Upload
trinhthu
View
219
Download
0
Embed Size (px)
Citation preview
Determinants of Sovereign Risk: Macroeconomic
Fundamentals and the Pricing of Sovereign Debt1
Jens Hilscher Yves Nosbusch
Brandeis University London School of Economics
First draft: November 2004
This version: July 2008
1Contact: Jens Hilscher, International Business School, Brandeis University, 415 South Street,Waltham MA 02453, USA. Phone 781-736-2261, email [email protected]. Yves Nosbusch, De-partment of Finance, London School of Economics, Houghton Street, London WC2A 2AE, UK. Phone+44-20-7955-7319, email [email protected]. An earlier version of this paper was circulated underthe title “Determinants of Sovereign Risk.” We are grateful to John Campbell, N. Gregory Mankiw,Kenneth Rogoff, and Jeremy Stein for their advice and suggestions. We thank Alberto Alesina, SilviaArdagna, Gregory Connor, Darrell Duffie, Nicola Fuchs-Schündeln, Amar Gande, David Hauner, DirkJenter, David Laibson, David Lesmond, Emi Nakamura, Ricardo Reis, Andrei Shleifer, Monica Singhal,Jón Steinsson, Federico Sturzenegger, Suresh Sundaresan, Glen Taksler, Sam Thompson, and seminarparticipants at Harvard University, the Federal Reserve Board of Governors, Warwick Business School,Oxford Saïd Business School, the Bank for International Settlements, Macalester, UC Irvine, BrandeisUniversity, Barclays Global Investors, Oak Hill Platinum Partners, the 2006 MMF conference, the 2006FMA conference, the Swiss National Bank, the 2007 Moody’s and Copenhagen Business School CreditRisk Conference, the 2007 EFA meetings, the 2007 EEA meetings, the Bank of England, and the 2008ESRC/WEF Warwick Conference on Incentives and Governance in Global Finance for helpful com-ments and discussions. We also thank Timothy Callen and James Morsink at the IMF, Gonca Okurat the World Bank, Carmen Reinhart, and Alvin Ying at J.P. Morgan for providing us with data andPavel Bandarchuk for research assistance.
Abstract
This paper investigates the effects of macroeconomic fundamentals on emerg-ing market sovereign credit spreads. We pay special attention to the volatil-ity of fundamentals in addition to their levels. In reduced form regressions,the volatility of terms of trade in particular has a statistically and econom-ically significant effect on spreads. We propose an empirically tractablemodel that addresses the distinct features of sovereign risk. Our model cap-tures a significant part of the empirical variation in spreads. It fits betterfor borrowers of lower credit quality. We also estimate default probabilitiesin a hazard model and link these probabilities to observed spreads.
JEL classifications: F34, G12, G13, G15
Keywords: sovereign spreads, credit risk, bond pricing, terms of trade, de-fault probabilities.
2
1 Introduction
There is tremendous variation in the interest rates emerging market governments pay
on their external debt. This is true both across countries and over time. A common
measure of a country’s borrowing cost in international capital markets is its yield spread,
which is defined as the difference between the interest rate the government pays on its
external U.S. dollar denominated debt and the rate offered by the U.S. Treasury on debt
of comparable maturity.
In this paper we consider how much of the variation in yield spreads across countries
and over time can be explained by macroeconomic fundamentals. We focus in partic-
ular on the explanatory power of the volatility of fundamentals. From a theoretical
perspective, the volatility of fundamentals should matter for the pricing of defaultable
debt, just as the level of these fundamentals does. All else equal, a country with more
volatile fundamentals is more likely to experience a severe weakening of fundamentals
which will force it into default. This risk should be reflected in a higher yield spread
on its bonds.1 While this idea is well understood in theory, it has been largely ignored
by the empirical literature on sovereign debt, which has tended to focus exclusively on
level variables.2 In contrast, the present paper includes both the level and the volatility
of macroeconomic fundamentals in its empirical analysis.
1This intuition corresponds to that of the structural bond pricing literature starting with Merton’s(1974) seminal model of risky corporate debt: an investor can view risky debt as a combination of safedebt and a short position in a put option. The value of the option depends on the volatility of theunderlying firm value; hence the value of the firm’s bonds also depends on this volatility. In the caseof a country, higher volatility of fundamentals increases the value of the default option and thereby thespread.
2Two exceptions are Edwards (1984), who includes variability of reserves and finds that it is insignif-icant, and Westphalen (2001), who finds some limited effect of changes in local stock market volatilityon changes in short term debt prices. In the corporate bond context, Campbell and Taksler (2003)find a strong empirical link between equity volatility and yield spreads.
3
There are three main parts to our analysis. In the first part we analyze sovereign
debt pricing in a reduced form regression framework. We measure yield spreads on
external U.S. dollar denominated debt using data from J.P. Morgan’s Emerging Market
Bond Index (EMBI) covering a set of 32 emerging market countries from 1994 to 2006.
We report results from reduced form regressions of these spread observations on a set of
macroeconomic fundamentals.
Motivated by previous studies (e.g. Edwards (1984, 1986)), we choose the ratio of
debt to GDP as our starting point. We then increase our set of explanatory variables,
paying particular attention to terms of trade. Terms of trade measure the price of
a country’s exports relative to its imports. Changes in the terms of trade should
therefore have a direct effect on a country’s ability to generate dollar revenue from
exports and thus make payments on its external dollar denominated debt. This point
was made by Bulow and Rogoff (1989). Mendoza (1997) presents evidence that terms
of trade volatility has negative effects on long run growth and Mendoza (1995) finds
that terms of trade volatility is important for explaining output variability at business
cycle frequencies. Another appealing feature of terms of trade is that they are plausibly
exogenous since they are calculated using world prices.
Our reduced form regression results show that spreads indeed tend to be higher for
countries that have recently experienced adverse terms of trade shocks, while countries
that have seen their terms of trade improve tend to have lower spreads. We also find that
the volatility of terms of trade has a highly significant effect on spreads, both statistically
and economically.
We also include the growth rate of GDP and the volatility of GDP growth in our
reduced form regression model and find that they are statistically and economically
4
significant and add overall explanatory power. In order to measure a country’s access to
liquid assets that can be used for debt repayment, we include the ratio of reserves to GDP
and find it to be significant. Finally, we add a measure of a country’s default history to
this set of explanatory variables. We find that, even after accounting for macroeconomic
fundamentals, spreads are higher for countries that have recently emerged from default.
This is consistent with Reinhart, Rogoff, and Savastano (2003), who argue that a key
predictor of future default is a country’s history of default.
In terms of economic significance, terms of trade turn out to be the most important
determinant of spreads, followed by the ratio of debt to GDP. Looking at explanatory
power in the cross-section and the time-series, the ratio of debt to GDP captures a large
share of the within-country time-series variation in spreads but has less explanatory
power in the cross-section. This suggests that what is important is whether a country’s
debt level is high relative to its own mean, not whether it is high relative to other
countries.
In the second part of the paper, we consider the link between debt prices and default
probabilities in the context of a structural form framework. Standard structural models
of risky corporate debt cannot be used when pricing sovereign debt. In these models,
the firm defaults and is taken over by creditors if firm asset value falls below liabilities.
In the sovereign case, there is no equivalent of firm asset value with an observable market
price. Moreover, even if one could agree on an underlying wealth measure, creditors
typically do not hold claims to a country’s assets in the event of default.
We propose an empirically tractable model in which the level and the volatility
of macroeconomic fundamentals determine the probability of default and debt prices.
The model addresses the distinct features of sovereign debt while retaining the familiar
5
intuition of structural corporate bond pricing models. In the model a country defaults if
fundamentals are too low or the debt burden is too high, i.e. if repayment becomes too
costly. Fundamentals determine both the probability of default and the recovery value
in the event of default.
Using data on country defaults and fundamentals from 1970 to 2003, we fit our model
to price bonds directly. We estimate the relative importance of fundamentals by running
a predictive regression of a default indicator on our set of explanatory variables. We
then use the estimated model implied parameters and default probabilities to predict
spreads and find that the model captures a large share of the variation in spreads.
In the third part of the paper, we estimate default probabilities in a reduced form
hazard model and use fitted probabilities to price debt. This approach effectively relaxes
the restrictions on functional form imposed by the structural model that we present
in the second part of the paper. We again use the sample period starting in 1970.
We estimate the conditional probability of default over the next period in a reduced
form logit model following the approach proposed by Shumway (2001) in the context
of predicting corporate bankruptcy. We find that the level of debt to GDP and the
volatility of terms of trade are important predictors of default, in line with their effect on
spreads. We use the reduced form default probabilities to calculate predicted spreads
and find an improvement in explanatory power relative to model implied spreads.
Both the structural model and the reduced form hazard model allow us to identify
systematic mispricing in the market, which is something reduced form spread regressions
like the ones presented in the first part of the paper cannot achieve since they fit the
overall mean of observed spreads by construction. We find that model implied and
reduced form spreads can account for 51% and 61% of the variation in observed spreads,
6
respectively.
We also group bond spread observations by their estimated default probability and
compute average levels of observed and model implied spreads for each group. We find
that model implied spreads are smaller than actual spreads for all the groups,3 but that
the proportion of the yield spread explained by our model increases significantly with the
probability of default. Huang and Huang (2003) find a similar pattern when calibrating
structural form models to U.S. corporate bond prices.
This paper adds to the large literature on the empirical determinants of sovereign
yield spreads. The literature varies widely in choice of variables and methodology. Some
papers concentrate on reduced form regressions of spreads on explanatory variables.
Examples of this line of research include Edwards (1986), Eichengreen and Mody (1998),
Min (1998), Beck (2001), and Ferrucci (2003), among others. These authors explore a
large set of macroeconomic variables to explain spreads.4 Although no clear consensus
on the determinants of spreads emerges from these studies, the level of debt to GDP is
generally found to be significant. As already mentioned, we pay special attention to
measures of volatility that are largely absent from this literature.
Duffie, Pedersen, and Singleton (2003) develop a flexible reduced form model of
sovereign yield spreads that builds on the framework of Duffie and Singleton (1999).
Their model accounts for the specific context in which a country may selectively default
3This is not surprising given that we only model the default component of the spread. Hund andLesmond (2007) provide evidence that a significant part of the spread on emerging market corporateand sovereign debt may be explained by liquidity. In related work, Chen, Lesmond, and Wei (2007)present evidence that liquidity is priced in U.S. corporate bonds. Longstaff, Mithal, and Neis (2005)find that the majority of the U.S. corporate default swap spread is explained by default risk and thatthe non-default component is time-varying and closely related to measures of liquidity.
4Some papers such as Cantor and Packer (1996), and Kamin and von Kleist (1999) instead use creditratings as a comprehensive measure.
7
on or restructure part of its debt. They estimate the model using weekly data on
Russian dollar-denominated debt and US swap yields between 1994 and 1998. Pan and
Singleton (2006) explore the term structure of credit default swap (CDS) spreads for
Mexico, Turkey, and Korea from 2001-2006 and consider the risk-neutral credit event
intensities and loss rates that best describe the CDS data. While these studies relate
spreads implied by a reduced from pricing model to political factors, foreign currency
reserves, oil prices, and measures of liquidity such as VIX, macroeconomic fundamentals
do not directly enter the estimation of the pricing model. In contrast, our focus is to
explore directly the extent to which variation in macroeconomic fundamentals determine
spreads and default probabilities across countries and over time.
There is also a large literature on predicting banking and currency crises and sov-
ereign default using macroeconomic fundamentals.5 Again, the focus of these studies
is on level variables, rather than volatility. There are some exceptions: in two recent
papers, Catão and Sutton (2002) and Catão and Kapur (2006) predict sovereign default
in a hazard model using a similar setup to the one used in the third part of this pa-
per. They identify volatility of terms of trade as an important predictor of default.
These studies on default and crisis prediction do not, however, relate the determinants
of default to the determinants of spreads.
The remainder of the paper is organized as follows. Section 2 describes the data.
Section 3 describes the results from using fundamentals to explain spreads in a reduced
form regression framework. Section 4 discusses why corporate bond pricing models
5Early contributions to this literature include Hajivassiliou (1987, 1994). Berg, Borensztein, Milesi-Ferretti, and Pattillo (1999), Kaminsky and Reinhart (1999), and Goldstein, Kaminsky, and Reinhart(2000) have concentrated on constructing what they refer to as “early warning systems.” Berg, Boren-sztein, and Pattillo (2004) provide an overview of this line of research.
