT •••••••••••••••••••
3-Year Commonwealth t-bonds SFE •••••••••••••••••••
Mini S&P 500 index CME •••••••••••••••
Crude oil – light sweet NYM •••••••••••••
Silver 5000 troy ounce NYM •
Japanese Yen ¥ CME •
British pound £ CME •
Wheat – hard red KCBT •
Table 1 – Relative contract liquidity [Futures liquidity – June 2010, (2010)]
135
The Capco Institute Journal of Financial TransformationA Stochastic Programming Model to Minimize Volume Liquidity Risk in Commodity Trading
the commodity futures markets – a problem that has not been rigorously
tackled before.
Stochastic programmingThis paper employs a stochastic programming approach in the devel-
opment of a model that determines portfolio positions with the aim of
minimizing price and volume risk. Fragnière and Gondzio (2005) demon-
strate why stochastic programming is the preferred optimization meth-
odology in the presence of uncertainties about the future. In the case of
deterministic modeling, a separate optimization is performed for each
scenario. In the end, the decision maker knows what the optimal deci-
sion is under each scenario. But not knowing which scenario will unfold
makes this information unusable. Stochastic programming combines all
the uncertain future scenarios in a single tree structure and the system
is then optimized with the aim of finding the optimal decision given all
possible scenarios.
Fragnière et al. (2010) explain the description of future conditions by
means of a multi-period stochastic tree, where the branches, referred
to as scenarios, represent the uncertainties in the future. An alternative
approach is the generation of scenario sample paths, where the only
branching point is the origin. The latter approach circumvents the prob-
lem of dimensionality when the number of stages is significant, a problem
known as the “curse of dimensionality” [Fragnière et al. (2010)]. Since the
number of effective decision stages in this paper’s stochastic program
is three, the stochastic tree structure is adopted because it is thought to
better represent the trader’s decision process. The trader makes deci-
sions depending on what the forward curve structure is. Since distant
maturities are associated with great uncertainty, while close maturities
converge to the spot price, the system is best represented by a stochas-
tic tree that branches out with maturity.
Statistical analysisThere is a lot of academic literature that deals with the relationships
among futures-related time series. The most profoundly researched re-
lationship is the one between volume and price variability (or return from
another point of view). Grammatikos and Saunders (1986) find a strong
positive relationship between trading volume and price volatility. More-
over, maturity has a strong effect on volume, but no effect on price vola-
tility. Karpoff (1987) finds a strong positive relationship between volume
and both the absolute value of price changes and price changes “per
se” in equity markets. He builds a simple volum-price change relation-
ship model. Wang and Yau (2000) also establish a positive relationship
between volume and price change.
This paper finds a strong positive relationship between volume and the
intraday price volatility, but no relationship between volume and settle-
ment price change. Also, there is strong evidence of an impact of maturity
on volume but no evidence of its impact on settlement price change. The
distributions of settlement price change for the ten future maturities are
almost identical. On the other hand, the paper finds a strong maturity ef-
fect on volume, with the distributions changing significantly with maturity.
The main advantage of the exploratory statistical analysis conducted in
this paper is its practicality. It is aimed at enhancing commodity trad-
ers’ knowledge about many of the important statistics pertaining to the
market of hard red winter wheat by adding rigor to their intuition. Even
though practicality takes priority over sophistication, all of the statistical
analysis is done rigorously.
Statistical analysisThis section is intended to provide an exploratory statistical analysis of
the first ten maturities of some of the most important time series that
futures traders face – volume and settlement price change. The first part
discusses some important seasonal patterns in the first ten maturities of
the time series. The second part analyses the relationships of volume and
settlement price change and studies the distributions of the time series.
The analysis and results should help place the concept of volume risk on
solid ground by analyzing volume in the context of other time series. In
addition, the results of the statistical analysis are directly used in the case
study of the stochastic model below. All of the analyses are performed
using futures trading data from 1 January 2005 to 1 June 2010.
PatternsThe analysis of the seasonal patterns of volume allows us to make the
following observations:
1. The average volume decreases strictly with maturity.
2. The standard deviation of the volume of the first maturity is always
lower than the standard deviation of the volume of the second matu-
rity. From then on, standard deviation decreases until the most dis-
tant maturity but not perfectly. It may have spikes.
3. There is a pattern in the distribution of volume with some delivery
months being above the average and some below the average: July
and December are above the average everywhere (except the ex-
treme back end of December); September is above the average only
for the nearest maturity and below the average everywhere else;
March and May are below the average everywhere with one excep-
tion for March; and May has the smallest average volumes of all ma-
turities.
4. There is also a pattern in the distribution of volume across the ma-
turity timelines of different months that the overall results are unable
to pick up. (1) The contracts expiring in March, May, and September
have large volumes of the first two maturities and thereafter volumes
decrease sharply. (2) The contracts expiring in July and December
have volumes that decrease much more gradually across the matu-
rity timeline. These also happen to be the contracts with the highest
volumes of all maturities.
136
5. The volumes of all maturities move in the same direction from one
delivery month to another. The pattern is as follows: March – de-
crease – May – increase – July – decrease – September – increase
– December – decrease – March. The last relation (December – de-
crease – March) does not hold for the two most distant maturity dates
(which could be due to insufficient data).
6. Standard deviation is not proportionate to average volume. The ratio
of the mean to the standard deviation of the volumes decreases with
each successive maturity on the maturity timeline (but not perfectly).
This is true both overall and for each delivery month. This means
that volumes of distant maturities are much more volatile around the
mean, which, of course, poses greater risk. The increased volatility
can be explained by the large number of zero values and the oc-
casional spikes to several tens or hundreds of traded contracts. It is
also noticeable that the volumes for some months are generally more
volatile around the mean than other months. For example July and
December have much more volatile volumes than May.
Given the findings above, let us turn to the shape of the volumes associ-
ated with the forward curve. Please refer to Figure 1 and 2.
As corroborated by the findings above, volumes associated with July
and December deliveries are always higher given any individual forward
curve and this is a rule rather than a coincidence. Thus, volumes for July
and December deliveries form “spikes,” or “humps,” in the forward curve
volumes. The shapes of forward curve volumes in Figures 1 and 2 il-
lustrate why volume liquidity risk is an important risk in commodity trad-
ing. The low trading volumes of the long term maturity contracts make
covering and reestablishing long term positions extremely difficult. The
model proposed in this paper tries to deal with volume liquidity risk in a
structured manner.
There are some interesting patterns associated with settlement price
change as well:
1. The standard deviation is much larger than the mean in absolute val-
ue. This would suggest that price change is very volatile around the
mean.
2. Looking at the overall results, the nearest maturity has a very slight
tendency to drop in price, while all other maturities have slight ten-
dencies to increase in price.
3. There is no visible seasonal pattern in the means. To be more precise,
means for different delivery months are not consistently and signifi-
cantly above or below the overall mean.
4. The standard deviations of the sixth maturity of March and May deliv-
eries are disproportionately high, which is also present in the overall
standard deviation.
5. The standard deviations for September, July, and December exhibit an
interesting pattern: July has increased standard deviations of maturi-
ties 2, 3 and 7, 8; September has increased standard deviations of ma-
turities 3, 4 and 8, 9; and December has increased standard deviations
of maturities 4, 5 and 9, 10. The sequences above look suspicious at
first glance but there could be an explanation. A quick mental exercise
reveals the following: the months associated with trading May as a
sixth maturity are March and April. The months associated with trad-
ing July as a second, third, seventh, and eighth maturity, September
as third, fourth, eighth, and nineth maturity, and December as fourth,
fifth, nineth, and tenth maturity are the same and they are December,
January, February, March, and April. Historical data from the Kansas
City Board of Trade reveals an enormous increase in volatility in the
beginning of 2008. Consequently, ignoring December in the two points
above, all of the months are in the beginning of the year.
Relationships and distributionsThis section analyzes some important relationships and approximates
volume and settlement price change by analytical distributions. The em-
pirical distributions are approximated with various degrees of precision
as described below. The results below help us in the construction of the
stochastic tree in the case study section. There are strong relationships
between the average maturity values of volume and other important se-
ries in commodity trading such as drawdown, the difference between
daily high and low and open interest. Settlement price change does not
0
1000
2000
3000
4000
5000
6000
7000
01/M
ar/0
9
01/M
ay/0
9
01/J
ul/09
01/S
ep/0
9
01/N
ov/0
9
01/J
an/1
0
01/M
ar/1
0
01/M
ay/1
0
01/J
ul/10
01/S
ep/1
0
01/N
ov/1
0
01/J
an/1
1
01/M
ar/1
1
01/M
ay/1
1
01/J
ul/11
Forwardcurve volumes
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
20000
01/D
ec/0
9
01/J
an/1
0
01/F
eb/1
0
01/M
ar/1
0
01/A
pr/
10
01/M
ay/1
0
01/J
un/1
0
01/J
ul/1
0
01/A
ug/1
0
01/S
ep/1
0
01/O
ct/1
0
01/N
ov/1
0
01/D
ec/1
0
01/J
an/1
1
01/F
eb/1
1
01/M
ar/1
1
01/A
pr/
11
01/M
ay/1
1
01/J
un/1
1
01/J
ul/1
1
Forwardcurve volumes
Figure 1 – Forward curve volumes, 9 Jan, 2009
0
1000
2000
3000
4000
5000
6000
7000
01/M
ar/0
9
01/M
ay/0
9
01/J
ul/09
01/S
ep/0
9
01/N
ov/0
9
01/J
an/1
0
01/M
ar/1
0
01/M
ay/1
0
01/J
ul/10
01/S
ep/1
0
01/N
ov/1
0
01/J
an/1
1
01/M
ar/1
1
01/M
ay/1
1
01/J
ul/11
Forwardcurve volumes
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
20000
01/D
ec/0
9
01/J
an/1
0
01/F
eb/1
0
01/M
ar/1
0
01/A
pr/
10
01/M
ay/1
0
01/J
un/1
0
01/J
ul/1
0
01/A
ug/1
0
01/S
ep/1
0
01/O
ct/1
0
01/N
ov/1
0
01/D
ec/1
0
01/J
an/1
1
01/F
eb/1
1
01/M
ar/1
1
01/A
pr/
11
01/M
ay/1
1
01/J
un/1
1
01/J
ul/1
1
Forwardcurve volumes
Figure 2 – Forward curve volumes, 23 Oct, 2009
137
The Capco Institute Journal of Financial TransformationA Stochastic Programming Model to Minimize Volume Liquidity Risk in Commodity Trading
exhibit a relationship with any of the series, not even volume. The re-
lationships between the average maturity values of volume on the one
hand and drawdown, daily high/low and open interest on the other hand
are very well described by power regressions. In all three cases the ad-
justed R-square is higher than 0.99.
The first six maturities of volume can be well approximated by analytical
distributions. Even though the best distribution is not Gamma in every
case, Gamma is usually the second or third best and has a slightly higher
AD. The seventh through tenth maturities cannot be well approximated
by analytical distributions. It is easily observed that the mode of the dis-
tribution moves closer and closer to zero with each successive maturity
due to the fact that distant maturities are associated with low volumes. In
other words, there is a strong maturity effect on volume.
The best distribution for each one of the maturities of settlement price
change is either logistic or log-logistic. Since the difference in AD be-
tween the logistic and the log-logistic distribution in every case is minute,
logistic distribution was fitted to all maturities. There is no evidence of a
maturity effect on settlement price change. The logistic distributions of
settlement price change are almost identical for all maturities.
MethodologyThis section is the crux of the paper. It explains how volume risk presents
itself – as the inability of a position to be covered or changed due to lack
of trading volume, which could lead to a monetary loss. It also explains
the interconnectedness of volume risk and price risk. A simple model
for quantifying volume risk is presented. This model is then included in
a stochastic optimization algorithm, whose aim is the construction of a
portfolio, which minimizes or avoids the risk of a volume liquidity trap.
The algorithm’s objective is the minimization of volume and price risk.
The minimization is subject to a number of core, optional, and mutually
exclusive conditions, which correspond to a trader’s preferences. These
include the establishment of a bull or bear spread, the specification of the
minimum and maximum number of contracts, etc. The predictions of the
stochastic program are tested in the case study section. The case study
employs results from the statistical analysis section.
Volume riskThe problem that this paper sets out to explore is a very important risk
associated with commodity trading. Volume risk is the risk that there
might not be sufficient volumes associated with distant maturities if a
trader decides to unwind a position. This could lead to a much larger loss
than the one associated simply with price changes. The problem is how
to quantify this risk.
Let us suppose a trader is long on the short term and short on the long
term (Figure 3). Assume the trader is making money on the short term
contract and losing money on the long term contract through the daily
account settlement. Further, assume that the trader wants to cover his
short position on the long term contract by going long on the same con-
tract. If he is unable to buy he will continue losing money on the contract.
Actually, there is a great chance that he will not be able to buy because
volumes in the back end of the forward curve are very low and many
times there is no trade at all. The loss that the trader will make depends
on when he will be able to cover his position, which depends on when
volume will be sufficient. Of course, there will not be any volume risk of
this kind associated with the first two or three maturities (depending on
the delivery month) because they have sufficient trading volume.
In order to quantify volume risk, we should have a subjective measure as
to what number of contracts traded on a given day is safe and does not
pose any risk. Then for any number of traded contracts lower than the
threshold level there will be risk.
Let SafeVol be the threshold level and V be the average number of con-
tracts actually traded and let V ≤ SafeVol. Let MaxPriceChangetok indi-
cate the price change in an unfavorable direction at a given confidence
level that the trader may experience from the time t when the cover has to
take place to the time ok when there will be sufficient volume associated
with this maturity. This time interval can be inferred from the statistical
analysis on volume presented above. The price change can be inferred
from the settlement price change distribution discussed above.
Then VolumeRisk = (SafeVol-V)
SafeVol × MaxPriceChange
tok
Model preliminariesFigure 4 is a diagrammatic representation of the forward curve that is
used in the stochastic model. In line with intuition, prices in the near end
are much more certain than prices in the back end. Every combination
of successive cases represents a possible combination of prices associ-
ated with the successive maturities. It should be mentioned here that
the maturities on the diagram do not coincide with the maturities dis-
cussed above. This is a simplified construction, where the first maturity
represents the short term maturities, the second maturity represents the
medium term maturities, and the third maturity represents the long term
short term long term
long position short position
Cash in
Cash out
Figure 3 – Futures account settlement
138
maturities of a real forward curve.
The price change distribution discussed above is obtained in the direc-
tion described by the solid black arrows in Figure 4, i.e., prices further
from the delivery date are subtracted from prices closer to the delivery
date. In order to construct the case system above, however, we need
price changes where prices closer to the delivery date are subtracted
from prices further from the delivery date, which is described by the
dashed black arrows. We can obtain the distribution described by the
dashed arrows by reflecting the distribution described by the solid arrows
around the y-axis.
Figure 5 presents the scenario development and highlights the regions
where there could be volume risk. Two regions are highlighted. The in-
ner region represents volume risk in the second stage. The outer region
represents volume risk in the third stage and includes the possibility that
volume risk does not disappear in the second stage. The first stage is
always assumed to have no volume risk.
Volume risk in the second stage takes into account the number of con-
tracts in the second stage, the probability of the leg occurring, the value
of the unfavorable price movement that would occur if the trader is unable
to unwind times its probability and the volume risk ratio as expressed by
the formula [(SafeVol – V)/(SafeVol)]. Volume risk in the third stage is more
complicated because it takes into account both the third and the second
stage. If the second stage has no volume risk, then volume risk in the
third stage is calculated in the exact same way as volume risk in the sec-
ond stage. If, however, there is volume risk in the second stage, volume
risk in the third stage is calculated by taking into account the [(SafeVol
– V)/(SafeVol)] ratio in the third stage times the unfavorable price move-
ment from the third to the second stage times its probability plus the
[(SafeVol – V)/(SafeVol)] ratio in the second stage times the unfavorable
price movement from the second to the first stage times its probability.
The sum is tested for being unfavorable or favorable as a whole. Only if
unfavorable, is it multiplied by the probability of the leg occurring and the
number of contracts in the third stage.
Stochastic modelAs explained in the literature review section, a stochastic program is
thought to better represent a trader’s decision process as it accounts
for the uncertainties associated with the prices of contracts with distant
maturities. Another important reason for the application of stochastic
programming is the problem of friction. First of all, positions on a given
forward curve are established simultaneously by taking into account
the fact that prices of distant maturities are uncertain. Decisions are not
made sequentially in each stage. Second, the presence of volume risk
means that a given position may not be covered instantly in order to ad-
just to changing circumstances. Consequently, optimization of individual
scenarios does not convey accurate information about the best spread
trade.
The stochastic model developed for this paper is essentially a historical
simulation. Consequently, it has the same shortcoming that all historical
simulations have – application of past data to predict the future. In order
to benefit from the model users need to understand the logic behind us-
ing a historical simulation to infer the future:
1. A given model gives a structured framework for thinking about risk
[Adapted from Jorion and Taleb (1997)]. In this sense, a model struc-
ture is superior to intuition because it lays out a systematic way of
Three decision stages
50%
50%
50%
50%
50%
50%
Maturity 1
Maturity 2
Maturity 3
Case 11
Case 21
Case 22
Case 32
Case 33
Case 34
Case 31
Figure 4 – Forward curve model
50%
50%
50%
50%
50%
50%
Maturity 1
Maturity 2
Maturity 3
Case 11
Case 21
Case 22
Case 31
Case 32
Case 33
Case 34
Figure 5 – Volume risk model
139
managing risk. Once the model structure is found to be invalid, it
should be subjected to review.
2. An optimization model gives the best trading strategy based on all
the available information. Intuitive establishment of positions can
rarely emulate the analytical precision of an optimization model.
3. Predictions made by a historical simulation model should hold rea-
sonably well provided that there are no abnormal market movements
[Adapted from Jorion (1997)]. Historical simulations do not have the
ability to predict abnormal market shocks, unless those are present
in the past data that were fed into the model.
Case study This section develops a case study designed to assess the usefulness and
predictive power of the stochastic program explained above. The results
of the case study show that the model predicts the best spread trade ac-
curately in 75 percent of cases. In the remaining quarter of the cases the
inaccuracy is due to the market shock present in the year 2008.
The case study comprises five years of data, from 1 January, 2006
through 1 June, 2010. In each case, the model is calibrated using data
from a given year n with the aim of predicting the best spread trade for
the following year n + 1. The predictions are then back-tested using data
from the year n +1.
The case study was carried out using the tools specifically designed for
the purpose. In order to make sensible comparisons, the requirements in
Table 2 hold in all cases.
Predictions and realizationsUsing data from 2006 to predict the best spread for 2007 produces very
good results. The type of contracts in each stage is predicted with ac-
curacy even though the number is exaggerated. Although the realized
risk is larger than the predicted one, the realized gain is also significantly
higher than the predicted one. Table 3 gives the predicted and realized
values of the risk, the gain and the number of contracts. Overall, predic-
tions are very good.
Table 4 reveals that when data from 2007 is used to predict the best
spread trade for 2008 the results are satisfactory as well. The realized
gain is slightly smaller than the predicted one, but still larger than the
required 1000. The realized risk is also smaller. Unlike in the previous
case, the sources of profit are not predicted with accuracy. According to
the predictions, the profit should come from the long positions. In reality,
however, the profit comes from the short positions. Nevertheless, predic-
tions can be classified as relatively good.
Table 5 shows that unlike in the previous two cases, when data from 2008
is used to predict the best spread for 2009 results are not as good as
The Capco Institute Journal of Financial TransformationA Stochastic Programming Model to Minimize Volume Liquidity Risk in Commodity Trading
Constraint or data Value
Required return 1000
Initial price 4000
Volume risk Safe volume level is 100 contracts
Table 2 – Case study constraints
2007 Predicted Realized
Minimized risk 2197.76 5637.08
Pure price risk 2197.76 5637.08
Volume risk 0 0
Gain 1960.5 3543.25
From long 3335.5 9137.25
From short -1375 -5594
Contracts
First stage 2 short 1 short
Middle stage 2 long 1 long
Third stage 2 long 1 long
Table 3 – Predictions and realizations for 2007
2008 Predicted Realized
Minimized risk 2818.54 1800.55
Pure price risk 2818.54 1800.55
Volume risk 0 0
Gain 1771.63 1065.38
From long 4568.63 -657.13
From short -2797 1722.5
Contracts
First stage 1 short 1 short
Middle stage 1 long No contracts
Third stage 1 long 1 long
Table 4 – Predictions and realizations for 2008
2009 Predicted Realized
Minimized risk 1141.15 1066.17
Pure price risk 1141.15 1066.17
Volume risk 0 0
Gain 2168.38 88.5
From long 445.88 -834
From short 1722.5 922.5
Contracts
First stage 1 short 1 short
Middle stage No contracts 1 short
Third stage 1 long 1 long
Table 5 – Predictions and realizations for 2009
140
expected. Even though bull spread is both the predicted and the realized
best spread trade, overall, the realized gain is much smaller than the re-
quired 1000. This is a consequence of the different shape of the forward
curve in 2009 as compared to the one in 2008. The small realized gain,
however, is mitigated by the lower value of any further possible losses as
confirmed by the realized risk. Overall, predictions for 2009 are poor.
Table 6 shows that the best spread for 2010 is not determined with ac-
curacy using data from 2009. However, even though the predicted and
the realized best spread trade do not coincide, this is due to a minute
difference in the risk. If a bull spread is applied to the data from 2010, the
risk is increased by only several units. Nevertheless, the predicted and
realized values of both the risk and the gain are of comparable magni-
tude. Results can be described as good.
The year 2008 contains an abnormal shock in terms of settlement price
changes. In addition to that, the forward curve changes its slope during
2008. These abnormal shocks compromise the predictive power of the
model, which can easily be seen above when data from 2008 is used to
estimate the best spread trade for 2009. Like all historical simulations,
this model’s predictive power is very good as long as the market does not
produce abnormal shocks.
The peculiar shapes of the price change distributions used above result
in no volume risk associated with any of the predictions and realizations.
To make a case for volume risk, we can also make predictions using a
price change distribution based on data from 2005 to 2010 (1 June). Es-
tablishing a spread trade with ten long contracts in both the short- and
the medium-term and ten short contracts in the long-term produces a
volume risk of 761.64. Normally volume risk will appear when the estab-
lished spread is not the most optimal one.
Volume riskThe values of the volume risk in Tables 3 through 6 above show that the
model was optimized in a way that volume risk was reduced to zero,
given all the constraints that were imposed. In other words, provided that
the proposed best spread trade strategy is established in each case, the
portfolio will not be exposed to volume risk. A case in point is the ex-
ample at the end, where the established spread trade is not the best one
possible. Unsurprisingly, volume risk is present. A worse spread strat-
egy will produce an even higher volume risk. As a comparison, the best
spread trade in the last example is ten short contracts in the short term
and ten long contracts in both the medium and long term. In addition to
satisfying all constraints, this portfolio is predicted to incur a price risk
of 1463.05 and a volume risk of only 14.09. Volume risk will generally be
present whenever the portfolio includes positions liable to unfavorable
price movements and insufficient volume.
As a rule of thumb, higher gain is generally associated with higher risk.
And even though analytical models can produce precise answers given
large quantities of data, it is always down to the risk manager’s judgment
to choose the best option [Fragnière and Sullivan (2006)]. The option to
be chosen depends on the risk manager’s risk appetite and his overall
strategy [Fragnière and Sullivan (2006)]. The same holds for the results of
the stochastic model presented in Tables 3 to 6. To facilitate comparison,
only the options with the lowest risk were presented. Other options may
have higher risk awarded by a higher gain.
ConclusionThis is the first paper that considered volume risk in the optimization of
a portfolio of commodity futures contracts. The volume liquidity risk is
due to the low trading volume in the back end of the forward curves. The
results of the case study suggest that the stochastic model is successful
in determining the most optimal spread trade as long as markets do not
behave abnormally. The case study provided a comparison of the predic-
tive power of the stochastic program during both normal and abnormal
market behavior. Like all historical simulations, the predictive power of
the model is significantly reduced in the second case. Nevertheless, con-
sidering the rareness of such market shocks and the fact that the model’s
predictions were valid in 75 percent of the cases, it could be classified as
successful. Moreover, in all cases the volume risk of the determined best
spread trade was reduced to zero. Even if the structure of the forward
curve does not allow the complete elimination of volume risk, the model’s
estimates will always produce the spread trade that involves the least
amount of volume risk.
The stochastic program developed for this paper has three decision
stages. The advantages of extending the program to more stages are
arguable. While it may theoretically lead to a more realistic description of
the forward curve, the model will become too complicated and unwieldy.
The number of scenarios grows at the powers of two with each succes-
sive stage and so does the number of constraints. Adding even one or
2010 Predicted Realized
Minimized risk 1521.88 1806.33
Pure price risk 1521.88 1806.33
Volume risk 0 0
Gain 1105.5 1189.5
From long -834 -1582.5
From short 1939.5 2772
Contracts
First stage 1 short 1 long
Middle stage 1 short 1 short
Third stage 1 long 1 short
Table 6 – Predictions and realizations for 2010
141
two more stages will make integer optimization impractically slow. A pos-
sible solution would be the implementation of the model in XPress. Even
so, the programmatic description of the model would be a challenge.
What is more, this would defeat the purpose of the model being a practi-
cal and easy to use application in Microsoft Excel.
