Sovereign Credit Risk and Contagion

Sovereign Credit Risk

and Contagion

Inaugural dissertation submitted in fulfillment of the requirements for the degree

of Doctor rerum oeconomicarum at the Faculty of Business, Economics and

Social Sciences of the University of Bern.

Submitted by

Georg Felix Brill

from Germany

2011

Original document saved on the web server of the University Library of Bern

This work is licensed under a

Creative Commons Attribution-Non-Commercial-No derivative works 2.5 Switzerland licence. To see the licence go to

http://creativecommons.org/licenses/by-nc-nd/2.5/ch/ or write to Creative Commons, 171 Second Street, Suite 300, San Francisco, California 94105,

USA. !

!

Copyright Notice

This document is licensed under the Creative Commons Attribution-Non-Commercial-No derivative works 2.5 Switzerland. http://creativecommons.org/licenses/by-nc-nd/2.5/ch/ !!!You are free:

to copy, distribute, display, and perform the work Under the following conditions:

Attribution. You must give the original author credit.

Non-Commercial. You may not use this work for commercial purposes.

No derivative works. You may not alter, transform, or build upon this work.. For any reuse or distribution, you must take clear to others the license terms of this work. Any of these conditions can be waived if you get permission from the copyright holder. Nothing in this license impairs or restricts the author’s moral rights according to Swiss law. The detailed license agreement can be found at: http://creativecommons.org/licenses/by-nc-nd/2.5/ch/legalcode.de !

!

!

The faculty accepted this work as dissertation on 20 October 2011 at the request

of the two advisors Prof. Dr. Klaus Neusser and Prof. Dr. Monika Bütler, without

wishing to take a position on the view presented therein.

Abstract

In the first chapter, the empirical relationship between CDSpremia and government bond spreads is examined for Portugal,Italy, Ireland, Greece, and Spain. The analysis yields some evi-dence of a long-term relationship between the two markets inthe sense of cointegration. In most cases, only CDS premia con-tribute to the price discovery process. In the other instances,both markets contribute more or less equally. This suggests thatbond spreads react only sluggishly to long-term imbalances, asmeasured by the cointegrating relationship, behaviour that maybe due – at least partially – to liquidity effects.

In the second chapter, a rolling-crisis-window approach for con-tagion testing is applied, derived from and enhancing an ap-proach proposed by Forbes and Rigobon (2002). The rolling-crisis-window approach helps account for crises of longer-than-usual duration, as is case for Greece since its crisis began in Octo-ber 2009. This rolling-crisis-window approach is applied to testwhether the co-movements of sovereign CDS premia increasedsignificantly after the Greek debt crisis started. The sample con-sists of daily data between October 2008 and July 2010 for 39countries from both emerging and industrialized countries. Thetest results indicate that there were periods of contagion for CDSmarkets during the Greek debt crisis, which contrasts with theresults of Forbes and Rigobon (2002) for equity markets duringthe East Asian crisis in 1997-98, the Mexican peso crisis in 1994,and the U.S. stock market crash in 1987, challenging their con-clusion of “no contagion, only interdependence.”

In the third chapter, the rolling-crisis-window approach is ap-plied to equity markets during these three crises and the re-sults are compared to those of Forbes and Rigobon (2002). Thesample consists of daily returns of 32 MSCI equity market in-dices in both local currencies and US dollars. In contrast to thestatic approach of Forbes and Rigobon (2002), the rolling-crisis-window approach yields ample evidence of contagion during thesecrises. This result is further supported by extensive robustnesstests that entailed altering the periods of relative stability andusing daily returns in US dollars instead of the local currency.

Contents

List of Figures vi

List of Tables viii

Abbreviations x

Acknowledgements xii

Preface 1

1 Did the CDS Market Push Up Risk Premiafor Sovereign Credit? 51.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.2 Introduction to the CDS Market . . . . . . . . . . . . . . . . 8

1.2.1 CDS in General . . . . . . . . . . . . . . . . . . . . . . 81.2.2 Special Features of the Sovereign CDS Market . . . . 91.2.3 History of the Sovereign CDS Market . . . . . . . . . 101.2.4 The Role of CDS in Financial Markets . . . . . . . . . 101.2.5 Outstanding Volumes of CDS Contracts . . . . . . . . 11

1.3 Empirical Analysis . . . . . . . . . . . . . . . . . . . . . . . . 141.3.1 Data Description . . . . . . . . . . . . . . . . . . . . . 161.3.2 Basic Analysis . . . . . . . . . . . . . . . . . . . . . . 181.3.3 Long-Run Relations . . . . . . . . . . . . . . . . . . . 201.3.4 Price Discovery . . . . . . . . . . . . . . . . . . . . . . 221.3.5 Liquidity Analysis . . . . . . . . . . . . . . . . . . . . 27

1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

CONTENTS v

2 Measuring Co-Movements of CDS PremiaDuring the Greek Debt Crisis 312.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312.2 Propagation of Shocks: Contagion vs. Interdependence . . . . 35

2.2.1 Theory and Literature Review . . . . . . . . . . . . . 362.2.2 A Bivariate Test of Contagion . . . . . . . . . . . . . . 39

2.3 Contagion during the Greek Debt Crisis . . . . . . . . . . . . 422.3.1 Between Countries . . . . . . . . . . . . . . . . . . . . 432.3.2 Between Regions . . . . . . . . . . . . . . . . . . . . . 50

2.4 Exploring the Common Factor . . . . . . . . . . . . . . . . . 542.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

3 Testing for Contagion with aRolling-Crisis-Window Approach 633.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633.2 The Rolling-Crisis-Window Approach . . . . . . . . . . . . . 663.3 Contagion During the East Asian Crisis . . . . . . . . . . . . 71

3.3.1 Contagion Stemming from Hong Kong . . . . . . . . . 733.3.2 Contagion Stemming from Thailand . . . . . . . . . . 773.3.3 Contagion Stemming from Indonesia . . . . . . . . . . 783.3.4 Contagion Stemming from Korea . . . . . . . . . . . . 82

3.4 Contagion During the Mexican Peso Crisis . . . . . . . . . . . 843.5 Contagion During the U.S. Stock Market Crash . . . . . . . . 893.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

A Approximate Critical Values for the Rolling FR-Tests 95

B Robustness Tests 99

References 117

List of Figures

1.1 Estimated Debt-to-GDP Ratio for 2010 in Percent . . . . . . 61.2 CDS Premia for Portugal, Italy, Ireland, Greece, and Spain . 71.3 Outstanding Gross Notional Volumes by Entities . . . . . . . 121.4 Outstanding Net Notional Volumes by Countries . . . . . . . 131.5 Government Bond Spreads and CDS Premia . . . . . . . . . . 181.6 Bid-Ask Spreads . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.1 Debt-to-GDP Ratio since 1950 . . . . . . . . . . . . . . . . . 322.2 CDS Premia in Basis Points . . . . . . . . . . . . . . . . . . . 352.3 Greek CDS Premia and Key Events During the Greek Debt

Crisis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452.4 Illustration of the Rolling FR-Test Approach . . . . . . . . . 462.5 Principal Component Analysis for PIGS Countries . . . . . . 57

3.1 CDS Premia in Basis Points . . . . . . . . . . . . . . . . . . . 643.2 Illustration of the Rolling Contagion Test Approach . . . . . 713.3 Stock Market Indices During the East Asian Crisis . . . . . . 723.4 Contagion Stemming from Hong Kong: Signals for Philippines 763.5 Contagion Stemming from Thailand: Signals for Malaysia . . 783.6 Contagion Stemming from Indonesia: Signals for Korea . . . 823.7 Contagion Stemming from Korea: Signals for China . . . . . 843.8 Stock Market Indices During the Mexican Peso Crisis . . . . 853.9 Contagion Stemming from Mexico: Signals for Argentina . . 883.10 Stock Market Indices During the U.S. Market Crash . . . . . 893.11 Contagion Stemming from the U.S.: Signals for Germany . . 91

A.1 Rolling Correlation Coefficient . . . . . . . . . . . . . . . . . 97A.2 Approximate Critical Value for α = 0.05 . . . . . . . . . . . . 97

List of Tables

1.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . 171.2 Correlation Coefficients . . . . . . . . . . . . . . . . . . . . . 191.3 Johansen Trace Test Statistics . . . . . . . . . . . . . . . . . 211.4 Contributions to Price Discovery . . . . . . . . . . . . . . . . 241.5 Granger Causality Test Results . . . . . . . . . . . . . . . . . 26

2.1 Descriptive Statistics of CDS Premia . . . . . . . . . . . . . . 442.2 Forbes and Rigobon Tests . . . . . . . . . . . . . . . . . . . . 482.3 Definition of Regional Aggregates . . . . . . . . . . . . . . . . 522.4 Correlation Coefficients and Contagion Signals for Regional

Aggregates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 532.5 Regional Aggregates - Principal Component Analysis . . . . . 56

3.1 East Asian Crisis: Contagion Stemming from Hong Kong . . 753.2 East Asian Crisis: Contagion Stemming from Thailand . . . . 793.3 East Asian Crisis: Contagion Stemming from Indonesia . . . 813.4 East Asian Crisis: Contagion Stemming from Korea . . . . . 833.5 Mexican Peso Crisis: Contagion Stemming from Mexico . . . 873.6 U.S. Stock Market Crash: Contagion Stemming from the U.S. 90

A.1 Approximate Critical Values for the Rolling FR-Tests . . . . 98

B.1 East Asian Crisis: Contagion Stemming from Hong Kong . . 100B.2 East Asian Crisis: Contagion Stemming from Hong Kong . . 101B.3 East Asian Crisis: Contagion Stemming from Thailand . . . . 102B.4 East Asian Crisis: Contagion Stemming from Thailand . . . . 103B.5 East Asian Crisis: Contagion Stemming from Indonesia . . . 104B.6 East Asian Crisis: Contagion Stemming from Indonesia . . . 105

LIST OF TABLES ix

B.7 East Asian Crisis: Contagion Stemming from Korea . . . . . 106B.8 East Asian Crisis: Contagion Stemming from Korea . . . . . 107B.9 Mexican Peso Crisis: Contagion Stemming from Mexico . . . 108B.10 Mexican Peso Crisis: Contagion Stemming from Mexico . . . 109B.11 U.S. Stock Market Crash: Contagion Stemming from the U.S. 110B.12 U.S. Stock Market Crash: Contagion Stemming from the U.S. 110B.13 East Asian Crisis: Contagion Stemming from Hong Kong . . 111B.14 East Asian Crisis: Contagion Stemming from Thailand . . . . 112B.15 East Asian Crisis: Contagion Stemming from Indonesia . . . 113B.16 East Asian Crisis: Contagion Stemming from Korea . . . . . 114B.17 Mexican Peso Crisis: Contagion Stemming from Mexico . . . 115B.18 U.S. Stock Market Crash: Contagion Stemming from the U.S. 116

Abbreviations

ABS Asset Backed Securities

AIC Akaike Information Criterion

BBA British Bankers Association

BIS Bank for International Settlements

CDS Credit Default Swap

CDS B CDS Index of Major Banks

CDO Collateralised Debt Obligations

CEE Central and Eastern Europe

Comp Component

Cov Covariance

Cum. Cumulative

DTCC Depository Trust and Clearing Corporation

EUM European Monetary Union

EU European Union

FR Forbes and Rigobon

FRN FR-Test Statistic With Non-Overlapping Data

FRO FR-Test Statistic With Overlapping Data

GDP Gross Domestic Product

GG Gonzalo Granger

G3 USA, Japan, Germany

GBS Government Bond Spread

H0 Null Hypothesis

H1 Alternative Hypothesis

HQIC Hannan Quinn Information Criterion

i.i.d. Independent and Identically Distributed

IMF International Monetary Fund

ISDA International Swaps and Derivatives Association

LATAM Latin America

ABBREVATIONS xi

Max. Maximum

ME Middle East

Min. Minimum

ν (Adjusted) Correlation Coefficient

ν Sample Estimator of ν

Obs. Observations

OECD Organisation for Economic Co-Operation and

Development

PASOK Panellinio Sosialistiko Kınima

PCA Principal Component Analysis

PIGS Portugal, Ireland, Greece, Spain

PIIGS Portugal, Italy, Ireland, Greece, Spain

QLR Quandt Likelihood Ratio

ρ (Unadjusted) Correlation Coefficient

ρ Sample Estimator of ρ

SBIC Schwartz Bayesian Information Criterion

S.D. Standard Deviation

σ2 Volatility

S&P Standard & Poor’s

Std. Err. Standard Error

Stks Stocks

Stks R Regional Stock Market Index

Stks W Global Stock Market Index

T Sample Size

V ar Variance

VAR Vector Autoregression

VECM Vector Error-Correction Model

VIX Volatility Index

Acknowledgements

I owe credit to many people who supported me in writing this dissertation.First and foremost, I want to thank my supervisor at the University ofBern, Prof. Dr. Klaus Neusser, for his accessibility, continuous support andvaluable comments. I am indebted to my co-advisor at the University of St.Gallen, Prof. Dr. Monika Butler, for her advice and support over the years.

Furthermore, I would like to thank Prof. Dr. Klaus W. Wellershoff forhis generous support and fruitful comments during the writing process of thisdoctoral thesis. It was both a pleasure and a challenge to pursue this projectwhile at the same time having the opportunity to contribute to building anew company. I am especially grateful to my colleagues at Wellershoff &Partners for their support and understanding when times were stressful.

I had the pleasure and privilege to co-author the first two chapters ofthis doctoral thesis with my fellow student colleague Sergio Andenmatten.His ideas and drive as well as our strong collaboration facilitated this jointproject.

Moreover, I would like to thank Dr. Dr. Doris Benz, Dr. Dirk Faltin, Dr.Mirko Jazbec, Dr. Daniel Kalt, and Dr. Alexander Kobler for encouragingmy decision to do this dissertation by sharing their experience and offeringvaluable advice and support.

Further, I am thankful to the participants of the Swiss Program forBeginning Doctoral Students in Economics at the Study Centre in Gerzenseefor making this program an unforgettable experience. In particular, I wouldlike to thank Tobias Duschl and Jan Imhof.

I would also like to acknowledge participants of the brown-bag seminarat the University of Bern and the anonymous referee from the Swiss Journalof Economics and Statistics for valuable comments.

Acknowledgements xiii

Further, I am thankful to Roy Greenspan for his fine sense of language whenediting this dissertation.

Finally, I would like to express my heartfelt gratitude to my partnerSylke, to my family and to my best friends. Without their unlimited andunfailing support, this dissertation, and a lot more than this, would not havebeen possible.

Preface

When George Papandreou became the prime minister of Greece, on October4, 2009, the Greek economy still faced severe repercussions from the 2008financial crisis. Little more than two weeks later, on October 20, the newgovernment announced that official statistics on Greek debt had previouslybeen fabricated. Instead of a public deficit estimated at 6% of gross domesticproduct (GDP) for 2009, the government now expected a figure twice ashigh to materialize. This revelation was the starting point of the Greekdebt crisis.

Since then, European policymakers have repeatedly reassured financialmarket participants that the situation could be contained. In reality, aftervarious measures and rescue packages, only brief periods of relief have beenachieved and the Greek debt crisis turned into a pan-European one. It waswidely assumed that a default by Greece would have contagion effects onthe euro area as a whole, with potentially severe consequences for the worldeconomy and global financial markets. Since October 2009, sovereign riskhas been a main driver for financial markets.

Sovereign credit risk has been repriced dramatically, as reflected, for in-stance, in the soaring premia on credit default swaps (CDSs). CDSs arebilateral contracts used to transfer risk among market participants; theyare basically defined by four parameters: the reference entity, the notionalamount, the price (spread or premium), and the maturity. One participantis the so-called protection buyer, who wants to buy insurance against thedefault of a specific entity, the so-called reference entity. The other party isthe protection seller, who writes the insurance on the reference entity. Tocompensate the seller of the insurance for the assumed risk, the protectionbuyer pays an initially fixed premium every year (or quarter) on the insurednotional value. If a credit event occurs, the CDS is triggered and the pro-

Preface 2

tection seller has to pay the difference between the insured notional valueand the recovery value.

In general, a CDS makes it possible to invest in the credit quality of acorporate or a sovereign entity. If an investor believes that the credit qualityof a corporate will decline in the coming months and believes this is not yetpriced into current spreads, he should buy protection. Once the premiumrises, the investor will profit because his insurance will increase in value. Hecan close the insurance contract whenever he wants and monetise his gains.

Thus, buying protection, for example, on Germany is not speculatingon the likelihood of that country going bankrupt. It merely means thepurchaser believes Germany’s credit quality will decrease in the future. Atthe same time, the seller of the protection on Germany believes that – givencurrent spreads – it is attractive to agree to the contract.

How dramatic the development in the CDS market has been since theemergence of the Greek debt crisis may be illustrated as follows: While theCDS premium for a Greek government bond with a 5-year maturity anda notional value of USD 10 million was 124 basis points on October 20,2009, it soared to 2150 basis points by July 6, 2011. This means that theinsurance annual cost of protection against a default of this particular Greekgovernment bond increased by a factor of more than 17, soaring from USD124,000 to USD 2.15 million.

At the same time, CDS premia for many other countries also increasedsharply, notably for Ireland, Portugal and Spain, countries characterisedeither by very high debt-to-GDP ratios, high public deficits, a high ratioof net debt interest payments to GDP, or fundamental structural economicproblems. Even though the specific set of problems differed in each, manyfinancial market participants and observers assessed the overall situation inthese three countries as all but unsustainable.

At some point, a public debate started on whether the widening of sov-ereign credit spreads and the worsening refinancing conditions were subjectto market speculation, or, even more worrisome, to market manipulation.Particularly in the case of Greece, many suspected the CDS market wasresponsible for the widening in the spread of the underlying governmentbonds.

Spurred by the important questions this debate addressed and by thefact that the relatively young European CDS market has been the subject

Preface 3

of only limited research to date, this thesis analyses the developments onCDS markets since the beginning of the Greek debt crisis.

In the first chapter, which is based on Andenmatten and Brill (2011a),the empirical relationship between CDS premia and government bond spreadsis described and examined for Portugal, Italy, Ireland, Greece, and Spain.The starting point for the analysis is the theoretical equivalence of CDSpremia and credit spreads, as derived by Duffie (1999) who showed thatunder certain conditions, the CDS premium should be approximately equalto the credit spread, that is, the yield minus the risk-free rates of the ref-erence bond of the same maturity. The analysis yields some evidence of along-term relationship between the two markets in the sense of cointegration.In most cases, only CDS premia contribute to the price discovery process.In the other instances, both markets contribute more or less equally. Thissuggests that bond spreads react only sluggishly to long-term imbalances,as measured by the cointegrating relationship, behaviour that may be due –at least partially – to liquidity effects.

In the second chapter, which is based on Andenmatten and Brill (2011b),a rolling-crisis-window approach for contagion testing is applied, derivedfrom and enhancing an approach proposed by Forbes and Rigobon (2002).The rolling-crisis-window approach helps account for crises of longer-than-usual duration, as is the case for Greece since its crisis began in October2009.

This rolling-crisis-window approach is applied to test whether the co-movements of sovereign CDS premia increased significantly after the Greekdebt crisis started. The sample consists of daily data between October2008 and July 2010 for 39 countries from both emerging and industrializedcountries. The test results indicate that there were periods of contagionfor CDS markets during the Greek debt crisis, which contrasts with theresults of Forbes and Rigobon (2002) for equity markets during the EastAsian crisis in 1997-98, the Mexican peso crisis in 1994, and the U.S. stockmarket crash in 1987, challenging their conclusion of “no contagion, onlyinterdependence.”

In the third chapter, which is based on Brill (2011), the rolling-crisis-window approach is applied to equity markets during these three crises andthe results are compared to those of Forbes and Rigobon (2002). The sampleconsists of daily returns of 32 MSCI equity market indices in both local

Preface 4

currencies and US dollars. In contrast to the static approach of Forbes andRigobon (2002), the rolling-crisis-window approach yields ample evidence ofcontagion during these crises. This result is further supported by extensiverobustness tests that entailed altering the periods of relative stability andusing daily returns in US dollars instead of the local currency.

Chapter 1

Did the CDS Market Push

Up Risk Premia for

Sovereign Credit?

1.1 Introduction

After the worst of the financial crisis seemed to be over and the recoveryunder way, financial markets started to focus on the fiscal situation of certaincountries. The financial crisis had caused the deficits of many countries toincrease substantially. Stimulus programs, bail-outs of financial institutionsand reduced tax revenues were the main drivers of the deteriorating fiscalconditions. For instance, the United States ran a budget deficit equivalentto 9.9% of GDP in 2009, the biggest since 1945. The total outstandingfederal debt is predicted to be approximately 90% of GDP in 2010.

But the U.S. was not alone. The UK almost doubled its debt-to-GDPratio and the euro area as a whole is expected to run a budget deficit ofaround 7% in 2010 (6.1% in 2009). This caused the average debt-to-GDPratio in the euro area to approach 84% (cf. Figure 1.1). Pro memoria: theMaastricht Treaty stipulates a maximum budget deficit for member statesof 3%, and a 60% ceiling for the debt-to-GDP ratio. Is this developmentsustainable? Most likely not. For instance, in their empirical study, Rein-hart and Rogoff (2010) show that a debt-to-GDP ratio of 90% is a critical

This chapter is based on Andenmatten and Brill (2011a). Both authors contributed equallyto this work.

1.1 Introduction 6

threshold. Above 90%, growth rates of real GDP fall significantly.Clearly, the increased budget deficits and the worsening fiscal conditions

of sovereign entities attracted the attention of financial market participants.After the contagion in the banking system, the next sources of trouble forglobal markets were localised. Consequently, activity on the sovereign CDSmarket increased and gained more attention among the financial marketcommunity. The focal points of this development were the PIIGS countries(Portugal, Italy, Ireland, Greece, Spain), which are characterised by veryhigh debt-to-GDP ratios, exceptionally high deficits, a high ratio of net debtinterest payments to GDP and fundamental structural economic problems.

The situation in Greece has received most attention: its high refinancingneeds, along with fabricated statistics and financial transactions designedto hide liabilities, as well as a weakening economic situation and increasingrefinancing costs fuelled the market’s fears of potential default. It was widelyassumed that such a default would have a contagion effect on the otherPIIGS countries and on the euro area as a whole. Hence, in the spring of2010, sovereign risk was a main driver for financial markets and the sovereignCDS market was an important stress indicator.

16.439.2

42.847.4

58.665.666.3

70.973.9

76.782.582.984.084.6

101.2116.7

124.9

60%−threshold according tothe Maastricht Treaty

0 20 40 60 80 100 120 140

LuxembourgSlovakiaSlovenia

FinlandCyprus

NetherlandsSpainMalta

AustriaGermany

FranceIreland

Euro areaPortugalBelgium

ItalyGreece

Source: European Commission

Figure 1.1: Estimated Debt-to-GDP Ratio for 2010 in Percent

1.1 Introduction 7

At some point, a public debate started on whether the widening of sover-eign credit spreads and the worsening refinancing conditions was subject tomarket speculation, or even worse, to market manipulation. Particularly inthe case of Greece, many suspected the CDS market was responsible for thewidening in the spread of the underlying government bonds.

0

100

200

300

400

Basi

s Po

ints

2007 2008 2009 2010

Portugal ItalyIreland GreeceSpain

Source: Bloomberg

Figure 1.2: CDS Premia for Portugal, Italy, Ireland, Greece, and Spain

Motivated by this debate and the fact that the European CDS market is arelatively young market which has not been the subject of a great deal ofresearch so far, we analyse the empirical relationship between sovereign CDSand the government bond market for the PIIGS countries. Most existingpapers on sovereign CDS deal with emerging markets, which were the birth-place of the sovereign CDS market. For instance, Longstaff, Pan, Pedersen,and Singleton (2011) explore the factors driving sovereign CDS, while Panand Singleton (2008) analyse the term structure of sovereign CDS. Whatmost of these studies have in common is that they do not cover a crisisin sovereign credit markets. Our analysis, however, focuses on the CDSmarkets of the PIIGS countries since their inception in 2007, and thus alsoincludes a period of crisis.