8
cannot be applied in the case of sovereign debt and introduces our model of sovereign
debt. We derive the model implied probability of default and spread. In section 5 we
fit our model to the data directly, and find that it accounts for a substantial fraction
of observed spread variation. Section 6 turns to reduced form default prediction and
compares model and reduced form predicted spreads to realized spreads. This section
also relates our results to recent findings in the corporate bond literature. Section 7
concludes.
2 Choice of variables and data description
Our empirical investigation concentrates on explaining variation in sovereign external
debt prices across countries and over time. We restrict ourselves to dollar-denominated
debt instruments issued or guaranteed by emerging market governments.6 Our measure
of yield spreads over U.S. Treasuries is J.P. Morgan’s Emerging Markets Bond Index
Global (EMBI Global). This series consists of daily data starting in January of 1994
and covers a sample of 32 emerging market countries. It includes a basket of Brady
bonds, loans, and Eurobonds. The average maturity of the debt instruments included
in the index is close to 12 years. In order to be included in the index, debt instruments
need to satisfy some basic liquidity requirements such as verifiable daily prices and cash
flows. We discuss these inclusion requirements in more detail in the data appendix.
Table 1 lists the set of countries with available EMBI spread data and groups coun-
tries into regions. The largest regional group is made up of countries in Latin America
with 12 countries. The data set also includes 5 countries in Africa, 6 in Eastern Europe
6In contrast, the yield on local currency debt tends to be driven mainly by local inflation risk.
9
and a total of 9 in the Middle East and Asia. The spread data begin in 1994, but not all
countries have data going back that far. We report the first year of available spread data
for each country. In 1994 the spread data series begin for Argentina, Brazil, Mexico,
Venezuela, Nigeria, Bulgaria, Poland, China, and South Korea. By 1996, nearly half of
the eventual set of countries have available data and since 1998 J.P. Morgan has been
constructing series for close to two thirds of the countries.
There is tremendous variation in spreads across countries at any given point in time.
To illustrate this, Table 1 reports median EMBI spreads for 2003 by country. For
countries that are not in default in 2003, median daily spreads vary from 31 basis points
for Hungary to 1188 basis points for Ecuador.7 In addition to the cross-sectional
variation in spreads there is substantial variation in spreads over time, which is apparent
from Figure 1 which plots average EMBI from 1994 to 2006.8
We start our investigation of spread determinants by considering some global evi-
dence that suggests that terms of trade are an important determinant of sovereign yield
spreads. Since many emerging markets are important commodity exporters, we expect
an improvement in the terms of trade and a resulting decline in spreads when commodity
prices have increased. We plot commodity prices as measured by the CRB commodity
index9 in Figure 1 and find that commodity prices and spreads in fact exhibit a strong
negative correlation of -0.49, which is apparent from the graph.
A natural interpretation is that when commodity prices are high, commodity ex-
7The spreads can be much higher for countries that are in default. The median daily spread forArgentina in 2003 is 5441 basis points. Côte d’Ivoire, Nigeria, and Uruguay are also in default in 2003.
8We weight the individual countries’ EMBI spreads by the governments’ total outstanding externaldebt levels.
9The data appendix provides more detailed information about the index which is constructed by theCommodity Research Bureau.
10
porters are more likely to repay their external debt, which reduces the yield spread they
face in international capital markets. In particular, the well-documented narrowing of
emerging market spreads over the last few years has been accompanied by an extraordi-
nary run-up in commodity prices. We use more detailed country specific terms of trade
data rather than a single global commodity index in our empirical implementation.
We now turn to a systematic analysis of spreads across countries and over time. For
this purpose we gather annual cross-country data on macroeconomic fundamentals from
a variety of sources, details of which can be found in the data appendix. External
dollar-denominated debt issued by emerging market sovereign governments commands
a yield spread over U.S. Treasury bonds primarily to compensate investors for the risk
of default.10 In our choice of variables we therefore focus on indicators that are related
to the risk of default.
In thinking about default risk in the case of a firm, the standard approach is to
compare firm asset value to liabilities. If a measure of asset value is not available one
could calculate the present discounted value of future revenues or dividends, compare it
to the level of liabilities, and calculate a proxy for leverage. The two components of
this calculation are a measure of liabilities and a measure of repayment capabilities.
In the case of a country, the natural analogues to these two measures are debt and
GDP. The ratio of debt to GDP is indeed arguably the most important explanatory
variable identified by existing empirical work on sovereign debt pricing. We follow the
previous literature and construct a country’s external debt to GDP ratio (Debt/GDP).
Another direct determinant of a country’s ability to repay often considered in the lit-
10There are of course other components of the spread, including liquidity and taxes, but these arenot the focus of our investigation. We discuss this point further in Section 5 when comparing actualand model implied spreads.
11
erature is its access to liquid assets, commonly measured by reserves. We include the
ratio of reserves (including gold) to GDP (Reserves/GDP) in our analysis.
We then add a more explicit measure of the government’s ability to generate dollar
revenue for repayment, namely the country’s terms of trade. The terms of trade measure
the price of a country’s exports relative to its imports. To see why this measure
is relevant for the pricing of debt, consider an oil-exporting country. The country
generates dollar revenue by exporting oil and spends dollars on imports. If the price
of oil rises, the country is in a better position to generate export revenue and repay
its dollar-denominated liabilities. Another reason why terms of trade is an appealing
measure is that it is plausibly exogenous. This is because terms of trade are calculated
using prices determined in world markets. Since terms of trade is a relative price series,
we measure the country’s terms of trade by constructing the percentage change in the
terms of trade over the past five years (Change in terms of trade). A positive number
means that a country’s exports have become more expensive relative to its imports.
Bondholders, however, do not only care about recent changes in the terms of trade
but also about the risk of a large adverse shock in the future, say a large decline in the
price of oil in our example. As a result, we expect the volatility of terms of trade to
influence debt prices. More specifically, volatility should be positively correlated with
spreads. We calculate the standard deviation of the percentage change in terms of trade
(Volatility of terms of trade) over the past ten years.11 We discuss the relation between
the volatility of fundamentals and debt prices in the context of a model in Sections 4
and 5.
11When less data is available, we require as few as 6 observations; the data appendix gives furtherdetails. For the Dominican Republic there is insufficient data to calculate volatility of terms of trade.
12
Table 1 reports the median levels of debt to GDP and the volatility of terms of trade.
There is considerable variation in the median levels of these variables across countries.
The median level of external debt ranges from 14% of GDP for China to 115% for Côte
d’Ivoire. Volatility of terms of trade varies substantially cross-sectionally from less than
five percent (China, Croatia, Hungary, Lebanon, Panama, and South Korea) to annual
volatilities above twenty percent (Nigeria and Venezuela). High terms of trade volatility
is to a large extent driven by oil prices. Crude petroleum and refined petroleum products
are among the top two export categories for the three countries with the highest terms
of trade volatilities. However, excluding countries with more than 20% of their exports
concentrated in oil does not affect our empirical results.
We also expect adverse changes in a country’s output to affect debt prices and default.
If a country has recently experienced slow or negative growth, it will likely find it more
difficult to meet payments on its external debt. We add the growth rate of GDP (GDP
growth) to our set of explanatory variables. As before we also expect the volatility of
output growth to be important. We calculate the standard deviation of annual GDP
growth (Volatility of GDP growth) using a rolling ten year backward looking window.
A country’s recent default history is another predictor of default risk. We count the
years since the country’s last year in default (Years since last default). If unrestricted,
this measure can become very large for countries that continue to stay out of default.
Because we expect additional years out of default to have incrementally less importance,
we cap the variable at 10 and set it equal to 11 if a country has never defaulted. It
is set equal to zero if a country is in default. This variable is motivated by the work
of Reinhart, Rogoff, and Savastano (2003), who argue that history of default is an
important predictor of future default. These authors have constructed annual default
13
indicators for different debt categories for a large sample of countries going back to the
early nineteenth century. We use their series for default on total debt, which includes
foreign currency bonds and foreign currency bank debt. This variable is also the basis
for our investigation of determinants of country default in Sections 5 and 6.
In order to summarize default activity for the set of 32 countries, we report the
number of defaults from 1970 to 2003 in Table 1. The first thing to note is that this
was not a quiet period. There are a total of 38 sovereign default episodes. A number
of countries defaulted on multiple occasions: Peru and Uruguay each have four defaults
since 1970. On the other hand, nine countries did not go into default in the sample
period.
3 Reduced form regressions of sovereign spreads on
macroeconomic fundamentals
We now explore how much of the variation in sovereign yield spreads can be explained by
macroeconomic fundamentals in a reduced form framework. We use a linear regression
model and run panel regressions of yield spreads on explanatory variables. For each
country we construct an annual spread series to match up with the annual frequency of
our measures of fundamentals. We calculate the annual mean of the daily EMBI spread
series from J.P. Morgan. We focus on spread observations while the country is not in
default.12 For an observation to be included in the regression we require that the full
12We make this restriction because our analysis considers determinants of the risk of sovereign defaultand its impact on debt prices. Nigeria is not included since it was in default throughout the sampleperiod.
14
set of explanatory variables is available.
Our regression sample consists of 177 country-year observations covering 30 countries
from 1994 to 2003.13 We restrict the sample to be the same for different specifications.
This has the advantage of allowing us to compare point estimates, significance levels, and
R2 across specifications. Table 2 reports summary statistics for the spread regression
sample. It is apparent that there is substantial variation in spreads and all of our
explanatory variables. For some of these variables, the minimum and maximum values
may look rather extreme. In order to explore whether or not this affects our results,
we winsorize the explanatory variables at the 5th and 95th percentile levels, replacing
values above the 95th percentile of the distribution with the 95th percentile and values
below the 5th percentile with the 5th percentile of the distribution.
Table 3 reports results from running regressions of a country’s spread on different
sets of explanatory variables. To make sure that outliers are not driving some of our
results we re-estimate all the reduced form regressions on the winsorized data. Our
main results are unaffected by these changes.
We start our regression analysis by including only the level of debt to GDP, the
main explanatory variable identified by the existing literature. We find that it is highly
significant but has rather limited ability to explain spread variation at an R2 of 0.11.
We next add the change in a country’s terms of trade and the volatility of its terms of
trade in column (2). Both variables have the expected sign and are significant at the
1% level. Countries that have benefited from improvements in their terms of trade tend
to have lower spreads and countries with higher terms of trade volatilities tend to have
13Although the spread data extend to 2006, some of our explanatory variables are only available until2003.
15
higher spreads. We find a substantial effect on explanatory power of introducing these
measures: the R2 increases from 0.11 to 0.47.
In column (3) we add the remaining measures of macroeconomic fundamentals, re-
serves to GDP, GDP growth, and the volatility of GDP growth. The coefficients on
all the variables have the expected sign and are significant at the 1% level, except for
the coefficient on GDP growth, which is significant at the 10% level. This regression of
spreads on fundamentals delivers an R2 of 0.59.
In our main specification reported in column (4), we include, in addition to our
set of macroeconomic fundamentals, the variable measuring the years since a country’s
last default. We find that, all else equal, the closer a country is to its most recent
default episode, the higher its spread. The coefficient on years since last default is
significant at the 1% level. This suggests that a country’s history of default matters
above and beyond what is accounted for by macroeconomic fundamentals, consistent
with the findings of Reinhart, Rogoff, and Savastano (2003). The coefficients on the
other explanatory variables remain significant at the 1% level, with the coefficient of
GDP growth now being significant at the 5% level. Although including years since last
default adds explanatory power beyond what is captured by underlying macroeconomic
fundamentals, the incremental improvement in R2 is rather modest at 2.3 percentage
points.
In addition to being statistically significant, we find that all seven explanatory vari-
ables are also economically significant. In Panel B of Table 3, we multiply the coefficients
from our main specification with each variable’s in-sample standard deviation. The two
variables with the highest economic impact on spreads are debt to GDP and the volatility
of terms of trade. A one standard deviation increase in debt to GDP is associated with
16
an increase of 106 basis points in the spread, while a one standard deviation increase of
volatility of terms of trade implies an increase of 150 basis points. The corresponding
effects for the other explanatory variables are —51 basis points for change in terms of
trade, —74 basis points for reserves to GDP, —42 basis points for GDP growth, +76 basis
points for the volatility of GDP growth, and —73 basis points for years since last default.
We investigate regional effects in column (5) of Table 3 by including regional dummy
variables. We find that spread levels are significantly lower in Eastern Europe and
South East Asia relative to Latin America (the Latin America dummy is omitted from
the regression).