References Boyd, M. S. and I. Kaastra, 1995, “Forecasting futures trading volume using neural networks,” •
Journal of Futures Markets, 15:8, 953-970
De Roon, F. A. and Y. V. Veld-Merkoulova, 2003, “Hedging long-term commodity risk,” Journal •
of Futures Markets, 23:2, 109-133
Fragnière, E. and G. Sullivan, 2006, Risk management: safeguarding company assets, Thomson •
NETg, Stamford, CT
Fragnière, E., J. Gondzio, N. S. Tuchschmid, and Q. Zhang, 2010, “Non-parametric liquidity •
adjusted VaR model: a stochastic programming approach,” Journal of Financial Transformation,
28, 109-116
Grammatikos, T. and A. Saunders, 1986, “Futures price variability: a test of maturity and volume •
effects,” Journal of Business, 59:2, 319-330
Jorion, P., 1997, Value at risk: the new benchmark for controlling market risk, McGraw-Hill, •
London
Jorion, P. and N. Taleb, 1997, “The Jorion-Taleb debate.” Derivatives strategy, available at: •
http://www.derivativesstrategy.com/magazine/archive/1997/0497fea2.asp
Kansas City Board of Trade, 2010, Historical data, available at: http://www.kcbt.com/historical_•
data.asp
Kansas City Board of Trade, 2010, “KCBT sets new daily trading volume record in HRW wheat •
futures,” KCBT subscription service
Karpoff, J. M., 1987, “The relation between price changes and trading volume: a survey,” •
Journal of Financial and Quantitative Analysis, 22:1, 109-126
Pedersen, L. H., 2008, “Liquidity risk and the structure of financial crises,” Presentation for the •
International Monetary Fund and Federal Reserve Board: New York University Stern School
of Business, New York, USA, Available at: http://pages.stern.nyu.edu/~lpederse/papers/
LiquidityRiskSlidesLHP.pdf
Traders.com, 2010, “Futures liquidity - June 2010,” Technical analysis of commodities and •
stocks, available at: http://www.traders.com/Documentation/FEEDbk_docs/2010/06/FutLiq.
html
Wang, G. H. K. and J. Yau, 2000 “Trading volume, bid-ask spread, and price volatility in futures •
markets,” Journal of Futures Markets, 20:10, 943-970
The Capco Institute Journal of Financial TransformationA Stochastic Programming Model to Minimize Volume Liquidity Risk in Commodity Trading
143
PART 2
The Organization of Lending and the Use of Credit Scoring Techniques in Italian Banks1
AbstractThis paper examines the results of a survey carried out in
2007 by the Bank of Italy concerning different characteristics
of the organization of lending activities. Between 2003 and
2006 the physical distance between the headquarters and the
branches increased, the limits to the decision-making process
of loan officers were eased, their mobility raised, and the use
of economic incentives to reward their activity expanded. The
huge heterogeneity in organizational structures persists even
within relatively homogenous size classes. The diffusion of
statistical models to assess credit risk (scoring) accelerated
recently particularly among large banks, boosted by the new
Basel Capital Accord. Scoring is either very important or de-
terminant in decisions on credit extension while it is rarely in-
fluential in setting interest rates, the duration of the credit, and
the amount and type of collateral required. The survey shows
that banks have been progressively adapting their organiza-
tional structure in order to incorporate the credit scoring tools
into their lending processes.
Giorgio Albareto — Structural Economic Analysis Department, Bank of Italy
Michele Benvenuti — Economic Research Unit, Florence Branch, Bank of Italy
Sauro Mocetti — Economic Research Unit, Bologna Branch, Bank of Italy
Marcello Pagnini — Economic Research Unit, Bologna Branch, Bank of Italy
Paola Rossi — Economic Research Unit, Milan Branch, Bank of Italy
The authors wish to thank Guglielmo Barone, Enrico Beretta, Luigi Cannari, 1
Xavier Freixas, Giorgio Gobbi, Giacinto Micucci and Paolo Emilio Mistrulli
for their comments. We are especially indebted to Guglielmo Barone for his
help in formulating the questionnaire that was distributed to the banks. A
special thanks also goes to Rino Ferlita and Angela Romagnoli, whose help
during the various phases of the survey proved indispensable, and to Paolo
Natile, for his assistance in the preparation of several programs. The survey
was made possible by the kind cooperation of the Bank of Italy’s branch
network. This paper is part of a research project at the Bank of Italy on
“Banking organisation and local credit markets” [Cannari et al., 2010)]. The
views expressed in this paper are those of the authors and do not involve
the responsibility of the Bank of Italy.
144
Introduction and main resultsDuring the 1990s, two major factors affected the Italian banking industry:
liberalization and an intensive wave of technological innovation originat-
ing in the ITC sector. As a result, the banking system underwent a pro-
cess of consolidation, banks expanded and entered new markets, and
internal decision-making processes for granting loans were completely
overhauled. The ways in which households and firms accessed credit
changed. In the wake of these transformations, banks grew in size and
organizational complexity; they now found themselves having to man-
age their presence in a number of different geographical and product
markets. Large banks were not the only ones affected by this trend, as
small- and medium-sized banks frequently joined larger banking groups
or expanded internally; in both cases, the leaps in size sometimes led
to organizational discontinuity. The rapid advances in ICT technologies
had a profound effect on the output of the entire banking industry. These
transformations imply that we need to have an updated and deeper
knowledge of how banks organize the many aspects of their lending ac-
tivities (customer screening, the terms and conditions of lending, moni-
toring of the borrower’s conduct, etc.).
The literature on bank-firm relationships generally treats banks as uni-
tary entities and neglects the characteristics of their internal structure.
Recently, however, the literature on organization has spawned several
papers that emphasize the importance of the strategic interaction among
managers in charge of various functions within the banking organization;
these managers have different information and are bearers of interests
that do not necessary coincide. It has been shown, both theoretically and
empirically, that the ways in which this interaction occurs can affect the
effectiveness of credit allocation, especially in the case of SMEs.
One of the main consequences of technological change for credit markets
was the introduction of credit scoring2 techniques based on standardized
data. Despite their increasing importance, including in the Italian market,
there are few studies on the diffusion and use of these procedures. To
collect data useful for understanding these changes, the Bank of Italy
conducted a survey in 2007 of over 300 banks, which represented the
universe of intermediaries of a certain minimum size and organizational
complexity. This report presents the results of that survey.
The analysis of banks’ internal organization revealed profound differences
among the Italian intermediaries from four standpoints: the geographical
distance between a bank’s headquarters and its branch network; the de-
cision-making autonomy of the branch managers, proxied by the amount
of small business credit that they are authorized to extend in proportion
to that which can be approved by the CEO; the length of the branch
managers’ tenure; and the use of incentives for their remuneration. Part
of this heterogeneity is accounted for by the size and institutional differ-
ences of the banks surveyed. However, some heterogeneity still exists
among homogenous groups of intermediaries. The results confirm how
the internal structure of lending activities adapts to specific circum-
stances and forms a crucial component of banks’ competitive strategies.
An implication of these findings is that an analysis of these phenomena
must employ a broader and richer taxonomy than the traditional one that
is based on the banks’ size. For most of the participating banks, the
distance between their headquarters and the branch network increased
between 2003 and 2006. Bank managers enjoyed greater mobility and
autonomy in decision-making, and economic incentives were more fre-
quently used for their remuneration. The results do not support the thesis
that the advent of new technologies greatly diminished the role of bank
managers with negative repercussions for banks’ direct interaction with
SMEs. On the contrary, it is possible that lower communication costs fa-
vored the greater autonomy of local managers in the periphery (branches
or local decision-making centers). Increased mobility could be the result
of events that were partially exogenous to the bank’s strategy (i.e., merg-
ers and acquisitions or tough competition in the local credit markets). It
could also be the result of an active policy by banks to reduce the costs
of opportunistic practices by local branch managers.
The survey has shown that credit scoring has spread among Italian in-
termediaries, with a sharp acceleration in recent years that is probably
related to the introduction of the new Capital Adequacy Accords (Ba-
sel II). The diffusion process was more pronounced among large banks,
which were in a position to exploit economies of scale. Credit scoring
techniques mostly process balance sheet data, which are historically
the most frequently used element for evaluating creditworthiness. Larger
banks generally use internally developed models and place great em-
phasis on qualitative data, such as borrowers’ corporate governance
and organization, and the project to be financed (for large firms espe-
cially). Although scoring techniques play a central role in the decision
to grant credit, they are not frequently used to determine the terms and
conditions of loans. The scores generated by the application of these
techniques appear more stringent for large banks than for smaller ones.
Overall, the results suggest that the new techniques, which banks are still
in the process of adapting, flank but do not entirely replace the previous
evaluation processes and data sources.
The organization of lending and the use of credit scoring: the main issuesBanks usually rely on a mix of data sources to assess firms’ creditwor-
thiness. Some information is easy to classify and transmit from a dis-
tance (hard information), whereas other types of information are acquired
through personal contacts and comprise qualitative elements that are
difficult to communicate to people other than those who collected them
We use this term to denote all of the automated techniques for assessing creditworthiness, 2
which are described later in the paper.
145
The Capco Institute Journal of Financial TransformationThe Organization of Lending and the Use of Credit Scoring Techniques in Italian Banks
(soft information). It is generally thought that qualitative data play a great-
er role in the evaluation of start-ups and small businesses, which are
prone to having more opaque information (for example, owing to fewer
years of experience) or less reliable processed data and are, in any event,
subject to less strict information requirements (such as accounting data)
than large corporations.
The collection of quantitative and especially qualitative information about
small firms is done though the branch manager. It is usually at this level
of the bank organization that the first contact is made with the small firm,
the assessment of creditworthiness is activated, and the relevant infor-
mation is transmitted for evaluation at the higher levels of the banking
structure. In some instances, the decision of whether or not to grant a
loan and the loan’s terms and conditions are made in complete autonomy
by the branch manager.
The literature on corporate organization acknowledges that it is possible
for the objectives of branch managers to diverge from those of the ulti-
mate control holders. It also emphasizes how specialization in the gather-
ing of data and information asymmetries between the headquarters and
the branch network within a complex organization can generate the need
for a two-way transfer of information along the hierarchy. The literature
on banking shows how the many distinctive elements featuring internal
organization (the organization chart, the extent to which decision-making
is centralized or decentralized, internal control systems, procedures for
communicating between the various organizational levels, etc.) have a
decisive influence on the strategies of the branch managers and, through
them, on the allocation of credit to small firms.
The central role played de facto by branch managers in lending to SMEs
has not been given sufficient attention by the literature, in part owing
to a lack of data. Some recent contributions have begun to fill this gap.
Liberti (2005) and Liberti and Mian (2006) show that as one goes higher
up in the hierarchy of the bank organization, and as the customer grows
more distant from the decision-making center, qualitative elements will
weigh less on loan decisions. This phenomenon shows how qualitative
information is in fact collected and stored at the lower hierarchical lev-
els and how its transmission costs increase with geographical and or-
ganizational distance (defined as the number of levels involved in the
decision-making process). Stein (2002) and Berger et al. (2005) broaden
the debate to include the effects that different organizational models can
have on the incentives for banks’ branch managers. In particular, Stein
uses a theoretical model to show how a large bank can discourage the
acquisition of soft information by a branch manager where data must be
communicated along multiple hierarchical layers and the transmission of
the data becomes extremely costly. This effect does not occur in the case
of small banks, where there is markedly less physical and hierarchical
distance between the headquarters and the branch network. As a result,
the branch managers of major banks may have a greater incentive to
collect hard data, which are more easily provided by large firms, whereas
smaller intermediaries can specialize in the acquisition of qualitative in-
formation and in small business lending.
For Japan, Uchida et al. (2006) show how branch managers’ traits are not
important for the purposes of accumulating soft information and explain
this result with reference to the likelihood that the strong social cohesion
of Japanese society reduces the costs of transmitting qualitative informa-
tion. In an empirical analysis of Italian data, Ferri (1997) shows the posi-
tive correlation between branch managers’ mobility and bank size and
explains this result by referring to the greater difficulties that large banks
face in limiting the moral hazard stemming from the potential for collu-
sion between branch managers and bank customers. The large physical
and organizational distance between the headquarters and branches of
major banks increases the costs of monitoring, driving these intermedi-
aries to use the mobility of banking managers as a tool for limiting their
opportunities to reap private benefits. By contrast, the geographical and
organizational proximity typical of small banks encourages them to maxi-
mize the benefits of the stability of local managers while at the same time
maintaining monitoring costs at reasonably low levels, given that the top
managers may belong to the same local community as the branch man-
ager. Moreover, if a large bank is specialized in lending to medium-sized
and large firms that are capable of providing hard data [Stein (2002)], it
follows that these banks will be less motivated to keep the local managers
at the same branch for a long time. In these cases, mobility can represent
a way of furthering local managers’ careers and of warding off excessive
inertia in the administration of local branches. Hertzberg et al. (2007) use
data on the turnover of the local bank managers of an Argentinean bank
and show that the bank utilizes mobility to persuade these managers to
report information on the creditworthiness of borrowers accurately. Scott
(2006) shows that for a sample of small U.S. firms, the turnover of local
managers increases the likelihood of credit rationing.
The literature previously surveyed brings out four key themes related to
the characteristics of branch managers’ activities:
Hierarchical and geographical distance between bank headquar-■■
ters and branch managers – as we have seen, distance can affect
the costs of transmitting qualitative information collected at the local
level and, accordingly, the incentive to acquire them. It can also deter-
mine the cost of monitoring branches at the central level.
Decision-making autonomy of the branch manager –■■ this variable
undoubtedly enhances the incentives for a branch manager to acquire
soft information, but at the same time, it increases the costs of control
and coordination at the central level.
Tenure of the branch manager –■■ a trade-off similar to the one
described earlier can also be generated: the higher stability of a
146
branch manager’s position may lead to more incentives to acquire
soft information, but the costs of control can also increase (due, for
example, to the moral hazard).
Incentives –■■ economic incentives for branch managers can help
reduce the moral hazard by aligning the objectives of peripheral
agents with those of the bank central management, with the danger,
however, of transferring excessive risk to the branch manager.
The role played by technological innovation in the development of proce-
dures for granting loans adopted by banks from the 1990s onward was
recalled earlier. One of the most important consequences of ITC advanc-
es in the banking industry was the sharp reduction in data processing
costs, that is, the use of data for administrative purposes at various orga-
nizational levels. The new regulations on minimum capital requirements
also provided strong incentives for the adoption of statistical techniques
for measuring credit risk; the methods vary, but their distinctive feature
consists of their ability to group customers within a finite number of cate-
gories, associating with each one a synthetic indicator that expresses the
probability of default and accordingly the degree of risk. The introduction
of these techniques can influence the role of branch managers in the al-
location of credit in various ways. Indeed, credit scoring can represent
an alternative means of assessing creditworthiness that contrasts with
decision-making processes that emphasize qualitative information and
the close interaction of branch managers with customers. At the same
time, the adoption of scoring techniques allows, at least partly, the trans-
formation of soft information into processed data and facilitates the con-
trol of branches. These issues, which are closely interrelated, cannot be
dealt with without further exploring the specific nature of credit scoring,
including its relatively recent introduction into the Italian banking system.
From this perspective, the analysis of credit scoring complements that of
organizational variables and central-peripheral relations.
Although these techniques have been used since the fifties, their mod-
eling has greatly evolved in the last decade [Hand and Thomas (2005),
Hand and Zhou (2009)]. At the very beginning, these models were aimed
at supporting accept-or-reject decisions; nowadays markets are more
mature, emphasis is moving from acquisition to retention, and the mod-
els tend to back a wide range of business objectives. Models are expect-
ed to drive the optimal decision in terms of price and credit conditions.
Furthermore, models are increasingly able to go beyond the individual
measure of risk by recognizing the portfolio dimension.
Quantitative methods include both credit scoring models, which distin-
guish cases of expected default from non-default using statistical tech-
niques such as discriminant analysis or logit and probit analysis, and
internal rating systems, which more or less automatically map individ-
ual borrowers (or, in the most sophisticated cases, the different lines of
credit of each borrower) on a scale of judgments [Allen et al. (2004)].
These technologies can allow, even before the formulation of any final
judgment, discretionary interventions by one or more persons to assess
qualitative elements not explicitly considered in the model. The degree of
flexibility in the use of credit scoring, therefore, varies depending both on
the characteristics of the procedures adopted and on their importance
in lending decisions and customer relationships. In this report, the term
“scoring” refers to all of these instruments indiscriminately. What follows
is a brief survey of the main questions related to the adoption and use of
credit scoring techniques, which will be described more fully below:
Banks’ characteristics and the adoption of credit scoring tech-■■
niques – in Europe, the introduction of credit scoring was more grad-
ual than in the U.S. and occurred later [Degryse and Ongena (2004)].
Recently, however, credit scoring has been widely adopted by Italian
banks. The literature concurs in emphasizing how the adoption of
credit scoring techniques is influenced both by the size of banks and
by their organization. Size acts in the usual ways; larger banks have
more resources to invest in new techniques, and the cost of invest-
ments is then distributed among a broader loan portfolio. Moreover,
large banks report greater diseconomies owing to distance, including
difficulties in transmitting soft information internally, in selecting and
monitoring loans, and in designing the right incentives for local man-
agers. The adoption of credit scoring techniques reduces the costs
of screening and monitoring firm activity and controlling branch man-
agers, which mitigates the problems of monitoring from a distance
[Berger et al. (2005)].
The characteristics of the scoring techniques –■■ scoring techniques
can differ both in their origin (developed internally or acquired exter-
nally) and in the datasets that are processed. The use of methods
developed internally by banks implies greater control and more flex-
ibility, in the sense that these methods are easier to modify if the
bank no longer considers them adequate. Moreover, broad recourse
by banks to externally developed techniques could lead to greater
homogeneity in the criteria for assessing creditworthiness by the vari-
ous intermediaries. These techniques mostly rely on quantitative data
inputs to formulate scores. As mentioned earlier, there is also concern
about the possible decline in the importance of soft information in
lending decisions, with the result that start-ups and smaller firms,
relying more on bank credit, could be adversely affected by an intense
and rigid use of these methods [Cowan and Cowan (2006)]. However,
the empirical evidence to date, above all that concerning the U.S.,
does not appear to justify these fears [Berger et al. (2005)].
The importance of scoring techniques and how they are used –■■
although scoring techniques have by now been widely adopted by
banks, their importance in assessing customers’ creditworthiness
is not a foregone conclusion. This issue is by no means secondary,
as the impact of these techniques on the extension of credit varies
depending on whether banks use them as the main instrument for
147
assessing creditworthiness or as a supplementary instrument along
with other evaluation techniques [Berger et al. (2005)]. Based on the
results of a recent survey conducted on a sample of U.S. banks, the
scores for small firms are considered to be less important than the
traditional indicators of creditworthiness, such as cash flow and avail-
able collateral [Cowan and Cowan (2006)].
In the next section, the distinction between intermediaries is based on
size and institutional set up. Given that information asymmetries, agency
problems, data transmission costs, and economies of scale linked to the
use of IT become more complex when we move from individual banks to
groups, the survey also distinguishes between stand-alone banks and
members of a group.
The surveyOur data are taken from a qualitative questionnaire (reproduced in the
Appendix). The purpose of the survey is to gather information about
the organizational aspects and scoring techniques used in the lending
process. The questions capture banks’ organizational structure, mean-
ing both their specialization by customer segment and the number of
hierarchical layers involved in the decision to lend to a firm. For each
hierarchical level, the survey establishes the degree of autonomy granted
to approve loan applications and possibly to establish their terms and
conditions. Finally, the existence and type of economic incentives for
branch managers and the length of their tenure are also considered. In
the second part of the questionnaire, the questions explore the adoption
of statistical-quantitative techniques for evaluating firms and their use in
setting terms and conditions of loans as well as in the monitoring of the
loans. Next, the characteristics of the models are surveyed, particularly
the data used to calculate the scores and the role of quantitative and
qualitative information in evaluating new loan applicants.
The questionnaire was submitted to intermediaries in 2007 through the
Bank of Italy’s branch network. The selection of the banks was aimed
at ensuring adequate coverage, both geographically and by the type of
bank (medium-sized and large banks, small stand-alone banks and oth-
ers belonging to groups, mutual banks). The sample design was based
on that used for a database including interest rates (TAXIA), which, until
the end of 2006, surveyed 215 reporting institutions selected by their
size (measured by total lending to customers), geographical location, and
share of loans reported to the Central Credit Register. The original sample
of 215 banks was modified in two ways. First, intermediaries specializing
in activities such as factoring, leasing, household mortgages, and credit
recovery were excluded, given that the questionnaire focused on lend-
ing to firms in traditional technical formats. Second, it was left to local
data collectors to submit the questionnaire to banks excluded from the
TAXIA sample, provided that the banks did not belong to the “minor” size
category.
The final sample included 333 banks and 322 responses. The missing
answers all refer to banks whose business lending was marginal or nil.
The composition of the sample by geographical area and size is shown in
Table 1. The sample accounts for 82.8 percent of the total amount of out-
standing loans to non-financial firms in 2007. Coverage is high through-
out the country, ranging from 81.0 percent in the regions of the North-
west to 90.4 percent in the South; among the size/institutional classes,
this percentage is reduced for small stand-alone banks (34.7 per cent),
a category that includes several branches of foreign banks that were not
surveyed. The average large bank in the sample has 519 branches and
more than 5,600 employees. It is about 10 times bigger than the average
small bank belonging to a group, 20 times that of a small stand-alone
bank, and 50 times that of a mutual bank.
The questionnaire contains mostly qualitative questions, and the data
have accordingly been aggregated in two separate ways: either as simple
frequencies of responses or as weighted frequencies where the weights
are equal to loans to SMEs or large firms, depending on which type of
borrower the question refers to. The significance assigned in the first
case is that of the diffusion of the phenomena (for example, what share
of the banks use credit scoring), and that assigned in the second case is
that of the likelihood of the borrower encountering a given phenomenon
(using the same example, the share of lending to SMEs, or large firms
granted by banks using scoring techniques).
Banks’ organization, SME lending, and the role of the branch managerDistance from headquarters and banking presence in local credit marketsSome recent literature has focused on the effects that the distance be-
tween a bank and a firm can have on access to credit and its terms and
conditions. This literature often treats the bank as a single entity, but if
instead, as noted above, we recognize that the bank has an articulated
internal structure, the question can be approached from two different
standpoints, i.e., one referring to the distance between the borrower firm
and the unit within the bank charged with lending decisions and the other
referring to the distance between that unit and the head office, where
the people who exercise control powers are located. This relationship
between center and branch may be influenced by agency problems that
Class of bank
North-
west
North-
east Center
South/
Islands Total
Medium-sized/large 17 9 7 4 37
Small group-member 27 21 24 20 92
Small stand-alone 10 9 10 3 32
Mutual 54 57 40 10 161
Total 108 96 81 37 322
Table 1 – Sample composition (number of banks)
The Capco Institute Journal of Financial TransformationThe Organization of Lending and the Use of Credit Scoring Techniques in Italian Banks
148
may be just as severe as those traditionally considered in bank-firm rela-
tions. Apart from the costs of monitoring the branch manager’s activity,
distance can also increase the costs of transmitting qualitative informa-
tion, lessen the incentives for the collection of this information, and in-
crease the effort required to pass on best practices.3 When instead of
geographical distance there is hierarchical distance, measured by the
number of organizational levels along the chain of command involved in
the decision, the implications are similar.
In what follows, we describe several indicators of the distance between
the center and periphery for the banks in our sample, the number of lo-
cal markets in which they do business, and how these factors evolved
between 2000 and 2006. To compute distance, we must first define the
relevant geographical units and devise a suitable gauge. To this end, we
define the location of the persons who ultimately control the bank as
the city in which the bank’s legal head office is established and that of
the branch manager as the locality of the branch, itself defined as the
main city within each of the “local labor market areas” mapped by Istat
in 2001. This procedure effectively balances the need for a geographi-
cal classification detailed enough to allow for precise measurement of
the distance between head office and branch with the need to capture
the differences between the market areas where the head office and the
branches are located.4 Once the local labor market area where the bank’s
legal office is located is determined, we calculate the distance between
that area and all of the other local labor markets where the bank has at
least one branch.5 The distances so obtained are weighted according to
the local labor market’s share in the bank’s total lending. Weighting is
necessary to prevent branches in far-off areas that do not account for a
significant portion of the bank’s assets from having a disproportionate
effect on the average distance.
The results are given in Table 2. The mean for this indicator in 2006 is
47 kilometers, and the median is 21 kilometers. The larger banks have
a more extensive branch network and accordingly higher average dis-
tances than medium-sized, small, and mutual banks. Statistical tests on
these averages indicate that the differences between classes of banks
are significantly different from zero. There are also significant inter-cate-
gory differences in the number of local labor systems in which the bank
has at least one branch.
The average distance is highly variable. The inter-quartile difference is
equal to 132 percent of the median for all banks and is also high even
within each of the size categories. Between 2000 and 2006, there was
a broad increase in the distance between the local labor system of the
head office and the branch. The median increased for all categories of
bank (except for mutual banks, for which it held constant), but the mean
diminished slightly among large banks while increasing for small and mu-
tual banks. Statistical tests on the averages indicate that the differences
are significantly different from zero for small and mutual banks but not
for the large banks. The average number of local labor market areas in
which banks had branches also increased over this period. In general,
the increase in distance mainly involved medium-sized banks, especially
for banks belonging to groups, and, to a lesser extent, small non-group
banks and mutual banks. In part, this pattern occurred because some
large banks involved in mergers reduced the number of branches and of
local labor systems in which they were present, shortening their range
of operations. Even excluding these banks, however, the increase of the
average distance for the large banks is small and less pronounced than
that recorded for the other size categories.
It is also interesting to examine the evolution of headquarters-branch
distances for banking groups, considering all of the local labor market
areas with at least one branch and computing the distance from the par-
ent bank’s head office. For such an evaluation, we exclude groups that
have only one bank and any branches or subsidiaries abroad. For the 33
banking groups so identified, the mean distance increases from 180 to
Berger and De Young (2001) have shown that a bank holding company’s ability to transfer 3
its own efficiency to the other banks in a group diminishes as the distance between the
former and latter increases. Deng and Elyasiani (2005) have found that the risk of the
holding company increases with the distance from its subsidiaries and interpret this in
terms of greater ability of the local manager to engage in opportunistic behavior when
operating in far-away markets.
Some of the foregoing observations on the relationship between the head office and local 4
branches and the related agency problems might suggest the adoption of a “continuous
space” and the definition of distance simply as the distance in kilometers between the
head office and the branch. However, other aspects of the relationship are better defined
by reference to the distance between the local markets of the two units. The first method
would give a better gauge of the physical distance; the second highlights the distance and
thus the differences between the broader areas within which the two units are located.
The local labor systems are identified by the geographical coordinates of their main 5
municipality; the distance is calculated by the “great circle” formula, assuming that the
earth is spherical. Another assumption concerns the distance to branches located within
the same local labor system as the head office. This distance could be set to zero or
otherwise calculated on the basis of the land area of that system. The results do not
change when the distance is set to zero.