This chapter proceeds as follows: In section 1.2 we introduce basic fea-tures of sovereign CDS markets and discuss the role of CDS in financial

1.2 Introduction to the CDS Market 8

markets as well as movements in the volumes of outstanding CDS contracts.In section 1.3 we examine the empirical relationship between CDS premiaand government bond spreads. Our analysis is based on the theoreticalequivalence of CDS premia and credit spreads, as derived by Duffie (1999)who shows that under certain conditions the CDS premium should be ap-proximately equal to the credit spread, that is, the yield minus risk-freerates of a reference bond of the same maturity. By applying cointegrationtechniques, Blanco, Brennan, and Marsh (2005) find support for Duffie’s the-oretical equivalence based on a sample of 33 corporate bonds and the CDSpremia for these bonds. Motivated by their findings, we apply this approachto CDS premia and government bond spreads for the PIIGS countries. Indoing so, we first test whether we are able to find any support for a long-run equilibrium in the sense of Duffie’s theoretical equivalence. Second, weanalyse potential deviations from this equilibrium and test whether one ofthe two markets might be inefficient with respect to the price discovery pro-cess. Third, we examine whether potential inefficiency in one of the marketsmight be related to measures of market liquidity. Section 1.4 concludes.

1.2 Introduction to the CDS Market

1.2.1 CDS in General

CDSs are bilateral contracts used to transfer risk between market parti-cipants and are basically defined by four parameters: the reference entity,the notional amount, the price (usually referred to as ‘spread’ or ‘premium’),and the maturity. One participant is the ‘protection buyer’ who wishes tobuy insurance against the default of a specific entity, the so-called ‘referenceentity’. The other party is the ‘protection seller’, who writes the insuranceon the reference entity. To compensate the seller of the insurance for the as-sumed risk, the protection buyer pays a spread (which is initially fixed) eachyear (or each quarter) on the insured notional value. If a credit event oc-curs, the CDS is triggered and the protection seller has to pay the differencebetween the insured notional value and the recovery value.1 Settlement is al-ways made by means of an auction and is mandatory (either cash or physical

1The so-called ISDA Credit Derivative Determination Committee – consisting of buy andsell side members – will decide whether the requirements for a credit event are fulfilled.The decision of the determination committee is binding for the whole market.


delivery), i.e. investors are signed up automatically for all auctions.A CDS is an easy way to invest in the credit quality of a corporate entity

or a country. If an investor believes that the credit quality will decrease infuture and that this is not yet priced into the current CDS premium, heshould buy protection. Once the premium increases, he will make moneybecause his insurance will increase in value. He can terminate the insurancecontract whenever he wants and monetise his gains. Thus, buying protectionon Germany is not speculating on the country going bankrupt. It merelymeans that somebody believes the credit quality of Germany will decreasein the future. At the same time, the seller of the protection on Germanybelieves that – given the current spreads – it is attractive to agree to thecontract. The view that credit quality will deteriorate is hard to applyto bonds, since shorting bonds is not always an easy endeavour. Sincethe CDS market makes betting on a deterioration in credit quality easy, ithas the potential to supplement and improve the price discovery process inunderlying sovereign bond markets.

1.2.2 Special Features of the Sovereign CDS Market

In the case of sovereign CDSs there are basically three credit events thatcan be triggered, based on the framework provided by the InternationalSwaps and Derivatives Association (ISDA). Ghosh, Hagemans, Leeming, andWillemann (2010) classify these events in the following way:2

1. Failure to pay: This event is recognised if the country has failed topay a minimum amount, usually USD 1 million.

2. Restructuring: This event is triggered if bonds with an outstandingvolume of at least USD 10 million are restructured.3

3. Repudiation or moratorium: This event is triggered if “an authorisedgovernment official disclaims, repudiates or rejects the validity of oneor more obligations or imposes a moratorium or standstill. In additionto this, there has to be a failure to pay or a restructuring event, notsubject to the minimum amounts given above, within 60 days or thenext bond payment date (whichever is later)” (Ghosh et al., 2010).

2In case of a credit event, however, the ISDA Credit Derivative Determination Committeemay interpret things differently.

3For more information regarding restructuring events cf., for example, Verdier (2004).


A further special feature of sovereign CDSs is the quotation. Sovereign CDSsare denominated in a different currency than the bulk of the outstandinggovernment debt, e.g., European CDSs are quoted in USD and vice versa.This is based on the assumption that, if a credit event has occurred, thelocal currency would depreciate significantly.

1.2.3 History of the Sovereign CDS Market

CDSs on the government debt of emerging markets have been used regularlysince the late 1990s. According to Ammer and Cai (2007), emerging marketsovereigns are among the largest high-yield borrowers in the world, typicallywith more bonds outstanding, longer maturities, larger issues, and moreliquidity than their corporate counterparts. At an early stage, CDS contractssatisfied market needs to insure against a default by these countries. In 1998the whole CDS market profited from the standardisation of contracts, whichled to a fast growing CDS market. In 2002, JP Morgan introduced the firstsovereign CDS index – the TRAC-X index – where the constituents werealmost exclusively emerging market sovereigns (Mexico, Russia and Brazilmade up more than 37% of the index). In 2003 only 10% of all sovereignCDS trades were on non-emerging market countries.

The financial crisis changed the situation as the level of public debt in-creased massively in industrialised countries. As a consequence, volumesof sovereign CDS contracts on developed countries began to grow. Thisincreased interest led to the introduction of the Western Europe SovereignCDS Index in September 2009. The outstanding net volume of this indexhas increased massively since the launch. The Bank for International Set-tlements (BIS) reports a downward trend in the outstanding gross volumeof CDS worldwide (-40% since the first half of 2008). However, accordingto data from the Depository Trust and Clearing Corporation (DTCC) thesubcategory of sovereign CDSs is still growing sharply and faster than therest of the CDS market.

1.2.4 The Role of CDS in Financial Markets

On the one hand, CDS may increase efficiency in the allocation of capital.Historically, investors who lend money to a company had to bear the creditrisk of that company. With the advent of CDSs it became possible for


investors to outsource some of the funding risks of a company to the market.As a result, companies can obtain more credit than they would otherwiseand on better terms. Furthermore, CDSs make financial markets potentiallymore efficient and transparent in price discovery as they increase liquidity.Stulz (2010) argues that despite huge and unexpected losses in underlyingproducts the CDS market remained fairly liquid for long periods during thefinancial crisis when the corporate bond market was totally illiquid.

On the other hand, CDS might create adverse incentives in the market.For example, a bank which lends money to a company and hedges itselfin the market has fewer incentives for monitoring the firm. Additionally,a hedged investor could prefer the bankruptcy of a company in financialdistress rather than working out a restructuring plan with the debtor.

1.2.5 Outstanding Volumes of CDS Contracts

The DTCC provides an electronic platform for banks and clients to confirmthe agreed contracts electronically. Virtually all electronically confirmedtransactions run through this platform. Since November 2008, the DTCChas provided weekly data for outstanding CDS positions on specific referenceentities and trading activity. This measure helps to increase transparencyin the market. The DTCC data is the only hard data available for the CDSmarket. The BIS, the ISDA and the British Bankers Association (BBA)reports are all based on surveys, provide only aggregated data, and arepublished less frequently. However, the DTCC is not representative for thewhole market, as more bespoke products like CDS on collateralised debtobligations (CDO) or asset backed securities (ABS) are not confirmed elec-tronically. There are different measures for the size of the CDS market andoutstanding positions on specific reference entities. Every measure tells adifferent story. As the DTCC and BIS data are the most important and themost frequently cited sources, the following concepts are crucial for under-standing the inner workings of the CDS market. Hence, in the following, werefer to the definitions used by DTCC and BIS.

Based on DTCC’s definition, ‘gross notional value’ measures the sum ofthe notional of all outstanding CDS contracts on a per trade basis. This canbe illustrated by the following example: Assume a transaction of USD 10million notional between buyer and seller of protection. DTCC reports thistransaction as one contract with a USD 10 million gross notional value, and


not as two contracts worth USD 20 million. The problem with gross data isthat from a risk perspective it overestimates the size of the market. To closean existing deal, an offsetting trade is often done. The actual risk of a defaultof the reference entity would be zero for the involved parties. However, thedeal actually closed would flow into the calculation of the gross notionalvolume twice. Because most CDS traders have a netting agreement in place,the systemic risk is not increased through this practice. Therefore, the grossnotional value overestimates the size of the market. However, for evaluatingthe trading activity, the gross value can be viewed as an indicator. Figure 1.3illustrates the movements in outstanding gross volumes for different entities.

0

1,000

2,000

3,000

USD

Billi

on

11,500

12,500

13,500

14,500

USD

Billi

on

09/2008 03/2009 10/2009 03/2010

Corporate (lhs)Sovereign (rhs)Other (rhs)

Source: DTCC

Figure 1.3: Outstanding Gross Notional Volumes by Entities

In addition to gross volumes, the BIS publishes “gross market values.” Thesevalues are defined as the sum of the total gross positive market value of con-tracts and the absolute value of the gross negative market value of contractswith non-reporting counterparties. Gross market values supply informationabout the potential scale of market risk in CDS markets.

Finally, the DTCC defines the “net notional value” for any single refer-ence entity as the sum of the net protection bought by net buyers or thesum of the net protection sold by net sellers, respectively. The aggregate netnotional value is calculated based on the concept of counterparty families,


which, for example, includes all of the accounts of a particular asset manager.Based on this, DTCC reports the aggregate net notional value as the sum ofnet protection bought, or equivalently sold, across all counterparty families.Accordingly, the net notional value for a particular reference entity indicatesthe maximum possible exchange between net sellers of protection and netbuyers of protection that could be required in case of a credit event. Figure1.4 illustrates the development of outstanding net volumes of sovereign CDScontracts for the PIIGS countries.

Stulz (2010) uses the example of the bankruptcy of the investment bankLehman Brothers to illustrate the difference between gross and net volumes:When Lehman went bankrupt, there were CDS records on Lehman for agross notional value of USD 72 billion registered at DTCC’s Trade Informa-tion Warehouse. According to the recovery rate that had been determinedin the auction process protection sellers had to pay 91.375 cents on the dol-lar to settle the contracts. The settlement for these contracts went withoutmany difficulties and on a net basis only USD 5.2 billion was exchangedthrough the DTCC. One important reason for both the smooth process andthe relatively small amount of net positions was that many institutions wereboth buyers and sellers of protection on Lehman. Accordingly, the grossnotional value had overstated the risks.

0

5

10

15

20

25

30

USD

Billi

on

09/2008 03/2009 10/2009 03/2010

Italy SpainGreece PortugalIreland

Source: DTCC

Figure 1.4: Outstanding Net Notional Volumes by Countries

1.3 Empirical Analysis 14

During the Greek debt crisis, a debate arose as to whether the CDS markethad been subject to manipulation that might have worsened the magnitudeof the crisis. According to Duffie (2010), one way to manipulate marketscould be that speculators progressively increase their protection on a certaincountry to push out CDS premia. Due to this practice, contracts boughtpreviously increase in value. Another possibility for market manipulationmight be achieved by the placement of large trades in the market with theaim of spreading market rumours. As Duffie (2010) argues, both activitiesshould manifest in an increase in outstanding net volumes.

However, we find no strong increase in outstanding net volumes in theDTCC data. As shown in Figure 1.4, net outstanding volumes for the PIIGScountries only increased slightly on average. In the case of Greece, therewas actually a drop in outstanding net volumes at the start of the crisis inNovember 2009. The net position for Greece was USD 8.7 billion in the firstweek of January 2010, and ranged between USD 8.5 billion and USD 9.2billion in the following months. This compares to a net position for Greeceof USD 7.4 billion at the beginning of 2009. Hence, the data suggests thatthere was no surge of interest in either 2009 or 2010 and that the movementin outstanding net volume does not signal any increase in speculative activityduring the Greek debt crisis.

1.3 Empirical Analysis

After introducing basic features of CDS markets and discussing the changesin the outstanding volumes of CDS contracts, we now turn to the empiricalrelationship between sovereign CDS premia and government bond spreads.The starting point for our analysis is the theoretical equivalence of CDSpremia and credit spreads as derived by Duffie (1999), who shows that, undercertain conditions,4 the CDS premium should be approximately equal to thecredit spread, that is, the yield minus risk-free rates of the reference bondof the same maturity.5

4E.g., market participants should be able to short risk-free bonds, which is equivalent toassuming that they can borrow at the risk-free rate. Also, market participants should beable to short the risky bonds, while counterparty default risk in a CDS is assumed to benegligible.

5Cf. also discussions in Hull, Predescu, and White (2004) and Zhu (2006).


According to Blanco et al. (2005), this can be illustrated as follows: Supposean investor buys an n-year par yield bond issued by a reference entity withy being the yield on this bond. In addition, suppose the investor buys creditprotection on that entity for n years in the CDS market at a cost of s. If s isexpressed annually as a percentage of the notional principle, then the annualreturn of the investor equals y − s. If r denotes the yield on an n-year paryield risk-free bond, the relationship r = y − s should hold approximately.If r is greater than y − s, then shorting the risky bond, writing protectionin the CDS market, and buying the risk-free bond would be a profitablestrategy for an arbitrageur. Similarly, if r is less than y−s, buying the riskybond, buying protection in the CDS market, and shorting the risk-free bondwould be a profitable arbitrage opportunity.

By applying cointegration techniques, Blanco et al. (2005) find supportfor Duffie’s theoretical equivalence based on a sample of 33 corporate bondsand the CDS premia for these bonds. The authors interpret this as a long-run equilibrium condition for the pricing of corporate credit risk. In ad-dition, the authors show that there are two forms of deviations from thelong-run equilibrium. One form of deviation is relatively long-lived and canbe explained by “imperfections in the contract specification of CDSs andmeasurement errors in computing the credit spread.” However, this formof deviation from the equilibrium is only apparent in three cases of theirsample. The other form of deviation is short-lived and arises due to “a leadfor CDS prices over credit spreads in the price discovery process.”

In what follows, we apply the approach by Blanco et al. (2005) to CDSand government bond markets in the PIIGS countries. Therefore, we firsttest whether we find support for a long-run equilibrium in the sense of thetheoretical equivalence derived by Duffie (1999). Second, we focus on thesecond form of deviation, i.e. short-run deviations from the equilibrium, andtest whether one of the two markets might be inefficient with respect to theprice discovery process. Finally, we examine whether potential inefficiencyin one of the markets might be related to measures of market liquidity.

In order to do this we proceed as follows: First, we briefly describe ourdata and present descriptive statistics. Second, we look at cross-correlationsbetween CDS premia and government bond spreads. Third, we analyse thepossible long-run equilibrium behaviour of the series by performing Johansencointegration tests. Fourth, we look into the price discovery process, using


vector error-correction models (VECM) of market prices and Granger caus-ality tests. Finally, we perform analyses to detect any differences betweenthe liquidity of the two markets, which might partly explain the lead-lagrelations in the price discovery process.

1.3.1 Data Description

Our sample is based on daily data that runs from January 1, 2007, throughApril 16, 2010. Table 1.1 lists basic descriptive information and the numberof observations for both CDS premia and government bond spreads in oursample. We use CDS premia from Bloomberg with a notional value of USD10 million. All prices are based on the standard ISDA contract for physicalsettlement with a constant 10-year maturity. For calculating the governmentbond spreads we use 10-year government bond yields from Thomson ReutersDatastream. As proxy for risk-free bonds we use German government bonds.However, we have to acknowledge that German government bonds are not anideal proxy for the unobservable risk-free rate. One reason is that Germany’sfiscal situation has also been deteriorating since 2007. Other reasons arerelated to government bonds in general, such as taxation treatment, repospecials, scarcity premia, and benchmark status (Blanco et al., 2005). Eventhough German government bonds are not an ideal proxy, they still seem tobe the best available in our context.

As we mentioned earlier, activity on CDS and government bond marketsincreased and gained more attention among the financial market communitydue to the situation surrounding Greece. On October 4, 2009, GeorgePapandreou became the new prime minister of Greece after his PanhellenicSocialist Movement (Greek: Panellinio Sosialistiko Kınima; PASOK) partywon the general election. At that time, the Greek economy was still facedwith the severe repercussions of the financial crisis. Around two weeks later,on October 20, officials of the new government announced that Greek debtstatistics had been forged in the past. Instead of a public deficit of 6% ofGDP for 2009, the government now expected twice as much to materialise.This was the starting point of the Greek debt crisis. Based on this, we de-cided not only to look at the whole sample (cf. Panel A of Table 1.1), butalso at two sub-samples. Accordingly, Panel B concentrates on the periodprior to the Greek problem, i.e. the period from January 1, 2007, to October19, 2009; Panel C on the period thereafter.


Table 1.1Descriptive Statistics

This table lists basic descriptive information and the number of observations for bothCDS premia and government bond spreads in our sample. We use daily CDS premiafrom Bloomberg with a constant 10-year maturity. For calculating the government bondspreads we use 10-year government bond yields from Thomson Reuters Datastream. Allspreads are based on German government bonds. The data run from January 1, 2007, toApril 16, 2010. Panel A shows the descriptive statistics for the whole sample. Panel Bconcentrates on the period prior to the Greek problem, Panel C on the period thereafter.

Panel A: January 1, 2007 until April 16, 2010

CDS Premia Government Bond Spreads

Obs. Mean S.D. Min. Max. Obs. Mean S.D. Min. Max.

Portugal 832 60.6 44.1 8.0 227.0 860 56.7 39.3 10.9 161.8Italy 850 70.8 50.8 11.0 205.0 860 61.8 36.9 16.1 155.9Ireland 593 132.0 82.2 22.0 365.0 860 81.0 79.9 -3.7 262.4Greece 835 114.6 99.1 10.0 396.0 860 116.7 98.7 16.2 424.4Spain 839 61.9 43.4 6.0 169.0 860 40.6 30.4 3.6 123.3

Panel B: January 1, 2008 until October 19, 2009



Portugal 468 68.6 28.9 26.0 157.0 470 70.0 36.3 21.3 161.8Italy 469 89.3 47.7 29.0 205.0 470 80.5 35.6 25.7 155.9Ireland 464 126.8 92.0 22.0 365.0 470 107.0 78.9 9.6 262.4Greece 469 122.3 68.1 30.0 282.0 470 128.1 75.6 29.8 298.5Spain 469 73.9 32.6 26.0 165.0 470 52.3 27.5 7.4 123.3

Panel C: October 20, 2009 until April 16, 2010



Portugal 129 118.9 40.4 60.0 227.0 129 86.9 28.4 47.7 161.7Italy 129 108.8 16.4 79.0 157.0 129 71.1 7.2 55.5 85.6Ireland 129 150.7 14.3 117.0 178.0 129 147.5 9.3 133.3 180.9Greece 129 268.0 70.5 134.0 396.0 129 260.2 78.2 131.4 424.4Spain 129 112.9 20.4 77.0 169.0 129 64.9 11.1 46.4 90.7


1.3.2 Basic Analysis

Figure 1.5 plots CDS premia and government bond spreads for Portugal,Ireland, Greece, and Spain. It is evident that the relationship between CDSpremia and the government bond spreads is very close. However, it is alsoobvious that there are periods when CDS premia and government bondspreads do not move in step. In Spain, for instance, CDS premia increasestrongly at the end of our sample while government bond spreads movesideways.

Portugal

0

50

100

150

200

250

Basi

s Po

ints

2007 2008 2009 2010

Ireland

0

100

200

300

400Ba

sis

Poin

ts

2007 2008 2009 2010

Government Bond Spreads CDS Premia

Greece

0

100

200

300

400

500

Basi

s Po

ints

2007 2008 2009 2010

Spain

0

50

100

150

200

Basi

s Po

ints

2007 2008 2009 2010

Source: Bloomberg, Thomson Reuters Datastream

Figure 1.5: Government Bond Spreads and CDS Premia

At first sight, this observation is supported by the correlation coefficientsbetween CDS premia and government bond spreads. For all five countriescorrelation is above 0.90 over the whole time period of the sample if calcu-lated in levels (cf. Panel A of Table 1.2).

However, if we focus on the period after October 2009, when the Greekproblem started, we find that correlation is lower in the cases of Italy andIreland. For Italy, for instance, the correlation coefficient drops to 0.24.Also, in the case of Spain, our observation based on Figure 1.5 is supported,as the correlation coefficient falls to 0.79 in the period between October 20,2009, and April 16, 2010. In the period January 1, 2008, to October 19,2009, the correlation coefficient was 0.93.


Table 1.2Correlation Coefficients

This table reports correlation coefficients between CDS premia and government bondspreads; the results in Panel A are based on calculations in levels, the results in PanelB on calculations in first differences. When using first differences, i.e., daily changes inlevels, we obtain stationary time series (we performed augmented Dickey-Fuller tests totest for unit roots). We distinguish between three time periods as in Table 1.1.

Panel A: In Levels

January 1, 2007 January 1, 2008 October 20, 2009until until until

April 16, 2010 October 19, 2009 April 16, 2010

Portugal 0.90 0.88 0.98Italy 0.94 0.94 0.24Ireland 0.94 0.94 0.44Greece 0.97 0.93 0.97Spain 0.95 0.93 0.79

Panel B: In First Differences



Portugal 0.50 0.39 0.63Italy 0.37 0.37 0.42Ireland 0.42 0.40 0.51Greece 0.68 0.40 0.82Spain 0.42 0.33 0.58

This approach might be problematic, however, if we deal with non-stationaryvariables as these variables usually show instability in the estimation ofcorrelation coefficients. Therefore, we calculate first differences, i.e. dailychanges, of all variables in order to transform the time series into station-ary ones. Augmented Dickey-Fuller tests confirm that the variables in firstdifferences are all stationary.

In Panel B of Table 1.2, we see that the correlation coefficients are muchlower for the whole time period of the sample than they are in Panel A.What is more, the correlation coefficients increase for all five countries if wefocus on the period after the Greek problem started. This is most apparentin the case of Greece. While the correlation coefficient for the period prior tothe Greek debt crisis is 0.40, it increases to 0.82 in the period from October20, 2009, to April 16, 2010. This might be an indication of contagion effects


as defined by Forbes and Rigobon (2002), i.e. a significant increase in cross-market linkages after a shock to one country.6

1.3.3 Long-Run Relations

The correlation coefficients indicate that there has indeed been a closerelationship between CDS premia and government bond spreads for thePIIGS countries since 2007. In the next step, we now test whether we findsupport for the theoretical equivalence between the two markets as derivedby Duffie (1999). In order to do this, we follow the approach of Blanco etal. (2005) by using cointegration techniques. The authors argue that thisapproach (and use of the term ‘long-run’) is valid, even though it mightappear inappropriate at first sight as their data set covers only 18 months.Our data set covers 28 months and is thus considerably longer. What ismore, Hakkio and Rush (1991) argue that it is not only the length of thedata set that matters but that the ratio of the length of the data set to thehalf-life of deviations is even more relevant. With a half-life of only a fewdays, our data set should allow us to use the cointegration approach.

We report Johansen trace test statistics7 for the number of cointegratingrelations between CDS premia and government bond spreads in Table 1.3.The test statistics are based on a model with a constant and up to three lags.The number of lags in the underlying VAR is optimised using the SchwartzBayesian Information Criterion (SBIC) for each entity. For selecting thenumber of lags to include in the VAR equations we also looked at the AkaikeInformation Criterion (AIC). On average, the AIC indicates 1.1 more lagsthan the SBIC. However, the Johansen trace test statistics signal anotherresult in only one of 15 cases if we use the AIC for optimising the underlyingVAR. Hence, the test statistics appear to be robust with respect to thesetwo information criteria. As in the previous sections we distinguish threetime periods.

If we focus on the whole sample, we find evidence of cointegration forItaly and Greece (as indicated by ∗). For these two countries, the CDSand government bond markets appear to price risk equally, on average, upto some constant term that might reflect mis-measurement of the risk-free

6Testing for contagion in CDS markets during the Greek debt crisis is beyond the scope ofthis paper. However, the interested reader is referred to Andenmatten and Brill (2011b).