In summary, we find that spreads are high if the level of debt to GDP is high,
reserves are low, terms of trade have deteriorated, GDP growth has been slow, and if
the volatility of terms of trade or the volatility of GDP growth are high. We find that
measures of volatility, in particular volatility of terms of trade, are highly statistically
and economically significant. Including measures of the volatility of fundamentals leads
to a large improvement in explanatory power relative to the measures used in the existing
literature.
3.1 Explanatory power in the cross-section and the time-series
In this section we investigate whether the explanatory variables which we found to
be significant in the previous section are identified from cross-sectional or time-series
variation. We focus in particular on volatility of terms of trade and debt to GDP.
Figure 2a plots average EMBI spreads against average volatility of terms of trade.
Countries with higher volatility of terms of trade tend to have higher spreads. The graph
17
reflects the strong positive correlation of 0.73 between the two. This is in contrast to
the weaker correlation of 0.30 between countries’ average spread and average level of
debt to GDP. This lack of correlation is apparent in Figure 2b. It seems, therefore,
that spreads are driven mainly by terms of trade volatility and less by debt to GDP
in the cross-section. Figure 2c plots country de-meaned spreads against debt to GDP.
We see that, within countries, debt to GDP moves together with spreads. At 0.50 the
two series are highly correlated, and it seems that debt to GDP is better at explaining
time-series variation in spreads within a country rather than variation across countries.
Next we explore these effects more rigorously by running panel regressions of spreads
on explanatory variables, and including year and country fixed effects, respectively.
Results are reported in Table 4. We find that in the time-series (country fixed effects)
the coefficient on debt to GDP is about twice as large as it is in the cross-section (year
fixed effects). Debt to GDP also explains a larger share of the variation in spreads by
itself in the time-series (R2 of 0.24) than it does in the cross-section (R2 of 0.14). In
contrast, the change in terms of trade and the volatility of terms of trade add more
explanatory power in the cross-section than in the time-series. When including both
variables in the cross-section, the R2 jumps from 0.14 to 0.50, compared to a much more
modest increase from 0.24 to 0.35 in the time-series.
In summary, the change in terms of trade and the volatility of terms of trade are
more successful at explaining the cross-section, whereas debt to GDP derives most of its
explanatory power from the time-series. The latter finding is consistent with Reinhart,
Rogoff, and Savastano’s (2003) idea that some countries are more intolerant to high debt
levels. In that case it matters if a country’s debt to GDP is high relative to the country
specific mean, but it does not matter if it is high relative to other countries.
18
3.2 Robustness checks
We perform several robustness checks on our main specification (Table 3 column (4))
of the reduced form regression analysis. We first test for the significance of several
time-series variables.14 As proxies for the riskless long and short term world interest
rates, we include the 10-year and 6-month U.S. Treasury rates. We also include the U.S.
default yield spread, defined as the spread of corporate bonds with a Moody’s rating of
Baa over U.S. Treasuries. We measure liquidity as the difference between the 3-month
Eurodollar rate and the 3-month Treasury rate.
None of these variables are significant at the 10% level. The only time-series variable
that is marginally significant is the default yield spread, with a p-value of 0.107. The
fact that U.S. interest rates are insignificant is consistent with the findings of Kamin
and von Kleist (1999) who could not identify a robust relationship between a number of
industrial country interest rates and emerging market spreads.15
We also perform several other robustness checks. As mentioned earlier, we first drop
large oil exporting countries from the regression. In particular we drop countries with
more than 20% of exports concentrated in crude petroleum and petroleum products (as
listed in Table 1), which means that Columbia, Ecuador, Venezuela, Russia, and Egypt
are dropped from the regression. The results are largely unaffected indicating that
fluctuations in oil prices are only part of the reason why terms of trade fluctuations
14Diaz Weigel and Gemmill (2006) argue that global factors are more important than country specificfundamentals in explaining the Brady bond prices of Argentina, Brazil, Mexico, and Venezuela, whileJostova (2006) provides evidence of predictability resulting from local factors in these countries.15More generally, there does not seem to be a consensus with regards to the sign of the relationship in
the existing empirical literature. For instance, Min (1998) finds a positive but insignificant effect of theU.S. interest rates on sovereign credit spreads, while Eichengreen and Mody (1998) report a significantnegative effect. Uribe and Yue (2006) find that an increase in the world interest rate causes a declinein emerging market spreads in the short run followed by an overall increase in spreads in the long run.
19
impact sovereign spreads. In order to rule out the possibility that outliers might be
driving our results, we next re-estimate all our regressions, winsorizing our variables at
the 5th and 95th percentiles. Third, we replace the volatility of terms of trade with its
one year lag. We then drop observations in years preceding a default. We also cluster
standard errors by country and by year. Our results are basically unaffected by these
changes.
Finally, we check whether there is evidence of co-movement in spreads not captured
by variation in country fundamentals (reflecting, for instance, some type of contagion
effect). First, we add year dummies to our regression and find that the coefficients on
the dummies are insignificant or only significant at the 5% level. Moreover, adding the
dummies leads only to a very modest increase in R2. Second, we include the implied
volatility of the S&P500 index (VIX) and find that it is insignificant. These results
may seem surprising in light of two recent studies which use higher frequency emerging
market credit default swap (CDS) spreads: Longstaff, Pan, Pedersen, and Singleton
(2007) find that at higher frequencies there is evidence of co-movement in sovereign CDS
spreads while Pan and Singleton (2008) find that the VIX is statistically significant in
explaining CDS spreads of Mexico, Turkey, and Korea.16 However, it is important
to note that these results can be perfectly consistent with our findings. The focus of
our study is on explaining low frequency differences in spread levels, while these other
two studies use higher frequency data and focus on changes in credit spreads. Thus
it seems to be the case that at high frequencies there is co-movement in CDS spread
changes, perhaps due to changes in aggregate liquidity or short run uncertainty, while
16In related work, Berndt, Douglas, Duffie, Ferguson, and Schranz (2005) find that the VIX hassubstantial explanatory power in a high frequency data set of U.S. CDS spreads for the period 2000-2004.
20
at low frequencies cross-country variation in the level of spreads is mostly explained by
differences in macroeconomic fundamentals and their volatility.17
4 A model of fundamentals and sovereign spreads
We now propose a simple and empirically tractable model for pricing sovereign debt.
In the previous section we presented evidence that macroeconomic fundamentals, in
particular measures of debt levels and repayment capabilities, can explain sovereign
yield spreads across countries and over time. Consistent with standard option and debt
pricing intuition, we found that volatility affects debt prices. The natural next step
would be to apply standard pricing models. However, it is not clear how to apply these
models to the case of a sovereign borrower.
In standard structural corporate bond pricing models,18 firm asset value jointly de-
termines default and recovery value. If firm assets fall below liabilities, the firm defaults.
The creditors take over the firm and receive the asset value. A bond is then priced as a
combination of safe debt and a short position in a put option on the asset value of the
firm with a strike price equal to the face value of debt. In such models, leverage and
volatility determine both the probability of default and the spread.
17Another issue with comparing our results with these studies is that some of the data on countryspecific macroeconomic fundamentals, particularly on terms of trade, are not available at the higherfrequencies used by these authors. It may therefore be difficult to fully account for the effects ofcountry specific macroeconomic fundamentals in high frequency data. Furthermore, our sample periodis different from both studies; Longstaff, Pan, Pedersen, and Singleton (2007) consider the period 2000-2007; Pan and Singleton (2008) use data from 2001-2006.18Merton (1974) is arguably the first modern structural bond pricing model. Many other models
have followed. Some examples are Black and Cox (1976), Leland (1994), Longstaff and Schwartz (1995),Anderson and Sundaresan (1996), Mella-Barral and Perraudin (1997) Collin-Dufresne and Goldstein(2001). Huang and Huang (2003) provide an overview of this literature. A common feature of thesemodels is that default is related to an underlying process of asset value.
21
Pricing sovereign debt is fundamentally different. First, there is no claim on a
country’s assets that is traded or that has an observable market price. Moreover, even
if one were to construct a measure of asset value, e.g. by discounting expected future
dollar revenues from exports, asset value is not the relevant measure for pricing sovereign
debt. Creditors do not take over the country in the event of default. Country wealth
therefore does not directly determine default: a country may default even if its wealth
exceeds its liabilities. This observation is consistent with the finding that, empirically,
emerging market economies often have far lower levels of debt to GDP than developed
countries yet default much more frequently (Reinhart, Rogoff and Savastano (2003)).
The inability of creditors to claim country assets also means that a country’s wealth
level may not determine the recovery rate following default.
We propose a simple model in which macroeconomic fundamentals determine debt
prices and default. A country defaults when the cost of repayment becomes “too high.”
Intuitively, the cost of repayment is high because debt levels are high or fundamentals
are low. Our model formalizes this intuition.
We consider a discrete time setup. We denote by Ft+1 a measure of fundamentals
for a country at time t+ 1 and by D a measure of the country’s debt level, suppressing
a country specific subscript for simplicity. We assume that both of these measures are
in comparable units, a point which we discuss in further detail in Section 5. We assume
that for a currently solvent country, default occurs next period if fundamentals are too
22
low relative to the level of debt:19
Ft+1 < D : default occurs at time t+ 1. (1)
We assume that fundamentals follow a martingale and that shocks to fundamentals are
lognormally distributed:
ft+1 = ft + φt+1, where φt+1 ∼ N¡0, σ2
¢, (2)
where lower case letters denote logs and φt+1 is the shock to log fundamentals with
standard deviation σ. Given these assumptions, the probability of default is equal to:
Pt (Ft+1 < D) = Φ
µ−ft − d
σ
¶, (3)
where Φ (.) denotes the standard normal c.d.f.
Default is less likely if fundamentals are high, debt levels are low, and volatility
is low. More precisely, default is less likely if the volatility scaled difference between
fundamentals and debt levels is high. Adopting the terminology of the corporate default
literature, we refer to this scaled difference as distance to default.20
This expression already illustrates the basic intuition of our model: the probability
of default depends on the level as well as the volatility of fundamentals. As discussed
earlier, this effect motivates the inclusion of measures of volatility in our empirical ex-
19In the model we distinguish between measures of debt levels and measures of fundamentals. Fun-damentals are subject to shocks - an adverse shock to fundamentals can trigger default. Debt levelsare known at time t. We therefore suppress the time subscript on D.20Duffie, Saita and Wang (2007) for instance use distance to default to predict bankruptcy. They
refer to it as a volatility adjusted measure of leverage.
23
ploration.
In order to derive bond prices, we need to make assumptions about the payoff to
bondholders in and out of default. We assume that if the country remains solvent,
bondholders receive face value. If the country defaults, we assume that the fractional
recovery value depends on the level of fundamentals, i.e. the worse fundamentals are at
the time of default, the lower the payoff to creditors. In Section 6, we also consider an
alternative assumption of a fixed recovery rate at the time of default. The payoff to
bondholders at time t+ 1, denoted by Bt+1, is given by
solvent : Bt+1 = B (4)
default : Bt+1 = δt+1B,
where B is the face value of the bond, and δt+1 is the fractional recovery rate.
To capture the idea that recovery depends on the level of fundamentals at the time
of default, we assume that δt+1 = γ exp (ft+1 − d): recovery rates are low if the country
has low fundamentals or high debt levels. The parameter γ governs the sensitivity of
changes in the fractional recovery rate to changes in fundamentals. Intuitively, γ will
have a value that makes the recovery rate δt+1 vary between 0 and 1.21
If we assume that investors are risk neutral, we can now calculate the price and yield
spread of the bond. The yield spread on debt is defined as st = − log¡Bt
B
¢− rf , where
Bt is the price of the bond. In the appendix to this paper we show that the yield spread
21The need for this parameter is related to the fact that default depends only on the distance todefault. This means that if we scaled up fundamentals and volatility, distance to default would remainunchanged. However, if we did not adjust the recovery rate sensitivity γ appropriately, recovery valueswould change. We discuss this issue and how it influences model estimation in more detail in Section5.
24
is given by:
st = − logµγ exp
µft − d+
1
2σ2¶Φ
µ−ft − d
σ− σ
¶+ Φ
µft − d
σ
¶¶(5)
≡ g (ft, d, γ, σ) .
The first part of this expression is equal to the expected payoff to bondholders conditional
on default multiplied by the probability of default. The second part is the probability
of the country staying solvent in which case creditors receive face value. As we would
expect, the spread depends negatively on fundamentals and recovery rates, and positively
on debt levels and volatility.
