Embed Size (px)
Citation preview
i
Price Discovery and Information Linkages in the Emission
Allowance and Energy Markets
John Edward Swieringa
February 2013
A thesis submitted for the degree of Doctor of Philosophy of The Australian National University
ii
Declaration
I hereby certify that this thesis is entirely the
work of the author and has not been submitted
to any other institution. Furthermore, all
sources used in the preparation of the thesis
have been acknowledged in the usual manner.
………………………………………...............
John Swieringa
15 February 2013
iii
Acknowledgements
I would like to thank my supervisors Emma Schultz and Tom Smith. Emma was an indefatigable
reader of drafts whose ruthless efficiency with a red pen was invaluable. She was the best sounding
board a PhD student could ask for and I am very grateful for her opinions, her judgement and her
enthusiasm. Tom provided key direction to the research, using his vast depth of experience and
knowledge to point out relevant literature and empirical techniques. I would also like to thank
Raymond Liu and Carole Comerton-Forde for insightful comments on drafts of the first chapter.
PhD students are always indebted to those who put up with them during their struggles and in that
regard I thank Gaurav Khemka, with whom I share an office and revelled in our daily coffee and
darts sessions. I thank my parents for their encouragement and for bravely attempting to read my
work. Most importantly, I would like to thank my wife Jess, who went through with marrying me in
the depths of this undertaking. Her love and support made this work possible.
iv
Abstract
We provide the first evidence on the catalysts for price discovery in the European Union Emissions
Trading System. Short-run return dynamics are analysed using a regression approach similar to
Fleming, Ostdiek and Whaley (1996), while the permanent contribution of securities to long-run
price equilibrium is examined by calculating Hasbrouck‘s (1995) information shares. By employing
high frequency data across a wide range of securities, we find that trading costs are a more
important determinant of price discovery than the implicit provision of leverage in securities such as
futures and options. Securities with low trading costs display greater price discovery than those with
high trading costs.
We also examine price discovery within the European markets for coal, natural gas and crude
oil. Results show that Brent crude oil futures display greater price discovery than a proxy for the
physical Brent market, while there is evidence that West Texas Intermediate futures still dominate
price discovery globally. In natural gas markets, UK natural gas futures display greater price
discovery than physical trading at North-West Europe‘s main natural gas hubs, though weak links
to the crude oil market remain. Due to a lack of liquidity and transparency, it remains difficult to
distinguish between coal securities. Overall, our results support the importance of futures contracts
as a source of price discovery in contrast with opaque over-the-counter physical trading.
Having established where price discovery is taking place in the European emission allowance
and energy markets, we examine volatility and information linkages between them by employing a
rational expectations framework similar to Fleming, Kirby and Ostdiek (1998). The model specifies
volatility linkages operating through common information and information spillover channels. We
estimate a representation of this model using GMM for bivariate pairings of emission allowances
with coal, natural gas and crude oil. We find that emission allowances are most strongly linked to
the crude oil market, in spite of more direct economic relationships with coal and natural gas.
v
Table of Contents
CHAPTER 1: INTRODUCTION ........................................................................................................................ 1
CHAPTER 2: PRICE DISCOVERY IN THE EUROPEAN UNION EMISSIONS TRADING SYSTEM ............................ 4
2.1 A BRIEF INTRODUCTION TO THE EU ETS ............................................................................................................. 8
2.1.1 Emission Allowances .......................................................................................................................... 8
2.1.2 Phases .............................................................................................................................................. 11
2.2 DATA ......................................................................................................................................................... 14
2.2.1 Series Selection ................................................................................................................................ 15
2.2.2 Measures of Trading Cost ................................................................................................................ 20
2.2.3 Return Series Construction .............................................................................................................. 21
2.2.4 Serial Correlation ............................................................................................................................. 23
2.2.5 Stationarity and Cointegration ........................................................................................................ 27
2.3 METHODOLOGY ........................................................................................................................................... 30
2.3.1 Basic Regression Specification ......................................................................................................... 30
2.3.2 Cointegration and Error Correction ................................................................................................. 33
2.3.3 The Final Model Specification .......................................................................................................... 35
2.3.4 Information Shares .......................................................................................................................... 36
2.4 RESULTS ..................................................................................................................................................... 39
2.4.1 Regression Results ........................................................................................................................... 39
2.4.2 Regression R-Squared and F-Statistics ............................................................................................. 44
2.4.3 Ordinal Ranking ............................................................................................................................... 48
2.4.4 Information Shares .......................................................................................................................... 53
2.4.5 Strength of Findings ......................................................................................................................... 57
2.5 CONCLUSION ............................................................................................................................................... 58
CHAPTER 3: PRICE DISCOVERY IN EUROPEAN ENERGY MARKETS ............................................................... 60
3.1 METHODOLOGY ........................................................................................................................................... 64
3.1.1 Regression Approach ....................................................................................................................... 64
3.1.2 Information Shares .......................................................................................................................... 65
3.2 COAL ......................................................................................................................................................... 67
3.3 NATURAL GAS ............................................................................................................................................. 73
3.4 CRUDE OIL.................................................................................................................................................. 82
3.4.1 Price Discovery in the Brent Crude Oil Complex ............................................................................... 82
3.4.2 Price Discovery in Brent and WTI Futures ........................................................................................ 88
3.5 CONCLUSION ............................................................................................................................................. 100
3.6 APPENDIX ................................................................................................................................................. 102
CHAPTER 4: INFORMATION LINKAGES BETWEEN THE EMISSION ALLOWANCE AND ENERGY MARKETS ... 108
4.1 EXISTING EVIDENCE ON MARKET INTERACTIONS ............................................................................................... 110
4.2 INFORMATION LINKAGES ............................................................................................................................. 112
4.3 METHODOLOGY ......................................................................................................................................... 117
4.3.1 Directionality of Emission Allowance and Energy Market Relationships ....................................... 117
4.3.2 Stochastic Volatility Model ............................................................................................................ 117
vi
4.4 DATA ....................................................................................................................................................... 121
4.4.1 Security Selection ........................................................................................................................... 121
4.4.2 Descriptive Statistics ...................................................................................................................... 123
4.4.3 Serial Correlation ........................................................................................................................... 125
4.4.4 Cross-Market Correlations ............................................................................................................. 126
4.5 RESULTS ................................................................................................................................................... 127
4.5.1 Regression Results ......................................................................................................................... 128
4.5.2 Information Linkages ..................................................................................................................... 128
4.6 CONCLUSION ............................................................................................................................................. 135
4.7 APPENDIX ................................................................................................................................................. 137
CHAPTER 5: CONCLUSION ........................................................................................................................ 143
REFERENCES ............................................................................................................................................. 146
1
CHAPTER 1: Introduction
There is widespread agreement that anthropogenic global warming is the result of long standing
market failures where, in the absence of defined property rights, firms combusting fossil fuels have
no incentive to restrict their resultant emissions of carbon dioxide and other greenhouse gases (see,
for example, Stern, 2006). Emissions trading is an attempt to redress this market failure along the
lines first suggested by Coase (1960) insofar as efficient outcomes are achieved by allocating
property rights and facilitating trade between the parties affected by an externality. Under a cap and
trade system, polluters who emit more than their allocation of carbon dioxide emissions must
purchase allowances from other market participants who emit less or pay significant penalties.
While scarcity of emission allowances encourages polluters who can abate their emissions at low
cost to do so and profit by selling their excess allowances, scarcity may also force polluters who
cannot afford to purchase allowances or abate their emissions to cease production altogether.
Emissions trading was first undertaken on a significant scale in the United States in the mid-
1990s to reduce acid rain. While the European Union Emissions Trading System (EU ETS) was
introduced a decade later, only being launched in 2005, it is many times larger in size. The EU ETS
places a cap on the total amount of carbon dioxide emissions Europe‘s large polluters are allowed to
emit each year, with emission allowances allocated or auctioned to polluters by Europe‘s
governments. The cap is reduced each year such that the right to pollute becomes increasingly
2
scarce over time. Emission allowances in the EU ETS are traded over-the-counter between
polluters, sometimes with the involvement of financial institutions, as well as on organised
exchanges that facilitate spot, futures and options trading. Given its importance in mitigating the
effects of global warming, understanding the dynamics of this relatively new emission allowance
market, and the energy markets more broadly, is paramount for policy makers, market participants
and academics alike. We contribute to the literature in this regard by assessing price discovery in
both the EU ETS and the main fossil fuel energy markets as well as information linkages between
the two.
Specifically, in Chapter 2 we examine price discovery in the main spot and futures securities
traded in the EU ETS. In doing so, we provide the first investigation into the catalysts for price
discovery in this market, paying particular attention to trading costs and leverage. We also consider
whether market segmentation between emission allowances created within the EU ETS or created
under the auspices of the United Nations impacts upon price discovery. Consistent with evidence on
price discovery in other markets, we find trading costs are more important than leverage or a
security‘s origin.
In Chapter 3, we broaden our focus to examine both short and long-run price discovery in the
main European fossil fuel energy markets, namely coal, natural gas and crude oil. Prior research
into the nature of price discovery in these energy commodities has been relatively sparse. These
markets are characterised by the operation of price reporting agencies who survey physical market
participants in order to construct benchmark prices. These prices then form the basis for long-term
supply contracts, which remain the most prevalent mode of exchange for all three commodities.
However, because these survey processes are often opaque and physical markets commonly suffer
liquidity problems, the financial layers of these markets, particularly futures trading, play an
important role in price discovery. We attempt to span both the physical and financial layers of these
markets and, to the best of our knowledge, we provide the first assessment of price discovery in the
European coal market and the broadest study of price discovery for European natural gas.
Thereafter, we make the first attempt to assess both the physical and financial layers of the Brent
3
market. Finally, in light of recent market dislocations, we assess linkages between the European
crude oil and natural gas markets and re-examine linkages between the major global crude oil
benchmarks, namely Brent and West Texas Intermediate. Overall, our results support the
importance of futures contracts as a source of price discovery in contrast with opaque over-the-
counter physical trading.
Having established price discovery both in the emission allowance and energy markets, we
analyse channels of interaction between them. Specifically, in Chapter 4 we employ a rational
expectations framework in the tradition of Tauchen and Pitts (1983), Fleming, Kirby and Ostdiek
(1998) and Kodres and Pritsker (2002) to analyse market interactions on the basis of responses to
commonly relevant information and the spillover effects of idiosyncratic information. By doing this,
we take account of complexities that have previously been overlooked, including the true impact of
operational and strategic considerations in the generation of electricity and the limitations imposed
by the current power generation mix in a given economy. These factors inhibit the extent to which
fuel switching is likely to be an observable short-term phenomena, let alone one that dictates
directional interactions between fuel input and emission allowance prices. We estimate a stochastic
volatility representation of this model and, in the absence of a priori expectations concerning the
directional relationships between the markets of interest, we assess information linkages using the
correlation of volatilities. Our results show that, despite the strong economic linkages between
emission allowances and coal and natural gas, emission allowances have the strongest information
linkages to the crude oil market, which is likely a product of strong common information linkages.
Our findings highlight the importance of information arrival in our understanding of markets
generally and, for the emission allowance and energy markets in particular, it is a reminder that
their interactions will be strongly influenced by information that is commonly relevant as they share
many fundamental determinants of value.
4
CHAPTER 2: Price Discovery in the European Union
Emissions Trading System
The EU ETS is a decentralised market place that includes over-the-counter trading of emission
allowances as well as spot, futures and option trading on nearly a dozen organised exchanges. With
such a wide dispersion of tradable securities and trading venues, identifying the security price most
reflective of current, relevant information—the source of price discovery—is important for market
participants and regulators alike. The large number of securities that are all essentially fungible with
one another also provides a good opportunity to examine which market frictions are important
drivers of price discovery, such as trading costs or leverage. Against this backdrop, we analyse
intraday return data for seven of the most traded securities in the EU ETS to establish which
security‘s returns have a greater tendency to lead the returns of the others. Thereafter, we look at
whether the relationships between securities are consistent with trading costs or the implicit
leverage in futures and options determining price discovery. In addition, we look at whether the
government allocated or auctioned European Union Allowances are better sources of price
discovery compared with external, project-based allowances such as the Certified Emission
Reduction units created under the auspices of the United Nations Clean Development Mechanism.
Price discovery in the EU ETS is a topic that has received attention in previous literature, but
prior studies rarely utilise high frequency data nor do they examine the underlying catalysts for
price discovery. As it is a relatively new market, much of this literature has concentrated on the
5
system‘s introductory phase, which is problematic because a collapse in prices late in this phase
largely resulted in a complete absence of meaningful price changes (see, for example, Uhrig-
Homburg and Wagner, 2008, and Benz and Hengelbrock, 2008). Several studies have looked at
price discovery during the second, better functioning phase of the EU ETS. However, by analysing
daily data, the important, granular aspects of the timeliness of price responses to information arrival
are somewhat obscured (see, for example, Chevallier, 2010a, Chevallier, 2010b, and Mizrach,
2010). These daily studies generally find that the most traded futures contract in the EU ETS is the
source of price discovery—the December expiry European Union Allowance futures contracts
traded on London‘s Intercontinental Exchange1. Rittler (2009) and Mizrach and Otsubo (2011)
undertake intraday studies during the system‘s second phase and also find evidence that the
Intercontinental Exchange futures are the source of price discovery but, like much of the other
literature, they do not attempt to examine what market frictions drive the price discovery process.
Differential trading costs between securities are a prominent form of market friction. These
costs can be explicit, such as brokerage and clearing fees, or implicit, such as the cost of a round
trip in buying and selling a security (the spread between bid and ask prices). The explicit costs of
transacting are difficult to measure as they will vary depending upon a market participant‘s
relationship to their particular broker or, in the case of brokers themselves, their clearing fees may
vary with their level of membership at a particular exchange. Regardless, these costs are often small
compared to the implicit costs of trading, particularly for less liquid securities which tend to have
wide bid-ask spreads, little market depth and for which trades have a large impact on the price level.
If several securities are identical in all characteristics save trading cost, a market participant looking
to profit by trading on new information will realise higher returns by trading the security with the
1 Note that these prior studies refer to this as the European Climate Exchange December futures contract.
Intercontinental Exchange purchased the parent company of the European Climate Exchange in April 2010.
2 The trader must keep the account supplied with funds above the maintenance margin after daily marking to
market, but also typically receives interest on these balances in the meantime.
3 See, for example, Fleming, Ostdiek and Whaley (1996) for evidence on US stocks, stock indices and stock
derivatives, Booth, So and Tse (1999) for evidence from Germany and Hsieh, Lee and Yuan (2008) for
evidence from Taiwan. 4 The EU ETS covers approximately 11,000 European installations owned by around 5,000 companies (World
Bank, 2010). According to the European Commission (2008), these installations account for approximately
6
lowest trading cost. Our ―Trading Cost Hypothesis‖ is that price discovery in the EU ETS takes
place in the securities with the least trading cost as measured by the bid-ask spread.
Another common market friction potentially impacting upon trader preferences is the relative
provision of leverage. While traders may fund long positions in spot markets by explicitly
borrowing funds or, in the case of a short sale by borrowing the security, futures and options, by
construction, provide leverage implicitly. Establishing a position in a futures market only requires
that a trader transfer an initial margin amount to their broker which is typically a fraction of what
the actual purchase price would be to achieve a similar exposure in the underlying asset‘s spot
market2. The initial outlay for an option takes the form of the call or put premium and these again
facilitate leverage inasmuch as premiums are only fractions of the underlying asset‘s price. In the
presence of limitations on the extent of a market participant‘s ability to seamlessly borrow funds or
assets, the leverage characteristics of futures and options would be attractive, giving a speculator a
preference for derivative securities when looking to profitably trade on new information. The
―Leverage Hypothesis‖ is that greater price discovery occurs in emission allowance futures and
options than in spot securities.
We also assess a third hypothesis that is unique to the EU ETS regarding whether market
segmentation by emission allowance type impacts upon price discovery. The EU ETS gives
installations a choice of surrendering European Union Allowances in abatement of their emissions,
which are allowances allocated or auctioned by European governments, or alternatively they can
choose to surrender allowances generated by projects that result in emission reductions in other
countries under the auspices of the United Nations. Unlike European Union Allowances, there are
limits imposed by various European governments on how many project-based allowances
installations can surrender for compliance each year. There is also a great deal of regulatory
uncertainty about whether particular project-based allowances will continue to be accepted at all.
On the basis of these limits and uncertainties, project-based allowances trade at a substantial price
2 The trader must keep the account supplied with funds above the maintenance margin after daily marking to
market, but also typically receives interest on these balances in the meantime.
7
discount compared to European Union Allowances. Market participants looking to exploit new
information relevant to the emissions market may forgo trading project-based allowances, in
preference for the costlier European Union Allowances effectively segmenting the EU ETS by
allowance type. The ―Market Segmentation Hypothesis‖ is that European Union Allowances display
greater price discovery than project-based allowances.
Our results show that the most traded futures contract in the EU ETS, the December expiry
Intercontinental Exchange European Union Allowance futures, is the source of price discovery.
This is unsurprising given this security has the least trading cost, provides market participants with
leverage and is not a project-based allowance. Unlike prior studies, we examine what factors drive
price discovery, with the results showing that low trading cost appears to be the most important
determinant. This is consistent with much of the literature concerning other markets3. The results on
market segmentation are mixed, while leverage is shown to be relatively unimportant.
Section 2.1 provides an introduction to the EU ETS, explains the differences in allowance types
and why these may constitute a segmentation of the market, and describes how problems with the
implementation phase of the EU ETS justify focusing on the second phase. Section 2.2 details the
construction of intraday return series, provides descriptive statistics and tests for autocorrelation,
stationarity and cointegration. The methodology employed is discussed in Section 2.3, with
reference to common methodologies used in prior literature on price discovery. The final regression
specification is a compromise between using leading, lagging and contemporaneous returns to
assess price discovery and the use of Vector Error Correction Models. Section 2.4 presents the
results of the regression analysis and attempts to distinguish between the relative strength of the
three hypotheses using measures of goodness of fit and the overall significance of the regression
coefficients. Hasbrouck‘s (1995) information shares are also calculated for robustness and to
establish which security makes the greatest contribution to long-run price equilibrium. Section 2.5
concludes.
3 See, for example, Fleming, Ostdiek and Whaley (1996) for evidence on US stocks, stock indices and stock
derivatives, Booth, So and Tse (1999) for evidence from Germany and Hsieh, Lee and Yuan (2008) for
evidence from Taiwan.
8
2.1 A Brief Introduction to the EU ETS
Since 2005, the EU has operated an emissions trading system to assist in achieving its commitment
to reduce greenhouse gas emissions. The EU ETS is a cap and trade system in which the quantity of
emissions that the EU‘s large polluters emit is capped and set by the European Commission. The
emission cap is lowered through time to meet emission reduction targets agreed internationally
under the Kyoto Protocol (the cap decreases annually by approximately 1.74 per cent). Polluters are
allocated allowances, either free or via auctions, which they surrender annually against their
assessed emissions. Where they have a surplus or deficit of allowances relative to their actual
emissions, polluters can trade with other institutions in the EU ETS either in bilateral over-the-
counter (OTC) transactions or in organized spot, futures and option markets facilitated by almost a
dozen exchanges. By making the right to pollute increasingly scarce, the market mechanism should
allocate emission rights to those with the highest value in continuing to pollute; those polluters for
whom the cost of reducing their emissions by other means is highest4,5
.
2.1.1 Emission Allowances
Three types of allowances can be used by polluters in the EU ETS: European Union
Allowances (EUAs), Certified Emission Reductions (CERs), and Emission Reduction Units
(ERUs). When surrendered, each allowance acts as abatement for emitting 1 metric tonne of carbon
dioxide (CO2) or the equivalent amount of another greenhouse gas into the atmosphere6.
4 The EU ETS covers approximately 11,000 European installations owned by around 5,000 companies (World
Bank, 2010). According to the European Commission (2008), these installations account for approximately
half of Europe‘s CO2 emissions and 40 per cent of Europe‘s total greenhouse gas emissions. Only a handful of
countries outside the EU have implemented compulsory emissions trading systems, such that the EU ETS
accounted for 97 per cent of the global emissions market by value in 2010 (World Bank, 2011).
5 See Hepburn (2006) and Stern (2006) on the efficacy of a carbon trading system versus taxes or mixed
systems.
6 Stern (2006, p.198) discusses how the global warming potential—the radiative forcing and lifespan—of
other greenhouse gases is used to calculate their carbon dioxide equivalence (termed CO2e). This metric
allows the global warming impact of different gases to be compared and, for the purposes of an ETS, gives an
exchange standard for compliance. For example, methane has 23 times the global warming potential of carbon
9
EUAs are the most common allowance type in the EU ETS and are allocated or auctioned by
governments to Europe‘s large polluters. These allowances are designed to be perfectly fungible
with the standard metric of emissions under the Kyoto protocol, the Assigned Amount Unit (AAU),
and are also fungible with CERs and ERUs, which are the Kyoto Protocol‘s project-generated
allowances.
The mechanisms for creating CERs and ERUs are designed to promote cross-border investment
in emission reduction and the transfer of clean technologies between countries. CERs are essentially
allowances generated when developed country organisations undertake emission reduction projects
located in developing countries. They are validated by the executive board of the Clean
Development Mechanism (CDM) under the United Nations Framework Convention on Climate
Change (UNFCCC). ERUs are similar to CERs but they are generated by a developed country
organisation undertaking an emission reduction project in another developed country and, as such,
they are termed Joint Implementation (JI) projects by the UNFCCC7. The majority of CERs come
from projects undertaken in China and India and, while far fewer ERUs have been created, those
that trade in the EU ETS mainly come from projects in Russia and the former Eastern Bloc
countries. Both types of project-based allowances trade in the EU ETS on the basis that, in terms of
the global warming effects, it is irrelevant where greenhouse gases are emitted. In accordance with
this, the European Commission established the EU Linking Directive to allow CERs and ERUs to
be surrendered by European polluters such that it encourages emission reduction schemes to be
undertaken in whatever part of the world in which such schemes are most cost effective. A number
of limitations apply to the use of these project-based allowances. The EU Linking Directive gives
member governments discretion over whether to cap the percentage of CERs and ERUs that can be
dioxide over a 100 year timeframe and therefore 23 allowances would have to be surrendered per metric tonne
of methane emitted.
7 There are several other subtle differences between CERs and ERUs. For example, when two developed
countries are involved in a JI project, they have both agreed to binding emission caps under the Kyoto
Protocol and so the project doesn‘t lead to the creation of new allowances, but rather existing AAUs
belonging to the project‘s host country are converted into ERUs which are then given to the organisation
undertaking the project. For more information on the distinguishing features of these project based
allowances, see the UNFCCC (http://unfccc.int/2860.php).
10
surrendered for compliance purposes by installations in their jurisdictions (European Commission,
2004). Many governments have chosen to impose such limits amid concerns that a large, externally-
generated supply of CERs and ERUs could flood the market and remove incentives for domestic
installations to take direct action to reduce emissions themselves. The average annual limit across
the EU ETS on the surrendering of project-based allowances is approximately 13.4 per cent8.
The European Commission also excludes particular types of CERs and ERUs from being
surrendered for compliance. The exclusion of certain allowance types is mainly due to concerns
about the methodologies used in calculating the future emission reductions stemming from
particular project types, but others are excluded on the basis of a project potentially having
counterproductive environmental impacts9. Examples of excluded allowances include those
generated by projects for land use, land use change and forestry, projects that involve moving from
fossil fuel to nuclear power generation and specific large hydroelectricity projects10
. From April
2013 onwards, projects reducing certain industrial gases, such as hydro fluorocarbons (HFCs) and
nitrous oxide (N2O), will also be excluded. This will rule out the use of 52 per cent of the CERs
currently in existence (European Commission, 2010). In addition, because some developing
countries are seen as frustrating attempts to reach a successor agreement to the Kyoto Protocol,
from 2013 the European Commission will also restrict the use of newly created CERs to be only
those originating from projects in Less Developed Countries (LDCs) unless a successor agreement
to Kyoto is subsequently reached11
.
Although eligible CERs and ERUs have exactly the same compliance value to a European
polluter in abating emissions as EUAs, the limitations and uncertainties surrounding their use mean
8See Mansanet-Bataller, Chevallier, Hervé-Mignucci and Alberola (2011).
9 For example, HFC-23, a by-product of making the refrigerant gas HCFC-22, has a global warming potential
11,700 times higher than CO2. Its destruction can be accomplished at a cost of as little as €0.17 per tonne.
This creates an incentive for existing installations in China and India to make more refrigerators containing
HCFC-22 just so the by-product can be destroyed in exchange for CERs worth €7.00 – €15.00 per tonne of
CO2e (European Commission, 2010).
10 European Commission website: http://ec.europa.eu/clima/policies/ets/linking_en.htm.
11 LDCs are defined as 33 African countries, 14 Asian countries and Haiti, few of which currently have any
emission reduction projects. The definition of developing countries is much broader (See the Economic and
Social Council of the United Nations: http://www.unohrlls.org/).
11
that they trade at a substantial discount (see Chart 2.1). As such, the market for emission allowances
may in fact be segmented along these lines. Traders utilising new information pertinent to all
emission allowances, such as energy prices, levels of industrial production in the EU or the effects
of weather on electricity demand, may be more inclined to trade EUAs than the project-based
allowances. This Market Segmentation Hypothesis is examined alongside the relative effects of
trading cost and leverage on price discovery.
As this study analyses trading of CERs and ERUs (along with EUAs), the focus is on the
secondary market. Primary market CER and ERU prices are typically even lower than in the
secondary market. This is because primary sales are often negotiated years in advance to secure
project funding and because buyers typically demand a discount given the risk that a project may
not meet the European Commission‘s compliance standards by the time the allowances are created.
2.1.2 Phases
The EU ETS has been implemented over three phases, with the first two timed to coincide with
the first implementation period agreed under the Kyoto Protocol. Phase 1 of the system ran from the
start of 2005 to the end of 2007. Designed to allow large polluters to ease into the new
arrangements, Phase 1 saw all of the emission allowances allocated free of charge to polluters.
From late April 2006, it became apparent that the assessed quantity of actual emissions by Europe‘s
large polluters was lower than the amount that had been allocated (i.e. the level of the cap). As
polluters were not permitted to bank allowances allocated during Phase 1 for use in later phases of
the EU ETS, the price of allowances began a descent towards zero in the months after the April
2006 emissions assessment (see Chart 2.1)12,13
.
12
Poland and France were exceptions to this, allowing very limited banking of Phase 1 allowances for later
use.
13 Other problems encountered in Phase 1 included legal action taken by the European Commission against
Poland and Estonia for overestimating their emissions—and thus increasing their national allocation of
allowances to the benefit of their polluting industries—and suspicions that the market was being used by
criminal syndicates for money laundering.
12
Chart 2.1
EUA and CER Prices
Last trade prices are sourced from Thomson Reuters Tick History sampled over 60-minute intervals between 2 May 2005 and 29 April 2011.
EUA and CER prices are December expiration annually maturing futures contract prices from Intercontinental Exchange in €/tCO2e.
0
5
10
15
20
25
30
35
0
5
10
15
20
25
30
35
May-05 May-06 May-07 May-08 May-09 May-10
EUA Price CER Price
2005 2006 2007 2008 2009 2010 2011
April 2006:
Over-allocation becomes apparent
March 2010: Hungarian CER
Recycling
January 2011:
Allowance TheftsMid-2009: Several countries
take action against VAT fraud
Phase 1 Phase 2
13
Phase 2, which will continue until the end of 2012, still largely involves the free allocation of
EUAs, but auctioning by several countries has progressively increased14
. With the deepening of the
financial crisis in late 2008, expectations of a slowdown in industrial production depressed
allowance prices amid fears that, similar to Phase 1, emissions would be well below the system cap.
However, the ability to bank EUAs from Phase 2 for use in future compliance periods ensured that
the price level of allowances did not collapse to zero. Despite this improvement, Phase 2 has been
subject to several problems, including Value Added Tax (VAT) fraud, CER recycling and
allowance thefts from national registries.
In mid-2009, significant VAT fraud was exposed in the allowance spot market. The fraud
involved buying allowances from counterparties in a country that did not include VAT in the
settlement price, but rather invoiced the buyer asking for payment 1 to 3 months later, then
simultaneously selling the allowances in a country that did include VAT in the settlement price and
then disappearing before the invoice for the VAT on the purchase was due to be paid15
. The
exposure of this fraud was the catalyst for several countries changing their tax codes (World Bank,
2010).
In March 2010, CERs surrendered by Hungarian companies in abatement of their emissions
found their way from the Hungarian national registry back into the BlueNext spot market. Instead of
retiring the CERs after they were surrendered, the Hungarian Ministry of Environment and Water
resold them to Hungarian Energy and Power, supposedly with the caveat that the buyer
acknowledged that they were ineligible for further use in the EU ETS (see The Economist, 2010,
and The World Bank, 2010). Japanese firms were reported to have subsequently offered them for
sale on BlueNext without any stipulation that they were ineligible in the EU ETS. As a result
BlueNext was closed from 17 to 19 March 2010 while more stringent checks were put in place by
14
Auctioning by Germany averages 9 per cent, the United Kingdom 7 per cent, the Netherlands 4 per cent and
1 per cent for both Austria and Ireland, according to the European Commission‘s website:
http://ec.europa.eu/clima/policies/ets/auctioning_second_en.htm.
15 Spot market allowance trades are subject to VAT in several countries, which define them as goods, whereas
futures and options are exempt on the basis that they are financial transactions.
14
the exchanges, and governments moved to introduce stricter rules to prevent double counting of
allowances.
On 19 January 2011, as a result of several reports of allowance thefts from national registries,
the European Commission announced a shutdown of spot markets for a minimum of two weeks or
until the registries were able to put better security systems in place. The thefts amounted to around
3 million allowances worth approximately €45 million from the national registries of Austria, the
Czech Republic and Greece (The Economist, 2011).
During Phase 3, which will run from the start of 2013 to the end of 2020, the EU ETS will
steadily move towards auctioning the majority of EUAs16
. It is expected that free allocations will
diminish from 80 per cent to 30 per cent of issued allowances by 2020. An exception will be made
for companies in trade exposed industries, such as cement and steel making, which will still receive
allocations for free so as to prevent the movement of these types of installations to countries not
subject to emission restrictions (thus preventing what is called ‗carbon leakage‘).
The next section describes the data used to study price discovery in the EU ETS. Due to the
collapse of prices during Phase 1 and the fact that CERs and ERUs were not accepted for
compliance during Phase 1, the analysis concentrates on Phase 2.
2.2 Data
This section describes the construction of intraday return time series for the most traded securities
in the EU ETS. Initially, the main exchanges and securities are identified, with the most traded of
these selected for analysis. Thereafter, a common date range is established and a common intraday
window chosen. Two return series are constructed for each security over the chosen window: one
based on actual trade prices; and, another on prices from the mid-point of the bid-ask spread.
Finally, tests for serial correlation, stationarity and cointegration are performed.
16
The European Commission has decided to extend the EU ETS to Phase 3 even though an international
agreement extending the implementation of the Kyoto Protocol beyond 2012 has not been reached.
15
2.2.1 Series Selection
The EU ETS is comprised of bilateral OTC trading as well as spot, futures and option trading of
allowances facilitated by exchanges. OTC trade price data are not available and, consequently, the
analysis necessarily focuses on trading facilitated by organized exchanges17
. Exchange trading of
emission allowances was originally conducted via specialist energy trading platforms that expanded
to encompass emission allowances with the advent of the EU ETS. However, consistent with the
increased consolidation of financial exchanges in the last decade, these emission exchanges are now
predominantly owned by large global exchange groups or consortiums of banks and brokers.
Table 2.1 lists the nine exchanges that have facilitated trade in EU ETS securities during Phase 2
together with the security types traded on each.
Hereafter, the securities listed in Table 2.1 will be referred to using a three letter abbreviation
corresponding to the exchange on which they are traded, followed by the unit type (EUA, CER or
ERU) and then the security type (spot, futures or options). For example, the EUA futures contract
traded on the Intercontinental Exchange is termed the ‗ICE EUA Futures‘ series and the CERs
traded spot on BlueNext are termed the ‗BNX CER Spot‘ series.
Price discovery is unlikely to take place in the use of securities for which there is little trade
activity. As such, our initial focus is on identifying the most traded securities by volume in the EU
ETS. To assess this, trade and quote data are sourced from Thomson Reuters Tick History. While
spot data series are easily constructed, futures series require combinations of contracts with
different expiration dates. Although most exchanges offer allowance futures contracts with a variety
of expiration frequencies, in the EU ETS trade volume is invariably concentrated in the annually
expiring futures. In light of this, our analysis is focused on these contracts.
17
The OTC market allows participants to trade smaller parcels of allowances than the standard 1,000 tCO2e
minimum trade size of most exchange-traded securities. According to the World Bank (2010), approximately
70-80 per cent of trade took place on an OTC basis during Phase 1, but this has since decreased to less than 50
per cent in Phase 2. Much of the remaining OTC trading is now cleared by the exchanges as market
participants seek greater transparency amidst incidents like the CER recycling described in the previous
section.
16
Table 2.1
Exchanges the in EU ETS
Information regarding the exchanges is sourced from their respective websites. Where there have been name
changes, these have been noted, with details of consolidation between exchanges gathered from press releases
issued by the exchanges.
When studying price discovery using frequent intraday price and quote observations it is
optimal to analyse changes in the prices or limit orders of tradable securities, rather than
synthetically created securities or indices. For this reason, futures chain series are not used as these
involve blending prices of overlapping expirations when the front contract approaches its maturity
date. Instead we use series that involve the use of the front annual expiry futures contract up until
the start of its expiration month, after which point the next closest annual expiry series is used18
.
18
Although the delivery window on emission allowances is typically only the last trading day, moving to the
next maturing contract approximately two weeks prior to maturity fits roughly with the behaviour of market
participants as futures traders typically roll into the new front contract prior to hitting the delivery date or
window. The efficacy of constructing a single futures series in this way is supported by the pattern of
declining trade volumes in front contracts in their final days before expiration and the increasing volume in
the next-to-front contract in these periods. This is the approach taken by Bessembinder (1992) in futures
series construction, though Carchano and Pardo (2009) point out that there is unlikely to be a significant
difference in results when series are constructed by rolling into the next contract at the start of the delivery
month or by rolling at the delivery date.
Exchange Name Abbreviation Primary Location EU ETS Security Types
BlueNext BNX Paris EUA Spot/Futures
- formerly PowerNext CER Spot/Futures
ERU Spot
Climex CLX Utrecht EUA Spot
- closed in 2010 CER Spot
Energy Exchange Austria EXA Vienna EUA Spot
European Energy Exchange EEX Leipzig EUA Spot/Futures/Options
CER Futures
Green Exchange GRX Chicago EUA Spot/Futures/Options
CER Futures/Options
Greenmarket Exchange GMX Munich EUA Spot
CER Spot
Intercontinental Exchange ICE London EUA Spot/Futures/Options
- formerly European Climate Exchange (ECX) CER Spot/Futures/Options
Multi Commodity Exchange MCX Mumbai EUA Futures
CER Futures
NASDAQ OMX Europe NOX Oslo EUA Spot/Futures/Options
- formerly Nord Pool CER Spot/Futures/Options
17
Table 2.2 shows the various EU ETS securities ranked by average trade volume and also
includes the dates during Phase 2 for which there is intraday trade and quote data available.
Initially, the first eight securities are selected for further analysis. Although this cut-off is somewhat
arbitrary, it is a relatively conservative choice considering the thin trading in some of the securities
included. Moreover, trading volume falls away sharply after this cut-off. The Thomson Reuters
trade and quote data were filtered to remove missing, zero and erroneous prices.
The common overlapping range of dates for which data is available for all eight series is too
short for meaningful analysis19
. As such, the selection is further reduced to seven series by the
removal of the GRX EUA Futures series for which only a relatively short sample of price, quote
and volume data is available (4 August 2010 to 29 April 2011)20
. With this series removed, the
sample date range for the analysis is set as the overlapping window starting on 1 July 2009 and
ending 30 December 201021
.
Holiday calendars have been examined for the four exchanges with dates removed from the
time series where one or more of the exchanges are closed. The three days from 17 to 19 March
2010 have been removed due to the closure of BlueNext following the Hungarian CER recycling
incident. Four days have also been removed where, though unexplained, one or more exchanges had
no trade or quote data. The sample thus contains 376 trading days over an 18-month window.
19
Many exchanges only introduced securities well after the start of Phase 2. These delayed introductions are
not surprising given the financial crisis and the impact of over-allocation of allowances in Phase 1, which
likely motivated exchanges to wait and see whether the reported emissions in April 2008 were above or below
the system cap. Likewise, having been launched, some securities failed to attract much interest and were
subsequently abandoned. In particular, a number of spot securities never reopened for trade after the European
Commission‘s two-week shutdown of the spot market following allowance thefts from national registries
starting 19 January 2011.
20 It should also be noted that these contracts hardly traded before 2011 anyway. Although in Table 2.2 the
GRX EUA Futures appear to have the fifth largest trade volume, there was hardly any trade in this security
until after the spot market closure in January 2011 when spot market participants were forced to seek out
derivative securities to manage their exposures. On average the GRX EUA Futures only traded about once
every five days between 4 August 2010 and 30 December 2010.
21 It is possible to start the overlapping window earlier, with the introduction of ICE EUA Spot in
March 2009, but this security had very little trade volume in its first months. Likewise, it would be possible to
extend the analysis to 19 January 2011 but, to avoid any association with the allowance thefts,
30 December 2010 is the chosen end-date.
18
Table 2.2
Trade Volume and Data Availability
Trade volume data is sourced from Thomson Reuters Tick History. Futures are annual expiration series.
Option volume data is aggregated across all strikes and option types (calls and puts). Trade volume is
measured in tCO2e, which are tonnes of carbon dioxide or an equivalent amount of another greenhouse gas.
The cut-off date for the analysis is 29 April 2011.
As well as a common overlapping range of dates, analysing price discovery requires the
assessment of the contemporaneity in price changes over a common overlapping intraday trading
Security Average Daily Volume
(000's tCO2e)
Average Monthly Volume
(000's tCO2e)
Data Availability (Phase 2)
ICE EUA Futures 5,459 114,497 2 Jan 08 - 29 Apr 11
BNX EUA Spot 2,400 49,162 20 Oct 08 - 29 Apr 11
ICE CER Futures 410 8,493 14 Mar 08 - 29 Apr 11
ICE EUA Spot 395 8,023 13 Mar 09 - 19 Jan 11
GRX EUA Futures 292 6,009 4 Aug 10 - 29 Apr 11
EEX EUA Futures 210 4,411 2 Jan 08 - 19 Jan 11
NOX EUA Futures 118 2,484 2 Jan 08 - 29 Apr 11
BNX CER Spot 85 1,749 20 Oct 08 - 29 Apr 11
MCX CER Futures 31 655 9 Jun 08 - 25 Feb 09
NOX CER Futures 29 605 2 Jan 08 - 29 Apr 11
NOX EUA Spot 22 470 2 Jan 08 - 29 Apr 11
ICE CER Spot 18 358 14 Mar 08 - 29 Apr 11
ICE EUA Options 7 146 19 Jan 09 - 29 Apr 11
BNX ERU Spot 7 130 3 Dec 10 - 29 Apr 11
MCX EUA Futures 3 62 21 Jan 08 - 13 Dec 08
GMX EUA Spot 2 37 20 Oct 09 - 30 Dec 10
GRX CER Futures 1 23 14 Jul 09 - 29 Apr 11
BNX EUA Futures 1 16 16 Oct 08 - 29 Apr 11
EEX CER Futures 1 14 26 Mar 08 - 29 Apr 11
EEX EUA Spot 0 7 22 Jan 09 - 29 Apr 11
NOX CER Spot 0 2 20 Nov 09 - 29 Apr 11
GMX CER Spot 0 1 20 Oct 09 - 30 Dec 10
EXA EUA Spot 0 0 22 Apr 08 - 21 Dec 10
GRX EUA Spot 0 0 14 Apr 11 - 29 Apr 11
BNX CER Futures 0 0 12 Oct 08 - 29 Apr 11
NOX EUA Options 0 0 7 Jun 10 - 29 Apr 11
EEX EUA Options 0 0 8 Apr 08 - 29 Apr 11
GRX EUA Options 0 0 7 May 10 - 29 Apr 11
NOX CER Options 0 0 7 Jun 10 - 29 Apr 11
ICE CER Options 0 0 12 Mar 08 - 29 Apr 11
GRX CER Options 0 0 7 May 10 - 29 Apr 11
19
window22
. This window is set to accommodate the security with the shortest opening hours
throughout the sample, which is from 8:00am to 4:00pm GMT. Table 2.3 shows the percentage of
trades occurring at different hours of the day expressed in GMT, with the shaded area indicating the
chosen intraday window23
. This window encompasses at least 73 per cent of activity across the
exchanges.