7Cf. Johansen (1991)


rate. For Portugal and Spain, however, cointegration is rejected, suggestingno long-term relationship between CDS premiums and government bondspreads.

If we concentrate only on the time period up to the starting point ofthe Greek debt crisis, there is stronger evidence for cointegration. We findsupport for cointegration in four out of the five entities. This is rathersurprising, given that the sample period is shorter than the one in Panel A.In the case of Ireland, one reason might be that this country already had itsown debt crisis at the end of 2009 due to the troubles encountered by theIrish banks.

Finally, if we focus on the time period after Greece’s problems started, weonly find evidence for cointegration in the case of Portugal. In the other fourcases we have to reject a long-term relationship in the sense of cointegration.This might be due to the relative shortness of the sample period (only aboutsix months). Another reason might be that the Greek debt crisis has led toan increased focus on the fiscal situation in other European countries, too.As we discussed earlier, this led to increased market activity in both theCDS and government bond markets. It may also have disrupted the pricingof risk in both markets as well as some short-term disconnection.

Table 1.3Johansen Trace Test Statistics

This table reports Johansen trace test statistics for the number of cointegrating relation-ships between CDS premia and government bond spreads. The test statistics are basedon a model with a constant and up to three lags. The number of lags in the underlyingVAR is optimised using the SBIC for each entity. The 5% critical values (as indicated by∗) for the trace statistics are 15.41 for none and 3.76 for at most one cointegrating vector.We distinguish three time periods as in Table 1.1.

Trace Statistics for the Number of Cointegrating Vectors



None At Most 1 None At Most 1 None At Most 1

Portugal 12.22∗ 1.55 24.50 5.03 20.28 1.10∗

Italy 15.48 3.02∗ 22.10 2.20∗ 9.89∗ 4.17Ireland 23.50 3.79 18.92 2.93∗ 16.69 5.17Greece 16.66 0.18∗ 25.02 3.49∗ 9.49∗ 0.39Spain 13.16∗ 2.68 17.29 2.90∗ 7.95∗ 2.18


1.3.4 Price Discovery

After we found some support for a long-run equilibrium between the sover-eign CDS and government bond market in the previous section we now turnto the dynamic behaviour of CDS premia and government bond spreadswith a focus on short-run deviations from the equilibrium. As Blanco etal. (2005) point out, an important function of financial markets is price dis-covery, which according to Luetkepohl (2005) can be defined as the efficientand timely incorporation in market prices of information that is implicit inthe trading of investors. The intuition behind this is straightforward. Letus assume there is only one place where an asset is traded. Then, by defin-ition, all price discovery must take place in this market location. However,if there are closely related assets that trade in different market places, thenusually there is fragmentation and the price discovery process is probablysplit among the different market locations.

As discussed earlier, the CDS and government bond markets are closelyrelated in terms of how credit risk is priced. Then, if CDS premia andgovernment bond spreads are cointegrated I(1) variables, the common factorcan be viewed as the implicit efficient price of credit risk (Blanco et al.,2005). Therefore, we first focus on those entities where we found evidencefor cointegration according to Table 1.3. In order to do that we rely on thebivariate VECM that we estimated for the Johansen trace test statistics,where the number of lags to include in the equations is identified again bythe SBIC. The specification of the VECM is as follows:

∆pCDS,t = λ1(pCDS,t−1 − α0 − α1pGBS,t−1)

+p∑

i=1

β1i∆pCDS,t−i +p∑

i=1

δ1i∆pGBS,t−i + ε1t (1.1)

and

∆pGBS,t = λ2(pCDS,t−1 − α0 − α1pGBS,t−1)

+p∑

i=1

β2i∆pCDS,t−i +p∑

i=1

δ2i∆pGBS,t−i + ε2t (1.2)


where ε1t and ε2t are i.i.d. shocks. Two important parameters for ourpurpose are λ1 and λ2. They can be interpreted as follows: If the governmentbond market is contributing significantly to the price discovery process, thenλ1 should be negative and statistically significant. The reason for this is thatin this case the CDS market adjusts to incorporate this information. Usingthe same line of argument, if λ2 is positive and statistically significant thenthe CDS market contributes significantly to the price discovery process. Ifboth coefficients are statistically significant, then both markets contributeto the price discovery process. According to the Granger representationtheorem, the existence of cointegration means that at least one market has toadjust (Engle & Granger, 1987). Adjusting to publicly available informationmeans, however, that this market is reacting more slowly than the other one.Blanco et al. (2005) conclude that the adjusting market is inefficient.

Table 1.4 reports λ1 and λ2 along with the respective p-values. In allseven cases where we found a long-run relation λ2 is positive and significant(at a 1% significance level except for Portugal in Panel C, where we find sig-nificance at the 5% level). This indicates that the CDS market contributesto the price discovery process. By contrast, only in two cases is λ1 negativeand significant at a 10% level of significance – an indication that the govern-ment bond market contributes to price discovery. Overall, we find that infive of the seven cases only the CDS market contributes to price discoverywhile in the other two cases both markets contribute.

According to Gonzalo and Granger (1995), we can use the relative mag-nitudes of the λ coefficients to determine which of the two markets leadsthe price discovery process. The contribution of the CDS market to pricediscovery can be calculated using the Gonzalo-Granger measure, which isdefined as follows.

GG =λ2

λ2 − λ1(1.3)

For the first five cases in Table 1.4, the Gonzalo-Granger measure producesa statistic of one or greater than one which is difficult to interpret. In noneof these cases, however, is λ1 statistically significant. Hence, without loss ofgenerality, we could replace the value of λ1 by zero. For the Gonzalo-Grangermeasure we would then obtain a statistic of one in all cases, which is equiva-lent to stating that only the CDS market contributes to price discovery. For


both Spain and Portugal in Panels B and C, respectively, we find significantλ coefficients with the expected sign. Accordingly, the Gonzalo-Grangermeasure yields values of less than one in both cases.

In the case of Spain we find a value of 0.54, i.e. the CDS market iscontributing 54% to price discovery. Hence, the CDS market is slightlymore dominant than the government bond market. In the case of Portugal,however, we find a value of 0.43, which means that the government bondmarket is contributing slightly more to price discovery than the CDS market.

Table 1.4Contributions to Price Discovery

This table reports various measures of the contribution to the price discovery process forthose entities where the results in Table 1.3 indicate a long-run relation between CDSpremia and government bond spreads. The parameters are estimated via a bivariateVECM. The Gonzalo-Granger measure shows the relative contribution of the CDS premiato the price discovery process. Standard errors are in brackets. For calculating the stand-ard errors of the Gonzalo-Granger measure we use the delta method. Panel A reports theresults for the whole sample since January 9, 2008. Panel B concentrates on the periodprior to the Greek problem, Panel C on the period thereafter.


Contribution of GBS Contribution of CDS Gonzalo-Granger

λ1 (Std. Err.) λ2 (Std. Err.) GG (Std. Err.)

Italy 0.002 (0.007) 0.021 (0.006) 1.08 (0.02)Greece 0.014 (0.013) 0.047 (0.013) 1.43 (0.02)




Italy 0.000 (0.011) 0.043 (0.010) 1.00 (0.01)Ireland 0.011 (0.010) 0.025 (0.006) 1.71 (0.05)Greece 0.014 (0.010) 0.044 (0.009) 1.44 (0.02)Spain -0.024 (0.014) 0.029 (0.010) 0.54 (0.01)




Portugal -0.147 (0.086) 0.112 (0.058) 0.43 (0.02)


Based on the Johansen trace test statistics in the previous section, how-ever, cointegration is rejected for 8 of the 15 cases and therefore the VECMrepresentation is not valid. Accordingly, we cannot use this approach forexamining the price discovery process in these cases. Instead, we rely onthe concept of Granger causality, which is motivated by the approach byBlanco et al. (2005) as well. Since one precondition for performing Grangercausality tests is that the variables are stationary, we use the transformedvariables in first differences. For selecting the number of lags to include inthe VAR equations we looked at three different information criteria: theAkaike Information Criterion (AIC), the Hannan Quinn Information Cri-terion (HQIC), and the Schwartz Bayesian Information Criterion (SBIC).We find that in 87% of the cases the HQIC and the SBIC yield the samenumber of lags. Only in 20% (27%) of the cases, however, does the AICyield the same number of lags as the SBIC (HQIC).

In all other cases the number of lags indicated by the AIC is significantlyhigher than indicated by the SBIC or the HQIC, respectively. As Luetkepohl(2005) demonstrates, the SBIC and the HQIC provide consistent estimatesof the true lag order, while minimising the AIC tends to overestimate thetrue lag order with positive probability. Therefore, we tend to rely either onthe SBIC or the HQIC, respectively. As discussed above, both informationcriteria yield the same lag order in most cases. The SBIC, for instance,yields at most two lags in the case of Greece, and only one lag in all othercases. The results of the Granger causality tests based on the SBIC aresummarised in Table 1.5.

We again distinguish three different time periods. First, we look at thewhole sample from January 9, 2008, to April 16, 2010. The results for thistime period are reported in Panel A of Table 1.5. CDS premia Granger-causegovernment bond spreads for four out the five entities (at a 1% significancelevel). Only in the case of Ireland are we unable to reject the null. However,in this case we found that government bond spreads Granger-cause CDSpremia instead. This is also the case for Portugal and Spain, indicating bi-directional causality. The results for the second time period – from January9, 2008, to October 19, 2009 – are reported in Panel B of Table 1.5. It isinteresting that only the results for Greece change. We now find Granger-causality in the opposite direction, i.e. from government bond spreads toCDS premia (at a 5% significance level).


Table 1.5Granger Causality Test Results

This table reports Granger causality test results. We use first differences, i.e., daily changesin levels, to obtain stationary time series (we performed augmented Dickey-Fuller tests totest for unit roots). For selecting the number of lags to include in the VAR-equations werely on the SBIC. In the case of Greece this yields two lags, in other cases one lag. PanelA reports the results for the whole sample since January 9, 2008. Panel B concentrateson the period prior to the Greek problem, Panel C on the period thereafter.


H0: CDS Do Not Cause GBS H0: GBS Do Not Cause CDS

χ2-Statistic p-Value χ2-Statistic p-Value

Portugal 13.98 0.000 18.46 0.000Italy 16.73 0.000 2.25 0.133Ireland 1.96 0.161 32.45 0.000Greece 11.35 0.001 0.70 0.403Spain 6.56 0.010 16.77 0.000










This changes again if we examine the third time period, from October 20,2009, to April 16, 2010 (Panel C). Moreover, only for two out of the five en-tities did we find that CDS premia Granger-cause government bond spreads.In contrast, government bond spreads Granger-cause CDS premia in fourout of five cases, at least at a 10% significance level.

We find this very interesting given the perception that the turbulencesurrounding the Greece debt crisis was, to a large extent, due to speculationin the CDS market. At least in terms of Granger-causality, this perceptionseems not to hold for the period October 2009 – April 2010.

Overall, the Granger causality test results signal that there is a lot ofpredictability in both instruments while the VECM analysis indicates thatCDS premia contribute more to the price discovery process in the event ofa long-run equilibrium. Also, the results do not appear to be very stableif we compare the two sub-samples of Panel B and Panel C. Still, we thinkthat our results are in line with the findings of Blanco et al. (2005), as theysuggest that bond spreads only react sluggishly to long-term imbalances asmeasured by the cointegrating relationship. In light of this we can concludethat CDS markets are in most cases leading markets if there is a long-runrelation between the CDS and government bond spread markets.

1.3.5 Liquidity Analysis

Now, we can ask whether the potential inefficiency of government bondmarkets in terms of the price discovery process might be related to measuresof market liquidity. The liquidity of a financial instrument is the cost ofopening and closing a position. According to Anderson (2010), liquidityin derivative markets is often better compared to the underlying marketsdue to the higher degree of standardisation. He argues that the liquidity ispositively influenced by “low bid/ask spreads”, “the ability to trade largequantities without having much price impact” and “the speed with which themarket absorbs a large trade.” What is more, the liquidity of sovereign CDSshas increased sharply in the past decade as the market benefited from thestandardisation of contract forms and definitions in 1998 and 1999 as well assuccessful executions in various defaults (e.g., Russia in 1998 and Argentinain 2001). The introduction of an ISDA auction process in 2005 further


smoothed processes in a default case.8 In emerging markets, sovereign CDSsare considered the most liquid credit derivative instruments. According toPacker and Suthiphongchai (2003), sovereign CDSs have the potential tosupplement and increase efficiency in underlying sovereign bond markets astheir liquidity increases. In general, the more liquid a sovereign CDS, themore it shows signs of financial stress. A relatively liquid CDS market is alsoan indication that there is agreement between market participants aboutthe present value, but disagreement about future value due to increaseduncertainty surrounding the country’s fiscal situation.

According to liquidity score data from Fitch Solutions, liquidity on thedeveloped market sovereign CDS index surpassed that of the emerging mar-ket sovereign CDS index for the first time in November 2009. This highlightsthe fact that, on average, the CDS market indicated more uncertainty withrespect to the fiscal situation of developed economies compared to the situ-ation of emerging countries. Although the 10 most liquid sovereign CDSmarkets are all from the emerging market index, overall liquidity in thisindex has only increased marginally compared with the significant increasein the developed market index. The increase in liquidity of the developedmarket index has been driven by persistent market uncertainty about thestrength of economic recovery and the sustainability of fiscal developmentson the back of fiscal stimulus packages and expected lower tax revenues. Forcountries considered safe, the government bond market is in general moreliquid than the sovereign CDS market. A good example for this relationshipis Germany, where the sovereign CDS market is less liquid than the highlyliquid bond market. Dresdner Kleinwort Wasserstein Research (2002) andJP Morgan Research (2001) have found that, generally, bid-ask spreads forcredit default swaps in the more liquid sovereign names are 10 to 20 basispoints wider than those observed in the bond market.

However, for countries in financial trouble, the bond market becomesmore illiquid than the sovereign CDS market. Hence, liquidity shifts towardsCDS markets during distress periods, making them more liquid. Accordingto consistent anecdotal evidence, during the financial crisis, the CDS mar-kets for most PIIGS countries were more liquid in certain phases than the

8So far, there has only been one example of the auction process being used to determinethe recovery rate for sovereign CDS: After the last default by Ecuador in 2008, the auctionsettled at a recovery rate of 31.4%.

1.4 Conclusion 29

equivalent bond market.9 An investment bank10 provided us with a timeseries of bid-ask prices for Greek government bonds. We compared the datawith market bid-ask prices for Greek CDS (CMAN prices). Figure 1.6 showsthat CDS instruments were consistently more liquid than government bondswith the same maturity.11 It is obvious that the more liquid market shouldbe the leading market, since higher liquidity enables market participantsto process information more efficiently (i.e. at lower costs). Hence, thethreshold for acting on new information is lower in the more liquid market.

0

50

100

150

200

Basi

s Po

ints

01/2008 07/2008 01/2009 07/2009 01/2010

Bond Bid−Ask Spread GGB 4.5%

CDS Bid−Ask Spread Hellenic Republic

Figure 1.6: Bid-Ask Spreads

1.4 Conclusion

Motivated by the dramatic developments on the sovereign CDS market inspring of 2010 and the subsequent discussion about the use and abuse ofthis market, we examined the empirical relationship between CDS premiaand government bond spreads in a time-series framework for Portugal, Italy,Ireland, Greece, and Spain. We found some evidence for a long-run relation-9According to traders from the Swiss National Bank and investment banks.10The bank is a major global market maker in the fixed income market. The bank explicitly

requested anonymity.11The pricing depends on the positioning of the individual bank. However, the result is

robust and is based on anecdotal evidence.

1.4 Conclusion 30

ship in the sense of cointegration for the two markets. In most cases (fiveout of seven), only CDS premia contribute to the price discovery process. Inthe remaining cases, both markets contribute more or less equally.

This suggests that bond spreads react only sluggishly to long-term im-balances, as measured by the cointegrating relationship. In light of this, wecan conclude that, in most cases, CDS markets are leading if there is a long-run relationship between the CDS and government bond spread markets.This may be partly due to liquidity effects. However, based on the Granger-causality tests, we also found a reaction to lagged differences between bondspreads and CDS premia, indicating that there is a lot of predictability inboth instruments. In this light, the cointegration-based evidence on marketinefficiency is less conclusive. We think that further research, which wouldinvolve extending the analysis (both the number of countries and the timeperiod), might offer valuable insights in this area.

Still, our results suggest that the sovereign CDS market is potentiallyan enrichment for the financial market community as it appears to be moreliquid than the underlying government bond market during periods of stress.However, it is important to note that, due to the relatively young EuropeanCDS market, our results are based only on the period from January 2007to April 2010. Consequently, the sample period is heavily influenced by theGreek debt crisis. It thus remains to be seen how the European sovereignCDS markets behave in ‘normal’ times.

Chapter 2

Measuring Co-Movements of

CDS Premia During the

Greek Debt Crisis

2.1 Introduction

Since the financial crisis of 2008, the sovereign CDS market in Europe hasbeen growing strongly. The financial crisis caused public deficits to increasemassively due to fiscal stimulus packages, bail-outs and reduced tax revenues.As can be seen in Figure 2.1, however, the trend of increasing public debthad started already in the 1970s. For instance, the average debt-to-GDPratio for the G7 countries had risen from a low of about 30% to around 90%in 2007. Since then, debt-to-GDP has increased by another 20 percentagepoints.

Accordingly, fears about the sustainability of this development, accom-panied by deteriorating credit quality for some countries, stimulated theneeds of market participants to hedge against sovereign default risk. Even-tually, this led to a significant increase in the demand for sovereign CDSs.Prior to 2008, sovereign CDSs were mainly traded for emerging markets.However, since then CDS markets for industrialized countries have also beendeveloped and have been in the limelight several times.

This chapter is based on Andenmatten and Brill (2011b). Both authors contributedequally to this work.

2.1 Introduction 32

0

20

40

60

80

100

120Pe

rcen

t

1940 1960 1980 2000

G7 Advanced EconomiesEmerging and Developing Economies World

Source: IMF

Figure 2.1: Debt-to-GDP Ratio since 1950

According to the Fitch Solutions liquidity index1, the liquidity in Europeansovereign CDS surpassed the liquidity of Latin American emerging econom-ies in several periods after November 2009. In general, as stated by Markit,the liquidity of a credit derivative asset increases when the underlying showssigns of financial stress in combination with a significant amount of debt out-standing and/or changes in its capital structure, including new issuance.2

Entities also tend to be more liquid when there is agreement about presentvalue but disagreement about future value due to heightened uncertaintysurrounding the entity.

Broadly, CDSs are bilateral contracts used to transfer risk among mar-ket participants and are basically defined by four parameters: the referenceentity, the notional amount, the price (spread or premium), and the matur-ity. One participant is the so-called protection buyer, who wants to buyinsurance against the default of a specific entity, the so-called reference en-tity. The other party is the protection seller, who writes the insurance onthe reference entity. To compensate the seller of the insurance for the as-sumed risk, the protection buyer pays an initially fixed spread every year1As stated by Fitch Solutions, the liquidity index measures “are derived from a proprietarystatistical model which produces a liquidity score for each credit derivative asset bymodelling a broad set of information taken from the CDS market.”

2Interestingly, CDS liquidity for France, Spain and Portugal has consistently been greaterthan for Ireland.

2.1 Introduction 33

(or each quarter) on the insured notional value. If a credit event occurs, theCDS is triggered and the protection seller has to pay the difference betweenthe insured notional value and the recovery value.3 Since 2005, an auctionprocess has been instituted and settlement is almost always made throughan auction (either cash or physical delivery), i.e. investors are signed upautomatically for all auctions.

A CDS makes it possible to invest in the credit quality of a corporateor a sovereign. If an investor believes that the credit quality of a corporatewill decrease in the coming months and that this is not yet priced into thecurrent spreads, he should buy protection. Once the spreads widen, he willprofit because his insurance will increase in value. He can close the insurancecontract whenever he wants and monetize his gains. Thus, buying protectionon Germany does not mean that somebody is speculating on the countrygoing bankrupt. It merely means that somebody believes the credit qualityof Germany will decrease in the future. At the same time, the seller ofthe protection on Germany believes that – given the current spreads – it isattractive to agree to the contract.

The so-called PIGS countries (i.e. Portugal, Ireland, Greece, Spain) havefrequently been in the focus of financial markets and the media, especiallysince October 2009 when Greek officials announced that debt statistics hadbeen forged. As a consequence, financial market participants respondedquickly to the deterioration in fiscal positions by requiring higher sovereigndefault risk premia not only for Greece but also for many other countries witha seemingly unsustainable fiscal situation. This supports previous studiesthat had shown that sovereign risk premium differentials tend to co-moveover time and are mainly driven by a common time-varying factor, which canbe interpreted as a repricing of global risk factors (cf., for instance, Codogno,Favero, & Missale, 2003; C. Favero, Pagano, & Thadden, 2007; and Geyer,Kossmeier, & Pichler, 2004). In line with that, in an extensive study of26 developed and emerging-market countries Longstaff et al. (2011) find forthe period 2000-2007 that sovereign credit risk premia were generally morerelated to the U.S. stock market and high-yield bond markets, global riskpremia, and capital flows than they were to local factors. Accordingly, an

3The so-called ISDA Credit Derivative Determination Committee – consisting of buy andsell side members – will decide whether the requirements for a credit event are fulfilled.The decision of the determination committee is binding for the whole market.

2.1 Introduction 34

investment in sovereign credit is to a large extent a compensation for bearingglobal risk and there is little or no country specific premium. However, theirstudy does not include a sovereign debt crisis or the sort of credit events weexperienced in the 1990s.

In contrast, Sgherri and Zoli (2009) find evidence that since October 2008markets have become progressively more concerned about the potential fiscalimplications of national financial sectors’ frailty and future debt dynamics,which would imply that sovereign credit risk premia are driven more bynational than global factors.

The recent focus of financial market participants on the fiscal situationin the PIGS countries provides an opportunity to study the developmentsof CDS premia for “hot-spot” countries. We now have a better data basisfor studying the co-movement of CDS premia across countries and regionsand testing whether the co-movement of sovereign CDS premia increasedsignificantly after the Greek debt crisis that started in October 2009 – adevelopment which is usually referred to as contagion. However, there doesnot seem to be any agreement on what contagion exactly means (Rigobon,2002) and how it manifests itself. For instance, Forbes and Rigobon (2001)declare that “there is no consensus on exactly what constitutes contagion orhow it should be defined.” In the following analysis, we will define contagionin the restrictive way proposed by Forbes and Rigobon (2002). They definecontagion as a significant increase in cross market linkages after a shock toone country.

In the remainder of the chapter we proceed as follows: In the next section,we discuss the concept of contagion and describe a bivariate test procedurethat was originally proposed by Forbes and Rigobon (2002). In section 2.3,we apply this test procedure to our sample of daily data between October2008 and July 2010 for 39 countries including both emerging and indus-trialized countries. In doing so, we test for contagion stemming from theGreek CDS market (country level analysis) and the regional CDS marketfor the PIGS countries (regional level analysis), respectively. In section 2.4,we attempt to explore the common factor and perform principal compon-ent analysis in order to analyse the degree of commonality of inner-regionalvariation of CDS premia. Finally, section 2.5 concludes.

2.2 Propagation of Shocks: Contagion vs. Interdependence 35

2.2 Propagation of Shocks: Contagion vs. Interdependence

On October 4, 2009, George Papandreou became the new prime ministerof Greece after his Panhellenic Socialist Movement (PASOK) party wonthe general election. At that time, the Greek economy was still faced withthe severe repercussions of the financial crisis. Around two weeks later, onOctober 20, officials from the new government announced that Greek debtstatistics had been forged in the past. Instead of a public deficit of 6% ofGDP for 2009 the government now expected twice as much to materialize.This was the starting point of the Greek debt crisis that led to radical aus-terity packages for the Greek economy and an international rescue package.

The crisis also led to a strong increase in risk premia for Greek sovereigndebt as reflected, for instance, in CDS premia. While the CDS premium fora Greek government bond with a 5-year maturity and a notional value ofUSD 10 million was 124 basis points on October 20, 2009, it soared to 1012basis points by May 7, 2010. This means that the insurance costs againsta default of this particular Greek government bond increased by a factor ofmore than 8 times, i.e. from USD 124,000 to more than USD 1 million. Atthe same time, CDS premia for many other countries increased strongly aswell.