One limitation of this result is that it assumes that investors are risk neutral, which
means that the correct discount rate for the bond is equal to the risk-free rate. In the
standard option pricing framework a measure of risk is not necessary since the assump-
tions of continuous trading and geometric Brownian motion allow for the construction
of a perfect hedge. The bond can then be priced using risk neutral probabilities. In
discrete time under lognormality, the no arbitrage condition implied by trading of the
underlying asset is sufficient to derive the price (Rubinstein (1976)). In our case, since
the underlying is not traded, we need a direct estimate of the riskiness of fundamentals.
Given an estimate of the covariance of the state variable with the stochastic discount
factor σfm, we can write down the spread on the bond for the case where investors are
risk averse:
st = − logµγ exp
µft − d+ σfm +
1
2σ2¶Φ
µ−ft − d+ σfm
σ− σ
¶+ Φ
µft − d+ σfm
σ
¶¶.
(6)
25
The common interpretation is that the covariance with the SDF changes the risk neutral
drift of fundamentals. The bond price is then given by the same formula, but with the
real default probabilities substituted by risk neutral default probabilities. We can see
that risk is priced in the way we would expect. If fundamentals tend to be low in bad
states of the world (from the marginal investor’s point of view), i.e. if the covariance with
the SDF is negative, the bond is more risky. Consequently, the risk neutral probability
of default is larger than the real probability of default and the spread is higher. In other
words, risk averse investors demand a higher spread if sovereign defaults tend to occur
in bad states of the world.
5 Model estimation using historical default data
In this section we take our model to the data and calculate model implied spreads.
We use historical data on defaults and explanatory variables to estimate model implied
default probabilities and parameters. The model specifies that the probability of a
country being in default in the next period is equal to the cumulative normal of a linear
combination of covariates today. Empirically we can estimate this conveniently in a
hazard model by running a probit on a set of explanatory variables.
We construct a measure of default which is equal to one in a given year if the country
is in default or goes into default over the course of the year. Specifically, we define a
default indicator variable Yi,t+1 that is equal to 1 if country i is in default in period
t+ 1, and 0 otherwise. We add to this a set of covariates. We use terms of trade as a
measure of fundamentals and repayment capabilities. Since we assume lognormality in
the model, we measure fundamentals as log terms of trade in period t relative to average
26
log terms of trade over the past ten years and denote it tti,t. Next we calculate the
volatility of log terms of trade σi,t over the past ten years. We use log debt to GDP dgi,t
to measure countries’ debt levels. To make sure that our analysis is not dominated by
extreme values, we winsorize the explanatory variables at the 5th and 95th percentile
levels.22 Since terms of trade and debt to GDP are not measured in the same units
we leave the coefficients on these variables as free parameters in the estimation. This
is in contrast to the firm case in which asset value and face value of debt are directly
comparable. We instead use model estimated parameters as measures of the relative
importance of the inputs in default prediction.
Country defaults are rare events. In order to estimate a hazard model, we need a
large enough sample with a sufficient number of defaults. We take advantage of the
fact that default data are available for a longer period than the spreads data used in the
previous section and extend the sample period to 1970. Although default data extend
back farther, terms of trade data are available from 1960. Since we are calculating
volatility over a ten year backward looking window, the sample with available covariates
begins in 1970.
Once a country enters default, economic activity does not cease and fundamentals
continue to be observable. The estimation sample therefore includes observations both
in and out of default. We include a dummy zi,t that is equal to one if the country is
currently in default because we expect that, for a given level of fundamentals, a country
is more likely to be in default next period if it is currently in default. Another reason
for including a default dummy is that it may pick up the effects of unobservable country
22This procedure is common when estimating hazard models of default and bankruptcy, e.g. Shumway(2001).
27
characteristics that are different in and out of default.23 We also include a constant term
to account for the fact that the relative magnitudes of our explanatory variables are not
directly comparable. Put differently, the fact that fundamentals lie below debt levels
does not necessarily trigger default. This is different from the case of corporate debt,
where default is conventionally triggered when asset value falls below the face value of
debt. Next, we divide through by volatility σi,t since in the model the difference between
fundamentals and debt levels is scaled by volatility. Finally, we add a country’s years
since last default denoted ytdi,t to capture the influence of recent default history on
the probability of default. The results from running a probit of default events on
explanatory variables are (z-statistics reported below the coefficients in parentheses):
P (Yi,t+1 = 1) = Φ ((−0.119 ∗ tti,t + 0.027 ∗ dgi,t + 0.119 ∗ zi,t − 0.163) /σi,t − 0.104 ∗ ytdi,t) .
(1.98) (1.67) (7.38) (2.43) (6.03)
All coefficients have the expected signs and are significant.
Before we can calculate model predicted spreads and compare those to actual spreads
we still need to scale the parameters. The reason is again related to the fact that there
are no natural units for the explanatory variables. Re-scaling both fundamentals and
the volatility of fundamentals does not affect the model implied probability of default.24
Measuring fundamentals in different units does, however, change recovery rates. We
23See, for example, De Paoli, Hoggarth, and Saporta (2006).24This feature is related to the fact that a standard probit estimation does not return an estimate of
the variance. Instead, the variance of the shock is assumed to be equal to 1.
28
can use this fact to scale the parameters. Recall that:
log (δit) = log γ + fi,t − di,t. (7)
This relation implies that recovery value moves one for one with fundamentals and debt
levels. We exploit this relation by regressing recovery values on fundamentals and
debt levels and imposing a coefficient of 1. We use spread observations in default to
calculate a measure of the market expected recovery rate. In particular, we assume that
the recovery rate is close to the current market price of the debt, i.e. δit+1 = 1
1+(sit+1+rf),
where we measure sit+1 using the country’s observed EMBI spread and rf denotes the
risk-free rate. We then re-scale the explanatory variables accordingly. From the
constant in the regression we also get an estimate of the recovery parameter γ̂. Recall
that st = g (ft, d, γ, σ) which means that we now have all the inputs necessary to calculate
model predicted spreads.25
5.1 Model implied spreads
Using the estimated model parameters we now calculate model implied spreads. Despite
a number of limitations we find that the model does surprisingly well. Figure 3 plots
actual spreads against model implied spreads. We calculate model implied spreads for
the EMBI regression sample of Tables 3 and 4. There is a strong positive relation
25Following the model in the previous section, we could, in principle, add more measures of funda-mentals and indebtedness to the analysis. Such an exercise involves estimating the relative importanceof different measures of fundamentals in predicting default. These estimates in turn affect the correctmeasure of volatility which means that the problem has to be estimated using maximum likelihood(instead of the probit setting we use here). In a previous version of the paper we discuss how toestimate such a model in more detail. One practical problem with this approach is that the maximumlikelihood surface is not concave.
29
between actual and model implied spreads which is reflected in a correlation of 71%. A
regression of actual on model implied spreads has an R2 of 51%.
In the next section we discuss the results from fitting our model to the data in more
detail. In that section, we use default probabilities estimated in a standard hazard
model to calculate reduced form spreads and compare reduced form and model implied
spreads. We also consider how well predicted spreads from both approaches can explain
variation in actual spreads.
We now turn briefly to some of the limitations we expect from this kind of an exercise.
A priori we would expect this type of model to have some difficulties matching observed
spreads closely for several reasons. First, we are only calculating the component of the
spread that is due to default risk. This means that, to the extent that there are other
spread components related to liquidity, risk aversion, and taxes (Elton, Gruber, Agrawal,
and Mann (2001), Hund and Lesmond (2007)), we expect observed spreads to be higher
on average than model implied spreads. Second, we know from the corporate bond
literature, that fitting structural form models to the data generally produces estimates
that range from less than a fraction of a basis point all the way up to unreasonably large
numbers. The reason is often that the implied default probabilities are extreme as well
(Eom, Helwege and Huang (2004), Huang and Huang (2003)). Third, unlike reduced
form models (including our analysis in Tables 3 and 4), our model estimation does not
use as inputs any of the observed spreads it is attempting to explain. This means that
there is no constraint that forces model implied and observed spreads to be broadly in
line.
Consistent with our expectations and findings in the corporate bond pricing litera-
ture, the model estimation gives some very low default probabilities. In the sample,
30
the 90th percentile of fitted distance to default is equal to 4, corresponding to a default
probability of 0.003% (the 10th percentile of the default probability distribution.) Such
low default probabilities imply low spreads and indeed the 10th percentile of the spread
distribution is equal to 0.05 bps. Since variation below 1 basis point is not really mean-
ingful, we set predicted spreads below 1 basis point equal to 1 basis point. In the figure
there is a corresponding local cluster of observations.26
We conclude that the model captures a surprisingly large share of the variation
in observed sovereign yield spreads. However, implicit nonlinearities in the assumed
functional form can lead to extreme values of predicted spreads, particularly for small
spreads.
6 Reduced form default probabilities and spreads
In this section we estimate default probabilities in a reduced form hazard model. This
effectively relaxes the functional form restriction imposed by the model in the previous
sections. We then use fitted default probabilities to calculate spreads. We compare
reduced form spreads to those predicted by our model of fundamentals and spreads.
We begin by constructing a set of explanatory variables. As in the previous spread
regression analysis we choose measures of debt levels, repayment capabilities, and volatil-
ity. We use the same set of variables as in Tables 3 and 4 measured over the extended
sample starting in 1970, again to increase power. We winsorize the data at the 5th
and 95th percentile to make sure that outliers do not drive our results. Conditional
26This feature is a result of the functional form assumed in the model. In particular, since thedifference between fundamentals and debt levels is scaled by volatility, the model can return extremevalues of distance to default if volatility is small.
31
on covariates, the sample has 587 country-year observations covering 31 countries from
1970-2002. This sample is much larger than the spread regression sample. However,
there are only 27 defaults which is less than 5% of the observations.
Table 5 presents summary statistics for our explanatory variables. We group obser-
vations into two groups: those where the country goes into default over the next year
and those observations for which it stays solvent. We calculate statistics for non-default
and default observations separately. For each of the seven explanatory variables the
mean and median of the default observations reflect noticeably less favorable conditions:
levels of debt to GDP are higher and levels of reserves are lower. Countries that are
about to default tend to have been exposed to negative terms of trade shocks and have
experienced low GDP growth. They tend to have high levels of volatility of terms of
trade and volatility of GDP growth and often have been in default more recently. The
levels of debt to GDP and volatility of terms of trade are particularly unfavorable in
years preceding default. For both measures, the mean and median of the default obser-
vation sample are about one standard deviation larger than the mean/median for the
non-default sample.
To investigate the predictive power of this set of explanatory variables more formally,
we run logit regressions of a forward looking default indicator on explanatory variables.
In particular, we estimate:
Pt (Yi,t+1 = 1) =1
1 + exp (−α− β0xi,t). (8)
The default indicator for a country is set equal to one in a given year if the country goes
into default over the course of the following year. It is set equal to zero if the country
32
is not in default the following year.
Estimation results for different specifications are reported in Table 6. We first
consider a specification that includes debt to GDP, volatility of terms of trade, and the
change in the terms of trade. We find that debt to GDP and volatility of terms of trade
are highly significant. This is consistent with the suggestive evidence from comparing
means in the summary statistics. The specification in column (1) explains variation
in default well and has a pseudo R2 of 17.3%. We also consider results from two
other specifications, one that adds the remaining variables measuring macroeconomic
fundamentals and one that also adds years since last default to measure a country’s
recent default history. Model fit improves to pseudo R2 of 20.3% and 21.4% respectively.
Debt to GDP and terms of trade volatility are significant at the 1% level in all three
specifications. However, none of the other explanatory variables are significant at the 5%
level although, with the exception of the change in terms of trade, they have the expected
signs. Reserves, GDP growth, and years since last default are significant at the 10%
level.27 We refer to the model in column (3) as our main specification. It includes the
measures of fundamentals and default history and corresponds to the main specification
in Table 3.28 We also consider economic significance of debt to GDP and volatility of
terms of trade for this specification. Both are highly economically significant. The
predicted probability of default when all explanatory variables are equal to their mean
values is equal to 4.6%. Increasing debt to GDP by one standard deviation increases the
27These results are consistent with independent work by Catão and Sutton (2002) and Catão andKapur (2006) who also investigate empirical determinants of default in a reduced form hazard modelsetting that is similar to the specification reported in Table 6.28In principle, we could consider additional explanatory variables that predict default in the sample
we consider. However, because of the sample size and the rather small number of default observations,our estimation only has a limited amount of power. For our investigation we therefore choose to usethe same set of variables as in the spread regression specifications.
33
predicted default probability to 8.4%; for a one standard deviation increase in volatility
of terms of trade the probability increases to 9.3%.