Table 2.3
Trade Occurrence by Hour of the Day
Trade data sourced from Thomson Reuters Tick History. Shaded area contains the chosen intraday window
used in subsequent analysis. Percentages calculated over the 1 July 2009 to 30 December 2010 sample period.
22
The four different exchanges operate in two different time zones: Central European Time (CET) for BNX,
EEX and NOX; and, Greenwich Mean Time (GMT) for ICE. It should be noted that the UK and Western
Europe both move in and out of daylight saving simultaneously and so, even though this varies their GMT
opening hours, it does not affect the contemporaneity of the analysis when time is measured in GMT.
23 Although it appears that a common overlapping window could be set from 6:00am to 5:00pm GMT, this is
not the common window across the entire sample period as some exchanges only increased their trading hours
as the EU ETS developed. EEX EUA Futures has the shortest intraday window in the early months of the
sample, with its first trades usually after 7:00am GMT and last trades typically just prior to 4:00pm GMT.
Time (GMT)
ICE EUA
Futures
BNX EUA
Spot
ICE CER
Futures
ICE EUA
Spot
EEX EUA
Futures
NOX EUA
Futures
BNX CER
Spot
0:00 - 1:00 0% 0% 0% 1% 0% 0% 0%
1:00 - 2:00 0% 0% 0% 0% 0% 0% 0%
2:00 - 3:00 0% 0% 0% 0% 0% 0% 0%
3:00 - 4:00 0% 0% 0% 0% 0% 0% 0%
4:00 - 5:00 0% 0% 0% 0% 0% 0% 0%
5:00 - 6:00 0% 0% 0% 0% 0% 0% 0%
6:00 - 7:00 6% 1% 9% 1% 3% 5% 1%
7:00 - 8:00 11% 5% 12% 4% 9% 11% 4%
8:00 - 9:00 11% 10% 10% 9% 11% 11% 9%
9:00 - 10:00 10% 13% 11% 9% 9% 8% 11%
10:00 - 11:00 8% 12% 7% 9% 7% 7% 9%
11:00 - 12:00 7% 11% 7% 9% 7% 7% 8%
12:00 - 13:00 8% 11% 8% 11% 9% 8% 9%
13:00 - 14:00 10% 11% 9% 12% 11% 10% 10%
14:00 - 15:00 11% 12% 10% 13% 12% 12% 14%
15:00 - 16:00 12% 10% 11% 12% 13% 11% 14%
16:00 - 17:00 5% 2% 5% 6% 8% 7% 3%
17:00 - 18:00 0% 0% 0% 1% 1% 3% 1%
18:00 - 19:00 0% 0% 0% 0% 0% 0% 0%
19:00 - 20:00 0% 1% 0% 0% 0% 0% 6%
20:00 - 21:00 0% 0% 0% 0% 0% 0% 0%
21:00 - 22:00 0% 0% 0% 0% 0% 0% 0%
22:00 - 23:00 0% 0% 0% 0% 0% 0% 0%
23:00 - 24:00 0% 0% 0% 1% 0% 0% 0%
Per cent Within Window 77% 91% 74% 84% 79% 73% 85%
20
In sum, the final sample includes seven of the eight most traded securities in Phase 2 of the EU
ETS, runs for 376 days from 1 July 2009 to 30 December 2010 and contains intraday prices
between 8:00am and 4:00pm GMT for each trading day.
2.2.2 Measures of Trading Cost
Assessing the strengths of the trading cost hypothesis necessitates measures of trading cost for
each security. The explicit cost of transacting in the EU ETS involves brokerage and clearing fees.
Like most brokerage arrangements, the amount of brokerage a market participant pays depends
upon their particular relationship to their broker and so cannot be directly assessed, though these are
likely to be small compared to the implicit costs of trading24
. Although a number of metrics could
be used to measure implicit trading costs such as those describing the market impact of trades, or
market depth and breadth, we employ the most commonly used measure of implicit trading costs in
the finance literature namely the bid-ask spread. Although what drives the size of the bid-ask
spread is interesting, for this study it is sufficient simply to measure the spread so that the actual
cost faced by traders can be used in evaluating whether trading costs dictate a preference to trade
particular securities25
. The spread is calculated as the ask price minus the bid price.
Table 2.4 ranks the instruments by the average size of the bid-ask spread sampled at 5-minute
intervals over the 1 July 2009 to 30 December 2010 period. Consistent with our expectations, the
seven securities are ranked in essentially the same order as that dictated by trade volume. Only the
ranking of the ICE CER Futures and ICE EUA Spot series switch places when ranking by implicit
trading cost, but it should be noted that these securities have fairly similar trade volumes. Ordering
24
The costs faced by brokers can be assessed. According to information provided by the exchanges, clearing
fees are typically between €0.003 and €0.010 per allowance (or €3 to €10 per 1,000 tCO2e trade lot). The
exchanges charge brokers different clearing fees depending on their level of membership on a particular
exchange and these fees are often comprised of fixed components per trade as well as variable, volume based
components.
25 There is a large literature examining different aspects of the bid-ask spread, from Demsetz (1968) to the
multitude of market microstructure papers since Copeland and Galai (1983) and Glosten and Milgrom (1985)
steered the focus toward the informational aspects of the spread. Although dissecting the spread into
components representing not just the inventory costs but also adverse selection costs in the presence of
informed traders is interesting, it is beyond the scope of this paper.
21
by spread gives the ranking of price discovery that would be expected if the Trading Cost
Hypothesis is a better explanation of what motivates traders to use particular instruments26
.
Table 2.4
Bid-Ask Spread, Trade Volume and Expected Ordinal Ranking by Hypothesis
Trade and quote data sourced from Thomson Reuters Tick History. Average bid-ask spread measured using
intraday price data sampled at a 5-minute frequency. All data measured over the 1 July 2009 to
30 December 2010 sample period. Note that the average daily volume data differs slightly from that reported
in Table 2.2 as this is now measured over the common sample period rather than over the securities‘
individual periods of data availability.
2.2.3 Return Series Construction
The arrival of new information may prompt market participants to respond by executing market
orders or by altering existing limit orders. The execution of market orders is also likely to affect the
limit order book insomuch as best bid or ask limit orders are partially or completely filled by these
trades. Both of these aspects of information arrival are examined by the creation of two types of
return series. More specifically, the first measure uses the last trade prices over a given interval,
while the second is created from changes in the mid-point of the bid-ask spread. In each case,
continuously compounded returns are calculated by taking log first differences in price levels.
26
On the other hand, if the Leverage Hypothesis is the better explanation, it would be expected that the
futures instruments would display greater price discovery than the spot instruments and, inasmuch as trading
cost is a secondary consideration, price discovery would follow the ranking given in the second last column of
Table 2.4. Finally, the rankings in the last column would be those expected under the Market Segmentation
Hypothesis, again with trading cost as the secondary determinant.
Average Average Average Price Price Price
Daily Daily Bid-Ask Discovery Discovery Discovery
Volume Volume Spread Ranked by Ranked by Ranked by
Security (000's tCO2e) (% of total) (€/tCO2e) Trading Cost Leverage Segmentation
ICE EUA Futures 6,772 71.10% 0.025 1 1 1
BNX EUA Spot 1,362 14.30% 0.055 2 5 2
ICE EUA Spot 448 4.70% 0.065 3 6 3
ICE CER Futures 518 5.44% 0.078 4 2 6
EEX EUA Futures 263 2.76% 0.081 5 3 4
BNX CER Spot 88 0.92% 0.122 6 7 7
NOX EUA Futures 74 0.78% 0.282 7 4 5
22
Bid and ask prices change with greater frequency than trades occur, giving the Mid Point return
series more non-zero return observations. Table 2.5 displays the percentage of intraday time
intervals without a price change (zero returns) for both types of return series.
Table 2.5
Percentage of Zero Returns
Returns calculated from trade and quote data sourced from Thomson Reuters Tick History. The zero returns
are the intraday intervals over which prices were unchanged. The number of these incidences are summed and
expressed as a percentage of the total number of intraday intervals in the sample period from 1 July 2009 to
30 December 2010.
Due to both the infrequency of trading when assessed over 1-minute intervals and the declining
ability to detect differences in price discovery over longer time intervals, much of the analysis
focuses on the 5-minute time series. Where there are notable differences in the analysis for the other
series, these differences are reported. For the 5-minute time series, there are 97 intraday return
observations for each of the 376 trading days, which gives a total of 36,472 observations in the
sample. Descriptive statistics for the 5-minute return time series are displayed in Table 2.6.
Large return observations have been investigated. Returns that appeared to be generated by
miss-reported prices were removed from the data, such as where it appeared that the trade volume
was reported in the price field27
. As shown in the maximum and minimum returns in Table 2.6,
some large return observations remain for the EEX EUA Futures Last Trade return series and the
BNX CER Spot Mid Point return series that appear out of line with the maximum and minimum
returns for the other securities. These returns were kept as they stem from infrequent trading and
27
This led to the further removal of three return observations from the Last Trade NOX EUA Futures series
and one each from the ICE CER Futures, EEX EUA Futures and BNX CER Spot.
Security 60-min 10-min 5-min 1-min 60-min 10-min 5-min 1-min
ICE EUA Futures 7% 22% 36% 72% 3% 12% 21% 53%
BNX EUA Spot 14% 54% 70% 92% 4% 15% 26% 60%
ICE EUA Spot 30% 76% 86% 97% 4% 16% 28% 61%
ICE CER Futures 29% 75% 86% 97% 5% 24% 37% 69%
EEX EUA Futures 60% 89% 94% 99% 10% 25% 35% 66%
BNX CER Spot 65% 92% 96% 99% 7% 32% 46% 74%
NOX EUA Futures 58% 90% 95% 99% 30% 66% 78% 94%
Panel B: Mid Point Return SeriesPanel A: Last Trade Return Series
23
genuine illiquidity driving large changes in the bid-ask spread28
. Consistent with Table 2.5, the
inter-quartile return observations also illustrate the infrequent trading, especially in the Last Trade
return series, where most of the 25th and 75
th percentile returns are zero, indicating that less than 50
per cent of the intervals in these series had observed price changes. This provides further
justification for the use of the Mid Point return series.
Table 2.6
Descriptive Statistics
Returns calculated from trade and quote data sourced from Thomson Reuters Tick History. Descriptive
statistics for each series are calculated for 36,472 return observations over the 1 July 2009 to
30 December 2010 sample period.
2.2.4 Serial Correlation
Serial correlation is often a problem for statistical analysis involving financial time series data.
Research in this area going back to Niederhoffer and Osborne (1966), Blume and Stambaugh
28
As a precaution, some of the analysis was conducted with these large observations removed but this had a
negligible effect on the results.
Panel A:
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
Mean 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
Standard Deviation 0.0019 0.0018 0.0018 0.0019 0.0021 0.0017 0.0022
Skewness -0.73 -0.72 -0.90 -0.42 2.28 -1.37 -0.47
Kurtosis 52.16 85.16 95.44 77.85 708.81 238.28 388.03
Maximum 0.0314 0.0350 0.0454 0.0401 0.1126 0.0452 0.0749
75th
Percentile 0.0007 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
25th
Percentile -0.0007 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
Minimum -0.0614 -0.0616 -0.0571 -0.0610 -0.0820 -0.0732 -0.0755
Panel B:
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
Mean 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
Standard Deviation 0.0016 0.0017 0.0017 0.0018 0.0017 0.0023 0.0032
Skewness -1.40 0.34 -1.22 -0.42 -1.27 2.30 -0.37
Kurtosis 100.86 122.75 94.69 79.20 157.22 320.10 78.18
Maximum 0.0317 0.0445 0.0291 0.0437 0.0393 0.1150 0.0587
75th
Percentile 0.0006 0.0004 0.0004 0.0004 0.0004 0.0004 0.0000
25th
Percentile -0.0004 -0.0004 -0.0004 -0.0004 -0.0004 -0.0004 0.0000
Minimum -0.0625 -0.0609 -0.0620 -0.0534 -0.0668 -0.0715 -0.0749
Last Trade Return Series
Mid Point Return Series
24
(1983) and Roll (1984) attributes the negative serial correlation observed in many financial time
series to bid-ask bounce or thin trading. Both of these factors may contribute to serial correlation in
the return series for the EU ETS securities. Table 2.7 contains autocorrelation coefficients ( ji )
for the 5-minute Last Trade and Mid Point return series out to 10 lags29
.
Table 2.7
Autocorrelation Functions
Returns calculated from trade and quote data sourced from Thomson Reuters Tick History. Autocorrelation
functions estimated over the 1 July 2009 to 30 December 2010 sample period.
29 Coefficients calculated as per:
jjR
RR
jT
t
j
i
jit
it
T
t
t
i 2
1 1
1
ˆ
ˆˆ
ˆ
where:
T
t
tRT 1
1
Panel A:
ρ1 ρ2 ρ3 ρ4 ρ5 ρ6 ρ7 ρ8 ρ9 ρ10
ICE EUA Futures -0.0571 -0.0360 -0.0089 -0.0039 0.0019 -0.0055 0.0042 0.0009 0.0073 0.0027
BNX EUA Spot 0.0203 0.0064 0.0023 -0.0137 -0.0022 -0.0094 -0.0106 0.0063 -0.0058 -0.0020
ICE EUA Spot 0.0009 0.0107 0.0042 0.0041 -0.0028 0.0044 -0.0031 0.0012 0.0132 0.0021
ICE CER Futures -0.0164 -0.0154 -0.0130 -0.0069 -0.0017 -0.0002 -0.0030 0.0088 -0.0006 -0.0076
EEX EUA Futures -0.0100 0.0122 0.0061 -0.0004 0.0032 -0.0054 0.0019 -0.0007 -0.0032 -0.0026
BNX CER Spot 0.0052 0.0074 0.0010 -0.0015 0.0008 -0.0003 -0.0005 0.0124 -0.0006 0.0065
NOX EUA Futures 0.0030 0.0006 -0.0219 0.0010 -0.0115 0.0066 -0.0289 -0.0008 0.0030 0.0048
Panel B:
ρ1 ρ2 ρ3 ρ4 ρ5 ρ6 ρ7 ρ8 ρ9 ρ10
ICE EUA Futures 0.0288 -0.0031 -0.0039 0.0102 0.0094 -0.0042 0.0090 0.0000 0.0104 -0.0048
BNX EUA Spot -0.0130 0.0083 0.0035 0.0065 0.0075 -0.0107 -0.0149 0.0011 0.0055 -0.0014
ICE EUA Spot 0.0122 0.0056 0.0008 0.0070 0.0090 0.0040 -0.0064 0.0077 0.0018 0.0037
ICE CER Futures -0.0543 -0.0071 -0.0049 -0.0052 0.0092 0.0015 -0.0011 0.0085 0.0111 -0.0009
EEX EUA Futures 0.0211 -0.0149 -0.0171 0.0087 0.0017 -0.0003 -0.0042 0.0032 0.0032 -0.0027
BNX CER Spot -0.1828 -0.0163 0.0217 -0.0561 0.0614 -0.0420 -0.0194 -0.0018 0.0132 -0.0046
NOX EUA Futures -0.1145 -0.0446 -0.0373 -0.0174 -0.0082 -0.0060 -0.0152 0.0030 -0.0140 0.0000
Mid Point Return Series
Last Trade Return Series
25
In order to test the overall significance of this serial correlation, joint test statistics are
calculated applying the weighting methodology in Richardson and Smith (1994) to the random walk
tests of Fama and French (1988) and Lo and MacKinlay (1988). Box and Pierce (1970) Q-statistics
are also calculated30
. In representing the Fama and French (1988) regression beta statistics and the
Lo and MacKinlay (1988) variance ratios, weights ( iD ) are applied to the autocorrelation
coefficients ( ji ). For the regression beta statistics, the intermediate autocorrelation coefficients
are more heavily weighted, while for the variance ratio the weights decline monotonically over
successive coefficients31
. Richardson and Smith (1994) use Hansen‘s (1982) result to show that the
asymptotic distribution of the estimated autocorrelation coefficients ( j ) is given by:
INjjTjTa
lk ,0~ˆˆˆ
(2.1)
Statistics are created from the sum of the weighted coefficients:
jDjD i
j
i
i ˆˆ1
(2.2)
Using (2.1) and (2.2) and the fact that linear combinations of normal distributions are also
normal, 2 test statistics are created as per:
21~ˆˆ
,0~ˆ
jT jDDDjDTJ
DDNjDT
(2.3)
30
Given the very large sample size (T ), Box-Pierce (1970) and Ljung-Box (1978) Q-statistics are practically
identical and so the choice of Box-Pierce (1970) is largely arbitrary. Test statistics calculated by:
j
i
ji jTQ1
22~ˆˆ
~
31 Weights are: 1iD
for the Box and Pierce (1970) Q-statistics, jijiDi 2,min for Fama and French
(1988) and: jijDi /2 for Lo and MacKinlay (1988), where the total number of autocorrelation
coefficients ( i ), is 10j for the Q and variance ratio statistics, but 9j for the regression beta statistics,
which require an odd number of lags to create the heaviest weighting on the central coefficient.
26
Depending on the test, Table 2.8 shows several series have statistically significant serial
correlation under the joint tests. Richardson and Smith (1994) conclude that tests utilizing declining
weights are more powerful against an alternative hypothesis of mean reversion. The variance ratio
tests fit this description. For the Last Trade return series, the variance ratio tests show the greatest
serial correlation in the ICE EUA Futures, which from Table 2.7 appears largely driven by the
negative first and second lag autocorrelation coefficients. Given the heavy trading of this security,
this result is likely symptomatic of bid-ask bounce. For the Mid Point return series, the variance
ratios also indicate serial correlation in BNX CER Spot and NOX EUA Futures. In accordance with
Table 2.7 showing strong autocorrelation at the first lag, the serial correlation in these securities‘
returns is likely induced by their being thinly traded.
Table 2.8
Joint Serial Correlation Tests
Weighted autocorrelation coefficient estimates ( jD ) and2 test statistics calculated as per the unified
approach in Richardson and Smith (1994). * and ** denote significance at the 5 and 1 per cent levels against 2 critical values with 9 degrees of freedom for Fama and French (1988) beta statistics and 10 degrees of
freedom for the Box and Pierce (1970) Q-statistics and the Lo and MacKinlay (1988) variance ratios.
Given the small magnitude of the coefficients in Table 2.7, the serial correlation evident from
the tests reported in Table 2.8 is unlikely to have a large impact on the regressions undertaken in
assessing price discovery. Nevertheless, to deal with bias induced by this serial correlation, the
Q-statistic Beta Statistic Variance Ratio Q-statistic Beta Statistic Variance Ratio
ICE EUA Futures 0.0048** -0.0324 -0.1757** 0.0013** 0.0239 0.0672
BNX EUA Spot 0.0009** -0.0177 0.0188 0.0007** -0.0004 -0.0059
ICE EUA Spot 0.0004 0.0122 0.0315 0.0004 0.0226 0.0522
ICE CER Futures 0.0009** -0.0230 -0.0809* 0.0033** -0.0054 -0.1068**
EEX EUA Futures 0.0004 0.0052 0.0074 0.0011** -0.0042 0.0015
BNX CER Spot 0.0003 0.0085 0.0259 0.0434** -0.0569** -0.3740**
NOX EUA Futures 0.0015** -0.0348 -0.0464 0.0173** -0.1008** -0.3743**
Panel A: Last Trade Return Series Panel B: Mid Point Return Series
27
regressions are run using the Newey-West (1987) method for calculating standard errors in the
presence of potential heteroskedasticity and autocorrelation in residuals.
2.2.5 Stationarity and Cointegration
The Methodology section that follows describes two techniques to assess price discovery. One
of these methodologies, the calculation of Hasbrouck‘s (1995) information shares, requires that we
establish that the series examined are I(1) variables—non-stationary in log price levels (following a
random walk), but stationary in returns (mean reverting). Augmented Dickey-Fuller (1979) tests for
the stationarity of log price levels and returns for the 5-minute series are displayed in Table 2.9.
Table 2.9
Augmented Dickey-Fuller Test Statistics
Panel A displays Augmented Dickey-Fuller test statistics for the 5-minute Last Trade log price level and
return series. Panel B presents results for the Mid Point log price level and return series. The unit root tests
are run with a constant ( ) and 10 lags of differenced dependent variables as explanatory variables
( 10k ) as per:
tjt
k
j
jtt yyy
1
1
The dependent variables ( ty ) are alternately differenced log price levels and differenced returns. The test
statistic ˆˆtZ is for 0:0 H , where is the standard error of . * and ** denote
significance at the 5 and 1 per cent levels against critical values from Fuller (1996) of -2.86 and -3.43,
respectively.
Panel A:
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
Levels -2.73 -2.72 -2.77 -2.55 -3.07* -3.29* -3.27*
Returns -58.35** -59.02** -56.66** -58.40** -57.61** -60.74** -61.76**
Panel B:
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
Levels -2.71 -2.62 -2.65 -2.37 -2.59 -2.69 -4.13**
Returns -56.66** -58.13** -56.77** -58.19** -58.24** -60.88** -61.88**
Last Trade Return Series
Mid Point Return Series
28
The results in Table 2.9 clearly show that all the securities‘ returns are stationary as expected
and that most are non-stationary in price levels. NOX EUA Futures is the only exception which is
likely a result of a lack of variation stemming from its being the most thinly traded. Similar results
are obtained for the data sampled at other frequencies32
.
Cointegration tests are also conducted on the log price level series as these are utilised in the
calculation of the information shares. The securities are expected to be cointegrated as their prices
are driven by the value the same asset. Given the augmented Dickey-Fuller (1979) test results
showing NOX EUA Futures as stationary in log price levels, there may be only five cointegrating
relations between the seven securities rather than the six that would be expected if all variables were
non-stationary in levels. Table 2.10 details the results of conducting Johansen (1995) tests for the
number of cointegrating relations between the log price levels sampled at a 5-minute frequency.
The trace statistics displayed in Table 2.10 show that the null hypothesis of no more than five
cointegrating relations cannot be rejected at the 1 per cent level of significance. These results are
unchanged when alternative cointegration tests are conducted using maximum-eigenvalue statistics
and the minimisation of information criteria33
. The results are also the same at other sampling
frequencies, with the exception of the hourly data series, which indicate there are four cointegrating
relations in the Last Trade series and only two in the Mid Point data series. These results for the
hourly series are of little concern as the calculation of information shares is most appropriate on
higher frequency data in which there is likely to be less contemporaneous correlation between error
terms. As such the information share calculations displayed in the results section are based on the
1-minute sampling of log price levels.
32
These results are largely the same for different lag specifications in the independent variables (up to 40
lags) and are unchanged by alternatively running Phillips-Perron (1988) unit root tests of up to 40 lags.
33 These results are available from the author on request.
29
Table 2.10
Cointegration Tests
The Johansen (1995) cointegration tests are run using a multivariate Vector Error Correction Model (VECM)
estimated by maximum likelihood with 10 lags ( 10n ) as explanatory variables to determine the number of
cointegrating relations ( r ) between the seven ( K ) variables:
tit
n
i
itt εyΓyβαy
1
1
1
Dependent variables ( ty ) are a 1K vector of differenced log price levels sampled at 5-minute intervals
from 1 July 2009 to 30 December 2010; ity are lagged dependent variables; α andβ are rK
parameter matrices in which the number of cointegrating equations is less than the number of I(1) variables
( Kr ); 11 ,, pΓΓ are KK matrices of parameters; and, tε is a 1K vector of normally distributed
and serially uncorrelated error terms with contemporaneous covariance matrix Ω . The null hypothesis for the
trace statistic is that there are no more than r cointegrating relations (i.e. the eigenvalues Kr ,,1 are
zero). ** denotes the rank at which the null hypothesis cannot be rejected at the 1 per cent level of statistical
significance.
Maximum Rank Eigenvalue Trace Statistic Eigenvalue Trace Statistic
r ≤ 0 2,660.09 1,633.16 133.57
r ≤ 1 0.03266 1,446.28 0.01988 899.00 103.18
r ≤ 2 0.02108 667.31 0.01886 202.92 76.07
r ≤ 3 0.00955 316.31 0.00333 81.10 54.46
r ≤ 4 0.00501 132.74 0.00117 38.21 35.65
r ≤ 5 0.00325 13.75** 0.00079 9.23** 20.04
r ≤ 6 0.00023 5.37 0.00019 2.25 6.65
r ≤ 7 0.00015 0.00006
Panel A: Last Trade Series Panel B: Mid Point Series 1% Critical
Value
30
2.3 Methodology
This section describes the methodology employed to investigate price discovery. Common
approaches to assessing short-run dynamics and contemporaneity involve regressing returns of one
series against leading, lagging and contemporaneous returns of another or the use of Vector Error
Correction Models (VECMs), which have the added benefit of addressing cointegration in price
levels and serial correlation in the dependent variable. Our methodology is a mixture of these two
approaches, though for robustness we also calculate Hasbrouck‘s (1995) information shares in a
multivariate VECM framework, which measures the contribution of each security‘s variance to
innovations in the long-run equilibrium price common to them all.
2.3.1 Basic Regression Specification
If two securities are perfect substitutes for one another, or are identically affected by the same
information, their prices should change simultaneously in a frictionless market. Similarly, given
minimal short-term changes in carrying costs, the prices of derivative securities should
simultaneously change to reflect information regarding the value of underlying assets. In such
markets, a regression of the returns of one of these securities against leads, lags and the
contemporaneous returns of another would be expected to show a regression beta close to one on
the contemporaneous return observations and zero on the leading and lagging returns (assuming
there is no serial correlation in the returns).
However, where market frictions such as transaction costs exist, or where the market is
otherwise segmented in some respect, this contemporaneous relationship may be weaker and
aspects of the market frictions or segmentation may determine one instrument‘s use in preference to
another‘s. By way of example, under the Trading Cost Hypothesis, securities with lower trading
cost are preferred by traders seeking to profitably exploit new information and it would be expected
that the prices of these low trading cost securities would thus be impounded with new information
more quickly. As such, returns of the security with the lowest trading cost would, on average, be
31
expected to lead those of higher trading cost securities. The same rationale could be applied to
preferences for leverage under the Leverage Hypothesis or the use of EUAs in preference to CERs
where limitations and uncertainty concerning the use of CERs segments the emission allowance
market (the Market Segmentation Hypothesis).
If only a small number of securities are examined, visual inspection will easily reveal whether
the leads have larger, more significant coefficients than the lags, making it straight forward to draw
conclusions about the relative strength and direction of price discovery between the securities34
.
This is largely the approach Stoll and Whaley (1990), Chan (1992) and Fleming et al. (1996) take in
looking at price discovery in stock indices and stock index derivatives. These models are broadly of
the following form35
:
tktB
k
ktA RR
,
10
10
,
(2.4)
As the number of instruments under examination increases, it is no longer practicable to
compare the bilateral regression results solely by visual inspection of the coefficients and their t-
statistics as it will likely become difficult to establish a distinct order in which securities have a
propensity to lead or lag others36
. Visual inspection is particularly problematic where there is not
perfect consistency in the order of these relationships. To deal with this problem, we want to use
measures of overall fit to describe the strength of the leading or lagging relationship. For each
dependent variable we then rank the independent variables in order of these measures of fit and
compare this to the order we would expect to find under the three hypotheses.
34
For example, if the number of securities ( n ) is three, they can be bilaterally compared by combining them
in only three different ways: 3)!!(!!)2,3(),( rnrnCrnC .
35 Although the specific models used in the cited literature differ in the number of leads and lags employed,
the model in (2.4) specifies 10k leads and lags, which allows for the assessment of price discovery of up to
50 minutes either side of the contemporaneous return when using data sampled over 5-minute intervals. The
choice of 10 lags is somewhat arbitrary, though it should be noted that by also running regressions on data
sampled at 1-minute, 10-minute and 60-minuite intervals we explore the appropriateness of this choice.
36 For example, comparing bilateral regressions of seven securities as per equation (2.4) requires comparing
21 different sets of lead, lag and contemporaneous coefficients: 21)!!(!!)2,7(),( rnrnCrnC .
32
Given the widespread use of R-squared as a goodness of fit measure, we too employ this as part
of the analysis. However, given our concerns regarding serial correlation, we also employ the robust
F-statistics generated from the Newey-West (1987) estimation of the variance-covariance matrix
when making comparisons. Newey-West (1987) estimation does not preclude the calculation of
traditional R-squared statistics from the total sum of squares and regression sum of squares,
however, these will not reflect nor take advantage of the robustness adjustments made to the
estimated variance-covariance matrix. The robust F-statistic is not calculated in the traditional
ordinary least squares (OLS) manner, but instead is a Wald statistic (W ) scaled by the number of
restrictions ( q ) imposed in jointly testing whether the estimated coefficients ( ) are zero (i.e.
rRH :0 for 0r )37
:
qWF
VrRRVRrRW
Statistic
ˆˆˆˆ 11
(2.5)
Where 1V is the inverse of the robust variance-covariance matrix of residuals,
is a
11 k vector of regression parameters, k is the number of estimated coefficients and R is a
1 kq matrix (see Wooldridge, 2009).
Utilising these statistics of the goodness of fit and joint significance of regression coefficients
poses another minor issue that requires a slight alteration to the model in equation (2.4). Because
the regressions in (2.4) contain both leading and lagging returns as independent variables, the use of
R-squared and F-statistics will not disentangle whether the relative strength of price discovery lies
37
Under OLS estimation of the residual variance-covariance matrix, this Wald statistic approach to
calculating F-statistics is identical to the traditional OLS approach:
1
knSSE
kRSSFStatistic
However, they will differ where adjustments are made to the variance-covariance matrix under Newey-West
(1987).
33
in the leading or the lagging coefficients (as they relate to the overall model). For example, the R-
squared from regressions of A on B and then B on A would be similar in magnitude. To address this
problem, we regress returns of a given series against only the contemporaneous and lagging returns
of another series, thereby omitting the leading returns, which are also incorporated in the models of
Stoll and Whaley (1990), Chan (1992) and Fleming et al. (1996). Formally, we fit the model:
tktB
k
ktA RR
,
0
10
,
(2.6)
If series A leads series B, the independent variables (B) will do a poor job of explaining the
returns of the dependent variable (A) and the regression will have a low R-squared and a low F-
statistic. However, when the regression is run in reverse, with series B returns as the dependent
variable and contemporaneous and lagging returns of series A as the independent variables, the R-
squared and F-statistic will be higher. Given our study focuses on price discovery in seven series,
our approach yields 42 permutations of the model in (2.6)38
.
2.3.2 Cointegration and Error Correction
Another common approach to analysing price discovery involves the use of VECMs (see, for
example, Schwarz and Szakmary, 1994, Harris, McInish, Shoesmith and Wood, 1995, Booth et al.
1999, and Hsieh et al. 2008). A simple VECM would simultaneously run the two regressions
described in the preceding section above (A on B and B on A) but would include lags of the
dependent variable as independent variables, to deal with any serial correlation, and would also
include an error correction variable to account for cointegration and long-run equilibrium dynamics
38
There are: 42)!!(!),( rnnrnP permutations of the bilateral regressions that can be run using (2.6). It
should be noted that, although the number of permutations in which the regressions can be run is twice the
number of combinations under the approach in equation (2.4), we can now assess price discovery using a
couple of statistics, rather than having to compare each coefficient and its t-statistic in every regression
individually. This is also less subjective than visual inspection.
34
in price levels39
. A typical VECM specification would also omit contemporaneous returns as
independent variables. An example of such a model is as follows:
ttzBktB
k
kBktA
k
kAtB
ttzAktA
k
kAktB
k
kBtA
zRRR
zRRR
1,,
1
10
,,
1
10
,,
1,,
1
10
,,
1
10
,,
(2.7)
The error correction term ( 1tz ) is calculated as the difference in log prices of the two series at
a one-period lag40
:
1,1,1 lnln tBtAt PPz
(2.8)
The prices of two securities that are perfect substitutes in frictionless markets would be
cointegrated and track one another‘s movements. Regressions involving such securities would have
zero coefficients on the error correction variable because prices would all be changing
contemporaneously. For the reasons discussed previously, this may not be the case if market
frictions or segmentation exists. The inclusion of lagging return observations facilitates
measurement of whether these frictions cause short-term price changes to be non-contemporaneous
in such cases. In light of this, the error correction terms may seem somewhat redundant alongside
lagged returns. Harris et al. (1995) argue that error correction is still necessary where there is a
possibility that one variable within the system of cointegrated variables is independent of the error
correction process. This seems unlikely for securities like those in the EU ETS which have such
close relationships in their use and fungibility but, to ensure consistency with the literature and in
39
More than two series can be simultaneously examined in the VECM framework; this example uses two
series to be consistent with the example in the last section.
40 Error correction terms at greater than one lag would also be applicable inasmuch as there is some
justification that the lag at which one price catches up with movements in another is expected to occur after
some more distant amount of time than a single interval. Hasbrouck (1995) points out that there are infinite
possible error correction representations as traders may respond to the ‗error‘ or price discrepancy at any
possible combination of lags. However, in reasonably liquid markets, it would be expected that trader
responses would be quite fast, either within the same contemporaneous interval or at most shortly afterwards.
35
light of the stationarity and cointegration test results concerning our thinly traded NOX EUA
Futures series, a one-period error correction term is included in our model. This term is expected to
be statistically significant, inasmuch as the emission allowance market is not perfectly frictionless,
and is expected to give an indication that there is a long-run equilibrium relationship between the
securities‘ price levels.
2.3.3 The Final Model Specification
Our model involves regressions of returns from one series against contemporaneous and lagged
returns of another as well as a one-period error correction term. In this respect, our model
specification is closest to that of Fleming et al. (1996). However, due to the large number of
securities examined in our study and the desire to isolate the ordinal ranking of leading and lagging
relationships, we depart from their methodology by excluding leading returns as independent
variables. Formally, our model is described as follows:
ttAzAktB
k
ktA zRR
1,,,
0
10
,
(2.9)
As previous analysis has indicated some serial correlation exists in our data, Newey-West‘s
(1987) method is employed in estimating the variance-covariance matrix to improve the robustness
of the standard errors. Specifically, the Newey-West (1987) standard errors in this analysis are
calculated for serial correlation up to 10 lags41
. The 42 permutations in which the returns of each
security can be used as an explanatory variable for each other security, combined with the use of
two distinct return metrics (Last Trade and Mid Point) and data calculated at four different intraday
intervals (1-minute, 5-minute, 10-minute and 60-minute), result in us running a total of 336
regressions (42×2×4 = 336).
41
A smaller number of lags could have been used, given that for all the series the autocorrelation coefficients
become very small after the first few lags (see Table 2.7), but to err on the side of conservatism, 10 was
chosen. Using 10 lags is close to the rule of thumb suggested by Newey-West (1987), which for our sample
is: 82.13472,3644 TL
36
2.3.4 Information Shares
For robustness we also calculate Hasbrouck‘s (1995) information shares measure under a
multivariate VECM framework. Where securities representing the same asset trade in different
markets, information shares measure the relative contribution of each market to the variance of
innovations in the common factor between them. As opposed to transitory deviations between
securities generated by frictions that are idiosyncratic to each securities‘ market, this common factor
is the permanent innovation in prices common to all the securities and is thus an unobserved, but
implicitly efficient, price (see Hasbrouck, 2002). This decomposition of actual prices ( tip , ) into an
unobservable common efficient price ( tm ), which follows a random walk ( ttt umm 1 ), and
idiosyncratic transitory factors ( tis , ), is given by Hasbrouck (2002) as:
tk
t
t
tk
t
t
s
s
m
p
p
,
,1
,
,1
1
1
y
(2.10)
Extending the bi-variate VECM from equation (2.7) to a multivariate specification similar to
that used in the Johansen (1995) cointegration tests gives:
tit
n
i
itt εyΓyβαy
1
1
1
(2.11)
The dependent variables ( ty ) are a 1K vector of first differences in log price levels ( tip , );
ity are lagged dependent variables; α andβ are rK parameter matrices in which the number
of cointegrating equations is less than the number of I(1) variables ( Kr ); 11 ,, pΓΓ are KK
matrices of parameters; and, tε is a 1K vector of normally distributed and serially uncorrelated
37
error terms with contemporaneous covariance matrix Ω 42. Hasbrouck (1995) represents equation
(2.11) as a Vector Moving Average (VMA), where LΨ is a matrix polynomial in the lag operator
L :
tt L Ψy
(2.12)
Following Baillie, Booth, Tse and Zabotina (2002), equation (2.12) is expressed in integrated
form as:
t
t
s
st L *ΨΨy
1
1
(2.13)
The moving average impact matrix 1Ψ is calculated as the sum of the moving average
coefficients, which is the long-run impact of price innovations that are common to all the series.
That is, the VMA is given by:
2211 tttt εψεψεy and 211 ψψIΨ is a
KK vector containing the sum of the ψ coefficients (see Hasbrouck, 2002). Though the
efficient price is unobservable, its variance can be related to the variance of actual prices by:
2
21
2 1 ψψ u as the long-run impact is common to all series, but the idiosyncratic
components are not. The rows of the impact matrix, 1Ψ , are identical. If we denote one of these
rows Ψ (a K1 row vector), we can express the variance of the common efficient price as:
ΨΨΩ 2
u
(2.14)
The information share ( iIS ) of a particular security is then its variance contribution (22
ii ) to
the variance of the common efficient price:
42
Note that the VECM specification in equation (2.11) is run without a time trend in the cointegrating
equations. Though a trend can be included when comparing the log prices of spot and futures securities to
deal with the steady decline in carrying costs through time for futures, the inclusion of a trend had a negligible
impact on the information shares such that they differed at no greater than the 3rd
decimal place compared
with excluding a trend term.
38
ΨΨΩ
22
iiiIS
(2.15)
The result in equation (2.15) relies on the absence of contemporaneous correlation in the error
terms from the VECM. This tends not to hold for strongly cointegrated time series and, even though
this can be mitigated to some extent by increasing the sampling frequency, the results are likely to
remain sensitive to the order in which the series are put into the log price vector ( ty ). Hasbrouck
(1995) uses a Cholesky factorisation of the error term covariance matrix: MMΩ (where M is
the lower triangular matrix) and cycles the ordering of the series in the log price vector to provide a
range for the information shares. This cycling will typically lead to the information share of a series
being greatest when its prices appear first in the log price vector and lowest when it is the last series
in the vector (the first and last positioning in the cycle thus being used as the information share
range for a series)43
. Following Baillie et al. (2002), this adjustment leads to information shares
given by:
ΨΨΩ
2
ii
MIS
(2.16)
For the purpose of utilising information shares as a robustness measure for our results and as a
representation of long-run price discovery, we will take the average information share calculated
from running the VECM in equation (2.11) seven times over which we completely cycle the order
in which the securities appear in the log price vector. All versions of equation (2.11) are run with
five cointegrating relations ( 5r ) commensurate with the results of the cointegration tests
displayed in Table 2.10.
43
This cycling of the order will not always provide the range purely from the first and last positioning if the
sum of the coefficients in the moving average impact matrix for a particular series are negative. As we report
the average of the information shares across all seven cycled positions this should not be an issue even if the
maximum and minimum for a series do not occur when it is ordered first and last, respectively.
39
2.4 Results
This section presents the results of running regressions to assess price discovery. As it is impractical
to report the results of all 336 regressions, a representative selection of the results are presented and
discussed. The regression R-squared and robust F-statistics are used to assess the extent to which
one security‘s returns explain subsequent returns in another. Thereafter, we use these findings on
short-run return dynamics to provide evidence on the three hypotheses regarding price discovery,
namely: the Trading Cost Hypothesis, the Leverage Hypothesis and the Market Segmentation
Hypothesis. The robustness of the regression results is confirmed by the relative contributions to
long-run price discovery evidenced in the information shares.
2.4.1 Regression Results
As noted previously, perfect simultaneous price discovery in a frictionless market would result
in regression coefficients on contemporaneous return observations that are close to one and small
regression coefficients that are insignificantly different from zero on any lagging return
observations. Although these perfectly synchronous relationships are not expected to be found in
the short-run return dynamics in the presence of market frictions, it is nonetheless expected that the
coefficients on contemporaneous returns will be positive and statistically different from zero.
Where market frictions cause traders to use one security in preference to another, the preferred
security‘s returns should lead the other‘s to a greater extent. Lagged returns from this preferred
security should have a positive, statistically significant relationship to the contemporaneous returns
of less preferred securities.
On the basis of all three hypotheses, the most traded security in the EU ETS (ICE EUA Futures)
is expected to be the preferred venue for price discovery, having lowest trading cost, facilitating
leverage and being an EUA security unencumbered by the limitations and uncertainties surrounding
the use of CERs. As such, it is expected to have a stronger leading relationship to the returns of all
the other securities than they do to it.
40
Table 2.11 presents the results of regressing contemporaneous and lagged Last Trade returns of
the ICE EUA Futures (independent variables) against the other six securities (dependent variables),
with the latter listed in order from lowest to highest trading cost. As expected, while the coefficients
on the contemporaneous return observations ( t ) are less than one, they are positive, large and
statistically significant. Coefficients on the lagged returns ( kt ), particularly the first few lags, are
also positive, large and statistically significant, indicating that the ICE EUA Futures contract often
has a strong, leading relationship to the other securities. The robust F-statistics are also large and
indicate that the coefficients are jointly statistically significant for all regressions.