Class of bank
2000 2006
Mean Median Difference2 Mean Median Difference2
Medium-sized/large 152 116 126 151 117 136
Small group-member 50 30 99 66 39 100
Small stand-alone 26 21 67 31 22 69
Mutual 15 14 64 16 14 68
Total 42 19 121 47 21 132
Source: Sample survey of 322 banks.
1 - Banks present in sample both in 2000 and in 2006. Distance is mean distance in
kilometers of head office from the local labor systems in which the bank is present,
weighted by the bank’s lending in that local labor system.
2 - Interquartile difference over median, in percent.
Table 2 – Distance head office-branch1 (kilometers, percentages)
149
217 kilometers, the median increases from 131 to 241 kilometers, and the
average number of local labor markets increases from 107 to 142.
The foregoing suggests several considerations. First, the variability in
banks’ size is reflected in huge differences in the extent of branch net-
works and in the distance between the center and peripheral branches.
However, even within the various size and institutional classes the differ-
ences remain, which means that the problem of the cost of transmission
of qualitative information, and of agency problems in general, between
branch managers and head offices differs in extent from bank to bank, as
do the organizational structures adopted to deal with them (see below).
Second, between 2000 and 2006, there was a general lengthening of
distances, owing in part to the introduction of information and commu-
nications technology, which greatly reduced the cost of acquiring, pro-
cessing, and transmitting information [Berger (2003)].6 Third, the increase
in distance was greater among small banks belonging to groups, among
small non-group banks and, to some extent, among mutual banks. Large
banks held the average distance of branches from the head office rough-
ly unchanged; their overall range of action was increased through their
group structure. This is in contradiction with some recent works showing
that small and minor banks have grown more than large banks [Bonac-
corsi di Patti et al. (2005)].
Organizational structure and layers of hierarchyThe previous section considered the distance between the head office
and branches in geographical terms. Now we consider it in terms of orga-
nizational structure, measuring the “distance” between the branch man-
ager and the top management of the bank by the organizational model
adopted and the number of layers of management between them. The
model may have a divisional structure, in which customers are segment-
ed by size, or a single structure that performs all lending activities. In a di-
visional structure, the middle management and/or the type of peripheral
unit varies with the size of the borrower firm. Along with the modification
of organizational roles, the delegation of powers, and the way in which
lending is planned and carried out, also change. The number of layers of
hierarchy is used to gauge the depth of the organizational structure.
The model of organization – the responses to the questionnaire indi-
cate that over 70 percent of the large banks are organized by divisions,
with customers segmented by size and typically divided into SMEs and
large firms. For small banks belonging to groups, this percentage falls
to 33 percent, and for small stand-alone banks, it falls to 24 percent; for
mutual banks, it is 10 percent. As a rule, the variable used for segment-
ing firms is sales, although in a significant number of cases, small banks
use loan size. The modal sales threshold dividing small from large firms is
€2.5 million. Among small banks, organization by divisions is a very recent
phenomenon – almost 50 percent of the small and mutual banks with
this model adopted it in the last three years, compared with only about
10 percent of the large banks. Typically, large banks have a greater dif-
ferentiation of products, markets, and customers. Accordingly, a divisional
organization can exploit specialization by assigning a single manager re-
sponsible for a given product or geographical area. A divisional structure
also makes it easier to adapt to the industrial or local context. The large
banks, thanks to potential scale of economy, can more readily sustain the
costs of divisional organization, which entails more staff and structures
engaged in similar activities. Among other potential costs, there is the
need for a closer coordination of relatively autonomous, diversified units.
Layers of hierarchy – the length of the chain of command is given by
the number of ranks between the branch manager and the CEO.7 These
ranks are hierarchically ordered, and each has specific powers in terms
of maximum loan amounts. In what follows, we consider the positions
involved in lending to SMEs.8 The average number of layers varies signifi-
cantly by institutional category: 5 for large banks, 4 for small group-mem-
ber banks, 3.6 for small stand-alone banks, and 2.8 for mutual banks.
Figure 1 shows the distribution of layers of hierarchy according to this
classification.
0 5
10 15 20 25 30 35 40 45 50
2 3 4 5 6 7 8 10
Medium-sized/large banks Small group-member banks Small stand-alone banks Mutual banks
Source: Sample survey of 322 banks.
1 - The number of layers is the number of hierarchically ranked positions from branch
manager through chief executive officer. The horizontal axis gives the number of layers
and the vertical axis gives the percentage shares of the number of layers for each class
of institution.
Figure 1 – Distribution of number of layers in hierarchy1 (numbers and percentages)
The Capco Institute Journal of Financial TransformationThe Organization of Lending and the Use of Credit Scoring Techniques in Italian Banks
Felici and Pagnini (2008) show that, with other things (including profitability) being equal, 6
the banks with larger ICT endowments increase their capacity to move into markets far
away from their existing localities.
We consider only the ranks that correspond to a significant rise in hierarchical level. This 7
excludes, for instance, deputies and auxiliary staff but includes all the grades assigned to
run an organizational unit.
We do not comment on the data on layers of hierarchy involved in lending to large firms 8
or the degree of decision-making autonomy enjoyed by branch managers in such lending.
This decision follows from the limited role of branch managers in this segment. Moreover,
few banks have different organizational structures for lending to SMEs and to large firms.
150
As bank size increases, the average number of hierarchical layers and
the variance of their distribution both increase. The large banks are those
with the greatest diversity in the length of the chain of command. One in
five has a simplified organization (with fewer than three layers), whereas
one in three has a highly complex structure (more than six layers). For
mutual banks, more than 40 percent have an elementary structure with
just two levels: branch manager and CEO. However, even in this class,
there are more complex organizations; one in five has four or more layers.
The number of layers of hierarchy has a significant impact on a bank’s
operation. On the one hand, more layers may mean higher costs for the
transmission of information from one level to another and longer decision
times. On the other hand, a flatter organizational chart for a given size
of staff implies larger units and thus a larger area to control, i.e., a larger
number of subordinates under each supervisor.
Branch managers’ decision powerThe branch manager obviously has a privileged position for acquiring
information relevant to loan decisions. However, the formal authority, i.e.,
the right of control, belongs to the top management of the bank. The
organizational problems posed by the lack of coincidence of these two
figures can be addressed by transferring the information to the person
endowed with the formal authority (central decision-making) or by as-
signing the power to decide to the person who has obtained the informa-
tion (decentralized decision-making).9 One solution creates problems of
information transmission, whereas the other creates problems of con-
trol.10 Between these two extremes, there is a continuum of degrees of
decentralization to resolve the trade off.
The literature on corporate organization has dealt with the problem of
measuring the degree of decentralization of the power of decision from
various standpoints.11 In this survey, the banks were asked to indicate,
for every hierarchical level, the maximum amount of credit that could
be granted on that level’s own authority. This information was used to
construct an indicator of the degree of decentralization. This section also
refers to lending to SMEs.12
The amount of credit that the branch manager can grant on his own pow-
er increases with the size of the bank (Table 3). The mean is €550,000
(and the median €250,000) for large banks, €200,000 for small banks
belonging to groups, €90,000 for small stand-alone banks, and €50,000
for mutual banks. However, these means conceal some variability within
the subgroups, as shown by the interquartile difference as a ratio to the
median. The greatest variability is found among mutual banks; in one-
fifth of them, the value is 0, and, in one-fifth, it is more than €100,000.
The amount of credit that the branch manager can extend autonomously
is larger for mortgage loans (€124,000) and smaller for uncollateralized
overdraft facilities (€62,000) or unsecured loans (€54,000). These differ-
ences reflect the role of collateral.
Comparing the loan authorization power of the branch manager with that
of the top management, we can construct an indicator to measure the
degree of decentralization. The branch manager (or head of the local unit
of the bank) and the CEO are the two figures that appear in virtually all of
the organizational charts. Comparing the powers delegated to them, we
can build an index of the branch manager’s autonomy with respect to the
powers of the CEO. This index equals 5 percent in large banks, 11 per-
cent in small banks belonging to groups, 14 percent in small stand-alone
banks, and nearly 20 percent in mutual banks. The index is negatively
correlated with bank size because as the bank becomes larger, the pow-
ers of the CEO increase more than proportionately with respect to those
of the branch manager. The CEO performs different functions depending
on the type of bank. In mutual banks, for example, the most important
See Christie et al. (2003). The distinction between real and formal authority is from Aghion 9
and Tirole (1997).
The complete centralization of the power of decision in the hands of top management could 10
lead to organizational failures as a result of the branch manager’s lack of incentive to acquire
information. Also, the transmission of information from lower to higher levels may entail a
loss of information or at any rate a lag between information acquisition and decision. Finally,
if agents are rationally constrained à la Simon, the bank’s top management might not be
capable of handling a large information flow. Decentralization, on the other hand, allows
the power of decision to be in the hands of the person with the information but leads to top
management’s loss of control over that person’s choices. The costs of delegating formal
authority are defined as agency costs and depend on the fact that the aims pursued by the
bank as such do not necessarily coincide with the personal objectives of the staff. Agency
problems typically involve collusion between the branch manager and the borrower firm or
the manipulation of the data that the branch manager has gathered.
Christie et al. (2003) and Acemoglu et al. (2006) identify the autonomy of decision with the 11
presence of profit centers within the corporation. This indicator reflects the observation that
a cost center controls either revenue or costs but not both, whereas a profit center makes
decisions involving both costs and revenue. The degree of decentralization is thus 0 when
the level immediately below the CEO is a cost center and 1 when it is a profit center.
The questionnaire also had a question about the autonomy of branch managers to price 12
loans; however, owing to the number of non-responses, we have elected not to comment
on these data.
Class of bank
Autonomous decision-making
power of branch manager,
thousands of euros1
Index of relative power
delegated2
Mean Median Difference3 Mean Median Difference3
Medium-sized/large 546 250 146 5.3 3.1 159
Small group-member 202 125 118 11.0 8.6 119
Small stand-alone 92 80 125 13.7 14.3 130
Mutual 53 30 217 19.2 16.7 122
Total 154 75 173 14.7 10.8 177
Source: Sample survey of 322 banks.
1 - Banks were asked: “Consider the granting of loans to non-financial companies applying
to your bank for the first time and which, based on the information available, have no
solvency problems. What is the maximum amount of credit (in thousands of euros) that
can be autonomously granted by ...”. The figure represents the power delegated to the
branch manager or head of local unit. Lending to SMEs.
2 - The relative index is the amount of power delegated to the branch manager normalized
with respect to that of the CEO.
3 - Interquartile difference over median, in percent.
Table 3 – Delegation of powers (thousands of euros and percent)
151
decisions on loans to SMEs are taken directly by the bank’s board or
council, whereas in large banks, they are taken lower down in the chain
of command. Rather than compare such drastically different types of
banks, again it is more meaningful to observe the variability within each
category. The ratio of the interquartile difference to the median again
shows great variability.
The evidence presented so far refers to the year 2006. The questionnaire
also asked about the trend in the past three years. Half of the banks re-
ported a tendency toward greater decentralization, whereas just 4 percent
declared that they had centralized their decision-making powers. This ten-
dency characterized all the banks but was most pronounced for the large
ones (about three quarters of which reported greater decentralization).
Several conclusions follow. First, the size of the bank is a major determi-
nant of the delegation of powers, in absolute value, to the branch man-
ager. Some large banks, which tend to be farther away from the local
credit markets and the borrower firms, could endow their branches with
considerable autonomy with a view to creating streamlined structures
that are relatively immune to the inertia and lack of dynamism typical
of large corporations and closer to the local community. The amount of
lending authority may also be affected by the type of customer served.
Banks with larger customers tend to have higher ceilings on their local
units’ lending powers. Second, decentralization is more pronounced in
small banks than in large banks. A greater geographical proximity of
local to central offices and less complex lending business may foster
the decentralization of powers of decision, thanks among other things
to a greater ease of control of top management over local managers.
However, by itself, bank size does not explain the variability in the de-
gree of decentralization. Within each of our size classes, the observed
variance would appear to indicate significant variety in organizational ar-
rangements. The responses further reveal a general tendency towards
decentralization. Together with the recent adoption of divisional models,
this tendency highlights a certain organizational dynamism, presumably
in connection with an increasing geographical distance between a head
office and branches and with the diffusion of ICT. Finally, the greater de-
centralization among large banks may be a response to the competition
of smaller institutions in lending to SMEs in local credit markets.
Branch managers’ tenureThe tenure of branch managers presents the bank with a trade off. Great-
er stability in the position facilitates contacts and familiarity with the local
market and hence the acquisition of soft information. However, it also
heightens the informational asymmetry between a branch manager and
the head office, possibly enabling the former to reap private benefits (by
collusion with local borrowers, say, or manipulation of the information
transmitted up the chain of command). The survey found that the mean
time for which branch managers held their position was nearly four years,
with a median of 38 months and a mode of 36 months (Table 4). These
figures are similar to those found in a similar survey by Ferri (1997). The
mode (three years) could depend on corporate routines and widespread
organizational models shared by many banks. The term of office is short-
er in the larger banks and longer in small stand-alone banks and mutual
banks. The standard statistical tests show that all of these differences
are significantly different from zero. Small banks belonging to banking
groups have values similar to large banks, suggesting that group organi-
zational strategies extend to smaller intermediaries. The degree of mobil-
ity of branch managers shows considerable variability not only for the
entire sample but also within each category of banks.
Regarding trends in tenure, nearly 40 percent of the sample banks report-
ed that tenure had shortened in the last three years, whereas 14 percent
reported a lengthening. The tendency towards greater mobility was also
Class of bank
Tenure in months1 Trend in the last three years1
Mean Median Difference2 Shortened Unchanged Lengthened
Medium-sized/large 32 32 33 39.4 45.5 15.1
Small group-member 38 36 50 36.5 54.0 9.5
Small stand-alone 49 38 63 26.7 60.0 13.3
Mutual 50 48 50 41.8 42.5 15.7
Total 45 38 71 38.5 47.7 13.8
Source: Sample survey of 322 banks.
1 - Respondents were asked: “Indicate the average length of tenure of branch managers,
in months (even an estimate). In the last three years has it lengthened, shortened or
remained unchanged?”
2 - Interquartile difference over median, in percent.
Table 4 – Average tenure of branch managers (absolute values and percentages)
0
10
20
30
40
50
60
70
80
90
Medium-sized/large banks
Small group-member banks
Small stand-alone banks
Mutual banks
Source: Sample survey of 322 banks.
1 - Percentage of banks that describe at least one of the factors shown in Table 5 as “very
important” for the compensation of branch managers. The horizontal line represents the
overall sample mean.
Figure 2 – Use of economic incentives for branch managers1
The Capco Institute Journal of Financial TransformationThe Organization of Lending and the Use of Credit Scoring Techniques in Italian Banks
152
broadly uniform within bank classes. The heightened mobility of branch
managers may be related to the introduction of ICT, which in practice
may have reduced the rents deriving from close local bank-firm relations.
It may also have led to the increase in head-branch distance, which in
turn presumably increases the costs of monitoring local managers’ ac-
tivities; to mergers and corporate restructuring, which have affected a
large number of banks since the 1990s; and to heightened competition
in credit markets, leading in turn to stiffer competition in the local bank
managers’ markets.
Branch managers’ incentivesThe potential costs connected with the distance between the head office
and local units and with the misalignment between the personal objec-
tives of local agents and those of the bank as such can be mitigated by
incentive systems linking agents’ compensation to results [Milgrom and
Roberts (1992)]. However, incentive systems can also entail an excessive
transfer of risk to the local agent.
Incentives are most common in the larger banks (Figure 2), with 83 per-
cent of the large banks in our sample stating that the use of incentives
in connection with lending for branch managers’ compensation is “very
important” compared with an overall sample mean of 57 percent. Mu-
tual banks make the least use of incentives (46 percent). This may be
explained by the higher incidence of agency costs in larger banks, i.e.,
a greater geographical and organizational distance between the center
and the periphery leads to a lower ability to monitor local agents’ activ-
ity. Consequently, incentives should help to align the branch managers’
goals with the bank’s objectives.
Regarding the factors on which incentives are based, the most com-
mon response was the overall profitability of the local unit (i.e., gross
income). This factor was especially important for large banks and small
group-member banks (Table 5). Practically nine-tenths of all banks using
incentives stated that the overall profitability of the branch was a very
significant factor. By contrast, small stand-alone banks and mutual banks
are much more sensitive to bad loan ratios or to variations in bad loans.
To simplify, larger banks and those belonging to groups tend to link lo-
cal managers’ incentives to the profitability of the branch and of its loan
portfolio, whereas other banks, including mutual banks, tend to stress
the containment of bad loans and limitation of credit risk.
Medium-
sized/
large
banks
Small
group-
member
banks
Small
stand-
alone
banks
Mutual
banks
Total
banks
Growth in lending 28.6 19.7 33.3 36.1 29.6
Bad debt and/or impaired
loan rate
5.7 16.4 37.5 53.0 32.0
Change in bad debts
and/or impaired loans
8.6 21.3 58.3 49.4 35.0
Net earnings on loan
portfolio
25.7 14.7 29.2 19.3 20.2
Overall profitability of unit
(i.e., gross income)
88.6 90.2 62.5 60.2 74.4
Average potential
riskiness of loan portfolio
11.4 18.0 20.8 32.5 23.2
Source: Sample survey of 322 banks.
1 - Percentage of banks that consider each factor “very important” for determining branch
managers” compensation. The choices were “very important,” “fairly important,” “not
very important” and “not important at all”. Sample limited to banks using incentives
linked to the factors specified.
Table 5 – Factors considered in determining incentives for branch managers1 (percentages)
Medium-
sized/
large
banks
Small
group-
member
banks
Small
stand-
alone
banks
Mutual
banks
Total
banks
Growth in lending 28.1 25.0 13.6 5.2 15.9
Bad debt and/or impaired
loan rate
34.6 33.3 31.6 48.6 40.1
Change in bad debts
and/or impaired loans
39.3 42.6 36.8 48.0 43.7
Net earnings on loan
portfolio
23.1 19.2 35.0 20.9 22.5
Overall profitability of unit
(e.g., gross income)
26.5 48.3 36.4 41.0 40.1
Average potential
riskiness of loan portfolio
48.3 33.3 33.3 43.6 40.3
Source: Sample survey of 322 banks.
1 - Percentage of banks reporting an increase in the last three years in the importance of
the factors indicated in branch managers” compensation. The possible answers were
“increased,” “essentially unchanged,” “decreased” and “not relevant”. Sample limited to
banks indicating a tendency.
Table 6 – Trend in use of incentives for branch managers1 (percentages)
46,9
56,5
83,8
42,2
59,4
64,1
97,3
0 20 40 60 80 100 120
Mutual banks
Small stand-alone banks
Small group-member banks
Medium-sized and large banks
SMEs
Large corporates
Source: Sample survey of 322 banks
Figure 3 – Diffusion of credit scoring (percentages; unweighted frequencies)
153
In the last three years, the banks have made greater use of incentive
schemes for branch managers’ compensation (Table 6). The factors
whose relative importance has increased have been gains in the overall
profitability of the branch and the containment of bad loans. The share
of banks reporting an increase in the relevance attached to these factors
was more than 40 percent compared to fewer than 5 percent that report-
ed a reduction. Again, the mutual banks paid the closest attention to the
incidence and variation of bad loans and substandard loan assets.
The importance of credit scoring techniques in assessing creditworthinessThe diffusion of credit scoring techniquesThe plunging cost of data processing in recent years has fostered banks
to introduce statistical techniques for measuring credit risk, supplement-
ing their external and internal sources of information. Our survey of the
Italian banking system shows the diffusion of credit scoring techniques
for business lending. Consistent with our hypotheses in the introduc-
tory section, and as other studies have shown [Bofondi and Lotti (2005)],
these techniques have been mainly adopted by larger banks with exten-
sive branch networks that can exploit economies of scale.
At the end of 2006, 57 percent of the sample banks had scoring tech-
niques in place to assess the creditworthiness of firms, whether large or
small. However, the distribution was not uniform by the type of intermedi-
ary; diffusion reached 97 percent for medium-sized and large banks, 64
percent for small banks belonging to groups, 59 percent for small stand-
alone banks, and just over 40 percent for mutual banks (Figure 3). More-
over, scoring techniques were systematically more common for lending
to SMEs. For large banks, the difference was more than 13 percentage
points (Figure 4), i.e., the prevailing tendency is to use these techniques
to reduce selection costs for smaller firms, even where the bank is less
specialized in this segment, thus freeing resources for screening in the
core business segment.
Italian banks began to adopt the scoring techniques substantially in
2003, and their introduction has been by degrees, accelerating sharply in
recent years. In 2000, less than 10 percent of the banks had such tech-
niques in place, whereas in 2003, almost 25 percent did, and in 2006,
more than half did (Figure 4).
The spread of scoring techniques in recent years is presumably related
to the new capital adequacy accords (Basel II), which link capital require-
ments more directly to customer risk with incentives for a more accurate
evaluation of the quality of the loan portfolio.13 The credit scoring tech-
niques adopted by Italian banks differ both in their origin (internal/exter-
nal) and in the data used. At the end of 2006, more than 50 percent of the
banks that had actively participated in the development and introduction
of a credit scoring methodology, either independently or in cooperation
with other institutions or consortia (Table 7). The degree of participation
is correlated with bank size. Nearly all of the larger banks had contributed
The Capco Institute Journal of Financial TransformationThe Organization of Lending and the Use of Credit Scoring Techniques in Italian Banks
After classifying their retail customers by risk under the internal ratings approach, banks 13
must estimate the main risk components for each class and then calculate, by the Basel
Committee’s methodology, specific capital requirements for each. Consequently, the
introduction of credit scoring may be highly advantageous to the banks, insofar as it can
lower capital charges [Jankowitsch et al. (2007)]. So far, very few banks, and only large
ones, have begun to adopt these methodologies for calculating capital requirements. To
do so, banks must meet stringent qualitative and quantitative requirements that are subject
to a complex process of supervisory validation. However, the possibility that the internal
models may be validated and recognized for supervisory purposes has nevertheless
fostered their diffusion [Bank for International Settlements (2005)] by stimulating studies on
the methodologies by specialized companies and by creating incentives for initiatives by
consortiums of institutions.
0
10
20
30
40
50
60
70
80
90
100
2000 2001 2002 2003 2004 2005 2006
Large corporates
0
10
20
30
40
50
60
70
80
90
100
2000 2001 2002 2003 2004 2005 2006
SMEs
Medium-sized and large banks
Small group-member banks
Small stand-alone banks
Mutual banks
Medium-sized and large banks
Small group-member banks
Small stand-alone banks
Source: Sample survey of 322 banks
Figure 4 – Introduction of credit scoring (percentages; unweighted frequencies)
154
actively to the development of the scoring method; the contribution from
small banks was less common.
The data used for credit scoringOne of the benefits derived from credit scoring involves the management
of the data available to banks. Banks can now fully exploit this informa-
tion, integrating and combining data for systematic, replicable use. How-
ever, accurate data extending over a suitably long period are indispens-
able to the reliability of the models’ forecasts. The new techniques also
impose a standardization of the documentation required for loan applica-
tions, which, among other things, facilitates subsequent securitization.
The adaptation of internal information systems originally designed for dif-
ferent purposes is one of the most serious problems and has slowed the
introduction of the new techniques.14 Consequently, the models focus on
the factors that have traditionally been used to assess creditworthiness
(firms’ financial statements and credit history), whereas other data, both
from external sources and available within the banking group, are used
less frequently (Table 8). The question on this matter was phrased in ordi-
nal terms, asking respondents to rank the various sources of information
by importance. Table 9 reconstructs the ranking based on the frequency
of the answers “very important” and “decisive.”
There are significant differences in the weights of the factors used in credit
scoring for SMEs. For mutual banks (and for the small banks), the most
important factor is the financial statement, followed by the credit history
with the bank and with the rest of the banking system. The larger banks, by
contrast, assign greater importance to the firm’s past credit performance
than to its accounting data (Table 9). Less importance is attached to the
firm’s economic sector and geographical area, which in fact are not even
considered in many cases (about a third of the models, accounting for 18
percent of loans). Other external data sources, including the interbank reg-
ister of bad checks and payment cards and the Chamber of Commerce’s
database, are of little importance and are often not used at all. Large banks
and mutual banks are the ones that most commonly ignore them (31 and
26 percent, respectively, corresponding to 38 and 19 percent of loans).
Qualitative information is generally included in the estimation models, al-
though with relatively modest weight in the bank’s overall assessment. Fi-
nally, large banks also consider any relations between the firm and other
members of the bank’s group; even so, about 40 percent of the large banks
do not include this information in their models.
In evaluating the creditworthiness of large firms, the importance of the
various sources of information used by the larger banks is similar to that
used for small firms, but there are some differences in the relative rank-
ing. In the former case, greater importance is attached to the company
accounts and especially to qualitative information generally relating to
organizational structure, the stature of the management, and the quality
of the investment project to be financed (Table 8). Overall, the models
appear to be more flexible, as they allow for more judgmental compo-
nents. For large firms, more attention is paid to the state of the borrower’s
existing relations with the other members of the bank’s group, although a
substantial 35 percent of large banks do not consider this factor.
The importance of credit scoring techniques in assessing creditworthinessStatistical scoring techniques have gained considerable importance in
the lending process, in particular in the decision of whether or not to grant
a loan. In most cases the score is decisive or very important (Table 10
and Figure 5). There are significant differences between banking groups,
and the relative importance of quantitative techniques is definitely greater
among the larger banks and decreases with the size of the institutions.
Generally, the rating/scoring influences the size of the loan, and for the
smaller firms, it also affects the amount of collateral requested.
Although scoring techniques are widely used, they are still rarely em-
ployed to determine interest rates and loan maturities, suggesting the
early stage of the evolution of scoring models depicted by Hand and
Thomas (2005) and Hand and Zhou (2009). This lack of an immediate
impact of the borrower’s credit score on the interest rate that the bank
charges would appear to indicate that the effect found by the literature
The numerous interbank mergers of recent years have also complicated the integration of 14
data from different sources.
Bank class and
firm size Internal
In
collaboration
with other
institutions
Purchase
from group
company
Purchase
from outside
company Other
Lending to SMEs
Medium-sized/
large
46.8 47.6 3.3 2.4 0.0
Small group-
member
24.4 24.6 22.5 26.0 0.8
Small stand-alone 21.0 35.8 0.2 41.4 1.5
Mutual 11.9 40.5 18.3 24.6 4.7
Lending to large firms
Medium-sized/
large
52.6 38.0 6.1 3.3 0.0
Small group-
member
18.6 21.7 19.6 25.7 14.4
Small stand-alone 33.6 35.5 0.6 30.3 0.0
Source: Sample survey of 322 banks.