We report model forecast accuracy in Table 5 Panel B. We divide fitted default
probabilities into quintiles for the different specifications. We then count the number
of defaults in the different quintiles. A perfect model would sort all 27 defaults into the
highest quintile. If the model had no power to predict default each bin would have 5.4
defaults. We find that our best model sorts 19 of 27, or 70% of defaults, into the top
quintile, with 24 of 27 defaults in the top two quintiles. There is also an increase in
forecast accuracy across specifications which corresponds to the increase in pseudo R2.
In Panel C we report summary statistics of fitted default probabilities. Fitted prob-
abilities are comparable across the specifications. The correlations of fitted default
probabilities of specifications (1) and (2) with the best model are 90% and 96% respec-
tively. In contrast to our structural model in the previous section, the reduced form
hazard model does not return extremely low default probabilities. It does, however,
predict some very high default probabilities.29 Overall, fitted probabilities are rather
stable and similar across different specifications.
6.1 Comparing reduced form and model implied spreads
We next calculate reduced form predicted spreads using fitted probabilities from the
hazard model. We assume that recovery rates are fixed, i.e. that creditors receive
a constant fraction of face value in the event of default. We use fitted probabilities
from our main specification in column (3). We set the fractional recovery rate equal to
29The two largest fitted values are for Côte d’Ivoire in 1982 and 1999, equal to 70.2% and 76.8%respectively. In both cases default occurred over the next year.
34
0.6, which is broadly in line with the average historical experience. We then calculate
reduced form predicted spreads and use them to explain variation in actual spreads.
This means that variation in fundamentals can affect spreads only to the extent that
fundamentals affect default probabilities. Also, predicted spreads are calculated without
using actual spreads as an input. This is in contrast to the earlier methodology of
regressing spreads on explanatory variables directly.
Figure 4 plots actual against reduced form spreads. Variation in predicted spreads
matches variation in actual spreads very well. We notice a strong positive relation which
is reflected in a correlation of 78%. We can compare explanatory power of reduced form
and model implied spreads by comparing Figures 3 and 4; both plot actual against
predicted spreads for the same sample. As discussed, model implied spreads are often
extremely small. We notice that reduced form predicted spreads do not have the same
characteristic. The minimum reduced form spread is equal to 2.7 basis points, the
10th percentile is 16 basis points. These relatively larger reduced form spreads are a
result of larger predicted default probabilities. The 10th percentile reduced form default
probability is 0.4% compared to the 0.003% model implied default probability. Table
7 Panel A reports summary statistics for model implied and reduced form spreads and
default probabilities together with actual EMBI spreads. Predicted spreads are smaller
than actual spreads: median reduced form and model implied spreads are 26 and 74 basis
points compared to a 411 basis points median EMBI spread. This difference reflects
other non-default spread components and is consistent with findings in the corporate
bond literature that default risk accounts for only part of the spread (Huang and Huang
(2003) and others).
We now ask how much of the variation in spreads is explained by variation in default
35
risk. We regress actual on model implied and predicted spreads. Predicted spreads
have a large share of extreme values. This is especially true for the reduced form spreads
which reach a maximum of 4653 basis points and have an in-sample kurtosis of 67. The
large predicted spread is less surprising when considering the maximum predicted default
probability of 76.8%. If we included these large values directly into the regression they
would dominate the results. We therefore winsorize model implied and reduced form
spreads at the 10th and 90th percentiles. Model implied spreads then run from 1 to
467 basis points compared to 16 to 460 basis points for reduced from spreads. The
in-sample standard deviations are reduced to 157 basis points and 144 basis points for
model implied and reduced form spreads respectively.
Results for spread regressions are reported in Table 7 Panel B. We find that our
model explains 53% of the variation in observed spreads and that a 1 basis point increase
in model implied spreads is associated with a 1.66 basis point increase in observed
spreads. Reduced form spreads can explain 55% of the actual spread variation. The
coefficient on predicted spreads is 1.85. These R2 are similar to those reported in
Section 3. This high level of explanatory power is rather surprising, keeping in mind
that our model estimation does not use as inputs any of the observed spreads it is
attempting to explain. For comparison, our main specification in Table 3 explained
62% of the variation. The large coefficients are not surprising once we consider the data
summary statistics: predicted spreads are smaller and less variable than actual spreads.
In contrast to predicted spreads, we do not winsorize actual spreads to make results
comparable to those in Section 3. This leads to a variance reduction of the winsorized
spread data which helps explains the coefficient size. We also note that the relative
magnitude of the coefficients is consistent with the relative size of the spread standard
36
deviations.
We next consider a different approach to deal with outliers which is to run regres-
sions of log actual spreads on log predicted spreads using the unwinsorized data. This
effectively compares percentage changes of actual and predicted spreads. It also min-
imizes the risk of large outliers affecting the regression. Running log-log regressions
corresponds to the regression lines of Figures 3 and 4 which use log scales. The regres-
sion of actual on model implied spreads has a coefficient of 0.27 and an R2 of 51%. For
reduced form spreads the coefficient is 0.51 and the regression can explain 61% of the
variation. This means that a 1% relative increase in reduced form spreads implies a
0.5% relative increase in actual spreads. The larger coefficient on reduced form spreads
is again consistent with the smaller variance of log reduced form spreads. This smaller
variance is also evident from Figures 3 and 4.
Comparing means of observed and model implied spreads, we already know that a
large component of the spread is not explained by our model. In order to explore if there
is variation in the explained fraction for different levels of credit risk, we group spreads
by their estimated default probability and compare average observed and fitted spreads
across groups. We choose break points using the distribution of fitted probabilities from
the entire estimation sample (1970-2002). Figure 5 plots average realized and implied
spreads by bins. We find that fitted spreads on good credit are much smaller than
observed spreads while for poor credit, the overall size of fitted spreads is a much larger
fraction of observed spreads.
These findings are consistent with patterns in the corporate bond market. Huang
and Huang (2003) show that the fraction of observed spreads accounted for by structural
form models decreases with credit quality. In other words, model implied spreads are
37
far too low on companies of good credit quality but they are closer to realized spreads
for companies of poor credit quality. The fact that credit risk only accounts for a
small fraction of the average spread in the sovereign debt market is thus consistent with
what Huang and Huang find for corporate bonds. In order to more accurately price
sovereign debt, it will be important to increase our understanding of the other spread
components.
To summarize, we find that both model implied and reduced form spreads have sig-
nificant explanatory power for observed sovereign spreads. We find that the default
component of spreads explains variation in observed spreads surprisingly well. Elimi-
nating the specific assumption of functional form in our structural model we find that
reduced form spreads do even better in explaining observed spread variation. We also
find that our model fits better for borrowers of lower credit quality.
7 Conclusion
This paper makes three contributions to the literature on sovereign risk and sovereign
debt pricing. First, we provide evidence that variation in country fundamentals explains
a large share of the variation in emerging markets’ sovereign debt prices. We consider
a set of 32 emerging market economies over the period 1994-2006 and use J.P. Morgan’s
Emerging Markets Bond Index (EMBI) to measure a country’s spread on its external
debt.
To explain spread variation we focus on measures of liabilities and repayment ca-
pabilities. In a departure from the existing empirical literature on sovereign debt we
pay special attention to the volatility of fundamentals. We demonstrate that, consis-
38
tent with standard bond and option pricing theory, the volatility of fundamentals is
an important determinant of a country’s spread on its external debt obligations. In
particular the volatility of terms of trade is both statistically and economically signifi-
cant in explaining spread variation. In a reduced form panel regression framework, a
one standard deviation increase in the volatility of terms of trade is associated with an
increase of 150 basis points in spreads, which corresponds to nearly half of the standard
deviation of observed spreads. The debt to GDP ratio has low explanatory power for
cross-country variation in spreads. Instead it adds substantial explanatory power in the
time-series. In other words, what matters is whether a country’s debt to GDP is high
relative to its own mean, not whether it is high relative to other countries. In addition,
even after controlling for fundamentals, a country’s default history has a significant im-
pact on spreads. Overall, we obtain a good fit, both across countries and over time,
with just a few variables.
Second, we propose a model of sovereign spreads to formalize the insight that a coun-
try’s spread is higher if its fundamentals are lower or if the volatility of fundamentals is
higher. Using historical data on actual country defaults we calculate predicted spreads.
Our model can explain half of the variation in observed spreads. This level of explana-
tory power is surprisingly large given the model’s specific assumption about functional
form and the fact that our model estimation only uses data on defaults and recovery
values to calculate predicted spreads.
Third, we demonstrate empirically the explicit link between sovereign debt prices
and default probabilities. We estimate a reduced form hazard model using historical
data on measures of fundamentals and default. Again, we find that the volatility of
terms of trade is an important explanatory variable of country default. We calculate
39
reduced form spreads using only fitted probabilities of default and find that predicted
spreads explain a large share of the variation in sovereign debt prices. In the literature
on corporate debt and default, the link between default probabilities and corporate CDS
spreads has been pointed out by Berndt , Douglas, Duffie, Ferguson, and Schranz (2005)
but this point has not, to the best of our knowledge, been made in the literature on
sovereign debt pricing.30
The findings in this paper have important implications for investors and policymak-
ers. We have identified factors explaining variation in spreads both across countries
and over time. Countries with larger fluctuations in their terms of trade, possibly due
to concentrated exports in commodities with volatile prices, are more susceptible to
external shocks. They may find it beneficial to address such risk factors in order to
reduce borrowing costs as well as the probability of a crisis or default. One possiblity
of course is to reduce debt levels or increase holdings of reserves. The recent run up
in commodity prices has provided an opportunity for some countries to move to a more
favorable fiscal position and to engage in large purchases of foreign liquid assets such
as U.S. debt obligations. Over the same period, debt markets have witnessed a large
run up in prices and an associated reduction in spreads. Our results suggest a further
potential avenue for risk reduction: countries could try to explicitly hedge exposures
to macroeconomic shocks in international financial markets. Caballero (2003), Shiller
(2003), and Merton (2005) advocate the use of innovative financial contracts to share
these macroeconomic risks more effectively.
30Cantor and Packer (1996) explore the link between debt prices and sovereign credit ratings.
40
A Appendix: Data sources and construction of vari-
ables
J. P. Morgan’s EMBI index includes U.S. dollar denominated debt instruments issued
by emerging market governments. It includes Brady bonds, loans, and Eurobonds. In
order for a country to be eligible for inclusion in the index, it needs to be classified as
low or middle income by the World Bank for the last two years, or have restructured
external or local debt in the past 10 years, or have restructured external or local debt
outstanding. For a debt instrument to be eligible, it needs to be issued by a sovereign
or quasi-sovereign entity (100% owned/guaranteed by the government), its issue size
needs to be greater than or equal to U.S.$500 million, its remaining maturity needs to
be greater than 2.5 years, it needs to have verifiable daily prices and cash flows, and it
has to fall under G7 legal jurisdiction.31
In our empirical analysis we calculate the annual mean of the daily EMBI spread
series. Since not all series begin in January we also include observations for which the
mean is constructed from less than a full year of spread data. However, we require at
least a full month of data (more than 20 trading days) for the observation not to be
dropped.
The CRB commodity index series is from Global Financial Data (GFD). It is con-
structed by the Commodity Research Bureau. The series is a continuation of the older
BLS index which was discontinued in 1981. According to the CRB, the index includes
energy 18% (crude oil, heating oil, natural gas), grains and oilseed 18% (corn, soybeans,
wheat), industrials 11% (copper, cotton), livestock 11% (live cattle, live hogs), precious
31Source: Emerging Markets External Debt Handbook for 2002-2003, J.P. Morgan.
41
metals 18% (gold, platinum, silver), softs 24% (cocoa, coffee, orange juice, sugar).32
We use annual terms of trade data provided by the Development Data Group at the
World Bank.33 Terms of trade indices are constructed as the ratio of an export price
index to an import price index. The underlying price and volume indices were compiled
by the United Nations Conference on Trade and Development (UNCTAD).
When constructing volatility of terms of trade we do not always have enough data to
calculate it over a ten-year window. For some countries, e.g. Croatia, Poland, Ukraine
and Russia, data on terms of trade data start later in the sample. To make sure this does
not significantly reduce the amount of observations available for the analysis, we reduce
the window size to calculate volatility down to a minimum window of six years when
necessary. In the case of the Dominican Republic there is insufficient terms of trade
data to construct a measure of volatility, even when reducing the calculation window
(as discussed in Section 2). Spread observations for the Dominican Republic therefore
does not enter the spread regression sample of Section 3.