In contrast, Table 2.12 presents regression results in which contemporaneous and lagged Last
Trade returns of BNX CER Spot are used as the independent variables against the other six
securities. BNX CER Spot is expected to be low in the preferences of traders given its high trading
cost, lack of leverage and the uncertainties surrounding CERs44
. Whilst the contemporaneous
coefficients are positive and statistically significant, they are smaller in magnitude than those for the
ICE EUA Futures. Moreover, there are few statistically significant coefficients against any of the
lagged returns of BNX CER Spot, indicating that its returns seldom lead those of the other
securities45
.
44
BNX CER Spot was chosen from the six other series to provide a contrast to the ICE EUA Futures
regressions because it is not expected to be the venue for price discovery under any of the hypotheses.
45 A comparison of the fifth regression in Table 2.11 (dependent variable: BNX CER Spot, independent
variables: ICE EUA Futures) and the first regression in Table 2.12 (dependent variable: ICE EUA Futures,
independent variables: BNX CER Spot) explicitly sheds light on the direction of price discovery between the
two securities. In Table 2.11, all of the coefficients on the lagged returns of ICE EUA Futures are significant
out to the 9th
lag, indicating that it frequently leads BNX CER Spot by up to 45 minutes. In Table 2.12, none
of the coefficients on lagged returns of BNX CER Spot are significant, save the slight significance on the 3rd
lag, indicating the returns of BNX CER Spot rarely lead ICE EUA Futures. The stronger joint significance of
the coefficients when ICE EUA Futures returns are the explanatory variables is also confirmed by the contrast
in the F-statistics: 20.98 versus 2.74. To varying degrees, explicit comparisons of bilateral regressions
involving ICE EUA Futures confirm its leading relationship to all the other securities, though for brevity we
do not present all 336 regression results.
41
Table 2.11
Regression Results: ICE EUA Futures
Table 2.11 presents the results of fitting model (2.9) using 5-minute continuously compounded returns calculated from Last Trade prices between 1 July 2009 and
30 December 2010 (36,462 return observations). In all cases, independent variables are contemporaneous and lagged returns of ICE EUA Futures (series B) and an error
correction term, with the response variable being one of the six remaining securities of interest in the study (series A). Formally:
ttAzAktB
k
ktA zRR
1,,,
0
10
,
(2.9)
The error correction terms are given by: 1,tAz ln( 1, tAp ) – ln( 1, tBp ), the one-period lag of the difference in log prices between two series. Square brackets [ ]
below coefficients contain t-statistics, while round brackets ( ) below F-statistics contain p-values. * and ** denote significance at the 5 and 1 per cent levels.
α βt βt-1 βt-2 βt-3 βt-4 βt-5 βt-6 βt-7 βt-8 βt-9 βt-10 βZ R-squared F-statistic
-0.0002** 0.4620** 0.1477** 0.0993** 0.0525** 0.0440** 0.0307** 0.0210** 0.0146** 0.0134** 0.0087 0.0009 -0.0276** 0.285 237.10**
[-19.16] [22.21] [21.14] [14.36] [8.46] [8.59] [5.49] [4.00] [2.57] [2.63] [1.55] [0.18] [-19.72] (0.000)
-0.0002** 0.2916** 0.1087** 0.0741** 0.0448** 0.0392** 0.0304** 0.0384** 0.0216** 0.0224** 0.0249** 0.0182** -0.0245** 0.143 43.43**
[-8.80] [18.40] [12.20] [10.17] [6.21] [5.90] [4.39] [4.05] [3.68] [3.53] [4.76] [3.51] [-9.44] (0.000)
-0.0001** 0.3465** 0.1237** 0.0752** 0.0501** 0.0449** 0.0279** 0.0317** 0.0188** 0.0211** 0.0175** 0.0054 -0.0008** 0.141 40.06**
[-4.46] [17.41] [14.65] [10.24] [7.38] [5.58] [4.25] [5.50] [2.57] [3.98] [3.16] [0.93] [-4.80] (0.000)
0.0000 0.1502** 0.0705** 0.0399** 0.0384** 0.0400** 0.0265** 0.0547 0.0338* 0.0211** 0.0146** 0.0107* -0.0071** 0.037 14.04**
[-0.87] [7.03] [7.09] [5.26] [4.68] [4.43] [4.30] [1.66] [2.36] [3.78] [2.72] [2.18] [-4.95] (0.000)
-0.0002** 0.1824** 0.0636** 0.0427** 0.0318** 0.0346** 0.0430** 0.0144* 0.0157** 0.0212** 0.0195** 0.0048 -0.0014** 0.050 20.98**
[-3.11] [5.94] [7.90] [5.61] [4.91] [6.08] [3.96] [1.98] [2.68] [2.92] [3.33] [0.78] [-3.33] (0.000)
0.0000 0.2885** 0.0508** 0.0286 0.0188 0.0128 0.0171 0.0238** 0.0136 -0.0036 0.0061 0.0090 -0.0687** 0.095 35.04**
[1.80] [8.24] [4.48] [1.64] [1.22] [0.79] [1.33] [3.36] [1.55] [-0.39] [0.53] [1.04] [-3.23] (0.000)
NOX EUA
Futures
Dep
en
den
t V
ari
ab
le
Last Trade Return Series
Independent Variable: ICE EUA Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
42
Table 2.12
Regression Results: BNX CER Spot
Table 2.12 presents the results of fitting model (2.9) using 5-minute continuously compounded returns calculated from Last Trade prices between 1 July 2009 and
30 December 2010 (36,462 return observations). In all cases, independent variables are contemporaneous and lagged returns of BNX CER Spot (series B) and an error
correction term, with the response variable being one of the six remaining securities of interest in the study (series A). Formally:
ttAzAktB
k
ktA zRR
1,,,
0
10
,
(2.9)
The error correction terms are given by: 1,tAz ln( 1, tAp ) – ln( 1, tBp ), the one-period lag of the difference in log prices between two series. Square brackets [ ]
below coefficients contain t-statistics, while round brackets ( ) below F-statistics contain p-values. * and ** denote significance at the 5 and 1 per cent levels.
α βt βt-1 βt-2 βt-3 βt-4 βt-5 βt-6 βt-7 βt-8 βt-9 βt-10 βZ R-squared F-statistic
0.0001** 0.1329** 0.0014 -0.0020 0.0098* -0.0032 0.0054 0.0158 -0.0063 -0.0015 -0.0033 0.0087 -0.0006** 0.037 2.74**
[3.03] [3.83] [0.19] [-0.35] [2.05] [-0.74] [0.96] [1.46] [-1.54] [-0.34] [-0.74] [1.78] [-3.29] (0.001)
0.0001** 0.1333** 0.0220** 0.0119** 0.0105 0.0031 0.0053 0.0196 -0.0016 0.0004 -0.0002 0.0074 -0.0006** 0.045 3.24**
[2.73] [3.69] [3.14] [2.83] [1.70] [0.92] [0.97] [1.95] [-0.48] [0.13] [-0.06] [1.34] [-3.00] (0.000)
0.0001** 0.0978** 0.0263** 0.0161** 0.0129** 0.0169** 0.0093* 0.0176* 0.0023 0.0030 0.0070* 0.0077 -0.0008** 0.030 5.07**
[3.24] [4.02] [3.68] [2.73] [3.34] [2.90] [2.37] [2.26] [0.55] [0.81] [2.16] [1.74] [-3.54] (0.000)
-0.0001** 0.1227** 0.0147* 0.0075 0.0095 0.0022 0.0116 0.0206* -0.0005 -0.0005 0.0012 -0.0023 -0.0071** 0.033 4.76**
[-4.71] [4.76] [2.09] [1.48] [1.51] [0.63] [1.89] [2.03] [-0.13] [-0.15] [0.19] [-0.49] [-6.29] (0.000)
0.0001** 0.0534** 0.0164* 0.0196 0.0157* 0.0147* 0.0071** 0.0359 0.0075 -0.0007 0.0060 0.0014 -0.0011** 0.009 3.60**
[3.57] [3.76] [2.36] [1.61] [2.42] [2.21] [2.59] [1.18] [1.78] [-0.32] [1.50] [0.59] [-3.87] (0.000)\
0.0002* 0.1200** 0.0351** 0.0134* 0.0194** 0.0103* 0.0046 0.0304* 0.0189 0.0033 -0.0019 0.0109 -0.0018* 0.030 4.32**
[2.42] [3.01] [3.27] [2.38] [2.89] [2.06] [1.81] [2.42] [1.71] [0.98] [-0.17] [1.87] [-2.48] (0.000)
ICE CER
Futures
Last Trade Return Series
Independent Variable: BNX CER Spot
Dep
en
den
t V
ari
ab
le
ICE EUA
Futures
ICE EUA
Spot
EEX EUA
Futures
NOX EUA
Futures
BNX EUA
Spot
43
The results for regressions utilising Mid Point rather than Last Trade returns are similar to those
just discussed, with two notable exceptions46
. Firstly, the coefficients on the contemporaneous
returns are larger and more significant for the Mid Point regressions than for the Last Trade
regressions. This indicates that there is generally a greater propensity for a change in a limit order in
one security to prompt market participants to alter an existing limit order in another security
compared to the propensity for market orders executed in one security to prompt trades in another.
The other difference is that the number of significant lags is generally fewer for the Mid Point
return regressions than for the Last Trade series, though the coefficients on the first few lagged
returns are still positive, large and statistically significant. For example, changes in the bid or ask
price of the ICE EUA Futures contract (i.e. changes in its Mid Point price) leads the other securities
to respond with bid or ask changes of their own within around 20 minutes compared with the results
in Table 2.9 which show that actual trades in ICE EUA Futures have a statistically significant
relationship with trades in other securities made anywhere up to 50 minutes later47
.
The error correction terms are significant at the 5 per cent level in all but 8 of the 336
regressions and all have negative coefficients as is expected of variables with cointegrated price
levels. Removing the error correction terms and re-running the regressions had little impact on the
R-squared and F-statistics. This indicates that there is little overlap between these price discovery
approaches which alternatively highlight short-run (proximate return coefficients) and long-run
(error correction in levels) dynamics.
46
The full regression results for both types of return series and the other sample frequencies are available
from the author on request.
47 When assessed over 10-minute intraday intervals this extends to as much as 70 minutes for several
securities. As expected, increasing the length of the intraday return interval leads to a lower number of
significant lagged return coefficients, while the significance of the contemporaneous coefficients rises. In
particular, the hourly series rarely display significance at the 1 per cent level beyond the first lag.
44
2.4.2 Regression R-Squared and F-Statistics
As noted previously, while regression comparisons conducted pair-by-pair provide insight into
which security has a leading relationship, the strength and direction of the relationships can more
readily be assessed over the large number of securities considered by comparing the R-squared and
F-statistics. The R-squared statistics displayed in Panel A of Table 2.13 are always larger for the
Mid Point returns. This is predominantly the case for the F-statistics presented in Panel B, with the
exception of the NOX EUA Futures. Although the Mid Point regressions generally have fewer
statistically significant lagged return coefficients, the larger R-squared and F-statistics are primarily
the result of larger coefficients on the contemporaneous return observations, indicating more rapid
short-run adjustment to limit order changes across markets than occurs as a result of trading
activity. However, it should noted that trade activity is likely to impact the limit order book
insomuch as market orders partially or completely erode the best bid or ask in the order book and,
as such, some of the speedy adjustment in the Mid Point returns may be capturing trade activity and
limit order changes as well as the interaction of the two. The relative size of this difference in R-
squared and F-statistics between the Last Trade and Mid Point return series (from Table 2.13) is
best illustrated by comparing Panel A with Panel B in Chart 2.2 and Chart 2.3.
The results presented numerically in Table 2.13 and graphically in Chart 2.2 and Chart 2.3 are
generally replicated in the statistics for the regressions employed at other intraday intervals. The
size of the R-squared and F-statistics generally increases for the longer intraday intervals. Once
again, this is driven by the larger, more significant coefficients on the contemporaneous return
observations, which is to be expected over longer, less frequent intraday intervals. However, as
these longer interval regressions have fewer significant lag coefficients, it is more difficult to
distinguish the direction of price discovery, which emphasises the advantages of using finer, more
granular intraday data in this kind of analysis.
45
Table 2.13
Regression R-Squared and F-Statistics
The reported R-squared and robust F-statistics are from regressions of 5-minute returns of one security (the
dependent variable) against contemporaneous and lagged returns of another and an error correction term (the
independent variables) as per equation (2.9). Returns are calculated from Last Trade (LT) prices and prices
calculated from the Mid Point (MP) of the bid-ask spread.
Panel A:
ICE EUA BNX EUA ICE EUA ICE CER EEX EUA BNX CER NOX EUA
Futures Spot Spot Futures Futures Spot Futures
LT 0.24 0.10 0.12 0.02 0.04 0.06
MP 0.60 0.71 0.49 0.59 0.22 0.08
LT 0.28 0.12 0.11 0.03 0.04 0.07
MP 0.60 0.56 0.37 0.44 0.22 0.07
LT 0.14 0.15 0.07 0.02 0.03 0.04
MP 0.72 0.55 0.40 0.48 0.20 0.07
LT 0.14 0.11 0.07 0.01 0.03 0.04
MP 0.49 0.37 0.40 0.34 0.20 0.05
LT 0.04 0.04 0.02 0.02 0.01 0.02
MP 0.60 0.44 0.48 0.35 0.17 0.06
LT 0.05 0.05 0.03 0.04 0.01 0.03
MP 0.21 0.21 0.18 0.20 0.16 0.03
LT 0.09 0.09 0.05 0.05 0.02 0.03
MP 0.10 0.09 0.08 0.05 0.07 0.04
Panel B:
ICE EUA BNX EUA ICE EUA ICE CER EEX EUA BNX CER NOX EUA
Futures Spot Spot Futures Futures Spot Futures
LT 56.38 29.04 29.81 4.07 2.74 5.46
MP 195.61 410.46 34.28 115.30 24.30 3.72
LT 237.10 28.20 33.59 4.49 3.24 6.78
MP 783.00 224.37 43.04 82.53 27.98 2.77
LT 43.43 28.61 33.05 6.16 5.07 7.60
MP 488.04 130.21 33.99 79.25 27.35 4.17
LT 40.06 33.16 27.12 4.04 4.76 4.40
MP 492.99 110.07 147.77 51.37 27.84 3.33
LT 14.04 10.55 8.60 8.16 3.60 5.20
MP 125.45 75.44 85.04 35.23 27.47 4.44
LT 20.98 18.74 14.75 16.92 3.69 4.70
MP 251.00 129.82 116.70 61.48 49.11 2.90
LT 35.04 28.12 17.75 20.98 5.08 4.32
MP 37.83 28.05 33.71 24.42 35.46 18.03
R-squared
Independent Variable
Dep
en
den
t V
ari
ab
le
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
F-statistics
Independent Variable
Dep
en
den
t V
ari
ab
le
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
46
Chart 2.2
Regression R-Squared
The reported R-squared statistics (from Table 2.13) are for regressions of 5-minute returns of one security
(dependent variable) against contemporaneous and lagged returns of another and an error correction term
(independent variables).
NOX EUA Futures
BNX CER Spot
EEX EUA Futures
ICE CER Futures
ICE EUA Spot
BNX EUA Spot
ICE EUA Futures
0.0000
0.2000
0.4000
0.6000
0.8000
1.0000
ICE EUA FuturesBNX EUA Spot
ICE EUA SpotICE CER Futures
EEX EUA FuturesBNX CER Spot
NOX EUA Futures
R-S
qu
are
d
Independent Variable
Depdendent Variable
NOX EUA Futures
BNX CER Spot
EEX EUA Futures
ICE CER Futures
ICE EUA Spot
BNX EUA Spot
ICE EUA Futures
0.0000
0.2000
0.4000
0.6000
0.8000
1.0000
ICE EUA FuturesBNX EUA Spot
ICE EUA SpotICE CER Futures
EEX EUA FuturesBNX CER Spot
NOX EUA Futures
R-S
qu
are
d
Independent Variable
Depdendent Variable
Panel A: Last Trade Return Series
Panel B: Mid Point Return Series
47
Chart 2.3
Regression F-Statistics
The reported robust F-statistics (from Table 2.13) are for regressions of 5-minute returns of one security
(dependent variable) against contemporaneous and lagged returns of another and an error correction term
(independent variables).
NOX EUA Futures
BNX CER Spot
EEX EUA Futures
ICE CER Futures
ICE EUA Spot
BNX EUA Spot
ICE EUA Futures
0
200
400
600
800
ICE EUA FuturesBNX EUA Spot
ICE EUA SpotICE CER Futures
EEX EUA FuturesBNX CER Spot
NOX EUA Futures
F-St
atis
tic
Independent Variable
Depdendent Variable
NOX EUA Futures
BNX CER Spot
EEX EUA Futures
ICE CER Futures
ICE EUA Spot
BNX EUA Spot
ICE EUA Futures
0
200
400
600
800
ICE EUA FuturesBNX EUA Spot
ICE EUA SpotICE CER Futures
EEX EUA FuturesBNX CER Spot
NOX EUA Futures
F-St
atis
tic
Independent Variable
Depdendent Variable
Panel A: Last Trade Return Series
Panel B: Mid Point Return Series
48
A casual look at Table 2.13 and both Chart 2.2 and Chart 2.3, which are ordered from lowest to
highest trading cost reveals that the size of R-squared and F-statistics diminish moving from left to
right in the table and charts. This is consistent with the notion that the strength of the price
discovery relationship explaining each dependent variable wanes as trading cost increases.
However, as the statistics are not diminishing in perfect order, this does not provide direct support
for the Trading Cost Hypothesis. Rather, visual inspection is inconclusive.
2.4.3 Ordinal Ranking
To help in interpreting results, it is useful to rank them based on the magnitude of their R-
squared and F-statistics. These rankings are displayed in Table 2.14. For each dependent variable,
the independent variables that yield the highest R-squared (or F-statistic) are ranked 1, the second
highest are ranked 2 and so on. The last row of each panel in Table 2.14 contains the average
ranking of these statistics for each independent variable. Table 2.15 displays the ranking that the
results would be expected to follow under each hypothesis and the difference between the average
actual rank and the average expected rank48
.
48
Given the binary choice of distinguishing between leveraged and unleveraged securities and also between
EUAs and CERs, for both the Leverage Hypothesis and Market Segmentation Hypothesis, trading cost is used
as a secondary determinant of the expected ordinal ranking. It should be noted that the expected average ranks
in Tables 2.14 and 2.15 are not monotonically rising whole numbers between 1 and 6 because the use of ICE
EUA Futures as a dependent variable means that one of the other six series necessarily achieves a rank of 1 as
the independent variable best explaining ICE EUA Futures returns and so on. As such, the average expected
rank rises monotonically in the following order: 1.00, 1.83, 2.67, 3.50, 4.33, 5.17, 6.00 as per whichever
hypothesis is being examined.
49
Table 2.14
Actual Rank
When a security is the dependent variable in a series of regressions, the explanatory power of the other six securities, based on
the R-squared and F-statistics, is ranked from largest (1) to smallest (6). For example, the rank of 1 in the 1st row, 2
nd column
of Panel A indicates that the regression in which contemporaneous and lagged returns of the BNX EUA Spot series are the
independent variables has the highest R-squared (best fit) in explaining the returns of ICE EUA Futures. In an attempt to make
visual comparisons with the expected rankings in Table 2.15 slightly easier, the shading follows the explanatory power and
goodness of fit measures, going from the best (darkest) to worst (lightest). R-squared and F-statistics ranks are reported
separately for Last Trade (LT) return series in Panels A and C, and for Mid Point (MP) return series in Panels B and D.
Panel A: Panel B:
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
ICE EUA
Futures1 3 2 6 5 4
ICE EUA
Futures2 1 4 3 5 6
BNX EUA
Spot1 2 3 6 5 4
BNX EUA
Spot1 2 4 3 5 6
ICE EUA
Spot2 1 3 6 5 4
ICE EUA
Spot1 2 4 3 5 6
ICE CER
Futures1 2 3 6 5 4
ICE CER
Futures1 3 2 4 5 6
EEX EUA
Futures1 2 4 5 6 3
EEX EUA
Futures1 3 2 4 5 6
BNX CER
Spot2 1 4 3 6 5
BNX CER
Spot1 2 4 3 5 6
NOX EUA
Futures1 2 3 4 6 5
NOX EUA
Futures1 2 3 5 4 6
Average
Rank1.33 1.50 3.17 3.33 6.00 5.17 4.00
Average
Rank1.00 2.33 2.33 4.00 3.67 5.17 6.00
Panel C: Panel D:
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
ICE EUA
Futures1 3 2 5 6 4
ICE EUA
Futures2 1 4 3 5 6
BNX EUA
Spot1 3 2 5 6 4
BNX EUA
Spot1 2 4 3 5 6
ICE EUA
Spot1 3 2 5 6 4
ICE EUA
Spot1 2 4 3 5 6
ICE CER
Futures1 2 3 6 4 5
ICE CER
Futures1 3 2 4 5 6
EEX EUA
Futures1 2 3 4 6 5
EEX EUA
Futures1 3 2 4 5 6
BNX CER
Spot1 2 4 3 6 5
BNX CER
Spot1 2 3 4 5 6
NOX EUA
Futures1 2 4 3 5 6
NOX EUA
Futures1 4 3 5 2 6
Average
Rank1.00 2.00 3.33 2.67 5.33 5.67 4.50
Average
Rank1.00 2.67 2.17 4.17 3.33 5.17 6.00
R-Squared (MP) R-Squared (LT)
Independent VariableIndependent Variable
F-Statistics (MP)
Dep
en
den
t V
ari
ab
le
Dep
en
den
t V
ari
ab
le
F-Statistics (LT)
Independent Variable
Dep
en
den
t V
ari
ab
le
Independent Variable
Dep
en
den
t V
ari
ab
le
50
Table 2.15
Expected Rank
Panel A contains the ordinal ranking of R-squared and F-statistics that would be expected if the Trading Cost Hypothesis explains relative price discovery—securities are ranked by the
size of the bid-ask spread. Panel B contains the ranking that would be expected if the Leverage Hypothesis explains relative price discovery—futures are ranked ahead of spot securities,
with trading cost a secondary determinant of ordering. Panel C contains the ordinal ranking that would be expected if the Market Segmentation Hypothesis explains relative price
discovery—EUA securities ranked before CER securities, with trading cost a secondary determinant of ordering. The differences in rank are the Actual Ranks for each independent
variable from the four panels in Table 2.14 minus the Expected Ranks in Panels A, B and C of this table. LT denotes Last Trade return series and MP denotes Mid Point return series.
Panel A: Panel B: Panel C:
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
ICE EUA
Futures1 2 3 4 5 6
ICE EUA
Futures4 5 1 2 6 3
ICE EUA
Futures1 2 5 3 6 4
BNX EUA
Spot1 2 3 4 5 6
BNX EUA
Spot1 5 2 3 6 4
BNX EUA
Spot1 2 5 3 6 4
ICE EUA
Spot1 2 3 4 5 6
ICE EUA
Spot1 5 2 3 6 4
ICE EUA
Spot1 2 5 3 6 4
ICE CER
Futures1 2 3 4 5 6
ICE CER
Futures1 4 5 2 6 3
ICE CER
Futures1 2 3 4 6 5
EEX EUA
Futures1 2 3 4 5 6
EEX EUA
Futures1 4 5 2 6 3
EEX EUA
Futures1 2 3 5 6 4
BNX CER
Spot1 2 3 4 5 6
BNX CER
Spot1 5 6 2 3 4
BNX CER
Spot1 2 3 6 4 5
NOX EUA
Futures1 2 3 4 5 6
NOX EUA
Futures1 4 5 2 3 6
NOX EUA
Futures1 2 3 5 4 6
Average
Rank1.00 1.83 2.67 3.50 4.33 5.17 6.00
Average
Rank1.00 4.33 5.17 1.83 2.67 6.00 3.50
Average
Rank1.00 1.83 2.67 5.17 3.50 6.00 4.33
R2 LT 0.33 -0.33 0.50 -0.17 1.67 0.00 -2.00 R
2 LT 0.33 -2.83 -2.00 1.50 3.33 -0.83 0.50 R
2 LT 0.33 -0.33 0.50 -1.83 2.50 -0.83 -0.33
R2 MP 0.00 0.50 -0.33 0.50 -0.67 0.00 0.00 R
2 MP 0.00 -2.00 -2.83 2.17 1.00 -0.83 2.50 R
2 MP 0.00 0.50 -0.33 -1.17 0.17 -0.83 1.67
F-stat LT 0.00 0.17 0.67 -0.83 1.00 0.50 -1.50 F-stat LT 0.00 -2.33 -1.83 0.83 2.67 -0.33 1.00 F-stat LT 0.00 0.17 0.67 -2.50 1.83 -0.33 0.17
F-stat MP 0.00 0.83 -0.50 0.67 -1.00 0.00 0.00 F-stat MP 0.00 -1.67 -3.00 2.33 0.67 -0.83 2.50 F-stat MP 0.00 0.83 -0.50 -1.00 -0.17 -0.83 1.67
Dep
en
den
t V
ari
ab
le
Dep
en
den
t V
ari
ab
le
Segmentation Hypothesis
Independent Variable
Dep
en
den
t V
ari
ab
le
Difference in Rank (Actual - Expected) Difference in Rank (Actual - Expected) Difference in Rank (Actual - Expected)
Trading Cost Hypothesis Leverage Hypothesis
Independent Variable Independent Variable
51
Having the largest F-statistics for all of the regressions in which they are used as
independent variables and the largest R-squared for all but two of the Last Trade regressions,
the ICE EUA Futures are on average ranked 1st in explaining the returns of all other series. The
BNX EUA Spot series is predominantly ranked 2nd
for the Last Trade series, lending some
support to the Trading Cost Hypothesis. However, looking at the Mid Point series BNX EUA
Spot‘s R-squared ranking is only equal 2nd
and its F-statistic ranking is 3rd
compared to ICE
EUA Spot which ranks 2nd
on average. Perhaps this indicates that, while actual trading activity
occurs in a security in which trading cost is lowest, the rationale for deploying limit orders is
somewhat different. In particular, institutions engaged in market making might be more inclined
to simultaneously deploy orders for different securities on a single exchange where they have
membership. In these circumstances, information that warrants a change in limit orders for one
security may soon result in changes in limit orders for other securities on the same exchange49
.
The results concerning the Market Segmentation Hypothesis, the relative ranking of EUA
versus CER securities, are mixed and somewhat inconclusive. ICE CER Futures are
predominantly ranked 3rd
for the Last Trade regressions by F-statistics and very close to 3rd
by
R-squared (average rank of 3.33 versus 3.17 for ICE EUA Spot). As ICE CER Futures would be
expected to place 6th under the Market Segmentation Hypothesis, these stronger rankings are
some evidence that market segmentation between EUAs and CERs may not be important. While
trading cost of ICE CER Futures is not vastly different from the adjacent security (see Table
2.4) and, thus its ranking is still largely in line with the Trading Cost Hypothesis, for the Market
Segmentation Hypothesis this result is somewhat surprising given the uncertainties and
limitations pertaining to the use of CERs are large enough that they trade at a significant
discount to EUAs. It would thus appear that this discount is adequate compensation for these
49
High membership fees may preclude multiple exchange memberships for some market participants.
However, a more plausible explanation might be that auto-quoting by designated market makers keeps
bids and offers across securities on the same exchange highly synchronized. This is most likely the case
for ICE which has a specific designated market maker program in which the exchange waives certain
transaction fees in return for the market maker agreeing to make two-sided, minimum spread quotes
(typical minimums are between €0.05 and €0.20/tCO2e depending on the security and/or contract
maturity). Market makers must make these minimum spreads at least 85 per cent of the time during
exchange opening hours. These results suggest that market makers on ICE may be using algorithms based
on the EUA futures price to update EUA spot market quotes with little latency.
52
risks and that this may not hinder price discovery in CER securities50
. It is, however, difficult to
draw firm conclusions, given that there are only two CER securities with enough trade volume
to warrant inclusion in the study and the other security, BNX CER Spot, is mostly ranked 6th for
both statistics, largely commensurate with all three hypotheses51
.
It is clear that the actual rankings of the R-squared and F-statistics do not conform perfectly
to what is expected under any of the hypotheses and so any evidence in favour of the Trading
Cost Hypothesis over the other hypotheses on the basis of these rankings must be treated with
some caution. Nonetheless, it is still worthwhile attempting to disentangle the relative strength
of support for the three hypotheses. As such, the differences between the actual and expected
rankings are measured and presented in the difference in rank rows at the bottom of Table 2.15.
The absolute value of these differences is taken as a measure of the deviation from the expected
rank and these are presented in Table 2.16.
The average of these absolute deviations show that, across the seven independent variables,
the deviations from what is expected are consistently lowest for the Trading Cost Hypothesis
followed by the Market Segmentation Hypothesis and, finally, the Leverage Hypothesis. This
result is robust to the use of different intraday return intervals and provides evidence in favour
of the Trading Cost Hypothesis. Interestingly, despite the predominant 3rd
rank of the Last Trade
returns for the ICE CER Futures, the Market Segmentation Hypothesis consistently has the
second smallest deviations, making the evidence on market segmentation unclear, though this
lack of clarity is in itself interesting given the limitations on CERs.
50
We note that the difference between the average price of ICE EUA Futures and ICE CER Futures over
the sample period was €1.84 per tCO2e, while the difference between BNX EUA Spot and BNX CER
Spot was €1.62 per tCO2e. The CER discount in the spot and futures markets are not the same
(€1.62/tCO2e versus €1.84/tCO2e) because over the sample period the EUA futures curve was generally
in contango, but the CER futures curve was often backwardated, despite positive carry for both securities.
The World Bank (2010) attributes the backwardation in the CER futures curve to lower than expected
delivery of CERs from the project-based pipeline. Having effectively sold their expected CER deliveries
forward (at least hedged via the establishment of short futures positions), pipeline intermediaries are
forced to buy them in the spot market to meet their obligations if deliveries fail to meet predicted
quantities (which The World Bank, 2010, claims inverts the CER curve).
51 It is predominantly ranked 6
th, except for the R-squared for Last Trade returns, where it is 5
th.
53
Table 2.16
Average Absolute Deviation in Rank
The absolute value of the differences in rank from the bottom of Table 2.15 are reported as a measure of
the deviation of the regression R-squared and F-statistics from what would be expected under the three
hypotheses. The average of these absolute deviations is reported in the final column. LT denotes Last
Trade return series and MP denotes Mid Point return series. The R-squared and F-statistics are from the
regressions estimated as per equation (2.9).
2.4.4 Information Shares
Hasbrouck‘s (1995) information shares are calculated to assess the extent to which each
security‘s variance contributes to innovations in the common efficient price between them.
While the regressions displayed in the previous sections describe the short-run dynamics
between the securities, the information shares are complementary in providing a measure of
each security‘s long-run contribution. Table 2.17 contains information shares for the Last Trade
and Mid Point log price level series sampled at 1-minute intervals, with the higher frequency
sample used in order to minimise contemporaneous correlation in the error terms obtained from
the multivariate VECM in equation (2.11). Some correlation is nonetheless still present in the
error terms, and so the Cholesky factorisation suggested by Hasbrouck (1995) is used, which
entails changing the order of the series in the log price vector to establish the range of
information shares for each series.
Hypothesis
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
Average
Deviation
Trading Cost 0.33 0.33 0.50 0.17 1.67 0.00 2.00 0.71
Leverage 0.33 2.83 2.00 1.50 3.33 0.83 0.50 1.62
Segmentation 0.33 0.33 0.50 1.83 2.50 0.83 0.33 0.95
Trading Cost 0.00 0.50 0.33 0.50 0.67 0.00 0.00 0.29
Leverage 0.00 2.00 2.83 2.17 1.00 0.83 2.50 1.62
Segmentation 0.00 0.50 0.33 1.17 0.17 0.83 1.67 0.67
Trading Cost 0.00 0.17 0.67 0.83 1.00 0.50 1.50 0.67
Leverage 0.00 2.33 1.83 0.83 2.67 0.33 1.00 1.29
Segmentation 0.00 0.17 0.67 2.50 1.83 0.33 0.17 0.81
Trading Cost 0.00 0.83 0.50 0.67 1.00 0.00 0.00 0.43
Leverage 0.00 1.67 3.00 2.33 0.67 0.83 2.50 1.57
Segmentation 0.00 0.83 0.50 1.00 0.17 0.83 1.67 0.71
Independent Variable
R-Squared (MP)
R-Squared (LT)
F-Statistics (LT)
F-Statistics (MP)
54
Table 2.17
Information Shares
Panel A contains Hasbrouck‘s (1995) information shares (i
IS ) calculated using log price levels for the
Last Trade series at a 1-minute sampling frequency between 1 July 2009 to 30 December 2010 (181,327
observations). Panel B contains information shares calculated for the Mid Point series. After estimating
the VECM described in equation (2.11), information shares are calculated as per:
ΨΨΩ
2
ii
MIS
(2.16)
The elements of the ( K1 ) row vector, Ψ , are the sums of each security‘s coefficients in the moving
average impact matrix ( i ). The lower triangular matrix from a Cholesky factorisation ( M ) of the error
term contemporaneous covariance matrix ( Ω ) from the VECM is used to construct the variance
contribution of each security, ([ M ]i )2, to the variance of the common efficient price ( ΨΨΩ ). The
order of the series in the VECM log price vector (t
y ) follows the order dictated by the Trading Cost
Hypothesis but this order is cycled such that the VECM is run with each series taking a turn at being the
first series in the log price vector. As the order is cycled and the VECM re-run, the information shares are
re-calculated. Shading indicates the information share that is typically, though not always, the maximum
out of the seven cycles because this cycle is the one in which that series is ordered first (as listed in the
first column of the table). The range of the information shares for each variable and the average
information share calculated across the seven cycles are displayed along with the ordinal ranking of these
averages. The multivariate VECM specification in equation (2.11) contains 10 lags in theit
y vector as
suggested by autocorrelation coefficients.
Panel A:
Series Ordered ICE EUA BNX EUA ICE EUA ICE CER EEX EUA BNX CER NOX EUA
First in Cycle Futures Spot Spot Futures Futures Spot Futures
ICE EUA Futures 0.867 0.041 0.010 0.058 0.017 0.003 0.003
BNX EUA Spot 0.538 0.269 0.052 0.125 0.008 0.009 0.000
ICE EUA Spot 0.644 0.023 0.119 0.192 0.003 0.017 0.003
ICE CER Futures 0.615 0.008 0.000 0.278 0.071 0.025 0.002
EEX EUA Futures 0.832 0.043 0.011 0.059 0.000 0.040 0.016
BNX CER Spot 0.823 0.040 0.010 0.057 0.016 0.040 0.015
NOX EUA Futures 0.848 0.043 0.011 0.059 0.016 0.004 0.019
Maximum 0.867 0.269 0.119 0.278 0.071 0.040 0.019
Minimum 0.538 0.008 0.000 0.057 0.000 0.003 0.000
Average 0.738 0.067 0.030 0.118 0.019 0.020 0.008
Rank 1 3 4 2 6 5 7
Last Trade Return Series
55
Table 2.17
Information Shares (continued)
The average information share of the Last Trade ICE EUA Futures indicates that its
variance explains 73.8 per cent of the innovations in the common efficient price for emission
allowances, while it explains 54.7 per cent using the Mid Point series. As ICE EUA Futures
drive the majority of innovations in the common efficient price, these results clearly support the
conclusion that this security is the main vehicle for price discovery in the EU ETS, though it
should be noted that these numbers are similar to its proportion of trade volume among this
group of seven more heavily traded securities (71.1 per cent as shown in Table 2.4).
The average information shares for the Mid Point series in Panel B of Table 2.17 conform
almost perfectly to the order expected under the Trading Cost Hypothesis, with the exception of
the switch in ranks for the BNX EUA Spot and ICE EUA Futures series. It is less clear which
hypothesis the ranking of the average information shares for the Last Trade series in Panel A
supports. It is again interesting that the ICE CER Futures seems to be a more prominent vehicle
for price discovery using the Last Trade series than would be expected under the Market
Segmentation Hypothesis accounting for 11.8 per cent and ranking 2nd
behind the ICE EUA
Panel B:
Series Ordered ICE EUA BNX EUA ICE EUA ICE CER EEX EUA BNX CER NOX EUA
First in Cycle Futures Spot Spot Futures Futures Spot Futures
ICE EUA Futures 0.853 0.069 0.034 0.041 0.002 0.000 0.001
BNX EUA Spot 0.147 0.549 0.218 0.084 0.002 0.000 0.000
ICE EUA Spot 0.204 0.034 0.618 0.132 0.009 0.003 0.000
ICE CER Futures 0.422 0.051 0.027 0.409 0.073 0.015 0.002
EEX EUA Futures 0.615 0.067 0.035 0.042 0.186 0.047 0.008
BNX CER Spot 0.759 0.067 0.034 0.042 0.002 0.080 0.017
NOX EUA Futures 0.830 0.069 0.035 0.041 0.002 0.000 0.023
Maximum 0.853 0.549 0.618 0.409 0.186 0.080 0.023
Minimum 0.147 0.034 0.027 0.041 0.002 0.000 0.000
Average 0.547 0.129 0.143 0.113 0.039 0.021 0.007
Rank 1 3 2 4 5 6 7
Mid Point Return Series
56
Futures (this information share is also much larger than its share of trade volume of 5.4 per cent
as shown in Table 2.4)52
.
Similar to the previous section, in order to disentangle the three hypotheses, we analyse the
deviations of the actual rank of the information shares from the ranks that would be expected
under each hypothesis. The absolute value of the deviations in the actual ranks from Table 2.17
to those that are expected under the three hypotheses are presented in Table 2.18, with the
average of these absolute deviations displayed in the last column.
Table 2.18
Absolute Deviation in the Ordinal Rank of Information Shares
Table 2.18 displays the actual ordinal ranking of the average information shares from the last rows of
both Panel A and Panel B in Table 2.17. The three ordinal rankings expected under each of the three
hypotheses are reported along with the absolute value of the deviation of the actual ranks from the
expected ranks. The averages of these absolute deviations for each hypothesis are displayed in the last
column.
52
Although the 1-minute series is used in an attempt to minimise contemporaneous cross correlations
between the error terms, the results from the 5-minute series are very similar in both the relative
magnitude of each security‘s information share and each security‘s rank relative to the others.
Panel A:
Hypothesis
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
Average
Deviation
Actual Rank 1 3 4 2 6 5 7
Expected Rank Trading Cost 1 2 3 4 5 6 7
Leverage 1 5 6 2 3 7 4
Segmentation 1 2 3 6 4 7 5
Trading Cost 0 1 1 2 1 1 0 0.86
Leverage 0 2 2 0 3 2 3 1.71
Segmentation 0 1 1 4 2 2 2 1.71
Panel B:
Hypothesis
ICE EUA
Futures
BNX EUA
Spot
ICE EUA
Spot
ICE CER
Futures
EEX EUA
Futures
BNX CER
Spot
NOX EUA
Futures
Average
Deviation
Actual Rank 1 3 2 4 5 6 7
Expected Rank Trading Cost 1 2 3 4 5 6 7
Leverage 1 5 6 2 3 7 4
Segmentation 1 2 3 6 4 7 5
Trading Cost 0 1 1 0 0 0 0 0.29
Leverage 0 2 4 2 2 1 3 2.00
Segmentation 0 1 1 2 1 1 2 1.14
Absolute
Deviation in
Rank
Last Trade Return Series
Absolute
Deviation in
Rank
Mid Point Return Series
57
In concordance with the average absolute deviations in rank from the regression results
(presented in Table 2.16), the average absolute deviations in rank for the three hypotheses are
smallest for the Trading Cost Hypothesis when constructed from information shares. This is the
case for both the Last Trade and Mid Point series, which supports the robustness of the
conclusion that low trading costs are a more important determinant of where price discovery is
likely to take place than the provision of leverage imbedded in futures or any segmentation of
the emission allowance market. At the margin, the Market Segmentation Hypothesis displays
the second lowest deviations and the Leverage Hypothesis the highest, though the two are equal
when the Last Trade prices are used.
2.4.5 Strength of Findings
Importantly, a number of the securities examined have significantly less trade volume than
the ICE EUA Futures. In light of this it may be unreasonable to expect that they should follow
any hypothesized preference scheme dictating their use, as it could be said that they are barely
used at all53
. The analysis leant towards conservatism in not wanting to forego the inclusion of
possible vehicles for price discovery when it was decided to include such securities in the first
place. Perhaps the true nature of price discovery in the EU ETS, or any market, is that one
security is predominantly used, the others track it, and any relationship between these other
securities is purely the product of lagging more or less than one another, with no hierarchy of
use related to trading cost. In this way, the appearance that other securities are sometimes used
for price discovery may simply be coincidental.