1 - Percentage frequency of responses, weighted by volume of lending to SMEs and to
large firms, respectively.
Table 7 – How scoring models were developed1 (percentages)
155
in other countries, i.e., an expansion of the volume of lending to mar-
ginal customers at higher interest rates, has not (yet) emerged in Italy.
However, as shown consistently with developments in risk management
and control, these methodologies have been widely used to monitor the
situation of firms and the state of loans and accounts. The differences
in practices concerning small and large firms do not appear to be pro-
nounced, although there is a somewhat greater tendency to use ratings
to determine the pricing of loans to the larger firms (Table 10).
The bank’s flexibility in using scoring techniques is highly variable. It
depends on the characteristics of the procedures chosen, which may
The Capco Institute Journal of Financial TransformationThe Organization of Lending and the Use of Credit Scoring Techniques in Italian Banks
0
10
20
30
40
50
60
70
80
90
100
Decision Amount Pricing Maturity Collateral Monitoring
Medium-sized and large banks
Small group-member banks
Small stand-alone banks
Mutual banks
Source: Sample survey of 322 banks
Figure 5 – Importance of scoring in lending to SMEs (frequency of “decisive” and “very important” answers, weighted by lending to SMEs)
SMEs Large firms
Information used2
Medium-sized/
large banks
Small group-
member banks
Small stand-alone
banks Mutual banks
Medium-sized/
large banks
Small group-
member banks
Small stand-alone
banks
Financial statement data 46.5 71.0 56.1 82.6 86.7 77.5 29.9
Geographical area and
economic sector
2.1 1.3 1.7 2.4 0.0 1.2 5.8
Credit relations with entire
system3
85.1 35.6 75.4 50.0 35.0 47.3 76.1
Other outside sources of data4 11.8 9.8 9.8 9.8 3.1 7.3 16.0
Relations with bank 50.4 63.9 54.0 46.9 27.3 50.0 47.3
Relations with bank’s group 3.2 6.9 0.0 0.0 8.7 9.3 0.0
Qualitative information5 4.7 13.4 5.5 3.8 33.7 8.8 22.6
Source: Sample survey of 322 banks.
1 - Sum of frequencies of responses indicating each source as one of the two most important, weighted by volume of lending to SMEs and large firms, respectively. Data for 2006.
2 - Banks were asked: “If you use statistical-quantitative methodologies for assessing firms’ creditworthiness, please rank by decreasing order of importance the data considered in your “calculation
engine” in assigning the overall score: 1 for the most important, 2 for the next most important, and so on. If you do not use any particular factor, answer NA.”
3 - Source: Central Credit Register or other credit bureau.
4 - Interbank register of bad checks and payment cards, Protest bulletin, etc.
5 - Codifiable data, as through special questionnaires, on organization of the firm, characteristics of the project to fund, and so on.
Table 8 – Information used in scoring models1 (percentages)
Medium-sized and large banks Mutual banks
1st State of credit relations with the
banking system
Income statement and balance sheet
2nd State of relations with the bank State of credit relations with the
banking system
3rd Income statement and balance sheet State of relations with the bank
4th Other outside sources Other outside sources
5th Qualitative information Qualitative information
6th Relations with the banking group Area and sector of activity
Source: Sample survey of 322 banks
Table 9 – Information sources included in scoring systems (ranking of information sources by importance)
0
10
20
30
40
50
60
70
80
90
100
Decisi
on
Amou
nt
Mon
itorin
g
Decisi
on
Amou
nt
Mon
itorin
g
Decisi
on
Amou
nt
Mon
itorin
g
Decisi
on
Amou
nt
Mon
itorin
g
very important
decisive
Medium-sized and large banks
Small group-member banks
Mutual banks Small stand-alone banks
Source: Sample survey of 322 banks
Figure 6 – Flexibility in the use of scoring techniques (percentage share of answers “decisive” and “very important” in loan approvals to SMEs, weighted)
156
Bank class and
firm size Lend/not Amount Pricing Maturity Collateral Monitoring
Lending to SMEs
Medium-sized/
large
91.8 57.0 14.7 15.8 42.9 70.0
Small group-
member
67.0 39.3 15.5 6.7 34.1 73.7
Small stand-
alone
54.1 31.4 21.4 13.3 34.0 71.0
Mutual 47.5 31.8 16.5 17.5 31.9 48.5
Lending to large firms
Medium-sized/
large
88.0 70.2 20.3 29.0 35.6 82.6
Small group-
member
67.6 34.8 17.3 9.0 27.9 61.4
Small stand-
alone
50.1 33.2 32.1 14.3 32.1 81.9
Source: Sample survey of 322 banks.
1 - Banks were asked to: “Rank from 1 to 5, in decreasing order of importance. 1=decisive,
2=very important, 3=fairly important, 4=not very important, 5=not important at all.
NA=not applicable.” The table gives the sum of the frequencies of answers 1 and 2
(decisive or very important), sample limited to banks that use statistical-quantitative
methods. Data for the end of 2006. Frequencies weighted by volume of lending to SMEs
and large firms respectively.
Table 10 – Importance of scoring models in lending decisions1 (percentages)
SMEs Large firms
Information used2
Medium-sized/
large banks
Small group-
member banks
Small stand-alone
banks Mutual banks
Medium-sized/
large banks
Small group-
member banks
Small stand-alone
banks
Statistical-quantitative methods 70.2 27.6 18.9 8.9 59.6 32.9 0.0
Financial statement data 95.2 85.7 95.2 96.5 100.0 95.2 98.2
Credit relations with entire
system3
82.6 86.7 97.2 89.5 72.0 92.5 98.2
Availability of guarantees4 28.3 51.8 45.7 42.0 3.9 24.4 33.5
Qualitative information5 35.8 33.5 33.4 49.9 69.2 48.8 61.8
First-hand information 16.3 15.9 9.6 15.0 3.9 4.7 2.9
Source: Sample survey of 322 banks.
1 - Sum of frequencies of responses indicating each source as one of the two most important, weighted by volume of lending to SMEs and large firms, respectively. Data for the end of 2006.
2 - The banks were asked: “For the granting of loans to non-financial firms that apply to you for the first time, please rank in decreasing order of importance the factors used in deciding whether or
not to grant the loan. 1 for the most important, 2 for the next most important, and so on. Two different factors cannot be given the same rank. If you do not use any particular factor, answer NA.”
3 - Source: Central Credit Register or other credit bureau.
4 - Interbank register of bad checks and payment cards, Protest bulletin, etc.
5 - Codifiable data, as through special questionnaires, on organization of the firm, characteristics of the project to fund, and so on.
Table 11 – Importance of factors in assessing creditworthiness of new loan applicant1 (percentages)
enable that adaptation of scores to account for elements of information
not included in the model, but it also depends on the actual importance
of the scores in lending decisions and management, i.e., whether they
are the main evaluation tool or a supplement to another method of as-
sessment. The models’ degree of flexibility was gauged with reference to
“decisive” as one of the possible responses to the importance of scor-
ing methods. “Decisive” was counted separately from “very important”
specifically to capture the possibility of loan officers derogating from the
automatic scores; the answer “decisive” was interpreted to mean practi-
cally no such flexibility. In all cases, the scores were more binding for
large banks than for small banks.
For decisions on whether to lend to SMEs, credit scoring tools, while im-
portant, are “decisive” for only 18 percent of the entire sample of banks
and for a third of the large banks; weighted by volume of loans, the fre-
quencies are higher (Figure 6), confirming that the discretionary power
of loan officers tends to diminish as the size of the bank (and of the bor-
rower firm) increases. Further, greater specialization in lending to SMEs
corresponds to lower weight assigned to scoring in loan decisions.
As we have seen, the likelihood of a bank developing its own scoring sys-
tem internally, at least in part, increases with bank size. The purchase of
a statistical procedure from the outside could reduce the bank’s control
over its own instrument for selecting borrowers, fostering the perception
of the system as a “black box,” both initially in relation to the algorithm
and database and then subsequently at times of revision (Frame et al.
(2001)]. Our survey shows that in each size class of banks, credit scoring
techniques for SMEs are more frequently decisive or very important in
lending decisions when they are developed by the bank itself.
157
The information used in deciding on loan applicationsTechnological change has affected the phase of selection of borrowers,
making it easier to build quantitative models that can sum up the data on
potential borrowers. Credit scoring techniques can be used to enhance
the information from other sources that banks ordinarily use in screen-
ing borrowers, these techniques can even replace the other scores and
become the main means of evaluation. One of our questions was what
importance the bank assigned to the various sources of information it
used when deciding whether or not to grant a loan to a first-time appli-
cant; the output of the statistical model is just one such source and is not
always the most important.
The results (Table 11) are in line with expectations. For loans to SMEs,
scoring methods are assigned high importance more frequently by larger
banks and less frequently by smaller ones, whereas qualitative informa-
tion is emphasized more commonly by mutual banks. In selecting large
corporate borrowers and not SMEs, the statistical models are less im-
portant and qualitative information is more so. Finally, small and mutual
banks tend to assign considerable weight to the loan applicant’s credit
relations with the entire system, a tendency that is accentuated when the
potential borrower is a larger firm.
ConclusionSince the 1990s, two significant drivers of change have been affecting
the Italian banking system: liberalization and the widespread adoption
of information and communication technologies. The two forces among
others generated an intense M&A activity and an extensive evolution of
the organization of financial intermediaries. This paper focuses on the lat-
ter topic. According to the recent banking literature organizational issues
do affect lending activity, especially in the case of SMEs, by influencing
the choice of the lending technology and by shaping banks’ propensity
and ability to deal with different types of information.
As a result of an ad-hoc survey, we find that Italian banks are very hetero-
geneous in organizational choices. Size explains only to a partial extent
such differences. Heterogeneity involves: the distance between a bank’s
headquarter and its branches network; the decision-making autonomy of
the branch managers; the length of their tenure, and the incentives they
face. A special consideration has been devoted to the adoption and use
of scoring/rating systems, boosted by the evolution of Capital Adequacy
Accords. The impact of these statistical models differs widely among
banks considering the type of information used to produce automatic
scores and the relevance of their role in credit decisions, such as grant-
ing credit, setting interest rates, and collateral requirements. All in all, our
paper suggests that analyzing lending activity requires a taxonomy that
is broader than the traditional one based on size only.
ReferencesAcemoglu, D., P. Aghion, C. Lelarge, J. Van Reenen and F. Zilibotti, 2006, “Technology, •
information and the decentralization of the firm,” mimeo
Aghion, P. and J. Tirole, 1997, “Formal and real authority in organizations,, Journal of Political •
Economy, 105, 1-29
Allen, L., G. DeLong, and A. Saunders, 2004, “Issues in the credit risk modelling of retail •
markets,” Journal of Banking and Finance, 28: 727-752
Bank for International Settlements, 2005, “Studies on the validation of internal rating systems, •
Basel Committee on Banking Supervision,” Working Paper, 14, May
Berger, A. N., 2003, “The economic effects of technological progress: evidence from the •
banking industry,” Journal of Money Credit and Banking, 35, 141-176
Berger, A. N. and R. DeYoung, 2001, “The effects of geographic expansion on bank efficiency,” •
Journal of Financial Services Research, 19, 163-84
Berger, A. N., W. S. Frame, and N. H. Miller, 2005, “Credit scoring and the availability, price, and •
risk of small business credit,” Journal of Money, Credit and Banking, 191-222
Berger, A. N., N. H. Miller, M. A. Petersen, R. G. Rajan, and J. C. Stein, 2005, “Does function •
follow organizational form? Evidence from the lending practices of large and small banks,”
Journal of Financial Economics, 76, 237-269
Bofondi, M. and F. Lotti, 2005, “Innovation in the retail banking industry: the diffusion of credit •
scoring,” Review of Industrial Organization, 28:4, 343-358
Bonaccorsi di Patti, E., G. Eramo, and G. Gobbi, 2005, “Piccole e grandi banche nel mercato •
del credito in Italia”, Banca Impresa Società, XXIV:1, 3-34
Cannari, L., M. Pagnini, and P. Rossi, (eds.), 2010, “Banks, local credit markets and credit •
supply,” collection of the papers presented at the Conference held in Milano on 24 March
Christie, A., M. P. Joyeb, and R. L. Watts, 2003, “Decentralization of the firm: theory and •
evidence,” Journal of Corporate Finance, 9, 3–36
Cowan, C. D. and A. M. Cowan, 2006, “A survey based approach of financial institution use of •
credit scoring for small business lending,” Working Paper, Office of Advocacy, United States
Small Business Administration, SBAH-04-Q-0021
Degryse, H. and S. Ongena, 2004, “The impact of technology and regulation on the •
geographical scope of banking,” Oxford Review of Economic Policy, 20, 571-590
Deng, S. and E. Elyasiani, 2005, “Geographic diversification, distance, BHC return and risk,” •
mimeo
Felici, R. and M. Pagnini, 2008, “Distance, bank heterogeneity and entry in local banking •
markets,” Journal of Industrial Economics, 56, 500-534
Ferri, G., 1997, “Mobilità dei dirigenti ed efficienza allocativa: banche locali e nazionali,” •
Quaderno di Moneta e Credito, 245-265
Frame, W. S., A. Srinivasan, and L. Woosley, 2001, “The effect of credit scoring on small •
business lending,” Journal of Money Credit and Banking, 33, 813-825
Hand, D. J. and L. C. Thomas, 2005, “A survey of the issues in consumer credit modelling •
research,” Journal of the Operational Research Society, 56, 1006-1015
Hand, D. J. and F. Zhou, 2009, “Evaluating models for classifying customers in retail banking •
collections,” Journal of the Operational Research Society, 61, 1540-1547
Hertzberg, A., J. M. Liberti, and D. Paravisini, 2007, “Information and incentives inside the firm: •
evidence from loan officer rotation,” mimeo
Jankowitsch, R., S. Pichler, and W. S. A. Schwaiger, 2007, “Modelling the economic value of •
credit rating systems,” Journal of Banking and Finance, 31:1, 181-198
Liberti, J. M., 2005, “How does organizational form matter? Distance, communication and soft •
information,” mimeo
Liberti, J. M. and A. Mian, 2006, “Estimating the effect of hierarchies on information use,” •
mimeo
Milgrom, P. and J. Roberts, 1992, Economics, organization and management, Prentice Hall, NJ•
Scott, J., 2006, “Loan officer turnover and credit availability for small firms,” Journal of Small •
Business Management, 44:4, 544–562
Stein, J. C., 2002, “Information production and capital allocation: decentralized versus •
hierarchical firms,” Journal of Finance, 57, 1891-1921
Uchida, H., G. F. Udell, and N. Yamori, 2006, “Loan officers and relationship lending,” mimeo•
The Capco Institute Journal of Financial TransformationThe Organization of Lending and the Use of Credit Scoring Techniques in Italian Banks
159
PART 2
Measuring the Economic Gains of Mergers and Acquisitions: Is it Time for a Change?
AbstractIn this paper we review the methods of measuring the eco-
nomic gains of mergers and acquisitions (M&A). We show
that the widely employed event study methodology, whether
for short or long event windows, has failed to provide mean-
ingful insight and usable lessons regarding the central ques-
tion of whether mergers and acquisitions create value. We
believe the right way to assess the success and therefore
the desirability of M&A is through a thorough analysis of
company fundamentals. This will require examining smaller
samples of transactions with similar characteristics.
Antonios Antoniou — FRT-C Consulting
Philippe Arbour — Director, Lloyds Bank Corporate Markets – Acquisition Finance
Huainan Zhao — Associate Professor of Finance, University of Nottingham1
The views expressed in this article are those of the authors and not 1
representative of the views of the authors’ affiliations.
160
The development of the market for mergers and acquisitions (M&A)
has gone hand in hand with the emergence of world capital markets.
The corporate landscape is perpetually being altered by M&A transac-
tions. On a macroeconomic level, mergers come in waves, with one of
the most memorable waves gaining momentum during the early 1990s
and crashing shortly after the turn of the millennium. During this period,
deregulation, a booming world economy, strong consumer and investor
confidence, combined with rich stock market valuations propelled the
economic significance of M&A to new heights. Indeed, during the late
1990s, the size, volume, and frequency of M&A transactions surpassed
anything the world had ever seen. On a microeconomic level, mergers
represent massive asset reallocations within and across industries, often
enabling firms to double in size in a matter of months. Because mergers
tend to occur in waves and cluster by industry, it is easily understood
that such transactions may radically and swiftly change the competitive
architecture of affected industries. It should, therefore, come as no sur-
prise that academics have been so intrigued by the merger debate in
recent years.
Examining the economic gains (value creation or destruction) of M&A
is one of the most coveted research areas in financial economics. The
spectacular growth of mergers has justifiably prompted many academ-
ics and practitioners to investigate whether such milestone transactions
are worth undertaking. More specifically, researchers have sought to find
out whether M&A create or destroy value and how the potential gains or
losses are distributed among transaction participants. If synergies truly
exist between bidders and their targets, M&A should have the potential of
representing value-creating corporate events. This question is of utmost
importance as its corresponding answer carries important policy implica-
tions for regulators and investors. Furthermore, it is vital to assess the
aftermath of these colossal transactions, as lessons learned may benefit
not only the corporate world, but also the society at large.
Although a plethora of research in financial economics has sought to ad-
dress the issue of M&A value creation generally, the investigation of how
value is created (or destroyed) and the examination of the question from
a company fundamentals standpoint has largely been ignored. The bulk
of the existing literature employs event study methodology as introduced
and popularized by Fama et al. (1969), which examines what impact, if
any, mergers and acquisitions have on stock prices. In accordance with
this methodology, a merger is branded successful if the combined en-
tity equity returns equal or exceed those predicted by some standard
benchmark model. For reasons argued below, this simplistic approach
too often leads to a Type II error (i.e., the null hypothesis is not rejected
when in fact it is false and should be rejected) with respect to the null
hypothesis that M&A are value-creating transactions. We invite readers
to review the evidence and arguments presented in this article and to
judge whether the event study is an appropriate tool, let alone worthy
of “gold-standard” status by some financial economists, for tackling the
question of whether M&A result in economic gains on an ex-post basis.
We emphasize that this article does not constitute an attack on the event
study in general, but rather an objection to the use of event studies as the
main academic investigative tool in assessing whether M&A represent
value-creating corporate events.
Short-window event study From a short-run perspective, the most commonly studied event window
encompasses the three or five days surrounding a merger announce-
ment. From a theoretical standpoint, and in the context of an efficient
market, changes in stock market valuations around merger announce-
ments should fully capture the economic gains (or losses) from merg-
ing. Following this argument, it is often argued that the abnormal returns
measured during the short-run event window (if any), represent reliable
predictors of the success (or failure) of the M&A transactions under
evaluation. The event study literature is unanimous in stating that target
firm shareholders enjoy a significant positive cumulative abnormal return
(CAR) around merger announcements. This finding, however, should be
expected and merely represents a statement of the obvious. Intuitively,
target firm shareholders expect to receive a premium if they are to hand
over their ownership stakes to the acquiring firm and/or if the bidding
firm is hoping, via the attractiveness of its bid, to persuade the target
firm’s board of directors to issue a public statement in recommendation
of the offer. It should consequently come as no surprise that positive
CARs accrue to target firm shareholders during the period surrounding
merger announcements. Provided the acquisition involves a target that is
believed to remain a going concern, the CARs earned by the target firm
shareholders will invariably be positive as long as a positive premium is
offered, irrespective of the identity of the bidder and perhaps even of the
synergy potential between the bidder and its target.
The effect of takeover announcements on the acquiring firms’ share pric-
es is far from clear. On the one hand, some short-window event studies
have found that no or small significant positive abnormal returns accrue
to acquiring firm shareholders around merger announcements [Dodd
and Ruback (1977); Asquith (1983); Asquith et al. (1983); Dennis and Mc-
Connell (1986); Bradley et al. (1988); Franks and Harris (1989)]. On the
other hand, others have reported that acquirers experience significant
but small negative abnormal returns over the same period [Firth (1980);
Dodd (1980); Sudarsanam et al. (1996); Draper and Paudyal (1999)]. In
short, the general picture that emerges is that, from a short-window
shareholder returns standpoint, M&A are clearly more beneficial to target
firm shareholders than to their respective suitors, a fact that is widely ac-
knowledged in the literature. This finding is not particularly encouraging,
however, as it is, after all, the acquiring firm which makes the investment
and ultimately continues on its journey.
161
The Capco Institute Journal of Financial TransformationMeasuring the Economic Gains of Mergers and Acquisitions: Is it Time for a Change?
Using a weighted-average approach based on firm size as measured
by market capitalization allows for assessing how combined (target and
acquirer) stock returns fare in the same event window. Many believe that
this type of study enables us to evaluate the net aggregate economic im-
pact from mergers and acquisitions. Indeed, some financial economists
contend that this type of analysis enables us to assess whether or not
M&A result in real economic gains or whether these transactions simply
involve a transfer of wealth from one entity to the other (i.e., a zero-sum
game). The literature broadly concludes that the combined entity earns
a small, albeit positive CAR around the merger announcement [Bradley
et al. (1988); Mulherin and Boone (2000); Andrade et al. (2001)]. But are
conclusions from such studies sufficient to draw high level inferences
about the true value creation potential or desirability of M&A? Many
believe so. Andrade et al. (2001) refer to short-window event studies
as: “The most statistically reliable evidence on whether mergers create
value for shareholders…”. They go on to conclude that: “[b]ased on the
announcement-period stock market response, we conclude that merg-
ers create value on behalf of the shareholders of the combined firms.”
This, we argue, is surely a premature conclusion as the hefty premiums
paid (which could turn out to be overpayments especially under agency
or hubris motives) blur the picture of whether M&A truly represent bene-
ficial corporate events. More specifically, the premiums offered to target
firm shareholders distort or bias weighted-average return calculations.
Let us now explain in more detail why we believe that short-window
event studies demonstrate very little with respect to the M&A value cre-
ation issue.
First, examining stock price movements around the merger announce-
ment tells us little about the sources of economic gains that arise from
combining the target and the acquirer (which is of course a high priority
question for M&A practitioners). Shelton (1988) writes: “value is created
when the assets are used more effectively by the combined entity than
by the target and the bidder separately.” Short-window event studies do
nothing to test this statement. In fact, short-window event studies and
their associated conclusions rely on strict assumptions in respect of mar-
ket efficiency. However, it is possible that investors systematically over- or
underreact to merger announcements, which could result in stock prices
temporarily deviating from their fundamental levels. If this is accepted to
be possible, the event study’s ability to distinguish real economic gains
from market inefficiency is compromised. As Healy et al. (1992) put it:
“[f]rom a stock price perspective, the anticipation of real economic gains
is observationally equivalent to market mispricing.” Indeed, the mount-
ing body of behavioral finance literature illustrates the need to approach
short-run event study results with skepticism. Furthermore, there is plen-
ty of evidence which suggests that stock market wealth may temporarily
move independently of fundamentals. The U.S.-centric dot.com stock
market bubble of the late 1990s and the more recent global post-Lehman
financial meltdown are irrefutable examples of this assertion.
Shojai (2009) points out that a significant level of information asymmetry
exists between the bidding firm M&A teams and the management of the
target firms that are in play. In the context of a bid for a listed company
(either by way of a “take-private” or straight M&A transaction involving
two publicly listed firms), bidders must contend with significant head-
winds including limited access to the target firms they are trying to buy.
As such, bidders will typically carry out their due diligence exercise from
the “outside-in” (i.e., management will rely on the work of advisors, bro-
kers, consultants, and other industry specialists in order to arrive at a rea-
sonable guesstimate about the likely economic state of the target), which
necessarily implies that bidders must take a leap of faith in completing
the acquisition.2 The ability to refine synergy estimates is a function of
time and access, with unrestricted access only becoming possible once
the deal is done. Post-acquisition, corporate development teams will fi-
nally gain unfettered access to the target’s books and to key target firm
personnel, allowing acquirers to finetune their original synergy estimates
and to establish targets and timescales for the realization of the syner-
gies. Bidding firm management will then prioritize synergy work streams
based on contribution potential, deliverability risk, level of operational
disruption, management time, and cost. In short, bidding firm manage-
ment will not have a clear picture of performance versus budgets until
they have traveled down the synergy extraction path. If bidders them-
selves do not know precisely what they are buying until after it is too late,
then is it reasonable to assume that the market should be equally capable
of predicting the outcome of a particular M&A transaction in the space of
a few days? If we accept the premise that markets at times struggle to
be efficient, then it follows that the short-window event study’s ability to
reliably measure real economic gains is compromised.
But even in the context of a reasonably efficient market, short-window
event study results remain problematic due to the substantial premiums
paid to target firm shareholders, which result in a bloated weighted-av-
erage CAR (WACAR) calculation and hence a potentially spurious result.
Put differently, weighted average calculations are almost guaranteed to
generate a positive result upon the inclusion of target firm abnormal re-
turns. Consequently, we believe that examining weighted-average com-
bined-entity CARs around merger announcements does not advance the
M&A value creation debate.
This issue can be illustrated with a simple example. Assume that a cash
tender offer is made by “Bidder Inc.” for the acquisition of “Target Inc.”
One month prior to the takeover announcement, the market value (MV)
of Bidder Inc. is $100,000,000, with a stock price trading at $2 per share,
Even where “data room access” is granted to the bidder, the problem of information 2
asymmetry still exists due to selective disclosure of key/sensitive documents by target
firm management and the vendors of the business (i.e., in the case of a major shareholder
selling its block of shares).
162
while Target Inc. has a market value of $10,000,000 and a stock price
of $1 per share. We employ the CAPM as our simple benchmark model
and assume a risk-free rate (rf) of 2 percent and a market risk premium of
6 percent. We also make the assumption that Bidder and Target have a
beta of one and two respectively. According to our selected benchmark
model, the annual expected returns for the Bidder and Target firms are
therefore 8 percent and 14 percent respectively,3 or 6.67 basis point (bp)
and 11.67 bp respectively during a three-day window.4 Based on this
information, the relative size of the target to the bidder is 10 percent,
which implies that weights in calculating the weighted-average cumula-
tive abnormal return (WACAR) for the combined entity would be Wb =
90.9 percent and Wt = 9.1 percent respectively.5 Now suppose that the
acquirer announces a cash offer at $1.40 per share for the target, which
represents a premium of 40 percent per share purchased. In a relatively
efficient market, the price of the target’s share price will adjust to the offer
price quickly. Hence, the market price of Target Inc.’s shares should shoot
up to the $1.40 range in the three days surrounding the announcement.