0

200

400

600

800

1000

1200

Basi

s Po

ints

10/2008 04/2009 10/2009 04/2010

Greece PortugalSpain Ireland

Source: Thomson Reuters Datastream

Figure 2.2: CDS Premia in Basis Points


Figure 2.2 illustrates this by comparing the development of the CDS premiafor the PIGS countries. As can be seen, the increase for Greek CDS premiawas strongest, but these dramatic movements were mirrored in the otherthree CDS markets as well.

This shows that dramatic events in one market can have strong im-pacts on other markets. The question then is whether a high degree of co-movement during times of crisis already constitutes contagion? Or does this,as Forbes and Rigobon (2002) argue, rather reflect the fact that global mar-kets are “so interdependent that they have similar high rates of co-movementin all states of the world?” Before we can discuss these questions for theGreek debt crisis in section 2.3, we first introduce the theoretical frameworkof what contagion constitutes as well as an empirical test procedure basedon the approach from Forbes and Rigobon (2002).

2.2.1 Theory and Literature Review

According to Dornbusch, Park, and Claessens (2000), reasons for contagioncan be divided into two groups: on the one hand fundamental-based reasonsand on the other hand investor behaviour-based reasons. While fundamental-based contagion works through real and financial linkages across countries,behaviour-based contagion is more sentiment-driven. It seems reasonableto assume that during a financial crisis both types of contagion are present:Firstly, fundamental-based because of the strong interrelationship of finan-cial sectors. For instance, during the Greek debt crisis it became apparentthat European banks had significant exposure to Greek government bonds.Hence, a potential restructuring of Greek bonds increased the probability ofbail-out packages in different European countries. Secondly, as discussed inDornbusch et al. (2000), investor behaviour-based contagion usually takeseffect through liquidity and incentive problems, as well as information asym-metries and coordination problems. As Dornbusch et al. (2000) stated:

In the absence of better information to the contrary, investorsmay believe that a financial crisis in one country could lead tosimilar crises in other countries. A crisis in one country maythen induce an attack on the currencies of other countries inwhich conditions are similar. This type of behaviour can reflectrational as well as irrational behaviour. If a crisis reflects and re-


veals weak fundamentals, investors may rationally conclude thatsimilarly situated countries are also likely to face such problems;such reasoning helps explain how crises become contagious. Thischannel presumes, of course, that investors are imperfectly in-formed about each country’s true characteristics and thus makedecisions on the basis of some known indicators, including thoserevealed in other countries, which may or may not reflect thetrue state of the subject country’s vulnerabilities. The inform-ation investors use may include the actions of other investors,which brings us to the effects of informational asymmetries oninvestor behaviour.

There is an extensive literature on the potential reasons and the transmis-sion channels of contagion4 as well as on theoretical modelling of contagion.5

However, little is yet known about the transmission channels and their rel-ative importance.

In addition, there is also a large body of literature that focuses on em-pirical tests for the existence of contagion in a certain stress period, thatis, if there are stronger cross-market linkages in times of crisis. This studybelongs to the latter group. In what follows, we will focus on testing for theexistence of contagion during the Greek debt crisis.

So far, however, no unifying framework of testing for the existence ofcontagion during financial crises has been agreed upon. Instead, a broadrange of different methodologies has been developed. Dungey, Fry, Gonzalez-Hermosillo, and Martin (2005) offer an extensive review of these methodo-logies as well as their empirical application for equity markets in 1997-98during the Asian crisis.6 According to the authors of this study, the factthat there are so many different methodologies in use makes the assessmentof contagion difficult. This seems to be particularly the case for assessingthe significance in transmitting crises between countries.

In what follows, we focus on the approach of Forbes and Rigobon (2002),which builds on a correlation analysis. The motivation for focusing on this

4Cf., for instance, Van Rijckeghem and Weder (2001) and Caramazza, Ricci, and Salgado(2000).

5Cf., for instance, Allen and Gale (2000), G. A. Calvo and Mendoza (2000), Chue (2002),Kodres and Pritsker (2002), and Kyle and Xiong (2001).

6Cf. also Dornbusch et al. (2000) and Pericoli and Sbracia (2003) for an overview of theliterature.


approach is based on the above-mentioned survey from Dungey et al. (2005).In their empirical application of different methodologies Dungey et al. clas-sify the approach from Forbes and Rigobon (2002) as a conservative testas it did not yield any evidence of contagion for equity markets during theAsian crisis in 1997-98. Accordingly, finding evidence for contagion duringthe Greek debt crisis with such a conservative test would then be a strongersignal than finding evidence for contagion with a less conservative test.

The basic idea of this approach is to test whether the correlation betweentwo variables increases significantly during a crisis period. However, onehas to be careful when comparing correlation coefficients between differentperiods because, as Boyer, Gibson, and Loretan (1997) and Forbes and Rigo-bon (2002) show, correlation coefficients between markets are conditional onvolatility. Hence, during times of increased volatility (i.e. in times of crisis)estimates of correlation coefficients are biased upward.7 If co-movementtests are not adjusted for that bias, contagion is too easily detected.

A good example for the misleading nature of an uncorrected bias is thepaper by King and Wadhwani (1990), which was the first major analysisfocusing on co-movement analysis. Here contagion is defined as a significantincrease in the correlation coefficient. The paper detects contagion betweeninternational stock markets after the U.S. market crash in 1987. However, asForbes and Rigobon (2002) show, when cross market correlation coefficientsare adjusted for heteroscedasticity, there is no longer a significant increasein these correlation coefficients.

Before discussing test procedures for contagion, we first examine poten-tial channels through which correlation between sovereign credit risk and,hence, CDS premia could arise. According to a framework presented byLongstaff et al. (2011) that builds on standard arbitrage arguments8 onecan distinguish between three different channels.

One channel might arise through the correlation between the arrival ratesof credit events. This might be induced by a deterioration in the economicsituation of the countries in question. For example, the financial crisis of2008 led to a severe economic slowdown around the globe, dragging manycountries into the most severe recession since World War II. In an attempt to

7Forbes and Rigobon (2002) use a numerical example to show how heteroscedasticity canbias a correlations estimator upward.

8As discussed, among others, in Duffie and Singleton (1999), Dai and Singleton (2003),and Pan and Singleton (2008).


stimulate their economies, many governments approved huge rescue packagesat the cost of soaring public deficits. Eventually, investors have started toquestion the ability of some governments to serve their debt. This could beinterpreted as an expected increase in the arrival rate of credit events forsome of the countries.

Another channel might arise through the correlation between loss ratesgiven a default. This could reflect a worsening of the bargaining situationof creditors due to a deterioration of the economic situation, political tur-moil, and legal disputes. As a consequence, governments might face higherrefinancing costs.

A third channel might arise through liquidity effects. This can alsobe illustrated by the recent financial crisis and the flight to quality thattook place after the collapse of the investment bank Lehman Brothers (cf.Longstaff (2004) for an analysis of flights to quality). Other forms of li-quidity effects might stem from higher trading costs on illiquid securities forwhich investors want to be compensated (cf. Amihud, Mendelson, & Ped-ersen, 2006) and from the possibility that investors are subject to marginrequirements (cf. Liu & Longstaff, 2004). As a consequence, these fundingproblems might lead to market illiquidity (Pedersen & Brunnermeier, 2007).

2.2.2 A Bivariate Test of Contagion

As mentioned in the previous section, correlation is conditional on volatility.Hence, in times of stress correlation coefficients are biased upward. Forbesand Rigobon (2002) present a statistical correction for this conditioning biasand the appropriate procedure to test for contagion (henceforth referred toas the FR-test), which we discuss in this section. We focus on two versionsof the bivariate FR-test: that originally developed by Forbes and Rigobon(2002) and another suggested by Dungey et al. (2005). We base our notationon that of Dungey et al. (2005).First, we define different sample periods:

x : a period before the crisis

y : a crisis period

z : the entire sample period


Moreover, we define the following parameters of volatility and correlation:

σ2x,i : volatility of country i′s CDS premia before the crisis (i = 1, 2)

σ2y,i : volatility of country i′s CDS premia in the crisis (i = 1, 2)

ρx : correlation between countries 1 and 2 in period x

ρy : correlation between countries 1 and 2 in period y

ρz : correlation between countries 1 and 2 in period z

Forbes and Rigobon (2002) show that the standard (unadjusted) correlationcoefficient is conditional on the variance in the two asset markets. Accord-ingly, if there is an increase in volatility in country 1 during times of crises, i.e.σ2

y,1 > σ2x,1, it would be misleading to suppose contagion if ρy > ρx. Hence,

we have to correct for the upward bias. Forbes and Rigobon quantify thisbias and show that the adjusted (unconditional) correlation is given by:9

νy =ρy√

1 + [(σ2y,1 − σ2

x,1)/σ2x,1](1 − ρ2

y)(2.1)

As can be seen from (2.1), the unconditional correlation coefficient, νy, isthe unadjusted correlation coefficient, ρy, scaled by a function of the changein volatility in asset returns of the source country over the high and lowvolatility periods. To illustrate this, assume that σ2

y,1 > σ2x,1, i.e. that the

volatility of asset returns in country 1 increases from period x to period y.Then, (σ2

y,1 − σ2x,1)/σ2

x,1 > 0 and for any ρy ∈ (−1; 1) it follows that

√1 + [(σ2

y,1 − σ2x,1)/σ2

x,1](1 − ρ2y) > 1 (2.2)

This implies that νy < ρy. From (2.1) it is also immediately apparent thatνy = ρx if there is no fundamental shift in the relationship between the twoasset markets from the low to the high volatility period.

Accordingly, we can formulate the null hypothesis and the alternativehypothesis, respectively, of a test that there is a significant increase in thecorrelation coefficient in the high volatility period, i.e. that there is conta-

9For a similar approach, cf. Boyer et al. (1997), Corsetti, Pericoli, and Sbracia (2001),Corsetti, Pericoli, and Sbracia (2005), and Loretan and English (2000); for alternativeapproaches, cf. Karolyi and Stulz (1996), and Longin and Solnik (1995).


gion, as follows:

H0 : νy = ρx (2.3)

H1 : νy > ρx (2.4)

The t-statistic can be used to test this hypothesis where the t-statistic isgiven by

t =νy − ρx√

V ar(νy − ρx)(2.5)

and where νy and ρx mark the sample estimators of νy and ρx, respectively.If we assume that the two samples are drawn from independent normaldistributions we can transform the standard error in (2.5) as follows:

V ar(νy − ρx) = V ar(νy) + V ar(ρx) − 2Cov(νy, ρx) (2.6)

= V ar(νy) + V ar(ρx) (2.7)∼= (1/Ty) + (1/Tx) (2.8)

where Tx (Ty) is the sample size of the low (high) volatility period. Toget to (2.7) we use the independence assumption and (2.8) follows from theassumption of normality and an asymptotic approximation.10 With that weget

FR =νy − ρx√

(1/Ty) + (1/Tx)(2.9)

In a next step, Forbes and Rigobon (2002) suggest using the Fisher trans-formation, as this improves the finite sample properties of the test statistic.11

This yields

FR =1/2[log((1 + νy)/(1 − νy)) − log((1 + ρx)/(1 − ρx))]√

(1/Ty − 3) + (1/Tx − 3)(2.10)

When applying (2.10) the respective low and high volatility periods are

10For this asymptotic approximation Dungey et al. (2005) refer to Kendall and Stuart(1969).

11According to Dungey et al. (2005), the Fisher transformation is valid for small values ofboth the unadjusted and the adjusted correlation coefficient.

2.3 Contagion during the Greek Debt Crisis 42

separated from each other, that is, the two periods do not overlap, as wassuggested by Dungey et al. (2005). However, in the original test statisticfrom Forbes and Rigobon (2002) for testing (2.3) against (2.4) the non-crisisperiod is defined as the whole sample period z, with the non-crisis and thecrisis periods overlapping. Accordingly, (2.9) would be formulated as

FR =νy − ρz√

(1/Ty) + (1/Tz)(2.11)

and the Fisher-adjusted version (2.10) as

FR =1/2[log((1 + νy)/(1 − νy)) − log((1 + ρz)/(1 − ρz))]√

(1/Ty − 3) + (1/Tz − 3)(2.12)

As shown, the test statistics (2.11) and (2.12) build on the assumption thatthe variances of νy and ρz are independent. This assumption is violated,however, if the two sample periods overlap. As a result, the covariance termin (2.6) is most likely not equal to zero and the step from (2.6) to (2.7)results in a standard error that is too large, as the negative covariance termis not taken into account. As stated by Dungey et al. (2005), this leadsto a downward bias in the t-statistic and, hence, to fewer rejections of thenull hypothesis. Nevertheless, in what follows we will apply both versions inorder to see if this downward bias leads to fundamentally different results.In terms of notation, we will refer to the overlapping version (2.12) suggestedby Forbes and Rigobon (2002) as the FRO-test, and to the non-overlappingversion suggested by Dungey et al. (2005) as the FRN -test.

2.3 Contagion during the Greek Debt Crisis

In this section, we apply both versions of the FR-test in order to test forcontagion during the Greek debt crisis: first, the original version (2.12) byForbes and Rigobon (2002) with overlapping data (FRO); and second, thealternative version (2.10) with non-overlapping data (FRN ) suggested byDungey et al. (2005). We perform the tests not only on a national butalso on a regional level after constructing various regional aggregates. Oursample is based on daily data running from October 1, 2008, through July27, 2010.


2.3.1 Between Countries

Starting on the national level, Table 2.1 lists basic descriptive informationand the number of observations for CDS premia for 39 countries. We useCDS premia from Thomson Reuters with a notional value of USD 10 million.All prices are based on the standard ISDA contract for physical settlementwith a constant 5-year maturity and are expressed in basis points. As canbe seen, the average values of the CDS premia range widely across countriesand the median is lower than the mean in all 39 cases. Argentina is thecountry with the highest mean of 2023 basis points, followed by Ukraineat 1718 basis points. At the other end of the range we find Germany andthe United States with means of just 15.3 and 16.8 basis points, respectively.What is more, both the standard deviations and minimum/maximum valuesindicate that many countries experienced a strong variation in their CDSpremia over time. For example, the CDS premia for Greece range from aminimum of 59.5 basis points to a maximum of 1037.6 basis points – morethan 17 times the minimum value.

We define October 20, 2009 as the start of the crisis. On this day, of-ficials from the new government in Greece announced irregularities in theGreek debt statistics. However, this choice is somewhat arbitrary, since thereliability of the debt statistics was widely doubted before the irregularitieswere officially confirmed. In general, defining an exact crisis period seemserratic based on the myriad events surrounding the Greek debt crisis. Sim-ilar issues were already recognized by Forbes and Rigobon (2002) during theAsian crisis in the late nineties. When did the crisis period start? Whendid it end? As Forbes and Rigobon (2002) state, there was “no single eventwhich acts as a clear catalyst behind this turmoil.”

To demonstrate these difficulties, we take a look at the developmentsin Greece since the outbreak of the crisis and their influence on the Greek5-year CDS. The key events are marked in Figure 2.3. In November 2009,Greece announced its update on the initial budget for 2008. The deficit was,as mentioned previously, more than twice as much as in the initial budgetpresented in December 2008. Against the background of spreading negat-ive market talk, CDS spreads soared in the following weeks, which led toincreasing refinancing costs on bond markets. After a period of reassurance,in spring 2010 the situation worsened again.


Table 2.1Descriptive Statistics of CDS Premia

This table lists basic descriptive information and the number of observations for CDSpremia in our sample. We use daily CDS premia from Thomson Reuters with a constant5-year maturity. The data run from October 1, 2008, through July 27, 2010. It is worthmentioning that in all cases the median is lower than the mean. Argentina is the countrywith both the highest mean and median, followed by Ukraine. At the other end of therange we find Finland, Germany and the United States.

Obs. Mean S.D. Min. Median Max.

Argentina 459 2023.4 1220.6 803.3 1536.1 4841.8Austria 460 96.7 44.5 19.2 83.0 265.0Belgium 459 72.2 32.1 30.0 62.8 158.0Brazil 460 204.9 104.3 109.3 140.7 606.3Bulgaria 460 340.4 135.2 174.5 300.6 692.7Chile 453 129.6 68.9 48.5 94.8 310.0China 460 111.2 57.5 57.5 82.0 284.0Columbia 460 231.3 108.4 123.7 168.8 668.7Denmark 458 56.8 32.7 23.2 41.9 146.0Estonia 453 313.9 194.6 90.0 236.0 732.5Finland 455 36.2 17.1 14.5 31.0 94.0France 459 48.6 20.9 21.0 44.0 98.7Germany 460 37.5 15.3 12.2 34.1 92.5Greece 460 297.3 234.3 59.5 223.3 1037.6Hungary 460 321.1 115.4 165.6 305.9 630.7Indonesia 458 352.9 228.9 147.5 215.0 1240.0Ireland 455 189.4 62.6 63.0 172.0 390.0Israel 454 145.3 44.5 99.0 122.5 282.5Italy 460 122.5 43.4 50.0 111.9 251.7Japan 453 63.0 21.2 18.0 64.0 120.0Kazakhstan 455 482.2 350.6 157.0 351.1 1634.1Malaysia 457 145.9 79.1 69.5 105.0 500.0Mexico 460 217.4 107.2 101.2 163.7 613.1Netherlands 458 51.6 26.4 25.0 42.9 130.0Peru 460 211.5 110.8 107.3 145.9 664.3Philippines 456 255.5 119.1 142.5 193.5 840.0Poland 455 175.8 74.1 86.0 146.8 421.0Portugal 460 127.0 85.8 45.0 93.5 466.5Qatar 460 152.4 75.6 76.2 115.0 390.0Romania 460 380.3 162.3 188.5 308.9 767.7Russia 454 355.1 237.4 124.0 264.2 1106.0South Korea 459 194.0 121.5 73.0 138.0 680.0Spain 460 117.1 51.9 47.2 100.6 270.2Sweden 457 64.3 31.8 21.0 52.2 159.0Thailand 460 158.4 76.6 75.0 120.8 500.0Turkey 455 276.4 126.1 155.5 208.8 835.0UK 460 82.6 27.0 27.5 79.1 165.0Ukraine 449 1718.0 1104.7 494.5 1363.0 5300.4USA 443 42.9 16.8 19.7 38.8 95.0


Budget update for 2008 deficit

EU announces EUR 750 billion safety net

S&P downgrades Greece to junk

Papandreou asks for activation of an EU/IMF aid package

EU signals willigness to intervene if bond markets freeze

EU/IMF announce EUR 110 billion bail−out for Greece

0

200

400

600

800

1000

1200Ba

sis

Poin

ts

10/2008 04/2009 10/2009 04/2010

Greek CDS Premia

Source: Bloomberg, Thomson Reuters Datastream

Figure 2.3: Greek CDS Premia and Key Events During the Greek Debt Crisis

A weak bond auction in April 2010 fuelled fears that bond issuing was closeto a standstill. This forced euro area leaders to signal their willingness tosupport Greece in case refinancing became blocked. This announcement,however, did not succeed in calming the markets. CDS spreads soon sky-rocketed, forcing Papandreou to activate the EU/IMF aid package.

On April 27, S&P downgraded Greek government bonds to junk. At thebeginning of May, the EU and the IMF were forced to announce a EUR 110billion bail-out for Greece. One week later, against the background of fur-ther rising spreads across European peripheral countries, the EU announceda EUR 750 billion safety net for potentially troubled states. This had a tem-porarily calming effect on markets. However, doubts on the sustainability ofthe bailout for the Greek economy and a general worsening outlook for theEuropean peripheral countries, the Greek CDS in the summer 2010 againsoared to new record levels.

This chronologically aggregated summary demonstrates the pulsatingnature of a typical crisis, where periods of reassurance and stress alternate.Hence, it is reasonable to assume that during 2009/2010 there were severalcontagious periods – not just one “enduring” period of contagion.

Based on the difficulty of defining a fixed crisis period, we suggest en-hancing the Forbes and Rigobon approach by applying the test on rollingwindows of periods of turmoil. In the case of the Hong Kong crash, Forbes


and Rigobon (2002) defined the period of relative stability as lasting almost22 months, namely from January 1, 1996 to October 16, 1997, and the periodof turmoil as the month starting on October 17, 1997. As discussed earlier,we put the start of the Greek debt crisis at October 20, 2009. Accordingly,our period of relative stability lasts from January 1, 2008 to October 19,2009, i.e. also almost 22 months.

However, the period of turmoil might have lasted from October 20, 2009,to the end of our sample, namely July 27, 2010. But even then, one mightargue, the debt crisis has not come to an end – as the events in Irelanddemonstrate. Instead, the discussion of key events above illustrates thatthe period after October 20, 2009 can be divided into various sub-periodsof relative turmoil and periods of relative calm. But even accounting forthis makes it almost impossible to accurately define fixed periods of turmoil.Therefore, we define rolling windows of relative turmoil lasting 20, 40, and 60days, respectively. The 20-day window corresponds to the 1-month periodof relative turmoil defined by Forbes and Rigobon (2002) as 20 business daysare a common proxy for a calendar month. The 40-day and 60-day windowsaccount for the long-lasting nature of the Greek debt crisis.

min

max

min time max

Period of Relative Stability

Crisis Windows

2-D

ay-M

ovin

g Av

erag

e of

D

aily

Cha

nges

in C

DS

Prem

ia

Period ofPotential Crisis

Figure 2.4: Illustration of the Rolling FR-Test Approach


Before applying the tests, we calculate first differences, that is, daily changes,of all variables in order to transform the time series into stationary ones.Augmented Dickey-Fuller tests confirm that the variables in first differencesare all stationary. Also, as is common practice when testing for contagion, wecalculate 2-day-moving averages of the daily changes in the CDS premia.12

This accounts for the problem that financial markets in different countriesare not open at the same time.

We then apply both versions of the FR-test on rolling crisis-windows:first, the original version with overlapping data (FRO); and second, thealternative version with non-overlapping data (FRN ) suggested by Dungeyet al. (2005). For every single crisis window we test for contagion on a 5%level of significance and count the number of such signals over the test period.Figure 2.4 illustrates this.

However, for identifying the signals, we cannot rely on the critical t-values of a standard one-sided t-test. Similar to testing for a structuralbreak at an unknown break date, where the so-called sup F-statistic13 is thelargest of many F-statistics and, hence, its distribution is not the same as anindividual F-statistic, the distribution of the FR-test statistic is not the sameas the standard t-distribution. Based on this, we used Monte Carlo methodsto find approximate critical t-values.14 The results are shown in Table A.1in the Appendix. As can be seen, the critical values for the FR-tests arelarger than the one for a standard one-sided t-test.

Table 2.2 reports the number of signals for contagion stemming from theGreek CDS market based on the transferred variables and the rolling-crisis-window approach. As can be seen, we obtain signals for contagion for bothversions of the FR-test as well as for the three different time windows, ifwe use the critical values from Table A.1. We obtain the largest number ofsignals, namely for 26 out of 38 cases, for the overlapping test, FRO, for the20-day window.

12For instance, cf. Corsetti et al. (2001), Dungey et al. (2005), and Forbes and Rigobon(2002).

13The term sup F-test is only one of many different ones that are in use for this approach.The idea was originally proposed by Quandt (1960) and, accordingly, the approach isoften called the Quandt likelihood ratio (QLR) test. Another term in use is sup-Waldtest. For an introduction to this approach, the interested reader is referred to Stockand Watson (2007). For more details, Perron (2005) offers a review of the literature ondealing with structural breaks.

14We describe the applied methodology in more detail in Appendix A.


Table 2.2Forbes and Rigobon Tests

This table records the number of contagion signals stemming from the Greek CDS marketbased on the bivariate Forbes and Rigobon (2002) approach. We test for rolling crisis-period windows starting on October 20, 2009. We apply two version of the test. First,the original version with overlapping data (FRO). Second, an alternative version withnon-overlapping data (FRN) suggested by Dungey et al. (2005). The signals are based ona 5% level of significance where we use the approximated critical values from Table A.1.