It could also be that the number of securities examined is too few to reliably disentangle the
number of hypotheses and that statistics calculated from deviations of ordinal ranks are devoid
of more important, nuanced information such as the relative strength of the fit of the various
regressions or the relative magnitude of the information shares. Inasmuch as this is the case, the
most reliable result is that price discovery in the EU ETS does in fact take place in the most-
traded, least-cost, leveraged, EUA security, as would be expected.
53
For instance, ICE EUA Futures are on average traded 92 times more frequently than NOX EUA
Futures over the sample period according to the data presented in Table 2.4.
58
Though these are valid concerns it should be noted that the Trading Cost Hypothesis is
supported by two distinct methodological approaches to detecting price discovery: one that
focuses on short-run return dynamics (regressions results), the other on the long-run equilibrium
price dynamics (information shares). It should also be noted that in using trading cost as a
secondary determinant of the ordering expected under the Leverage and Market Segmentation
Hypotheses, the analysis of deviations from expected rankings has subtly biased the results for
these two hypotheses in the event that trading cost is the more important determinant as has
been shown. This points to the actual degree to which trading cost is a better determinant of
price discovery being stronger than is belied by comparing the average absolute deviation in
rankings. Both of these factors add to the robustness of the support for the Trading Cost
Hypothesis.
2.5 Conclusion
By employing high frequency data across a wide range of securities, this study provides the first
evidence on the catalysts, and not simply the source, for price discovery in the EU ETS. The
results indicate that trading cost is a more important determinant of whether a security displays
greater price discovery than whether the security implicitly provides the trader with leverage or
whether the security is an EUA or a CER. These results are in line with much of the literature
that examines price discovery in other markets where low trading costs are overwhelmingly
shown to be associated with price discovery.
While the lowest trading cost security, the Intercontinental Exchange December expiry
EUA futures, is predominantly the source of price discovery in the EU ETS, as a futures
contract it also entails the implicit provision of leverage. As such, we disentangled the benefits
of low trading costs from the provision of leverage by examining the leading and lagging
relationships between a range of other highly traded securities in the EU ETS. We also
compared the security‘s respective information shares. We find support for the Trading Cost
Hypothesis on the basis of the ordinal ranking of the goodness of fit and the joint statistical
59
significance of the coefficients in regressions run on high frequency intraday return data and on
the basis of the information shares attributable to each security.
Though trading cost is also shown to be a more important determinant than any
segmentation of the emission allowance market between EUAs and CERs, the ICE CER Futures
were highly ranked under several measures, indicating a not insignificant contribution to price
discovery, especially in comparison to trade volumes. This is particularly interesting given that
there are enough uncertainties about the continuing role of CERs in the EU ETS as well as
annual limitations on their acceptance for regulatory compliance such that they trade at a
substantial discount. In this sense, the discount appears to be adequate compensation for these
risks, with the speculative use of CERs unimpeded.
60
CHAPTER 3: Price Discovery in European Energy Markets
Considerable effort has been made to understand price discovery in stock markets (see, for
example, Hasbrouck, 1995, and Fleming, Ostdiek and Whaley, 1996) and bond markets (see, for
example, Fleming and Remolona, 1999, and Brandt and Kavajecz, 2004). However, the area of
price discovery in energy markets remains relatively unexplored. The main reasons for this are
that there is little price transparency in the physical markets, which are dominated by long-term
supply contracts, while associated derivative securities often lack liquidity. Notwithstanding
this, gaining a better understanding of price discovery in these markets is important for market
participants, regulators and researchers alike. We provide evidence in this regard for the coal,
natural gas and crude oil markets in Europe.
Energy markets have a complicated mix of financial and physical layers in which prices are
discovered either by the outcome of trading activity on organised exchanges or through surveys
of over-the-counter market participants. Price discovery amongst the securities that inhabit these
financial and physical layers is important because benchmark prices derived from them
ultimately underpin the value of vast quantities of coal, natural gas and crude oil transacted
under long-term supply contracts, which remain a prevalent mode of exchange for these
commodities. While the analysis of price discovery in these markets is thus of great importance
to market participants, it is also important for researchers, and for studies that involve temporal
evaluations in particular. For example, in research that involves causality between energy
commodities themselves or their interaction with other economic indicators, the timeliness of
61
security price responses to information is vital. In order to obtain unbiased results, it is
imperative to select security prices that best incorporate relevant information in each commodity
market of interest. Further investigation of price discovery is also warranted in light of the
concerns raised by regulators as to the impartiality of the price reporting agencies in setting
benchmark prices by surveying physical transactions (see IOSCO, 2012). Despite its
importance, the existing literature in this area is somewhat sparse.
To the best of our knowledge there have not been any explicit studies of price discovery in
the European coal market, while studies involving European natural gas have either focussed on
interactions with natural gas prices in other regions of the world (Mazighi, 2005, Siliverstovs,
L‘Hégaret, Neumann and von Hirschhausen, 2005, and Kao and Wan, 2009) or with crude oil
prices (Asche, Osmundsen and Sandsmark, 2006, and Panagiotidis and Rutledge, 2007)54
. For
crude oil itself, several studies explain the complicated interaction between the financial and
physical layers of North Sea crude oil markets (see, for example, Fattouh, 2011, and Barret,
2012), however, none have attempted to quantitatively assess price discovery within this
complex of securities. In the absence of quantitative work on price discovery in these energy
commodity markets, we contribute to the literature using two distinct methodologies that focus
on short-run and long-run aspects of the price formation process, respectively.
We examine short-run dynamics in the relative speed of information absorption in the
physical and financial markets by employing regression analysis to assess the contemporaneity
of returns within each class of energy commodity. These regressions assess the strength of the
return relationships between different securities that essentially represent ownership of the same
underlying commodity and look to establish whether the movements in one security price tend
to lead or lag price movements in others. We also examine which securities‘ innovations
contribute most to the long-run equilibrium in each commodity market by calculating
Hasbrouck‘s (1995) information shares. While securities in the same market may tend to track
one another over the short-run, the information shares methodology looks to ignore any
54
We note that Neumann, Siliverstovs and von Hirschhausen (2006) study convergence in day-ahead
natural gas prices from the National Balancing Point (UK) and Zeebrugge (Belgium) hubs using daily
data from 2000 to 2005, but do not directly investigate price discovery.
62
transitory components of price innovations and to capture the extent to which a security‘s price
movements permanently impact upon market equilibrium. Although European coal, natural gas
and crude oil markets display commonalities insofar as they all involve complex interactions of
financial securities and physical over-the-counter spot, forward and swap transactions, these
markets differ markedly in their specific structures and their respective degrees of liquidity and
transparency. As such, each market is considered individually.
Coal is the least liquid and least transparent of the energy markets analysed in this study,
with the main sources of price information being the assessment of contract prices through
surveys conducted by several price reporting agencies and the intermittent trading in futures
markets. In the absence of suitable and timely price reporting agency assessments for coal, our
study focuses on observations of prices and returns in the futures market. Although the
regression results show that it is hard to distinguish between the securities on the basis of short-
term return dynamics, the information shares indicate a greater proportion of long-run price
discovery takes place amongst the monthly expiry futures traded on the Intercontinental
exchange.
The better established natural gas pricing hubs in the UK, Belgium and the Netherlands
display greater liquidity and transparency than the coal futures market. However, regression
analysis reveals that, despite the web of interconnecting gas pipelines across Europe, the short-
term return linkages between the different markets are somewhat weak. In fact, there is some
evidence that return interactions are stronger at similar points on the forward curve than
interactions between returns specific to the geographic location of hubs, with short-dated gas
prices displaying similar demand inelasticity to electricity prices. The information share results
show that the greatest contribution to long-run equilibrium clearly comes from the monthly
expiry UK natural gas futures traded on the Intercontinental Exchange. In addition,
cointegration tests indicate that natural gas prices remain weakly linked to the crude oil market,
contrary to the findings of Asche et al. (2006), but supporting those of Panagiotidis and
Rutledge (2007).
The analysis of crude oil markets is split between an examination of the financial and
physical layers of the Brent complex of securities and an examination of whether the Brent
63
market responds to the prices of West Texas Intermediate crude oil. Although the physical Brent
market remains opaque, we examine a proxy constructed from the Exchange-for-Physicals
market. The Intercontinental Exchange Brent crude oil futures are shown to dominate price
discovery compared with this proxy, but there is some evidence of bi-directional leadership in
short-run return dynamics. This is the first quantitative evidence in the literature of the dominant
role of the futures market in the determination of Brent crude oil prices and supports the
postulations in Fattouh (2011) and Barret (2012).
We also re-examine the relationship between Brent and West Texas Intermediate futures
prices, the two most important global benchmarks, in light of recent dislocations caused by
structural issues at the pricing point for West Texas Intermediate in Cushing, Oklahoma (see
Montepeque, 2012, and Sen, 2012). Despite the large number of prior studies that conclude
crude oil prices are determined in a unified global market (Adelman, 1984, 1992, Gülen, 1997,
1999, Bachmeier and Griffin, 2006, Bentzen, 2007, and Kaufmann and Ullman, 2009), we find
only minor evidence that these markets are cointegrated at standard levels of statistical
significance. Indeed, the degree of cointegration declines through our sample coincident with
the increasing structural problems depressing the price of West Texas Intermediate. At the
margin, West Texas Intermediate tends to display greater price leadership in short-term return
dynamics. The information shares also point to West Texas Intermediate futures making a
greater contribution to long-run price equilibrium, though there are several sub-periods in which
Brent futures are shown to be relatively more important. These results are largely consistent
with the findings of prior studies by Brunetti and Gilbert (2000), Lin and Tamvakis (2001),
Hammoudeh, Ewing and Thompson (2008) and Kaufmann and Ullman (2009).
The remainder of this paper is organised as follows: Section 3.1 describes the regression and
information share methodologies; Sections 3.2, 3.3 and 3.4 examine the data and results for the
coal, natural gas and crude oil markets, respectively; and, Section 3.5 concludes.
64
3.1 Methodology
We employ two distinct methodologies to assess price discovery in both short and long-term
contexts. Specifically, we use a regression approach to focus on the short-term return dynamics
between the securities, and utilise Hasbrouck‘s (1995) information shares measure to provide
evidence on the contribution of each security to the long-run price equilibrium. We discuss each
methodology in greater detail below.
3.1.1 Regression Approach
This paper adopts the methodology of Chapter 2, which makes small alterations to the
approach in Fleming et al. (1996), and is a compromise between Vector Error Correction
Models (VECMs) and ordinary least squares (OLS) regression. Specifically, contemporaneous
returns of one series ( tAR , ) are regressed against the contemporaneous and lagged returns of
another series ( ktBR , ) and an error correction term equal to the one-period lag of the difference
in log prices between the two series, or 1,1,1, lnln tBtAtA PPz , to determine which series
appears to react more quickly to information. Formally, the following model is used:
ttAzAktB
jk
ktA zRR
1,,,
0
,
(3.1)
The number of return lags ( j ) used as independent variables is chosen with reference to the
best available frequency for the respective energy security price data, the selection of which is
discussed for each commodity in full below. All regressions use the method proposed by
Newey-West (1987) to estimate variance-covariance matrices that are robust to autocorrelation
and heteroskedasticity and which allow for the calculation of robust F-statistics55
. The number
of permutations in which these regressions are run depends on the number of viable contending
55
The lag length ( L ) for the Newey-West (1987) technique is chosen with reference to their rule of
thumb that there be : 4 TL lags and is thus dependent on the number of observations ( T ), which in
turn is the product of the frequency of sampling and the number of days in common between the
respective sets of return series.
65
securities that are potential sources of price discovery in their respective markets. Conclusions
are drawn from visual inspection of regression coefficients and, where these are inconclusive,
from the comparison of the adjusted R-squared and robust F-statistics.
3.1.2 Information Shares
Hasbrouck‘s (1995) information shares measure the relative contribution of each security to
the variance of innovations in the common factor between them. Decomposing actual prices
( tip , ) into an unobservable common efficient price ( tm ), which follows a random walk
( ttt umm 1 ), and idiosyncratic transitory factors ( tis , ), Hasbrouck (2002) gives:
tk
t
t
tk
t
t
s
s
m
p
p
,
,1
,
,1
1
1
y
(3.2)
Using a multivariate VECM specification gives56
:
tit
n
i
itt εyΓyβαy
1
1
1
(3.3)
Where: ty is a 1K vector of first differences in log price levels ( tip , ); ity are lagged
dependent variables; α and β are rK
parameter matrices in which the number of
cointegrating equations is less than the number of I(1) variables ( Kr ); 11 ,, pΓΓ are
KK matrices of parameters; and, tε is a 1K vector of normally distributed and serially
uncorrelated error terms with contemporaneous covariance matrix Ω . Hasbrouck (1995)
represents equation (3.3) as a Vector Moving Average (VMA), where LΨ
is a matrix
polynomial in the lag operator, L :
56
Note that a trend term is used in the cointegrating equation of the VECM in circumstances in which
spot and futures securities are being compared (i.e. are components of the vector t
y ) to account for
carrying costs.
66
tt L Ψy
(3.4)
Following Baillie et al. (2002), equation (3.4) can be expressed in an integrated form as
follows:
t
t
s
st L *ΨΨy
1
1
(3.5)
The moving average impact matrix, 1Ψ , is calculated as the sum of the moving average
coefficients, which is the long-run impact of price innovations that are common to all the series.
The rows of the impact matrix, 1Ψ , are identical. If we denote one of these rows Ψ (a K1
row vector), we can express the variance of the common efficient price as:
ΨΨΩ 2
u
(3.6)
The information share ( iIS ) of a particular security is then calculated as its variance
contribution (22
ii ) to the variance of the common efficient price:
ΨΨΩ
22
iiiIS
(3.7)
The specification in equation (3.7) relies on the absence of contemporaneous correlation in
the error terms from the VECM. Because correlation is typically present, Hasbrouck (1995) uses
a Cholesky factorisation of the error term covariance matrix, Ω , equal to MM where M is
the lower triangular matrix. Thereafter, the ordering of the series in the log price vector is cycled
to provide a range for the information shares. Thus, Hasbrouck‘s (1995) information shares are
given by:
ΨΨΩ
2
ii
MIS
(3.8)
We take the average information share calculated from running the VECM in equation (3.3)
as many times as necessary to completely cycle the order of the securities in the log price
67
vector. All versions of (3.3) are run with the appropriate number of cointegrating relations
commensurate with Johansen (1995) cointegration tests (detailed in the Appendix to this
chapter).
In the analysis that follows, all return data for the energy commodities is calculated as the
logarithm of the first differences in prices sourced from Thomson Reuters Tick History in
respect of the 4-year period from 2 January 2008 to 30 December 2011, inclusive. All days in
this period where securities are traded or settlement prices are recorded are used. In light of time
zone differences intraday windows are all expressed in Greenwich Mean Time (GMT). Futures
securities used are front contracts up until the day prior to expiry, at which time we switch into
the next-to-front contract splicing the series so as to avoid the return effects of the contract
change57
.
3.2 Coal
Coal is the least liquid of the markets examined in this study. It is predominantly traded via
long-term, bilateral contracts under which counterparties periodically renegotiate prices, with
little external price transparency. The main sources of price information for coal transactions are
the indices compiled by price reporting agencies such as McCloskey and Argus or the futures
prices from the exchanges facilitating futures trading. The McCloskey and Argus index prices
are proxies for long-term contract prices as they are gathered from regular surveys which take
an average of the prices observed by dozens of market participants for coal to be delivered in
the next 90 days. Unfortunately, until very recently, these indices were only available on a
weekly basis and, at this frequency, would not yield enough data for robust analysis. In addition,
prices gleaned in this fashion may not reflect current market circumstances as the long-term
57
In addition, the analysis in the paper was also run rolling all the futures contracts two weeks prior to
expiry. The results were not substantially different, but we note that the crude oil information share
results in particular were somewhat more ambiguous. We determined that rolling the futures contracts one
day prior to expiry was more consistent with the persistence of greater volume in the front contract up
until this time, though we note that some well known commodity funds that invest via futures sometimes
roll their positions earlier than this (for example, Stoll and Whaley, 2010, describe how two prominent
commodity index funds—the S&P Goldman Sachs Commodity Index and the Dow Jones/UBS
Commodity Index—often roll into the next-to-front contract several weeks before the front contract
expires).
68
contracts to which they pertain may have been agreed in the past58
. Although for these reasons
coal price index data are not used directly in this study, it should be noted that most coal futures
contracts are nonetheless cash settled against the level of the API2 index published in the
Argus/McCloskey Index Report59
.
The main European exchanges facilitating coal futures trading are the Intercontinental
Exchange (ICE) and the European Energy Exchange (EEX). The liquidity in these contracts is
very poor and, in the absence of actual trade activity, reported daily settlement prices are
typically averages of bid and ask prices or the result of the exchanges surveying market
participants for indicative levels, which admittedly is not dissimilar to the approach taken by the
price reporting agencies. The lack of trading in coal futures markets necessitates analysis being
conducted on a daily basis as dependable higher frequency data is not available. We examine
price discovery between the coal futures contracts that actually trade every day or so: the
monthly and quarterly expiry ICE Rotterdam coal futures and the monthly and annual expiry
EEX ARA coal futures60
.
Descriptive statistics for the coal securities are detailed in Table 3.1 and are similar across
the four futures contracts. With the exception of the EEX monthly series, all the series display
significant autocorrelation, largely driven by the large, positive autocorrelation coefficients on
their first lagged returns (as detailed in the Appendix). The presence of this autocorrelation,
which is likely a product of thin trading, justifies the use of Newey-West (1987) estimation of
the residual variance-covariance matrix under the regression approach and the inclusion of at
58
In addition, the daily McCloskey price series available since February 2009 appears to be either
released at a one-day lag or alternatively the survey responses may be influenced by previous day
settlement prices in the futures market. For example, the contemporaneous return correlation of the daily
McCloskey series with the ICE monthly expiry coal futures is only 13.3 per cent, while the correlation
when the ICE futures are lagged one day is 57.0 per cent (while the correlation if the McCloskey series is
lagged is only 1.3 per cent). These features make the use of the McCloskey series in studies such as
Keppler and Mansanet-Bataller (2010) and Hintermann (2010) a somewhat curious choice, particularly
for analysis looking at causality in which timing is so important.
59 The API2 index is the average of the Argus Rotterdam price assessment and the McCloskey Northwest
Europe steam coal marker. These prices are on a Cost, Insurance and Freight (CIF) basis for coal arriving
in the ports of Amsterdam, Rotterdam and Antwerp (ARA). See the McCloskey website for details of the
index calculation methodology: http://www.ihs.com/products/coal-information/coal-methodology-
guide.aspx
60 Intraday annual expiry coal futures data sourced from Thomson Reuters for the commodities brokerage
GFI were also examined as well as daily settlement prices for the annual expiry ICE coal futures, but
these were not found to be sources of price discovery. Although these results are not reported for the sake
of brevity, they are available from the author on request.
69
least one lag in the VECM estimated to calculate information shares. Augmented Dickey-
Fuller (1979) stationarity tests of the log price levels and continuously compounded returns
reveal that the securities are all I(1) variables, stationary in returns but not in price levels, and
are cointegrated. The results of these tests are detailed in the Appendix.
Table 3.1
Descriptive Statistics for Coal Returns
Table 3.1 displays descriptive statistics of the continuously compounded returns for coal futures sampled
at daily intervals between 2 January 2008 and 30 December 2011 (961 observations).
Table 3.2 displays the results of running regressions for the coal futures as per equation
(3.1)61
. According to the regressions, the differences between coal securities are slight, which is
most likely because of the necessarily low sampling frequency and daily settlement prices being
determined via exchanges surveying some of the same market participants on days without
trading. Visual inspection of the coefficients in the twelve regressions provides no strong
evidence on the direction of price discovery. All series display very large, positive and
statistically significant contemporaneous return coefficients. However, while many regressions
also have significant coefficients on the first lagged returns, there isn‘t a consistent pattern that
points to greater short-run return leadership residing in a particular security. Because visual
inspection is inconclusive, we also examine the average R-squared and F-statistics from these
regressions to see whether these statistics are clearly stronger for any of the securities. We also
61
A lag length ( j ) of 5 periods was chosen as ample given that the data are sampled on a daily basis.
ICE Monthly Futures ICE Quarterly Futures EEX Monthly Futures EEX Annual Futures
Frequency Daily Daily Daily Daily
Mean -0.0001 -0.0002 0.0002 -0.0002
Standard Deviation 0.0172 0.0193 0.0151 0.0184
Skewness -1.45 -0.71 -0.88 -0.62
Kurtosis 18.92 7.10 12.16 7.15
Maximum 0.1090 0.0713 0.0819 0.0864
75th
Percentile 0.0061 0.0093 0.0049 0.0090
25th
Percentile -0.0049 -0.0083 -0.0039 -0.0081
Minimum -0.1655 -0.1043 -0.1068 -0.1088
70
note that the coefficients on the error correction terms are seldom statistically significant
indicating that any cointegration between the securities is weak at best62
.
Table 3.2
Price Discovery Regressions for European Coal
Table 3.2 presents the results of fitting model (3.1) using continuously compounded returns calculated
from daily settlement prices between 2 January 2008 and 30 December 2011 (956 return observations). In
Panel A, the independent variables are contemporaneous and lagged returns of monthly expiry ICE
Rotterdam coal futures (series B) and an error correction term, with the response variable being one of the
three other series of interest (series A). Panels B, C and D are set out similarly. Formally:
ttAzAktB
k
ktA zRR
1,,,
0
5
,
(3.1)
The error correction terms are given by: 1,tA
z ln(1, tA
p ) – ln(1, tB
p ), the one-period lag of the
difference in log prices between the two series. Square brackets [ ] below coefficients contain t-statistics,
while round brackets ( ) below F-statistics contain p-values. * and ** denote significance at the 5 and
1 per cent levels, respectively.
62
This conforms with the cointegration test results in the Appendix (Table A3), which show three
cointegrating relations between the series at the 5 per cent level of significance, but only two relations at
the 1 per cent level.
α βt βt-1 βt-2 βt-3 βt-4 βt-5 βZ Adj R-sqr F-stat
Panel A:
0.0001 0.8034** 0.1135** -0.0289 0.1056* 0.0041 -0.0234 -0.0019 0.517 16.86**
[0.21] [7.58] [3.26] [-0.85] [2.24] [0.10] [-0.52] [-0.97] (0.000)
0.0002 0.5517** 0.1141** -0.0015 0.0296 0.0562 -0.0046 0.0004 0.412 12.03**
[0.81] [6.84] [3.18] [-0.05] [0.77] [1.29] [-0.09] [0.55] (0.000)
-0.0002 0.5647** 0.1136* -0.0443 0.0532 0.0528 0.0203 -0.0019 0.289 13.16**
[-0.33] [4.38] [2.27] [-1.30] [1.12] [1.09] [0.46] [-0.93] (0.000)
Panel B:
-0.0005 0.6283** 0.0210 0.0075 -0.0400 0.0208 0.0258 0.0019* 0.500 52.40**
[-1.09] [16.10] [0.54] [0.28] [-1.14] [0.44] [0.75] [1.98] (0.000)
-0.0001 0.4471** 0.0727* 0.0188 -0.0366 0.0684 -0.0041 0.0015 0.364 22.71**
[-0.17] [8.42] [1.99] [0.53] [-0.90] [1.92] [-0.12] [1.51] (0.000)
0.0000 0.8156** 0.0113 -0.0181 -0.0281 0.0296 0.0435* -0.0004 0.737 94.32**
[0.00] [21.14] [0.45] [-0.92] [-1.22] [1.19] [2.55] [-0.32] (0.000)
Panel C:
-0.0003 0.7289** 0.1382** 0.0080 -0.0315 -0.0778 0.1049 -0.0003 0.421 63.94**
[-0.89] [15.34] [3.75] [0.28] [-0.96] [-1.73] [1.87] [-0.19] (0.000)
-0.0008 0.7714** 0.1689** 0.0381 0.0281 -0.0778 0.0217 -0.0026 0.369 40.34**
[-1.02] [12.08] [4.28] [0.79] [0.61] [-1.62] [0.51] [-0.85] (0.000)
-0.0010 0.6924** 0.0818* 0.0211 0.0076 -0.0333 0.0903* -0.0022 0.325 24.48**
[-1.06] [10.24] [2.14] [0.40] [0.15] [-0.65] [2.04] [-0.83] (0.000)
Panel D:
-0.0008 0.4844** 0.0768 0.0382 -0.0349 0.0335 -0.0230 0.0028 0.283 12.10**
[-1.31] [6.15] [1.73] [0.89] [-0.75] [0.56] [-0.58] [1.87] (0.000)
-0.0002 0.8990** 0.0543** 0.0247 0.0188 -0.0240 -0.0553** 0.0007 0.739 138.09**
[-0.56] [25.63] [2.88] [1.00] [0.79] [-1.04] [-2.98] [0.58] (0.000)
-0.0002 0.4561** 0.0482 0.0463 -0.0348 0.0727* -0.0359 0.0018 0.330 21.19**
[-0.44] [9.03] [1.66] [1.28] [-0.97] [2.51] [-1.03] [1.29] (0.000)
Independent Variable: EEX Annual
Dep
en
den
t V
ari
ab
le ICE
Monthly
ICE
Quarterly
EEX
Monthly
Dep
en
den
t V
ari
ab
le ICE
Monthly
ICE
Quarterly
EEX
Annual
Independent Variable: ICE Monthly
Dep
en
den
t V
ari
ab
le ICE
Quarterly
EEX
Monthly
EEX
Annual
Independent Variable: ICE Quarterly
Dep
en
den
t V
ari
ab
le ICE
Monthly
EEX
Monthly
EEX
Annual
Independent Variable: EEX Monthly
71
Table 3.3 presents the average R-squared and F-statistics and ranks them in order from
highest to lowest. Although there is little difference between the securities in relation to their
ability to explain the returns of other securities, the ICE quarterly coal futures appear to have a
marginally better explanatory power, while the EEX annual futures have the better joint
significance of regression coefficients. These two securities are also more heavily traded than
the monthly securities (albeit within the limited trading of coal futures in general)63
. This
provides some evidence that these securities are marginally better sources of price discovery in
the European coal market in terms of short-term return dynamics.
Table 3.3
Summary Regression Statistics and Ranking for Coal
Table 3.3 reports the average R-squared and robust F-statistics from the regression results
in Table 3.2. Averages are for the case in which each security is the independent variable in
the regressions. The average summary statistics are ranked from largest (1) to smallest (4).
Table 3.4 presents Hasbrouck‘s (1995) information shares for the coal securities. The results
show that, contrary to the short-term return dynamics, the variance contribution to long-run
equilibrium price innovations common to the series is greatest for the ICE monthly coal futures.
This presents something of a quandary as to which security is the better indicator of price
discovery in the European coal market as it cannot unequivocally be said which of the
dynamics, short or long-run, are more important. However, in the absence of more frequent
data, which would facilitate a more comprehensive consideration of the short-run return
dynamics between the securities, we are inclined to give greater weight to the long-run
dynamics captured by the information shares.
63
The Thomson Reuters data indicate that for the front contracts in our 2008-2011 sample period 525
contracts of the ICE quarterly futures were traded, 91 contracts of the EEX annual futures, 65 contracts of
the ICE monthly futures, while there were no trades recorded of the EEX monthly futures.
ICE Monthly ICE Quarterly EEX Monthly EEX Annual
Average R-squared 0.41 0.53 0.37 0.45
F-statistic 14.02 56.48 42.92 57.13
Rank R-squared 3 1 4 2
F-statistic 4 2 3 1
Independent Variable
72
Table 3.4
Information Shares for European Coal
Hasbrouck‘s (1995) information shares (i
IS ) are calculated using the log of daily settlement price levels
between 2 January 2008 and 30 December 2011 (960 observations) for the four coal futures as per:
ΨΨΩ
2
ii
MIS
(3.8)
The elements of the ( K1 ) row vector Ψ are the sums of each security‘s coefficients in the moving
average impact matrix (i
). The lower triangular matrix from a Cholesky factorisation ( M ) of the
VECM‘s contemporaneous error term variance-covariance matrix ( Ω ) is used to construct the variance
contribution of each security, ([ M ]i )2, to the variance of the common efficient price ( ΨΨΩ ). The
multivariate VECM specification from equation (3.3) contains 2 lags in theit
y vector as suggested by
the autocorrelation coefficients and Schwarz‘s Bayesian Information Criterion. The order of the series in
the VECM log price vector (t
y ) is cycled such that the VECM is run with each series taking a turn at
being the first series in the log price vector. As the order is cycled and the VECM re-run, the information
shares are re-calculated. Shading indicates the information share that is typically, though not always, the
maximum out of the four cycles because this cycle is the one in which that series is ordered first (as listed
in the first column of the table). The range of the information shares for each variable and the average
information share calculated across the four cycles are displayed along with the ordinal ranking of the
averages.
Series Ordered First in Cycle ICE Monthly Futures ICE Quarterly Futures EEX Monthly Futures EEX Annual Futures
ICE Monthly Futures 0.661 0.275 0.029 0.035
ICE Quarterly Futures 0.357 0.271 0.349 0.023
EEX Monthly Futures 0.611 0.033 0.356 0.000
EEX Annual Futures 0.630 0.270 0.049 0.051
Maximum 0.661 0.275 0.356 0.051
Minimum 0.357 0.033 0.029 0.000
Average 0.565 0.212 0.196 0.027
Rank 1 2 3 4
73
3.3 Natural Gas
In Europe natural gas is predominantly priced under two different systems best understood by
comparing Europe‘s two largest gas markets, namely Germany and the UK. In Germany, natural
gas is mostly priced under long-term, oil-indexed supply contracts. These agreements are
complicated, but typically run for 10 to 30 years and include price renegotiation clauses together
with a limited amount of volume flexibility at the discretion of the gas purchaser. The most
transparent price in this system is the German Border Price, which is a volume-weighted
average price, published monthly. The German Border Price is not a suitable indicator for
natural gas price discovery as it is infrequently released and largely linked to oil prices through
formulas specified in the supply contracts (usually crude oil, gasoil and/or heavy fuel oil
assessed at a 6 to 9-month lag and sometimes combinations of these that also include an
inflation indexation component)64
.
In contrast, the UK natural gas market has been characterised by market pricing at the
National Balancing Point (NBP) hub, a virtual trading location, since the mid-1990s. Market
pricing for gas contracts has since spread and become increasingly important in continental
Europe, particularly since bi-directional pipelines connecting the UK‘s NBP to Belgium‘s
Zeebrugge hub and to the Dutch Title Transfer Facility (TTF) became operational in 1998 and
2006, respectively65
. Market pricing now stretches into the rest of Europe through
interconnecting pipelines with major hubs in France at Point d‘Exchange de Gaz (PEG) Nord
and Sud and in Germany at Gaspool and Netconnect Germany (NCG). However, the spot (day-
ahead), forward (month-ahead) and associated derivative securities at the more established
pricing hubs in the UK, Belgium and the Netherlands remain the most liquid markets and the
likely points of price discovery for natural gas in Europe. As such, we examine price discovery
64
Asche, Osmundsen and Tveterås (2002) comprehensively describe these long-term take-or-pay
contracts.
65 Melling (2010) provides a thorough overview of the development of European natural gas markets and
the factors underlying the increasing prevalence of market pricing in continental Europe.
74
between NBP day-ahead and month-ahead prices, Zeebrugge day-ahead and month-ahead
prices, TTF day-ahead prices and monthly expiry ICE UK natural gas futures contracts66
.
The natural gas prices considered display much greater liquidity than the coal prices
considered in the previous section and, as such, they warrant a more informative, high
frequency analysis to determine the best venue for price discovery67
. The TTF day-ahead series
are converted from euro into Great British pence using intraday EBS exchange rates, which
were also sourced from Thomson Reuters, though interestingly the results are not significantly
different when the regressions are run with each series in its native currency68
. Descriptive
statistics for the six natural gas return series are displayed in Table 3.5. The range of returns for
the natural gas securities is much larger than for the coal securities reflecting their much greater
volatility.
Table 3.5
Descriptive Statistics for Natural Gas Returns
Table 3.5 reports descriptive statistics calculated in respect of continuously compounded returns for the
natural gas securities sampled at 10-minute intervals from 2 January 2008 to 30 December 2011.
66
Other securities examined but unreported here for brevity and which were shown not to be sources of
price discovery, include: PEG Nord day-ahead, ICE TTF futures, ICE Gaspool futures and ICE NCG
futures.
67 The regressions were run on an hourly, 10-minute, 5-minute and 1-minute basis. The reported results
are for the 10-minute series as there were too many zero return observations in the higher frequency series
and the hourly series did not fully utilise the frequency of return observations available. Even so, the
results did not differ substantially at these other intervals. The regressions were run for 10 lagged returns
as independent variables (i.e. capturing leading and lagging behaviour of just over an hour and a half).
68 As shown in the Appendix, Table B1, all the series, except the ICE natural gas futures, display
significant autocorrelation. This justifies the use of Newey-West (1987) in the regression approach.
According to the augmented Dickey-Fuller (1979) test statistics in Table B2, all of the securities are I(1)
variables; non-stationary in levels but stationary in returns. Johansen (1995) tests show that they are all
cointegrated at the 1 per cent level of statistical significance (Table B3).
NBP NBP TTF Zeebrugge Zeebrugge ICE
Month Day Day Month Day Monthly
Ahead Ahead Ahead Ahead Ahead Futures
Frequency 10-minute 10-minute 10-minute 10-minute 10-minute 10-minute
Mean 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
Standard Deviation 0.0055 0.0093 0.0081 0.0052 0.0096 0.0050
Skewness 12.20 -1.80 -2.44 8.16 -5.80 15.31
Kurtosis 826.05 566.10 134.95 624.64 457.13 1,000.59
Maximum 0.3466 0.5431 0.1598 0.3091 0.2944 0.3443
75th
Percentile 0.0000 0.0000 0.0008 0.0000 0.0000 0.0000
25th
Percentile 0.0000 0.0000 -0.0008 0.0000 0.0000 0.0000
Minimum -0.1823 -0.4844 -0.2984 -0.1493 -0.4715 -0.1491
75
The results in Table 3.6 show that the lagged return coefficients are most often positive and
significant when the NBP month-ahead, TTF day-ahead or ICE monthly futures series are the
explanatory variables, but less so for the other series. The strongest relationship exists between
the NBP month-ahead and ICE monthly futures returns, which is unsurprising given their
similar tenor and the physical deliverability of the futures.
As expected, in light of the more granular intraday interval, the contemporaneous return
coefficients, R-squared and robust F-statistics are lower than in the coal futures regressions.
Notwithstanding this, these measures are also very small in their own right, with the exception
of the aforementioned linkages between NBP month-ahead natural gas and the ICE monthly
futures. This weakness between the physical markets gives some indication that significant
market frictions exist between the various natural gas hubs in Europe, potentially resulting from
different storage and pipeline capacities at and between the hubs and the time it takes to move
natural gas from one market to another. However, if these frictions were geographically
specific, it would be expected that there would be stronger contemporaneous return coefficients
between securities on the same hub (for example, the Zeebrugge day-ahead and month-ahead
regressions). Puzzlingly, this does not seem to be the case. Rather, there appear to be stronger
forward curve relationships, with day-ahead series having higher contemporaneous coefficients
when regressed against other day-ahead series and month-ahead series similarly having higher
coefficients with other month-ahead series. This suggests the possibility that short-term supply
disruptions or demand spikes might drive different dynamics at the front of the curve compared
with further out and that these dynamics are felt in common at various hubs with some degree of
simultaneity. For example, perhaps the impact of unanticipated cold weather could drive a spike
in natural gas demand at several North West European gas hubs at approximately the same time,
affecting day-ahead markets more than further out the curve. In this sense, the day-ahead
markets display demand inelasticity that is similar to electricity prices, which is not surprising
given that demand for power and natural gas are both related to the demand for heating.
Consistent with their greater susceptibility to short-term factors, the volatilities of all the day-
ahead series, as displayed in Table 3.5, are roughly twice that of the month-ahead and futures
securities.
76
Table 3.6
Price Discovery Regressions for European Natural Gas
Table 3.6 presents the results of fitting model (3.1) using 10-minute continuously compounded returns between 8:00am and 4:00pm GMT from 2 January 2008 to
30 December 2011 (47,078 return observations). In Panel A, the independent variables are contemporaneous and lagged returns of NBP month-ahead natural gas
(series B) and an error correction term, with the response variable being one of the five other series of interest (series A). Panels B, C, D, E and F contain results from
using the other five securities as independent variables. Formally:
ttAzAktB
k
ktA zRR
1,,,
0
10
,
(3.1)
The error correction terms are given by: 1,tA
z ln(1, tA
p ) – ln(1, tB
p ), the one-period lag of the difference in log prices between the two series. Square brackets [ ]
below coefficients contain t-statistics, while round brackets ( ) below F-statistics contain p-values. * and ** denote significance at the 5 and 1 per cent levels.