This necessarily implies that the three-day CAR for the target would be
just shy of 40 percent.6 Because the target in this example is small rela-
tive to the acquirer, it is not unreasonable to assume that the acquirer
should earn the expected rate of return in the three days surrounding
the announcement.7 Thus, on a weighted-average net aggregate basis,
the combined entity’s CAR (WACAR) would be around 3.63 percent.8 Ac-
cording to short-window event study proponents, there is no question
that the acquisition in this example would be branded as having been
value-creating and therefore desirable.
Changing the example slightly, assume that Bidder Inc.’s shareholders
do not share the same optimism regarding the union because they be-
lieve that their management is paying too much to acquire Target Inc., or
because they fear that the acquisition signals the beginning of a buying
spree by the bidder. Thus, shareholders may decide to sell Bidder Inc.’s
shares, which may result in the acquiring firm earning a negative CAR
around the merger announcement (assume CAR= -2.00 percent).9 Us-
ing the same weights and premium as described above, the WACAR for
the combined entity will still be approximately 1.81 percent. That is, the
negative bidder CAR of -2.00 percent corresponds to $2,000,000 of bid-
ding firm market value destruction and yet event study proponents would
continue to designate this acquisition as having been value creating. All
else being equal, the acquiring firm shareholders in this example could
earn a negative return as large as 3.92 percent (which corresponds to a
market value loss of $3,920,000) and the acquisition would still have been
considered a success owing to the large target premium which supports
the positive result of the weighted-average calculation.11 This illustrates
what we call the premium exacerbation problem. The practical interpreta-
tion of this example, however, is that the acquiring firm shareholders of
the going concern entity (the acquirer) have suffered a -3.92 percent loss
in value in the three days surrounding the merger announcement (-470
percent annualized)12 which compares to a benchmark annual expected
return of 8 percent for the bidder and despite this, event study propo-
nents would have the audacity to claim that this acquisition has been
“value creating;” such a conclusion is clearly misguided and is illustrative
of the reasons why decades of academic research into M&A based on
the event study methodology have failed to make it anywhere near cor-
porate boardrooms [Shojai (2009)].
In reality, Moeller et al. (2004) report that the mean premium paid for over
12,000 U.S. takeover transactions with announcement dates between
1980 and 2001 was 68 percent for large firms and 62 percent for small
firms, which necessarily implies that, according to our illustration, most
(if not all) M&A transactions evaluated using a short-window event study
approach will generate a positive weighted-average CAR and will therefore
be branded as being successful, or value creating, in spite of the potential-
ly significantly negative returns accruing to acquiring firms during the same
period. The WACAR is almost invariably positive, and the problem is com-
pounded when the relative size of the target to the acquirer increases.
Indeed, premiums offered may easily represent overpayments [Roll
(1986)]. During times of high M&A activity, firms believed to be potential
takeover targets are likely to carry a substantial takeover premium as
part of their market capitalization. Hence, during M&A waves, acquiring
firms can end up paying a premium on top of what may already be an
overvalued target share price. Ironically, the WACAR approach to evalu-
ating M&A tends to reward overpayments, which invariably leads to the
conclusion that M&A are value-creating events.
Mitchell et al. (2004) examine price pressure effects around merger
announcements and find that on average, acquirers earn a significant
Bidder expected return based on CAPM: E(R3 b) = rf + b (Rm – rf) = 2% + 1(6%) = 8%. Target
return calculated on the same basis although beta = 2.
The three-day window expected return for the bidder is: (3/360)*8% = 6.67bp; Target: 4
(3/360)*14% = 11.67bp. For simplicity, we do not adjust the beta estimates for the short-
window returns calculation.
Weights are determined as follow: W5 t = MVt /(MVt+MVb) = Wt = 10,000,000 /(110,000,000) =
9.09%, and Wb = 1–Wt = 90.91%.
CAR = actual return – expected return = 40.00% -0.12% = 39.88%. 11.67bp is rounded to 6
0.12 percent for simplicity.
Prior research shows that on average acquirers break even in the few days surrounding 7
merger announcements. See, for instance, Asquith (1983), Dennis and McMconnell (1986),
Bradley et al. (1988), Andrade et al. (2001), etc.
WACAR = W8 t (CARt) + Wb (CARb) = (9.09%)*(39.88%) + (90.91%)*(0%) = 3.63%.
In three-day window, Andrade et al. (2001) find that the abnormal return is approximately 9
0 percent for bidders, regardless of the benchmark model used. An abnormal return of
–2 percent therefore represents an exaggerated estimate to prove our point.
WACAR = W10 t (CARt) + Wb (CARb) = (9.09%)*(39.88%) + (90.91%)*(–2%) = 1.81%. Bidder
market capitalization of $100,000,000 * -2% = -$ 2,000,000 of lost bidding firm market
value.
We use the goal seek function in Excel to solve for the bidder return (-3.92 percent) which, 11
holding the remainder of the example constant, will yield a WACAR of 0 percent. On a
base market capitalization of $100,000,000, this level of return translates into a bidding firm
market capitalization loss of $3,920,000.
-3.92% x (360/3) = -470.40%. 12
163
The Capco Institute Journal of Financial TransformationMeasuring the Economic Gains of Mergers and Acquisitions: Is it Time for a Change?
negative abnormal return of -1.2 percent in the three days surrounding
the announcement. However, they find that a substantial proportion of
this negative return is due to merger arbitrage short-selling, rather than
information, thereby contradicting the premise that stock returns solely
reflect changes in expectations regarding the present value of future cash
flows. After controlling for price pressure effects, however, acquirers’ an-
nouncement period abnormal returns shrink to -0.47 percent and be-
come statistically insignificant. Their findings demonstrate that conven-
tional short-run event study results may reflect more than just investors’
expectations regarding the desirability of mergers taking place.
Moeller et al. (2005) examine acquiring firm returns in recent merger
waves. In addition to testing abnormal percentage returns, they also
measure aggregate dollar returns.13 Strikingly, they find that between
1998 and 2001, acquiring firms’ three-day announcement period (-1; +1)
average CAR is 0.69 percent, while the aggregate dollar return14 measure
indicates that acquiring firm shareholders lose a total of U.S.$240 billion
over the window spanning from -2 to +1 days around the merger an-
nouncement (a result which is explained by the significant dollar losses
generated by some of the larger-size wealth-destroying acquisitions in
the sample). Upon further investigation, they also find that the losses
to acquirers exceeded the gains to targets, resulting in a net aggregate
dollar loss of U.S.$134 billion during the same window. These findings
provide interesting but rather painful evidence that if we merely rely on
the short-run event study result (i.e., the three-day CAR 0.69 percent), we
run the risk of concluding that the sample merger transactions are value-
creating, despite the massive aggregate shareholder losses experienced
during the same event window.15 The latter authors also show that be-
tween 1980 and 2001, the average three-day announcement period CAR
for acquirers is positive every year except for 2 out of 21 years, while the
three-day aggregate dollar returns are negative for 11 out of 21 years.
Once again, the three-day CARs tell us that mergers create value for
acquirers in almost every year between 1980 and 2001 regardless of the
massive dollar losses realized in over half of the period. If short-window
event studies fail to shed clarity on the “shareholder wealth effect” asso-
ciated with M&A, then their ability to answer the bigger-picture question
of whether M&A represent desirable corporate events hardly fills us with
confidence.
According to short-window event studies, mergers and acquisitions are
value-creating transactions. So why is there so much controversy sur-
rounding the desirability of M&A? That is, why do event study results
stand in such sharp contrast with the growing rhetoric that creating value
through M&A is easier said than done?16 And why do event study re-
sults stand in such stark contrast with investor experience? It is widely
acknowledged that many recent mergers have proven to be total disas-
ters. Even consultancy firms, which derive significant income in advising
companies on M&A-related matters, have documented the widespread
nature of these failures [Lajoux and Weston (1998)]. Academic studies
have also exposed the high prevalence of divestitures after acquisitions
[Kaplan and Weisbach (1992)]. If mergers are truly value-creating transac-
tions due to real economic gains and not market mispricings, it is highly
unlikely that acquirers would divest recent purchases at such a high fre-
quency, nor would there be so much controversy surrounding the desir-
ability of M&A generally.
Hitherto, we have shown that it may be naïve to conclude that M&A are
value-creating transactions based solely on the prevalence of positive
CARs or even weighted-average abnormal returns around merger an-
nouncements. In many cases, target firm shareholders may be the only
ones who gain anything from the transactions, and possibly to the detri-
ment of acquiring firms. But in retrospect, examining how target share-
holders fared around the bid announcement has very little relevance
to the questions being asked by M&A practitioners. Undeniably, at the
merger announcement and the few days surrounding it, we know very
little about any future possible negative drift in the acquirer’s share price,
or whether acquiring firm managers will succeed at unlocking synergies.
We ought to be much more concerned about the firm that makes the
investment and ultimately carries on: the acquiring firm.
Long-window event studyA second strand of the literature examines the long-run post-merger
stock performance of the acquirer and its absorbed target, or the “com-
bined entity.” In general, this strand of the literature converges on the
notion that acquiring firms underperform their benchmark returns in the
post-merger period – this is often referred to in the literature as the “long-
run merger underperformance anomaly.” Some researchers have asked
whether “overpayments” could be responsible for the long-run nega-
tive drift in share price after acquisitions.17 In one study [Antoniou et al.
(2008a)], however, the authors were unable to establish such a relation-
ship statistically. Although we concur that it makes more sense to focus
on bidder firm results and bidder shareholder returns to ascertain the
desirability of M&A on an ex-post basis, we argue that long-run event
Malatesta (1983) argues that the widely used percentage abnormal returns do not capture 13
the real wealth changes of acquiring firm shareholders. However, dollar returns capture the
wealth change of acquiring firm shareholders.
Moeller et al. (2005) define aggregate dollar returns as the sum of the acquisition dollar 14
returns (change in market capitalization from day –2 to day +1) divided by the sum of the
market capitalization of acquiring firms two days before the acquisition announcements, in
2001 dollars.
The sharp contrast between the CAR and the shareholder wealth effect signals the 15
possibility of a Type II error (if the CAR is used) in respect of the null hypothesis that M&A
are value-creating corporate events.
“Evidence suggests that the majority of acquisitions do not benefit shareholders in the 16
long term. Valuations and premiums tend to be excessively high and targets impossible to
achieve.” Financial Times 2004.
Schwert (2003) states: “…One interpretation of this evidence (post merger 17
underperformance) is that bidders overpay and that it takes the market some time to
gradually learn about this mistake.”
164
studies also fail to provide interesting insights due to the following short-
comings.
First and foremost is the methodological problem associated with long-
run event studies. For instance, bad model problems imply that it is sim-
ply not possible to accurately forecast or measure expected returns, thus
rendering futile the analysis of long-run abnormal returns, particularly as
the event window lengthens. In addition to the bad model problem, a
number of researchers have pointed out that the process used in calcu-
lating and testing long-run abnormal returns is in itself biased. For ex-
ample, Barber and Lyon (1997) and Kothari and Warner (1997) address
misspecification problems in long-horizon event studies. They argue that
commonly employed methods for testing for long-run abnormal returns
yield misspecified test statistics, and invite researchers to use extreme
caution in interpreting long-horizon test results.
In order to mitigate bad model problems and biased test statistics, Bar-
ber and Lyon (1997) and Lyon et al. (1999) advocate the use of a single
control firm or a carefully-constructed reference portfolio approach. The
idea is to select a matching firm or a portfolio of matching firms that have
approximately the same size (MV18) and book-to-market ratio (BE/ME19)
as the sample firms. This approach has been shown to eliminate some
well-known biases and results in better test statistics.20 Although some
recent M&A studies have applied this updated approach (listed below), it
remains problematic for the following reason.
Until the adoption of the Statement of Financial Accounting Standard
(SFAS) 141 in 2001 in the U.S. (i.e., shortly after the crash of the important
M&A wave of the late 1990s), mergers were accounted for using either
the purchase or the pooling of interests (pooling) methods.21 At least con-
ceptually, the use of the pooling method was meant to be reserved for
transactions involving the “merger of equals.” Due to a favorable impact
on reported earnings (rather than economic earnings), however, acquirers
began manipulating M&A transactions in order to “qualify” for the use of
the pooling method, partly based on the belief that the higher level of re-
ported earnings would result in higher stock market valuations,22 or in cer-
tain cases, to obfuscate the quarter-by-quarter comparability of results.23
As such, the choice of permissible merger accounting methods would
eventually become the subject of widespread controversy and regulatory
scrutiny, culminating in the eventual elimination of the pooling method
under both U.S. Generally Accepted Accounting Principles (GAAP) and
eventually International Financial Reporting Standards (IFRS).
But the very fact that acquirers previously enjoyed discretion over their
choice of merger accounting method compromises the ability of financial
economists to select relevant and unbiased control firms based on BE/
ME as advocated in Lyon et al. (1999). For instance, the two account-
ing techniques are fundamentally different in that under the purchase
method, the offer price for the target is compared to the fair market value
(FMV) of the net assets of the target, with the difference being capitalized
on the balance sheet in the form of Goodwill,24 which was previously
required to be amortized over a period not exceeding 40 years (20 years
under IAS GAAP), resulting in a lower level of reported earnings under
this method.25 Under the pooling method, however, operating results and
prior-period accounts (i.e., assets and liabilities) were simply added to-
gether at their stated book values (irrespective of whether different ac-
counting methods were historically used by the bidder and its target26),
with the concepts of FMV and Goodwill playing no part in the process.
Under the purchase method, the acquirer’s post-merger equity book
value27 will remain equal to its original book equity, unless new equity
is issued to finance the acquisition.28 Under the pooling method, how-
ever, the book value of equity will be equal to the sum of the historical
book values of the acquirer and its target. In summary, if an acquisi-
tion is financed with equity, the purchase method should usually result
in a higher book equity value because the offer price for the target will
typically exceed the FMV of net assets of the target. But even under the
same method (i.e., purchase), the resulting book equity account for the
combined firm will vary depending on whether the acquisition is financed
with cash or with stock.29
Matters are further complicated by the timing issue for the application
of each method. The pooling method consolidates target and acquiring
Market value of equity.18
Book value of equity to market value of equity.19
Barber and Lyon (1997) identify three biases: the new listing bias, the rebalancing bias, and 20
the skewness bias.
Note that the pooling method was also previously allowed under International Accounting 21
Standards (IAS) 22, although IAS 22 was superseded by IFRS 3 in 2004, which no longer
allows the pooling method.
The choice of accounting method should not impact cash flow. See, for example, Hong et 22
al. (1978) and Davis (1990).
The pooling method requires the restatement of historical financial accounts (in some cases 23
well before the closing of the acquisition), as if the bidder and its target had been one firm
all along. This makes deciphering the performance of each individual business very difficult.
In the late 1990s, Tyco International Ltd., for instance, came under fire after pursuing a
series of acquisitions which were accounted for under the pooling method, making it nearly
impossible to compare one quarter to the next. Businessweek 2001.
Under the purchase method, a positive differential between the offer price and the FMV 24
of the target’s net assets would first be allocated to identifiable intangibles (i.e., licenses,
in-process research and development, or patents), with the balance allocated to Goodwill.
As a compromise, when the pooling method was eventually abolished under U.S. GAAP, 25
the Financial Accounting Standards Board (FASB) issued statement no. 142, which
specified that Goodwill would no longer be allowed to be amortized but rather would
become the subject of an annual impairment test.
For instance, U.S. GAAP allows both last-in-first-out (LIFO) and first-in-first-out (FIFO) 26
inventory accounting methods.
Equity book value is defined as the sum of shareholder’s equity and retained earnings.27
Under the U.S. GAAP variant, in-process research and development is required to be 28
written off as part of the transaction, thereby reducing the retained earnings of the acquirer
and impacting book value of equity.
Book equity values should generally be larger for stock-financed transactions (i.e., where 29
an equity issuance has taken place) relative to acquisitions that are paid for in cash (i.e.,
financed either by a debt issuance or cash on balance sheet).
165
firm accounts from the beginning of the year (historical bidder and target
results are also added together), regardless of when the merger is actu-
ally completed. However, under the purchase method, target and acquir-
ing firm results are only added together from the date of the transaction
completion onwards.
The picture that emerges is that one cannot bundle into the same sample
deals that were accounted for under different merger accounting meth-
ods while continuing to expect a justified benchmark return and therefore
an unbiased test result. Unless the aforementioned factors are controlled
for in the research design, the selection of control firms and therefore
the calculation of expected returns becomes tricky and probably flawed.
This, in turn, compromises the calculation of long-run abnormal returns,
as well as the cross-sectional comparability of results. Although merger
accounting intricacies go beyond the scope of this article, it is easily
understood that such issues must be carefully analyzed when applying
popular long-run event study methodology. The literature has largely
failed to control for these key issues.
The above-mentioned merger accounting discussion significantly weak-
ens the use of the so-called “state-of-the-art” bootstrap approach30 ad-
vocated by Ikenberry et al. (1995), Kothari and Warner (1997), and Lyon
et al. (1999) and applied by Rau and Vermaelen (1998) in mergers and
acquisitions. Indeed, the 1000 pseudo-portfolios matched in size and
book-to-market ratio at the time of merger completion do not control for
the aforementioned accounting issues, thereby calling into question the
validity of obtained matches and thus the empirical distribution of abnor-
mal returns generated under the approach.
But even if we control for merger accounting differences and all other
possible sources of misspecification,31 we are still far from obtain-
ing an accurate and reliable long-run test result. In one attempt, Lyon
et al. (1999) recommend two general approaches that control for com-
mon misspecification problems in long-run event studies. Despite the
authors’ positive intentions, however, their simulated results confirm that
well-specified test statistics (i.e., where empirical rejection levels are con-
sistent with theoretical rejection levels) are only guaranteed in random
samples, while misspecified test statistics are pervasive in non-random
samples. We also know that mergers and acquisitions cluster in time and
by industry, which necessarily implies that well-specified test statistics in
long-run M&A event studies should hardly exist [Antoniou et al. (2008b)].
The central message in their study, however, is that: “the analysis of long-
run abnormal returns is treacherous.”
Second, on a more general framework, Viswanathan and Wei (2004) pro-
vide a mathematical proof that the usual abnormal return (CAR/BHAR32)
calculated in event studies has a negative expectation. They prove that,
in any finite sample, the expected event abnormal return will invariably be
negative and becomes more negative as the event window is lengthened.
The implication of utmost importance here is that these negative results
do not discriminate between successful or unsuccessful transactions,
suggesting that the so-called long-run M&A underperformance anomaly
may not be anomalous at all.
In addition, Viswanathan and Wei go on to examine the above problem
in infinite samples. They prove that, asymptotically, the event abnormal
return converges to zero and hence they conclude that the negative
long-run event abnormal return is simply a small sample problem. This
again offers a reasonable explanation as to why some of the larger M&A
studies have reported insignificant results. If the small sample problem is
the long-run event study’s only flaw, then one can probably get around
this issue by increasing the sample size. But what would such studies
contribute to our understanding of M&A? By averaging abnormal returns
across a very large number of cross-sectional observations, we end up
with what might resemble a near-normal distribution of abnormal returns
with a mean of zero. Consequently, nothing can be concluded from this
result apart from the fact M&A have a 50/50 probability of creating or
destroying value, and can therefore be likened to a crapshoot. Surely, this
type of conclusion is of no use to regulators, practitioners, or investors.
Finally, the recent development of a series of new methodologies has giv-
en rise to a “new wave” of long-run event studies.33 Mitchell and Stafford
(2000), for instance, reexamine the long-run anomaly associated with
corporate takeovers.34 Their results suggest that acquirers (combined
firms) earn the expected rate of return over the long run, thereby implying
that mergers do not create nor destroy value, an idea which is consistent
with that found in the previous paragraph.
But even if we were able to overcome all the methodological problems
associated with long-window event studies, what level of insight could
be gained from such a “perfect” study? We know that stock price is for-
ward looking and that, in a relatively efficient market, the price of an asset
should reflect expectations regarding the underlying asset’s future cash
flows, based on information which is available today. Consequently, the
returns observed under long-window event studies (particularly in the lat-
ter years of the sample), ought to be discounting anticipated events that
The Capco Institute Journal of Financial TransformationMeasuring the Economic Gains of Mergers and Acquisitions: Is it Time for a Change?
As noted in Ikenberry et al. (1995), the bootstrap approach avoids many problematic 30
assumptions associated with conventional t-tests over long time horizons, namely
normality, stationarity, and the time independence of sample observations.
Lyon et al. (1999) document that the misspecification of test statistics can generally be 31
traced to five sources: the new listing bias, the rebalancing bias, the skewness bias, cross-
sectional dependence, and bad models problems.
Buy and hold abnormal returns (BHARs).32
For these new methodologies, refer to Barber and Lyon (1997), Lyon et al. (1999), Brav 33
(2000), and Mitchell and Stafford (2000).
For these long-run event studies, see, for example, Mitchell and Stafford (2000), Brav et al. 34
(2000), Eckbo et al. (2000), and Boehme and Sorescu (2002).
166
reach far beyond the merger or acquisition under analysis. For example,
stock returns generated five years after an M&A transaction (t+5 years)
should discount what is expected to happen in periods t+6, t+7, etc. But
this extends so far beyond the actual transaction that occurred in year
t that we fail to see the relevance of this analysis. We refer to this as the
long-window forward expectations trap. Adding to event study’s general
malaise, confounding events (whether exogenous or endogenous) may
further distort the inferences from long-run stock returns. All in all, long-
term event studies fail to provide a means of identifying and isolating the
effects of the actual merger that has previously taken place and thus do
not provide much relevant insight about the micro or macroeconomic
impact of M&A.
In light of the arguments presented thus far, we believe that event study
methodology, for both short and long event windows, falls short of offer-
ing an economically sound tool for measuring merger performance on an
ex-post basis. Ironically, it is the assumption regarding market efficiency
that is the downfall of both long-term and short-term event studies, but in
different ways: in the case of the former, a reasonable degree of market ef-
ficiency should imply that long-run stock returns are probably irrelevant to
the analysis of individual events at a fixed point in time, while in the case of
the latter, markets are insufficiently efficient to reliably predict the outcome
of a particular M&A transaction. In our view, financial economists have
overindulged in event studies, which have largely yielded results which
are biased, unreliable, and lacking insight. Although we can appreciate the
merits of tackling the M&A value-creation debate from an investor experi-
ence standpoint,35 we believe the debate is best served when approached
from a different perspective: that of company fundamentals.
Fundamental analysis“Unlocking synergies” is the most commonly cited managerial motivation
for undertaking M&A [Walter and Barney (1990)]. If synergies truly exist,
economic gains from mergers should thus show up in the combined firm’s
fundamentals. Coming back to Shelton’s definition of value creation, it is
clear that the concept has less to do with share price movements (at
least in terms of aetiology), and more to do with asset reorganizations,
and improvements in a number of financial performance metrics and key
performance indicators relevant to the firm/sector under analysis. Jarrell
et al. (1988) postulate that gains to shareholders must be real economic
gains via the efficient rearrangement of resources. Consistent with basic
finance principles, improvements in company fundamentals should drive
capital gains. As such, we believe that the analysis of the desirability of
M&A begins not with stock returns (the symptom), but rather with the un-
derlying factors/fundamentals which drive cash flows (the cause), which
in turn power shareholder returns.
In addition to short- and long-window event studies, there is a small body
of literature that examines pre- and post-merger operational performance
of acquiring firms. In short, a value-creating transaction is one that is as-
sociated with some measurable improvement (a relative improvement at
a minimum) in company fundamentals. Furthermore, if purported syner-
gies are real, and cash flows do improve, we should be able to identify
the sources of any such real economic gains. Managers undertaking M&A
for synergistic reasons rather than for hubris-related motives must have
identified possible sources of economic gains before proceeding with the
transaction. These types of studies are more conducive to performance
evaluation on an ex-post basis in our view and they are also more likely
to contain information that can be used and applied by business school
students and M&A practitioners generally.
In a 1992 journal article, Healy et al. find that the 50 largest mergers
between U.S. public industrial firms completed between 1979-1984 ex-
perienced higher operating cash flow returns, mainly due to post-merger
improvements in asset productivity. They also report that such improve-
ments do not come at the expense of cutbacks in capital expenditure
or research and development (R&D) spending, thereby undermining the
claim that the improvement in post-merger cash flows is achieved at the
expense of the acquiring firms’ long-run viability/competitiveness. These
results are similar to Kaplan (1989), although the latter author finds that
merged firms reduce capital expenditure in the three years following the
merger. In short, both studies indicate that merging was probably a de-
sirable course of action, as evidenced by the fact that acquiring firms
appeared to enjoy better economic strength relative to peers during the
post-acquisition period. This point illustrates the importance of using
carefully-selected, industry-adjusted benchmarks in interpreting M&A
study results. If the aforementioned results are pervasive on a time se-
ries and cross-sectional basis, then we feel more comfortable with the
blanket statement that M&A are generally desirable corporate events.
However, a large literature gap remains to be filled in this particular area
of research.
But there is no Holy Grail solution to this debate. Healy et al.’s analysis
and many of its kind also suffer from their own methodological problems
stemming from complexities in analyzing and interpreting financial state-
ments and accounting data generally. As highlighted earlier, the existence
of different permissible merger accounting methods renders time series
and cross-sectional comparisons difficult. For instance, in their raw form,
some pre- and post-merger accounting results and ratios, particularly
those involving both P&L and balance sheet accounts,36 are simply not
comparable under the purchase method. This is because the purchase
method consolidates bidder and target firm accounts from the acquisition
Unlike CARs, Buy and hold abnormal returns (BHARs) provide a reasonable measure of 35
“investor experience” [Barber and Lyon (1997)].
Goodwill, for example, should be excluded for better comparability of results across time 36
periods and across sample firms.