FRO FRN

20 days 40 days 60 days 20 days 40 days 60 days

Argentina 0 0 0 0 0 0Austria 21 12 1 17 3 0Belgium 6 4 6 6 0 0Brazil 4 0 0 0 0 0Bulgaria 0 0 0 0 0 0Chile 3 0 0 0 0 0China 0 0 0 0 0 0Columbia 4 0 0 0 0 0Denmark 10 6 7 7 3 4Estonia 0 0 0 0 0 0Finland 6 3 3 3 0 0France 7 9 10 5 5 5Germany 15 15 15 6 2 3Greece 0 0 0 0 0 0Hungary 3 0 0 0 0 0Indonesia 0 0 0 0 0 0Ireland 12 13 2 1 1 0Israel 0 0 0 0 0 0Italy 22 3 0 21 2 0Japan 0 0 0 0 0 0Kazakhstan 18 27 17 5 13 0Malaysia 0 0 0 0 0 0Mexico 4 0 0 0 0 0Netherlands 19 7 8 9 0 0Peru 4 0 0 0 0 0Phlippines 0 0 0 0 0 0Poland 2 3 9 0 0 0Portugal 15 2 0 9 0 0Qatar 3 4 4 0 0 0Romania 0 5 0 0 0 0Russia 5 5 5 1 0 0South Korea 0 0 0 0 0 0Spain 23 6 0 13 3 0Sweden 4 5 6 0 0 0Thailand 0 0 0 0 0 0Turkey 5 2 3 0 0 0UK 1 0 0 1 0 0Ukraine 1 0 0 0 0 0USA 4 4 4 0 0 0


If we focus on the relative frequency of the signals across countries, we findthat European countries dominate. For instance, for Spain, Portugal, andIreland, we get 23, 15, and 12 signals, respectively. However, it is not onlythe CDS markets of the PIGS countries that seem to be affected by contagionstemming from the Greek market, but also the CDS markets of Germany,France, Italy, the Netherlands, and Austria. Outside the European Union,we obtain most signals for countries in Central and Eastern Europe, such asKazakhstan, Turkey, and Russia. In Latin America, the number of signalsis on average much lower. For Asian countries, we get only a few signals onaverage and for some countries such as South Korea and the Philippines nosignals at all.

If we focus on the 40- and 60-day windows, we find that the number ofsignals decreases across countries but the regional pattern seems to be similarto the 20-day window. With the exceptions of Kazakhstan and Qatar, wefind that the signals almost completely break down for countries outside theEuropean Union. Accordingly, one conclusion to draw from the FRO-testis that we not only find evidence for contagion stemming from the GreekCDS market but also a strong regional pattern. This observation holds truefor the FRN -test as well if we focus on the 20-day window. The FRN -testseems to be more restrictive compared to the FRO-test, as the number ofcountries for which we obtain at least one signal decreases from 26 to 14,and the average number of signals for countries with at least one signaldeclines from 8.5 to 7.4. As discussed in the previous section, this had tobe expected as the standard errors of the FRO-test are likely to be biasedupward due to the overlapping nature of the periods of relative turmoil andrelative stability. This becomes even more apparent if we concentrate onthe 40-day and 60-day windows. For instance, for the 60-day window theFRN -test only yields signals for 3 countries while the FRO yields signals for15 countries.

Overall, our results for CDS markets during the Greek debt crisis con-trast with the results from Forbes and Rigobon (2002) for equity marketsafter the Hong Kong crash and their conclusion of “no contagion, only inter-dependence.” With the term “interdependence”, Forbes and Rigobon (2002)refer to a situation where there is no significant increase in the adjustedcorrelation coefficient between two markets but a continued high level of co-movement. In their view, this “continued high level of market co-movement


suggests strong real linkages between the two economies.” Especially forEuropean countries we would instead conclude “both contagion and inter-dependence.”

It seems that during the Greek debt crisis there were not only periods ofinterdependence but also periods which were characterized by a significantincrease in the co-movement of sovereign credit risk as measured in CDSpremia. This is especially interesting as the approach of Forbes and Rigobon(2002) is, according to Dungey et al. (2005), a conservative test. Theirfindings are based on a comparison of various contagion tests during theAsian crisis. What is more, even Forbes and Rigobon (2002) state that theirresult is “controversial”.

We think that finding evidence for contagion for such a restrictive and,hence, controversial test is a very strong result. One underlying reasonfor finding evidence for contagion for CDS markets during the Greek debtcrisis might be that the developments after the collapse of the investmentbank Lehman Brothers had demonstrated how interconnected financial mar-kets and the world economy are nowadays. Accordingly, the risk of Greecebecoming a new “Lehman” might have led to contagion rather than justinterdependence.

2.3.2 Between Regions

Motivated by the regional pattern of contagion signals we found in the previ-ous section, we also aim to test for contagion on a regional level. We base ouranalysis on the findings of Longstaff et al. (2011), who performed a clusteranalysis to identify significant commonality in sovereign credit spreads on anaggregated level. However, in contrast to Longstaff et al. (2011), we do notconstruct clusters ex post based on the pairwise correlations in our sample.Instead, we construct ex ante regional aggregates. The motivation of thisapproach is threefold.

First, our data sample covers more countries from various regions thanthe one in Longstaff et al. (2011). Also, our sample is more balanced betweenindustrial and emerging countries. While in the sample of Longstaff et al.(2011) Japan is the only industrialized country out of 26 countries, in oursample 17 out of 39 countries are industrialized. This allows us to construct abroad range of both geographical and political aggregates. Second, Longstaffet al. (2011) find that there is a strong regional structure in their cluster


analysis. In particular, the first cluster is dominated by Eastern Europeand the Mediterranean, the second by Asian countries, the third again byEastern Europe, the fourth by Latin America, and the sixth by the MiddleEast. Thus it seems plausible to construct aggregates along regional lines.Finally, the focus of investors during the Greek debt crisis first on the PIGScountries and later on the euro area as a whole indicates a strong interestin particular aggregates which we would like to address directly.

Accordingly, we construct various aggregates based both on geographicaland political criteria. These aggregates are constructed as unweighted aver-ages of CDS premia of countries belonging to the selected region. Table 2.3gives an overview of these aggregates and the countries that were included forconstructing them. The aggregates are the following: the European Union(EU), the European Monetary Union (EMU), the PIGS countries, Centraland Eastern Europe (CEE), the Middle East (ME), Asia, Latin America(LATAM), and an aggregate for the USA, Japan, and Germany (G3).

Based on the national-level approach, we calculate first daily changesof the regional aggregates and then 2-day-moving averages. Panel A ofTable 2.4 reports correlation coefficients between the CDS premia in firstdifferences of the regional aggregates for the whole sample. We excludedthe PIGS countries for the regional aggregates of the EU, the EMU, andthe world aggregate. In general, one can see that the pairwise correlationcoefficients between most regional aggregates are very high. For instance,correlation coefficients for the EU range from 0.68 with Asia to 0.87 withthe world aggregate.

Panel B of Table 2.4 reports the number of signals for contagion stem-ming from the PIGS aggregate based on the same approach as on the na-tional level. Accordingly, we test for rolling crisis-period windows starting onOctober 20, 2009. We apply both versions of the FR-test: first, the originalversion with overlapping data (FRO); and second, the alternative versionwith non-overlapping data (FRN ) suggested by Dungey et al. (2005). Thesignals are based on a 5% level of significance where we use the approximatedcritical values from Table A.1.

As can be seen, we also obtain many signals for contagion on the regionallevel. The regional pattern we found on the national level seems to beconfirmed by the regional analysis.


Table 2.3Definition of Regional Aggregates

This table provides an overview about the regional aggregates and the countries thatwere included for constructing the aggregates. EU stands for European Union; EMU forthe European Monetary Union; PIGS for Portugal, Ireland, Greece, and Spain; CEE forCentral and Eastern Europe; ME for Middle East; LATAM for Latin America; and G3for USA, Japan, and Germany.

EU EMU PIGS CEE ME Asia LATAM G3

Argentina xAustria x xBelgium x xBrazil xBulgaria x xChile xChina xColumbia xDenmark xEstonia x xFinland x xFrance x xGermany x x xGreece x x xHungary x xIndonesia xIreland x x xIsrael xItaly x xJapan xKazakhstan xMalaysia xMexico xNetherlands x xPeru xPhlippines xPoland x xPortugal x x xQatar xRomania x xRussia xSouth Korea xSpain x x xSweden xThailand xTurkey xUK xUkraine xUSA x


Table 2.4Correlation Coefficients and Contagion Signals for Regional Aggregates

Panel A reports correlation coefficients between the CDS premia in first differences ofselected regional aggregates for the whole sample. The regional aggregates are constructedas unweighted averages of CDS premia of countries belonging to the selected region. EUstands for European Union (ex PIGS countries); EMU for the European Monetary Union(ex PIGS countries); PIGS for Portugal, Ireland, Greece, and Spain; CEE for Central andEastern Europe; ME for Middle East; LATAM for Latin America; G3 for USA, Japan,and Germany; and World for the world aggregate (ex PIGS countries).

Panel A: Correlation Coefficients for the Whole Sample Period

EU EMU PIGS CEE ME Asia LATAM G3 World

EU 1.00EMU 0.74 1.00PIGS 0.68 0.86 1.00CEE 0.90 0.59 0.53 1.00ME 0.79 0.59 0.53 0.82 1.00Asia 0.68 0.61 0.53 0.65 0.76 1.00LATAM 0.63 0.57 0.51 0.62 0.63 0.68 1.00G3 0.66 0.80 0.68 0.59 0.60 0.65 0.50 1.00World 0.87 0.73 0.66 0.89 0.88 0.84 0.84 0.71 1.00

Panel B records the number of contagion signals stemming from the PIGS aggregate basedon the bivariate Forbes and Rigobon (2002) approach. We test for rolling crisis-periodwindows starting on October 20, 2009. We apply two version of the test. First, the originalversion with overlapping data (FRO). Second, an alternative version with non-overlappingdata (FRN) suggested by Dungey et al. (2005). The signals are based on a 5% level ofsignificance where we use the approximated critical values from Table A.1.

Panel B: Signals for Contagion Stemming from the PIGS Aggregate

FRO FRN


EU 42 57 73 39 53 70EMU 15 5 7 8 0 2CEE 0 5 0 1 4 0ME 21 23 8 18 11 0Asia 0 0 0 0 0 0LATAM 0 0 0 0 0 0G3 0 0 0 0 0 0World 0 0 0 0 0 0

2.4 Exploring the Common Factor 54

While the European aggregates produce the largest number of signals forboth versions of the test as well as for the different time windows, the testsalso return signals for Central and Eastern Europe and the Middle East. Incontrast, we find neither evidence of contagion for Latin America nor forAsia. If we further compare the results of the regional analysis with thoseof the national level, we find that the FRN is again more restrictive exceptfor the EU aggregate.

Overall, the regional analysis supports the findings of the analysis onthe national level. For Europe in particular, we can again conclude thatthere was “both contagion and interdependence” stemming from the PIGScountries. To our knowledge, testing for contagion with regional aggregatesis a fairly new approach. Hence, it might be interesting also to apply thisapproach to past crises such as the Asian crisis at the end of the 1990s.

2.4 Exploring the Common Factor

In the previous section we found strong evidence for contagion in CDS mar-kets during the Greek debt crisis both on a national and a regional levelby applying two versions of the FR-test to rolling crisis windows. Thesefindings are in contrast to the conclusion of “no contagion, only interde-pendence” from Forbes and Rigobon (2002) for equity markets after theHong Kong crash. We argued that one reason for this might be that thedevelopments after the collapse of the investment bank Lehman Brothershad demonstrated the close interconnection of financial markets with theworld economy. Accordingly, the risk of Greece becoming a new “Lehman”might have led to contagion rather than just interdependence.

Another possible explanation might be that the so-called “common factor”between CDS markets is stronger than it is between equity markets. Corsettiet al. (2001) argue that the strong conclusions from Forbes and Rigobon(2002) follow “from arbitrary assumptions on the variance of the country-specific noise in the market where the crisis originates – assumptions thatbias the test towards the null hypothesis of interdependence.”

They use a standard factor model of stock market returns to show thatthe ratio of the variance of common factors to country-specific factors hasan influence on test results for contagion. Accounting for this, they findevidence of contagion for at least five countries in the case of the Hong


Kong crash. However, the bivariate correlation approach we used in theprevious section does not allow the inclusion of country-specific factors. Also,applying the framework from Corsetti et al. (2001) to our sample lies beyondthe scope of this paper. Nevertheless, we think it is worth trying to get anidea of the role of the common factor between CDS markets.

To achieve that, we perform principal component analysis (PCA) for theregional aggregates of the CDS premia in first differences, i.e. we perform ananalysis to establish whether the patterns of correlations between sovereignCDS premia of a particular aggregate can be explained in terms of a smallernumber of common factors. This analysis is motivated by Longstaff et al.(2011). Table 2.5 summarizes the main results of the PCA analysis for thewhole sample (Panel A) as well as the two sub-samples (Panels B and C).

Like the findings of Longstaff et al. (2011), the results of the PCA ana-lysis indicate that there is a large amount of commonality in the intraregionalvariation of CDS premia. When all observations are used, we find that thefirst principal component captures almost half of the variation in the cor-relation matrix of the world aggregate, i.e. when all countries are included.This value increases to 78% in the case of the PIGS aggregate and 80% inthe case of the Asian aggregate. Furthermore, the first three principal com-ponents collectively explain 62% of the variation in the correlation matrixof the world aggregate. This is even higher than the 53% that Longstaff etal. (2011) find in their analysis.

The analysis of the two sub-samples indicates that the amount of com-monality in the intraregional variation of CDS premia is even larger whenwe only focus on the period after October 20, 2009 (Panel C). The first prin-cipal component now explains between 47% (world) and 84% (PIGS) of thevariation in the correlation matrix. Similarly, the first three components nowcollectively explain more than in Panels A and B. For instance, the shareincreases to 65% in the case of the world aggregate. These observationsindicate that the common factor plays a dominant role in CDS markets. Forinterpreting the first principal component, which might be interpreted as thecommon factor, we compute time series of the first principal components forthe different regional aggregates. Therefore, we take a weighted average ofthe daily changes in the sovereign CDS premia, where the weight for sover-eign i equals the i-th principal component weight divided by the sum of allthe principal component weights of the particular regional aggregate.


Table 2.5Regional Aggregates - Principal Component Analysis

This table reports principal components of inner-regional CDS premia. In addition we listcorrelation coefficients of the first principal component with selected financial variables.Cum. stands for cumulative; Stks R for the regional stock market index; Stks W for theworld stock market index; VIX for the volatility index; and CDS B for the average CDSpremium of 18 international banks. Significance of an increase of the adjusted correlationin Panel C compared to the correlation in Panel B at the 1%, 5%, 10% level is denotedby ∗∗∗, ∗∗, ∗, respectively.

Panel A: October 1, 2008 until July 27, 2010

Principal Components Correlation of Comp1 with

Comp1 Comp2 Comp3 Cum. Stks R Stks W VIX CDS B

EU 0.56 0.10 0.05 0.71 -0.58 -0.45 0.37 0.69EMU 0.63 0.08 0.05 0.77 -0.51 -0.38 0.32 0.63PIGS 0.78 0.09 0.08 0.95 -0.50 -0.36 0.34 0.60CEE 0.59 0.11 0.08 0.78 -0.58 -0.53 0.44 0.61ME 0.59 0.17 0.16 0.92 -0.54 -0.56 0.46 0.56Asia 0.80 0.07 0.05 0.92 -0.56 -0.37 0.25 0.44LATAM 0.74 0.11 0.09 0.95 -0.65 -0.66 0.55 0.55G3 0.57 0.26 0.18 1.00 -0.29 -0.33 0.25 0.60World 0.46 0.10 0.06 0.62 -0.56 0.45 0.70

Panel B: October 1, 2008 until October 19, 2009

Principal Components Correlation of Comp1 with


EU 0.56 0.11 0.04 0.71 -0.57 -0.42 0.27 0.64EMU 0.63 0.07 0.06 0.76 -0.51 -0.38 0.23 0.57PIGS 0.71 0.11 0.09 0.91 -0.48 -0.37 0.23 0.54CEE 0.56 0.12 0.08 0.77 -0.55 -0.49 0.35 0.55ME 0.61 0.16 0.15 0.92 -0.53 -0.53 0.43 0.51Asia 0.80 0.08 0.05 0.93 -0.55 -0.37 0.23 0.41LATAM 0.74 0.12 0.08 0.94 -0.69 -0.64 0.52 0.51G3 0.59 0.24 0.17 1.00 -0.25 -0.30 0.18 0.58World 0.47 0.10 0.05 0.62 -0.53 0.37 0.65

Panel C: October 20, 2009 until July 27, 2010

Principal Components Adjust. Correlation of Comp1 with


EU 0.58 0.09 0.06 0.73 -0.63∗∗ -0.57∗∗ 0.48∗∗∗ 0.79∗∗∗

EMU 0.65 0.09 0.06 0.80 -0.58∗∗ -0.47 0.39∗∗ 0.72∗∗∗

PIGS 0.84 0.07 0.06 0.97 -0.58∗ -0.38 0.31 0.61CEE 0.71 0.08 0.07 0.86 -0.81∗∗∗ -0.80∗∗∗ 0.71∗∗∗ 0.84∗∗∗

ME 0.55 0.22 0.18 0.95 -0.74∗∗∗ -0.83∗∗∗ 0.75∗∗∗ 0.85∗∗∗

Asia 0.79 0.08 0.05 0.92 -0.70∗∗∗ -0.52∗∗ 0.41∗∗ 0.65∗∗∗

LATAM 0.80 0.09 0.07 0.96 -0.61 -0.91∗∗∗ 0.85∗∗∗ 0.84∗∗∗

G3 0.54 0.28 0.18 1.00 -0.53∗∗ -0.53∗∗∗ 0.45∗∗∗ 0.72∗∗∗

World 0.47 0.10 0.08 0.65 -0.76∗∗∗ 0.66∗∗∗ 0.88∗∗∗


−10

−8

−6

−4

−2

0

2

4

6

8

10St

anda

rd D

evia

tion

10/2008 04/2009 10/2009 04/2010

Portugal IrelandGreece SpainFirst Principal Component

Figure 2.5: Principal Component Analysis for PIGS Countries

Figure 2.5 illustrates this by plotting the daily changes of the CDS premiaof the PIGS countries, and, in addition, the time series of the first principalcomponent of the PIGS aggregate. As can be seen, there is a large amountof commonality in the variation within the PIGS aggregate. In addition,and for a better understanding of the large amount of commonality withinthe regions, we explore the correlation of the first principal component withvarious financial variables. These variables are as follows:

• Index of regional stock market returns (Stks R): We construct indicesof daily regional stock market returns as an unweighted average ofdaily returns of national MSCI equity market indices that belong tothe regional aggregate in question.

• Index of worldwide stock market returns (Stks W): We construct anindex of daily worldwide stock market returns as an unweighted aver-age of daily returns of all national MSCI equity market indices in oursample.

• U.S. equity market volatility (VIX): The volatility of the U.S. equitymarket serves as a proxy for global nervousness of financial marketsand is expressed by the popular VIX index, which measures the impliedvolatility of S&P 500 index options.


• Index of CDS premia for international banks (CDS B): We constructan index of daily changes in CDS premia of banks as an unweightedaverage of CDS premia for 18 international banks such as GoldmanSachs or UBS.

The results are reported in Table 2.5 as well. In general, the signs of thecorrelation coefficients for the selected financial variables with the first prin-cipal component of the regional aggregates appear intuitive and consistent.For instance, a positive correlation coefficient with the index for bank CDSpremia is intuitive: the higher the risk premia for international banks, thehigher the implicit risk for a default of one of these international banks.This, in turn, should increase the implicit risk of negative spillovers to thesovereign level as the collapse of the investment bank Lehman Brothers hasclearly demonstrated. Similarly, the higher the risk of default is on a sover-eign level the higher the risk premia for banks should be, as they are usuallyheavily involved in financing the sovereign debt.

Looking at Panel A, we find that the correlation of the first principalcomponent is generally highest with the daily changes of the index of CDSpremia for international banks when all observations are included. Moreover,correlation is, in absolute terms, also very high for daily changes of theregional and worldwide stock market indices. The correlation coefficients ofthe stock indices are similar to the one Longstaff et al. (2011) find for theU.S. stock market returns. The results for the VIX index are usually lowercompared to the stock market returns. This is most apparent in the case ofthe Asian aggregate, where the correlation coefficient with the VIX index isonly at 0.25 while the correlation coefficient with the regional stock marketreturns is at -0.56.

What is more, we find that most adjusted (unconditional) correlationcoefficients15 increase significantly if we only focus on the time period afterthe debt crisis in Greece started (Panel C). For instance, correlation betweenthe first principal component for the world aggregate and the VIX index isnow at 0.66 while it was only at 0.37 in Panel B. This is more consistent withthe results from Longstaff et al. (2011), who find a correlation coefficient of0.659 with the VIX index.

15We apply formula (2.1) to adjust for higher volatility during the Greek debt crisis.

2.5 Conclusion 59

Overall, these results indicate that the principal source of variation of theCDS premia across the sovereigns of a particular region appears to be veryhighly correlated with global financial variables such as stock market returnsand stock market volatility. These results are consistent with Longstaff etal. (2011) and Pan and Singleton (2008) who likewise find a strong relationbetween sovereign credit spreads and financial market volatility measuredin the form of the VIX index. What is more, the adjusted correlation coef-ficients indicate that the co-movement increased significantly during theGreek debt crisis. This may help to explain why we find evidence for conta-gion even with the restrictive approach from Forbes and Rigobon (2002).

2.5 Conclusion

The recent focus of financial market participants on the fiscal situation in thePIGS countries provided us with the possibility to study the developments ofCDS premia for “hot-spot” countries. The difficult fiscal situation in Greeceled to a strong increase of risk premia for Greek sovereign debt as measuredin CDS premia. At the same time, CDS premia for many other countriesincreased strongly as well. While the increase for Greek CDS premia wasstrongest, these dramatic movements were also mirrored in the other CDSmarkets.

This shows that dramatic events in one market can have strong impactson other markets. The question is, however, whether a high degree of co-movement during times of crisis already constitutes contagion? The aim ofthis paper was to analyse this question for the Greek debt crisis. Thereforewe discuss the theoretical framework of what contagion constitutes as well asan empirical test procedure based on the approach from Forbes and Rigobon(2002).

However, given the difficulty of defining a fixed crisis period, we suggestenhancing the Forbes and Rigobon approach by applying the test on rollingwindows of periods of turmoil. We use rolling windows of relative turmoillasting 20, 40, and 60 days, respectively. As 20 business days are a commonproxy for a calendar month, the 20-day window corresponds with the 1-month period of relative turmoil that Forbes and Rigobon (2002) use fortesting for contagion after the Hong Kong crash in 1997. The 40-day and60-day windows account for the long-lasting nature of the Greek debt crisis.

2.5 Conclusion 60

Our results indicate that there were periods of contagion for CDS marketsduring the Greek debt crisis, which is in contrast to the results from Forbesand Rigobon (2002) for equity markets after the Hong Kong crash and theirconclusion of “no contagion, only interdependence.” Especially for Europeancountries we would instead conclude “both contagion and interdependence.”It seems that during the Greek debt crisis there were not only periods ofinterdependence but also periods characterized by a significant increase inthe co-movement of sovereign credit risk as measured in CDS premia. Thisis especially interesting as the approach of Forbes and Rigobon (2002) is,according to Dungey et al. (2005), a conservative test. What is more, evenForbes and Rigobon (2002) state that their result is “controversial”. Wethink that finding evidence for contagion for such a restrictive and, hence,controversial test is a very strong result. One underlying reason for findingevidence for contagion for CDS markets during the Greek debt crisis mightbe that the developments after the collapse of the investment bank LehmanBrothers had demonstrated the close interconnection of financial marketsand the world economy. Accordingly, the risk of Greece becoming a new“Lehman” might have led to contagion rather than just interdependence.