α βt βt-1 βt-2 βt-3 βt-4 βt-5 βt-6 βt-7 βt-8 βt-9 βt-10 βZ R-squared F-statistic
Panel A:
0.0001** 0.4382** 0.1124** 0.0843** 0.0537** 0.0359** 0.0293** 0.0136 0.0271** 0.0156* 0.0179* -0.0041 -0.0041** 0.074 8.18**
[4.37] [5.43] [5.19] [5.10] [3.60] [3.06] [3.00] [1.11] [3.00] [2.10] [2.24] [-0.43] [-4.17] (0.000)
0.0000 0.2290** 0.0691** 0.0477** 0.0308* 0.0241** 0.0224* 0.0378** 0.0278* 0.0208** -0.0003 0.0034 -0.0042** 0.029 12.47**
[1.02] [3.83] [4.89] [4.77] [2.23] [2.64] [2.34] [2.79] [2.29] [2.61] [-0.04] [0.47] [-5.87] (0.000)
-0.0001** 0.2482** 0.0486** 0.0244** 0.0161 0.0213** 0.0227** 0.0319 0.0162* 0.0165** 0.0164** 0.0231** -0.0099** 0.072 7.49**
[-3.86] [3.37] [3.89] [4.66] [1.84] [3.40] [4.80] [1.73] [2.35] [2.75] [2.92] [3.84] [-6.62] (0.000)
0.0001* 0.1688** 0.0483** 0.0324** 0.0375** 0.0373** 0.0369** 0.0251* 0.0393** 0.0203* 0.0502* 0.0283** -0.0050** 0.014 7.17**
[2.50] [2.81] [4.12] [3.25] [4.00] [4.58] [3.57] [2.04] [3.67] [2.40] [2.16] [2.96] [-4.78] (0.000)
0.0001 0.6101** 0.0864** 0.0390** 0.0111* 0.0055 0.0048 0.0038 0.0062 0.0002 0.0018 0.0016 -0.3455** 0.500 33.44**
[6.29] [11.48] [7.69] [4.00] [2.53] [1.23] [1.27] [0.94] [1.88] [0.07] [0.52] [0.44] [-10.22] (0.000)
Panel B:
-0.0001** 0.1519** 0.0155** 0.0149** 0.0054 -0.0072 0.0021 0.0002 0.0025 0.0023 -0.0004 0.0012 -0.0022** 0.067 6.23**
[-3.63] [5.97] [3.83] [2.75] [1.09] [-1.29] [0.82] [0.07] [0.76] [0.80] [-0.16] [0.35] [-5.16] (0.000)
-0.0002** 0.2360** 0.0291** 0.0238** 0.0141 -0.0009 0.0055 0.0056 0.0054 0.0003 0.0077 -0.0005 -0.0093** 0.078 12.57**
[-6.74] [6.35] [3.12] [2.84] [1.59] [-0.08] [0.79] [0.71] [0.92] [0.06] [1.34] [-0.07] [-8.07] (0.000)
-0.0001* 0.0664** 0.0105** 0.0091** 0.0053 0.0049 0.0044 -0.0042 0.0076* 0.0064* 0.0084** 0.0077** -0.0013** 0.015 6.29**
[-2.35] [3.61] [3.40] [3.22] [1.05] [1.37] [1.58] [-0.47] [2.27] [2.00] [2.74] [2.79] [-3.45] (0.000)
-0.0002** 0.3057** 0.0394** -0.0309 0.0127 0.0336* 0.0362** -0.0162 0.0253* 0.0149 0.0360** 0.0208 -0.0149** 0.096 10.15**
[-4.54] [6.06] [3.81] [-0.51] [0.84] [2.38] [2.98] [-0.66] [2.33] [1.40] [3.27] [1.66] [-5.46] (0.000)
-0.0001** 0.1132** 0.0154** 0.0054* 0.0060 -0.0056 0.0034 0.0020 -0.0008 0.0002 0.0009 0.0036 -0.0022** 0.045 5.67**
[-3.44] [4.87] [3.36] [2.12] [1.32] [-1.24] [1.27] [0.67] [-0.30] [0.09] [0.39] [1.10] [-5.37] (0.000)
Independent Variable: NBP Day Ahead
Dep
end
ent
Va
ria
ble
NBPMonth
Ahead
TTF Day
Ahead
Zeebrugge
Month Ahead
Zeebrugge
Day Ahead
ICE Monthly
Futures
Independent Variable: NBP Month Ahead
Dep
end
ent
Va
ria
ble
NBP Day
Ahead
TTF Day
Ahead
Zeebrugge
Month Ahead
Zeebrugge
Day Ahead
ICE Monthly
Futures
77
Table 3.6
Price Discovery Regressions for European Natural Gas (Continued)
α βt βt-1 βt-2 βt-3 βt-4 βt-5 βt-6 βt-7 βt-8 βt-9 βt-10 βZ R-squared F-statistic
Panel C:
0.0000 0.1092** 0.0507** 0.0135** 0.0119** 0.0043 0.0030 0.0049 0.0007 0.0057 0.0053 0.0037 -0.0024** 0.030 8.84**
[-0.74] [3.61] [5.19] [3.53] [2.60] [0.78] [0.90] [1.84] [0.22] [1.73] [1.70] [1.10] [-5.20] (0.000)
0.0003** 0.3250** 0.1195** 0.0656** 0.0380* 0.0075 0.0348** 0.0118 0.0353** 0.0126* 0.0015 0.0024 -0.0105** 0.090 15.81**
[6.11] [7.37] [5.41] [5.69] [2.20] [0.61] [4.10] [1.29] [4.37] [2.11] [0.18] [0.16] [-5.07] (0.000)
0.0000 0.0617* 0.0113** 0.0095** 0.0047 -0.0007 0.0121** -0.0010 0.0113* 0.0091* 0.0125** 0.0071* -0.0014** 0.010 4.09**
[-1.07] [2.53] [2.75] [3.02] [1.18] [-0.13] [3.76] [-0.11] [2.04] [2.51] [3.01] [2.06] [-3.29] (0.000)
0.0003** 0.2316** 0.0545** 0.0606** 0.0332** 0.0367** 0.0489** 0.0281 0.0513** 0.0309** 0.0344** 0.0485** -0.0254** 0.047 11.44**
[6.26] [5.14] [5.45] [5.38] [2.97] [3.71] [5.46] [1.55] [6.12] [3.27] [3.34] [3.36] [-7.71] (0.000)
0.0000 0.1002** 0.0210** 0.0076* 0.0087* 0.0052 0.0008 0.0032 0.0017 0.0032 0.0020 0.0048 -0.0022** 0.027 7.74**
[-0.66] [3.34] [5.27] [1.98] [2.10] [0.99] [0.26] [1.09] [0.57] [1.23] [0.75] [1.57] [-5.00] (0.000)
Panel D:
0.0002** 0.2704** 0.0186 0.0088 0.0022 0.0038 0.0001 -0.0008 -0.0047 0.0120* -0.0016 0.0021 -0.0172** 0.075 11.88**
[5.85] [3.01] [1.96] [1.71] [0.31] [0.57] [0.02] [-0.15] [-1.02] [2.43] [-0.40] [0.47] [-11.36] (0.000)
0.0002** 0.2114** 0.0379 0.0119 0.0167 0.0141 0.0077 0.0130 0.0056 0.0186* 0.0040 0.0042 -0.0044** 0.016 3.60**
[4.62] [3.55] [1.34] [1.54] [1.50] [1.35] [0.67] [1.23] [0.68] [1.98] [0.54] [0.39] [-4.43] (0.000)
0.0001** 0.1448** 0.0118 0.0114 0.0215* -0.0015 0.0132 0.0118 0.0041 0.0014 -0.0025 -0.0009 -0.0047** 0.011 6.06**
[2.60] [2.68] [1.37] [1.33] [1.96] [-0.15] [1.21] [1.44] [0.47] [0.20] [-0.34] [-0.15] [-6.42] (0.000)
0.0001** 0.2314** 0.0117 0.0161 0.0229** 0.0100 0.0110 -0.0019 0.0042 0.0032 0.0126 0.0020 -0.0056** 0.019 4.88**
[4.06] [3.93] [0.59] [1.20] [2.83] [1.14] [1.06] [-0.14] [0.62] [0.36] [1.58] [0.23] [-5.15] (0.000)
0.0001** 0.2512** 0.0157** 0.0048 0.0067 0.0035 -0.0014 0.0051 0.0016 0.0077 0.0012 0.0002 -0.0152** 0.076 13.13**
[5.58] [2.84] [2.89] [0.88] [1.05] [0.83] [-0.20] [1.04] [0.42] [1.71] [0.30] [0.05] [-12.22] (0.000)
Independent Variable: Zeebrugge Month Ahead
Dep
en
den
t V
ari
ab
le
NBPMonth
Ahead
NBP Day
Ahead
TTF Day
Ahead
Zeebrugge
Day Ahead
ICE Monthly
Futures
Independent Variable: TTF Day Ahead
Dep
en
den
t V
ari
ab
le
NBPMonth
Ahead
NBP Day
Ahead
Zeebrugge
Month Ahead
Zeebrugge
Day Ahead
ICE Monthly
Futures
78
Table 3.6
Price Discovery Regressions for European Natural Gas (Continued)
α βt βt-1 βt-2 βt-3 βt-4 βt-5 βt-6 βt-7 βt-8 βt-9 βt-10 βZ R-squared F-statistic
Panel E:
-0.0000* 0.0556** 0.0098 0.0118** -0.0029 0.0003 -0.0004 0.0057* -0.0007 -0.0019 -0.0020 0.0051 -0.0029** 0.012 4.94**
[-2.25] [2.74] [1.53] [2.94] [-0.54] [0.08] [-0.16] [2.18] [-0.29] [-0.60] [-0.62] [1.62] [-6.74] (0.000)
0.0002** 0.2874** 0.0360 -0.0015 0.0184 0.0099 0.0089 0.0042 -0.0014 0.0038 0.0060 -0.0052 -0.0174** 0.094 6.47**
[5.41] [5.72] [1.58] [-0.18] [1.24] [0.89] [1.82] [0.43] [-0.25] [0.60] [0.89] [-0.34] [-4.81] (0.000)
-0.0003** 0.1556** 0.0068 0.0175** 0.0185* 0.0068 0.0166** 0.0112 -0.0002 0.0077 0.0082 0.0064 -0.0250** 0.047 21.00**
[-7.78] [5.09] [0.94] [2.70] [2.31] [1.34] [3.67] [1.82] [-0.03] [1.86] [1.48] [1.30] [-15.08] (0.000)
-0.0000* 0.0701** 0.0126** 0.0155** 0.0135* 0.0025 0.0096* 0.0031 0.0068* 0.0020 0.0051** 0.0024 -0.0016** 0.018 5.03**
[-2.05] [3.89] [4.29] [3.29] [2.08] [1.03] [2.27] [0.40] [2.16] [0.97] [2.67] [0.91] [-4.05] (0.000)
-0.0000* 0.0463* 0.0040 0.0039 -0.0007 -0.0034 0.0006 0.0021 -0.0025 0.0018 0.0004 0.0027 -0.0027** 0.010 4.25**
[-2.08] [2.22] [1.68] [1.35] [-0.14] [-1.05] [0.20] [0.98] [-0.97] [0.68] [0.19] [1.04] [-6.57] (0.000)
Panel F:
-0.0001** 0.7609** 0.1942** 0.0670** 0.0372** 0.0403 0.0197** 0.0133* 0.0093 0.0114** -0.0021 0.0135** -0.3722** 0.503 173.06**
[-6.36] [18.52] [6.48] [4.54] [2.75] [1.89] [2.85] [2.22] [1.49] [2.56] [-0.51] [3.20] [-9.94] (0.000)
0.0001** 0.3813** 0.1117** 0.1089** 0.0908** 0.0360 0.0496** 0.0185 0.0289** 0.0232** 0.0058 0.0033 -0.0042** 0.055 8.29**
[4.32] [4.44] [4.54] [5.62] [4.19] [1.68] [4.29] [1.10] [3.11] [2.79] [0.58] [0.31] [-4.27] (0.000)
0.0000 0.2494** 0.0887** 0.0492** 0.0565** 0.0332** 0.0459** 0.0263 0.0253* 0.0176* 0.0005 0.0030 -0.0041** 0.033 12.54**
[0.94] [3.48] [5.22] [4.59] [2.90] [2.88] [3.05] [1.61] [2.45] [1.98] [0.06] [0.41] [-5.79] (0.000)
-0.0001** 0.2705** 0.0253** 0.0239** 0.0124 0.0264** 0.0168** 0.0444* 0.0181** 0.0121 0.0135* 0.0213** -0.0095** 0.072 7.97**
[-3.81] [3.30] [4.33] [4.18] [1.29] [3.63] [3.53] [2.08] [2.73] [1.75] [2.55] [3.08] [-6.54] (0.000)
0.0001* 0.1655* 0.0492** 0.0365** 0.0470** 0.0488** 0.0418** 0.0163 0.0432** 0.0278* 0.0350* 0.0217* -0.0050** 0.013 7.48**
[2.40] [2.29] [3.56] [3.14] [4.44] [4.50] [4.12] [1.14] [3.60] [2.54] [2.40] [2.03] [-4.74] (0.000)
Independent Variable: ICE Monthly Futures
Dep
en
den
t V
ari
ab
le
NBPMonth
Ahead
NBP Day
Ahead
TTF Day
Ahead
Zeebrugge
Month Ahead
Zeebrugge
Day Ahead
Independent Variable: Zeebrugge Day Ahead
Dep
en
den
t V
ari
ab
le
NBPMonth
Ahead
NBP Day
Ahead
TTF Day
Ahead
Zeebrugge
Month Ahead
ICE Monthly
Futures
79
Given the large number of comparisons involved in visual inspection of regression
coefficients, we also provide the average R-squared and F-statistics. Table 3.7 displays the
average fit and joint significance statistics from the regressions in Table 3.6 together with their
ordinal ranking. These indicate that both the NBP month-ahead and ICE monthly futures returns
are strong explanators of the short-term return behaviour of the other securities. However, given
the relative ranking of their average R-squared and F-statistics, it is difficult to distinguish
between them and reach a conclusion regarding which of the two securities is the better security
for short-term price discovery.
Table 3.7
Summary Regression Statistics and Ranking for Natural Gas
Table 3.7 reports the average R-squared and robust F-statistics from the natural gas regression results
presented in Table 3.6. Averages are for the case in which each security is the independent variable in the
regressions. The average summary statistics have been ranked from largest (1) to smallest (6).
Table 3.8 presents the information shares for the natural gas securities. The ICE monthly
futures clearly have the greatest variance contribution to innovations in the long-run equilibrium
price, on average accounting for 68.1 per cent of the common innovations. The next most
important contributions come from the NBP month-ahead and TTF day-ahead markets but these
averages are only 10.4 per cent and 10.3 per cent, respectively. This indicates that while these
different regional markets may only be weakly linked in terms of short-term responsiveness of
prices, they are nonetheless cointegrated (tests detailed in the Appendix), with innovations in
prices in the mature UK natural gas futures market contributing most to the ultimate, common
equilibrium between the markets. It would also appear that for all the volatility in the various
day-ahead prices, much of this is of a transitory nature, with the somewhat more stable futures
trading activity better incorporating information relevant to natural gas prices in Europe.
NBP NBP TTF Zeebrugge Zeebrugge ICE Monthly
Month-Ahead Day-Ahead Day-Ahead Month-Ahead Day-Ahead Futures
Average R-squared 0.138 0.060 0.041 0.039 0.036 0.135
F-statistic 13.75 8.18 9.58 7.91 8.34 41.87
Rank R-squared 1 3 4 5 6 2
F-statistic 2 5 3 6 4 1
Independent Variable
80
Table 3.8
Information Shares for European Natural Gas
Hasbrouck‘s (1995) information shares (i
IS ) are calculated from log price levels for the six natural gas
securities sampled at 10-minute intervals between 8:00am and 4:00pm GMT from 2 January 2008 to
30 December 2011 (47,085 observations) as per:
ΨΨΩ
2
ii
MIS
(3.8)
The elements of the ( K1 ) row vector Ψ are the sums of each security‘s coefficients in the moving
average impact matrix (i
). The lower triangular matrix from a Cholesky factorisation ( M ) of the
VECM‘s contemporaneous error term variance-covariance matrix ( Ω ) is used to construct the variance
contribution of each security, ([ M ]i )2, to the variance of the common efficient price ( ΨΨΩ ). The
multivariate VECM specification from equation (3.3) contains 4 lags in theit
y vector in order to deal
with any minor autocorrelation. This lag length was selected using Schwarz‘s Bayesian Information
Criterion. The order of the series in the VECM log price vector (t
y ) is cycled such that the VECM is run
with each series taking a turn at being the first series in the log price vector. As the order is cycled and
the VECM re-run, the information shares are re-calculated. Shading indicates the information share that
is typically, though not always, the maximum out of the six cycles because this cycle is the one in which
that series is ordered first (as listed in the first column of the table). The range of the information shares
for each variable and the average information share calculated across the six cycles are displayed along
with the ordinal ranking of the averages.
Commonalities in large natural gas and crude oil market participants, oil and gas
fundamentals and linkages between gas and oil prices in long-term supply contracts mean there
is some prospect that even the more liquid natural gas hub prices may track movements in crude
oil prices. Consistent with this, Asche et al. (2006) find the UK natural gas and Brent crude oil
markets were cointegrated between 1995 and 1998, with Brent displaying price leadership over
this time. However, their testing suggests that this relationship broke down after the
Interconnector pipeline between Bacton (UK) and Zeebrugge (Belgium) opened in 1998.
Unfortunately, the Asche et al. (2006) study employs a very small data set containing only 42
NBP NBP TTF Zeebrugge Zeebrugge ICE Monthly
Month-Ahead Day-Ahead Day-Ahead Month-Ahead Day-Ahead Futures
NBP Month-Ahead 0.524 0.027 0.102 0.034 0.002 0.311
NBP Day-Ahead 0.015 0.062 0.139 0.118 0.007 0.660
TTF Day-Ahead 0.023 0.000 0.146 0.124 0.003 0.703
Zeebrugge Month-Ahead 0.018 0.038 0.076 0.141 0.000 0.727
Zeebrugge Day-Ahead 0.022 0.041 0.077 0.015 0.002 0.843
ICE Monthly Futures 0.023 0.037 0.078 0.016 0.001 0.845
Maximum 0.524 0.062 0.146 0.141 0.007 0.845
Minimum 0.015 0.000 0.076 0.015 0.000 0.311
Average 0.104 0.034 0.103 0.075 0.002 0.681
Rank 2 5 3 4 6 1
Series Ordered First in
Cycle
81
monthly observations. A more extensive study by Panagiotidis and Rutledge (2007) shows
monthly UK natural gas and Brent crude prices were cointegrated throughout the 1996-2003
period.
Against this backdrop, we examine the linkages between markets by testing for
cointegration between the monthly expiry ICE UK natural gas futures prices and monthly expiry
ICE Brent crude oil futures at 10-minute intervals between 2008 and 2011. The Johansen (1995)
test results in Table 3.9 provide further evidence that the security prices are cointegrated at the
5 per cent level of statistical significance and that some linkages between these markets remain.
Table 3.9
Cointegration Test: Natural Gas and Crude Oil
Table 3.9 displays Johansen (1995) cointegration tests of log prices of monthly expiry ICE UK natural
gas futures and monthly expiry ICE Brent crude oil futures sampled at 10-minute intervals from 8:00am
to 4:00pm GMT between 2 January 2008 and 30 December 2011 (47,038 observations). Tests are run
using a multivariate VECM estimated by maximum likelihood with 2 lags ( 2n ) as explanatory
variables to determine the number of cointegrating relations ( r ) between the ( K ) variables:
tit
n
i
itt εyΓyβαy
1
1
1
(3.3)
Dependent variables (t
y ) are a 1K vector of differenced log price levels; it
y are lagged dependent
variables; α and β are rK parameter matrices in which the number of cointegrating equations is less
than the number of I(1) variables ( Kr ); 11
,,p
ΓΓ are KK matrices of parameters; and, tε is a
1K vector of normally distributed and serially uncorrelated error terms. The null hypothesis for the
trace statistic is that there are no more than r cointegrating relations (i.e. the eigenvalues Kr ,,
1
are
zero). * and ** denote the rank at which the null hypothesis cannot be rejected at the 5 and 1 per cent
levels of statistical significance, respectively (critical values are from the tables in Johansen, 1995).
Maximum Rank Eigenvalue Trace Statistic
5% 1%
r ≤ 0 14.40** 12.21 16.16
r ≤ 1 0.00023 3.45* 4.14 7.02
r ≤ 2 0.00007
Critcal Values
82
3.4 Crude Oil
Our analysis of price discovery in European crude oil markets initially focuses on the complex
of securities related to North Sea crude oils (―Brent‖). We then provide evidence on a topic
receiving significant attention in the literature in recent years, namely the relationship between
the two key global crude oil benchmarks. Specifically, we consider whether Brent and West
Texas Intermediate (WTI) are cointegrated and, if so, which displays greater price leadership.
Our evidence is particularly relevant given the growing dislocation between these markets due
to structural issues at the WTI pricing point in Cushing, Oklahoma.
3.4.1 Price Discovery in the Brent Crude Oil Complex
Market pricing has overwhelmingly dominated crude oil transactions since the mid-1980s,
with long-term contracts usually pegged to the market prices of certain key benchmark crude oil
grades69
. The key benchmarks are typically constructed by price reporting agencies from over-
the-counter spot, forward and swap prices70
. Futures contracts are linked to these physical
markets by settlement against an index of forward prices. Futures are more liquid as they allow
for smaller trade sizes than the physical market and facilitate hedging and speculation
unimpeded by the logistical concerns of physical delivery.
The key benchmark for crude oil pricing in Europe is dated BFOE, which is interchangeably
referred to as dated Brent, though technically these differ71
. Dated BFOE refers to light, sweet
crude oil from a number of defined North Sea fields and, although it is often called a spot
69
The various grades of crude oil trade at different prices to reflect viscosity and sulphur content as these
yield differing quantities of consumable petroleum products and require different amounts of distillation
and refining. Long-term contracts may specify fixed spreads or specify reliance on spreads from price
reporting agencies such as Platts that are appropriate for the particular grade contracted for, relative to the
benchmark grade price.
70 Note that what are often termed ‗spot‘ transactions in North Sea crude oil markets necessarily involve
degrees of forwardness as required by the logistical considerations of physical delivery. Contracted ‗spot‘
prices are frequently set at the time oil cargoes are loaded, with reference to a particular benchmark.
71 Brent is the original name for oil from specific fields collected through a pipeline system that is
connected to a terminal at Sullom Voe in the Shetland Islands. Declining supplies from the original Brent
field first led to comingling with oil from the Ninian field and has subsequently led to an expansion of the
benchmark definition such that it now includes oil from the Forties, Oseberg and Ekofisk fields (thus the
acronym BFOE). Because the quality of oil from these different fields varies, the poorest quality oil sets
the price for dated BFOE, which since 2007 has usually been oil from the Buzzard field (part of the
Forties).
83
market, it is an oil cargo with a specific loading slot for delivery in the next 10 to 25 days72
.
Fattouh (2011) postulates that ICE Brent futures drive forward BFOE prices through the
exchange-for-physicals swap market. These markets in turn influence (or are influenced by)
dated BFOE prices through the contract-for-difference market according to transactions during
the daily price assessment window of the major price reporting agencies (Platts and Argus).
Barret (2012) describes in detail how the price reporting agencies have come to have a large
impact on physical market interactions. The price transparency of the agencies‘ daily
assessment windows draws in trading volume and creates liquidity, though there have recently
been suggestions that the major oil companies are drawn to trade in these windows in order to
influence the benchmark prices that may apply to their long-term contractual crude oil sales
(IOSCO, 2012). In this way, the assessment window is not unlike a daily auction for forward,
contract-for-difference and dated transactions, the end result of which is a dated BFOE price
published at 4:30pm London time.
While dated BFOE is the key benchmark grade, the ICE Brent futures contracts are likely the
more important source of price discovery, especially outside the daily assessment window used
by most price reporting agencies between 4:00 and 4:30pm London time. Unfortunately, testing
this hypothesis is difficult because dated, forward and contract-for-difference BFOE are over-
the-counter markets for which high frequency data is not publically available. An indication of
prices in these markets can be gleaned from price reporting agency data such as Platts and
Argus, but this data is only available on a daily basis, which makes for an unappealing
72
The dated BFOE price applicable in most long-term contracts is itself a price determined by a price
reporting agency, such as Platts, that is backed-out from trades in the forward BFOE and contract-for-
difference markets during a daily assessment window (Barret, 2012). Thus the physical BFOE markets,
which in combination set reported dated BFOE prices, are potentially important sources of price
discovery, despite trading being much less frequent than in the futures market. As well as differences in
trade frequency, the financial and physical markets also differ greatly in trade size. The minimum
physical trade size is large at 100,000 barrels (a partial cargo), while most futures trades are at the 1,000
barrel minimum. To put these figures in context, average daily production for Norway and the UK in
2011 was approximately 3.15 million barrels per day (Blas, 2012a). Although the minimum shipment size
acts as a prohibitive barrier to physical BFOE market entry, such that there are typically less than a dozen
market participants at any given time, these participants, the major global oil companies, are some of the
best informed, particularly on matters concerning supply (see Fattouh, 2011).
84
comparison considering that the ICE Brent futures contract is one of the most liquid securities in
the world73
.
Instead of using low frequency survey price data from the price reporting agencies, we use a
proxy for the dated BFOE market calculated by Thomson Reuters. The dated BFOE proxy
series adjusts the near contract ICE Brent futures price by adding an exchange-for-physicals
(EFP) contract and cash market inter-month spread74
. As the major global oil companies, who
are the dominant participants in the actual dated, forward and contract-for-difference BFOE
markets, also transact between hedged futures positions and forward BFOE in the EFP market,
this proxy at least provides some basis for assessing the impact of those important physical
market participants upon price discovery.
Descriptive statistics for the dated BFOE proxy and ICE Brent futures sampled at 10-second
intervals between 7:00am and 5:00pm GMT from 2 January 2008 to 30 December 2011 are
detailed in Table 3.10, along with the descriptive statistics for the Chicago Mercantile Exchange
(CME) WTI futures used in the analysis in the next section75
. The range and volatility of returns
for the crude oil securities is much lower than those presented for the natural gas securities in
the previous section. Though it is economically unimportant, the statistically significant
autocorrelation indicated by the tests in the Appendix, Table C2, justifies the use of Newey-
West (1987) in the regression approach and the inclusion of a large number of lags in the
VECM specification used to calculate the information shares76
.
73
It should be noted that the Platts assessment process itself is timely when published and observable by
oil market participants in real time should trades and quotes be posted during this half-hour period (see
Barret, 2012).
74 Proxies for forward BFOE prices could similarly be constructed by either omitting or adjusting the cash
market inter-month spread. However, these spreads are only altered on a daily basis (if at all) and so high
frequency analysis of a forward BFOE proxy alongside the Thomson Reuters dated BFOE proxy would
not be a fruitful exercise as these only differ by a constant, albeit one the is adjusted every day or so. In
this sense, we are using the intraday Thomson Reuters dated BFOE series, which is mainly an EFP
construction, as a proxy for physical market activity in general.
75 Numerous studies term these the New York Mercantile Exchange (NYMEX) WTI futures, however,
we term them the CME WTI futures following the CME‘s acquisition of NYMEX in 2008. Analysis of
this data at 1-minute intervals was also conducted, but these results are largely the same as the 10-second
results presented here and, as such, only the results from the narrower intraday interval are detailed.
76 The augmented Dickey-Fuller (1979) test results displayed in the Appendix, Table C3, indicate that
both series are I(1) variables, stationary in returns but non-stationary in price levels, while the Johansen
(1995) cointegration test results in Table C4 confirm that the series are strongly cointegrated.
85
The regression results presented in Table 3.11 point to the ICE Brent crude oil futures
leading the physical market inasmuch as the Thomson Reuters‘ EFP construction (our dated
BFOE proxy) is a valid representation of physical market activity. This is clear from visual
inspection of the coefficients as well as from the higher robust F-statistics. Most of the lagged
ICE Brent futures coefficients are positive and statistically significant determinants of the dated
BFOE proxy‘s contemporaneous returns, while only a couple of the lagged coefficients of the
dated BFOE proxy are significant when the regressions are run the other way around. The fact
that there are significant, positive coefficients at lags of 10 and 20-seconds when the dated
BFOE proxy is the independent variable suggests that price discovery is bi-directional and there
may be occasions when pertinent information from the EFP market subsequently impacts the
futures market. Interestingly, the contemporaneous return coefficients are very high, even at the
10-second sampling frequency, which attests to the depth of liquidity in these markets in stark
contrast to the liquidity of the coal and to some extent the natural gas markets77
.
Table 3.10
Descriptive Statistics for Crude Oil Returns
Table 3.10 displays descriptive statistics of continuously compounded returns for the crude oil securities
sampled at 10-second intervals between 2 January 2008 and 30 December 2011.
77
Even at such a high frequency as this 10-second sample, only 42.6 per cent of the return observations
for the ICE Brent futures are unchanged prices (zero returns), while for the dated BFOE proxy this figure
is 50.5 per cent.
ICE Dated ICE CME
Brent BFOE Brent WTI
Futures Proxy Futures Futures
Intraday Window (GMT) 7:00am-5:00pm 7:00am-5:00pm 7:00am-9:00pm 7:00am-9:00pm
Frequency 10-second 10-second 10-second 10-second
Observations 3,680,221 3,680,221 5,151,901 5,151,901
Mean 0.0000 0.0000 0.0000 0.0000
Standard Deviation 0.0004 0.0005 0.0004 0.0004
Skewness 2.14 13.66 0.70 0.76
Kurtosis 1,751.38 4,777.32 759.90 745.71
Maximum 0.0746 0.1201 0.0760 0.0811
75th
Percentile 0.0001 0.0000 0.0001 0.0001
25th
Percentile -0.0001 0.0000 -0.0001 -0.0001
Minimum -0.0569 -0.0708 -0.0420 -0.0487
86
Table 3.11
Price Discovery Regressions for European Crude Oil
Table 3.11 presents the results of fitting model (3.1) using continuously compounded returns sampled at 10-second intervals between 7:00am and 5:00pm GMT from
2 January 2008 to 30 December 2011 (3,683,820 return observations). The independent variables in (3.1) are alternately contemporaneous and lagged 10-second returns of the
ICE Brent futures and dated BFOE proxy (series B) as well as an error correction term, with the response variable being the other series of interest (series A). Formally:
ttAzAktB
k
ktA zRR
1,,,
0
10
,
(3.1)
The error correction terms are given by: 1,tA
z ln(1, tA
p ) – ln(1, tB
p ), the one-period lag of the difference in log prices between the two series. Square brackets [ ] below
coefficients contain t-statistics, while round brackets ( ) below F-statistics contain p-values. * and ** denote significance at the 5 and 1 per cent levels.
α βt βt-1 βt-2 βt-3 βt-4 βt-5 βt-6 βt-7 βt-8 βt-9 βt-10 βZ Adj. R-sqr F-stat
0.0000 0.6502** 0.1018** 0.0267** 0.0130** 0.0066** 0.0049** 0.0030** 0.0036** 0.0023** 0.0017* 0.0033** -0.0000** 0.374 689.79**
[-1.70] [47.35] [46.69] [24.25] [13.94] [7.85] [5.69] [3.54] [4.62] [3.09] [2.16] [4.18] [-2.63] (0.000)
0.0000 0.5653** 0.0215** 0.0035** 0.0007 0.0001 -0.0010 0.0003 0.0004 -0.0007 -0.0001 -0.0011 0.0000 0.365 265.99**
[0.83] [55.06] [18.07] [4.84] [1.10] [0.20] [-1.39] [0.41] [0.63] [-0.89] [-0.12] [-1.79] [-1.43] (0.000)Dep
en
den
t V
ari
ab
le Independent Variable: Brent Futures
Dated
BFOE Proxy
Independent Variable: Dated BFOE Proxy
Brent
Futures
87
Table 3.12 contains information shares and shows that the Brent futures variance
contribution to the long-run equilibrium price was greater than that of the dated BFOE proxy
though, again, with the caveat that such a proxy may not be an adequate representation of the
complexity of the physical markets. This result is even stronger in 6-month sub-samples over
the 2008-2011 period. The ICE Brent futures dominate price discovery in all eight of the 6-
month sub-periods, with their average information shares ranging from as high as 87.5 per cent
in the first half of 2008 to a low of only 67.4 per cent in the second half of 2010 (these
additional results are available on request). Both the regression and information share results
point to the ICE Brent crude oil futures being the more important security for price discovery.
Table 3.12
Information Shares for European Crude Oil
Hasbrouck‘s (1995) information shares (i
IS ) are calculated from log price levels sampled at 10-second
intervals between 2 January 2008 and 30 December 2011 (3,680,207 observations). Formally:
ΨΨΩ
2
ii
MIS
(3.8)
The elements of the ( K1 ) row vector Ψ are the sums of each security‘s coefficients in the moving
average impact matrix (i
). The lower triangular matrix from a Cholesky factorisation ( M ) of the
VECM‘s contemporaneous error term variance-covariance matrix ( Ω ) is used to construct the variance
contribution of each security, ([ M ]i )2, to the variance of the common efficient price ( ΨΨΩ ). The
VECM specification displayed in equation (3.3) contains 15 lags in theit
y vector in order to deal with
any minor autocorrelation. This lag length was selected using Schwarz‘s Bayesian Information Criterion.
The order of the series in the VECM log price vector (t
y ) is cycled such that the VECM is run with each
series taking a turn at being the first series in the log price vector. The average information shares
calculated across the two cycles are displayed along with the ordinal ranking of these averages.
Series Ordered First in Cycle Brent Futures Dated BFOE Proxy
Brent Futures 0.874 0.126
Dated BFOE Proxy 0.285 0.715
Average 0.579 0.421
Rank 1 2
88
3.4.2 Price Discovery in Brent and WTI Futures
There has been a long running debate in the energy market literature concerning whether
crude oil prices are determined regionally or whether the various regional physical benchmark
and futures prices are determined in a unified global market. Although the work on correlations
by Weiner (1991) finds support for regionalisation, most subsequent studies find that, within the
approximate limits of transportation costs, arbitrage opportunities tend to prevent large
divergences in the various crude oil prices and conclude that the global crude oil market is
unified78
. Within this unified market, most studies find that WTI prices are the more important
source of price discovery. For example, Hammoudeh et al. (2008) estimate a threshold
cointegration model on daily closing prices for four global crude oil benchmarks and find that
WTI prices lead Brent prices between 1990 and 2006. However, Kao and Wan (2012) calculate
rolling information shares from daily prices and find Brent futures have displayed greater price
discovery than WTI futures since around 2004. We re-examine this question in light of the
growing dislocation of WTI prices from the other global crude oil benchmarks. In addition, we
examine this question using high frequency intraday futures price data rather than the daily,
weekly or monthly data employed in most previous studies. We also note that low frequency
studies may bias results by comparing closing prices from the WTI and Brent markets at
different times of the day, potentially inducing spurious leading or lagging relationships.
Similar to the BFOE grades, WTI is a light, sweet crude oil, though of a slightly higher
quality. It flows through pipelines from wells in Texas, New Mexico, Kansas and Oklahoma to
the storage facilities at Cushing79
. WTI‘s importance as a global benchmark stems from the fact
that it is the main grade physically deliverable into the CME‘s light sweet crude oil futures
contract (in fact these contracts are typically just called WTI futures). Over time this has
resulted in many long-term crude oil import contracts being pegged to WTI, though dislocations
78
Though utilising different methodologies, this is generally the conclusion reached in Adelman (1984,
1992), Gülen (1997, 1999), Bachmeier and Griffin (2006), Bentzen (2007) and Kaufmann and
Ullman (2009).
79 Unlike BFOE which is waterborne, physical WTI is a pipeline crude such that trades can take place for
smaller parcels than the partial or full cargoes in the physical BFOE market, with typical trade sizes of
around 30,000 barrels (Fattouh, 2011). This means there are fewer barriers to entry in the physical WTI
market and greater diversity in market participation compared with the physical BFOE market.
89
between WTI and other benchmark crude prices has progressively eroded its use for this
purpose.
WTI futures are frequently affected by expectations of storage and pipeline capacity
constraints at Cushing80
. In the past this has been the result of logistical difficulties getting
enough crude oil to the pricing point at Cushing and has served to reinforce WTI‘s traditional
premium over BFOE. However, the recent expansion in shale oil production in North Dakota‘s
Bakken fields and oil sands production in Canada, both of which flow south to Cushing, has
resulted in the opposite problem. Although WTI should trade at a premium to BFOE due to its
higher quality, the lack of outward flowing pipeline capacity from Cushing to the refineries on
the US Gulf Coast has seen it trade at a substantial discount (see Chart 3.1)81
. In the face of
pipeline constraints, WTI-Brent differentials of around US$8-12 per barrel reflect the
approximate cost of getting crude oil out of Cushing and to the US Gulf Coast by rail (see Sen,
2012). On the basis of this dislocation we would expect to find that any degree of cointegration
between WTI and Brent futures becomes weaker toward the end of our sample period82
.
80
We look at WTI futures and not WTI spot prices because the majority of previous studies show that,
within the complex of WTI securities, futures are more often found to lead spot prices, though there is
some contention depending on the approach taken. Schwarz and Szakmary (1994) use an error-correction
model and the Garbade and Silber (1983) model and find that WTI futures lead spot prices from 1984 to
1991. Moosa (2002) estimates the Garbade and Silber (1983) model and finds that, relative to spot prices,
WTI futures account for approximately 60 per cent of price discovery from 1985 to 1996. For the same
period, Silvapulle and Moosa (1999) find similar results for linear causality tests, but find support for bi-
directional causality when non-linear tests are run. Bekiros and Diks (2008) also found bi-directional,
non-linear relationships in their 1991 to 2007 sample. Despite this evidence, it should be noted that this
debate has become less meaningful over time. In recent years trading in the Platts WTI Cash Window has
virtually ceased and what are referred to as WTI ‗spot‘ prices are either prices posted by major oil
companies such as ConocoPhillips (referred to as posting plus or P-Plus, which are wellhead prices plus
delivery costs into Cushing) or are based on differentials to the New York Mercantile Exchange Calendar
Monthly Average market which are largely driven by WTI futures anyway (see Fattouh, 2011).
81 The discount narrowed somewhat after the announcement that the Seaway pipeline from the US Gulf
Coast to Cushing would have its direction reversed to help alleviate the problem, however, reversed flows
may not reach full capacity until 2013. The prospect that these pressures will be alleviated by the
construction of the Keystone XL pipeline, effectively linking Canada to the US Gulf Coast, remains
uncertain as environmental concerns continue to delay the project‘s approval. Montepeque (2012)
highlights how factors related to investment vehicles add to the persistence of this discount. More
specifically, the expectation that the storage facilities at Cushing may reach capacity creates a steep
contango in the futures curve. This makes rolling forward expensive and discourages investment in crude
oil via WTI futures. Montepeque (2012) notes that both the Goldman Sachs Commodity Index and the
UBS/Dow Jones Index have recently decreased their WTI futures weights in favour of increased Brent
futures weightings.
82 Both the ICE Brent futures and CME WTI futures contracts trade around 23 hours per day on
weekdays. However, in order to avoid any effects of non-concurrent shifts on and off daylight savings in
the different geographic trading locations of the contracts, we have chosen a narrower 14 hour intraday
window. We examine the contracts at 10-second intervals between 7:00am and 9:00pm GMT daily,
90
Chart 3.1
Dislocation between Brent and WTI Futures Prices
Monthly expiry ICE Brent crude oil futures and monthly expiry CME WTI futures prices sampled on a daily
basis between 2 January 2008 and 30 December 2011. Differential is the WTI futures price minus the Brent
futures price.
which captures 98 per cent of Brent futures trades and 97 per cent of the WTI futures trades (see the
Appendix, Table C1). We also conduct the analysis for a 1-minute sample but, with little difference in the
results, in what follows we focus on the more granular 10-second data.
0
20
40
60
80
100
120
140
160
0
20
40
60
80
100
120
140
160
Brent Futures WTI Futures
US$/bbl US$/bbl
2008 2009 2010 2011
-30
-25
-20
-15
-10
-5
0
5
10US$/bbl
2008 2009 2010 2011
Panel A: Brent and WTI Futures Prices
Panel B: WTI – Brent Futures Price Differential
91
Cointegration tests on the two securities are very sensitive to the lag specification in the
VECM as shown by the contrasting results for 1-lag compared with 15-lags in Table 3.13.
Although the autocorrelation in the sample is not economically important and is likely driven by
bid-ask bounce, as indicated by tiny negative autocorrelation coefficients, in such a large sample
(5,151,901 observations) it is nonetheless statistically significant, which justifies putting greater
reliance on the VECM specification with a larger number of lags (autocorrelation coefficients
and test statistics are presented in the Appendix, Table C2). On the basis of the 15-lag VECM
specification, the Johansen (1995) test indicates that Brent and WTI futures were not
cointegrated when sampled at 10-second intervals between 2008 and 201183
. This result is
intuitively surprising, given the obvious, close relationship over most of the sample period
depicted in Chart 3.1. As such, we also examine cointegration over 6-month windows in the
sample period, with results displayed in Table 3.14.
The cointegration test results run over 6-month windows remain sensitive to the lag
specification in the VECM. In the more appropriate 15-lag specification, Brent and WTI futures
are shown to have been cointegrated only in the first half of 2008 and the second half of 2011.
Interestingly, even in the 1-lag specification, cointegration begins to wane towards the end of
the sample as evidenced by both the increasing trace statistics for one or fewer cointegrating
relationships ( 1r ) and the lower trace statistics for zero cointegrating relationships ( 0r ).
This is in line with the growing dislocation of WTI from the other crude benchmarks in 2011
which is observable in Chart 3.1. Although the evidence that the securities are cointegrated over
the whole 2008-2011 sample—and even over the smaller 6-month sub-samples—is weak at
best, we nonetheless examine price leadership using the regression and information share
approaches. We first examine whether the increasing prominence of Brent futures has led to a
reversal in the price leadership of WTI over Brent futures observed in Brunetti and Gilbert
83
These results are the same for the 1-minute sample. Interestingly, tests on daily data fail to find a
cointegrating relationship between the two securities between 2008 and 2011, despite the close
relationship that appears in Chart 3.1. However, daily tests run from 2008 to 2010 indicate a cointegrating
relationship at the 1 per cent level of statistical significance (though this is not so for the same period
using the higher frequency data).
92
(2000), Lin and Tamvakis (2001), Hammoudeh et al. (2008) and Kaufmann and Ullman (2009)
using the regression approach.
The results of the two regressions presented in Table 3.15 are quite similar. The
contemporaneous return coefficients are large, positive and statistically significant even at the
fine 10-second sampling interval, which indicates a surprisingly high degree of short-term co-
movement considering the poor long-term cointegration results. Lagged return observations of
both securities have some explanatory power in determining the contemporaneous returns of the
other, indicating a degree of bi-directional price discovery occurring across the securities. The
differences in the number of statistically significant lagged coefficients and their size are minor,
making it hard to distinguish between the two regressions. However, when the lagged WTI
futures are the independent variables, the regression has a higher robust F-statistic, which entails
some marginal evidence that the WTI futures display slightly more short-term price leadership.
Table 3.13
Cointegration Test: Brent and WTI Futures
Table 3.13 displays Johansen (1995) cointegration tests of the log prices of monthly expiry ICE Brent
and CME WTI crude oil futures sampled at 10-second intervals from 7:00am and 9:00pm GMT between
2 January 2008 and 30 December 2011 (5,151,901 observations). Tests are run using a multivariate
VECM estimated by maximum likelihood with alternately 15 lags ( 15n ) and 1 lag ( 1n ) as
explanatory variables to determine the number of cointegrating relations ( r ) between the ( K ) variables:
tit
n
i
itt εyΓyβαy
1
1
1
(3.3)
Dependent variables ( ty ) are a 1K vector of differenced log price levels;
ity are lagged dependent
variables; α and β are rK parameter matrices in which the number of cointegrating equations is less
than the number of I(1) variables ( Kr ); 11
,,p
ΓΓ are KK matrices of parameters; and, tε is a
1K vector of normally distributed and serially uncorrelated error terms. The null hypothesis for the
trace statistic is that there are no more than r cointegrating relations (i.e. the eigenvalues Kr ,,
1
are
zero). * and ** denote the rank at which the null hypothesis cannot be rejected at the 5 and 1 per cent
levels of statistical significance, respectively (critical values are from the tables in Johansen, 1995).