167
date onwards, and prior accounts are not restated. Conversely, particular
in the case of serial acquirers adopting the pooling method, it is nearly
impossible to isolate the like-for-like performance of component firms
due to the restatement of historical accounts by the acquirer after each
new acquisition. Without making the relevant adjustments to the data, a
research design which involves performing ratio analysis to compare pre-
and post-merger results must therefore be applied with care.
Another problem arises in the computation of meaningful industry aver-
ages due to the fact that firms operate across multiple industries and ge-
ographies, which obscures the process of identifying relevant pure-play
benchmarks for calculating industry-adjusted results and ratios. The se-
lected benchmark should ideally be limited to relevant comparable firms
that have chosen not to merge. However, we know and recognize that
mergers come in waves and cluster by industry, which poses a research
design challenge given the currently statistically-focused foundation of
academic research into M&A.37 If our call for new research methods is
heard and answered, we would not be surprised if this meant the end of
the large-sample statistically-oriented studies in favor of smaller, almost
“case-study-like” research that focuses on select peer groups, with ana-
lytical emphasis placed on “softer” elements of post-merger integration
including managerial compensation, the creation of M&A oversight com-
mittees for combined firms, through to the robustness of 180-day post-
acquisition plans and “heat maps” of where synergies may lie between a
bidder and its target. Although these studies are likely to suffer from their
own methodological weaknesses, they can only represent an improve-
ment relative to the current crop of event studies which have dismally
failed to produce any meaningful and usable insights on the issue of M&A
value creation.
Conclusion: is it time for a change?Since its birth in 1969, the event study has held a virtual monopoly over
academia’s various attempts to shed light on the M&A value creation
debate. We argue that the event study has stagnated in terms of its
incremental insights into M&A. Among other things, the event study’s
shortcomings include the hefty premiums offered to target shareholders,
which exacerbate the short-run weighted-average CARs for combined
entities, and this in turn has misled many researchers into concluding that
most M&A transactions represent value-creating events. Furthermore,
we show that cumulative abnormal returns observed around merger an-
nouncements can produce poor estimates of shareholder wealth effects.
We also discuss various problems inherent to long-window event studies
and we conclude that such studies are also unsuitable for measuring the
economic gains of M&A due to various methodological problems and the
forward-looking nature of stock returns. In short, both short- and long-
window event studies provide biased results and undependable insights
regarding the question of M&A value creation.
In the wake of multiple merger waves that appear to be growing in
strength and size every time the tide comes in, financial economists
and finance students can no longer afford to use research methods that
do little more than reinforce the obsolete work of their academic peers,
whilst continuing to completely sidestep the key questions being asked
by investors, regulators, and M&A practitioners. Advancing the M&A val-
ue creation debate has never been more critical. Since the first version
of this article in 2004, we have argued that new methods are very much
needed to advance our knowledge on this vital issue and we continue to
believe that fundamentals-based approaches represent more promising
avenues for new and future research. Despite the flexibility offered by
GAAP, accounting data (when analyzed with the appropriate level of skill
and competence) remains the best proxy of company economic perfor-
mance available to investors, analysts, and academics alike. Encourag-
ingly, we expect for the divide between IFRS and U.S. GAAP to continue
to narrow over time, in line with the globalization of the investing practice
and of the investor base. As such, we believe that the M&A value creation
puzzle can be better understood by returning to fundamentals. If the very
businesses that are studied in the academic research merely ensure their
own survival by continuously reinventing themselves to successfully
meet the evolving needs of consumers, while continuously facing up to
the ever-lurking threats of competition, complacency, and obsolescence,
then why should research into M&A not be subject to the same Darwin-
ian forces?
References:Andrade, G., M. Mitchell, and E. Stafford, 2001, “New evidence and perspectives on mergers,” •
Journal of Economic Perspectives, 15, 103-120
Antoniou, A., P. Arbour, and H. Zhao, 2008a, “How much is too much: are merger premiums too •
high?” European Financial Management 14, 268-287
Antoniou, A., P. Arbour, and H. Zhao, 2008b, “The effects of cross-correlation of stock returns •
on post merger stock performance,” ICFAI Journal of Mergers and Acquisitions, 5, 7-29
Asquith, P., 1983, “Merger bids, uncertainty, and stockholder returns,” Journal of Financial •
Economics, 11, 51-83
Asquith, P., R. F. Bruner, and D. W. Mullins, Jr., 1983, “The gains to bidding firms from merger,” •
Journal of Financial Economics, 11, 121-139
Barber, B. M. and J. D. Lyon, 1997, “Detecting long-run abnormal stock returns: the empirical •
power and specification of test statistics,” Journal of Financial Economics, 43, 341-372
Boehme, R. D. and S. M. Sorescu, 2002, “The long-run performance following dividend •
initiations and resumptions: underreaction or product of chance?” Journal of Finance, 57,
871-900
Bradley, M., A. Desai, and E. H. Kim, 1988, “Synergistic gains from corporate acquisitions and •
their division between the stockholders of target and acquiring firms,” Journal of Financial
Economics, 21, 3-40
Brav, A., 2000, “Inference in long-horizon event studies: a Bayesian approach with application •
to initial public offerings,” Journal of Finance, 55, 1979-2016
Brav, A., C. Geczy, and P. Gompers, 2000, “Is the abnormal return following equity issuances •
anomalies?” Journal of Financial Economics, 56, 209-249
Davis, M. L., 1990, “Differential market reaction to pooling and purchase methods,” The •
Accounting Review 65, 696-709
Dennis, D. K. and J. J. McConnell, 1986, “Corporate mergers and security returns,” Journal of •
Financial Economics, 16, 143-187
The Capco Institute Journal of Financial TransformationMeasuring the Economic Gains of Mergers and Acquisitions: Is it Time for a Change?
The small-sample bias is often regarded as a shortcoming for studies in financial 37
economics.
168
Dodd, P., 1980, “Merger proposals, management discretion and stockholder wealth,” Journal of •
Financial Economics, 8, 105-137
Dodd, P. and R. Ruback, 1977, “Tender offers and stockholders returns: An empirical analysis,” •
Journal of Financial Economics, 5, 351-373
Draper, P. and K. Paudyal, 1999, “Corporate takeovers: Mode of payment, returns and trading •
activity,” Journal of Business Finance and Accounting, 26, 521-558
Eckbo, B. E., R. Masulis, and O. Norli, 2000, “Seasoned public offerings: resolution of the new •
issue puzzle,” Journal of Financial Economics, 56, 251-291
Fama, E. F., L. Fisher, M. C. Jensen, and R. Roll, 1969, “The adjustment of stock prices to new •
information, International Economic Review, 10, 1-21
Firth, M., 1980, “Takeovers, shareholder returns, and the theory of the firm,” Quarterly Journal •
of Economics, 94:2, 235-260
Franks, J. R. and R. S. Harris, 1989, “Shareholder wealth effects of corporate takeovers: the •
U.K. experience 1955-1985,” Journal of Financial Economics, 23, 225-249.
Healy, M. H., K. G. Palepu, and R. S. Ruback, 1992, “Does corporate performance improve •
after mergers?” Journal of Financial Economics, 31, 135-175
Hong, H., R. S. Kaplan, and G. Mandelker, 1978, “Pooling vs. purchase: the effects of •
accounting for mergers on stock prices,” The Accounting Review, 53, 31-47
Ikenberry, D., J. Lakonishok, and T. Vermaelen, 1995, “Market underreaction to open market •
share repurchases,” Journal of Financial Economics, 39, 181-208
Jarrell, G. A., J. Brickley, and J. Netter, 1988, “The market for corporate control: The empirical •
evidence since 1980,” Journal of Economic Perspectives, 2, 49-68
Kaplan, S., 1989, “The effects of management buyouts on the operating performance and •
value,” Journal of Financial Economics, 24, 217-254
Kaplan, S. and M. S. Weisbach, 1992, “The success of acquisitions: Evidence from •
divestitures,” Journal of Finance, 47, 107-138
Kothari, S. P. and J. B. Warner, 1997, “Measuring long-horizon security price performance,” •
Journal of Financial Economics, 43, 301-339
Lajoux, A. R. and J. F. Weston, 1998, “Do deals deliver on postmerger performance?” Mergers •
and Acquisitions, 33:2, 34-37
Loughran, T. and A. M. Vijh, 1997, “Do long-term shareholders benefit from corporate •
acquisitions?” Journal of Finance, 52, 1765-1790
Lyon, J. D., B. M. Barber, and C. Tsai, 1999, “Improved methods for tests of long-run abnormal •
stock returns,” Journal of Finance, 54, 165-201
Malatesta, P. H., 1983, “The wealth effect of merger activity and the objective functions of •
merging firms,” Journal of Financial Economics, 11, 155-181
Mitchell, M. and E. Stafford, 2000, “Managerial decisions and long-term stock price •
performance,” Journal of Business, 73, 287-329
Mitchell, M., T. Pulvino, and E. Stafford, 2004, “Price pressure around mergers,” Journal of •
Finance 58, 31-63
Moeller, S. B., F. P. Schlingemann, and R. M. Stulz, 2004, “Firm size and the gains from •
acquisitions,” Journal of Financial Economics, 73, 201–228
Moeller, S. B., F. P. Schlingemann, and R. M. Stulz, 2005, “Wealth destruction on a massive •
scale? A study of acquiring-firm returns in the recent merger wave,” Journal of Finance, 60,
757–782
Mulherin, J. H. and A. L. Boone, 2000, “Comparing acquisitions and divestitures,” Journal of •
Corporate Finance, 6, 117-139
Rau, P. R. and T. Vermaelen, 1998, “Glamour, value and the post-acquisition performance of •
acquiring firms,” Journal of Financial Economics, 49, 223-253
Roll, R., 1986, “The hubris hypothesis of corporate takeovers,” Journal of Business, 59, 197-•
216
Schwert, G. W., 2003, “Anomalies and market efficiency,” in Constantinides, G., M. Harris, and •
R. Stulz (eds.), Handbook of the economics of finance, North-Holland
Shelton, L. M., 1988, “Strategic business fit and corporate acquisitions,” Strategic Management •
Journal, 9, 279-287
Shojai, S., 2009, “Economists’ hubris - the case of mergers and acquisitions,” Journal of •
Financial Transformation, 26, 4-12
Sudarsanam, S., P. Holl, and A. Salami, 1996, “Shareholder wealth gains in mergers: effect of •
synergy and ownership structure,” Journal of Business Finance and Accounting, 23, 673-698
Viswanathan, S. and B. Wei, 2004, “Endogenous events and long run returns,” Working Paper, •
Fuqua School of Business, Duke University
Walter, G. A. and J. B. Barney, 1990, “Management objectives in mergers and acquisitions,” •
Strategic Management Journal, 11, 79-86
169
PART 2
Mobile Payments Go Viral: M-PESA in Kenya
AbstractM-PESA is a small-value electronic payment and store of
value system that is accessible from ordinary mobile phones.
It has seen exceptional growth since its introduction by mo-
bile phone operator Safaricom in Kenya in March 2007: it has
already been adopted by 14 million customers (correspond-
ing to 68 percent of Kenya’s adult population) and processes
more transactions domestically than Western Union does
globally. M-PESA’s market success can be interpreted as
the interplay of three sets of factors: (i) preexisting country
conditions that made Kenya a conducive environment for a
successful mobile money deployment; (ii) a clever service
design that facilitated rapid adoption and early capturing of
network effects; and (iii) a business execution strategy that
helped M-PESA rapidly reach a critical mass of customers,
thereby avoiding the adverse chicken-and-egg (two-sided
market) problems that afflict new payment systems.
Ignacio Mas — Senior Advisor, Financial Services for the Poor (FSP) team, Bill and Melinda Gates Foundation
Dan Radcliffe — Program Officer, Financial Services for the Poor (FSP) team, Bill and Melinda Gates Foundation1
From “Yes Africa can: success stories from a dynamic continent,” World 1
Bank, August 2010.
170
M-PESA in a nutshell2
M-PESA was developed by mobile phone operator Vodafone and
launched commercially by its Kenyan affiliate Safaricom in March 2007.
M-PESA (“M” for mobile and “PESA” for money in Swahili) is an elec-
tronic payment and store of value system that is accessible through mo-
bile phones. To access the service, customers must first register at an
authorized M-PESA retail outlet. They are then assigned an individual
electronic money account that is linked to their phone number and ac-
cessible through a SIM card-resident application on the mobile phone.3
Customers can deposit and withdraw cash to/from their accounts by
exchanging cash for electronic value at a network of retail stores (often
referred to as agents). These stores are paid a fee by Safaricom each
time they exchange these two forms of liquidity on behalf of customers.
Once customers have money in their accounts, they can use their phones
to transfer funds to other M-PESA users and even to non-registered us-
ers, pay bills, and purchase mobile airtime credit. All transactions are
authorized and recorded in real time using secure SMS, and are capped
at around U.S.$800.
Customer registration and deposits are free. Customers then pay a flat
fee of around U.S. 35¢4 for person-to-person (P2P) transfers and bill pay-
ments, U.S. 30¢ for withdrawals (for transactions less than US $30), and
U.S. 1.1¢ for balance inquiries. Individual customer accounts are main-
tained in a server that is owned and managed by Vodafone, but Safari-
com deposits the full value of its customers’ balances on the system in
pooled accounts in two regulated banks. Thus, Safaricom issues and
manages the M-PESA accounts, but the value in the accounts is fully
backed by highly liquid deposits at commercial banks. Customers are
not paid interest on the balance in their M-PESA accounts. Instead, the
foregone interest is paid into a not-for-profit trust fund controlled by Sa-
faricom (the purpose of these funds has not yet been decided).
M-PESA is useful as a retail payment platform because it has extensive
reach into large segments of the population. Figure 1 shows the size of
various retail channels in Kenya.6 Note that there are over five times the
number of M-PESA outlets than the total number of PostBank branches,
post offices, bank branches, and automated teller machines (ATMs) in the
country. Using existing retail stores as M-PESA cash-in/cash-out outlets
reduces deployment costs and provides greater convenience and lower
cost of access to users.
A snapshot of M-PESA after four yearsM-PESA is going from strength to strength. Safaricom reached the
10 million customer mark in just over three years and is now serving
14 million customers, of which the majority are active. This corresponds
to 81 percent of Safaricom’s customer base and 68 percent of Kenya’s
adult population.7
Other key developments and figures reported by Safaricom as of April
30, 2011 are:8
28,000 retail stores at which M-PESA users can cash-in and cash-out, ■■
of which nearly half are located outside urban centers.
U.S.$415 million per month in person-to-person (P2P) transfers. On ■■
an annualized basis, this is equal to roughly 17 percent of Kenyan
gross domestic product (GDP).9 Although transactions per customer
have been on a rising trend, they remain quite low, probably still under
two P2P transactions per month.
The average transaction size on P2P transfers is around U.S.$33, but ■■
Vodafone has stated that half the transactions are for a value of less
than U.S.$10.
U.S.$94 million in annual revenue in FY2010. This is equal to 9 percent ■■
of Safaricom revenues.
For more detailed accounts of the M-PESA service, see Hughes and Lonie (2009) for a 2
historical account, Mas and Morawczynski (2009) for a fuller description of the service, and
Mas and Ng’weno (2009) for the latest accomplishments of M-PESA.
The Subscriber Identification Module (SIM) card is a smart card found inside mobile phones 3
that are based on the GSM family of protocols. The SIM card contains encryption keys,
secures the user’s PIN on entry, and drives the phone’s menu. The Short Messaging
Service (SMS) is a data messaging channel available on GSM phones.
We assume an exchange rate of U.S.$1 = 85 Kenyan Schillings.4
Data from this table was pulled from the Central Bank of Kenya, Kenya Post Office Savings 5
Bank, and Safaricom websites.
Kenya has a total population of nearly 40 million, with 78 percent living in rural areas and a 6
GDP per capita of U.S.$1,600. 19 percent of adults have access to a formal bank account.
See FSDT (2009a) for financial access data derived from the FinAccess survey, a nationally
representative survey of 6,600 households conducted in early 2009.
Population figures are from the United Nations (2010) http://data.un.org/CountryProfile.7
aspx?crName=Kenya.
M-PESA performance statistics are as of December 31, 2010 (http://www.safaricom.co.ke/8
index.php?id=1073). Additional figures are taken from Safaricom’s FY2010 results for the
period ending May 31, 2010 and Central Bank of Kenya reports.
GDP figure is from the World Development Indicators database, World Bank (July 2010).9
440800 840
1,510
16,900
100,000
1
10
100
1,000
10,000
100,000
PostBank branches
Total post of�ces
Bank branches
ATMs M ‐Pesa stores
Airtime resellers
Figure 1 – Outlets offering financial services in Kenya5
171
The Capco Institute Journal of Financial TransformationMobile Payments Go Viral: M-PESA in Kenya
There are at least 27 companies using M-PESA for bulk distribu-■■
tion of payments. Safaricom itself used it to distribute dividends on
Safaricom stock to 180,000 individual shareholders who opted to
receive their dividends into their M-PESA accounts, out of a total of
700,000 shareholders.
Since the launch of the bill pay function in March 2009, there are at ■■
least 75 companies using M-PESA to collect payments from their
customers. The biggest user is the electric utility company, which now
has roughly 20 percent of its one million customers paying through
M-PESA.
At least two banks (Family Bank and Kenya Commercial Bank) are ■■
using M-PESA as a mechanism for customers to either repay loans or
withdraw funds from their banks accounts.
In May 2010, Equity Bank and M-PESA announced a joint venture, M-
KESHO, which permits M-PESA users to move money between their M-
PESA mobile wallet and an interest-bearing Equity Bank account. While
several hundred thousand customers have opened M-KESHO accounts,
only a fraction of these are actively being used and it is unclear how ag-
gressively Equity Bank and Safaricom are promoting this jointly-branded
product.
Customer perspectives on M-PESAWilliam Jack of Georgetown University and Tavneet Suri of MIT recently
released results from a panel survey that queried 2016 Kenyan house-
holds in August 2008 and resurveyed them in December 2009 [Jack and
Suri (2010)]. The results show that M-PESA is steadily propagating down
market, reaching a majority of Kenya’s poor, unbanked, and rural popula-
tions:
Adoption of M-PESA has continued to march ahead, going from 44 ■■
percent of Kenyan households in 2008 to 70 percent in 2009.
M-PESA has propagated down market: the share of poor households ■■
that are registered M-PESA users has gone from 28 percent in 2008 to
51 percent in 2009. (Here, the poor are defined as the poorest 50 per-
cent of Kenyan households who earn, on average, about U.S.$3.40
per capita per day.) Similarly, the percent of rural households using
M-PESA has gone from 29 percent to 59 percent, and the percent of
unbanked households using M-PESA has gone from 25 percent to
50 percent.
Customers’ perceptions of M-PESA are steadily improving: the per-■■
centage of users who trust their agent was 95 percent in Round 2,
compared to 65 percent in Round 1, even while the number of agents
quadrupled during the period from 4,000 to 16,000. Customers
reporting delays in withdrawing money fell from 22 percent to 16
percent, and the share of delays attributed to agents running out of
liquidity fell from 70 percent to 30 percent. When asked about the
hypothetical impact of M-PESA closing down, 92 percent of custom-
ers said that it would have a large and negative effect on their lives,
up from 85 percent.
M-PESA users are increasingly using it to save: the percentage of ■■
users who say they use M-PESA to save has gone from 76 percent
to 81 percent, and the percentage who say they save for emergen-
cies has gone from 12 percent to 22 percent. Most of the uptick in
saving behavior is due to early adopters saving more over time. This
indicates that as users get familiar with the product, they are more
likely to use it as a savings tool.
M-PESA helps users deal with negative shocks: the researchers find ■■
that households who have access to M-PESA and are close to an
agent are better able to maintain their level of consumption expen-
ditures, and in particular food consumption, in the face of negative
income shocks, such as job loss, livestock death, bad harvest, busi-
ness failure or poor health. On the other hand, households without
access to M-PESA are less able to absorb such adverse shocks. The
researchers have been careful to rule out explanations based on mere
correlation and are currently investigating the precise mechanisms
that underlie this ability to spread risk.
M-PESA’s service evolutionM-PESA’s original core offering was the P2P payment – enabling custom-
ers to send money to anyone with access to a mobile phone. It opened up
a market for transactions which previously were handled largely informal-
ly – through personal trips, friends, and public transport networks. That
is represented by the set of transactions labeled “personal networks” in
the middle of Figure 2. Many P2P transactions can be characterized as
scheduled payments (such as sending a portion of salary earned at the
end of the month to relatives back home), but many represent a basic
form of finance, where people can draw on a much broader network of
family members, friends, and business associates to access money as
and when required. Thus, M-PESA not only introduces a large measure
of convenience to transactions that were already occurring, but it also
enables a basic form of financial protection for a large number of users
by creating a network for instant, “on demand” payments.
In recent months, Safaricom has increasingly opened up M-PESA to in-
stitutional payments, enabling companies to pay salaries and collect bill
payments. In the future, Safaricom envisions increased use of M-PESA
for in-store purchases. Thus, Safaricom intends for M-PESA to become
a more pervasive retail payments platform, a strategy represented by the
downward arrow in Figure 2. The challenge remains for M-PESA to be-
come a vehicle for delivery of a broader range of financial services to
the bulk of the Kenyan population, represented by the upward arrow in
Figure 2. While some users are using M-PESA to save, there is likely
to be a need to develop more targeted savings products that balance
customers’ preference for liquidity and commitment, and which connect
into a broader range of financial institutions. This is the journey M-PESA
172
must be on for it to deliver on its promise of addressing the challenge of
financial inclusion in Kenya. Safaricom will need to develop appropriate
service, commercial, and technical models for M-PESA to interwork with
the systems of other financial service providers. We return to this topic in
the concluding section of this paper.
The broader significance of M-PESABefore examining why M-PESA achieved such dramatic growth, we dis-
cuss briefly three top-line lessons that have emerged from M-PESA’s
success. Firstly, M-PESA has demonstrated the promise of leveraging
mobile technology to extend financial services to large segments of un-
banked poor people. This is fundamentally because the mobile phone is
quickly becoming a ubiquitously deployed technology, even among poor
segments of the population. Mobile penetration in Africa has increased
from 3 percent in 2002 to 51 percent today, and is expected to reach
72 percent by 2014.10 And, happily, the mobile device mimics some of
the key ingredients needed to offer banking services. The SIM card in-
side GSM phones can be used to authenticate users, thereby avoiding
the costly exercise of distributing separate bank cards to low-profitabil-
ity poor customers. The mobile phone can also be used as a point of
sale (POS) terminal to initiate financial transactions and securely com-
municate with the appropriate server to request transaction authoriza-
tion, thus obviating the need to deploy costly dedicated devices in retail
environments. Secondly, M-PESA has demonstrated the importance of
designing usage- rather than float-based revenue models for reaching
poor customers with financial services. Because banks make most of
their money by collecting and reinvesting deposits, they tend to distin-
guish between profitable and unprofitable customers based on the likely
size of their account balances and their ability to absorb credit. Banks
thus find it difficult to serve poor customers because the revenue from
reinvesting small-value deposits is unlikely to offset the cost of serving
these customers. In contrast, mobile operators in developing countries
have developed a usage-based revenue model, selling prepaid airtime to
poor customers in small increments, such that each transaction is profit-
able on a stand-alone basis. This is the magic behind the rapid penetra-
tion of prepaid airtime into low-income markets: a card bought is profit
booked, regardless of who bought the prepaid card. This usage-based
revenue model is directly aligned with the model needed to sustain-
ably offer small-value cash-in/cash-out transactions at retail outlets and
would make possible a true mass-market approach, with no incentive
for providers to deny service based on minimum balances or intensity
of use. Thirdly, M-PESA has demonstrated the importance of building
a low-cost transactional platform which enables customers to meet a
broad range of their payment needs. Once a customer is connected to
an e-payment system, s/he can use this capability to store money in a
savings account, send and receive money from friends and family, pay
bills and monthly insurance premiums, receive pension or social welfare
payments, or receive loan disbursements and repay them electronically.
In short, when a customer is connected to an e-payment system, his or
her range of financial possibilities expands dramatically.
Putting these elements together, M-PESA has prompted a rethink on the
optimal sequencing of financial inclusion strategies. Where most finan-
cial inclusion models have employed “credit-led” or “savings-led” ap-
proaches, the M-PESA experience suggests that there may be a third
approach – focus first on building the payment “rails” on which a broader
set of financial services can ride.
Accounting for M-PESA’s success: three perspectivesThe rest of this paper explores M-PESA’s success from three angles.
First, we examine the environmental factors in Kenya that set the scene
for a successful mobile money development. Then, we examine the ser-
vice design features that facilitated the rapid adoption and frequent use
of M-PESA. And, finally, we examine the elements in Safaricom’s execu-
tion strategy that helped M-PESA rapidly reach a critical mass of custom-
ers. In so doing, we draw extensively on a sequence of four papers which
readers can refer to for more detailed accounts of the M-PESA story:
Heyer and Mas (2009) on the country factors that led to M-PESA’s suc-
cess, Mas and Morawczynski (2009) on M-PESA’s service features, Mas
and Ng’weno (2010) on Safaricom’s execution, and Mas (2009) on the
economics underpinning branchless banking systems.
Beyond the compelling marketing, cold business logic and consistent
execution of M-PESA, its success is a vivid example of how great things
Wireless Intelligence (www.wirelessintelligence.com).10
Pushing and pulling money across time
“Just payments”
M-PESA’s role in promoting fuller financial inclusion
M-PESA as a fuller retail payments platform
Formal financial products
Savings, credit, insurance
Informal service providers
Pawnbroker, money lender
“On-demand” payments Personal networks
“Scheduled” payments
Remote B2C/C2B institutional payments Salaries, bill pay, G2P,
online/e-commerce
In-store merchant payments
For goods and services
Figure 2 – Potential range of transactions supported by M-PESA
173
The Capco Institute Journal of Financial TransformationMobile Payments Go Viral: M-PESA in Kenya
happen when a group of leaders from different organizations rally around
common challenges and ideas. The story of M-PESA straddles the social
and the commercial, the public and the private, powerful organizations
and determined individuals:
Kenya country factors: unmet needs, favorable market conditionsThe growth of M-PESA is a testament to Safaricom’s vision and execu-
tion capacity. However, Safaricom also benefited from launching the
service in a country which contained several enabling conditions for a
successful mobile money deployment, including: strong latent demand
for domestic remittances, poor quality of available financial services, a
banking regulator which permitted Safaricom to experiment with different
business models and distribution channels, and a mobile communica-
tions market characterized by Safaricom’s dominant market position and
low commissions on airtime sales.