Motivated by the regional pattern of contagion signals we found on thenational level, we also aimed at testing for contagion on a regional level.We base our analysis on the findings from Longstaff et al. (2011), who per-formed a cluster analysis to identify significant commonality in sovereigncredit spreads on an aggregated level. Applying the same methodologyas on the national level yields many signals for contagion on the regionallevel as well. The regional pattern we found on the national level seemsto be confirmed by the regional analysis. While we get most signals forthe European aggregates, the tests also return many signals for Centraland Eastern Europe and the Middle East. In contrast, we find almost noevidence of contagion for Latin America and no signals at all for Asia.

Accordingly, the regional analysis supports the findings of the analysison the national level. Especially for Europe we again can conclude thatthere was “both contagion and interdependence” stemming from the PIGScountries. To our knowledge, testing for contagion with regional aggregatesis a fairly new approach. Hence, it might be interesting also to apply thisapproach to past crises such as the Asian crisis at the end of the 1990s.

2.5 Conclusion 61

Finally, we explore the common factor by conducting a principal componentanalysis. This analysis is motivated by Corsetti et al. (2001) who arguethat the strong conclusions from Forbes and Rigobon (2002) follow “fromarbitrary assumptions on the variance of the country-specific noise in themarket where the crisis originates – assumptions that bias the test towardsthe null hypothesis of interdependence.”

Our results indicate that there is a large amount of commonality in theintraregional variation of CDS premia. In addition, we find that the principalsource of variation of the CDS premia across the sovereigns of a particularregion appears to be very highly correlated with regional and global financialvariables such as stock market returns and stock market volatility. Theseresults are consistent with Longstaff et al. (2011) and Pan and Singleton(2008), who likewise find a strong relation between sovereign credit spreadsand financial market volatility measured in the form of the VIX index. Whatis more, the adjusted correlation coefficients indicate that the co-movementincreased significantly during the Greek debt crisis. This could explain whywe found evidence for contagion even with the restrictive approach of Forbesand Rigobon (2002).

Chapter 3

Testing for Contagion with a

Rolling-Crisis-Window

Approach

3.1 Introduction

Ever since the Greek debt crisis began, in October 2009, European policy-makers have attempted to reassure financial market participants that thesituation could be contained. However, there have been only temporaryperiods of relief after various measures and rescue packages had been an-nounced. Instead, the Greek debt crisis turned into a European one withfears of contagion steadily growing stronger.

This escalating worry can be best illustrated by the development of riskpremia for sovereign debt, for instance, in the premia of CDSs. While theCDS premium for a Greek government bond with a 5-year maturity anda notional value of USD 10 million was 124 basis points on October 20,2009, it soared to 2150 basis points by July 6, 2011. This means that thecost of insurance against a default of this particular Greek government bondincreased by a factor of more than 17, from USD 124,000 to USD 2.15 millionper year. At the same time, CDS premia for many other countries alsoincreased sharply. Figure 3.1 illustrates this by comparing the developmentof the CDS premia for Portugal, Ireland, Greece, and Spain. It is apparent

This chapter is based on Brill (2011).

3.1 Introduction 64

that the increase and the volatility for Greek CDS premia were strongest, butthese dramatic movements were mirrored in the other three CDS marketsas well. The question this paper seeks to answer is whether developmentslike these already constitute contagion.

0

500

1000

1500

2000

2500

CD

S Pr

emia

in B

asis

Poi

nts

01/2008 07/2008 01/2009 07/2009 01/2010 07/2010 01/2011

Portugal

Ireland

Greece

Spain

Source: Bloomberg

Figure 3.1: CDS Premia in Basis Points

Before answering this question, it is necessary to define what is meant bythe term “contagion”. Is it already sufficient to observe a high degreeof co-movement of financial markets across assets and countries during acrisis? The literature reveals little consensus on the definition of contagion(Rigobon, 2002). One reason for that lack of unity may be that the literatureon contagion is so extensive and the approaches to tackle the issue are verybroad. According to Dornbusch et al. (2000), the literature on contagioncan be separated conceptually into two broad categories.

On the one hand, much of the literature focuses on potential causes andtransmission channels of contagion. Dornbusch et al. (2000) distinguish twomain categories, namely fundamental causes and those related to investors’behaviour. Fundamental causes of contagion might be due to macroeco-nomic shocks on a global scale and local shocks that might be transmittedthrough trade, currency devaluations, and financial links.1 In addition to

1For macroeconomic common shocks cf., for instance, Chuhan, Claessens, and Mamingi

3.1 Introduction 65

fundamental causes, investors’ behaviour, whether rational or not, may fa-cilitate the transmission of shocks from one market to another.2 In any case,it appears that the causes and transmission channels and their relative im-portance are little understood and remain topics for useful future research.

On the other hand, there is extensive literature focusing on empiricaltests for the existence of contagion during crises. Many approaches forthis testing have been proposed, and no unifying framework has prevailed.Dungey et al. (2005) review various methodologies – including the correla-tion analysis of Forbes and Rigobon (2002), the VAR approach of C. A. Faveroand Giavazzi (2002), the probability model of Eichengreen, Rose, and Wyplosz(1995; 1996), the co-exceedance approach of Bae, Karolyi, and Stulz (2003),and that of Corsetti et al. (2001, 2005) and Bekaert, Harvey, and Ng (2005)based on a latent factor model – and compare the outcomes of these ap-proaches when applied to the East Asian crisis in 1997-98. Dungey et al.(2005) find “that the Forbes and Rigobon adjusted correlation test is a con-servative test, whereas the contagion test of Favero and Giavazzi tends toreject the null of no contagion too easily. The remaining tests investigatedyield results falling within these two extremes.”

Returning to the Greek debt crisis and the question whether there isempirical evidence for contagion, Andenmatten and Brill (2011b) focusedon the approach of Forbes and Rigobon (2002) and proposed to enhance it,taking into account the difficulties of defining the exact time span of a crisis,as acknowledged by Forbes and Rigobon (2002). Instead of testing for a fixedcrisis period, Andenmatten and Brill (2011b) applied the test procedureon rolling windows of potential crises, using this approach on CDS premiaover the period from 2008 until mid 2010. The results found evidence ofcontagion despite the conservative nature of the original approach of Forbesand Rigobon (2002).

(1998), Corsetti, Pesenti, and Roubini (1998), Radelet and Sachs (1998a, 1998b); for localshocks cf., for instance, Corsetti, Pesenti, Roubini, and Tille (1999).

2Cf. Pritsker (2001) for a classification of different types of investors’ behaviour relatedcauses. These include liquidity and incentive problems (e.g., G. A. Calvo & Mendoza,2000; Garber, 1998; Kaminsky & Reinhart, 2000; Kodres & Pritsker, 2002; Schinasi& Smith, 2000), information asymmetries and coordination problems (e.g., Agenor &Aizenman, 1998; Banerjee, 1992; Bikhchandani, Hirshleifer, & Welch, 1992; G. A. Calvo& Mendoza, 2000; Scharfstein & Stein, 1990; Shiller, 1995; Wermers, 1999), multipleequilibriums (e.g., Diamond & Dybvig, 1983; Drazen, 1999; Gerlach & Smets, 1995;Jeanne, 1997; Masson, 1998), and changes in the rule of the game (e.g., Dornbusch, 1997;Park, 1998).

3.2 The Rolling-Crisis-Window Approach 66

Motivated by this result, the present paper aims at challenging the robust-ness of the results of Forbes and Rigobon (2002) by applying the proposedenhancement of Andenmatten and Brill (2011b) on some 32 MSCI equitymarket indices during the same crises Forbes and Rigobon (2002) focus on:the East Asian crisis of 1997-98, the Mexican peso crisis of 1994, and the U.S.stock market crash of 1987. Accordingly, this paper relies on the definition ofcontagion formulated by Forbes and Rigobon (2002) as a significant increasein the unconditional correlation coefficient in a period of high volatility (acrisis) compared to a period of relative low volatility.

The rest of the chapter proceeds as follows: In section 3.2, the approachof Andenmatten and Brill (2011b) is developed by first introducing the ori-ginal test procedure from Forbes and Rigobon (2002) and then discussingthe proposed enhancement to apply the original test procedure on rollingwindows of turmoil. In section 3.3, the rolling test approach is applied onthe East Asian crisis of 1997-98 and the results are compared to those ofForbes and Rigobon (2002). The potential source of contagion is alteredalong the lines of Forbes and Rigobon (2002), from Hong Kong to Thailand,Indonesia, and Korea, respectively. In sections 3.4 and 3.5, the focus thenshifts to the Mexican peso crisis of 1994, and then to the U.S. stock marketcrash of 1987. Finally, section 3.6 presents the conclusions drawn from theseanalyses.

3.2 The Rolling-Crisis-Window Approach

The basic idea of the approach of Forbes and Rigobon (2002) was to testwhether the correlation between two variables increases significantly duringa crisis period compared to a period of relative stability. However, careneeds to be taken when comparing correlation coefficients between differentperiods because, as Boyer et al. (1997) and Forbes and Rigobon (2002) show,correlation coefficients between markets are conditional on volatility. Hence,during periods of increased volatility (in times of crisis) estimates of correl-ation coefficients are biased upward. If co-movement tests are not adjustedfor that bias, contagion is too easily detected, as was the case in the analysisof King and Wadhwani (1990).

Forbes and Rigobon (2002) present a statistical correction for this condi-tioning bias and the appropriate procedure to test for contagion (henceforth


referred to as the FR-test), which will be introduced in the following.3 First,different sample periods will be used:

x : a period of relative stability (low volatility)

y : a crisis period (high volatility)

z : the entire sample period

Moreover, the following parameters of volatility and correlation need to bedefined:

σ2x,i : volatility of country i′s equity market returns before the crisis (i = 1, 2)

σ2y,i : volatility of country i′s equity market returns in the crisis (i = 1, 2)

ρj : correlation between countries 1 and 2 in period j (j = x, y, z)

Forbes and Rigobon show that the unconditional correlation that adjusts forthe above-mentioned volatility bias in correlation coefficients is given by:4

νy =ρy√

1 + [(σ2y,1 − σ2

x,1)/σ2x,1](1 − ρ2

y)(3.1)

As can be seen from (3.1), the unconditional correlation coefficient, νy, isthe unadjusted correlation coefficient, ρy, scaled by a function of the changein volatility in asset returns of the source country over the high and lowvolatility periods. To illustrate this, assume that σ2

y,1 > σ2x,1, i.e., that the

volatility of asset returns in country 1 increases from period x to period y.Then, (σ2

y,1 − σ2x,1)/σ2

x,1 > 0 and for any ρy ∈ (−1; 1) it follows that

√1 + [(σ2

y,1 − σ2x,1)/σ2

x,1](1 − ρ2y) > 1 (3.2)

This implies that νy < ρy. From (3.1) it is also apparent that νy = ρx

3The notation is based on that of Andenmatten and Brill (2011b) and Dungey et al. (2005),respectively.

4For a similar approach, cf. Boyer et al. (1997), Corsetti et al. (2001, 2005), and Loretanand English (2000); for alternative approaches, cf. Karolyi and Stulz (1996), and Longinand Solnik (1995).


if there is no fundamental shift in the relationship between the two assetmarkets from the low to the high volatility period.

Accordingly, the null hypothesis and the alternative hypothesis, respect-ively, of a test that there is a significant increase in the correlation coefficientin the high volatility period, i.e. that there is contagion, can be formulatedas follows:

H0 : νy = ρx (3.3)

H1 : νy > ρx (3.4)

The t-statistic can be used to test this hypothesis where the t-statistic isgiven by

t =νy − ρx√

V ar(νy − ρx)(3.5)

and where νy and ρx denote the sample estimators of νy and ρx, respect-ively. Assuming that the two samples are drawn from independent normaldistributions, the standard error in (3.5) can be transformed as follows:

V ar(νy − ρx) = V ar(νy) + V ar(ρx) − 2Cov(νy, ρx) (3.6)

= V ar(νy) + V ar(ρx) (3.7)∼= (1/Ty) + (1/Tx) (3.8)

where Tx (Ty) is the sample size of the low (high) volatility period. To getto (3.7) the independence assumption is used and (3.8) follows from theassumption of normality and an asymptotic approximation.5 Using (3.8) in(3.5) results in

FR =νy − ρx√

(1/Ty) + (1/Tx)(3.9)

In a next step, Forbes and Rigobon (2002) suggest using the Fisher trans-formation, as this improves the finite sample properties of the test statistic.6

5For this asymptotic approximation Dungey et al. (2005) refer to Kendall and Stuart(1969).

6According to Dungey et al. (2005), the Fisher transformation is valid for small values ofboth the unadjusted and the adjusted correlation coefficient.


This yields

FR =1/2[log((1 + νy)/(1 − νy)) − log((1 + ρx)/(1 − ρx))]√

(1/Ty − 3) + (1/Tx − 3)(3.10)

When applying (3.10) the respective low and high volatility periods areseparated from each other, that is, the two periods do not overlap, as wassuggested by Dungey et al. (2005). However, in the original test statisticfrom Forbes and Rigobon (2002) for testing (3.3) against (3.4) the non-crisisperiod is defined as the whole sample period z, with the non-crisis and thecrisis periods overlapping. Accordingly, (3.9) would be formulated as

FR =νy − ρz√

(1/Ty) + (1/Tz)(3.11)

and the Fisher-adjusted version (3.10) as

FR =1/2[log((1 + νy)/(1 − νy)) − log((1 + ρz)/(1 − ρz))]√

(1/Ty − 3) + (1/Tz − 3)(3.12)

As shown, the test statistics (3.11) and (3.12) build on the assumption thatthe variances of νy and ρz are independent. This assumption is violated,however, if the two sample periods overlap. As a result, the covariance termin (3.6) is most likely not equal to zero and the step from (3.6) to (3.7)results in a standard error that is too large, as the negative covariance termis not taken into account. As stated by Dungey et al. (2005), this leadsto a downward bias in the t-statistic and, hence, to fewer rejections of thenull hypothesis. Nevertheless, both versions might be applied in order tosee if this downward bias leads to fundamentally different results. In termsof notation, the remainder of this paper refers to the overlapping version(3.12) suggested by Forbes and Rigobon (2002) as the FRO-test, and to thenon-overlapping version suggested by Dungey et al. (2005) as the FRN -test.

Given the difficulties of defining a fixed period for the Greek debt crisis,Andenmatten and Brill (2011b) suggested enhancing the Forbes and Rigo-bon approach by applying the test to so-called “rolling windows” of turmoil.They defined rolling windows of relative turmoil lasting 20, 40, and 60 days.The 20-day window corresponds to a 1-month period of relative turmoilsince 20 business days are a common proxy for a calendar month. This


is the length of the crisis period Forbes and Rigobon (2002) suggest using.The 40- and 60-day windows account for a longer lasting crisis, such as theGreek debt crisis turned out to be. Even longer lasting crisis windows couldbe considered.

Before applying the tests, differences should first be calculated, that is,the daily changes of all variables in order to transform the time series intostationary ones. Augmented Dickey-Fuller tests might be used to confirmthat the variables in first differences are all stationary. Also, as is com-mon practice when testing for contagion, 2-day-moving averages of the dailychanges in the variables should be calculated.7 This accounts for the prob-lem that financial markets in different regions are not open at the sametime.

After transforming the variables, both versions of the FR-test might beapplied on rolling windows of turmoil: first, the original version with over-lapping data (FRO); second, the alternative version with non-overlappingdata (FRN ). The approach of Andenmatten and Brill (2011b) suggests us-ing incremental steps of a day to move the crisis window forward, therebytesting step by step for contagion on a 5% level of significance and countingthe number of such signals over the whole test period. Figure 3.2 illustratesthis approach.

However, Andenmatten and Brill (2011b) show that for identifying thesignals – comparing the test statistics with the critical t-values – the crit-ical t-values of a standard one-sided t-test are unreliable. Similar to testingfor a structural break at an unknown break date, where the so-called supF-statistic8 is the largest of many F-statistics and, hence, its distribution isnot the same as an individual F-statistic. The distribution of the FR-teststatistic differs from the standard t-distribution. The authors use MonteCarlo methods to find approximate critical t-values. The results are shownin Table A.1. As can be seen, the critical values for the rolling FR-testsare larger than those for a standard one-sided t-test, which are approxim-ately 1.282, 1.646, and 2.330 for significance levels of 10%, 5%, and 1%,respectively.

7For instance, cf. Corsetti et al. (2001), Dungey et al. (2005), and Forbes and Rigobon(2002).

8For an introduction to this approach, readers are referred to Stock and Watson (2007).For more details, Perron (2005) offers a review of the literature on dealing with structuralbreaks.

3.3 Contagion During the East Asian Crisis 71

min

max

min time max

Period of Relative Stability

Crisis Windows2-

Day-M

oving

Ave

rage

of

Daily

Cha

nges

of V

ariab

les

Period ofPotential Crisis

Figure 3.2: Illustration of the Rolling Contagion Test Approach

3.3 Contagion During the East Asian Crisis

In the previous section the rolling-crisis-window approach of Andenmattenand Brill (2011b) was introduced. In this section, this approach is appliedto the East Asian crisis that started in 1997 and the results are comparedto those from Forbes and Rigobon (2002).

In order to better understand what happened during the East Asiancrisis, a brief overview of key events, based on the chronology of BIS (1998),is presented.9

The crisis started in early 1997 when Thai authorities intervened heavilyin currency markets to counter pressure on the Thai baht. Neither theseinterventions nor additional measures managed to alleviate the pressure.On July 2, 1997, a devaluation and the floating of the Thai baht occurred.Thereafter, the pressure spread to the Philippine peso, the Malaysian ringgitand the Indonesian rupiah, which first led to a widening of the fluctuationbands of the Philippine peso and the Indonesian rupiah, on July 11, 1997,and then, on August 14, to the floating of the Indonesian rupiah. After somerelatively calm weeks, financial turbulence revived in October 1997, peaking

9For detailed chronologies of the East Asian crisis, cf. also Baur and Fry (2006) andKaminsky and Schmukler (1999).


with the speculative attack on the Hong Kong dollar when the Hang SengIndex fell by 23% within three days. These developments are illustrated inFigure 3.3.

20

40

60

80

100

120

140

MSC

I Ind

ex in

Loc

al C

urre

ncy,

Reb

ased

06/1997 08/1997 10/1997 12/1997 02/1998

Hong Kong

Indonesia

Korea

Thailand

Brazil

Mexico

Canada

Germany

UK

U.S.


Figure 3.3: Stock Market Indices During the East Asian Crisis

Thereafter, the pressure on the Korean won mounted and the fluctuationband had to be widened on November 20, 1997. On the next day, Koreaapplied for an IMF standby credit. On December 4, 1997, the IMF standbycredit was approved and December 16 saw the floating of the Korean won.In the following weeks, the situation remained difficult in many East Asiancountries and only on August 31, 1998, with the re-pegging of the Malaysianringgit, could the end of the East Asian crisis be declared (Dungey et al.,2005).

The severity of the crisis is reflected by the many significant changes un-dertaken in its aftermath: numerous financial institutions were closed or re-structured in most crisis countries, including Indonesia, Korea, Malaysia andThailand; capital controls were imposed in Malaysia in 1998; political lead-ership was changed in all countries except Malaysia; sovereign ratings weredramatically reduced for all crisis countries; and Indonesia, Thailand andKorea all sought IMF assistance packages (Dungey, Fry, & Martin, 2006).

This brief review of key events shows the crisis evolved over time and


across markets. The dynamic nature of this crisis was acknowledged byForbes and Rigobon (2002), who wrote: “One difficulty in testing for conta-gion during this period is that there is no single event which acts as a clearcatalyst behind this turmoil.” They argue that the East Asia crisis onlymade headlines after the October crash in Hong Kong, making this, in theirview, the most likely event for driving contagion. Hence, in their study, theydefine Hong Kong as the source of potential contagion and the crisis periodas the month starting on October 17, 1997. What is more, they performrobustness tests by altering both the crisis and the non-crisis periods as wellas the source of contagion.

Forbes and Rigobon (2002) find almost no evidence for contagion duringthe East Asian crisis and argue that their key finding – “no contagion, onlyinterdependence” – is unaffected by the period definition. However, theresults from Andenmatten and Brill (2011b) suggest that this conclusionmay be challenged by the rolling-crisis-window approach.

This challenge will be explored in the following pages by applying therolling-crisis-window approach on 2-day moving averages of daily returns ofMSCI equity market indices in local currencies, with data from ThomsonReuters Datastream. With this approach, the potential source of contagionshifts from Hong Kong to Thailand, Indonesia, and Korea. In the base case,the period of relative stability starts on January 1, 1996 and the rolling-crisiswindows run from June 1, 1997, until August 31, 1998. Accordingly, usingrolling crisis windows already incorporates the robustness tests of Forbesand Rigobon (2002) with regard to the definition of the crisis period. Therobustness tests of altering the start of the period of relative stability arecovered separately and the results are presented in Appendix B. Moreover,the tests are replicated using daily returns in US dollars instead of localcurrency. These results are also presented in Appendix B.

3.3.1 Contagion Stemming from Hong Kong

Starting with Hong Kong as the source of contagion, Table 3.1 reports thenumber of signals based on 2-day-moving-average returns of the MSCI indexin local currency. The period of relative stability starts on January 1, 1996and the rolling crisis windows run from June 1, 1997, until August 31, 1998.Both versions of the FR-test are applied for rolling crisis windows of 20, 40,and 60 days. The signals are based on a 5% level of significance where the


approximated critical values from Andenmatten and Brill (2011b) are used(cf. Table A.1).

In Table 3.1, in contrast to the static approach of Forbes and Rigobon(2002), the rolling-crisis-window approach of Andenmatten and Brill (2011b)yields evidence for contagion for both FR-tests as well as for the three dif-ferent crisis windows. The largest number of signals – 27 out of 31 cases –is observed for the overlapping test, FRO, for the 20-day window. Focusingon the relative frequency of the signals across countries reveals that Asiancountries dominate. For instance, for Malaysia, Thailand, and Taiwan, 35,33, and 22 signals are obtained, respectively.

It was not only Asian countries that seem to be affected by contagionstemming from the Hong Kong, but also the equity markets of many otheremerging countries and those of many industrialized countries. In LatinAmerica, for instance, all four countries of the sample were affected by con-tagion stemming from the Hong Kong equity market. Among OECD coun-tries, only Belgium, Canada, and the UK seem unaffected, noting that thenumber of signals is much lower on average outside Asia.

Focusing on the 40- and 60-day windows, signals for fewer countrieswere observed but the regional pattern resembles that of the 20-day window.Accordingly, one conclusion from the FRO-test is that there is not onlyevidence for contagion stemming from the Hong Kong equity market butalso a strong regional pattern. This observation holds true for the FRN -testas well. However, the FRN -test seems more restrictive than the FRO-test,as the number of countries for which at least one signal is obtained decreasesfrom 27 to 7 for the 20-day window, from 23 to 6 for the 40-day window andfrom 23 to 5 for the 60-day window. What is more, the average number ofsignals for countries with at least one signal declines from 10 to 5 for the 20-day window. As discussed in the previous section, this was to be expected,as the standard errors of the FRO-test are likely to be biased upward dueto the overlapping nature of the periods of relative turmoil and stability.

Figure 3.4 illustrates this pattern for the equity markets of Hong Kongand the Philippines. The areas shaded in grey highlight the periods ofcontagion stemming from the Hong Kong equity market for both versions ofthe FR-test as well as for the three different crisis windows. The upper rowshows the results for the FRO-tests and the lower row for the FRN -tests.