Maximum Rank Eigenvalue Trace Statistic Eigenvalue Trace Statistic 5% 1%
r ≤ 0 9.32** 24.06 12.21 16.16
r ≤ 1 0.00000 2.34 0.00000 1.60** 4.14 7.02
r ≤ 2 0.00000 0.00000
Panel B: 1 Lag Critcal ValuesPanel A: 15 Lags
93
Table 3.14
Cointegration Test: Brent and WTI Futures, 6-month Sub-samples
Table 3.14 displays Johansen (1995) cointegration tests of log prices for monthly expiry ICE Brent and CME WTI crude oil futures sampled at 10-second intervals from 7:00am
to 9:00pm GMT in 6-month windows between 2 January 2008 and 30 December 2011. Tests are run using a multivariate VECM estimated by maximum likelihood. In Panel A,
15 lags ( 15n ) are used as explanatory variables to determine the number of cointegrating relations ( r ) between the ( K ) variables, while in Panel B only 1 lag ( 1n ) is
used:
tit
n
i
itt εyΓyβαy
1
1
1
(3.3)
Dependent variables (t
y ) are a 1K vector of differenced log price levels; it
y are lagged dependent variables; α and β are rK parameter matrices in which the
number of cointegrating equations is less than the number of I(1) variables ( Kr ); 11
,,p
ΓΓ are KK matrices of parameters; and,
tε is a 1K vector of normally
distributed and serially uncorrelated error terms. The null hypothesis for the trace statistic is that there are no more than r cointegrating relations (i.e. the eigenvalues
Kr ,,
1
are zero). * and ** denote the rank at which the null hypothesis cannot be rejected at the 5 and 1 per cent levels of statistical significance, respectively (critical
values are from Johansen, 1995, and are displayed in Table 3.13 above).
Panel A:
Maximum
Rank Eigenvalue
Trace
Statistic Eigenvalue
Trace
Statistic Eigenvalue
Trace
Statistic Eigenvalue
Trace
Statistic Eigenvalue
Trace
Statistic Eigenvalue
Trace
Statistic Eigenvalue
Trace
Statistic Eigenvalue
Trace
Statistic
r ≤ 0 37.71 4.30** 11.90** 6.92** 21.07 6.53** 8.76** 15.82
r ≤ 1 0.00006 0.03** 0.00001 0.92 0.00002 1.77 0.00001 2.60 0.00002 9.66 0.00001 0.77 0.00001 2.67 0.00002 1.62**
r ≤ 2 0.00000 0.00000 0.00000 0.00000 0.00002 0.00000 0.00000 0.00000
Panel B:
r ≤ 0 226.18 19.33 24.97 23.96 27.49 15.63 9.65** 23.27
r ≤ 1 0.00036 0.02** 0.00003 0.01** 0.00004 1.17** 0.00003 3.83** 0.00004 4.36** 0.00002 1.05** 0.00001 3.22 0.00003 2.84**
r ≤ 2 0.00000 0.00000 0.00000 0.00001 0.00001 0.00000 0.00001 0.00000
15-Lag Specification
2011: H1 2011: H22010: H2
1-Lag Specification
2008: H1 2008: H2 2009: H1 2009: H2 2010: H1
94
Table 3.15
Price Discovery Regressions: Brent versus WTI Futures
Table 3.15 presents the results of fitting model (3.1) using continuously compounded returns observed at 10-second intervals between 7:00am and 9:00pm GMT from
2 January 2008 to 30 December 2011 (5,151,891 observations). The independent variables are alternately contemporaneous and lagged returns of the ICE Brent and CME WTI
futures (series B) as well as an error correction term, with the response variable being the other series of interest (series A). Formally:
ttAzAktB
k
ktA zRR
1,,,
0
10
,
(3.1)
The error correction terms are given by: 1,tA
z ln(1, tA
p ) – ln(1, tB
p ), the one-period lag of the difference in log prices between the two series. Square brackets [ ] below
coefficients contain t-statistics, while round brackets ( ) below F-statistics contain p-values. * and ** denote significance at the 5 and 1 per cent levels.
α βt βt-1 βt-2 βt-3 βt-4 βt-5 βt-6 βt-7 βt-8 βt-9 βt-10 βZ Adj. R-sqr F-stat
0.0000 0.6769** 0.0611** 0.0147** 0.0100** 0.0046** 0.0036** 0.0017** 0.0008 0.0019** -0.0002 0.0005 -0.0000** 0.445 4,115.70**
[1.76] [153.19] [58.94] [20.28] [11.96] [6.80] [5.75] [2.70] [1.31] [3.07] [-0.29] [0.76] [-4.00] (0.000)
-0.0000** 0.6575** 0.1086** 0.0186** 0.0051** 0.0025** 0.0015* -0.0002 0.0007 0.0006 0.0004 0.0012* -0.0000** 0.454 6,464.82**
[-2.56] [161.69] [74.98] [28.43] [8.20] [3.93] [2.42] [-0.30] [1.14] [0.91] [0.74] [2.14] [-5.19] (0.000)Dep
en
den
t V
ari
ab
le
WTI
Futures
Brent
Futures
Independent Variable: WTI Futures
Independent Variable: Brent Futures
95
Although over the 2008-2011 period WTI futures typically attracted a greater daily trade
volume than Brent futures, this gap has been narrowing (see Montepeque, 2012, and Blas,
2012b). For the front month contracts used in this study, Brent futures trade volume was greater
than WTI futures volume on 3.9 per cent of days in 2008, 5.5 per cent of days in 2009, 6.3 per
cent of days in 2010 and had risen to being greater on 15.4 per cent of days by 2011. Given the
recent dislocation of the WTI market and the narrowing of the trade volume gap, we would
expect to find greater price leadership attributed to Brent futures, if at all, towards the end of the
sample period. To examine this we run the regressions for 6-month windows over the sample.
The results in Table 3.16 largely conform to those in Table 3.15: the robust F-statistics are
predominantly larger when WTI futures are the independent variables; but, the evidence does
not overwhelmingly distinguish between the securities. Despite this, the 6-month sub-sample
regressions do reveal two points of interest. Firstly, the declining degree of cointegration is
evidenced by the progressively smaller t-statistics on the error correction terms, which are not
significant in the first half of 2011. Secondly, though the F-statistics have decreased markedly
in magnitude over the sample as the relationship between the Brent and WTI futures weakens,
there are two periods—the second half of 2009 and the first half of 2011—in which a higher F-
statistic in regressions with Brent futures as the independent variables provides some evidence
that there are periods in which Brent futures are leading the short-run return dynamics.
96
Table 3.16
Price Discovery Regressions: Brent versus WTI Futures, 6-month Sub-samples
Table 3.16 presents the results of fitting model (3.1) using continuously compounded returns observed at 10-second intervals between 7:00am and 9:00pm GMT in 6-month
windows from 2 January 2008 to 30 December 2011. In Panel A, the independent variables are contemporaneous and lagged returns of the ICE Brent futures (series B) as well
as an error correction term, with the response variable being the CME WTI futures (series A). In Panel B, the CME WTI futures are the independent variables. Formally:
ttAzAktB
k
ktA zRR
1,,,
0
10
,
(3.1)
The error correction terms are given by: 1,tA
z ln(1, tA
p ) – ln(1, tB
p ), the one-period lag of the difference in log prices between the two series. Square brackets [ ] below
coefficients contain t-statistics, while round brackets ( ) below F-statistics contain p-values. * and ** denote significance at the 5 and 1 per cent levels, respectively.
α βt βt-1 βt-2 βt-3 βt-4 βt-5 βt-6 βt-7 βt-8 βt-9 βt-10 βZ Adj. R-sqr F-stat
Panel A:
2008 -0.0000** 0.6345** 0.0628** 0.0139** 0.0065** 0.0035* 0.0001 0.0032* 0.0015 -0.0005 0.0004 -0.0015 -0.0003** 0.404 726.07**
H1 [-14.29] [87.28] [19.17] [7.44] [3.54] [1.98] [0.05] [2.01] [0.86] [-0.27] [0.28] [-1.00] [-16.23] (0.000)
2008 -0.0000** 0.6580** 0.0590** 0.0125** 0.0100** 0.0029* 0.0034* 0.0008 0.0015 0.0022 -0.0012 0.0020 -0.0000** 0.468 919.98**
H2 [-5.94] [61.93] [27.05] [7.68] [6.19] [2.02] [2.55] [0.59] [1.15] [1.61] [-0.86] [1.51] [-5.86] (0.000)
2009 0.0000** 0.6701** 0.0694** 0.0220** 0.0156** 0.0090** 0.0058** 0.0022 0.0005 0.0038** 0.0003 -0.0008 -0.0000** 0.400 970.37**
H1 [3.88] [73.77] [28.66] [13.59] [6.97] [5.86] [3.99] [1.59] [0.38] [2.66] [0.21] [-0.57] [-3.59] (0.000)
2009 0.0000* 0.7011** 0.0586** 0.0172** 0.0076** 0.0045** 0.0048** 0.0032** 0.0008 0.0010 0.0007 0.0017 -0.0000** 0.509 2,040.42**
H2 [2.45] [129.96] [40.13] [14.25] [6.46] [4.18] [4.62] [3.16] [0.75] [0.93] [0.72] [1.66] [-6.07] (0.000)
2010 0.0000 0.7240** 0.0641** 0.0141** 0.0072** 0.0036** 0.0037** 0.0014 0.0004 0.0012 0.0009 -0.0005 -0.0000** 0.523 903.81**
H1 [-1.63] [77.72] [32.27] [10.28] [5.69] [2.93] [3.05] [1.24] [0.33] [1.08] [0.79] [-0.47] [-4.89] (0.000)
2010 0.0000** 0.6873** 0.0635** 0.0122** 0.0066** 0.0050** 0.0018 0.0012 0.0017 -0.0011 0.0011 0.0004 -0.0000** 0.480 1,844.02**
H2 [3.07] [115.44] [38.02] [9.79] [5.73] [4.48] [1.60] [1.04] [1.48] [-1.07] [0.98] [0.42] [-4.27] (0.000)
2011 0.0000 0.6691** 0.0440** 0.0034* 0.0042* 0.0008 -0.0011 -0.0011 -0.0027 0.0029* 0.0002 -0.0011 0.0000 0.415 471.81**
H1 [1.42] [70.83] [22.07] [2.22] [2.43] [0.49] [-0.61] [-0.79] [-1.92] [2.16] [0.16] [-0.80] [-1.82] (0.000)
2011 0.0000** 0.7553** 0.0558** 0.0113** 0.0074** 0.0029 0.0059** 0.0036* 0.0014 0.0000 -0.0005 0.0011 -0.0000** 0.427 270.56**
H2 [2.94] [52.40] [19.87] [6.25] [4.69] [1.94] [3.72] [2.45] [0.94] [0.03] [-0.32] [0.82] [-2.79] (0.000)
Dep
end
ent
Va
ria
ble
: C
ME
WT
I F
utu
res
Independent Variable: ICE Brent Futures
97
Table 3.16
Price Discovery Regressions: Brent versus WTI Futures, 6-month Sub-samples (Continued)
α βt βt-1 βt-2 βt-3 βt-4 βt-5 βt-6 βt-7 βt-8 βt-9 βt-10 βZ Adj. R-sqr F-stat
Panel B:
2008 0.0000** 0.6350** 0.1413** 0.0240** 0.0113** 0.0066** 0.0032* 0.0044** 0.0040** 0.0025 0.0028 0.0009 -0.0003** 0.420 3,607.64**
H1 [14.62] [130.47] [57.32] [13.13] [6.76] [4.09] [2.10] [2.90] [2.92] [1.66] [1.88] [0.60] [-15.54] (0.000)
2008 0.0000** 0.7116** 0.1242** 0.0274** 0.0103** 0.0054** 0.0013 -0.0013 0.0029* 0.0014 0.0017 0.0025* -0.0000** 0.479 2,565.88**
H2 [5.55] [83.88] [30.42] [17.57] [6.97] [3.65] [0.93] [-0.99] [2.17] [0.87] [1.33] [1.99] [-6.81] (0.000)
2009 -0.0000** 0.5957** 0.1002** 0.0129** 0.0002 -0.0007 0.0030* -0.0003 -0.0023 0.0005 -0.0019 0.0021 -0.0000** 0.406 1,065.93**
H1 [-2.70] [62.42] [37.65] [9.38] [0.14] [-0.48] [2.13] [-0.23] [-1.86] [0.41] [-1.54] [1.73] [-2.84] (0.000)
2009 -0.0000* 0.7294** 0.1052** 0.0150** 0.0014 0.0014 -0.0007 -0.0016 0.0000 -0.0011 -0.0007 -0.0007 -0.0000** 0.517 894.20**
H2 [-2.26] [99.29] [42.52] [12.57] [1.22] [1.27] [-0.64] [-1.58] [0.02] [-1.02] [-0.63] [-0.66] [-6.16] (0.000)
2010 0.0000 0.7223** 0.0942** 0.0143** 0.0034** -0.0011 0.0000 0.0005 0.0004 -0.0014 0.0013 0.0016 -0.0000** 0.528 3,343.98**
H1 [1.01] [126.62] [31.81] [9.55] [2.76] [-0.86] [-0.01] [0.46] [0.34] [-1.27] [1.20] [1.47] [-3.97] (0.000)
2010 -0.0000** 0.7003** 0.1162** 0.0158** 0.0046** 0.0020 0.0005 0.0007 -0.0028** 0.0024* -0.0004 -0.0014 -0.0000** 0.489 1,848.11**
H2 [-3.03] [138.61] [51.32] [12.77] [3.93] [1.93] [0.47] [0.70] [-2.70] [2.40] [-0.34] [-1.36] [-5.03] (0.000)
2011 0.0000 0.6220** 0.0954** 0.0152** 0.0054** 0.0035** 0.0034 0.0014 0.0020 -0.0014 0.0004 -0.0014 0.0000 0.424 331.75**
H1 [-0.75] [53.29] [28.81] [10.08] [3.89] [2.59] [1.95] [1.10] [1.44] [-1.08] [0.28] [-1.07] [-1.29] (0.000)
2011 -0.0000** 0.5667** 0.0789** 0.0115** 0.0025 0.0020 -0.0018 0.0001 0.0011 -0.0014 0.0016 -0.0021* -0.0000** 0.433 830.64**
H2 [-2.77] [72.08] [29.47] [9.40] [1.83] [1.85] [-1.69] [0.06] [1.05] [-1.27] [1.49] [-1.96] [-2.77] (0.000)
Dep
en
den
t V
ari
ab
le:
ICE
Bre
nt
Fu
ture
s
Independent Variable: WTI Futures
98
The relative contribution to long-run equilibrium between the securities is given by the
information shares in Table 3.17. The results indicate that WTI futures accounted for
approximately 58 per cent of price discovery between 2008 and 2011. This is despite anecdotal
reports of Brent‘s increasing importance as a global benchmark for long-term and spot oil
contracts84
. These results are consistent with the location of the majority of trade activity,
though we note that this is not necessarily a determinant of price discovery. Similar to the
cointegration tests, we examine the information shares over 6-month sub-samples to see
whether there is any discernable trend in long-run price discovery.
Table 3.17
Information Shares: Brent versus WTI Futures
Hasbrouck‘s (1995) information shares (i
IS ) calculated from log price levels of ICE Brent and CME
WTI crude oil futures sampled at 10-second intervals between 7:00am and 9:00pm GMT from
2 January 2008 to 30 December 2011 as per:
ΨΨΩ
2
ii
MIS
(3.8)
The elements of the ( K1 ) row vector Ψ are the sums of each security‘s coefficients in the moving
average impact matrix (i
). The lower triangular matrix from a Cholesky factorisation ( M ) of the
VECM‘s contemporaneous error term variance-covariance matrix ( Ω ) is used to construct the variance
contribution of each security to the variance of the common efficient price ( ΨΨΩ ). The VECM
specification from equation (3.3) contains 15 lags in theit
y vector in order to deal with any minor
autocorrelation. This lag length was selected using Schwarz‘s Bayesian Information Criterion. The order
of the series in the VECM log price vector (t
y ) is cycled such that the VECM is run with each series
taking a turn at being the first series in the log price vector. The average information shares calculated
across the two cycles are displayed along with the ordinal ranking of these averages.
84
Though difficult to quantify, various market commentators purport that Brent prices are the benchmark
for between 50 and 70 per cent of international oil transactions (see, for example, Fattouh, 2011, and
Barret, 2012).
Series Ordered First in Cycle Brent Futures WTI Futures
Brent Futures 0.769 0.231
WTI Futures 0.076 0.924
Average 0.423 0.577
Rank 2 1
99
Table 3.18
Information Shares: Brent versus WTI Futures, 6-month Sub-samples
Hasbrouck‘s (1995) information shares (i
IS ) calculated from log price levels of ICE Brent and CME WTI crude oil futures sampled at 10-second intervals between 7:00am and
9:00pm GMT over 6-month windows from 2 January 2008 to 30 December 2011 as per:
ΨΨΩ
2
ii
MIS
(3.8)
The elements of the ( K1 ) row vector Ψ are the sums of each security‘s coefficients in the moving average impact matrix (i
). The lower triangular matrix from a Cholesky
factorisation ( M ) of the VECM‘s contemporaneous error term variance-covariance matrix ( Ω ) is used to construct the variance contribution of each security to the variance of
the common efficient price ( ΨΨΩ ). The VECM specification from equation (3.3) contains 15 lags in theit
y vector in order to deal with any minor autocorrelation. This lag
length was selected using Schwarz‘s Bayesian Information Criterion. The order of the series in the VECM log price vector ( ty ) is cycled such that the VECM is run with each
series taking a turn at being the first series in the log price vector. The average information shares calculated across the two cycles are displayed along with the ordinal ranking
of these averages.
Brent
Futures
WTI
Futures
Brent
Futures
WTI
Futures
Brent
Futures
WTI
Futures
Brent
Futures
WTI
Futures
Brent
Futures
WTI
Futures
Brent
Futures
WTI
Futures
Brent
Futures
WTI
Futures
Brent
Futures
WTI
Futures
Brent Futures 0.595 0.405 0.878 0.122 0.226 0.774 0.602 0.398 0.950 0.050 1.000 0.000 0.255 0.745 0.215 0.785
WTI Futures 0.020 0.980 0.170 0.830 0.055 0.945 0.005 0.995 0.653 0.347 0.495 0.505 0.038 0.962 0.074 0.926
Average 0.307 0.693 0.524 0.476 0.141 0.859 0.304 0.696 0.802 0.198 0.747 0.253 0.147 0.853 0.144 0.856
Rank 2 1 1 2 2 1 2 1 1 2 1 2 2 1 2 1
Series Ordered
First in Cycle
2008: H1 2008: H2 2009: H1 2009: H2 2011: H1 2011: H22010: H1 2010: H2
100
Consistent with the results for the whole sample period, the results in Table 3.18 show that
for the majority of the 6-month sub-samples WTI futures are dominant in long-run price
discovery. However, of the eight semi-annual periods, the second half of 2008 and the whole of
2010 show greater price discovery in the Brent futures, though we note the results for the first
half of 2008 are fairly close (at 52.4 per cent). While there have been significant periods in
which Brent futures were the more important source of price discovery, it is interesting that
these sub-periods do not line-up with the period of greatest dislocation between the securities in
201185
. We note that the interpretation of these results should be tempered by the fact that the
cointegration of the securities, which underpins the information shares methodology, is largely
lacking over the 2008-2011 period as indicated by the Johansen (1995) test results in Tables
3.13 and 3.1486
.
3.5 Conclusion
Price discovery in the complex of financial and physical layers commonly found in energy
markets is important because these layers determine key benchmarks used as reference points to
value vast volumes of commodity transactions through long-term supply contracts. Although
examining price discovery in the European markets for coal, natural gas and crude oil is made
difficult by the lack of transparency and liquidity in over-the-counter transactions, we
nonetheless attempt to establish which prices better reflect information in these markets on the
basis of both short-term return dynamics and long-run price equilibrium.
85
Results from calculating information shares using data sampled at 1-minute intervals are very similar to
those presented here for 10-second intervals. In addition, similar results are obtained using Gonzalo and
Granger‘s (1995) common factor weights, which is an alternative price discovery methodology that also
focuses on the relative contributions to long-run equilibrium. For the sample as a whole, the common
factor weights attribute 51.5 per cent of the long run price discovery to the WTI futures. Similar to the
information shares, the common factor weights for 6-month sub-samples only support Brent futures as the
dominant source of price discovery in the first half of 2008 and for the whole of 2010 (these results are
available on request).
86 Information shares calculated using daily data from 2008 to 2010, a period for which there is
statistically significant cointegration (see footnote 83), also shows WTI futures display greater price
discovery, though their dominance is somewhat marginal, with an information share of 54.3 per cent
(these results are available on request). We note that this analysis is over a small sample size, however,
intraday timing issues are not a concern as the daily sample is constructed from contemporaneous
observations at 4:00pm GMT each day; a time at which both securities are near their highest intraday
activity (see the Appendix, Table C1).
101
The coal market remains somewhat obscured by the lack of liquidity and transparency in
both physical and financial transactions. However, at the margin, innovations in the prices of the
Intercontinental Exchange monthly coal futures appear to have the greatest impact on the
equilibrium price in the coal market, while the more traded Intercontinental Exchange quarterly
futures and European Energy Exchange annual futures display greater short-run return
leadership.
Monthly expiry UK natural gas futures traded on the Intercontinental Exchange display
greater price discovery than physical trading at the major hubs in North-West Europe,
particularly in their contribution to long-run equilibrium. There is evidence that short-run
interactions are stronger at similar points on the forward curve than interactions between
securities specific to geographic locations, which is likely related to the common impact of
weather events on near-term gas demand and constraints that impede the instantaneous transfer
of natural gas between hubs. In this way, inelasticity of demand means that short-dated natural
gas prices behave like electricity prices and can quickly become volatile. In addition,
cointegration tests indicate that natural gas prices remain weakly linked to the crude oil market.
To the limited extent that it is possible to distinguish the financial and physical layers of the
Brent complex of securities, we find evidence that the Intercontinental Exchange Brent crude oil
futures contract leads the price discovery process. However, the regression results do show
some evidence of bi-directionality, with trading in the physical layers—specifically exchange-
for-physicals in this study—at times briefly leading the futures market.
Adding to the debate on the regional versus global determination of crude oil prices and in
light of recent structural issues with WTI pricing, we find only weak evidence that Brent and
WTI futures remain cointegrated, with their relationship deteriorating further towards the end of
our 2008-2011 sample. The regression results point to WTI futures leading short-run return
dynamics, but this evidence is very marginal and there are two 6-month sub periods in which
the evidence favours Brent futures leading the short-run dynamics. Similarly, the information
shares point to WTI futures making the greater contribution to long-run equilibrium, but there
are several sub-periods for which Brent futures are relatively more important.
102
3.6 Appendix
Tests for autocorrelation, stationarity and cointegration for each category of energy
commodity—coal, natural gas and crude oil—are presented in the following sections.
A. Coal
Table A1
Autocorrelation Coefficients and Test Statistics for Coal Securities
Weighted autocorrelation coefficient estimates ( jD ) and2
test statistics calculated as per the unified
approach in Richardson and Smith (1994) for Box and Pierce (1970) Q-statistics, Fama and French
(1988) beta statistics and Lo and MacKinlay (1988) variance ratios. * and ** denote significance at the 5
and 1 per cent levels against 2
critical values with 5 degrees of freedom. Autocorrelation coefficients
and statistics calculated from continuously compounded returns for coal securities sampled daily from
2 January 2008 to 30 December 2011 (961 observations).
Table A2
Augmented Dickey-Fuller Stationarity Tests for Coal Securities
Table A2 displays Augmented Dickey-Fuller (1979) test statistics for log price levels and continuously
compounded returns for coal at daily intervals between 2 January 2008 and 30 December 2011 (956
observations). The unit root tests are run with a constant ( ) and 5 lags of differenced dependent
variables as explanatory variables ( 5k ) as per:
tjt
k
j
jtt yyy
1
1
Dependent variables (t
y ) are alternately differenced log price levels and differenced returns. Test
statistics, ˆˆ
tZ , are for 0:
0H , where
is the standard error of . * and ** denote
significance at the 5 and 1 per cent levels against critical values from Fuller (1996) of -2.86 and -3.43,
respectively.
ρ1 ρ2 ρ3 ρ4 ρ5 Q-statistic Beta Statistic Variance Ratio
ICE Monthly Futures 0.0010 0.0449 -0.0965 0.0133 0.0674 0.0160** -0.0349 -0.0164
ICE Quarterly Futures 0.1969 0.0406 0.0609 0.0295 0.0147 0.0452** 0.1782* 0.4243**
EEX Monthly Futures -0.0650 0.0448 -0.0426 0.0589 0.0180 0.0118* 0.0109 -0.0608
EEX Annual Futures 0.0857 0.0150 0.0100 0.0548 0.0790 0.0169** 0.1114 0.1850
Autocorrelation Coefficients Test Statistics
ICE Monthly ICE Quarterly EEX Monthly EEX Annual
Levels -1.17 -1.39 -1.08 -1.40
Returns -13.13** -12.11** -12.20** -12.05**
103
Table A3
Johansen Cointegration Test for Coal Securities
Table A3 displays Johansen (1995) cointegration tests of the daily log price levels for the four coal
securities ( 4K ) between 2 January 2008 and 30 December 2011 (960 observations). Tests are run
using a multivariate VECM estimated by maximum likelihood with 2 lags ( 2n ) as explanatory
variables to determine the number of cointegrating relations ( r ) between the ( K ) variables:
tit
n
i
itt εyΓyβαy
1
1
1
Dependent variables (t
y ) are a 1K vector of differenced log price levels; it
y are lagged dependent
variables; α and β are rK parameter matrices in which the number of cointegrating equations is less
than the number of I(1) variables ( Kr ); 11
,,p
ΓΓ are KK matrices of parameters; and, tε is a
1K vector of normally distributed and serially uncorrelated error terms. The null hypothesis for the
trace statistic is that there are no more than r cointegrating relations (i.e. the eigenvalues Kr ,,
1
are
zero). * and ** denote the rank at which the null hypothesis cannot be rejected at the 5 and 1 per cent
levels of statistical significance, respectively (critical values are from Johansen, 1995).
B. Natural Gas
Table B1
Autocorrelation Coefficients and Test Statistics for Natural Gas Securities
Weighted autocorrelation coefficient estimates ( jD ) and2
test statistics calculated as per the unified
approach in Richardson and Smith (1994) for Box and Pierce (1970) Q-statistics, Fama and French
(1988) beta statistics and Lo and MacKinlay (1988) variance ratios. * and ** denote significance at the 5
and 1 per cent levels against 2
critical values with 5 degrees of freedom. Autocorrelation coefficients
and statistics calculated from continuously compounded returns for the natural gas securities sampled at
10-minute intervals from 2 January 2008 to 30 December 2011 (47,088 observations).
Maximum Rank Eigenvalue Trace Statistic 5% Critical Value 1% Critical Value
r ≤ 0 69.56 39.71 46.00
r ≤ 1 0.0327 37.67 24.08 29.19
r ≤ 2 0.0240 14.38** 12.21 16.16
r ≤ 3 0.0112 3.54* 4.14 7.02
r ≤ 4 0.0037
ρ1 ρ2 ρ3 ρ4 ρ5 Q-statistic Beta Statistic Variance Ratio
NBP Month-Ahead -0.0726 -0.0015 -0.0084 -0.0016 -0.0003 0.0053** -0.0348** -0.1253**
NBP Day-Ahead -0.0390 -0.0086 -0.0078 0.0115 -0.0103 0.0019* -0.0223** -0.0744**
TTF Day-Ahead -0.0626 -0.0229 -0.0477 -0.0338 -0.0110 0.0080** -0.1100** -0.1793**
Zeebrugge Month-Ahead -0.0214 0.0008 -0.0254 0.0004 -0.0094 0.0012** -0.0349** -0.0534**
Zeebrugge Day-Ahead -0.0401 -0.0368 -0.0264 -0.0114 0.0032 0.0038** -0.0708** -0.1340**
ICE Monthly Futures 0.0052 -0.0020 0.0008 -0.0041 -0.0048 0.0001 -0.0031 0.0049
Autocorrelation Coefficients Test Statistics
104
Table B2
Augmented Dickey-Fuller Stationarity Tests for Natural Gas Securities
Table B2 displays Augmented Dickey-Fuller (1979) test statistics for log price levels and continuously
compounded returns for natural gas securities sampled at 10-minute intervals between 8:00am and
4:00pm GMT from 2 January 2008 to 30 December 2011 (47,085 observations). The Augmented
Dickey-Fuller tests are run with a constant ( ) and 4 lags of differenced dependent variables as
explanatory variables ( 4k ) as per:
tjt
k
j
jtt yyy
1
1
Dependent variables (t
y ) are alternately differenced log price levels and differenced returns. Test
statistics, ˆˆ
tZ , are for 0:
0H , where
is the standard error of . * and ** denote
significance at the 5 and 1 per cent levels against critical values from Fuller (1996) of -2.86 and -3.43,
respectively.
Table B3
Johansen Cointegration Test for Natural Gas Securities
Table B3 displays the results of a Johansen (1995) cointegration test on the log price levels of the six
natural gas securities ( 6K ) sampled at 10-minute intervals between 8:00am and 4:00pm GMT from
2 January 2008 to 30 December 2011 (47,085 observations). Tests are run using a multivariate VECM
estimated by maximum likelihood which includes 4 lags ( 4n ) as explanatory variables to determine
the number of cointegrating relations ( r ) between the ( K ) variables:
tit
n
i
itt εyΓyβαy
1
1
1
Dependent variables (
ty ) are a 1K vector of differenced log price levels;
ity are lagged dependent
variables; α and β are rK parameter matrices in which the number of cointegrating equations is less
than the number of I(1) variables ( Kr ); 11
,,p
ΓΓ are KK matrices of parameters; and, tε is a
1K vector of normally distributed and serially uncorrelated error terms. The null hypothesis for the
trace statistic is that there are no more than r cointegrating relations (i.e. the eigenvalues Kr ,,
1
are
zero). * and ** denote the rank at which the null hypothesis cannot be rejected at the 5 and 1 per cent
levels of statistical significance, respectively (critical values are from Johansen, 1995).
NBP NBP TTF Zeebrugge Zeebrugge ICE
Month Day Day Month Day Monthly
Ahead Ahead Ahead Ahead Ahead Futures
Levels -1.64 -2.84 -2.28 -1.62 -2.77 -1.66
Returns -51.56** -53.73** -55.89** -53.29** -55.94** -51.73**
Maximum Rank Eigenvalue Trace Statistic 5% Critical Value 1% Critical Value
r ≤ 0 6,347.32 82.61 91.12
r ≤ 1 0.0921 1,797.32 59.24 66.71
r ≤ 2 0.0218 759.57 39.71 46.00
r ≤ 3 0.0100 286.33 24.08 29.19
r ≤ 4 0.0047 62.74 12.21 16.16
r ≤ 5 0.0013 2.33** 4.14 7.02
r ≤ 6 0.0001
105
C. Crude Oil
Table C1
Percentage of Trades by Time of Day for Brent and WTI Futures
Trade data sourced from Thomson Reuters Tick History. Shaded area contains the chosen intraday
window used in the analysis. Percentages calculated over the 2 January 2008 to 30 December 2011
sample period, which contains approximately 75 million Brent futures trades and 127 million WTI futures
trades87
.
87
Percentages of trades by time of day for the other securities in this study are available on request.
Along with trading hours information provided by the exchanges, these statistics were used in deciding
the intraday windows over which price discovery is assessed. They are explicitly provided here for the
Brent and WTI futures because of the large time differences between the UK and the US and to illustrate
that the hours with the greatest volumes of trade are nonetheless similar for both securities. This also
highlights how daily studies using traditional close of day prices for each market, respectively, may be
capturing very different market conditions. Specifically, a traditional 5pm London close price is captured
at a time of much greater liquidity than a 5pm Chicago close (around 11pm GMT depending on daylight
saving).
Time (GMT) ICE Brent Futures CME WTI Futures
0:00 - 1:00 0.10% 0.34%
1:00 - 2:00 0.18% 0.38%
2:00 - 3:00 0.21% 0.33%
3:00 - 4:00 0.20% 0.27%
4:00 - 5:00 0.19% 0.29%
5:00 - 6:00 0.21% 0.29%
6:00 - 7:00 0.67% 0.52%
7:00 - 8:00 2.52% 1.06%
8:00 - 9:00 6.84% 1.53%
9:00 - 10:00 5.99% 1.47%
10:00 - 11:00 5.53% 1.55%
11:00 - 12:00 4.89% 2.05%
12:00 - 13:00 5.96% 4.74%
13:00 - 14:00 9.87% 12.48%
14:00 - 15:00 12.46% 16.41%
15:00 - 16:00 14.91% 15.18%
16:00 - 17:00 11.07% 11.66%
17:00 - 18:00 7.06% 10.25%
18:00 - 19:00 7.40% 11.82%
19:00 - 20:00 3.08% 5.45%
20:00 - 21:00 0.42% 1.14%
21:00 - 22:00 0.14% 0.30%
22:00 - 23:00 0.03% 0.24%
23:00 - 24:00 0.08% 0.25%
Per cent Within Window 98.00% 96.79%
106
Table C2
Autocorrelation Coefficients and Test Statistics for Crude Oil Securities
Weighted autocorrelation coefficient estimates ( jD ) and2
test statistics calculated as per the unified approach in Richardson and Smith (1994) for Box and Pierce (1970)
Q-statistics, Fama and French (1988) beta statistics and Lo and MacKinlay (1988) variance ratios. * and ** denote significance at the 5 and 1 per cent levels against 2
critical
values with 10 degrees of freedom for the Q-statistics and variance ratios and 9 degrees of freedom for the beta statistics. Autocorrelation coefficients and statistics calculated
from continuously compounded returns for the crude oil securities sampled at 10-second intervals from 2 January 2008 to 30 December 2011 (ICE Brent Futures1 and Dated
BFOE Proxy are sampled daily between 7:00am and 5:00pm GMT yielding 3,680,221 observations, while ICE Brent Futures2 and CME WTI Futures are sampled daily
between 7:00am and 9:00pm GMT yielding 5,151,901 observations).
ρ1 ρ2 ρ3 ρ4 ρ5 ρ6 ρ7 ρ8 ρ9 ρ10 Q-statistic Beta Statistic Variance Ratio
ICE Brent Futures1
-0.0444 -0.0060 -0.0022 -0.0030 -0.0007 -0.0022 -0.0012 -0.0003 -0.0006 -0.0019 0.0020** -0.0184** -0.0996**
Dated BFOE Proxy -0.0235 -0.0030 -0.0014 -0.0018 -0.0031 -0.0016 -0.0016 -0.0010 -0.0009 0.0004 0.0006** -0.0141** -0.0571**
ICE Brent Futures2
-0.0513 -0.0087 -0.0068 -0.0034 -0.0026 -0.0033 -0.0029 -0.0017 -0.0005 -0.0017 0.0028** -0.0283** -0.1276**
CME WTI Futures -0.0237 -0.0039 -0.0011 -0.0014 -0.0004 -0.0010 -0.0027 0.0001 -0.0011 -0.0016 0.0006** -0.0111** -0.0551**
Autocorrelation Coefficients Test Statistics
107
Table C3
Augmented Dickey-Fuller Stationarity Tests for Crude Oil Securities
Table C3 displays Augmented Dickey-Fuller (1979) test statistics for log price levels and continuously
compounded returns for the crude oil securities sampled at 10-second intervals from 2 January 2008 to
30 December 2011. The unit root tests are run with a constant ( ) and 15 lags of differenced dependent
variables as explanatory variables ( 15k ) as per:
tjt
k
j
jtt yyy
1
1
Dependent variables (t
y ) are alternately differenced log price levels and differenced returns. Test
statistics, ˆˆ
tZ , are for 0:
0H , where
is the standard error of . * and ** denote
significance at the 5 and 1 per cent levels against critical values from Fuller (1996) of -2.86 and -3.43,
respectively.
Table C4
Johansen Cointegration Test for Crude Oil Securities
Table C4 displays Johansen (1995) cointegration tests of the log price levels of the ICE Brent futures and
dated BFOE proxy ( 2K ) sampled at 10-second intervals between 2 January 2008 and 30 December
2011. Tests are run using a multivariate VECM estimated by maximum likelihood, which includes
15 lags ( 15n ) as explanatory variables to determine the number of cointegrating relations ( r ) between
the ( K ) variables:
tit
n
i
itt εyΓyβαy
1
1
1
Dependent variables (
ty ) are a 1K vector of differenced log price levels;
ity are lagged dependent
variables; α and β are rK parameter matrices in which the number of cointegrating equations is less
than the number of I(1) variables ( Kr ); 11
,,p
ΓΓ are KK matrices of parameters; and, tε is a
1K vector of normally distributed and serially uncorrelated error terms. The null hypothesis for the
trace statistic is that there are no more than r cointegrating relations (i.e. the eigenvalues Kr ,,
1
are
zero). * and ** denote the rank at which the null hypothesis cannot be rejected at the 5 and 1 per cent
levels of statistical significance, respectively (critical values are from Johansen, 1995).
ICE Brent Futures Dated BFOE Proxy ICE Brent Futures CME WTI Futures
Intraday Window (GMT) 7:00am-5:00pm 7:00am-5:00pm 7:00am-9:00pm 7:00am-9:00pm
Levels -1.35 -1.30 -1.41 -1.58
Returns -586.45** -584.22** -699.75** -691.66**
Maximum Rank Eigenvalue Trace Statistic 5% 1%
r ≤ 0 22.11 12.21 16.16
r ≤ 1 0.00001 1.72** 4.14 7.02
r ≤ 2 0.00000
Critical Values
108
CHAPTER 4: Information Linkages between the Emission
Allowance and Energy Markets
Research on the interactions between emission allowances and energy markets commonly
supposes that fuel switching between coal and natural gas in power generation is the marginal
form of emissions abatement and that this should lead to a positive (negative) relationship
between natural gas returns (coal returns) and emission allowance returns88
. However, a number
of theoretical and practical considerations suggest observing any directional relationship
assumed between emission allowances and energy securities is likely spurious. Specifically,
while supply-side shocks may lead to the price behaviour assumed in prior research, a demand-
side shock that increases price and quantity demanded in either the natural gas or coal market,
should tend to raise prices and the quantity demanded in the other market due to substitution
effects. In turn, the price of emission allowances, which are complementary to both, would also
be expected to increase due to the expected rise in the quantity of fossil fuels combusted. In the
absence of a priori expectations for price and return relationships between these securities an
alternative approach to simple regression analysis is warranted.
We employ a rational expectations framework similar to that of Tauchen and Pitts (1983),
Fleming, Kirby and Ostdiek (1998) and Kodres and Pritsker (2002) that relates securities based
on their response to common information and the spillover of idiosyncratic information across
88
See Mansanet-Bataller, Pardo and Valor (2007), Alberola, Chevallier and Chèze (2008), Alberola,
Chevallier and Chèze (2009), Bonacina, Creti and Cozialpi (2009), Keppler and Mansanet-Bataller
(2010), Bredin and Muckley (2011), Creti, Jouvet and Mignon (2011) and Mansanet-Bataller, Chevallier,
Hervé-Mignucci and Alberola (2011).
109
markets. This allows for a more complete characterisation of the dynamics between the markets
of interest. The emission allowance and energy markets in question are expected to have strong
common information linkages as they share sensitivities to factors such as economic growth,
industrial production and the impact of extreme weather on power demand. They are also
expected to experience strong volatility spillovers driven by cross-market hedging as well as
spillovers driven by economic linkages; specifically the aforementioned relationships of coal
and natural gas as substitutes for one another and emission allowances as a complement to both.
We follow Fleming et al. (1998) in estimating a stochastic volatility representation of the
rational expectations model using GMM. Contrary to our expectations, emission allowances
exhibit the strongest linkages to the crude oil market. This is surprising given most combustion
activities related to crude oil occur in transportation, a sector not covered by the EU ETS. Given
their omission from the trading system, and the resultant decrease in direct economic linkages
between emission allowances and crude oil, there is less potential for information spillover
effects. Thus, we would predict a weaker relationship between the two markets. As such, the
strength of the linkages between these markets likely reflects strong common information
linkages. Despite the direct economic relationships between emission allowances and coal and
natural gas, which are combusted for heat and power generation and are covered by the EU
ETS, linkages with these markets are weaker.
The remainder of this paper is organised as follows: Section 4.1 discusses the limitations of
previous studies; against this backdrop, Section 4.2 presents the rational expectations model of
the information and volatility linkages between the markets; Section 4.3 details the regression
methodology employed in assessing directional relationships and the bivariate stochastic
volatility representation of the model used in Fleming et al. (1998); Section 4.4 describes the
data selection process and presents statistics for our final sample; Section 4.5 presents the
results of regression analysis and GMM estimation; and, Section 4.6 concludes.
110
4.1 Existing Evidence on Market Interactions
From a theoretical perspective, the price of carbon emission allowances should equal the
marginal cost of abatement (see Stern, 2006). This, after all, is the point of an emissions trading
system: place a cap on the total volume of greenhouse gases that polluters are allowed to emit
each year, making the right to pollute a scarce commodity, the cost of which must be
internalised. This should force marginal polluters, those who cannot remain profitable if
required to purchase allowances, either to abate their emissions by altering their current
production processes or to cease production altogether. As more than half of the emissions
covered by the EU ETS come from power generation, much of the literature on abatement looks
at the inter-relationship between emission allowance and energy prices in the context of fuel
switching in electricity generation89
.