Our measure of government external debt is the Total Debt series from the World
Bank Global Development Finance (GDF) data set. GDP data are from the Interna-
tional Monetary Fund World Economic Outlook (WEO).34 When GDP growth data
is not available we follow the same procedure as in the case of volatility of terms of
trade: we calculate the standard deviation using a shorter window but again (as when
32Source: documentation describing CRB Commodity Index in GFD33In an earlier version of this paper we used terms of trade data from 1960 to 2002. In these data, all
series are normalized to 100 for 1995. Since we calculate the volatility of the percent change in termsof trade this normalization has no effect on the values of our variables. We have subsequently addedthe percent change in the terms of trade 2003 using an updated series.34We carefully examine the GDP data series in order to eliminate any obvious data errors. Since
WEO data is only available starting in 1970, we use data from WDI for 1960-1969 to calculate volatilityof GDP growth.
42
calculating volatility of terms of trade) require a minimum of six years of observations.
Reserves refers to the Total Reserves Including Gold series from the World Development
Indicators (WDI).
The 10-year U.S. Treasury yield, 6-month Treasury rate, and 3-month Eurodollar
rate are from the Federal Reserve Bank of St. Louis. Yields on Baa U.S. corporate
bonds are from Amit Goyal’s website. The VIX measure of volatility of the S&P500
index is provided by the Chicago Board Options Exchange.
43
B Appendix: Derivation of the model implied spread
The price of the bond in period t depends on the expected payoff to the bond and the
appropriate risk adjusted discount rate rt:
Bt = exp (−rt)Et [Bt+1]
= exp (−rt)B [Pt (Ft+1 > D) + Pt (Ft+1 < D)Et [δt+1|Ft+1 < D]]
= exp (−rt)B∙Pt (Ft+1 > D) + Pt (Ft+1 < D)Et
∙γFt+1
D|Ft+1 < D
¸¸,
since we assume that the recovery rate is given by: δt+1 = γ Ft+1D.
We assume that Ft+1 is conditionally lognormally distributed:
ft+1 = ft + φt+1
φt+1 ∼ N¡0, σ2
¢.
The conditional probability of the country defaulting next period is given by Pt (Ft+1 < D) =
Pt (ft+1 < d) , where d = log (D). Given our distributional assumption, the probability
of default is equal to:
Pt (ft+1 < d) = Φ
µd− ftσ
¶,
where Φ (.) denotes the c.d.f. of the normal distribution.
If investors are risk neutral and the risk-free rate is constant, the price of the bond
is given by:
Bt = exp (−rf)B∙Pt (Ft+1 > D) + Pt (Ft+1 < D)Et
∙γFt+1
D|Ft+1 < D
¸¸.
44
Exploiting the assumption that Ft+1 is lognormally distributed and making a change of
variable, we obtain the following expression for the price of the bond after some tedious
but straightforward calculations:
Bt = exp (−rf)B∙γ exp
µ−d+ ft +
1
2σ2¶Φ
µd− ftσ− σ
¶+ Φ
µ−d+ ft
σ
¶¸.
Bond prices are commonly quoted in yields and yield spreads. With a constant interest
rate, the bond’s yield and yield spread are given by
θt = − logµBt
B
¶, st = − log
µBt
B
¶− rf ,
where θt is the yield on the bond, and st is the spread.
With risk neutral investors, the spread is thus given by:
st = − logµγ exp
µ−d+ ft +
1
2σ2¶Φ
µd− ftσ− σ
¶+ Φ
µ−d+ ft
σ
¶¶st = g (ft, d, γ, σ) .
If investors are risk averse, we can rewrite the bond price as:
Bt = Et
∙Mt+1
½Pt (Ft+1 > D)B + Pt (Ft+1 < D)Et
∙Ft+1
DγB|Ft+1 < D
¸¾¸,
where Mt+1 is the stochastic discount factor. Then we can risk neutralize to find the
45
price:
Bt = BEt [exp (mt+1)min (exp (0) , exp (ft+1 − d) γ)]
Bt = exp (−rf)EQt [min (exp (0) , exp (ft+1 − d) γ)]
ft+1 ∼ N¡ft + σfm, σ
2¢, under Q.
With risk averse investors, the spread is given by:
st = − log
⎛⎜⎝ γ exp¡−d+ ft + σfm +
12σ2¢Φ¡d−ft−σ
σ− σ
¢+ Φ
¡−d+ft+σσ
¢⎞⎟⎠
st = g (ft; d, γ, σ, σfm) .
where σfm denotes the covariance of the willingness to pay index with the stochastic
discount factor.
46
References
Anderson, Ronald, and Suresh Sundaresan, 1996, “Design and Valuation of Debt Con-
tracts,” Review of Financial Studies 9, 37-68.
Anderson, Ronald, and Suresh Sundaresan, 2000, “A Comparative Study of Struc-
tural Models of Corporate Bond Yields: An Exploratory Investigation,” Journal
of Banking and Finance 24, no. 12, 255-269.
Beck, Roland, 2001, “Do Country Fundamentals Explain EmergingMarket Bond Spreads?”
CFSWorking Paper, No. 2001/02, Frankfurt, Germany: JohannWolfgang Goethe-
Universität.
Beim, David O., and CharlesW. Calomiris, 2001, Emerging Financial Markets, McGraw-
Hill, New York, NY.
Berg, Andrew, Eduardo Borensztein, and Catherine Pattillo, 2004, “Assessing Early
Warning Systems: How Have They Worked in Practice?” IMF Working Paper No.
04/52, Washington, DC: International Monetary Fund.
Berndt, Antje, Rohan Douglas, Darrell Duffie, Mark Ferguson, and David Schranz,
2005, “Measuring Default-Risk Premia from Default Swap Rates and EDFs,” un-
published paper, Tepper School of Business, Carnegie Mellon University and Grad-
uate School of Business, Stanford University.
Black, Fischer and John C. Cox, 1976, “Valuing Corporate Securities: Some Effects of
Bond Indenture Provisions,” Journal of Finance 31, 351-367.
Black, Fischer and Myron Scholes, 1973, “The Pricing of Options and Corporate Lia-
bilities,” Journal of Political Economy 81, 637-654.
47
Bulow, Jeremy, and Kenneth Rogoff, 1989, “A Constant Recontracting Model of Sov-
ereign Debt,” Journal of Political Economy 97, 155-178.
Bulow, Jeremy, and Kenneth Rogoff, 1989, “Sovereign Debt: Is to Forgive to Forget?”
American Economic Review 79, 43-50.
Caballero, Ricardo J., 2003, “The Future of the IMF,” American Economic Review
Papers and Proceedings 93, 31-38.
Campbell, John Y., and Glen B. Taksler, 2003, “Equity Volatility and Corporate Bond
Yields,” Journal of Finance 58, 2321-2349.
Cantor, Richard, and Frank Packer, 1996, “Determinants and Impact of Sovereign
Credit Ratings,” Federal Reserve Bank of New York Economic Policy Review,
October, 37-54.
Carr, Peter, and Liuren Wu, 2006, “Theory and Evidence on the Dynamic Interactions
Between Sovereign Credit Default Swaps and Currency Options,” forthcoming,
Journal of Banking and Finance.
Catão, Luis, and Sandeep Kapur, 2006, “Volatility and the Debt-Intolerance Paradox,”
IMF Staff Papers 53, 195-218.
Catão, Luis, and Bennett Sutton, 2002, “Sovereign Defaults: The Role of Volatility”,
IMF Working Paper No. 02/149, Washington, DC: International Monetary Fund.
Chen, Long, David A. Lesmond, and Jason Wei, 2007, “Corporate Yield Spreads and
Bond Liquidity,” Journal of Finance 62, 119-149.
Collin-Dufresne, Pierre and Robert S. Goldstein, 2001, “Do Credit Spreads Reflect
Stationary Leverage Ratios?” Journal of Finance 56, 1929-1957.
48
De Paoli, Bianca, Glenn Hoggarth, and Victoria Saporta, 2006, “Costs of Sovereign
Default,” Bank of England Financial Stability Paper no.1.
Diaz Weigel, Diana and Gordon Gemmill, 2006, “What Drives Credit Risk in Emerg-
ing Markets? The Roles of Country Fundamentals and Market Co-Movements,”
Journal of International Money and Finance 25, 476-502.
Dooley, Michael, and Mark R. Stone, 1993, “Endogenous Creditor Seniority and Ex-
ternal Debt Values,” IMF Staff Papers 40, 395-413.
Duffee, Gregory R., 1998, “The Relation Between Treasury Yields and Corporate Bond
Yield Spreads,” Journal of Finance 53, 2225-2241.
Duffie, Darrell, Lasse Heje Pedersen, and Kenneth J. Singleton, 2003, “Modeling Sov-
ereign Yield Spreads: A Case Study of Russian Debt,” Journal of Finance 58,
119-159.
Duffie, Darrell, and Kenneth J. Singleton, 1999, “Modeling Term Structures of Default-
able Bonds,” Review of Financial Studies 12, 687-720.
Duffie, Darrell and Kenneth J. Singleton, 2003, Credit Risk: Pricing Measurement, and
Management (Princeton University Press, Princeton, NJ).
Duffie, Darrell, Leandro Saita, and Ke Wang, 2007, “Multi-Period Corporate Failure
Prediction with Stochastic Covariates,” Journal of Financial Economics 83, 635—
665.
Edwards, Sebastian, 1984, “LDC Foreign Borrowing and Default Risk: An Empirical
Investigation, 1976-80,” American Economic Review 74, 726-734.
49
Edwards, Sebastian, 1986, “The Pricing of Bonds and Bank Loans in International
Markets, An Empirical Analysis of Developing Countries’ Foreign Borrowing,”
European Economic Review 30, 565-589.
Eichengreen, Barry, and Ashoka Mody, 1998, “What Explains Changing Spreads on
Emerging-Market Debt: Fundamentals or Market Sentiment?” NBER Working
Paper Number 6408.
Elton, Edwin J., Martin J. Gruber, Deepak Agrawal, and Christopher Mann, 2001,
“Explaining the Rate Spread on Corporate Bonds,” Journal of Finance 56, 247-
277.
Emerging Market Bond Index, (J.P. Morgan, New York, NY).
Emerging Markets External Debt Handbook for 2002-2003, New York, NY: J.P. Morgan.
Eom, Young Ho, Jean Helwege, and Jing-Zhi Huang, 2004, “Structural models of cor-
porate bond pricing: an empirical investigation,” Review of Financial Studies 17,
499-544.
Ferrucci, Gianluigi, 2003, “Empirical Determinants of Emerging Market Economies’
Sovereign Bond Spreads,” Bank of England Working Paper no. 205.
Gibson, Rajna, and Suresh M. Sundaresan, “A Model of Sovereign Borrowing and
Sovereign Yield Spreads,” unpublished paper, Universität Zürich, Columbia Uni-
versity.
Goldstein, Morris, Graciela L. Kaminsky, and Carmen M. Reinhart, 2000, Assessing Fi-
nancial Vulnerability, An Early Warning System for Emerging Markets (Institute
for International Economics, Washington, DC).
50
Hajivassiliou, Vassilis A., 1987, “The External Debt Repayments Problems of LDC’s,
An Econometric Model Based on Panel Data,” Journal of Econometrics 36, 205-
230.
Hajivassiliou, Vassilis A., 1994, “A Simulation Estimation Analysis of the External Debt
Crises of Developing Countries,” Journal of Applied Econometrics 9, 109-131.
Heston, Alan, Robert Summers and Bettina Aten, Penn World Table Version 6.1,
Center for International Comparisons at the University of Pennsylvania (CICUP),
October 2002.
Huang, Jing-Zhi and Ming Huang, 2003, “How Much of the Corporate-Treasury Yield
Spread is Due to Credit Risk?” unpublished paper, Smeal College of Business,
Penn State University and Graduate School of Business, Stanford University.
Hund, John and David A. Lesmond, 2007, “Liquidity and Credit Risk in Emerging Debt
Markets,” unpublished paper, University of Texas - Austin and Tulane University.
Global Development Finance, Washington, DC: World Bank, various issues.
Jostova, Gergana, 2006, “Predictability in Emerging Sovereign Debt Markets,” Journal
of Business 79, 527-565.
Kamin, Steven B., and Karsten von Kleist, 1999, “The Evolution and Determinants of
Emerging Market Credit Spreads in the 1990’s,” Board of Governors of the Federal
Reserve System Internatinal Finance Discussion Papers, Number 653.
Kaminsky, Graciela L., and Carmen M. Reinhart, 1999, “The Twin Crises: The Causes
of Banking and Balance-of-Payments Problems,” American Economic Review 89,
473-500.