Table 3.1East Asian Crisis: Contagion Stemming from Hong Kong

This table records the number of contagion signals stemming from the Hong Kong equitymarket based on 2-day-moving-average returns in local currency. The period of relativestability starts on January 1, 1996, and the rolling crisis windows run from June 1, 1997,until August 31, 1998. Both the FRO and the FRN test versions are applied for rollingcrisis windows of 20, 40 and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN


East AsiaIndonesia 2 0 0 0 0 0Japan 12 7 5 1 0 0Korea 19 30 25 0 0 0Malaysia 35 51 47 5 0 0Philippines 18 20 32 2 6 0Singapore 4 1 20 0 0 0Taiwan 22 47 93 7 44 79Thailand 33 26 44 0 7 1

Latin AmericaArgentina 1 6 8 0 0 0Brazil 4 3 4 0 0 0Chile 14 6 6 2 3 3Mexico 1 2 0 0 0 0

OECDAustralia 9 10 10 0 0 0Belgium 0 0 0 0 0 0Canada 0 0 0 0 0 0France 6 8 10 0 0 0Germany 2 2 2 0 0 0Italy 5 7 9 0 0 0Netherlands 2 2 3 0 0 0Spain 7 7 8 0 0 0Sweden 5 6 7 0 0 0Switzerland 2 0 0 0 0 0UK 0 0 0 0 0 0USA 6 7 10 0 0 0

Other European MarketsGreece 7 11 7 0 0 0Ireland 5 0 0 0 0 0Portugal 6 0 3 0 0 0

Other Emerging MarketsChina 20 44 60 11 22 39India 11 15 12 7 8 8Russia 0 0 0 0 0 0South Africa 1 2 2 0 0 0


0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

Hong KongPhilippinesContagion

FRO: 20-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRO: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRO: 60-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 20-Day Window

0

20

40

60

80

100M

SCI I

ndex

, Reb

ased

01/96 01/97 01/98 01/99

FRN: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 60-Day Window

Figure 3.4: Contagion Stemming from Hong Kong: Signals for Philippines

As can be seen from the three columns, the FRN -test is more restrictive thanthe FRO-test, i.e., the periods of contagion are shorter or even non-existentfor the FRN -test in the case of the 60-day window, for example. However,both versions of the FR-test yield consistent results and the periods of con-tagion based on the FRN -test match those of the FRO-test. What is more,Figure 3.4 not only illustrates that there were periods of contagion but alsothat these periods are distinct from the crisis period Forbes and Rigobon(2002) relied on, namely the month starting October 17, 1997. Instead, theFRO-test based on the 20-day window yields two periods of contagion, onecovering July 1997 and the other starting in February 1998 and lasting untilMay 1998.

It is interesting to note that, in both cases, equity markets tended up-ward. In July 1997, contagion emerged during the final phase of the equitybull market in Hong Kong; in early 1998, it appeared during a recoveryphase after equity markets had plunged by almost 60% in Hong Kong andalmost 50% in the Philippines since August 1997. One reason for this mightbe that, as a group, investors reacted more to good news than to bad newsabout the speculative attack on Hong Kong in October 1997. As can be seenin Figure 3.4, the MSCI index for the Philippines had already strongly cor-rected in the weeks prior the start of the speculative attack on Hong Kong.Hence, it may be concluded that the bad news during the early phase of the


East Asian crisis was perceived more locally in each affected country andhad less broad impact than did the recovery phase in 1998, when investorseverywhere were relieved that an end of the crisis might be in sight.

3.3.2 Contagion Stemming from Thailand

After testing for contagion stemming from Hong Kong in the previous sec-tion, this section focuses on assessing Thailand as a potential source ofcontagion. The period of relative stability starts again on January 1, 1996and the rolling crisis windows run from June 1, 1997, until August 31, 1998.

As can be seen in Table 3.2, the results are similar to those of Table3.1 as the rolling-crisis-window approach of Andenmatten and Brill (2011b)again yields evidence of contagion for both versions of the FR-test as wellas for the three different crisis windows. The greatest number of signalsis observed – 27 out of 31 cases – for the FRO-test for both the 20- and60-day windows. Focusing on the relative frequency of the signals acrosscountries, it is again apparent that Asian countries dominate. Indonesia isthe only Asian country evidencing no contagion at all. Outside of Asia theFRO-test yields evidence of contagion for most countries. In Latin Americaall four countries of the sample are also affected by contagion stemmingfrom Thailand’s equity market. Among the OECD countries, only Italyand Switzerland seem to have been unaffected, based on the 20-day crisiswindow.

The main difference to the results from the previous section is that theFRN -tests yield more signals. The number of countries for which at leastone signal is obtained increases from 7 to 22 for the 20-day window andfrom 6 and 5 to 16 for the 40- and 60-day windows, respectively. Also, theaverage number of signals increases for the 20-day and 40-day windows from5 and 15 to 11 and 18, respectively.

Figure 3.5 illustrates this observation. The areas shaded in grey high-light the periods of contagion from Thailand’s equity market to Malaysia’sfor both versions of the FR-test as well as for the three different crisis win-dows. While similar to Figure 3.4, it is apparent that the FRN -test is morerestrictive than the FRO-test. It seems that in this case the difference isnot as large as in the previous one. Moreover, the two equity markets seemto have been more in sync from August 1997 onward than they had beenduring the period of relative stability. This reflects the chronology of the


crisis, as Thailand was one of the first countries forced to fight the pressureon its currency during the crisis. Apparently, this turbulence was passed onto Thailand’s equity market much earlier than to the other equity marketsin the region.

The broad evidence for contagion stemming from Thailand’s equity mar-ket again contrasts with the results of Forbes and Rigobon (2002). They findno evidence for contagion at all for the two crisis periods they use for test-ing for contagion stemming from Thailand’s equity market, which run fromJanuary 6, 1997, to June 30, 1997, and from August 7, 1997, to September6, 1997, respectively.

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

ThailandMalaysiaContagion

FRO: 20-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRO: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRO: 60-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 20-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 60-Day Window

Figure 3.5: Contagion Stemming from Thailand: Signals for Malaysia

3.3.3 Contagion Stemming from Indonesia

In this section, the focus shifts to Indonesia as a potential source of con-tagion. As in the previous sections, the rolling-crisis-window approach ofAndenmatten and Brill (2011b) is applied for both versions of the FR-testand crisis windows of 20, 40, and 60 days on 2-day-moving-average returnsof national MSCI indices in local currency. The period of relative stabilityresumes on January 1, 1996, and the rolling crisis windows run from June1, 1997, until August 31, 1998.


Table 3.2East Asian Crisis: Contagion Stemming from Thailand

This table records the number of contagion signals stemming from Thailand’s equity mar-ket based on 2-day-moving-average returns in local currency. The period of relative sta-bility starts on January 1, 1996, and the rolling crisis windows run from June 1, 1997,until August 31, 1998. Both the FRO and the FRN test versions are applied for rollingcrisis windows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN


East AsiaHong Kong 31 20 39 9 14 11Indonesia 0 0 0 0 0 0Japan 25 46 56 14 29 20Korea 51 49 51 60 78 105Malaysia 22 34 47 15 25 71Philippines 7 3 5 0 0 0Singapore 12 19 20 3 6 0Taiwan 35 31 9 6 11 0






Table 3.3 reports the number of signals for contagion on a 5% level of sig-nificance. The results are similar to those of Tables 3.1 and 3.2, as therolling-crisis-window approach again yields evidence for contagion for bothversions of the FR-test as well as for the three different crisis windows. Thelargest number of signals, namely for 23 out of 31 cases, is observed forthe FRO-test for the 60-day window. This compares to 23 and 27 caseswhen Hong Kong and Thailand, respectively, were the potential source ofcontagion.

With the focus on the relative frequency of the signals across countries,the data again suggests that Asian countries dominate but less so than inthe previous cases. For the FRO-test based on the 60-day window, forinstance, the average number of signals was 38 for East Asian countrieswhen testing for contagion stemming from Hong Kong while it was only 6and 7 for Latin American and OECD countries, respectively. The averagenumber of signals is 26 for East Asian countries and 24 for Latin Americanand 17 for OECD countries. Additionally, more East Asian countries seemnot to have been at all affected by contagion from the Indonesian equitymarket. Interestingly, these countries are Hong Kong and Thailand. In fact,no signals for contagion were observed in either direction: Indonesia wasthe only East Asian country unaffected by contagion from Hong Kong andThailand, as noted in the two previous sections.

Looking at the three columns for the FRN -test for Indonesia, it is ap-parent that fewer signals were obtained than in the previous section, whenThailand was the potential source of contagion. Indeed, not a single countryrecorded any signal for contagion stemming from Indonesia’s equity marketfor the 60-day FRN -test window. On the other hand, when the same testwas applied to Thailand’s equity market, some 16 countries evidenced signalsfor contagion (cf. Table 3.2).

Figure 3.6 reinforces this observation. The areas shaded in grey highlightthe periods of contagion from Indonesia’s equity market to Korea’s for bothversions of the FR-test as well as for the three different crisis windows. Thepattern is similar to that seen in Figure 3.4, namely, the FRN -test is morerestrictive than the FRO-test, which is apparent in columns 2 and 3 of thefigure: While the FRO-test finds evidence for contagion for both the 40-dayand the 60-day windows, the FRN -test finds no evidence at all for these twocrisis windows.


Table 3.3East Asian Crisis: Contagion Stemming from Indonesia

This table records the number of contagion signals stemming from Indonesia’s equitymarket based on 2-day-moving-average returns in local currency. The period of relativestability starts on January 1, 1996, and the rolling crisis windows run from June 1, 1997,until August 31, 1998. Both the FRO and the FRN test versions are applied for rollingcrisis windows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN


East AsiaHong Kong 0 0 0 0 0 0Japan 24 31 39 0 0 0Korea 20 8 23 11 0 0Malaysia 3 4 4 0 0 0Philippines 5 10 15 0 0 0Singapore 0 2 2 0 0 0Taiwan 31 40 70 0 2 0Thailand 0 0 0 0 0 0






Overall, the results from the approach of Andenmatten and Brill (2011b)yield ample evidence for contagion stemming from Indonesia’s equity market.Again, this contrasts with the results of Forbes and Rigobon (2002), whofound no such evidence for the crisis period August 7 to September 6, 1997,the month immediately after the Indonesian equity market plummeted.

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

IndonesiaSouth KoreaContagion

FRO: 20-Day Window

0

20

40

60

80

100M

SCI I

ndex

, Reb

ased

01/96 01/97 01/98 01/99

FRO: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRO: 60-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 20-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 60-Day Window

Figure 3.6: Contagion Stemming from Indonesia: Signals for Korea

3.3.4 Contagion Stemming from Korea

In this section the source of contagion is changed from Indonesia to Korea.As can be seen from Table 3.4, there is again broad evidence for contagionand the results are similar to those from the previous sections.

Once again, the FRN -test is more restrictive than the FRO-test as boththe number of countries for which at least one signal is obtained and theaverage number of signals for these countries is lower for each of the threedifferent time windows.

In addition, it is interesting to note that Latin American and OECDcountries seemed less affected by contagion than in the previous sections.For instance, the FRO-test based on the 20-day window finds evidence onlyfor 4 out of the 12 OECD countries – Belgium, Canada, the Netherlands, andSwitzerland – while there was evidence for 9 and 10 OECD countries whentesting for contagion stemming from Hong Kong, Thailand, and Indonesia.


Table 3.4East Asian Crisis: Contagion Stemming from Korea

This table records the number of contagion signals stemming from Korea’s equity marketbased on 2-day-moving-average returns in local currency. The period of relative stabilitystarts on January 1, 1996, and the rolling crisis windows run from June 1, 1997, untilAugust 31, 1998. Both the FRO and the FRN test versions are applied for rolling crisiswindows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN


East AsiaHong Kong 10 3 8 7 0 0Indonesia 0 0 1 0 0 1Japan 8 3 0 0 0 0Malaysia 1 0 0 1 0 0Philippines 2 0 2 1 0 0Singapore 11 5 8 10 7 0Taiwan 0 0 0 0 0 0Thailand 8 17 16 12 32 63





3.4 Contagion During the Mexican Peso Crisis 84

Figure 3.7 illustrates the test results of contagion spreading from Korea’sequity market to China’s. Again, it is apparent that the FRN -test is morerestrictive than the FRO-test, but that the periods match for the respectivetime windows if they exist for both test versions.

Overall, the test results for contagion stemming from Korea contrastto the results of Forbes and Rigobon (2002), who found no evidence forcontagion at all for the crisis period from October 23 to December 22, 1997,which was the period after the speculative attack on Hong Kong, when theKorean won came under pressure.

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

KoreaChinaContagion

FRO: 20-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRO: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRO: 60-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 20-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/96 01/97 01/98 01/99

FRN: 60-Day Window

Figure 3.7: Contagion Stemming from Korea: Signals for China

3.4 Contagion During the Mexican Peso Crisis

After applying the rolling-crisis-window approach of Andenmatten and Brill(2011b) to the East Asian crisis of 1997-98 in the previous section, thissection focuses on the second crisis Forbes and Rigobon (2002) use for theiranalysis, namely the Mexican peso crisis in 1994. According to Whitt (1996),the Mexican government played a key role in causing the collapse of the Mex-ican peso in December 1994. In order to avoid an economic slowdown, thegovernment attempted to stimulate the economy by maintaining the quasi-pegged exchange rate of the peso, at the same time limiting its monetary


tightening with massive sterilized interventions. Ultimately, this resulted ina balance of payment crisis that caused the peso to collapse and the Mexicanstock market to crash (Forbes & Rigobon, 2002). Following these dramaticevents, fears of contagion increased rapidly, especially in Latin America.

These fears became reality, as S. Calvo and Reinhart (1996) illustrate.Argentina lost 18% of the deposits in its banking system and about 50%of its foreign exchange reserves and Brazil had to implement measures tostimulate capital inflows. However, from April 1995 on, foreign capital beganreturning and equity markets started to recover, as can be seen from Figure3.8.

20

40

60

80

100

120

MSC

I Ind

ex in

Loc

al C

urre

ncy,

Reb

ased

09/1994 11/1994 01/1995 03/1995 05/1995

Mexico

Argentina

Brazil

Chile

Hong Kong

Indonesia

Canada

Germany

UK

U.S.


Figure 3.8: Stock Market Indices During the Mexican Peso Crisis

Forbes and Rigobon (2002) argue that testing for contagion during thiscrisis “is more straightforward than that during the East Asian crises dueto the existence of one clear catalyst (the collapse of the peso) driving anycontagion.”

For their base case, Forbes and Rigobon define the period of relative sta-bility running from January 1, 1993, to December 31, 1995, and the crisisperiod lasting from December 19 to December 31, 1994. For calculating thedaily returns, they rely on equity indices in US dollars. Based on this ap-proach, they find almost no evidence for contagion. This result is supported


by robustness tests where they alter the period definitions and use equityreturns in local currency instead of US dollars.

Similar to the previous section, the rolling-crisis-window approach ofAndenmatten and Brill (2011b) is applied in order to test whether the res-ults from Forbes and Rigobon (2002) are as robust as the authors argue.Therefore, the period of relative stability is set to start on January 1, 1993,and the rolling crisis windows to run from December 1, 1994, until Decem-ber 31, 1995. Again, both versions of the FR-test are applied for rollingcrisis windows of 20, 40, and 60 days. The signals are based on a 5% levelof significance where the approximated critical values from Andenmattenand Brill (2011b) are used (cf. Table A.1). Table 3.5 presents the numberof signals for contagion stemming from Mexico’s equity market based on2-day-moving-average returns in local currency. The sample countries werereduced by four, as China, India, Russia and South Africa were omitted dueto lack of adequate data.

Once again, in contrast to the static approach of Forbes and Rigobon(2002), the rolling-crisis-window approach of Andenmatten and Brill (2011b)yields evidence for contagion stemming from Mexico during the Mexicanpeso crisis in both versions of the FR-test and for the three different crisiswindows. The largest number of signals – 23 out of 27 cases – is observedfor the FRO-test for the 20-day window. Focusing on the relative frequencyof the signals across countries, it is apparent that Latin American countriesdominate. For Brazil, Argentina, and Chile, the FRO-test based on the20-day window yields 48, 24, and 17 signals, respectively. Nor is it only theequity markets of the Latin American countries that seem to be affectedby contagion stemming from the Mexican equity market: so were the equitymarkets of many other emerging countries and many industrialized countries.In East Asia, for instance, 8 out of 9 countries are affected by contagionstemming from the Mexican equity market. Among the OECD countries,only Belgium and the Netherlands seem unaffected. However, the numberof signals is, on average, lower outside Latin America.

Focusing on the 40- and 60-day windows, signals for fewer countriesare recorded but the regional pattern seems similar to the 20-day window.Accordingly, one conclusion that can be drawn from the FRO-test is thatnot only is there evidence for contagion stemming from the Mexican equitymarket but also a strong regional pattern.


Table 3.5Mexican Peso Crisis: Contagion Stemming from Mexico

This table records the number of contagion signals stemming from Mexico’s equity marketbased on 2-day-moving-average returns in local currency. The period of relative stabilitystarts on January 1, 1993, and the rolling crisis windows run from December 1, 1994, untilDecember 31, 1995. Both the FRO and the FRN test versions are applied for rollingcrisis windows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN


East AsiaHong Kong 22 18 9 10 6 0Indonesia 21 16 4 6 7 0Japan 12 23 33 13 29 49Korea 29 48 44 20 30 24Malaysia 0 0 0 0 0 0Philippines 4 17 16 0 0 0Singapore 12 8 0 9 3 0Taiwan 17 12 6 13 8 0Thailand 3 11 5 0 0 0

Latin AmericaArgentina 24 44 66 15 26 10Brazil 48 63 69 49 61 69Chile 17 11 0 16 1 0




This observation holds true for the FRN -test as well. However, the FRN -test seems more restrictive than the FRO-test, as the number of countriesfor which at least one signal is obtained decreases from 23 to 17 for the20-day window, from 20 to 13 for the 40-day window and from 18 to 7 forthe 60-day window. This confirms the results from the East Asian crisis.

Figure 3.9 illustrates the results for the equity markets of Mexico andArgentina. As can be seen from the three columns, the FRN -test is slightlymore restrictive than the FRO-test, but both versions of the FR-test yieldconsistent results, i.e. the periods of contagion based on the FRN -test matchthose of the FRO-test.

Figure 3.9 not only illustrates that not only were there periods of conta-gion but also that these periods were distinct from the crisis period Forbesand Rigobon (2002) relied on, namely the two weeks following December 19,1994, the day the exchange rate regime was abandoned. Instead, all panelsof Figure 3.9 indicate that the period of contagion only started in mid-1995,when the recovery was under way. This is similar to the findings of the testsfor contagion stemming from Hong Kong during the East Asian crisis.

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/93 01/94 01/95 01/96

MexicoArgentinaContagion

FRO: 20-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/93 01/94 01/95 01/96

FRO: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/93 01/94 01/95 01/96

FRO: 60-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/93 01/94 01/95 01/96

FRN: 20-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/93 01/94 01/95 01/96

FRN: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/93 01/94 01/95 01/96

FRN: 60-Day Window

Figure 3.9: Contagion Stemming from Mexico: Signals for Argentina

3.5 Contagion During the U.S. Stock Market Crash 89

3.5 Contagion During the U.S. Stock Market Crash

In this section, the rolling-crisis-window approach of Andenmatten and Brill(2011b) is applied on the U.S. stock market crash in October 1987 – a periodthat is the focus of many of the early empirical papers testing for contagionin equity markets.10

On October 19, 1987, also referred to as Black Monday, U.S. stock mar-kets crashed sending shock waves around the globe. The S&P 500 stockmarket index, for instance, fell by about 20% on that day alone and so didmany other stock markets, as illustrated in Figure 3.10.

40

60

80

100

120

140

MSC

I Ind

ex in

Loc

al C

urre

ncy,

Reb

ased

06/1987 08/1987 10/1987 12/1987 02/1988

U.S.

UK

Australia

Canada

France

Germany

Hong Kong

Japan

Netherlands

Switzerland


Figure 3.10: Stock Market Indices During the U.S. Market Crash

At first glance, this seems to be a perfect example of contagion as the (unad-justed) cross-market correlations increased significantly – a result that wasprominently advocated by King and Wadhwani (1990), among others. Thisview especially tempting since a clear catalyst for contagion can be identi-fied, namely, the weakness of trading systems at that time. For instance,Carlson (2006) argues: “The market crash of 1987 is a significant event notjust because of the swiftness and severity of the market decline, but also be-cause it showed the weaknesses of the trading systems themselves and how10Cf., for instance, Bertero and Mayer (1990), Hamao, Masulis, and Ng (1990), King and

Wadhwani (1990), and Lee and Kim (1993).


they could be strained and come close to breaking in extreme conditions.The problems in the trading systems interacted with the price declines tomake the crisis worse.”

However, using data from the ten largest stock markets, Forbes andRigobon (2002) find virtually no evidence for contagion after adjusting thecorrelation coefficient for the higher volatility during the crash. For theirbase case, Forbes and Rigobon define the period of relative stability runningfrom January 1, 1986, to the crash in October 1987, and the crisis periodlasting from October 19 to December 4, 1987. For calculating the dailyreturns, they rely on equity indices in US dollars. Their result – neatlysummarized as “no contagion, just interdependence” – is supported by ro-bustness tests where they alter the period definitions and use equity returnsin local currency instead of US dollars.

Applying the rolling-crisis-window approach of Andenmatten and Brill(2011b), the period of relative stability starts on January 1, 1986 and therolling crisis windows run from October 1, 1987, until September 30, 1988.

Table 3.6U.S. Stock Market Crash: Contagion Stemming from the U.S.

This table records the number of contagion signals stemming from the U.S. equity marketbased on 2-day-moving-average returns in local currency. The period of relative stabilitystarts on January 1, 1986, and the rolling crisis windows run from October 1, 1987, untilSeptember 30, 1988. Both the FRO and the FRN test versions are applied for rollingcrisis windows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN


Australia 0 0 0 1 1 1Canada 19 39 55 20 45 55France 11 23 36 9 20 44Germany 25 37 67 23 26 63Hong Kong 31 44 61 29 42 66Japan 3 7 2 0 2 0Netherlands 0 8 28 0 27 42Switzerland 6 11 0 6 14 0UK 0 9 8 1 22 53


Once again, both versions of the FR-test are applied for rolling crisis win-dows of 20, 40, and 60 days, and the signals are based on a 5% level ofsignificance using the approximated critical values from Andenmatten andBrill (2011b) (cf. Table A.1). Table 3.6 reports the number of signals forcontagion stemming from the U.S. equity market based on 2-day-moving-average returns in local currency. The sample of countries is reduced to theten largest stock markets at that time (including that of the U.S.).

Testing for contagion stemming from the U.S. after the U.S. stock marketcrash, the rolling-crisis-window approach of Andenmatten and Brill (2011b)again yields evidence for contagion for both versions of the FR-test and forthe three different crisis windows. Australia is the only country where oneof the two test versions, namely the FRO-test, yields no signal for contagionfor the different time windows. It is worth noting that the average numberof signals increases with the length of the time window for both versions ofthe FR-test. What is more, the FRN -tests seem to be less restrictive thanin the previous section. In fact, with 46 signals, the FRN -test based onthe 60-day window yields the largest average number of signals for countrieswhere at least one signal is recorded.

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/86 01/87 01/88 01/89

USAGermanyContagion

FRO: 20-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/86 01/87 01/88 01/89

FRO: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/86 01/87 01/88 01/89

FRO: 60-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/86 01/87 01/88 01/89

FRN: 20-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/86 01/87 01/88 01/89

FRN: 40-Day Window

0

20

40

60

80

100

MSC

I Ind

ex, R

ebas

ed

01/86 01/87 01/88 01/89

FRN: 60-Day Window

Figure 3.11: Contagion Stemming from the U.S.: Signals for Germany

3.6 Conclusion 92

Figure 3.11 illustrates the results for the equity markets of the U.S. andGermany. As can be seen in the three columns, both versions of the FR-testidentify almost identical periods of contagion. However, it is also apparentthat no contagion is identified for the period right after the stock marketcrash, in late October 1987. Instead, the periods of contagion start onlyin the beginning of 1988. So, on the one hand, the rolling-crisis-windowapproach confirms the result from Forbes and Rigobon (2002) – namely “nocontagion, only interdependence” – for the crisis period they analysed. Onthe other hand, it once again suggests that the results from Forbes andRigobon (2002) might not be robust with regard to shifting the period ofturmoil. Moreover, Figure 3.11 illustrates a similar pattern as seen in theprevious sections, namely that signals for contagion are obtained mainlyduring periods of recovery and not around the peak of a crisis.