These studies commonly assume that fuel switching informs the directional relationship
between emission allowance and energy prices. Moreover, they argue that, if energy input prices
are determined exogenously to emission allowance prices, an increase in the price of coal
relative to natural gas prompts greater gas-fired generation. This, in turn, leads to less demand
for emission allowances and a decrease in emission allowance prices. For example, Mansanet-
Bataller, Chevallier, Hervé-Mignucci and Alberola (2011) document a positive relationship
between both natural gas and crude oil price changes and emission allowance price changes
overall, but a negative dependence of the latter on coal price movements. In explaining their
results Mansanet-Bataller et al. (2011) argue: “this implies that when the coal price increases,
industries have an incentive to use less CO2-intensive fuels, which decreases the demand and
the price of CO2 allowances.”90
89
Given that electricity and heat production account for 24 per cent of the EU‘s total emissions and given
that the EU ETS covers around 40 per cent of the EU‘s total greenhouse gas emissions (encompassing
only the large, easily assessable polluters), electricity and heat production accounts for approximately 60
per cent of the EU ETS (European Environment Agency, 2011). Fuel switching from higher emission
brown (lignite) coal plants to lower emission black (bituminous) coal plants is also a form of abatement,
though coal-to-gas switching opportunities are thought to be more common.
90 Similar results and conclusions are drawn by Alberola et al. (2008) and Alberola et al. (2009),
Hintermann (2010) and Chevallier (2012).
111
However, this argument ignores the differing effect of demand-side and supply-side shocks,
an important distinction, given the substitutability of coal and natural gas and the fact that
emission allowances are a complementary good to both. For example, if there is a positive
demand shock in the coal market, the resultant increase in the coal price will likely trigger
increased demand for natural gas as the power generation sector substitutes towards the
relatively cheaper fuel which will, in turn, push up the price of natural gas. If increased demand
for both coal and gas causes market participants to believe there will be increased emissions, a
likely outcome where the usefulness of these commodities is almost entirely in their
combustion, the coal demand shock will result in an increase in the emission allowance price
and quantity demanded91
. A demand shock in the natural gas market would lead to similar
unambiguous price and quantity outcomes for emission allowances. These are not the
relationships predicted by the fuel switching literature, whose expectations are more consistent
with supply-side shocks in one market. For example, an expected decrease in natural gas supply
would see a corresponding price increase and fall in demand. Conversely, given its
substitutability for natural gas, coal will consequently experience an increase in demand and
price. The net effect of these opposite pressures on emission allowances would depend upon
income and demand elasticities and so are potentially ambiguous. However, this type of natural
gas supply shock would most likely prompt an increase in the quantity and price of emission
allowances due to the higher emission factor for combusting coal than natural gas. Thus, the
dominance of supply-side effects in energy markets might induce spurious correlations that only
appear consistent with a fuel switching argument. In this context, the introduction of a
methodological approach not solely predicated on directional price and return relationships
represents an important contribution.
In addition to incompletely describing fuel switching, previous studies suffer from a number
of practical problems with how fuel switching variables are constructed and analysed. Firstly,
91
There is one caveat on the demand-side effects described above. This relates to circumstances in which
the demand shock is driven by activity in a sector outside the EU ETS, for example in non-European
countries, as the increased emissions from higher foreign demand will not be related to European
emission allowance demand in the fashion described. It is not entirely certain to what extent European
coal and natural gas security prices are determined exogenously on world markets or endogenously within
Europe.
112
they ignore the obvious multicollinearity between switching price variables and the energy
prices from which they are constructed when both are used as explanatory variables92
. Secondly,
most fuel switching variables are created from dark and spark spreads—proxies for the profit
accruing to coal and gas generators, respectively—which frequently fail to incorporate
important operational and maintenance costs. These costs may induce strategic behaviour in
generators that is inconsistent with fuel switching. Lastly, fuel switching is unlikely to be
observed during peak electricity demand periods when most of a system‘s installed capacity is
running. Even if fuel switching is a source of short-term abatement during off-peak periods,
changes to a system‘s installed capacity can only be implemented in the long-run and so the
current diversity of the energy generation mix may limit the extent to which it occurs. We
elaborate on these problems in the Appendix, where we also provide a brief overview of EU
electricity markets and common approaches to the construction of fuel switching variables.
4.2 Information Linkages
The emission allowance and energy markets are driven by many common sources of
macroeconomic information. Further, idiosyncratic information in one market can influence
return volatility in another as a result of cross-market hedging or because the information
prompts market participants to trade substitute or complementary securities. These complexities
are more adequately reflected in our rational expectations specification, which is similar to
Tauchen and Pitts (1983), Fleming et al. (1998) and Kodres and Pritsker (2002). In the absence
of a priori expectations of directional relationships between the markets, we focus on this
model‘s implications for volatility linkages.
92
This occurs in Mansanet-Bataller et al. (2007), Alberola et al. (2008), Alberola et al. (2009), Bonacina
et al. (2009), Keppler and Mansanet-Bataller (2010), Bredin and Muckley (2011), Creti et al. (2011) and
Mansanet-Bataller et al. (2011).
113
Starting with the Tauchen and Pitts (1983) model, a speculative trader‘s demand for a
security at time ( t ), such as coal futures ( tcQ , ), is determined by the difference between their
individual valuation of that security (*
,tcp ) and its price ( tcp , )93
:
tctctc ppQ ,
*
,,
(4.1)
The constant ( ) represents factors (inversely) affecting traders‘ speculative demand, such
as their risk aversion ( ) and the variance of their expected profits (2
c ). Expanding equation
(4.1) to account for these factors gives:
2
,
*
,
,2 c
tctc
tc
ppQ
(4.2)
In the absence of liquidity traders, the market price is simply the average of individual
speculative valuations. When this average differs from a given trader‘s individual valuation,
they will transact so as to maximise their expected profit. Specifically, traders will take a long
position if 0,
*
, tctc pp , entering a short position if 0,
*
, tctc pp . As new information
arrives, traders reassess their valuations, trade accordingly and a new equilibrium price is
reached94
. For a trader limited to the coal futures market, their demand for coal futures increases
with expected profits, but decreases with increasing risk aversion or increasing expected profit
volatility.
Fleming et al. (1998) generalise Tauchen and Pitts‘ (1983) model to allow for the effects of
information arrival on a trader‘s demand in more than one market. If the trader from our
93
Tauchen and Pitts (1983) treat the prices in equation (4.1) as the trader‘s reservation price versus the
market price as we do, while Fleming et al. (1998) treat this as the difference between the expected future
spot price and the actual futures price, though they note that inasmuch as spot securities can be used for
hedging, their model is not specific to futures markets. Transaction costs are assumed zero in both
models.
94 Tauchen and Pitts (1983) treat the sequence of information arrival as triggering a series of distinct
Walrasian equilibria. This is similar to the implications in Ross (1989) in that the variance of price
changes is related to the rate of information flow, except the Tauchen and Pitts (1983) model is in discrete
and not continuous time.
114
previous example now speculates in both natural gas ( g ) and coal futures ( c ), we can express
their demand function in both markets as:
c
gc
tctc
cg
tgtg
g
g
cg
tgtg
gc
tctc
c
ppppQ
ppppQ
2
,
*
,
2
,
*
,
2
,
*
,
2
,
*
,
22
22
(4.3)
In this generalised model, c is the slope coefficient from a regression of the trader‘s
expected coal profits on their expected natural gas profits; 2
gc represents the variance of this
regression‘s error terms; and, g and 2
cg are analogously defined by a regression of expected
natural gas profits on expected coal profits. For each demand function in equation (4.3), the first
term represents the sensitivity of speculative demand for the commodity to changes in both
common and idiosyncratic information. The second term measures the change in hedging
demand for a commodity in response to changing expectations in the other market, or the
indirect spillover effect of information95
.
While Fleming et al. (1998) apply their model to a mean-variance optimising portfolio
manager operating in multiple markets, such as both the stock and bond markets, its application
is equally valid in describing a diversified power generator who must manage multiple fuel
input price exposures96
. Moreover, its application in the case of the latter allows fuel inputs to
be recognised as substitutes and fuel inputs and emission allowances to be seen as complements,
with these relationships characterised as driving potential spillovers between markets. In this
sense, tractable extensions of the generalised model in (4.3) could incorporate a larger number
of energy commodities and the addition of emission allowances.
95
Kodres and Pritsker (2002) form a similar rational expectations model in order to explain contagion
across markets in terms of a portfolio rebalancing channel. However, their specification highlights the
role of asymmetric information between developed and emerging markets, which seems less appropriate
for the markets under consideration in this study.
96 For other direct applications of the Fleming et al. (1998) model see Treepongkaruna and Gray (2009)
and Treepongkaruna, Brooks and Gray (2012) for evidence in the foreign exchange market and
Fleischer (2003) for evidence across different countries‘ stock, bond and money markets.
115
Like many asset pricing models, the usefulness of Fleming et al.‘s (1998) framework is not
in its direct empirical application; demand functions cannot be estimated as speculative traders‘
expected profits are inherently unobservable. Notwithstanding this, the model is valuable
insofar as it facilitates our understanding of the measureable linkages between the emission
allowance and energy markets. One example of a measureable linkage is the correlation in the
volatilities of security returns. The model predicts that, absent any market frictions, the direct
and indirect impact of information arrival should drive a perfect correlation in volatility. While
frictions including transaction costs, leverage constraints and illiquidity, will impair the
volatility linkages between the markets, strong linkages will remain where the securities are
driven by common information and given frequent inter-market spillovers resulting from
portfolio diversification benefits or economic linkages.
The literature provides some guidance regarding the identity of common sources of
information for both the emission allowance and energy markets. Specifically, research
invariably agrees on the importance of changing expectations for industrial production97
and the
effect of unanticipated weather events on power demand98
. The exact role of these fundamentals
is subject to some debate in the literature, although this is unsurprising given the noisy proxies
employed.
As noted previously, each market is also driven by a wide range of idiosyncratic factors
expected to have some indirect spillover into other markets through either hedging demand or
by virtue of the substitutability/complementarity of fuels and allowances. Examples of
idiosyncratic information satisfying this definition include: The impact of prospective conflict in
97
Studies attempting to examine the relationship between emission allowances, energy prices and
industrial production often have inconsistent results. This may result from a failure to acknowledge that it
is factors affecting expectations for future industrial production that are important and that backward
looking macroeconomic data may be a poor proxy for this (though admittedly past data might inform
expectations). Also problematic in prior studies, is the tendency to create daily industrial production
variables by simple linear interpolation of monthly data, which seems grossly inappropriate (see, for
example, Alberola et al. 2009, and Bredin and Muckley, 2011).
98 Most studies show that extreme cold weather has a greater effect on heating and power demand, and
thus emission allowances, than is the case for extreme hot weather, which drives demand for air
conditioning, though both are frequently found to be statistically significant. See, for example, Mansanet-
Bataller et al. (2007), Alberola et al. (2008), Fezzi and Bunn (2009) and Keppler and Mansanet-Bataller
(2010). Some studies, such as Hintermann (2010), also include rainfall or reservoir level variables due to
their impact on hydroelectricity, which is especially relevant for Scandinavia. While weather data are
prolifically available at very timely frequencies, it is difficult to say how a variable representative of and
encompassing the large geographic region covered by the EU ETS should be constructed.
116
the Middle East, which may prompt traders to revise their expectations for crude oil prices;
Reports concerning shale gas exploration and discoveries, which impact upon expectations for
natural gas prices; Changes in coal demand from steel mills, which directly affects expectations
for coal prices; and, Announcements regarding the level of verified emissions, which affect
emission allowance prices99
.
Importantly, the impact of common information and information spillover on market
volatilities cannot be distinguished from one another. Nonetheless, these theoretical distinctions
inform our expectations for the strength of the observable linkages overall. More specifically,
because the power generation sector makes up the largest component of the EU ETS and
emissions from power generation are predominantly from coal and natural gas combustion, we
would expect stronger spillover effects between the coal, natural gas and emission allowance
markets. Conversely, we predict limited spillover effects from crude oil to emission allowances
markets, given the ultimate combustion of oil-related fuels largely occurs in the transportation
sector, most of which is currently outside the EU ETS100
. The expectation that linkages between
coal, natural gas and emission allowances will be stronger than those between crude oil and
emission allowances constitutes the „Spillover Chanel Hypothesis‟.
Alternatively, the greater depth and liquidity of the crude oil market suggests that coal,
natural gas and emission allowance prices may closely follow crude oil market developments
inasmuch as information relevant to common fundamentals in their pricing will be impounded
into prices in this vastly more liquid market much more quickly and completely. That emission
allowances have stronger linkages to the crude oil market than to the coal and natural gas
markets constitutes the „Common Information Chanel Hypothesis‟.
99
Mansanet-Bataller and Pardo (2009) conduct event studies of regulatory announcements in the EU
ETS, such as those relating to National Allocation Plans and verified emission reports, and find they have
a significant impact on emission allowance prices. Interestingly, they also detect some evidence
consistent with a small amount of insider trading ahead of these announcements.
100 Within electricity and heat production, about 45 per cent of emissions come from solid fuels (coal),
38 per cent come from natural gas and only 5 per cent come from liquid fuels (oil) according to the
European Environment Agency (2011). While these figures are for emissions, this is also illustrated in the
breakdown of electricity generation by fuel type for the EU-27, which is: nuclear 27.8 per cent; solid fuels
(coal) 26.7 per cent; natural gas 24.0 per cent; renewables 16.8 per cent; and, oil 3.1 per cent (European
Commission, 2011).
117
4.3 Methodology
Initial testing seeks to better understand the direction of relationships between the energy and
emission allowance markets. To this end, we regress emission allowance returns against
contemporaneous and lagged returns for coal, natural gas and crude oil securities. Thereafter,
we employ the stochastic volatility model of Fleming et al. (1998) to understand the volatility
linkages between markets. In this context, our main statistic for gauging the relative strength of
the linkages between the respective markets is the correlation in cross-market log information
flows (volatilities).
4.3.1 Directionality of Emission Allowance and Energy Market Relationships
We regress daily emission allowance returns, eR , against contemporaneous and lagged
daily returns for coal, cR , natural gas, gR , and crude oil, oR , in order to observe any
directional relationships between these markets. To deal with autocorrelation and
heteroskedasticity in the returns, we use the Newey-West (1987) technique for estimating the
residual variance-covariance matrix101
. Formally, we fit the following regression model:
t
l
toto
l
tgtg
l
tctcte RRRR
1
,,
1
,,
1
,,,
(4.4)
In addition to contemporaneous energy security returns, the regression is run for several ( l )
lagged energy security returns as independent variables in order to rule out emission allowances
having a delayed response to returns in these markets.
4.3.2 Stochastic Volatility Model
Fleming et al. (1998) formulate a stochastic representation of the rational expectations
trading model in equation (4.3). We use Hansen‘s (1982) generalised method of moments
101
We employ up to 5 lags in the Newey-West (1987) specification which appears adequate given the
autocorrelation statistics presented in the next section.
118
(GMM) to estimate the bivariate specification of the Fleming et al. (1998) model, and, in doing
so, we derive the contemporaneous correlations of the log information flows (volatilities) as
measures of the strength of the information linkages between the markets of interest.
The returns used in this study are log first differences in daily prices, 1,,, ln tktktk ppR ,
for the respective emission allowance and energy securities ( k coal, natural gas, crude oil and
emission allowances). Information is assumed to arrive randomly through the trading day
generating incremental price changes, tik , . The sum of these incremental changes constitutes
the unpredictable component of daily returns generated by the daily number of information
events, tkI , :
tkI
i
tiktktkR,
1
,,,
(4.5)
The predictable component of daily returns is the conditional expected value, tk , , while
the incremental intraday returns are assumed to be normal, independent and identically
distributed variables with a zero mean and variance of 2
,k . Decomposing the second term in
equation (4.5) yields:
tkI
i ktiktktk
tktkktktk
Iz
zIR
,
1 ,,
2/1
,,
,
2/1
,,,,
1
(4.6)
Under the central limit theorem, as the number of information events increases ( tkI , ),
the distribution of tkz , approaches a standard normal. This, in turn, implies that returns are
approximately normally distributed with a mean of tk , and a variance of tkk I ,
2
, . In line with
Ross (1989), the volatility of returns, 21
,, tkk I , is proportional to the number of information
events, tkI , , with more information flows leading to greater return volatility. While equation
(4.6) defines the return generating process, Fleming et al. (1998) model the stochastic volatility,
tkktk Ih ,
2
,, ln , as AR(1) given the empirical support for this structure in the literature. The
119
joint stochastic process by which returns and volatility are determined is then expressed in terms
of tkh , as:
tktkkhkhtk
tktktktk
uhh
zhR
,1,,,,
,,,, 5.0exp
(4.7)
The residuals from the AR(1) volatility process, tku , , are assumed to have a mean of zero
and be independent of tkz , , implying that, while information arrival incrementally determines
returns, the actual number of information events itself does not. GMM moment restrictions are
formulated from the unpredictable return component, tktktk Rr ,,, , and its volatility, which
is:
2
,,
2
, lnln tktktk zhr
(4.8)
In line with Fleming et al. (1998) we first remove seasonality from the returns by regressing
them against day-of-the-week and post-public holiday dummy variables and using the residuals,
tkr , , to construct the series 2
,ln tkr . Similarly, we remove volatility seasonality by regressing
this series against Monday (i.e. post-weekend) and post-public holiday dummy variables.
Re-arranging (4.8), we define the estimated volatility series as:
2
,
2
,, lnln tktktk zEry
(4.9)
In which 93.4,27.1~ln 2
, Nz tk , given 1,0~, Nz tk . The series tky , is estimated with
error: tktktk hy ,,, , that is 2
,
2
,, lnln tktktk zEz , and thus 93.4,0~, Ntk and is
independent of tkh , . This allows for the definition of univariate moment conditions for
estimating the mean, kh, , variance, 2
,kh , and AR(1) coefficient, kh, , values for each market‘s
volatility series tky , :
120
tkkhtktk
tktktk
tktk
hyy
hy
hEyE
,,,,
,,,
,,
var,cov
varvarvar
(4.10)
The bivariate model estimates (4.10) for two securities as well as cross-market linkages
between them measured by the correlation of tih , and tjh , . Including the correlation between the
error terms ti , and tj , , cross-market restrictions are given by:
tjtiihtjti
tjtijhtjti
tjtitjtitjti
hhyy
hhyy
hhyy
,,,,,
,,,,,
,,,,,,
,cov,cov
,cov,cov
,cov,cov,cov
(4.11)
Six distinct bivariate pairings of the four securities of interest are considered, namely:
emission allowances and coal; emission allowances and natural gas; emission allowances and
crude oil; coal and natural gas; coal and crude oil; and, natural gas and crude oil. For each
bivariate pairing, the GMM disturbance vector derived from equations (4.10) and (4.11) is:
ijjhtjihtiihjhtjihti
ijjhtjihtijhjhtjihti
ijjhihijhjhtjihti
tjtjjhtjtjtitj
jhtjtj
tjtj
titiihtitititi
ihtiti
titi
t
yyyy
yyyy
yy
yyy
y
y
yyy
y
y
e
,
2
,,,,
2
,,,,,
,
2
,,,,
2
,,,,,
,
2
,,,,,,,
22
,,
2
,,,,,
22
,
2
,,
,,
22
,,
2
,,,,,
22
,
2
,,
,,
(4.12)
The vector of unknown parameters for a bivariate pairing of securities i and j is
ijijhjhjhjhihihih ,,,
2
,,,
2
,, ,,,,,,, , where ij, is the correlation between the error
terms, ti , and tj , , and ijh, is the correlation between the log information flows (volatilities),
121
tih , and tjh , . Equation (4.12) is estimated using a variable number of autocorrelation
restrictions, l,2,1 , which range in value from 1l to 40 . Thus, in the bivariate system
in (4.12), there are 54 l equations with eight unknowns. As per Hansen (1982), the
parameters are estimated by minimising TT gSg 1ˆ for
lT
t tT elTg1
1 102.
J-statistic tests for over-identifying restrictions are calculated for each bivariate pairing and are
distributed 2
34 l .
4.4 Data
In this section we explain which securities are selected for analysis from their respective
markets. We also present descriptive statistics, serial correlation tests and cross-market
correlations for these securities. All data are sourced from Thomson Reuters.
4.4.1 Security Selection
Securities are selected for use in the analysis on the basis that they are the most likely to
reflect pertinent information for the emission allowance and the energy markets of interest.
Information regarding these likelihoods in the context of the EU ETS and the European energy
markets is inferred from the analysis of short and long-run price discovery in Chapter 2 and
Chapter 3, respectively.
In terms of emission allowances, the main type of securities traded in the EU ETS are
European Union Allowances (EUAs). These are traded over-the-counter and in spot, futures and
option markets facilitated by approximately nine organised exchanges. Intraday data from the
main spot and futures securities traded in the EU ETS are examined in Chapter 2 using a
102
The variance-covariance matrix, S , is estimated using quadratic spectral weights to adjust for
conditional heteroskedasticity and autocorrelation rather than the Parzens weights Fleming et al. (1998)
employ. We note that the choice of weighting produces minimal changes in the cross-market correlation
estimates, with the resultant correlations produced by the use of quadratic spectral weights being
approximately 1 - 2 per cent lower for all pairings than if Parzens weights are used.
122
regression approach to analyse short-term return dynamics and Hasbrouck‘s (1995) information
shares to assess their relative contribution to long-run price equilibrium. The results show that
the annual expiry Intercontinental Exchange (ICE) EUA futures contracts are overwhelmingly
the most important security for price discovery and, as such, they are used in this study (see also
Chevallier, 2010a; Chevallier, 2010b; and, Mizrach and Otsubo, 2011).
In Europe, coal, natural gas and crude oil are priced on a mixture of long-term contracts and
market pricing mechanisms based on over-the-counter physical trading and derivative trading
on organised exchanges. These markets differ markedly in their depth, liquidity and
transparency. For example, the market for coal arriving in the ports of Amsterdam, Rotterdam
and Antwerp is dominated by long-term bilateral contracts for which prices are not directly
observable. Instead, price reporting agencies survey market participants in order to assess
contract prices. However, these are not always available in a timely fashion. Moreover, while
derivative securities for coal are available from several European exchanges, these contracts
trade infrequently and reported settlement prices are often determined by averages of bid and
ask prices or from surveys of market participants. At the other end of the liquidity and
transparency spectrum, the ICE Brent crude oil futures contracts are some of the most heavily
traded securities in the world and thus present few problems in assessing relevant prices. From
the results of an investigation into price discovery in these markets contained in Chapter 3 we
select the ICE monthly expiry Rotterdam coal futures, the ICE monthly expiry UK natural gas
futures and the ICE monthly expiry Brent crude oil futures as the securities most reflective of
current information in their respective energy markets.
Due to the lack of liquidity in the coal futures market, analysis involving all three energy
securities and the emission allowance futures is, by necessity, conducted on a daily basis. In
constructing the daily return time series, each day‘s price observation is gathered at the time
closest to the time stamp of the coal futures so as to minimise any intraday lack of
contemporaneity in the observations103
. The prices of the coal futures, Brent futures and natural
103
Over the four-year sample period, the average latency with which the ICE Brent crude oil futures price
is observed compared to the time stamp on the ICE Rotterdam coal futures price is 12 seconds, for ICE
UK natural gas futures it is 38 minutes and for the ICE EUA futures it is 39 minutes. These differing
latencies are an indication of the relative frequency of trading in these markets around the time settlement
123
gas futures are all converted into euro using an Electronic Broking Services (EBS) tick
exchange rate data set, once again linked to the intraday time stamps on the coal futures price
observations. Chart 4.1 displays the price of the four securities converted from their usual
quotation units into euro per tonne for comparability.
Returns are calculated as log first differences in prices over the 4-year period from
2 January 2008 to 30 December 2011, inclusive104
. Dates in this period are only used when all
comparable securities are traded or settlement prices are recorded. The futures are front
contracts up until the day prior to expiry except for the EUA futures which are rolled two weeks
prior to expiry as this largely accords with the timing of the shift in volume to the next-to-front
contracts in each of these markets. The return impact of the switch into the next-to-front
contract is stripped out by indexing the units back into those of the first contract in each series.
4.4.2 Descriptive Statistics
Descriptive statistics for all securities we consider are displayed in Table 4.1. Examination
of the table reveals that the ICE UK natural gas futures returns are the most volatile, having the
largest standard deviation and the widest inter-quartile range, while the ICE coal futures are the
least volatile105
. The large maximum and minimum return observations were investigated and
found to be genuine reflections of the volatility in these markets, with many of the extreme
observations occurring during the crucial months of the financial crisis in late 2008 and early
2009.
prices in the coal futures market are published, which is typically around 4:00pm - 5:00pm London time
each trading day.
104 The beginning of Phase II of the EU ETS is chosen as the start date to avoid the problems associated
with the oversupply of allowances and prohibitions on banking across phases that led to a price collapse
in Phase I.
105 Note that the average return for the ICE Brent crude oil futures is negative even though the price
displayed in Chart 4.1 for the last day of the sample is higher than for the first day of the sample. This is
because the futures contracts are indexed back to the original contract such that the return effect of rolling
contracts a day prior to expiry each month is stripped out. This discrepancy between visual inspection of
Chart 4.1, which contains prices before indexation, and the return statistics in Table 4.1 is supported by
the fact that the front and next-to-front contracts in the Brent futures curve were in contango more often
than backwardation between 2008 and 2011.
124
Chart 4.1
Emission Allowance and Energy Prices
ICE monthly expiry Rotterdam coal futures converted from US dollars/tonne into euro/tonne using intraday EBS exchange rates; ICE
monthly expiry UK natural gas futures converted from Great British pence/therm into euro/tonne using EBS exchange rates and a
factor of 396.53 therms/tonne; ICE monthly expiry Brent crude oil futures converted from US dollars/barrel into euro/tonne using
EBS exchange rates and a factor of 7.64 barrels/tonne (utilising the API gravity of the Forties blend of 40.3 degrees as this North Sea
grade typically sets the price of BFOE crude—see Platts, 2012); and, ICE annual expiry EUA futures price in euro/tonne of CO2e .
0
100
200
300
400
500
600
700
800
0
100
200
300
400
500
600
700
800
ICE Rotterdam Coal Futures ICE UK Natural Gas Futures ICE Brent Crude Oil Futures ICE EUA Futures
€/t €/t
2008 2009 2010 2011
125
Table 4.1
Descriptive Statistics
Table 4.1 presents descriptive statistics of continuously compounded returns for ICE monthly expiry
Rotterdam coal futures, ICE monthly expiry UK natural gas futures, ICE monthly expiry Brent crude oil
futures and ICE annual expiry EUA futures sampled daily from 2 January 2008 to 30 December 2011
(959 observations).
4.4.3 Serial Correlation
Table 4.2 displays the results of three distinct tests for serial correlation on our series of
interest. The Box and Pierce (1970) Q-statistic tests indicate statistically significant serial
correlation in the Brent crude oil futures and the EUA futures at the 5 per cent level. Though the
very small size of the autocorrelation coefficients renders this economically unimportant, the
presence of statistically significant serial correlation supports the use of heteroskedasticity and
autocorrelation consistent standard errors in econometric testing. Consistent with this, we
employ the Newey-West (1987) method of standard error calculation when estimating
equation (4.4) and we employ GMM in the estimation of the Fleming et al. (1998) stochastic
volatility model.
ICE Monthly ICE Monthly ICE Monthly ICE Annual
Rotterdam Coal UK Natural Gas Brent Crude EUA
Futures Futures Futures Futures
Mean 0.0000 -0.0028 -0.0001 -0.0013
Standard Deviation 0.0180 0.0297 0.0221 0.0263
Skewness -1.00 -0.14 0.37 0.55
Kurtosis 18.01 5.87 8.69 10.87
Maximum 0.1295 0.1299 0.1793 0.2126
75th
Percentile 0.0081 0.0117 0.0114 0.0130
25th
Percentile -0.0068 -0.0181 -0.0117 -0.0142
Minimum -0.1736 -0.1684 -0.0821 -0.0982
126
Table 4.2
Serial Correlation Tests
Table 4.2 presents autocorrelation coefficients and test statistics of continuously compounded returns for
ICE monthly expiry Rotterdam coal futures, ICE monthly expiry UK natural gas futures, ICE monthly
expiry Brent crude oil futures and ICE annual expiry EUA futures sampled daily between 2 January 2008
and 30 December 2011 (959 observations). Weighted autocorrelation coefficient estimates ( jD ) and
2
test statistics are calculated as per the unified approach in Richardson and Smith (1994). * and **
denote significance at the 5 and 1 per cent levels against 2
critical values with 5 degrees of freedom for
Box and Pierce (1970) Q-statistics, Fama and French (1988) beta statistics and Lo and MacKinlay (1988)
variance ratios.
4.4.4 Cross-Market Correlations
Table 4.3 reports cross-market correlations for returns and two volatility proxies, namely
absolute returns and returns squared. All the return correlations in Panel A are positive and
range between 6.3 per cent and 30.3 per cent. The correlations in Panel A show a stronger
directional relationship between emission allowance and crude oil returns over the 2008-2011
sample (26.0 per cent) than between emission allowances and natural gas (21.4 per cent),
despite most oil-related combustion activities falling outside the scope of the EU ETS. In
contrast to the arguments in the fuel switching literature reviewed in Section 4.1, there is also a
small positive return correlation between emission allowances and coal (6.3 per cent).
In terms of the proxies for volatility, the correlations in Panels B and C are all smaller than
the correlations between returns, with the exception of those between crude oil and natural gas
and between emission allowances and coal. Emission allowances have the highest volatility
correlation with natural gas for both absolute returns and returns squared (15.6 per cent and 10.1
per cent, respectively). This is in line with spillover effects between these securities resulting
from their complementary relationship. However, as pointed out by Fleming et al. (1998),
absolute returns and returns squared are only noisy approximations of volatility such that their
correlations may not capture the depth of the linkages between the markets of interest. In this
ρ1 ρ2 ρ3 ρ4 ρ5 Q-statistic Beta Statistic Variance Ratio
ICE Rotterdam Coal Futures 0.0199 -0.0154 -0.0956 0.0180 0.0261 0.0108 -0.0785 -0.0559
ICE UK Natural Gas Futures 0.0790 -0.0446 -0.0469 0.0206 -0.0159 0.0111 -0.0419 0.0436
ICE Brent Crude Oil Futures 0.0785 -0.0511 -0.0326 -0.0566 0.0258 0.0137* -0.0696 0.0156
ICE EUA Futures 0.0616 -0.0749 0.0260 -0.0416 -0.0303 0.0127* -0.0412 0.0128
Autocorrelation Coefficient Test Statistics
127
sense they serve as a benchmark against which to judge the more precise estimates of the
volatility linkages presented in the next section.
Table 4.3
Cross-Market Correlations
Table 4.3 presents the cross-market correlations for ICE monthly expiry Rotterdam coal futures, ICE
monthly expiry UK natural gas futures, ICE monthly expiry Brent crude oil futures and ICE annual expiry
EUA futures sampled daily between 2 January 2008 and 30 December 2011 (959 observations). Panel A
contains correlations of continuously compounded returns. Panel B contains correlations of the absolute
value of returns. Panel C contains correlations of returns squared.
4.5 Results
We present the results of regressing emission allowance returns against contemporaneous and
lagged coal, natural gas and crude oil returns between 2008 and 2011. However, given that
directional relationships are not necessarily expected a priori, we estimate the stochastic
volatility model in Fleming et al. (1998) using GMM. These results show that the correlation of
log information flows for emission allowances are most highly correlated with the crude oil
market, supporting our Common Information Channel Hypothesis.
Panel A:
Coal Natural Gas Crude Oil Emission Allowances
Coal 1.000
Natural Gas 0.303 1.000
Crude Oil 0.195 0.145 1.000
Emission Allowances 0.063 0.214 0.260 1.000
Panel B:
Coal Natural Gas Crude Oil Emission Allowances
Coal 1.000
Natural Gas 0.175 1.000
Crude Oil 0.184 0.161 1.000
Emission Allowances 0.111 0.156 0.147 1.000
Panel C:
Coal Natural Gas Crude Oil Emission Allowances
Coal 1.000
Natural Gas 0.065 1.000
Crude Oil 0.097 0.227 1.000
Emission Allowances 0.033 0.101 0.054 1.000
Correlation of Returns - ρ(R)
Correlation of Absolute Returns - ρ(|R|)
Correlation of Returns Squared - ρ(R2)
128
4.5.1 Regression Results
The results of running the regressions as per equation (4.4) are presented in Table 4.4. In
regression (A), only contemporaneous energy market returns are used as independent variables
and, similar to the return correlations in Table 4.3, the coefficients are positive for crude oil and
natural gas, while the coefficient for coal returns is not significantly different from zero. With
the exception of coal, the coefficients are of a similar magnitude to the return correlations in
Panel A of Table 4.3.
Regressions (B) and (C) include one and two-day lagged energy market returns as
independent variables, respectively106
. In regressions (B) and (C) lagged crude oil returns are
significant at one and two-days, but with much smaller (and negative) coefficients and only at
the 5 per cent level. The inclusion of the lagged coefficients does little to improve the fit of the
model, with the adjusted R-squared only rising from 10 to 11 per cent. These results point to
most of the return relationship between emission allowances and the energy market securities
being captured contemporaneously.
4.5.2 Information Linkages
The GMM parameter estimates of Fleming et al.‘s (1998) stochastic volatility model are
presented in Table 4.5. These estimates are generated by fitting the moment restrictions in
equation (4.12) using the log, squared return series described in equation (4.9), with weekday
and public holiday seasonality removed as previously noted. Panel A details the estimates for
bivariate pairings of emission allowances with coal, natural gas and crude oil, respectively.
Panel B details the three bivariate pairings amongst the energy securities themselves. For all six
bivariate pairings, the J-statistic tests for over-identifying restrictions indicate that the models
106
None of the coefficients are significant for lags greater than two days. Several regressions were also
run with the coal, natural gas and crude oil security returns as the dependent variables and with the
emission allowance returns as independent variables. In these regressions the contemporaneous emission
allowance coefficients are similarly positive and statistically significant in explaining crude oil and
natural gas returns, but not significantly different from zero against coal returns. In addition, lagged
emission allowance return coefficients are not significant at standard levels, which indicates that on a
daily basis emission allowance returns do not lead the returns of the energy securities considered (these
results are available on request).
129
are not miss-specified. Furthermore, the mean, variance and autocorrelation parameters are all
statistically significant at the 1 per cent level.
Table 4.4
Regression Results
Table 4.4 presents the results of fitting model (4.4) using continuously compounded returns for ICE
annual expiry EUA futures, ICE monthly expiry Rotterdam coal futures, ICE monthly expiry UK natural
gas futures and ICE monthly expiry Brent crude oil futures sampled daily between 2 January 2008 and
30 December 2011 (959 observations). In regression (A), the independent variables are contemporaneous
returns for the coal (tc
R,
), natural gas (tg
R,
) and crude oil (to
R,
) securities, with the response variable
being emission allowance returns (te
R,
). The independent variables in regression (B) additionally contain
returns lagged by one-day ( 1l ), while regression (C) additionally contains returns lagged by two-days
( 2l ). Formally:
ttototgtgtctcte
lll
RRRR
1,,
1,,
1,,,
(4.4)
The square brackets [ ] below coefficients contain t-statistics, while round brackets ( ) below F-statistics
contain p-values. * and ** denote significance at the 5 and 1 per cent levels, respectively.
(A) (B) (C)
α -0.001 -0.001 -0.001
[-1.01] [-1.21] [-1.43]
βcoal,t -0.060 -0.049 -0.037
[-1.16] [-0.91] [-0.69]
βcoal,t-1 -0.021 -0.010
[-0.43] [-0.20]
βcoal,t-2 -0.032
[-0.78]
βnatural gas,t 0.170** 0.173** 0.170**
[4.85] [4.95] [4.85]
βnatural gas,t-1 -0.049 -0.045
[-1.33] [-1.14]
βnatural gas,t-2 -0.052
[-0.81]
βcrude oil,t 0.286** 0.293** 0.287**
[5.58] [5.85] [5.71]
βcrude oil,t-1 -0.088* -0.080
[-2.02] [-1.92]
βcrude oil,t-2 -0.074*
[-2.06]
Adj-R2
0.098 0.106 0.113
F-stat 21.70** 15.50** 11.19**
(0.000) (0.000) (0.000)
130
The mean log information flow estimates, kh, , indicate that coal futures have the lowest
average volatility followed by crude oil and emission allowances, with natural gas having the
highest average volatility. This is consistent with the ordering of the return standard deviations
presented in Table 4.1107
. These mean log information flows are not dissimilar to those reported
in Fleming et al. (1998) for stocks and bonds, though they are all generally higher than reported
estimates for Treasury bills, which is to be expected given return volatility is typically quite low
in the Treasury bill market.
The estimated variance of the log information flows, 2
,kh , is typically highest for the
emission allowance futures. In general, the variance parameters in Table 4.5 are higher than
those reported in Fleming et al. (1998), suggesting greater kurtosis in the distribution of
emission allowance and energy market returns than in stocks, bonds and bills. However, it
should be noted that this is likely a product of the sample period used in this study rather than
necessarily being an intrinsic feature of these markets. More specifically, our 2008-2011 sample
covers a period in which volatility greatly increased across all markets and was sustained for a
considerable period of time due to the financial crisis. This volatility created many extreme
return observations and, thus, return distributions exhibit substantial excess kurtosis and the
variances of the log information flows are mostly higher than those reported in previous studies.
The autocorrelation parameters for the log information flows, kh, , which are in a tight
range from 0.983 to 1.005, indicate a very high degree of persistence in volatility
autocorrelation and support the use of the autoregressive structure. Though the reported log
information flow autocorrelation parameters are for an estimated lag length ( l ) of 40, they are
little changed in other specifications in which the lag length is greater than one.
107
As the mean log information flows are in units of log, squared returns, they are negative numbers.
While this makes it difficult to readily interpret them, by taking the square root of the inverse of the
natural logarithm, the average volatilities can be brought back into units of return variance.
131
Table 4.5
GMM Results
Table 4.5 presents the GMM parameter estimates from fitting the Fleming et al. (1998) bivariate stochastic volatility model for the moment restrictions in equation (4.12). The
seasonally adjusted volatility series tk
y,
is constructed as per equation (4.9) from continuously compounded returns for ICE annual expiry EUA futures, ICE monthly expiry
Rotterdam coal futures, ICE monthly expiry UK natural gas futures and ICE monthly expiry Brent crude oil futures sampled daily from 2 January 2008 to 30 December 2011
(959 observations). In Panel A, the bivariate pairings ( jik , ) are for the emission allowance and energy markets: (1) emission allowances and coal; (2) emission allowances
and natural gas; and, (3) emission allowances and crude oil. In Panel B, the bivariate pairings contain the linkages between the energy markets: (4) coal and natural gas; (5) coal
and crude oil; and, (6) natural gas and crude oil. The parameter estimates are the mean (kh ,
), variance (2
,kh ) and AR(1) coefficient (
kh , ) of the log information flows as well
as the correlations of the log information flows (ijh ,
) and the correlations between the error terms (ij,
). Reported results are for a lag length of 40l . Round brackets ( )
below coefficients are standard errors. * and ** denote significance at the 5 and 1 per cent levels, respectively. Over-identifying J-statistics are distributed 2
34 l
.