51
Klingen, Christoph, Beatrice Weder and Jeromin Zettelmeyer, 2004, “How Private
Creditors Fared in Emerging Debt Markets, 1970-2000,” IMF Working Paper No.
04/13, Washington, DC: International Monetary Fund.
Leland, Hayne E., 1994, “Corporate Debt Value, Bond Covenants, and Optimal Capital
Structure,” Journal of Finance 49, 1213-1251.
Longstaff, Francis A., Sanjay Mithal, and Eric Neis, 2005, “Corporate Yield Spreads:
Default Risk or Liquidity? New Evidence from the Credit Default Swap Market,”
Journal of Finance 60, 2213-2253.
Longstaff, Francis A., Jun Pan, Lasse H. Pedersen, and Kenneth J. Singleton, 2007,
“How Sovereign is Sovereign Credit Risk?” unpublished paper, UCLA Anderson
School, MIT Sloan School, NYU Stern School, and Stanford Graduate School of
Business.
Longstaff, Francis A., and Eduardo S. Schwartz, 1995, “A Simple Approach to Valuing
Risky Fixed and Floating Rate Debt,” Journal of Finance 50, 789-819.
Mella-Barral, Pierre, and William Perraudin, 1997, “Strategic Debt Service,” Journal
of Finance 52, 531-556.
Mendoza, Enrique G., 1995, “The Terms of Trade, The Real Exchange Rate and Eco-
nomic Fluctuations,” International Economic Review 36, 101-137.
Mendoza, Enrique G., 1997, “Terms of Trade Uncertainty and Economic Growth,”
Journal of Development Economics 54, 323-356.
Merton, Robert C., 1974, “On the Pricing of Corporate Debt: The Risk Structure of
Interest Rates,” Journal of Finance 29, 449-470.
52
Merton, Robert C., 2005, “Swapping Your Country’s Risks,” Harvard Business Review
83, no. 2.
Min, Hong G., 1998, “Determinants of Emerging Market Bond Spreads: Do Economic
Fundamentals Matter?” World Bank Policy Research Paper, Number 1899.
Obstfeld, Maurice, and Kenneth Rogoff, 2000, “New Directions for Stochastic Open
Economy Models,” Journal of International Economics 50, 117-153.
Pan, Jun, and Kenneth J. Singleton, 2008, “Default and Recovery Implicit in the Term
Structure of Sovereign CDS Spreads,” forthcoming, Journal of Finance.
Reinhart, Carmen M., Kenneth S. Rogoff and Miguel A. Savastano, 2003, “Debt In-
tolerance,” in William Brainard and George Perry (eds.), Brookings Papers on
Economic Activity 1, 1-74.
Rubinstein, M., 1976, “The Valuation of Uncertain Income Streams and the Pricing of
Options,” Bell Journal of Economics and Management Science 7, 407-25.
Shiller, Robert J., 2003, The New Financial Order: Risks in the 21st Century (Prince-
ton University Press, Princeton, NJ).
Shumway, Tyler, 2001, “Forecasting Bankruptcy More Accurately: A Simple Hazard
Model,” Journal of Business 74, 101-124.
United Nations Conference on Trade and Development, New York, NY: United Nations,
various issues.
Uribe, Martin, and Vivian Z. Yue, 2006, “Country spreads and emerging countries:
who drives whom?” Journal of International Economics 69, 6-36.
53
Vassalou, Maria and Yuhang Xing, 2004, “Default Risk in Equity Returns,” Journal of
Finance 59, 831-868.
Westphalen, Michael, 2001, “Determinants of Sovereign Bond Credit Spreads Changes,”
unpublished paper, Universite de Lausanne.
World Economic Outlook. Washington, DC: International Monetary Fund, September
2003.
54
Latin America Argentina ARG 1994 5441 39.1 9.4 2 Crude petroleum (10%), Feeding stuff for animals (10%)Brazil BRA 1994 836 34.9 10.1 1 Aircraft (6%), Iron ore and concentrates (5%)Chile CHI 1999 126 47.9 9.4 1 Copper (27%), Base metals ores (13%)Colombia COL 1997 506 33.9 10.2 0 Crude petroleum (26%), Coffee and substitutes (8%)Dominican Rep. DMR 2001 690 29.9 . 1 Men's outwear non-knit (17%), Under garments knitted (13%)Ecuador EQU 1995 1188 70.0 13.3 2 Crude petroleum (41%), Fruit, nuts, fresh, dried (18%)El Salvador ESD 2002 338 29.5 13.8 0 Coffee and substitutes (16%), Paper and paperboard, cut (6%)Mexico MEX 1994 246 33.0 9.0 1 Passenger motor vehicles, exc. bus (10%), Crude petroleum (8%)Panama PAN 1996 367 78.2 2.3 1 Fish, fresh, chilled, frozen (20%), Fruit, nuts, fresh, dried (20%)Peru PER 1997 428 53.9 14.6 4 Gold, non-monetary (17%), Feeding stuff for animals (13%)Uruguay URU 2001 902 31.6 8.9 4 Meat, fresh, chilled, frozen (16%), Leather (10%)Venezuela VEN 1994 1004 41.4 21.6 3 Crude petroleum (59%), Petroleum products, refined (25%)
Africa Côte d'Ivoire CDI 1998 2748 115.2 20.0 2 Cocoa (28%), Petroleum products, refined (18%)Morocco MOR 1998 273 58.5 7.3 2 Women's outwear non-knit (11%), Men's outwear non-knit (8%)Nigeria NIG 1994 1004 67.2 26.8 1 Crude petroleum (99.6%)South Africa SAF 1995 175 18.2 6.9 3 Pearl, prec., semi-prec. stones (13%), Special transactions (12%)Tunisia TUN 2002 189 54.6 5.6 0 Men's outwear non-knit (17%), Women's outwear non-knit (10%)
Eastern Europe Bulgaria BUL 1994 228 75.7 7.8 1 Petroleum products, refined (11%), Copper (7%)Croatia CRO 1996 118 49.1 1.6 1 Ships, boats etc. (15%), Petroleum products, refined (8%)Hungary HUN 1999 31 58.4 3.0 0 Int. combust. piston engines (9%), Automatic data proc. eq. (7%)Poland PLD 1994 98 43.2 5.0 1 Furniture and parts thereof (7%), Ships, boats etc. (4%)Russia RUS 1998 315 47.2 13.7 1 Crude petroleum (24%), Gas, natural and manufactured (17%)Ukraine UKR 2000 359 27.0 12.9 1 Iron, steel primary forms (13%), Iron, steel shapes etc. (8%)
Southeast Asia China CHN 1994 57 13.9 2.4 0 Telecom equip., parts, acces. (5%), Automatic data proc. eq. (5%)Malaysia MAL 1996 151 39.2 7.2 0 Transistors, valves etc. (19%), Office, adp. mach. parts (12%)Philippines PHI 1998 453 64.0 7.8 1 Transistors, valves etc. (42%), Automatic data proc. eq. (13%)South Korea SKO 1994 106 34.6 4.6 0 Transistors, valves etc. (12%), Passgr. motor vehicl. exc. bus (7%)Thailand THA 1997 91 35.5 7.9 0 Office, adp. mach. parts (9%), Transistors, valves etc. (8%) Egypt EGY 2001 204 48.9 10.6 1 Petroleum products, refined (39%), Crude petroleum (10%)Lebanon LEB 1998 511 25.9 4.0 0 Gold, silver ware, jewelry (9%), Gold, non-monetary (7%)Pakistan PAK 2001 317 44.3 12.3 1 Textile articles (16%), Textile yarn (12%)Turkey TUR 1996 627 34.1 5.4 2 Outer garments knit non-elastic (6%), Under garments knitted (6%)
Top two exports (SITC group) and their percentage of total exports
EMBI refers to the median daily spread in 2003 of J.P. Morgan's Emerging Market Bond Index Global (EMBI), measured in basis points over U.S. Treasuries. We report the first year for which EMBI data is available. Debt/GDP is the median level of external debt divided by GDP. Volatility of terms of trade is the median standard deviation of the percent change in terms of trade and is calculated using annual data and a rolling ten year backward-looking window. Number of defaults is the number of times a country has gone into default between 1970 and 2003. This is also the sample for the macroeconomic explanatory variables. Export groups are for 2000-2001 (UNCTAD).
Table1: Summary statistics by country
Middle East & South Asia
Debt/ GDP
Volatility of terms of trade
Number of
defaults
Country Country code
EMBI spread
Start of EMBI series
Mean 488.9 48.9 0.5 5.7 16.7 5.3 14.4 7.3Median 411.0 47.2 -1.1 4.7 14.2 4.7 13.8 9St. Dev. 360.2 20.9 13.0 4.3 11.8 14.5 7.4 3.5
Min 30.6 11.8 -33.5 0.7 1.5 -33.9 2.0 1p5 87.5 16.7 -15.2 1.3 5.0 -20.9 5.0 1p95 1268.5 81.6 21.3 14.8 37.2 31.1 28.5 11Max 1679.1 104.7 78.8 21.8 82.4 65.4 44.5 11
Observations: 177
EMBI refers to spreads on J.P. Morgan's Emerging Market Bond Index, debt/GDP denotes US dollar denominated debt to GDP, change in terms of trade is the percent change in terms of trade over the past five years. Volatility of terms of trade is the standard deviation of the annual percent change in terms of trade. It is calculated using annual data from a ten year rolling window. Years since last default counts the number of years since the last year in which the country was in default. It is capped at 10, equal to zero if the country is in default and set equal to 11 if the country has never defaulted. It is constructed using a default indicator variable from Reinhart, Rogoff, and Savastano (2003). The sample corresponds to the regression sample used in Table 3 and Table 4. p5 and p95 refer to the 5th and 95th percentiles of the distribution respectively.
Years since last default
Table 2: Summary statistics for EMBI spread regression sample
GDP growth Volatility of GDP growth
EMBIspread
Debt/GDP
Change in terms of trade
Volatility of terms of trade
Reserves/GDP
(1) (2) (3) (4) (5)5.659 4.440 5.780 5.054 5.114
(4.61)** (4.60)** (6.23)** (5.41)** (5.66)**-4.341 -3.791 -3.893 -2.982
(2.72)** (2.69)** (2.84)** (2.22)*52.756 42.789 35.232 28.333
(10.76)** (9.41)** (6.99)** (5.62)**-8.841 -6.282 -4.659
(5.43)** (3.52)** (2.50)*-2.250 -2.923 -2.133(1.81) (2.37)* (1.79)11.205 10.304 13.036(4.45)** (4.17)** (5.16)**
-20.923 -26.323(3.15)** (3.83)**
-33.796(0.56)
-212.770(4.01)**-109.992(2.08)*19.534(0.31)
212.016 -27.448 -38.576 166.622 235.953(3.25)** (0.49) (0.66) (1.93) (2.72)**
Observations 177 177 177 177 177R-squared 0.108 0.466 0.594 0.617 0.660
Absolute value of t-statistics in parentheses* significant at 5%; ** significant at 1%
Panel B: Economic significance of coefficients in specification (4)
Debt/GDP 5.1 20.9 105.8Change in terms of trade -3.9 13.0 -50.8Volatility of terms of trade 35.2 4.3 149.8Reserves/GDP -6.3 11.8 -73.8GDP growth -2.9 14.5 -42.2Volatiliy of GDP growth 10.3 7.4 76.1Years since last default -20.9 3.5 -73.3
Regression coefficient
standard deviation
predicted change
Table 3: Regressions of sovereign spreads on fundamentals
Debt/GDP
Reserves/GDP
Change in terms of trade
Volatility of terms of trade
This table reports results from regressions of mean annual EMBI spreads for a set of 30 countries from 1994-2003 on a set of explanatory variables. We only include observations for which a country is not in default.