Strong latent demand for domestic remittances Safaricom based the initial launch of the M-PESA service on the “send
money home” proposition, even though it also allows the user to buy
and send airtime, store value, and, more recently, to pay bills. Demand
for domestic remittance services will be larger where migration results in
splitting of families, with the breadwinner heading to urban centers and
the rest of the family staying back home. This is the case in Kenya, where
17 percent of households depend on remittances as their primary income
source [FSD-Kenya (2007a)].
In her study of M-PESA, Ratan (2008) suggests that the latent demand
for domestic remittances is related to urbanization ratios. More propitious
markets will be those where the process of rural-urban migration is suffi-
ciently rooted to produce large migration flows, but not so advanced that
rural communities are hollowed out. Countries with mid-range urbaniza-
tion ratios (20 percent to 40 percent), especially those that are urbanizing
at a rapid rate, are likely to exhibit strong rural-urban ties requiring trans-
fer of value between them. This is the case in many African countries like
Kenya and Tanzania, where the urbanization ratios are 22 percent and
25 percent, respectively.11 In the Philippines and Latin America, where
urbanization ratios exceed 50 percent, remittances are more likely to be
triggered by international rather than domestic migration patterns.
Where entire nuclear families move, remittances will be stronger where
there is cultural pressure to retain connection with one’s ancestral village.
In Kenya, migrants’ ties with rural homes are reinforced by an ethnic (rather
than national) conception of citizenship. These links are expressed through
burial, inheritance, cross-generational dependencies, social insurance, and
other ties, even in cases where migrants reside more or less permanently
in cities.12 In other settings, a greater emphasis on national as opposed to
local or ethnic identity may have diminished the significance of the rural
“home” and hence dampened domestic remittance flows.
Poor quality of existing alternativesLatent demand for e-payments must be looked at in the context of the
accessibility and quality of the alternatives. If there are many good alter-
natives to mobile payments (as is typically the case in developed coun-
tries), it will be difficult to convince users to switch to the new service. In
the Philippines, for example, the G-Cash and Smart Money mobile pay-
ment services experienced low take-up in part due to the availability of a
competitive alternative to mobile payments – an extensive and efficient
semi-formal retail network of pawnshops which offered domestic remit-
tance services at 3 percent.
U.N. Population Division: World Urbanization Prospects (2007). 11
For fuller analyses of the use of mobile money for domestic remittances in Kenya, see 12
Ratan (2008) and Morawczynski (2008).
The individuals and institutions behind M-PESA
The idea of M-PESA was originally conceived by a London-based
team within Vodafone, led by Nick Hughes and Susie Lonie. This
team believed that the mobile phone could play a central role in
lowering the cost of poor people accessing financial services.
The idea was seized by the Safaricom team in Kenya, led by CEO
Michael Joseph and Product Manager Pauline Vaughn. They toyed
with the idea, convinced themselves of its power, developed it
thoroughly prior to the national launch, and oversaw a very focused
execution.
The Central Bank of Kenya (CBK), and in particular its Payments
System group led by Gerald Nyoma, deserves much credit for being
open to the idea of letting a mobile operator take the lead in providing
payment services to the bulk of the population. The CBK had
recently been made aware of the very low levels of bank penetration
in the country by the first FinAccess survey in 2006, and they were
determined to explore all reasonable options for correcting the access
imbalance. The CBK worked in close partnership with Vodafone and
Safaricom to assess the opportunities and risks involved prior to
the launch and as the system developed. They were conscious that
premature regulation might stifle innovation, so they chose to monitor
closely and learn, and formalize the regulations later.
Finally, the U.K.’s Department for International Development (DfID)
played an instrumental role, first by funding the organizations that
made the FinAccess survey possible, the Financial Sector Deepening
Trust in Kenya, the FinMark Trust in South Africa, and then by
providing seed funding to Vodafone to trial its earliest experiments
with M-PESA. DfID’s role in spotlighting the need for mobile
payments and funding the early risk demonstrates good roles for
donor funding.
174
In Kenya, the most common channel for sending money before M-PESA
was informal bus and matatu (shared taxi) companies. These compa-
nies are not licensed to transfer money, resulting in considerable risk that
the money will not reach its final destination. And Kenya Post, Kenya’s
major formal remittance provider, is perceived by customers as costly,
slow, and prone to liquidity shortages at rural outlets. Meanwhile, Ke-
nya’s sparse bank branch infrastructure (840 branches) is far too limited
to compete with M-PESA’s 28,000 cash-in/cash-out outlets. Figure 3
below illustrates how Kenyan households sent money before and after
M-PESA [FSD-Kenya (2007a and 2009a)]. Note the dramatic reduction
in the use of informal bus systems and Kenya Post to transfer money
between 2006 and 2009.
As noted above, M-PESA’s early adopters were primarily banked custom-
ers, which suggests that M-PESA did not acquire its initial critical mass
through competition with the formal sector but rather as a complement
to formal services for clients who were wealthier, more exposed to formal
financial service options, and less risk-averse. As services move deeper
into the market, unbanked users will likely drive M-PESA’s expansion,
due to the competitive advantages of formal mobile offers over other op-
tions. This is one reason why Africa, with its high population of unbanked,
is seen as such a promising market for mobile money deployments.
A supportive banking regulatorRegulation of mobile money can help to secure trust in new mobile mon-
ey schemes. At the same time, regulation may constrain the success of
a mobile money deployment by limiting the scheme operator’s degrees
of freedom in structuring the business model, service proposition, and
distribution channels. In the case of M-PESA, Safaricom had a good
working relationship with the Central Bank of Kenya (CBK) and was given
regulatory space to design M-PESA in a manner that fit its market. The
CBK and Safaricom worked out a model that provided sufficient pru-
dential comfort to the CBK. The CBK insisted that all customer funds be
deposited in a regulated financial institution, and reviewed the security
features of the technology platform. In turn, the CBK allowed Safaricom
to operate M-PESA as a payments system, outside the provisions of the
banking law.13
Safaricom has had to pay a certain price for this arrangement. For in-
stance, interest earned on deposited balances must go to a not-for-profit
trust and cannot be appropriated by Safaricom or passed on to custom-
ers. There are also limits on transaction sizes (subsequently relaxed) to
address anti-money laundering concerns. But, fundamentally, Safaricom
was able to design the service as it saw fit, without having to contort its
business model to fit within a prescribed regulatory model.
The CBK has continued to support M-PESA’s development, even in the
face of pressure from banks. In late 2008, after a lobbying attack from the
banking industry seeking to shut down the service, the Central Bank did
an audit of the M-PESA service at the request of the Ministry of Finance
and declared it safe and in line with the country’s objectives for financial
inclusion.14 So far, the Central Bank appears justified in its confidence
in M-PESA as there have been no major reports of fraud. And system
downtime, although frequent, has not been catastrophic.
A dominant mobile operator and low airtime commissionsThe chances of a mobile money scheme taking root depend also on the
strength of the mobile operator within its market. Market share is an im-
portant asset because it is associated with a larger customer base for
cross-selling the mobile money service, a larger network of airtime resell-
ers which can be converted into cash-in/cash-out agents, stronger brand
recognition and trust among potential customers, and larger budgets to
finance the heavy up-front market investment needed to scale a deploy-
ment. With a market share of around 80 percent, Safaricom enjoyed each
of these benefits when it launched M-PESA.
A mobile money deployment will also have greater chance of success in
countries where the commissions mobile operators pay airtime resellers
are relatively low. This is because if commissions are too high, resellers
will not be attracted by the lower commissions of the incipient cash-in/
cash-out business. In Safaricom’s case, airtime commissions total 6 per-
cent, of which 5 percent are passed on to the retail store. A 1 to 2 per-
cent commission on a cash-in/out transaction is plausibly attractive – the
store need only believe that the cash business may be five times as big
The Central Bank of Kenya Act was amended in 2003 to give CBK broad oversight 13
mandate over payment systems, but the operational modalities for its regulatory powers
over payments systems have not been implemented, pending approval of a new National
Payments System Bill which has languished in Parliament.
The results of the survey are explained in Okoth (2009).14
0
10
20
30
40
50
60
70
M-PESA Hand Bus Post office
Direct deposit
Money transfer service
Other
2006
2009
Figure 3 – Money transfer behavior before and after M-PESA
175
The Capco Institute Journal of Financial TransformationMobile Payments Go Viral: M-PESA in Kenya
as the airtime business in volume terms. This seems reasonable, con-
sidering that the bulk of airtime sales are of low denominations (around
U.S. 25¢).
A reasonable base of banking infrastructureFinally, the ability of M-PESA stores to convert cash to e-value for cus-
tomers depends on how easily they can rebalance their liquidity portfo-
lios. This will be more difficult to achieve if bank branch penetration is
too low, as this will force the agent channel to develop alternative cash
transport mechanisms. Thus, an agent network will need to rely on a
minimal banking retail infrastructure. (This qualifies our earlier point that
lack of access to formal services indicates a strong market opportunity.
There appears to be a branch penetration “sweet spot” for mobile mon-
ey, where penetration is not so high that it hampers demand for mobile
money services, but not so low that agents are unable to manage their
liquidity.) Kenya is reasonably well supplied with rural liquidity points due
to the branch networks of Equity Bank and other banks and MFIs. Even
so, shortage of cash or electronic value for M-PESA agents is a prob-
lem both in country and city. Other countries face more serious liquidity
constraints, especially in rural areas, which is likely to be a major factor
affecting the success of mobile services in specific country contexts.
M-PESA’s service design: getting people onto the systemWhile M-PESA’s explosive growth was fueled by certain country-specific
enabling conditions, the success of such an innovative service hinged
on the design of the service. Conducting financial transactions through
a mobile phone is not an intuitive idea for many people, and walking to a
corner shop to conduct deposits and withdrawals may not at first seem
natural to many. To overcome this adoption barrier, Safaricom had to de-
sign M-PESA in a way that (i) helped people grasp immediately how they
might benefit from the service, (ii) removed all barriers that might prevent
people from experimenting with the service; and (iii) fostered trust in the
retail outlets who would be tasked with promoting the service, registering
customers, and facilitating cash-in/cash-out services.
A simple message targeting a big pain point M-PESA was originally conceived as a way for customers to repay mi-
croloans. However, as Safaricom market-tested the mobile money prop-
osition, the core proposition was shifted from loan repayment to helping
people make P2P transfers to their friends and family. From its commer-
cial launch, M-PESA has been marketed to the public with just three pow-
erful words: “send money home.” This message was well adapted to the
Kenyan phenomenon of split families discussed above and tapped into a
major pain point for many Kenyans – the risks and high cost associated
with sending money over long distances. This basic “e-remittance” prod-
uct became the must-have “killer” application that continues to drive ser-
vice take-up and remains the main (though not only) marketing message
four years later. Although people have proved creative in using M-PESA
for their own special needs, sending money home continues to be one of
the most important uses: the number of households receiving money in
Kenya has increased from 17 percent to 52 percent since M-PESA was
introduced.15
A simple user interfaceThe simplicity of M-PESA’s message has been matched by the simplicity
of its user interface. The M-PESA user interface is driven by an applica-
tion that runs from the user’s mobile phone. The service can be launched
right from the phone’s main menu, making it easy for users to find. The
menu loads quickly because it resides on the phone and does not need
to be downloaded from the network each time it is called. The menu
prompts the user to provide the necessary information, one prompt at a
time. For instance, for a P2P transfer, the user will be asked to enter the
destination phone number, the amount of the transfer, and the personal
identification number (PIN) of the sender. Once all the information is gath-
ered, it is fed back to the customer for final confirmation. Once the cus-
tomer hits “OK,” it is sent to the M-PESA server in a single text message.
Consolidating all information into a single message reduces messaging
costs, as well as the risk of the transaction request being interrupted
half-way through. A final advantage is that the application can use the
security keys in the user’s SIM card to encrypt messages end-to-end,
from the user’s handset to Safaricom’s M-PESA server.
Removing adoption barriers: free to register, free to deposit, no minimum balancesSafaricom designed the scheme to make it as easy as possible for cus-
tomers to try the new service. There is a quick and simple process for
customer registration, which can be done at any M-PESA retail outlet.
Customers pay nothing to register and the clerk at the outlet does most
of the work during the process. First, the clerk provides a paper registra-
tion form, where the customer enters his or her name, ID number (from
Kenyan National ID, Passport, Military ID, Diplomatic ID, or Alien ID), date
of birth, occupation, and mobile phone number. The clerk then checks
the ID and inputs the customer’s registration information into a special
application in his mobile phone. If the customer’s SIM card is an old one
that is not preloaded with the M-PESA application, the clerk replaces it.
The customer’s phone number is not changed even if the SIM card is.
Safaricom then sends both the customer and outlet an SMS confirming
the transaction. The SMS gives customers a four-digit “start key” (one-
time password), which they use to activate their account. Customers en-
ter the start key and ID number, and they are then asked to input a secret
PIN of their choice, which completes the registration process. In addition
FinAccess Survey, FSDT (2009a), p 16.15
176
to leading customers through this process, retail outlets explain how to
use the application and the tariffs associated with each service. Such
agent support early in the process is particularly important in rural areas,
where a significant percentage of the potential user base is illiterate or
unfamiliar with the functioning of their mobile phone.
The minimum deposit amount was originally set at around U.S.$1.20 but
has since been halved, and there is no minimum balance requirement.
Customers can deposit money for free, so there is no immediate bar-
rier to taking up the service. M-PESA charges customers only for “doing
something” with their money, such as making a transfer, withdrawal, or
prepaid airtime purchase.
Being able to send money to anyoneM-PESA customers can send money to non M-PESA customers, includ-
ing any person with a GSM mobile phone in Kenya, whether they are
subscribers of Safaricom or of any of the other three competing net-
works (Airtel, Orange, and Yu). Under this service, money is debited from
the sender’s account, and the recipient gets a code by SMS which she
can use to claim the monetary value at any M-PESA store. Thus, it is an
account-to-cash service, with the receiver’s experience being similar to
how Western Union works today. The pricing on this service is interesting:
customers pay a higher (roughly triple) P2P charge when sending money
to a non-customer, but at the other end cashing out is free for a non-
customer, whereas registered customers pay a cash-out fee of at least
U.S.$0.30. Why “penalize” the customer rather than the non-customer?
Safaricom understood that the sender had power over the recipient, so
it chose to put pressure on the sender to require the recipient to register
with M-PESA. Furthermore, non-customers got a great first experience
with M-PESA when they received money for free, which Safaricom hoped
would convince them to register for M-PESA.
Building trust in the retail networkSafaricom recognized that M-PESA would not achieve rapid adoption
unless customers had enough trust in the M-PESA retail network that
they were willing to conduct cash-in/cash-out transactions through those
outlets. Safaricom employed several measures to build that trust. Firstly,
it closely linked the M-PESA brand to customers’ affinity with and trust
in Safaricom’s strong corporate brand. As the mobile operator in Kenya
with a dominant share (over 80 percent at M-PESA’s launch and scarce-
ly less today), Safaricom was already a broadly respected and trusted
brand, even among low-income customers. (M-PESA retail outlets are
required to paint their store “Safaricom green,” which not only builds cus-
tomers confidence that the store is acting on behalf of Safaricom, but
also makes it easier for customers to locate cash-in/cash-out points.)
Secondly, Safaricom ensured that customers can walk into any autho-
rized retail outlet and have a remarkably similar experience. This has
helped to build trust in the platform and the outlets, and gives customers
a consistently positive view of the service. Safaricom maintains this con-
trol over the customer experience by investing heavily in store training
and on-site supervision. Safaricom chose to centralize these functions in
a single third-party vendor (Top Image) rather than relying on its channel
intermediaries (i.e., master agents) to cascade these functions to retail
shops. A Top Image representative visits each outlet at minimum on a
monthly basis and rates each store on a variety of criteria, including vis-
ibility of branding and the tariff poster, availability of cash and M-PESA
electronic value to accommodate customer transactions, and the quality
of recordkeeping.
Thirdly, customers receive instant SMS confirmation of their transac-
tion, helping customers learn by experience to trust the system. The
confirming SMS constitutes an electronic receipt, which can be used in
dispute resolution. The receipt confirming a money transfer details the
name and number of the recipient and the amount transferred. This al-
lows the sender to confirm instantly that the money was sent to the right
person – the most common source of error. Finally, Safaricom requires
its outlets to record all cash-in/cash-out transactions in a paper-based,
Safaricom-branded logbook. For each transaction, the store clerk en-
ters the M-PESA balance, the date, agent ID, transaction ID, transaction
type (customer deposit or withdrawal), value, customer phone number,
customer name, and the customer’s national ID number. Customers are
then asked to sign the log for each transaction, which helps discour-
age fraud and also gives agents a way to offer first-line customer care
for customers querying previous transactions. Each page in the log is in
triplicate. The top copy is kept by the retail outlet for his own records, a
second is passed on to the store’s master agent, and the third is sent to
Safaricom. Recall that all information contained in the agent log (except
for the customer signature) is captured electronically by Safaricom when
the transaction is made and is available to the master agents via their
web management system. Hence, the main purpose of the agent log is
not for recordkeeping, but to provide comfort to customers who are used
to having transactions recorded on paper.
Simple and transparent pricingM-PESA pricing is made transparent and predictable for users. There
are no customer charges for the SMSs that deliver the service, and in-
stead fees are applied to the actual customer-initiated transactions. All
customer fees are subtracted from the customer’s account, and outlets
cannot charge any direct fees. Thus, outlets collect their commissions
from Safaricom (through their master agents) rather than from customers.
This reduces the potential for agent abuses. Customer fees are uniform
nationwide, and they are prominently posted in all outlet locations.
M-PESA chose to specify its fees in fixed currency terms rather than as
a percentage of the transaction. This makes it easier for customers to
understand the precise cost of each transaction and helps them think of
177
the fee in terms of the transaction’s absolute value (i.e., sending money
to grandmother). It also helps them compare the transaction cost against
alternative and usually costlier money-transfer arrangements (i.e., the
bus or matatu fare plus travel time).
Deposits are free to customers. Withdrawals under U.S.$30 cost around
0.30¢. Withdrawal charges are “banded” (i.e., larger transactions incur
a larger cost) so as not to discourage smaller transactions. ATM with-
drawals using M-PESA are slightly more expensive than at a retail out-
let (35¢ versus 30¢). P2P transfers cost a flat rate of around U.S.35¢.
This is where Safaricom makes the bulk of its revenue. Thus, for a purely
electronic transfer, customers pay more than double what they pay for
the average cash transaction (15¢), despite the cost to provide being
lower for purely electronic transactions than those involving cash. This
reflects a notion of optimal pricing that is less based on cost and more
on customer willingness to pay: enabling remote payments is the biggest
customer pain point which M-PESA aims to address. M-PESA is cheaper
than the other available mechanisms for making remote payments, such
as money transfer by the bus companies, Kenya Post’s Postapay, or
Western Union.16
It is noteworthy that M-PESA largely maintained the same pricing for
transactions in its first three years, despite the significant inflation expe-
rienced during the period. This helped establish customer familiarity with
the service. Early on, Safaricom changed the pricing for two customer
requests that do not involve a financial transaction: balance inquiries (be-
cause the initial low price generated an overly burdensome volume of
requests) and PIN changes (because customers were far more likely to
remember their PIN if the fee to change it was higher). The volume of both
types of requests decreased substantially after these price changes. As
noted earlier, the SMS confirmation of a transaction contains the avail-
able balance, which also helps cut down on the number of balance inqui-
ries. More recently, M-PESA introduced new pricing tiers for very small
(U.S.$ 0.60-1.20) and large (U.S.$ 400-800) transactions.
Liquidity of last resort at bank branches and ATMsFrom very early on, M-PESA signed up banks as agents, so that any
M-PESA customer could walk into the branches of several banks to con-
duct cash-in/cash-out transactions. One year after its launch, M-PESA
went further and partnered with PesaPoint, one of the largest ATM ser-
vice providers in Kenya. The PesaPoint network includes over 110 ATMs
scattered all over the country, giving them a presence in all eight prov-
inces. Customers can now retrieve money from any PesaPoint ATM. To
do so, they must select “ATM withdrawal” from their M-PESA menu. They
then receive a one-time ATM authorization code, which they enter on the
ATM keyboard to make the withdrawal. No bank card is needed for this
transaction. By accessing the PesaPoint ATM network, M-PESA custom-
ers can now make withdrawals at any time, day or night.
Yet M-PESA’s liquidity system is not without its challenges. Due to cash
float constraints, M-PESA retail outlets cannot always meet requests for
withdrawals, especially large withdrawals. Furthermore, the agent com-
mission structure discourages outlets from handling large transactions.
As a result, customers are sometimes forced to split their transactions
over a few days, taking money out in bits rather than withdrawing a lump
sum, adding both cost and inconvenience. It also undermines customer
trust in M-PESA as a mechanism for high-balance, long-term saving. Us-
ing bank branches and ATMs to give customers a sort of liquidity mecha-
nism of last resort bolstered the credibility of the M-PESA system.
Execution: getting to critical mass, quicklyWith a strong service design in place, Safaricom then set about develop-
ing its execution plan. It recognized that it would be difficult to scale M-
PESA incrementally as it had to overcome three significant hurdles that
are common to any new electronic payment system, namely:
Adverse network effects –■■ the value to the customer of a pay-
ment system depends on the number of people connected to and
actively using it. The more people on the network, the more useful it
becomes.17 While network effects can help a scheme gain momentum
once it reaches a critical mass of customers, they can make it difficult
to attract early adopters in the early phase when there are few users
on it.
Chicken-and-egg trap –■■ in order to grow, M-PESA had to attract
both customers and stores in tandem. It is hard to sell the proposition
to customers while there are few stores to serve them, and equally
hard to convince stores to sign up while there are few customers to
be had. Thus, the scheme needed to drive both customer and store
acquisition aggressively.
Trust –■■ customers have to gain confidence in the reliability of a new
system. In this case, customers had to be comfortable with three ele-
ments that were new at the time in Kenya: (i) a payment system that
was operated by a mobile operator, (ii) going to non-bank retail outlets
to meet their cash-in/cash-out needs, and (iii) accessing their account
and initiating transactions through their mobile phone.
These problems reinforce each other in the early-stage development of a
payments system, creating a significant hurdle to growth. We suspect this
The Capco Institute Journal of Financial TransformationMobile Payments Go Viral: M-PESA in Kenya
In her field research, Olga Morawczynski finds that sending KSh 1,000 through M-PESA is 16
27 percent cheaper than the post office’s PostaPay, and 68 percent cheaper than sending
it via a bus company. See Morawczynski and Pickens (2009).
It has become habitual to illustrate network effects with reference to fax machines: the 17
first set of people who bought a fax machine did not find it very useful as they could
not send faxes to many people. As more people bought fax machines, everyone’s faxes
became more and more useful. Network effects are sometimes referred to as demand-
side economies of scale, to emphasize that scale affects the value of the service to
each customer. This distinguishes it from supply-side economies of scale, which refer to
situations where average costs per customer fall as volume increases. Davidson (2009)
discusses implications of network effects for mobile money.
178
hurdle helps explain why many other mobile money deployments remain
sub-scale. M-PESA overcame this hurdle through very forceful execution
on two key fronts: (i) Safaricom made significant up-front investments in
building a strong service brand for M-PESA; and (ii) Safaricom effectively
leveraged its extensive network of airtime resellers to build a reliable,
consistent retail network that served customers’ liquidity needs.
Aggressive up-front investment in promoting the M-PESA brandFrom the beginning, Safaricom sought to foster customer trust in the new
payment mechanism and relied on existing customers to be the prime
mechanism to draw in new customers. This was all the more difficult be-
cause Safaricom was introducing not only a new product, but an entirely
new product category to a market that had little experience with formal
financial services. The internal launch target for M-PESA was 1 million
customers within one year, equal to 17 percent of Safaricom’s customer
base of about 6 million customers at that time.18
National launch at scale – after small pilots involving less than 500 cus-
tomers,19 M-PESA launched nationwide, increasing the likelihood that the
service could reach a critical mass of customers in a short time frame.
At launch, Safaricom had 750 stores and had made sure to cover all of
Kenya’s 69 district headquarters. It was a massive logistical challenge
that led to a great deal of customer and store confusion and, in the first
months after launch, several days’ delays to reach customer service
hotlines. User and store errors were frequent since everyone was new to
the service. But the gamble paid off. Logistical problems subsided after
a few months, leaving strong brand recognition and top-of-mind aware-
ness among large segments of the population. The service outran first-
year growth targets, quickly turning network effects in their favor as new
customers begat more customers and turned M-PESA into a compelling
business proposition for more stores.
An appropriate marketing mix – initial marketing featured and target-
ed the wealthier city dweller with the need to “send money home.” This
choice of the richer urban dweller as the initial customer created an as-
pirational image for M-PESA and avoided the impression that it was a
low-value product aimed at the poor. Over time, the marketing moved
from young, upmarket urban dwellers with desk jobs to more ordinary
Kenyans from lower-paid professions.
While M-PESA’s launch was associated with significant up-front invest-
ment in above-the-line marketing via TV and radio,20 there was also intense
outreach through road shows and tents that traveled around the country
signing people up, explaining the product, and demonstrating how to use
it. Over time, as people became more familiar with the product and how
to use it, it was no longer necessary to do this kind of hands-on outreach.
TV and radio were largely replaced by the omnipresent M-PESA branding
at all outlets, supported with a few large billboards. Newer ads feature a
general emotional appeal, with a wider range of services indicated.
A scalable distribution channelSafaricom understood that the primary role of the mobile phone is to en-
able the creation of a retail outlet-based channel for cash-to-digital value
conversion. And, for this cash-to-digital conversion to be broadly avail-
able to the bulk of the population, it had to develop a channel structure
that could support thousands of M-PESA stores spread across a broad
geography. To achieve this, Safaricom built four elements into its channel
management execution strategy: (i) engaging intermediaries to help man-
age the individual stores, thereby reducing the number of direct contacts
it had to deal with; (ii) ensuring that outlets were sufficiently incentiv-
ized to actively promote the service; (iii) maintaining tight control over the
customer experience; and (iv) developing several different methods for
stores to re-balance their stocks of cash and e-value.