(1) Emissions (i) Coal (j) (2) Emissions (i) Natural Gas (j) (3) Emissions (i) Crude Oil (j)
μh,i -10.403** μh,j -11.345** μh,i -10.428** μh,j -10.147** μh,i -10.111** μh,j -10.471**
(0.057) (0.061) (0.059) (0.056) (0.054) (0.048)
σ2
h,i 0.918** σ2
h,j 0.780** σ2
h,i 0.902** σ2
h,j 0.788** σ2
h,i 0.893** σ2
h,j 0.631**
(0.047) (0.046) (0.046) (0.047) (0.046) (0.052)
φh,i 0.983** φh,j 0.989** φh,i 0.983** φh,j 0.987** φh,i 0.989** φh,j 0.992**
(0.004) (0.004) (0.004) (0.004) (0.003) (0.004)
ρh,ij 0.520** ρh,ij 0.246** ρh,ij 0.769**
(0.059) (0.080) (0.054)
ρξ,ij 0.003 J-statistic 153.68 ρξ,ij 0.091** J-statistic 165.91 ρξ,ij 0.034 J-statistic 146.07
(0.020) p-value 0.560 (0.021) p-value 0.298 (0.017) p-value 0.724
Panel A: Emission Allowance and Energy Market Linkages
132
Table 4.5
GMM Results (Continued)
(4) Coal (i) Natural Gas (j) (5) Coal (i) Crude Oil (j) (6) Natural Gas (i) Crude Oil (j)
μh,i -11.463** μh,j -9.924** μh,i -11.344** μh,j -10.420** μh,i -9.912** μh,j -10.516**
(0.060) (0.057) (0.059) (0.050) (0.056) (0.057)
σ2
h,i 0.782** σ2
h,j 0.774** σ2
h,i 0.781** σ2
h,j 0.652** σ2
h,i 0.781** σ2
h,j 0.651**
(0.053) (0.047) (0.047) (0.051) (0.047) (0.053)
φh,i 0.991** φh,j 0.994** φh,i 0.991** φh,j 1.005** φh,i 0.989** φh,j 0.993**
(0.004) (0.003) (0.003) (0.003) (0.004) (0.004)
ρh,ij 0.576** ρh,ij 0.825** ρh,ij 0.493**
(0.064) (0.050) (0.063)
ρξ,ij -0.016 J-statistic 152.65 ρξ,ij -0.065** J-statistic 146.47 ρξ,ij 0.007 J-statistic 151.01
(0.021) p-value 0.583 (0.017) p-value 0.716 (0.016) p-value 0.620
Panel B: Energy Market Linkages
133
The parameters of most interest in the bivariate model are the estimated correlations
between the log information flows, ijh, , which we use as measures of the information linkages
across markets. In comparison to the correlation of absolute returns and returns squared, which
are noisy proxies for correlations in volatilities (ranging from 3.3 to 22.7 per cent), the
correlations between the log information flows are much higher, ranging from 24.6 to 82.5 per
cent. These correlation parameter estimates appear reasonably accurate, with standard errors
between 5.0 and 8.0 per cent, and are all significant at the 1 per cent level. The relative size of
these information flow correlations has implications for our competing hypotheses on the
linkages between the emission allowance and energy markets.
The Spillover Chanel Hypothesis predicts that the strong economic linkages between
emission allowances and fuels that are commonly combusted for power generation—coal and
natural gas—will result in information that creates volatility in one of these markets spilling
over into the other markets. This spillover channel exists on the basis that emission allowances
are a necessary complementary good for fuels whose combustion is subject to the EU ETS.
Under our theoretical specification, this implies linkages through the second terms in
equation (4.3). On the other hand, the combustion of refined products related to crude oil occurs
largely in the transportation sector, which is outside the EU ETS and does not require the
surrendering of emission allowances in abatement. As such, the Spillover Chanel Hypothesis
predicts that the correlation between the information flows relevant to crude oil and emission
allowances will be low, as there are fewer direct economic linkages.
The results in Table V clearly contradict the expectations of the Spillover Chanel
Hypothesis. The correlation between the log information flows (volatilities) for emission
allowances and coal is 52.0 per cent, between emission allowances and natural gas it is only
24.6 per cent, while the correlation between emission allowances and crude oil is the highest at
76.9 per cent. These results support the Common Information Channel Hypothesis because, in
the absence of a strong direct economic relationship between emission allowances and crude oil
that would prompt interaction via the spillover channel, the strength of the volatility linkages is
likely driven by common information. However, support for the Common Information Channel
134
Hypothesis more generally across all the securities cannot be inferred simply from the lack of
support for the Spillover Channel Hypothesis.
The Common Information Channel Hypothesis posits that the linkages between these
securities will be the result of them commonly sharing sensitivities to particular types of
information, including those relating to economic growth or industrial production. Where one
market has fewer frictions, implying information may be more quickly and completely
impounded into prices, that market should have stronger common information linkages with the
other markets. In the theoretical specification in equation (4.3), this implies linkages occur
through the first term. Our a priori expectation would be that the much greater depth and
liquidity of the Brent crude oil futures market would lead it to have the strongest linkages with
each of the other markets.
Looking at the overall strength of the linkages in Table 4.5, the average correlation of log
information flows across the three bivariate pairings is in fact highest for crude oil (69.6 per
cent) followed by coal (64.0 per cent), emission allowances (51.2 per cent) and then natural gas
(43.8 per cent). Although these results are for a limited sample of only four energy securities, it
does confirm our a priori expectations in that crude oil has the highest average correlation of log
information flows, which supports our Common Information Channel Hypothesis. That is, as
prices in the deep and liquid Brent crude oil futures market respond to information, volatility is
frequently observed in the other energy markets contemporaneously.
The strength of the linkages with the coal market reported in Table 4.5 is somewhat
surprising, given that it is the market most subject to frictions. Transactions in the main coal
futures contracts take place only every day or so, with reported daily settlement prices
frequently established from averages of bid and ask prices or indicative surveys of market
participants. Coal market illiquidity is a sizeable friction that would be expected to impair its
timely responsiveness to common information and any benefits from cross-market hedging. If,
in the absence of actual trade activity, coal futures settlement prices are determined by
surveying market participants, a tendency may exist to report indicative coal prices on the basis
of movements in other markets, and movements in the crude oil market may be seen as relevant
135
in this context. This could perhaps account for the very strong information linkages between the
coal and crude oil futures (82.5 per cent).
At the other end of the spectrum, natural gas has the weakest information linkages with the
other markets. This is unexpected given the results in Chapter 3 show that the natural gas and
crude oil markets are cointegrated between 2008 and 2011, albeit only weakly. However, the
daily sampling frequency employed in this paper, necessitated by the illiquidity of the coal
futures, may make these results incomparable with those in Chapter 3.
4.6 Conclusion
Much of the existing literature characterises interactions between emission allowances and
energy markets as a fuel switching relationship in which emission allowance returns are
positively (negatively) related to natural gas (coal) returns. However, we argue that these studies
incompletely characterise relationships between the markets of interest and ignore other
complexities in modelling. In this context, we describe a rational expectations model in the
tradition of Tauchen and Pitts (1983), Fleming et al. (1998) and Kodres and Pritsker (2002) in
which cross-market linkages are observable in the correlations of information flows
(volatilities). In this setting, volatility may occur in a number of markets simultaneously due to
common sensitivities to particular types of information or because the arrival of information
idiosyncratic to one market prompts a spillover of volatility into others. These spillovers occur
because of cross-market hedging demand or, more appropriately for an emission allowance and
energy market specification, because of economic linkages based on their relationships as
substitutes and complements. The model predicts a perfect correlation of volatilities in the
absence of market frictions such as trading costs, leverage constraints or illiquidity. Even in the
presence of frictions, the model predicts strong linkages because emission allowance and energy
securities are commonly sensitive to many types of information such as economic growth,
industrial production and the impact of unanticipated weather events.
136
We formulate two competing a priori expectations. Firstly, that the linkages between
emission allowances and the main fuel inputs to power generation (coal and natural gas) will be
stronger than the linkage between emission allowances and crude oil; a fuel whose combustion
occurs largely outside the EU ETS. Because emission allowances, coal and natural gas share
strong economic linkages as substitutes or complements, they should thus experience strong
volatility spillovers in the absence of market frictions. We call this the Spillover Channel
Hypothesis. Secondly, we postulate that, because of its greater depth and liquidity, the crude oil
market should impound information more quickly and completely, and thus should have
stronger linkages to the other markets. Where crude oil is less subject to the EU ETS, the
strength of its linkage to the emission allowance market will be predominantly on the basis of
common information. We call this the Common Information Channel Hypothesis. We employ
Fleming et al.‘s (1998) bivariate stochastic volatility representation of the rational expectations
model to estimate the strength of the cross-market linkages between emission allowances, coal,
natural gas and crude oil.
The results of estimating the Fleming et al. (1998) model using GMM show that, for
emission allowances, the strongest correlation of log information flows (volatilities) is with the
crude oil market (76.9 per cent). This is higher than for the other fuels, with correlations for
emission allowances and coal of 52.0 per cent and emission allowances and natural gas of only
24.6 per cent. These are much higher than the correlations between absolute returns and returns
squared, which are noisy proxies for volatility (and range between 3.3 to 22.7 per cent). These
results clearly contradict our expectations under the Spillover Channel Hypothesis and provide
support for our Common Information Channel Hypothesis.
137
4.7 Appendix
EU electricity markets have been increasingly liberalised since the introduction of the EU
Electricity Directive and subsequent legislation (see European Commission, 1996). Despite
some differences, these liberalised markets all broadly facilitate competition between wholesale
generators selling to electricity retailers, with transmission and distribution executed by a
regulated monopoly. A merit order for meeting expected power demand can be constructed by
lining up power generators from the one with the lowest marginal cost (price) in base-load
power generation to the one with the highest marginal cost108
. The merit order is used as a
marginal cost curve over which an impartial transmission system operator sequentially
dispatches the generating units109
. For a given amount of electricity demand, the offer price of
the marginal generation unit sets the electricity price. At any given time, a generator‘s offer
price will reflect fuel input costs, emission allowance prices, relative plant efficiencies,
operational and maintenance expenses and a profit margin. Ignoring operational and
maintenance costs, a rough approximation of generator profit is given by a coal generator‘s
clean dark spread ( CDS ) and a gas generator‘s clean spark spread ( CSS ):
gCOgge
cCOcce
IpppCSS
IpppCDS
2
2
(A1)
These spreads are the difference between the base-load electricity price ( ep ) received and
the costs of generation: fuel input prices for coal ( cp ) and natural gas ( gp ) scaled by the
thermal efficiency of typical coal ( c ) and natural gas ( g ) plants less emission allowance
costs (2COp ) adjusted for the emission intensity of typical coal ( cI ) and natural gas ( gI )
108
The marginal costs are in reference to base-load generation because the merit order is less meaningful
during peak periods when it is likely that all generation units are in use irrespective of their marginal costs
(see, for example, Delarue and D‘Haeseleer, 2007). Nuclear and hydro plants generally have the lowest
marginal costs, with coal, natural gas and oil-fired plants having progressively higher marginal costs.
109 The operator also keeps some capacity in reserve to meet unanticipated demand (spinning reserve).
Europe has 41 transmission system operators across 34 countries according to the European Network of
Transmission System Operators for Electricity (https://www.entsoe.eu/home/).
138
combustion110
. Equilibrating the clean dark and clean spark spreads and cancelling out the
electricity price allows us to form a rough estimate of the emission allowance price ( switchp ) that
will prompt a switch in the merit order from coal-fired to gas-fired generation:
cg
g
g
c
cswitch II
ppp
(A2)
A comparison of the actual European Union Allowance (EUA) price and the switching price
calculated as per equation (A2) using the Caisse des Dépôts efficiency and emission intensity
factors is displayed in Chart A1. According to this specification, noting the caveats regarding
the differing efficiency and emission intensity of individual plants together with the exclusion of
operational and maintenance costs, the actual emission allowance price was too low to support a
switch from coal-fired base-load generation to gas-fired generation on 69.3 per cent of days in
the 2008 to 2011 period.
The first problem with many prior studies arises because in practice switching prices
formulated as per equation (A2) are predominantly driven by the high volatility of natural gas
prices. In fact, between 2008 and 2011 returns for the two have a correlation of 89 per cent111
.
This compares with a return correlation between the switching price and the coal price of minus
5 per cent. Where studies use changes in these switching price variables in combination with
natural gas returns as independent variables for regression analysis the results will be affected
by a high degree of multicollinearity. In spite of this concern, much of the data-driven literature
undertakes analysis in this fashion112
.
110
Fuel prices are in euro per megawatt hour (€/MWh). Although they differ widely for individual power
plants, many studies use net thermal efficiency figures for conventional coal and gas-fired plants of
around 40 per cent and 55 per cent, respectively. Typical emission intensity factors are 86 per cent for
coal plant combustion and 36 per cent for natural gas combustion (see, for example, Caisse des Dépôts:
http://www.caissedesdepots.fr/).
111 Note that in calculating this return correlation, a constant of €10/tCO2e was added to the daily
switching price variable to make all observations positive such that returns could be calculated. This is
necessary because, as shown in Chart A1, the switching price can be negative (it was as low as
-€5.13/tCO2e at the end of August 2009).
112 These include: Alberola et al. (2008), Alberola et al. (2009), Bonacina et al. (2009), Keppler and
Mansanet-Bataller (2010), Bredin and Muckley (2011), Creti et al. (2011) and Mansanet-Bataller et al.
(2011). Mansanet-Bataller et al. (2007) construct a simple ratio of gas to coal price changes as a switching
price variable, which is also likely to induce a degree of multicollinearity.
139
Chart A1
Switching Price versus Actual Abatement Price
The EUA price (2
COp ) is the annual expiry ICE futures contract in €/tCO2e. The fuel switching price is calculated as per equation (A2)
using monthly expiry ICE Rotterdam coal futures (c
p ), month-ahead UK National Balancing Point natural gas (g
p ) and the Caisse
des Dépôts plant efficiency ( 40.0c
, 55.0g
) and emission intensity ( 86.0c
I , 36.0g
I ) figures.
-10
0
10
20
30
40
50
60
70
-10
0
10
20
30
40
50
60
70
ICE Annual EUA Futures Switching Price
€/tCO2e €/tCO2e
2008 2009 2010 2011
140
The second problem with analysing clean dark spreads, clean spark spreads and switching
points in the merit order is that these ignore other important variable costs, some of which are
endogenously determined contingent upon positioning in the merit order itself113
. Notably, the
contribution of operational and maintenance costs to a generator firm‘s total costs may alter the
strategic bidding behaviour of generating firms, particularly if those firms have diversified
portfolios of generating units (i.e. they own various different types of plant) or are otherwise able to
exert market power.
Extra operational costs are incurred as a result of the marginal generation unit having to be
cycled, or ramped-up and down to meet actual electricity demand conditions. Ramping up a plant
requires greater fuel use than simply running a plant already close to maximum capacity, with the
cost level dependent on whether the plant is starting cold or is being restarted after recent
operation114
. In addition, frequent cycling inevitably leads to higher maintenance costs due to the
impact of temperature and pressure changes on the plant itself. These costs are generally much
higher and the process more time consuming for coal-fired generators, particularly older plants
which were originally designed as base-load generators, compared with modern gas-fired plants
built for cycling. Where an electricity firm possesses a variety of different types of generating units,
as is common among the large power utilities in the EU, keeping plants with lower cycling costs on
the margin, like gas turbines and hydro plants, either by submitting higher prices for the output of
these units or by simply holding them in reserve, can reduce the firm‘s total costs (see Denny and
113
We note that some studies include operational and maintenance costs in their formulation of dark and
spark spreads and short-run marginal cost functions, such as Laurikka and Koljonen (2006) and Sijm, Bakker,
Chen, Harmsen and Lise, (2005), however, many do not (those listed in the previous footnote for example).
114 In a case study of the Irish power generation sector, Denny and O‘Malley (2009) examine the effect of the
EU ETS on cycling costs. They estimate the ramp-up costs to cold start a 285MW Irish coal plant would be:
€32,164 in coal, assuming a price of €2.20/GJ (equivalent to €7.92/MWh or €64.46/t); and, €39,000 in
emission allowances, assuming ramp-up requires 14,620GJ of energy producing 1,300t of CO2 priced at
€30/tCO2e. On the other hand, Rosnes (2008), who notes that heavy fuel oil is often used to start thermal
power plants, estimates the range in start up costs for a 400MW plant to be between €1,330 and €8,662 for hot
and cold starts, respectively (lower input costs are used in forming these estimates: coal at €50/t, heavy fuel
oil at €203/t and emission allowances at only €5.40/tCO2e).
141
O‘Malley, 2009)115
. These faster starting plants may also be kept in reserve in order to maximise
profits from unanticipated spikes in power demand. In either of these circumstances, the
characterisation of emission allowance prices being driven by relative energy input prices due to
fuel switching is likely not evident or, at best, substantially distorted116
. Similar arguments apply
when a system has a number of plants of a type in which producing electricity is a secondary
consideration, such as waste incineration plants and combined heat and power (CHP) plants. The
demand for emission allowances of these ‗must-run‘ facilities may not be influenced by changes in
fuel input prices in the way predicted by the fuel switching literature117
.
The last problem with modelling a relationship between emission allowance and energy prices
using a theoretical switching point is that fuel switching is irrelevant during peak generation
periods, which constitute significant portions of most weekdays. During peak periods most of an
electricity system‘s installed capacity will be running irrespective of relative fuel prices due to the
inelasticity of short-term power demand (see Delarue and D‘Haeseleer, 2007). Even in off-peak
periods, limits in the diversity of the installed generation mix may prevent fuel switching from
being a frequent or meaningful form of abatement, let alone an activity detectable in the short-term
interactions between emission allowance and energy prices. If switching fuel is indeed the marginal
form of abatement for the power sector, this will be more evident in the long-run capital
expenditure decisions of firms in which there is a gradual migration from building coal-fired plants
115
Many European electricity markets are somewhat oligopolistic, with a few generating firms possessing a
high degree of horizontal and/or vertical integration. For example, Scheepers, Wals and Rijkers (2003) give a
description of the concentration of power producers in North-West European electricity markets, showing
particularly high concentration in France and Belgium compared to Germany and the Netherlands. Similar
results are obtained by Percebois (2008) whose Hirschmann-Herfindahl Index calculations show high
concentration in France and Belgium and low concentration in Germany and the UK (with Italy and Spain in
between). Moreover, Percebois (2008) shows that 9 companies accounted for 83 per cent of electricity sales in
the EU-15 in 2006.
116 In fact, if fuel switching were to take place as hypothesised, it may also frustrate the system‘s purpose of
reducing emissions because, if coal-fired plants are forced to cycle as the marginal generation unit, the greater
combustion of fuel during the ramp-up phase may increase total emissions (see Denny and O‘Malley, 2009).
Rosnes (2008) forms similar conclusions based on the interaction of renewable energy targets and emissions
trading where the intermittence of wind generation leads to greater cycling by fossil fuel based generators.
117 Regulatory policies, such as renewable energy targets and feed in tariffs, may similarly distort the merit
order and obfuscate expected fuel switching relationships based on emission allowance and energy prices
alone. In fact, these policies are often seen as undercutting the effectiveness of the EU ETS (Blyth, Bunn,
Kettunen and Wilson, 2009, and Blyth and Bunn, 2011, model the effects of various policy scenarios).
142
towards gas-fired or renewable energy facilities. This is only likely to be observed over a span of
decades118
.
These theoretical and practical problems in the existing literature warrant the alternative
characterisation of emission allowance and energy market interactions presented in this paper; one
that better reflects the role of information in the price formation process.
118
Analysing long-run marginal costs, Sijm et al. (2005) show that new combined-cycle gas turbines are
competitive with new coal-fired installations even in the absence of the EU ETS, indicating that, although this
gradual migration may be supported by high emission allowance prices, fuel switching in the form of long-
term capacity changes may be underway regardless. They also find that, in addition to long-run marginal costs
favouring gas-fired plants as a replacement for facilities at the end of their lives, if emission allowance prices
were to be sustained above €23.20/tCO2e, there would also be an incentive to build new combined-cycle gas
turbines in preference to the continued use of many existing coal plants. However, they note this estimate is
very sensitive to cost assumptions and regional differences.
143
CHAPTER 5: Conclusion
Emissions trading is an attempt to redress the market failures associated with greenhouse gas
pollution. As the world‘s largest emission trading system, the EU ETS is expected to have a large
impact upon Europe‘s energy generation in the coming decades. As such, it is very important for
policy makers, market participants and academics to understand both the dynamics of this relatively
new market and its interaction with Europe‘s energy markets. In this context, we contribute to the
literature in this area by studying price discovery in the EU ETS and its catalysts, price discovery in
the main fossil fuel energy markets and information linkages between emission allowance and
energy securities.
We commence our investigation by studying price discovery in the EU ETS and, in doing so,
we provide the first evidence on its catalysts. In particular, we consider the impact of market
frictions that inhibit this price discovery process, namely trading cost and leverage, as well as
market segmentation between EUAs and CERs. In line with much of the literature concerning other
markets, trading costs are found to be the most important determinant of which securities are traded.
Interestingly, the results concerning CER futures indicate an absence of market segmentation, given
its contribution to price discovery in comparison to its small share of trade volume. This is an
144
indication that the substantial discount at which CERs trade relative to EUAs is adequate
compensation for their associated risks, with their speculative use unimpeded.
Next we examine price discovery in the main European fossil fuel energy markets, namely coal,
natural gas and crude oil spanning both the physical and financial layers of these markets. Despite
the lack of liquidity and transparency in coal transactions, the futures traded on the Intercontinental
Exchange are shown to be marginally better sources of short and long-run price discovery.
Similarly, UK natural gas futures traded on the Intercontinental Exchange display greater price
discovery than physical trading at the major natural gas hubs in North-West Europe, particularly in
their contribution to long-run equilibrium. In addition, there is evidence that short-run interactions
are stronger at similar points on the natural gas forward curve than interactions between securities
specific to natural gas hub locations. This is likely related to the inelasticity of demand for short-
dated natural gas in the presence of extreme (cold) weather, which can induce high volatility at the
front of the curve felt commonly across hubs. Thus, variations in longer dated prices reflect fewer
transitory changes and contribute more to long-run equilibrium between natural gas security prices,
though we note that there are still linkages between these securities and the crude oil market. In the
European crude oil market, we find that the Brent futures contract leads the price discovery process,
though there is some evidence that trading in the physical layers—specifically exchange-for-
physicals—can, in turn, impact upon the futures market. In comparing global benchmarks, we find
only weak evidence that Brent and WTI futures remain cointegrated, with their relationship
deteriorating further towards the end of our sample. To the extent that there is any price leadership
between them, both short and long-run analysis reveals that price discovery more often resides with
WTI than with Brent futures, despite growing volumes of Brent futures trading and the recent
dislocations at WTI‘s pricing point in Cushing, Oklahoma.
Having established where price discovery is taking place in the European emission allowance
and fossil fuel energy markets, we analyse the interactions between them. Given the miss-
characterisation of the relationship in prior research, we describe a rational expectations model in
which volatility may occur in a number of markets simultaneously due to common sensitivities to
145
particular types of information or from the spillover of information that is idiosyncratic to one
market into others. The results show that, despite the strong economic linkages between emission
allowances and coal and natural gas that are supposed on the basis of the potential for fuel
switching, emission allowances have the strongest information linkages to the crude oil market.
This relationship is likely a product of strong common information linkages. Overall, our results not
only reinforce the importance of information in the determination of security prices, but also serve
as a reminder that fuel switching in the presence of emissions trading is a long-term process
affecting an economy‘s energy generation mix over decades.
146
REFERENCES
[1] Adelman, M. A., 1984, International oil agreements, The Energy Journal 5, 1-9.
[2] Adelman, M. A., 1992, Is the world oil market 'one great pool'?-comment, The Energy Journal 13, 157-158.
[3] Alberola, Emilie, Julien Chevallier and Benoît Chèze, 2008, Price drivers and structural breaks
in European carbon prices 2005-2007, Energy Policy 36, 787-797.
[4] Alberola, Emilie, Julien Chevallier and Benoît Chèze, 2009, Emissions compliances and carbon
prices under the EU ETS: A country specific analysis of industrial sectors, Journal of Policy Modeling 31, 446-462.
[5] Asche, Frank, Petter Osmundsen and Maria Sandsmark, 2006, The UK market for natural gas,
oil and electricity: Are the prices decoupled?, The Energy Journal 27, 27-40.
[6] Asche, Frank, Petter Osmundsen and Ragnar Tveterås, 2002, European market integration for
gas? Volume flexibility and political risk, Energy Economics 24, 249-265.
[7] Bachmeier, Lance J. and James M. Griffin, 2006, Testing for market integration: Crude oil,
coal, and natural gas, The Energy Journal 27, 55-71.
[8] Baillie, Richard T., G. Geoffrey Booth, Yiuman Tse and Tatyana Zabotina, 2002, Price
discovery and common factor models, Journal of Financial Markets 5, 309-321.
[9] Barret, Christophe, 2012, Brent prices: Impact of PRA methodology on price formation, The
Oxford Institute for Energy Studies.
[10] Bekiros, Stelios D. and Cees G. H. Diks, 2008, The relationship between crude oil spot and
futures prices: Cointegration, linear and nonlinear causality, Energy Economics 30, 2673-2685.
[11] Bentzen, Jan, 2007, Does OPEC influence crude oil prices? Testing co-movements and
causality between regional crude oil prices, Applied Economics 39, 1375-1385.
[12] Benz, Eva and Jördis Hengelbrock, 2008, Liquidity and price discovery in the European CO2
futures market: An intraday analysis, Working Paper (Bonn Graduate School of Economics).
[13] Bessembinder, Hendrik, 1992, Systematic risk, hedging pressure, and risk premiums in futures
markets, Review of Financial Studies 5, 637-667.
[14] Blas, Javier, 2012a, Physical oil traders focus on North Sea, The Financial Times (London).
[15] Blas, Javier, 2012b, WTI loses ground in physical oil market, The Financial Times (London).
[16] Blume, Marshall E. and Robert F. Stambaugh, 1983, Biases in computed returns: An application
to the size effect, Journal of Financial Economics 12, 387-404.
[17] Blyth, William and Derek Bunn, 2011, Coevolution of policy, market and technical price risks
in the EU ETS, Energy Policy 39, 4578-4593.
[18] Blyth, William, Derek Bunn, Janne Kettunen and Tom Wilson, 2009, Policy interactions, risk
and price formation in carbon markets, Energy Policy 37, 5192-5207.
[19] Bonacina, Monica, Anna Creti and Simone Cozialpi, 2009, The European carbon market in the
financial turmoil: Some empirics in early Phase II, Working Paper (Bocconi University).
[20] Booth, G. Geoffrey, Raymond W. So and Yiuman Tse, 1999, Price discovery in the German
equity index derivatives markets, Journal of Futures Markets 19, 619-643.
[21] Box, G. E. P. and David A Pierce, 1970, Distribution of residual autocorrelations in
autoregressive-integrated moving average time series models, Journal of the American
Statistical Association 65, 1509-1526.
[22] Brandt, Michael W. and Kenneth A. Kavajecz, 2004, Price discovery in the U.S. Treasury
market: The impact of orderflow and liquidity on the yield curve, Journal of Finance 59, 2623-
2654.
[23] Bredin, Don and Cal Muckley, 2011, An emerging equilibrium in the EU emissions trading
scheme, Energy Economics 33, 353-362.
[24] Brunetti, Celso and Christopher L. Gilbert, 2000, Bivariate FIGARCH and fractional
cointegration, Journal of Empirical Finance 7, 509-530.
[25] Carchano, Óscar and Ángel Pardo, 2009, Rolling over stock index futures contracts, Journal of Futures Markets 29, 684-694.
147
[26] Chan, Kalok, 1992, A further analysis of the lead-lag relationship between the cash market and
stock index futures market, Review of Financial Studies 5, 123-152.
[27] Chevallier, Julien, 2010a, EUAs and CERs: vector autoregression, impulse response function
and cointegration analysis, Economics Bulletin 30, 558-576.
[28] Chevallier, Julien, 2010b, A note on cointegrating and vector autoregressive relationships
between CO2 allowances spot and futures prices, Economics Bulletin 30, 1564-1584.
[29] Chevallier, Julien, 2012. Econometric Analysis of Carbon Markets (Springer).
[30] Coase, Ronald H., 1960, The problem of social cost, The Journal of Law & Economics 3, 1-44.
[31] Copeland, Thomas E. and Dan Galai, 1983, Information effects on the bid-ask spread, Journal
of Finance 38, 1457-1469.
[32] Creti, Anna, Pierre-André Jouvet and Valérie Mignon, 2011, Carbon price drivers: Phase I
versus Phase II equilibrium, Working Paper (CEPII).
[33] Delarue, Erik D. and William D. D'haeseleer, 2007, Price determination of ETS allowances
through the switching level of coal and gas in the power sector, International Journal of Energy
Research 31, 1001-1015.
[34] Demsetz, Harold, 1968, The cost of transacting, The Quarterly Journal of Economics 82, 33-53.
[35] Denny, Eleanor and Mark O'Malley, 2009, The impact of carbon prices on generation-cycling
costs, Energy Policy 37, 1204-1212.
[36] Dickey, David A. and Wayne A. Fuller, 1979, Distribution of the estimators for autoregressive
time series with a unit root, Journal of the American Statistical Association 74, 427-431.
[37] The Economist, 2010, The wrong sort of recycling, The Economist. [38] The Economist, 2011, Green fleeces, red faces, The Economist.
[39] European Commission, 1996, Directive 96/92/EC of the European Parliament and of the
Council of 19 December 1996 concerning common rules for the internal market in electricity,
Official Journal of the European Communities L27, 20-29.
[40] European Commission, 2004, Directive 2004/101/EC of the European Parliament and of the
Council of 27 October 2004 amending directive 2003/87/EC establishing a scheme for
greenhouse gas emission allowance trading within the community, in respect of the Kyoto
Protocol's project mechanisms Official Journal of the European Union 47, 18-23.
[41] European Commission, 2008, Questions and answers on the revised EU Emissions Trading
System, Press Release 08/796 (Brussels).
[42] European Commission, 2010, Accompanying document to the Commission decision on
applying use restrictions on international credits (from HFC-23 and N20 projects) pursuant to
Article 11a(9) of Directive 2009/29/EC, European Commission (Brussels).
[43] European Commission, 2011, Europe's energy position: 2010 annual report, the Market
Observatory for Energy (Publications Office of the European Union, Luxembourg).
[44] European Environment Agency, 2011, Annual European Union greenhouse gas inventory 1990-
2009 and inventory report 2011: Submission to the UNFCCC Secretariat (Publications Office of
the European Union, Luxembourg).
[45] Fama, Eugene F. and Kenneth R. French, 1988, Permanent and temporary components of stock
prices, Journal of Political Economy 96, 246-273.
[46] Fattouh, Bassam, 2011, An anatomy of the crude oil pricing system, The Oxford Institute for Energy Studies.
[47] Fezzi, Carlo and Derek W. Bunn, 2009, Structural interactions of European carbon trading and
energy prices, The Journal of Energy Markets 2, 53-69.
[48] Fleischer, Petra, 2003, Volatility and information linkages across markets and countries,
Australian Journal of Management 28, 251-272.
[49] Fleming, Jeff, Chris Kirby, and Barbara Ostdiek, 1998, Information and volatility linkages in
the stock, bond, and money markets, Journal of Financial Economics 49, 111-137.
[50] Fleming, Jeff, Barbara Ostdiek, and Robert E. Whaley, 1996, Trading costs and the relative
rates of price discovery in stock, futures, and option markets, Journal of Futures Markets 16,
353-387.
[51] Fleming, Michael J. and Eli M. Remolona, 1999, Price formation and liquidity in the U.S.
Treasury market: The response to public information, Journal of Finance 54, 1901-1915.
[52] Fuller, W. A., 1996. Introduction to Statistical Time Series (Wiley, New York).
148
[53] Garbade, Kenneth D. and William L. Silber, 1983, Price movements and price discovery in
futures and cash markets, The Review of Economics and Statistics 65, 289-297.
[54] Glosten, Lawrence R. and Paul R. Milgrom, 1985, Bid, ask and transaction prices in a specialist
market with heterogeneously informed traders, Journal of Financial Economics 14, 71-100.
[55] Gonzalo, Jesús and Clive W. J. Granger, 1995, Estimation of common long-memory
components in cointegrated systems, Journal of Business and Economic Statistics 13, 27-36.
[56] Gülen, S. Gürcan, 1997, Regionalization in the world crude oil market, The Energy Journal 18,
109-126.
[57] Gülen, S. Gürcan, 1999, Regionalization in the world crude oil market: Further evidence, The
Energy Journal 20, 125-139.
[58] Hammoudeh, Shawkat M., Bradley T. Ewing and Mark A. Thompson, 2008, Threshold
cointegration analysis of crude oil benchmarks, The Energy Journal 29, 79-95.
[59] Hansen, Lars Peter, 1982, Large sample properties of generalized method of moments
estimators, Econometrica 50, 1029-1054.
[60] Harris, Frederick H. deB, McInish, Thomas H., Shoesmith, Gary L., and Robert A. Wood, 1995,
Cointegration, error correction, and price discovery on informationally linked security markets,
Journal of Financial and Quantitative Analysis 30, 563-579.
[61] Hasbrouck, Joel, 1995, One security, many markets: Determining the contributions to price
discovery, Journal of Finance 50, 1175-1199.
[62] Hasbrouck, Joel, 2002, Stalking the "efficient price" in market microstructure specifications: An
overview, Journal of Financial Markets 5, 329-339.
[63] Hepburn, Cameron, 2006, Regulation by prices, quantities or both: A review of instrument
choice, Oxford Review of Economic Policy 22, 226-247.
[64] Hintermann, Beat, 2010, Allowance price drivers in the first phase of the EU ETS, Journal of Environmental Economics and Management 59, 43-56.
[65] Hsieh, Wen-Liang G., Chin-Shen Lee and Shu-Fang Yuan, 2008, Price discovery in the options
markets: An application of put-call parity, Journal of Futures Markets 28, 354-375.
[66] IOSCO, 2012, IOSCO consults on oil price reporting agency oversight, International
Organization of Securities Commissions, Media Release 04/2012.
[67] Johansen, Søren, 1995. Likelihood-based inference in cointegrated vector autoregressive
models (Oxford University Press, Oxford).
[68] Kao, Chung-Wei and Jer-Yuh Wan, 2009, Information transmission and market interactions
across the Atlantic—An empirical study on the natural gas market, Energy Economics 31, 152-
161.
[69] Kao, Chung-Wei and Jer-Yuh Wan, 2012, Price discount, inventories and the distortion of WTI
benchmark, Energy Economics 34, 117-124.
[70] Kaufmann, Robert K. and Ben Ullman, 2009, Oil prices, speculation, and fundamentals:
Interpreting causal relations among spot and futures prices, Energy Economics 31, 550-558.
[71] Keppler, Jan Horst and Maria Mansanet-Bataller, 2010, Causalities between CO2, electricity,
and other energy variables during phase I and phase II of the EU ETS, Energy Policy 38, 3329-
3341.
[72] Kodres, Laura E. and Matthew Pritsker, 2002, A rational expectations model of financial
contagion, Journal of Finance 62, 769-799.
[73] Laurikka, Harri and Tiina Koljonen, 2006, Emissions trading and investment decisions in the
power sector—a case study in Finland, Energy Policy 34, 1063-1074.
[74] Lin, Sharon Xiaowen and Michael N. Tamvakis, 2001, Spillover effects in energy futures
markets, Energy Economics 23, 43-56.
[75] Ljung, G. M. and G. E. P. Box, 1978, On a measure of lack of fit in time series models,
Biometrika 65, 297-303.
[76] Lo, Andrew W. and A. Craig MacKinlay, 1988, Stock market prices do not follow random
walks: Evidence from a simple specification test, Review of Financial Studies 1, 41-66.
[77] Mansanet-Bataller, Maria and Ángel Pardo, 2009, Impacts of regulatory announcements on CO2
prices, The Journal of Energy Markets 2, 75-107.
[78] Mansanet-Bataller, Maria, Angel Pardo and Enric Valor, 2007, CO2 prices, energy and weather,
The Energy Journal 28, 73-92.
149
[79] Mansanet-Bataller, Maria, Julien Chevallier, Morgan Hervé-Mignucci and Emilie Alberola,
2011, EUA and sCER Phase II price drivers: Unveiling the reasons for the existence of the
EUA-sCER spread, Energy Policy 39, 1056-1069.
[80] Mazighi, Ahmed El Hachemi, 2005, Henry hub and national balancing point prices: What will
be the international gas price reference?, OPEC Review 29, 219-230.
[81] Melling, Anthony J., 2010, Natural gas pricing and its future: Europe as the battleground,
Carnegie Endowment for International Peace (Washington DC).
[82] Mizrach, Bruce, 2010, Integration of the global carbon markets, Working Paper (Rutgers
University).
[83] Mizrach, Bruce and Yoichi Otsubo, 2011, The market microstructure of the European Climate
Exchange, Working Paper (Rutgers University).
[84] Montepeque, Jorge, 2012, Oil price benchmarks in international trade, Oxford Energy Forum
87, 3-6.
[85] Moosa, Imad A., 2002, Price discovery and risk transfer in the crude oil futures market: Some
structural time series evidence, Economic Notes 31, 155-165.
[86] Neuman, Anne, Boriss Siliverstovs and Christian von Hirschhausen, 2006, Convergence of
European spot market prices for natural gas? A real-time analysis of market integration using
the Kalman filter, Applied Economics Letters 13, 727-732.
[87] Newey, Whitney K. and Kenneth D. West, 1987, A simple, positive semi-definite,
heteroskedasticity and autocorrelation consistent covariance matrix, Econometrica 55, 703-708.
[88] Niederhoffer, Victor and M. F. M. Osborne, 1966, Market making and reversal on the stock
exchange, Journal of the American Statistical Association 61, 897-916.
[89] Panagiotidis, Theodore and Emilie Rutledge, 2007, Oil and gas markets in the UK: Evidence
from a cointegrating approach, Energy Economics 29, 329-347.
[90] Percebois, Jacques, 2008, Electricity liberalization in the European Union: Balancing benefits
and risks, The Energy Journal 29, 1-19.
[91] Phillips, P. C. B. and P. Perron, 1988, Testing for a unit root in time series regression,
Biometrika 75, 335-346.
[92] Platts, 2012, Methodology and specifications guide: Crude oil, updated January 2012 (McGraw
Hill Companies).
[93] Richardson, Matthew and Tom Smith, 1994, A unified approach to testing for serial correlation
in stock returns, Journal of Business 67, 371-399.
[94] Rittler, Daniel, 2009, Price discovery, causality and volatility spillovers in European Union
Allowances Phase II: A high frequency analysis, Discussion Paper Series (University of
Heidelberg).
[95] Roll, Richard, 1984, A simple measure of the effective bid-ask spread in an efficient market,
Journal of Finance 39, 1127-1139.
[96] Rosnes, Orvika, 2008, The impact of climate policies on the operation of a thermal power plant,
The Energy Journal 29, 1-22.
[97] Ross, Stephen A., 1989, Information and volatility: The no-arbitrage martingale approach to
timing and resolution, Journal of Finance 44, 1-17.
[98] Scheepers, M.J.J., A.F. Wals and F.A.M. Rijkers, 2003, Position of large power producers in
electricity markets of North Western Europe, Report for the Dutch Energy Council on the
electricity markets in Belgium, France, Germany and The Netherlands, Energy Research Centre
of the Netherlands (Amsterdam).
[99] Schwarz, Thomas V. and Andrew C. Szakmary, 1994, Price discovery in petroleum markets:
Arbitrage, cointegration, and the time interval of analysis, Journal of Futures Markets 14, 147-
167.
[100] Sen, Amrita, 2012, Oil benchmarks in international trade, Oxford Energy Forum 87, 19-22.
[101] Sijm, J.P.M, S.J.A. Bakker, Y. Chen, H.W. Harmsen and W. Lise, 2005, CO2 price dynamics:
The implications of EU emissions trading for the price of electricity, Energy Research Centre of
the Netherlands (Amsterdam).
[102] Siliverstovs, Boriss, Guillaume L'Hégaret, Anne Neumann and Christian von Hirschhausen,
2005, International market integration for natural gas? A cointegration analysis of prices in
Europe, North America and Japan, Energy Economics 27, 603-615.
150
[103] Silvapulle, Param and Imad A. Moosa, 1999, The relationship between spot and futures prices:
Evidence from the crude oil market, Journal of Futures Markets 19, 175-193.
[104] Stern, Nicholas, 2006, The economics of climate change: The Stern Review, HM Treasury
(Cambridge University Press).
[105] Stoll, Hans R. and Robert E. Whaley, 1990, The dynamics of stock index and stock index
futures returns, Journal of Financial and Quantitative Analysis 25, 441-468.
[106] Stoll, Hans R. and Robert E. Whaley, 2010, Commodity index investing and commodity futures
prices, Journal of Applied Finance 20, 7-46.
[107] Treepongkaruna, Sirimon, Robert Brooks and Stephen Gray, 2012, Do trading hours affect
volatility links in the foreign exchange market?, Australian Journal of Management 37, 7-27.
[108] Treepongkaruna, Sirimon and Stephen Gray, 2009, Information and volatility links in the
foreign exchange market, Accounting and Finance 49, 385-405.
[109] Tauchen, George E. and Mark Pitts, 1983, The price variability-volume relationship on
speculative markets, Econometrica 51, 485-506.
[110] Uhrig-Homburg, Marliese and Michael Wagner, 2009, Futures price dynamics of CO2 emission
allowances - An empirical analysis of the trial period, Working Paper (University of Karlsruhe).
[111] Weiner, Robert J., 1991, Is the world oil market 'one great pool'?, The Energy Journal 12, 95-
107. [112] Wooldridge, Jeffery M., 2009. Introductory Econometrics: A Modern Approach (Cengage
Learning, Mason).
[113] World Bank, 2010, State and trends of the carbon market 2010, Carbon Finance at the World
Bank (Washington DC).
[114] World Bank, 2011, State and trends of the carbon market 2011, Carbon Finance at the World
Bank (Washington DC).