Panel A: EMBI spread regressions
GDP growth
Volatility of GDP growth
Constant
Years since last default
Eastern Europe
Southeast Asia
Middle East & South Asia
Africa
(1) (2) (3) (4) (5) (6) (7) (8)6.235 4.981 5.941 5.180 12.182 11.120 10.799 11.145
(5.13)** (5.28)** (6.58)** (5.61)** (6.82)** (6.45)** (6.18)** (6.40)**-3.924 -3.264 -3.281 -2.784 -2.749 -2.603(2.57)* (2.39)* (2.46)* (1.90) (2.00)* (1.91)51.054 41.514 33.938 49.313 36.324 25.899
(10.79)** (9.32)** (6.65)** (4.62)** (3.49)** (2.22)*-8.492 -6.493 -11.933 -10.150(5.25)** (3.75)** (3.81)** (3.13)**-2.133 -3.020 -1.080 -0.966(1.60) (2.25)* (0.97) (0.88)10.512 10.175 11.422 10.600(4.26)** (4.21)** (2.77)** (2.58)*
-19.456 -20.101(2.86)** (1.93)
Year fixed effects Yes Yes Yes YesCountry fixed effects Yes Yes Yes Yes
Observations 177 177 177 177 177 177 177 177R-squared 0.137 0.496 0.609 0.628 0.242 0.351 0.462 0.476
Absolute value of t-statistics in parentheses* significant at 5%; ** significant at 1%
Years since last default
Volatility of terms of trade
Reserves/GDP
GDP growth
Volatility of GDP growth
Debt/GDP
Change in terms of trade
Table 4: Determinants of spreads across countries and over timeThis table reports results from regressions of mean annual EMBI spreads on explanatory variables. Columns (1) to (4) include year fixed effects, columns (5) to (8) include country fixed effects. Coefficients are within estimators of the panel regressions. R-squared is the within R-squared. We only include observations for which a country is not in default.
cross-section time-series
Non-default observations (not in default next period):
Mean 40.8 -0.3 9.0 11.3 10.0 12.3 9.3Median 39.0 -0.8 8.0 9.5 9.8 10.6 11St. Dev. 18.8 15.6 5.5 7.4 11.8 6.0 3.1Min 12.1 -33.7 2.2 1.6 -17.7 5.0 1Max 104.3 35.6 23.9 27.6 32.3 29.8 11Observations: 560
Default observations (in default next period):
Mean 58.8 -2.1 13.9 9.2 3.1 14.8 7.4Median 59.8 -3.6 14.6 5.2 0.9 14.5 11St. Dev. 25.1 21.0 6.5 7.7 14.9 6.6 4.5Observations: 27
Table 5: Summary statistics for default prediction
Debt/GDP
Change in terms of
trade
Volatility of terms of
trade
Reserves/GDP
GDP growth
Volatility of GDP growth
Years since last
default
This table reports summary statistics for our set of 7 macroeconomic explanatory variables. We consider the sample for which all variables are available. Variables are winsorized at the 5th and 95th percentile levels. We report summary statistics of the group of observations for which the country does not enter default over the following year and for the sample where it does enter default. The sample corresponds to the sample used for logit regressions in Table 6.
GDP growth
Volatility of GDP
growth
Years since last
default
Debt/GDP
Change in terms of
trade
Volatility of terms of
trade
Reserves/GDP
Panel A: Logit regressions on fundamentals(1) (2) (3)
0.043 0.037 0.033(4.47)** (3.62)** (3.06)**0.005 0.015 0.015(0.42) (1.22) (1.19)0.135 0.128 0.132
(4.28)** (3.88)** (4.00)**-0.057 -0.066(1.75) (1.92)-0.031 -0.032(1.85) (1.85)0.029 0.020(0.87) (0.60)
-0.093(1.64)
Constant -6.672 -5.848 -4.701(8.91)** (6.73)** (4.25)**
Observations 587 587 587# of defaults 27 27 27
Pseudo R-squared 0.173 0.203 0.214Absolute value of z-statistics in parentheses* significant at 5%; ** significant at 1%
Panel B: Model forecast accuracy phat > p80 17 18 19p80 >= phat > p60 5 5 5p60 >= phat > p40 2 3 2p40 >= phat > p20 3 1 1p20 >= phat 0 0 0
Panel C: Predicted probabilities of defaultMean 4.6 4.6 4.6Median 2.3 2.2 2.0St. Dev. 7.1 8.0 8.2Min 0.3 0.2 0.1Max 69.8 77.9 76.8Observations: 587
Table 6: Reduced form default prediction
Change in terms of trade
GDP growth
Panel A reports results from logit regressions of a forward looking default indicator on explanatory variables. The default indicator is from Reinhart, Rogoff, and Savastano (2003). It is equal to one if the country goes into default during the next year and zero if it does not enter default in the following year. Summary statistics for the sample are reported in Table 5. Panel B counts the number of actual defaults for parts of the fitted default probability distribution. Panel C reports summary statistics for fitted default probabilities.
Years since last default
Volatility of GDP growth
Debt/GDP
Reserves/GDP
Volatility of terms of trade
Panel A: Summary statistics of predicted spreads and default probabilities
Mean 488.9 127.5 198.2 6.4 4.3Median 411.0 26.5 74.1 1.4 1.8St. Dev. 360.2 203.7 438.2 9.6 8.0Min 30.6 0.0 2.7 0.0 0.1Max 1679.1 864.8 4653.1 37.5 76.8Observations: 177
Panel B: Regression of realized on predicted spreads(1) (2) (3) (4)
dependent variable: EMBI EMBI log(EMBI) log(EMBI)
1.663 0.265(13.99)** (13.42)**
1.850 0.510(14.54)** (16.38)**
305.556 227.153 5.064 3.635(13.41)** (8.86)** (66.79)** (25.33)**
Observations 177 177 177 177R-squared 0.528 0.547 0.507 0.605
Absolute value of t-statistics in parentheses* significant at 5%; ** significant at 1%
Model implied probability
Constant
Table 7: Comparing actual and predicted spreadsPanel A reports summary statistics for EMBI, model implied and reduced form predicted spreads, and model implied and reduced form default probabilities. The sample corresponds to the spread regression sample of Table 3. Panel B reports estimates from regressions of actual spreads on model implied spreads and actual spreads on reduced form predicted spreads, both in levels and in logs. Model implied spreads below 1 basis point are set equal to 1 basis point. For level regressions model implied and reduced form spreads are winsorized at the 10th and 90th percentiles.
EMBI Model implied spread
Reduced form
spread
Reduced form
probability
Reduced form spread
Model implied spread
Figure 1: EMBI spreads and Commodity Prices
100
300
500
700
900
1100
1300
1500
1994
m119
94m7
1995
m119
95m7
1996
m119
96m7
1997
m119
97m7
1998
m119
98m7
1999
m119
99m7
2000
m120
00m7
2001
m120
01m7
2002
m120
02m7
2003
m120
03m7
2004
m120
04m7
2005
m120
05m7
2006
m1
date
EMB
I spr
ead
(bas
is p
oint
s ov
er U
.S. T
reas
urie
s)
180
240
300
360
CR
B C
omm
odity
pric
e in
dex
EMBI
Commodity_Index
E
MB
I spr
ead
(bas
is p
oint
s)
Figure 2a: Volatility of terms of trade in the cross sectionVolatility of terms of trade (%)
0 5 10 15 20
0
500
1000
1500
ARG
BRABUL
CDI
CHICHN
COL
CRO
EGY
EQU
ESD
HUN
LEB
MAL
MEX MOR
PAK
PAN
PER
PHI
PLD
RUS
SAF
SKOTHATUN
TUR
UKR
URU
VEN
EM
BI s
prea
d (b
asis
poi
nts)
Figure 2b: Debt/GDP in the cross-sectionDebt/GDP (%)
20 40 60 80 100
0
500
1000
1500
ARG
BRABUL
CDI
CHICHN
COL
CRO
EGY
EQU
ESD
HUN
LEB
MAL
MEX MOR
PAK
PAN
PER
PHI
PLD
RUS
SAF
SKO THATUN
TUR
UKR
URU
VEN
E
MB
I spr
ead
(bas
is p
oint
s)
Figure 2c: Debt/GDP in the time-seriesDebt/GDP (%)
-40 -20 0 20 40
-500
0
500
1000
ARG
ARG
ARG
ARG
ARG
ARGARG
BRA
BRA
BRA
BRA
BRA
BRA
BRA
BRA
BRA
BUL
BUL
BULBULBUL
BULBUL
BUL
BUL
CDICHICHI CHI CHICHI
CHNCHNCHNCHN
CHNCHNCHNCHNCHN
CHN
COL
COL
COLCOLCOL
COL
COL
CRO
CRO
CRO CRO
EGYEGY
EGYEQU
EQU
EQU
EQUEQU
EQUESD ESD
HUNHUN
HUNHUNHUN
LEB LEBLEB
LEB
LEB
LEB
MAL MAL
MAL
MAL
MAL MALMAL
MALMEX
MEX
MEX
MEX
MEXMEX
MEXMEXMEXMEX
MORMOR
MORMOR
MOR
MOR
PAK
PAK
PAK
PAN
PANPANPAN PANPAN
PANPER PER
PER
PERPER
PER
PHI
PHI
PHIPHI
PHIPHI
PLD
PLDPLD
PLD PLDPLDPLD PLD
PLD
RUS
RUS
RUS
SAF
SAFSAF
SAFSAF
SAFSAFSAF
SAFSKOSKO SKOSKO
SKO
SKOSKO THA
THA
THATHATHA
THATHA
TUN
TUN
TURTUR
TURTUR
TUR
TUR
TUR
TUR
UKR
UKR
UKR
URU
URU VEN
VEN
VEN
VENVEN
VENVEN
E
MB
I spr
ead
(bas
is p
oint
s)
Figure 3: Actual and model implied spreadsModel implied spread
EMBI Fitted_values
1 5 10 50 100 500 1000 2500
10
50
100
500
1000
2500
SAF
SAF
SAF
SAFSAF
SAF
SAF
SAFSAF
MAL
MAL
MAL
MAL
MAL
MAL
MALMAL
TUN
TUN
CHNCHNCHN
CHN
CHNCHNCHN
CHN
CHNCHN
CRO
CROCRO
CRO
HUN
HUNHUN
HUNHUN
URU
URU
SKOSKO
SKO
SKO
SKO
SKO
SKO
PANPANPANPAN
PAN
PANPAN
COL
COLCOL
COLCOL
COLCOL
THATHA
THATHA THA
THA
THAEGY
EGYEGYESDESD
LEB
LEB
LEB LEB
LEBLEB
TUR
TUR
TUR
TURTUR
TURTUR
TUR
MORMOR
MORMOR
MOR
MOR
PLD PLD
PLD
PLDPLD
PLD
PLD
PLDPLD
MEXMEX
MEXMEXMEX
MEX
MEX
MEX
MEXMEX PHI
PHI
PHIPHI
PHIPHI
CHICHICHI CHI
CHI
ARG ARG
ARG
ARGARG ARG
ARG
BRABRA
BRA
BRA
BRA
BRA
BRABRA BRA
BUL
BULBUL
BUL
BULBUL BUL
BULBUL
PERPER
PER
PERPER PER
UKR
UKR
UKR
RUS
RUS
RUSPAK
PAK
PAK
VENVENVEN
VENVEN VEN
VENEQUEQU EQUEQU
EQU
EQU
CDI
EM
BI s
prea
d (b
asis
poi
nts)
Figure 4: Actual and reduced form spreadsReduced form spread
EMBI Fitted_values
1 5 10 50 100 500 1000 2500
10
50
100
500
1000
2500
SAF
SAF
SAF
SAFSAF
SAF
SAF
SAFSAF
MAL
MAL
MAL
MAL
MAL
MAL
MALMAL
TUN
TUN
CHNCHNCHN
CHN
CHNCHNCHN
CHN
CHNCHN
CRO
CROCRO
CRO
HUN
HUNHUN
HUNHUN
URU
URU
SKOSKO
SKO
SKO
SKO
SKO
SKO
PANPANPANPAN
PAN
PANPAN
COL
COLCOL
COLCOL
COLCOL
THATHA
THATHATHA
THA
THAEGY
EGYEGYESDESD
LEB
LEB
LEBLEB
LEBLEB
TUR
TUR
TUR
TURTUR
TURTUR
TUR
MORMOR
MORMOR
MOR
MOR
PLD PLD
PLD
PLDPLD
PLD
PLD
PLDPLD
MEXMEX
MEXMEXMEX
MEX
MEX
MEX
MEXMEX PHI
PHI
PHIPHI
PHIPHI
CHICHICHI CHI
CHI
ARGARG
ARG
ARGARGARG
ARG
BRABRA
BRA
BRA
BRA
BRA
BRABRABRA
BUL
BULBUL
BUL
BUL BUL BUL
BULBUL
PERPER
PER
PERPER PER
UKR
UKR
UKR
RUS
RUS
RUSPAK
PAK
PAK
VENVENVEN
VEN VENVEN
VENEQUEQU EQUEQU
EQU
EQU
CDI
Ave
rage
spr
ead
(bas
is p
oint
s)
Figure 5: Spreads by average default probabilityAverage probability of default
0
300
600
900
(mean) EMBI (mean) Reduced_Form (mean) Model_Implied
0.4 1.1 2.0 4.1 14.1