Two-tier channel management structure – Safaricom created a two-
tier structure with individual stores (sub-agents, in Safaricom’s parlance)
who depended on master agents (referred to by Safaricom as Agent Head
Offices [HO]). Agent HOs maintain all contact with Safaricom, and perform
two key functions: (i) liquidity management (buying and selling M-PESA
balance from Safaricom and making it available to individual stores un-
der their responsibility), and (ii) distributing agent commissions (collecting
the commission from Safaricom based on the overall performance of the
stores under them and remunerating each store). Individual stores may be
directly owned by an agent HO or may be working for one under contract.
Incentivizing stores – retail outlets will not maintain sufficient stocks of
cash and e-money unless they are adequately compensated for doing
so. Hence, Safaricom pays commissions to agent HOs for each cash-
in/cash-out transaction conducted by stores under their responsibility.
Safaricom did not prescribe the commission split between agent HOs
and stores, though most agent HOs pass on 70 percent of commissions
to the store.21 For deposits under U.S. $30, Safaricom pays U.S. 11.8¢
in total commissions (pre-tax), of which U.S. 6.5¢ goes to the store after
tax. For withdrawals, Safaricom pays U.S. 17.6¢ to the channel, of which
U.S. 9.8¢ goes to the store. So, assuming equal volumes of deposits
and withdrawals, the store earns U.S. 8.2¢ per transaction. Assuming
the store conducts 60 transactions per day, it earns around U.S. $5.70 –
almost twice the prevailing daily wage for a clerk in Kenya.
Safaricom company results for the year ending March 2007. 18
The earliest pilot project conducted in 2004/05 revolved around microloan repayments, 19
and involved the Commercial Bank of Africa, Vodafone, Faulu Kenya, and MicroSave, in
addition to Safaricom.
A survey of 1210 users in late 2008 revealed that 70 percent of survey respondents claimed 20
that they had first heard about M-PESA from advertisements, TV or radio. FSDT (2009b), p. 6.
Safaricom wants the split to be 20 percent/80 percent, thus passing more of the 21
commission down to the retail outlet.
179
Recall that Safaricom charges customers U.S. 30¢ (US 29.4¢ to be exact)
on a round-trip savings transaction (free deposit plus U.S. 29.4¢ with-
drawal), which is, in fact, equal to what the channel gets (U.S. 11.8¢ on
the deposit + U.S. 17.6¢ on the withdrawal). So, assuming equal vol-
umes of deposits and withdrawals, Safaricom does not make any money
on cash transactions. It merely “advances” commissions to the chan-
nel when customers deposit, and recoups it when customers withdraw.
By charging U.S. 35¢ on electronic P2P transactions (which are almost
costless to provide), Safaricom opted to generate the bulk of its revenue
from the service for which there is highest customer willingness to pay –
remote P2P payments.
Because store revenues are dependent on the number of transactions
they facilitate, Safaricom was careful not to flood the market with too
many outlets, lest it depress the number of customers per agent. Instead,
it maintained a balanced growth in the number of outlets relative to the
number of active customers, resulting in an incentivized and committed
agent base.
Maintaining tight control over the customer experience – Safaricom
also recognized that customers need to have a good experience at the
retail points, where the bulk of transactions take place. To ensure that
it maintained control over the customer experience, Safaricom did not
rely on the broad base of agent HOs to perform all channel management
functions. Instead (as mentioned above), it concentrated the evaluation,
training, and on-site supervision of stores in a single outsourcing part-
ner, Top Image. Thus, we see that Safaricom delegated the more routine,
desk-bound, non-customer-facing store support activities (i.e., liquidity
management, distribution store commissions) to a larger pool of agent
HOs. At the same time, through its contract with Top Image, it retained
direct, centralized control over the key elements of the customer experi-
ence (i.e., store selection, training, supervision).
Developing multiple store liquidity management methods – by far the
biggest challenge faced by M-PESA stores is maintaining enough liquidity
in terms of both cash and e-float to be able to meet customer requests for
cash-in and cash-out. If they take too many cash deposits, stores will find
themselves running out of e-float with which to facilitate further deposits. If
they do too many withdrawals, they will accumulate e-float but will run out
of cash. Hence, they frequently have to rebalance their holdings of cash
versus e-float. This is what we refer to as liquidity management.
The M-PESA channel management structure was conceived to offer
stores three methods for managing liquidity. Two of these place the agent
HO in a central role, with the expectation that the agent HO will “recycle”
e-float between locations experiencing net cash withdrawals (i.e., accu-
mulating e-float) and locations with net cash deposits (i.e., accumulating
cash). We discuss each of these methods in turn:
Agent HO provides direct cash support to stores –■■ under this
option, the store clerk comes to the agent HO’s head office to deliver
or offload cash, or the agent HO sends cash runners to the store to
perform these functions (not very common).
Agent HO and stores use their respective bank accounts –■■ under
this option, if the store has excess cash and wants to buy M-PESA
e-float from the agent HO, the store will deposit the cash into the
account of the agent HO at the nearest bank branch or ATM. Once
the agent HO confirms receipt of the funds into its account, the HO
transfers M-PESA e-float to the store’s M-PESA account. If the store
wants to sell e-float to get cash, the store transfers M-PESA e-float to
the agent HO. The agent HO then deposits (or transfers) money into
the store’s account at the branch of the store’s bank. The store can
then withdraw the cash at the nearest branch or ATM.
Stores interact directly with a bank that has registered as an ■■
M-PESA “superagent” – under this option, the agent HO does not
get involved in liquidity management. Instead, stores open an account
with a participating “superagent” bank. To rebalance their cash,
stores deposit and withdraw cash against their bank account at the
nearest branch or ATM of the bank. The store then electronically buys
and sells e-float in real time against their bank account. From a store’s
perspective, one drawback of the bank-based superagent mechanism
is that it can only use it during banking business hours. This presents
a problem for stores in the evenings and on weekends.
The e-float-cash nexus will remain the key constraint to the further de-
velopment of M-PESA since it requires the physical movement of cash
around the country and is thus the least scalable part of the system.
M-PESA’s future evolutionThe experience of M-PESA demonstrates how powerful a payment net-
work that offers convenience at an affordable cost can be once a critical
mass of customers is reached. It also shows that achieving critical mass
requires both a service design that removes as many adoption barriers
as possible and significant investment in marketing, branding, and agent
network management. The Kenyan experience also suggests that several
country-level environmental factors need to align to set the scene for a
successful mobile money development, including the labor market profile
(demand for remittances generated by rural-urban migration), the quality
of available financial services, support from the banking regulator, and
the structure of the mobile communications market (dominant mobile op-
erator and low airtime commissions).
Yet, while M-PESA has been successful beyond what anyone could have
imagined at its launch, the model still has substantial room to develop
further. Our wish list for M-PESA is three-fold: (i) the mainstreaming of
M-PESA’s regulatory treatment; (ii) pricing that opens up a much larger
market of micro-transactions; and (iii) building of a much more robust
The Capco Institute Journal of Financial TransformationMobile Payments Go Viral: M-PESA in Kenya
180
ecosystem around M-PESA that enables customers to access a broader
range of financial services. We address each of these below, before offer-
ing some concluding thoughts on how M-PESA offers a rekindled vision
for achieving financial inclusion in developing countries.
Mainstreaming M-PESA’s regulatory treatmentM-PESA’s regulatory treatment as a payments vehicle needs to be for-
malized so that it can become regulated in the most appropriate way. To
this end, the CBK has been trying to get a new payments law enacted
by Parliament, but the draft has not yet been approved. The intention is
for M-PESA to be covered in future by regulations emanating from this
payments law. The CBK issued new agent banking regulations in early
2010 that allowed commercial banks to use retail outlets as a delivery
channel for financial services. This gave banks the possibility of replicat-
ing the M-PESA service themselves. However, the requirements on both
banks and their agents are more onerous than what applies to Safaricom
(a non-bank) and its agents. In early 2011, the CBK issued e-money and
payment service provider guidelines which incorporate less restrictive
agent regulations. By applying for agents under these guidelines rather
than as banking agents, banks can now deploy agents on terms similar
to Safaricom’s.
Pricing that enables smaller paymentsM-PESA’s current pricing model is not conducive to small transactions. A
U.S.$10 P2P transfer plus withdrawal, for example, costs around 6.5 per-
cent of the transaction size (U.S.0.35¢ for the transfer plus U.S.0.30¢ for
the withdrawal). We see two advantages to adjusting M-PESA’s current
pricing model to make it work for smaller-denomination transactions:
It would make the service accessible to a poorer segment of the ■■
population, for whom pricing is now too high given their transactional
needs. This would allow Safaricom to maintain customer growth once
saturation starts to set in at current pricing.
It would allow customers to use M-PESA for their daily transaction ■■
needs, and in particular to save on a daily basis when they are paid
daily.
A reduction in customer prices could come about in several ways:
For electronic transactions, the current P2P charge of U.S. 35¢ allows ■■
for substantial scope for price reductions. But let us be careful. There
is a compelling logic behind the current model of extracting value
from remote payments (for which there is substantial customer will-
ingness to pay), while maintaining tight pricing on cash transactions
(for which customers are less willing to pay). But we do believe there
is room for “tranching” the P2P fee so that the price works for smaller
(i.e., daily) transactions.
For cash transactions, one way to enable lower fees would be to ■■
create a category of street-level sub-agents, characterized by lower
costs and commissions than store-based agents. Sub-agents would
be a kind of “e-susu collector,” operating with small working capital
balances in order to aggregate small customer transactions. Sub-
agents would use normal M-PESA retail outlets to rebalance their
cash and M-PESA stored value. The key principle here is that seg-
mentation of customers needs to go hand-in-hand with segmentation
of agents.
Linking with banks and other institutional partners to offer a fuller range of financial servicesWhile some customers use M-PESA as a savings device, it still falls short
of being a useful savings proposition for most poor people. According
to the January 2009 CBK audit of M-PESA, the average balance on M-
PESA accounts was around U.S.$3. This is partly a “large number” prob-
lem: if 900,000 people used M-PESA to save, that would “only” be 10
percent of users and their savings would be diluted within an “average”
savings balance. But the fundamental problem is that there is still a lot of
conversion of electronic value back into cash, say following receipt of a
domestic remittance. We attribute this to a combination of factors:
Lack of marketing –■■ Safaricom does not want to publicly promote
the savings usage of M-PESA for fear of provoking the Central Bank
into tighter regulation of M-PESA.
Customer pricing –■■ there is a flat fee of around U.S. 30¢ for with-
drawals under U.S.$30, which means that small withdrawals carry a
large percent fee.
Product design –■■ M-PESA works very much like an electronic check-
ing account, and does not offer structured saving products which may
help people build discipline around savings.
Inflation –■■ M-PESA does not pay interest. In an environment with 15
percent inflation (during its first full year of operation in 2008), this may
be too onerous for savings.
Trust –■■ deposits are not supervised by the Central Bank. And unlike
payments, where trust can be validated experientially in real time, sav-
ings requires trust over a longer period of time.
Privacy –■■ people may want more privacy in their savings behavior
than an agent provides.
Excess liquidity –■■ 23,400 cash-in points are also 23,400 cash-out
points. The ubiquity of M-PESA agents may make it too easy for cus-
tomers to cash-out their funds, thus limiting their ability to accumulate
large balances.
Rather than expecting Safaricom to develop and market richer savings
services, we believe that M-PESA should support savings propositions
by linking into banks. M-PESA would then become a massive transaction
acquisition network for banks rather than an alternative to them. Safa-
ricom is beginning to connect with banks. In May 2010, for example,
181
Equity Bank and M-PESA announced a joint venture, M-KESHO, which
permits M-PESA users to move money between their M-PESA mobile
wallet and an interest-bearing Equity Bank account.
M-PESA would also benefit from establishing further linkages with insti-
tutions beyond banks, such as billers, distributors, and employers. By
promoting M-PESA as a mechanism for distributing salaries and social
welfare payments, enabling payments across supply chains, and paying
bills, the need for cash-in and cash-out would be minimized, and, as a
result, a key component of transaction costs could be reduced. We also
suspect savings balances would be higher if people received payments
directly into their account rather than in cash, and if they had more useful
things they could do with their money in electronic form.
Concluding thoughts: how M-PESA can reinvigorate visions around financial inclusionImagine a world where banks are nowhere near where you live. The near-
est branch is 10 kilometers away, but it takes you almost an hour to get
there by foot and bus because you do not have your own wheels. With
waiting times at the branch, that is a round-trip of two hours – a quarter
or so of your working day gone. The bus fare is only 50 cents, but that is
one quarter of what you make on a good day. So each banking transac-
tion costs you the equivalent of almost half a day’s wages. It would be
like an ATM charging us something like U.S.$50 for each transaction,
given what we earn.
Then, imagine a world without credit instruments or electronic payments.
No checks, no cards, no money orders, no direct debits, no internet bank-
ing. All your transactions are done in cash or, worse, by bartering goods.
All exchanges are physical, person-to-person, hand-to-hand. Consider
the hassle and the risk of sending money to distant relatives, business
partners, or banks.
How would you operate in such a world? A recent book, “Portfolios of
the poor”, [Collins et al. (2009)] has documented how poor people cope.
How they save to “push” some excess money from today to tomorrow,
how they borrow to “pull” tomorrow’s money to fund some needed ex-
pense today. You store some cash in the home to meet daily needs, you
park it with a trusted friend for emergencies, you buy jewelry because
that represents a future for your children, you pile up some bricks for the
day when you can build an extra room in your house. You make regular
contributions to a savings group with a circle of friends to build up a pot,
and one day it will be your turn to take that pot home to buy new clothes.
You also borrow from friends, seek advances from your employer, pawn
some of your jewelry, and go to the moneylender.
The authors of “Portfolios of the poor” document some poor families
across India, Bangladesh, and South Africa using up to 14 different
mechanisms to manage their financial lives. They have few options, but
you need to deploy all your ingenuity to use them all, precisely because
none are very good. Some are not very safe because of their sheer physi-
cality. If you save by storing your grain or buying goats, when your village
hits hard times you may not be able to find ready buyers for your grain
or goats. Forget about getting loans from neighbors during hard times.
The local moneylender runs a quasi-monopoly in the village because it
is too costly for you to go to other moneylenders in other villages, and in
any case they do not know you there. So you end up paying dearly for
a loan.
We estimate that over 2 billion people need to cope with such circum-
stances. The lack of good financial options is undoubtedly one of the
reasons why poor people are trapped in poverty. They cannot sustain or
even aspire to higher income because they are not able to invest in better
farming tools and seeds to enhance their productivity, start a microen-
terprise, or even take the time to search for better paying employment
opportunities. Their income is volatile, often fluctuating daily, so without
reliable ways of pushing and pulling money between good days and bad
days they may have to face the stark decision to pull the kids out of
school or put less food on the table during bad patches. And without
good financial tools they may not be able to cope with shocks that set
them back periodically. Most of these shocks are foreseeable if not en-
tirely predictable: a drought, ill-health, lifecycle events such as marriage
and death.
Cash is the main barrier to financial inclusion. As long as poor people
can only exchange value in cash – or, worse, physical goods – they will
remain too costly for formal financial institutions to address in significant
numbers. Collecting low-value cash deposits and redeeming their sav-
ings back into small sums of cash requires a costly infrastructure which
few banks are willing to make extensive in low-income or rural areas. But
once poor people have access to cost-effective electronic means of pay-
ments such as M-PESA, they could, in principle, be profitably marketable
subjects by a range of financial institutions.
M-PESA itself does not constitute financial inclusion. But it does give us
glimpses of a commercially sound, affordable and effective way to offer
financial services to all.
References Camner, G. and E. Sjöblom, 2009, “Can the success of M-PESA be repeated? A Review of •
implementations in Kenya and Tanzania,” Valuable Bits note, July
Collins, D., J. Morduch, S. Rutherford, and O. Ruthven, 2009, Portfolios of the poor: how the •
world’s poor live on $2 a day,Princeton University Press
Davidson, N., 2009, “Tactics for tipping markets: influence perceptions and expectations,” GSM •
Association, Mobile money for the unbanked blog, November 15
Financial Sector Deepening Trust, FSDT, 2007a, “Finaccess Kenya 2006: results of a financial •
survey on access to financial services in Kenya”
Financial Sector Deepening Trust [FSDT], 2007b, “Key findings of the FinScope survey in •
Tanzania in 2006”
The Capco Institute Journal of Financial TransformationMobile Payments Go Viral: M-PESA in Kenya
182
Financial Sector Deepening Trust [FSDT], 2009a, “FinAccess national survey 2009: dynamics of •
Kenya’s changing financial landscape,” June
Financial Sector Deepening Trust [FSDT], 2009b, “Research on mobile payments experience: •
M-PESA in Kenya,” Unpublished draft, December
GSM Association, 2009, “Wireless intelligence database,” Available at www.wirelessintelligence.•
com
Heyer, A. and I. Mas, 2009, “Seeking fertile grounds for mobile money,” unpublished paper•
Hughes, N. and S. Lonie, 2009, “M-PESA: mobile money for the unbanked,” Innovations, •
special edition for the Mobile World Congress 2009, MIT Press
Isaacs, L., 2008, IAMTN presentation, MMTA conference Johannesburg, May•
Jack, B. and T. Suri, 2011, “The economics of M-PESA,” Working paper, http://www.mit.•
edu/~tavneet/M-PESA-Final.pdf
Juma, V., 2009, “Family bank offers new service linking accounts to M-PESA,” Business •
Daily,18 December
Kimenyi, M. and N. Ndung’u, 2009, “Expanding the financial services frontier: lessons from •
mobile phone banking in Kenya,” Brookings Institution, October
Kinyanjui, K., 2009, “Yu launches new mobile cash transfer platform,” Business Daily, 16 •
December
Mas, I., 2008, “M-PESA vs. G-Cash: accounting for their relative success, and key lessons for •
other countries,” CGAP, unpublished, November
Mas, I., 2008b, “Realizing the potential of branchless banking: challenges ahead,” CGAP focus •
note 50
Mas, I., 2009, “The economics of branchless banking,” Innovations, Volume 4, Issue 2, MIT •
Press, Spring
Mas, I. and O. Morawczynski, 2009, “Designing mobile money services: lessons from M-PESA,” •
Innovations, Vol. 4, Issue 2, MIT Press, Spring
Mas, I. and A. Ng’weno, 2010, “Three keys to M-PESA’s success: branding, channel •
management and pricing,” Journal of Payments Strategy and Systems, Vol. 4, No. 4
Mas, I. and S. Rotman, 2008, “Going cashless at the point of sale: hits and misses in developed •
countries,” CGAP focus note 51
Morawczynski, O., 2008, “Surviving in the dual system: how M-PESA is fostering urban-to-rural •
remittances in a Kenyan slum,” HCC8 conference proceedings, Pretoria
Morawczynski, O., 2009, “Exploring the usage and impact of transformational m-banking: the •
case of M-PESA in Kenya,” Unpublished draft
Morawczynski, O. and G. Miscione, 2008, “Examining trust in mobile banking transactions: the •
case of M-PESA in Kenya,” in: C. Avgerou, M. Smith, and P. van den Besselaar (eds.) “Social
dimensions of information and communication technology policy: proceedings of the 8th
International Conference on Human Choice and Computers HCC8,” International Federation for
Information Processing TC 9, Pretoria, South Africa, September 25-26, Volume 282, 287–298
Morawczynski, O. and M. Pickens, 2009, “Poor people using mobile financial services: •
observations on customer usage and impact from M-PESA,” CGAP Brief, August
Okoth, J., 2009, “Regulator gives M-PESA a clean bill of health,” The Standard, 27 January•
Okuttah, M., 2009, “Safaricom changes method of recruiting M-pesa agents,” Business Daily, •
23 December
Ratan, A. L., 2008, “Using technology to deliver financial services to low-income households: •
a preliminary study of Equity Bank and M-PESA customers in Kenya,” Microsoft Research
Technical Report, June
Safaricom, 2011, “M‐PESA key performance statistics,” available on (http://www.safaricom.•
co.ke/index.php?id=1073) and various financial reports available on www.safaricom.co.ke
183
The Capco Institute Journal of Financial TransformationA General Structural Approach For Credit Modeling Under Stochastic Volatility
184
Guidelines for Manuscript Submissions
Manuscript guidelinesAll manuscript submissions must be in English.
Manuscripts should not be longer than 7,000 words each. The maximum
number of A4 pages allowed is 14, including all footnotes, references, charts
and tables.
All manuscripts should be submitted by e-mail directly to the editor@capco.
com in the PC version of Microsoft Word. They should all use Times New
Roman font, and font size 10. Accompanying the Word document should be a
PDF version that accurately reproduces all formulas, graphs and illustrations.
Please note that formatting in italics or underline will not be reproduced, and
that bold is used only in subtitles, tables and graphs.
Where tables or graphs are used in the manuscript, the respective data
should also be provided within a Microsoft Excel spreadsheet format.
The first page must provide the full name(s), title(s), organizational affiliation
of the author(s), and contact details of the author(s). Contact details should
include address, phone number, fax number, and e-mail address.
Footnotes should be double-spaced and be kept to a minimum. They should
be numbered consecutively throughout the text with superscript Arabic
numerals.
For monographs
Aggarwal, R., and S. Dahiya, 2006, “Demutualization and cross-country
merger of exchanges,” Journal of Financial Transformation, Vol. 18,
143-150
For books
Copeland, T., T. Koller, and J. Murrin, 1994, Valuation: Measuring and
Managing the Value of Companies. John Wiley & Sons, New York, New York
For contributions to collective works
Ritter, J. R., 1997, Initial Public Offerings, in Logue, D. and J. Seward, eds.,
Warren Gorham & Lamont Handbook of Modern Finance, South-Western
College Publishing, Ohio
For periodicals
Griffiths, W. and G. Judge, 1992, “Testing and estimating location vectors
when the error covariance matrix is unknown,” Journal of Econometrics 54,
121-138
For unpublished material
Gillan, S. and L. Starks, 1995, Relationship Investing and Shareholder
Activism by Institutional Investors. Working Paper, University of Texas
Guidelines for authorsIn order to aid our readership, we have established some guidelines to
ensure that published papers meet the highest standards of thought leader-
ship and practicality. The articles should, therefore, meet the
following criteria:
1. Does this article make a significant contribution to this field of research?
2. Can the ideas presented in the article be applied to current business mod-
els? If not, is there a road map on how to get there.
3. Can your assertions be supported by empirical data?
4. Is my article purely abstract? If so, does it picture a world that can exist in
the future?
5. Can your propositions be backed by a source of authority, preferably
yours?
6. Would senior executives find this paper interesting?
Subjects of interest
All articles must be relevant and interesting to senior executives of the leading
financial services organizations. They should assist in strategy formulations.
The topics that are of interest to our readership include:
• Impactofe-financeonfinancialmarkets&institutions
• Marketing&branding
• Organizationalbehavior&structure
• Competitivelandscape
• Operational&strategicissues
• Capitalacquisition&allocation
• Structuralreadjustment
• Innovation&newsourcesofliquidity
• Leadership
• Financialregulations
• Financialtechnology
Manuscript submissions should be sent to
Prof. Shahin Shojai, Ph.D.
The Editor
Capco
Broadgate West
9 Appold Street
London EC2A 2AP
Tel: +44 207 426 1500
Fax: +44 207 426 1501
185
Request for Papers Deadline October 14th, 2011
The world of finance has undergone tremendous change in recent years.
Physical barriers have come down and organizations are finding it harder
to maintain competitive advantage within today’s truly global market place.
This paradigm shift has forced managers to identify new ways to manage
their operations and finances. The managers of tomorrow will, therefore,
need completely different skill sets to succeed.
It is in response to this growing need that Capco is pleased to publish
the “Journal of financial transformation.” A journal dedicated to the
advancement of leading thinking in the field of applied finance.
The Journal, which provides a unique linkage between scholarly
research and business experience, aims to be the main source of
thought leadership in this discipline for senior executives, management
consultants, academics, researchers, and students. This objective
can only be achieved through relentless pursuit of scholarly integrity
andadvancement. It is for this reason that we have invited some of the
world’s most renowned experts from academia and business to join
our editorial board. It is their responsibility to ensure that we succeed in
establishing a truly independent forum for leading thinking in this new
discipline.
You can also contribute to the advancement of this field by submitting
your thought leadership to the Journal.
We hope that you will join us on our journey of discovery and help shape
the future of finance.
Prof. Shahin Shojai
For more info, see opposite page
2010 The Capital Markets Company. VU: Prof. Shahin Shojai,
Prins Boudewijnlaan 43, B-2650 Antwerp
All rights reserved. All product names, company names and registered trademarks in
this document remain the property of their respective owners.
186
Layout, production and coordination: Cypres – Daniel Brandt, Kris Van de Vijver and
Pieter Vereertbrugghen
Graphic design: Buro Proper – Bob Goor
Photographs: Bart Heynen
© 2011 The Capital Markets Company, N.V.
All rights reserved. This journal may not be duplicated in any way without the express
written consent of the publisher except in the form of brief excerpts or quotations
for review purposes. Making copies of this journal or any portion there of for any
purpose other than your own is a violation of copyright law.
MSc in Insurance and Risk Management
Cass is one of the world’s leading academic centresin the insurance field. What's more, graduatesfrom the MSc in Insurance and Risk Managementgain exemption from approximately 70% of theexaminations required to achieve the AdvancedDiploma of the Chartered Insurance Institute (ACII).
For applicants to the Insurance and RiskManagement MSc who already hold a CII Advanced Diploma, there is a fast-track January start, giving exemption from the first term of the degree.
To find out more about our regular informationsessions, the next is 10 April 2008, visitwww.cass.city.ac.uk/masters and click on'sessions at Cass' or 'International & UK'.
Alternatively call admissions on:
+44 (0)20 7040 8611
With a Masters degree from Cass Business School,you will gain the knowledge and skills to stand out in the real world.
Minimise risk,optimise success
The Capco Institute
Journal of Financial Transformation #32 08.2011
JournalThe Capco Institute Journal of Financial Transformation
#32Applied Finance
CApCo.Com
AmsterdamAntwerp
BangaloreChicago
FrankfurtGenevaLondon
New Yorkparis
San FranciscoToronto
Washington, D.C.Zurich
Recipient of the Apex Awards for Publication Excellence 2002-2011
08.2011