3.6 Conclusion

Motivated by the analysis of Andenmatten and Brill (2011b), this paperaimed at challenging the robustness of the results of Forbes and Rigobon(2002) by testing for contagion with the rolling-crisis-window approach dur-ing the same crises Forbes and Rigobon (2002) focused on, namely the EastAsian crisis in 1997-98, the Mexican peso crisis in 1994, and the U.S. stockmarket crash in 1987. In contrast to the result of Forbes and Rigobon (2002)summarised as “no contagion, only interdependence,” applying the rolling-crisis-window approach of Andenmatten and Brill (2011b) on 2-day-movingaverages of daily returns of MSCI equity market indices in local currencyyields broad evidence for contagion for both versions of the FR-test and forthe three different crisis windows during these crisis. This result is suppor-ted by extensive robustness tests, presented in Appendix B, that altered theperiods of relative stability and used daily returns in US dollars instead ofthe local currency.

Examining the details of the test results, three observations stand out:First, there seems to be evidence not only of contagion but also of a strongregional pattern during the East Asian crisis in 1997-98 and the Mexicanpeso crisis in 1994. In other words, Asian and Latin American countriesdominated in the relative frequency of contagion signals during the respect-ive crisis. Such a regional pattern is impossible to observe during the U.S.

3.6 Conclusion 93

stock market crash in 1987 due to the smaller sample of countries.The second observation is that the FRN -test is more restrictive com-

pared to the FRO-test, because, in most cases, the number of countrieshaving at least one signal decreases. As discussed in section 2, this was tobe expected, as the standard errors of the FRO-test are likely to be biasedupward due to the overlapping nature of the periods of relative turmoiland stability. This pattern is illustrated by Figures 3.4, 3.5, 3.6, 3.7, and3.9, where it can be seen that, compared to the FRO-test, the periods ofcontagion are shorter or even non-existent for the FRN -test. However, bothversions of the FR-test yield consistent results, with the periods of contagionbased on the FRN -test matching those of the FRO-test.

The third notable observation, visible in Figures 3.4, 3.5, 3.6, 3.7, 3.9,and 3.11, is that not only did periods of contagion occur but they alsoseemed to be distinct from the crisis periods defined by Forbes and Rigobon(2002). Thus, the rolling-crisis-window approach both confirms the resultfrom Forbes and Rigobon (2002) of “no contagion, only interdependence”for the crisis periods they defined, and it also suggests their results may notbe robust with regard to shifting the period of turmoil. Moreover, the rolling-crisis-window approach reveals a common pattern, namely, that signals forcontagion are obtained mainly during periods of recovery and not aroundthe peak of a crisis.

Appendix A

Approximate Critical Values

for the Rolling FR-Tests

Similar to testing for a structural break at an unknown break date, wherethe so-called sup F-statistic is the largest of many F-statistics and, hence,its distribution is not the same as an individual F-statistic, the distribu-tion of the FR-test statistic is not the same as the standard t-distribution.Andrews (1993) derived the asymptotic distribution for a wide class of testsfor structural change, among them the sup F-test. However, the asymptoticdistribution of these tests is non-standard and depends both on the numberof restrictions being tested, i.e. the number of coefficients that are beingallowed to break, and the range of the subsample over which the F-statisticsare computed. Hansen (1997) developed computationally convenient approx-imations to the asymptotic p-value functions for the Andrews asymptoticdistributions.

Applying the approach of Andrews (1993) and Hansen (1997), respect-ively, to the rolling FR-test approach is out of scope of this paper. Instead,Monte Carlo methods are used to find approximate critical values for therolling FR-test statistics.

In order to do that we assume that the samples of daily returns are drawnfrom a bivariate normal distribution. Then, for a given correlation betweenthe two variables, we draw a sample of the bivariate normal distribution. Inorder to match the characteristics of our sample of CDS premia, we definethe non-crisis period as the first 260 observations. This should be sufficientas Essaadi, Jouini, and Khallouli (2007) show that the minimum for a stable

Appendix A 96

rolling correlation coefficient is 224 observations (on a 5% percent level ofsignificance). This is illustrated by figure A.1.

We then run both versions of the FR-test for the three different timewindows and store the respective test statistics. We repeat this procedureuntil we have collected for each of the different test versions 10,000 teststatistics. Figure A.2 illustrates that 10,000 observations should be sufficientto determine stable approximate critical values.

Now, we can determine the approximate critical values for the correla-tion structure we used for drawing the repeated samples from the bivariatenormal distribution. We do that for three different levels of significance,namelely 1%, 5%, and 10%, by using the respective percentiles of the ob-served distribution of the test statistic.

Finally, we repeat this procedure for different correlation coefficients.The results are presented in table A.1.

Appendix A 97

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0Co

rrela

tion

Coef

ficie

nt

0 200 400 600 800 1000Number of Observations

Figure A.1: Rolling Correlation Coefficient

2.0

2.5

3.0

3.5

Appr

oxim

ate

Critic

al V

alue

for �

=0.0

5

0 10,000 20,000 30,000Number of Tests

Figure A.2: Approximate Critical Value for α = 0.05

Appendix A 98

Table A.1Approximate Critical Values for the Rolling FR-Tests

This table shows approximate critical values for the different versions of the rolling FR-testbased on Monte Carlo simulations with 10,000 repeated tests.

Panel A: α = 0.01

FRO FRN

Correlation coefficient 20 days 40 days 60 days 20 days 40 days 60 days

0.00 3.595 3.081 2.581 3.688 3.375 3.0460.10 3.827 3.445 3.147 3.937 3.916 3.7580.20 3.877 3.267 3.073 4.125 3.736 3.4390.30 3.411 3.209 3.047 3.404 3.248 3.2970.40 3.584 3.499 3.224 3.684 3.712 3.5740.50 3.635 3.462 3.467 3.877 3.839 4.2180.60 3.389 3.093 2.509 3.465 3.188 2.8130.70 3.380 3.031 2.583 3.528 3.257 2.9070.80 3.342 3.019 2.830 3.439 3.185 3.1030.90 3.125 3.016 2.567 3.171 3.056 2.686

Panel B: α = 0.05

FRO FRN


0.00 2.542 2.147 1.959 2.610 2.349 2.2400.10 2.679 2.241 2.066 2.747 2.418 2.3620.20 2.700 2.493 2.251 2.851 2.780 2.6240.30 2.433 2.255 2.135 2.451 2.422 2.3660.40 2.735 2.477 2.440 2.805 2.683 2.6240.50 2.628 2.405 2.363 2.721 2.602 2.6890.60 2.588 2.107 1.836 2.616 2.221 2.0260.70 2.433 2.159 1.980 2.530 2.342 2.1510.80 2.494 2.203 2.059 2.563 2.351 2.3650.90 2.382 2.078 1.866 2.371 2.200 2.023

Panel C: α = 0.10

FRO FRN


0.00 1.958 1.647 1.528 2.019 1.819 1.8130.10 2.030 1.688 1.617 2.100 1.829 1.8330.20 2.152 1.956 1.889 2.270 2.203 2.1480.30 1.954 1.791 1.737 1.990 1.858 1.8160.40 2.217 2.004 1.953 2.298 2.120 2.1500.50 2.065 1.868 1.787 2.131 1.972 1.9410.60 2.034 1.633 1.467 2.090 1.749 1.6330.70 2.000 1.731 1.586 2.040 1.823 1.7430.80 2.048 1.795 1.720 2.111 1.981 1.9520.90 1.953 1.645 1.494 1.963 1.706 1.551

Appendix B

Robustness Tests

Tables B.1 to B.18 report the results of extensive robustness tests of theanalysis in Chapter 3. The robustness tests were performed with respect tothe stability period (Tables B.1 to B.12) and with respect to using equitymarket indices in US dollar instead of local currency (Tables B.13 to B.18).

Appendix B 100

Table B.1East Asian Crisis: Contagion Stemming from Hong Kong

This table records the number of contagion signals stemming from the Hong Kong equitymarket based on 2-day-moving-average returns in local currency. The period of relativestability starts on January 1, 1995, and the rolling crisis windows run from June 1, 1997,until August 31, 1998. Both the FRO and the FRN test versions are applied for rollingcrisis windows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN







Appendix B 101


This table records the number of contagion signals stemming from the Hong Kong equitymarket based on 2-day-moving-average returns in local currency. The period of relativestability starts on January 1, 1993, and the rolling crisis windows run from June 1, 1997,until August 31, 1998. Both the FRO and the FRN test versions are applied for rollingcrisis windows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN







Appendix B 102

Table B.3East Asian Crisis: Contagion Stemming from Thailand


FRO FRN







Appendix B 103



FRO FRN







Appendix B 104

Table B.5East Asian Crisis: Contagion Stemming from Indonesia


FRO FRN







Appendix B 105



FRO FRN







Appendix B 106

Table B.7East Asian Crisis: Contagion Stemming from Korea


FRO FRN







Appendix B 107



FRO FRN







Appendix B 108

Table B.9Mexican Peso Crisis: Contagion Stemming from Mexico


FRO FRN






Appendix B 109



FRO FRN






Appendix B 110

Table B.11U.S. Stock Market Crash: Contagion Stemming from the U.S.


FRO FRN





FRO FRN



Appendix B 111


This table records the number of contagion signals stemming from the Hong Kong equitymarket based on 2-day-moving-average returns in US dollar. The period of relative sta-bility starts on January 1, 1996, and the rolling crisis windows run from June 1, 1997,until August 31, 1998. Both the FRO and the FRN test versions are applied for rollingcrisis windows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN







Appendix B 112


This table records the number of contagion signals stemming from Thailand’s equity mar-ket based on 2-day-moving-average returns in US dollar. The period of relative stabilitystarts on January 1, 1996, and the rolling crisis windows run from June 1, 1997, untilAugust 31, 1998. Both the FRO and the FRN test versions are applied for rolling crisiswindows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN







Appendix B 113


This table records the number of contagion signals stemming from Indonesia’s equity mar-ket based on 2-day-moving-average returns in US dollar. The period of relative stabilitystarts on January 1, 1996, and the rolling crisis windows run from June 1, 1997, untilAugust 31, 1998. Both the FRO and the FRN test versions are applied for rolling crisiswindows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN







Appendix B 114


This table records the number of contagion signals stemming from Korea’s equity marketbased on 2-day-moving-average returns in US dollar. The period of relative stability startson January 1, 1996, and the rolling crisis windows run from June 1, 1997, until August31, 1998. Both the FRO and the FRN test versions are applied for rolling crisis windowsof 20, 40, and 60 days, respectively. The signals are based on a 5% level of significancewhere the approximated critical values from Andenmatten and Brill (2011b) are used (cf.Table A.1).

FRO FRN







Appendix B 115


This table records the number of contagion signals stemming from Mexico’s equity marketbased on 2-day-moving-average returns in US dollar. The period of relative stabilitystarts on January 1, 1993, and the rolling crisis windows run from December 1, 1994, untilDecember 31, 1995. Both the FRO and the FRN test versions are applied for rollingcrisis windows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN






Appendix B 116


This table records the number of contagion signals stemming from the U.S. equity marketbased on 2-day-moving-average returns in US dollar. The period of relative stabilitystarts on January 1, 1986, and the rolling crisis windows run from October 1, 1987, untilSeptember 30, 1988. Both the FRO and the FRN test versions are applied for rollingcrisis windows of 20, 40, and 60 days, respectively. The signals are based on a 5% level ofsignificance where the approximated critical values from Andenmatten and Brill (2011b)are used (cf. Table A.1).

FRO FRN



References

Agenor, P.-R., & Aizenman, J. (1998). Contagion and volatility withimperfect credit markets. IMF Staff Papers, 45 (2), 207-235.

Allen, F., & Gale, D. (2000). Financial contagion. The Journal of PoliticalEconomy , 108 (1), 1-33.

Amihud, Y., Mendelson, H., & Pedersen, L. H. (2006). Liquidity and assetprices. now Publishers Inc.

Ammer, J., & Cai, F. (2007). Sovereign CDS and bond pricing dynamicsin emerging markets: Does the cheapest-to-deliver option matter? (Inter-national Finance Discussion Papers No. 912). Board of Governors of theFederal Reserve System (U.S.).

Andenmatten, S., & Brill, F. (2011a). Did the CDS market push up riskpremia for sovereign credit? Swiss Journal of Economics and Statistics,147 (3), 275-302.

Andenmatten, S., & Brill, F. (2011b). Measuring co-movements of CDSpremia during the Greek debt crisis (Discussion Papers No. dp1104). Uni-versity of Bern, Department of Economics.

Anderson, R. W. (2010). Credit default swaps: What are the social benefitsand costs? Banque de France Financial Stability Review , 14 , 1-14.

Andrews, D. W. K. (1993). Tests for parameter instability and structuralchange with unknown change point. Econometrica, 61 (4), 821-856. (Cor-rigendum, 71(1), 395-397)

Bae, K.-H., Karolyi, G. A., & Stulz, R. M. (2003). A new approach tomeasuring financial contagion. The Review of Financial Studies, 16 (3),717-763.

References 118

Banerjee, A. V. (1992). A simple model of herd behavior. The QuarterlyJournal of Economics, 107 (3), 797-817.

Baur, D., & Fry, R. (2006). Endogenous contagion - a panel data ana-lysis (CAMA Working Paper Series No. 9/2006). The Australian NationalUniversity. Centre for Applied Macroeconomic Analysis.

Bekaert, G., Harvey, C. R., & Ng, A. (2005). Market integration andcontagion. The Journal of Business, 78 (1), 39-69.

Bertero, E., & Mayer, C. (1990). Structure and performance: Global inter-dependence of stock markets around the crash of October 1987. EuropeanEconomic Review , 34 (6), 1155-1180.

Bikhchandani, S., Hirshleifer, D., & Welch, I. (1992). A theory of fads,fashion, custom, and cultural change as informational cascades. Journal ofPolitical Economy , 100 (5), 992-1026.

Blanco, R., Brennan, S., & Marsh, I. W. (2005). An empirical analysis ofthe dynamic relation between investment-grade bonds and credit defaultswaps. Journal of Finance, 60 (5), 2255-2281.

Boyer, B. H., Gibson, M. S., & Loretan, M. (1997). Pitfalls in tests forchanges in correlations (International Finance Discussion Papers No. 597).Board of Governors of the Federal Reserve System (U.S.).

Brill, F. (2011). Testing for contagion with a rolling-crisis-window ap-proach. (Unpublished Working Paper)

Calvo, G. A., & Mendoza, E. G. (2000). Rational contagion and theglobalization of securities markets. Journal of International Economics,51 (1), 79-113.

Calvo, S., & Reinhart, C. (1996). Capital flows to Latin America: Is thereevidence of contagion effects? (Policy Research Working Paper Series No.1619). The World Bank.

Caramazza, F., Ricci, L. A., & Salgado, R. (2000). Trade and financialcontagion in currency crises (Working Paper No. 00/55). InternationalMonetary Fund (IMF).

References 119

Carlson, M. (2006). A brief history of the 1987 stock market crash witha discussion of the Federal Reserve response (Finance and Economics Dis-cussion Series No. 2007-13). Board of Governors of the Federal ReserveSystem (U.S.).

Chue, T. K. (2002). Time-varying risk preferences and emerging marketco-movements. Journal of International Money and Finance, 21 (7), 1053-1072.

Chuhan, P., Claessens, S., & Mamingi, N. (1998). Equity and bond flowsto Latin America and Asia: The role of global and country factors. Journalof Development Economics, 55 (2), 439-463.

Codogno, L., Favero, C., & Missale, A. (2003). Yield spreads on EMUgovernment bonds. Economic Policy , 18 (37), 503-532.

Corsetti, G., Pericoli, M., & Sbracia, M. (2001). Correlation analysis offinancial contagion: What one should know before running a test (WorkingPaper No. 822). Economic Growth Center, Yale University.

Corsetti, G., Pericoli, M., & Sbracia, M. (2005). ‘Some contagion, someinterdependence’: More pitfalls in tests of financial contagion. Journal ofInternational Money and Finance, 24 (8), 1177-1199.

Corsetti, G., Pesenti, P., & Roubini, N. (1998). What caused the Asiancurrency and financial crisis? part I: A macroeconomic overview (WorkingPaper No. 6833). National Bureau of Economic Research (NBER).

Corsetti, G., Pesenti, P., Roubini, N., & Tille, C. (1999). Competitive de-valuations: A welfare-based approach (Working Paper No. 6889). NationalBureau of Economic Research (NBER).

Dai, Q., & Singleton, K. (2003). Term structure dynamics in theory andreality. The Review of Financial Studies, 16 (3), 631-678.

Diamond, D. W., & Dybvig, P. H. (1983). Bank runs, deposit insurance,and liquidity. Journal of Political Economy , 91 (3), 401-419.

Dornbusch, R. (1997). Brazil’s incomplete stabilization and reform. Brook-ings Papers on Economic Activity , 1997 (1), 367-404.

References 120

Dornbusch, R., Park, Y. C., & Claessens, S. (2000). Contagion: Un-derstanding how it spreads. The World Bank research observer , 15 (2),177-198.

Drazen, A. (1999). Political contagion in currency crises (Working PaperNo. 7211). National Bureau of Economic Research (NBER).

Dresdner Kleinwort Wasserstein Research. (2002). Credit default swaps:A product overview (Research Paper). Dresdner Kleinwort WassersteinResearch.

Duffie, D. (1999). Credit swap valuation. Financial Analysts Journal , 55 ,73-87.

Duffie, D. (2010). Is there a case for banning short speculation in sovereignbond markets? Banque de France Financial Stability Review , 14 , 55-60.

Duffie, D., & Singleton, K. J. (1999). Modeling term structures of default-able bonds. The Review of Financial Studies, 12 (4), 687-720.

Dungey, M., Fry, R., Gonzalez-Hermosillo, B., & Martin, V. L. (2005).Empirical modelling of contagion: A review of methodologies. QuantitativeFinance, 5 (1), 9-24.

Dungey, M., Fry, R., & Martin, V. L. (2006). Correlation, contagion, andAsian evidence. Asian Economic Papers, 5 (2), 32-72.

Eichengreen, B., Rose, A. K., & Wyplosz, C. (1995). Exchange marketmayhem: The antecedents and aftermath of speculative attacks. EconomicPolicy , 10 (2), 251-312.

Eichengreen, B., Rose, A. K., & Wyplosz, C. (1996). Contagious currencycrises (Working Paper No. 5681). National Bureau of Economic Research(NBER).

Engle, R. F., & Granger, C. W. J. (1987). Co-integration and error cor-rection: Representation, estimation, and testing. Econometrica, 55 (2),251-276.

Essaadi, E., Jouini, J., & Khallouli, W. (2007). The Asian crisis conta-gion: A dynamic correlation approach analysis (Working Paper No. 0725).Groupe d’Analyse et de Theorie Economique (GATE).

References 121

Favero, C., Pagano, M., & Thadden, E.-L. von. (2007). How does liquidityaffect government bond yields? (Working Paper No. 323). IGIER (Inno-cenzo Gasparini Institute for Economic Research), Bocconi University.

Favero, C. A., & Giavazzi, F. (2002). Is the international propagation offinancial shocks non-linear? Evidence from the ERM. Journal of Interna-tional Economics, 57 (1), 231-246.

Forbes, K., & Rigobon, R. (2001). Contagion in Latin America: Definitions,measurement, and policy implications. Economia, 1 (2), 1-46.

Forbes, K., & Rigobon, R. (2002). No contagion, only interdependence:Measuring stock market comovements. Journal of Finance, 57 (5), 2223-2261.

Garber, P. M. (1998). Derivatives in international capital flows (WorkingPaper No. 6623). National Bureau of Economic Research (NBER).

Gerlach, S., & Smets, F. (1995). Contagious speculative attacks. EuropeanJournal of Political Economy , 11 (1), 45-63.

Geyer, A., Kossmeier, S., & Pichler, S. (2004). Measuring systematic riskin EMU government yield spreads. Review of Finance, 8 (2), 171-197.

Ghosh, A., Hagemans, R., Leeming, M., & Willemann, S. (2010). Standardcorporate CDS handbook, ongoing evolution of the CDS markets (CreditDerivative and Quantitative Strategy). Barclays Capital.

Gonzalo, J., & Granger, C. W. J. (1995). Estimation of common long-memory components in cointegrated systems. Journal of Business & Eco-nomic Statistics, 13 (1), 27-35.

Hakkio, C. S., & Rush, M. (1991). Cointegration: How short is the longrun? Journal of International Money and Finance, 10 (4), 571-581.

Hamao, Y., Masulis, R. W., & Ng, V. (1990). Correlations in price changesand volatility across international stock markets. The Review of FinancialStudies, 3 (2), 281-307.

Hansen, B. E. (1997). Approximate asymptotic p values for structural-change tests. Journal of Business & Economic Statistics, 15 (1), 60-67.

References 122

Hull, J., Predescu, M., & White, A. (2004). The relationship betweencredit default swap spreads, bond yields, and credit rating announcements.Journal of Banking & Finance, 28 (11), 2789-2811.

Jeanne, O. (1997). Are currency crises self-fulfilling? A test. Journal ofInternational Economics, 43 (3-4), 263-286.

Johansen, S. (1991). Estimation and hypothesis testing of cointegrationvectors in Gaussian vector autoregressive models. Econometrica, 59 (6),1551-1580.

JP Morgan Research. (2001). Emerging market credit derivatives (ResearchPaper). JP Morgan.

Kaminsky, G. L., & Reinhart, C. M. (2000). On crises, contagion, andconfusion. Journal of International Economics, 51 (1), 145-168.

Kaminsky, G. L., & Schmukler, S. L. (1999). What triggers market jit-ters? A chronicle of the Asian crisis. Journal of International Money andFinance, 18 (4), 537-560.

Karolyi, G. A., & Stulz, R. M. (1996). Why do markets move together? Aninvestigation of U.S.-Japan stock return comovements. Journal of Finance,51 (3), 951-986.

Kendall, M., & Stuart, A. (1969). The advanced theory of statistics (Vol. 1).Charles Griffin and Co.

King, M. A., & Wadhwani, S. (1990). Transmission of volatility betweenstock markets. The Review of Financial Studies, 3 (1), 5-33.

Kodres, L. E., & Pritsker, M. (2002). A rational expectations model offinancial contagion. Journal of Finance, 57 (2), 769-799.

Kyle, A. S., & Xiong, W. (2001). Contagion as a wealth effect. Journal ofFinance, 56 (4), 1401-1440.

Lee, S. B., & Kim, K. J. (1993). Does the October 1987 crash strengthenthe co-movements among national stock markets? Review of FinancialEconomics, 3 (1), 89-102.

References 123

Liu, J., & Longstaff, F. A. (2004). Losing money on arbitrage: Optimaldynamic portfolio choice in markets with arbitrage opportunities. TheReview of Financial Studies, 17 (3), 611-641.

Longin, F., & Solnik, B. (1995). Is the correlation in international equityreturns constant: 1960-1990? Journal of International Money and Finance,14 (1), 3-26.

Longstaff, F. A. (2004). The flight-to-liquidity premium in U.S. Treasurybond prices. The Journal of Business, 77 (3), 511-526.

Longstaff, F. A., Pan, J., Pedersen, L. H., & Singleton, K. (2011). Howsovereign is sovereign credit risk? American Economic Journal: Macroe-conomics, 3 (2), 75-103.

Loretan, M., & English, W. B. (2000). Evaluating “correlation breakdowns”during periods of market volatility (International Finance Discussion Pa-pers No. 658). Board of Governors of the Federal Reserve System (U.S.).

Luetkepohl, H. (2005). New introduction to multiple time series analysis.Heidelberg: Springer.

Masson, P. R. (1998). Contagion-monsoonal effects, spillovers, and jumpsbetween multiple equilibria (Working Paper No. 98/142). International Mon-etary Fund (IMF).

Packer, F., & Suthiphongchai, C. (2003). Sovereign credit default swaps.BIS Quarterly Review , 2003 , 79-88.

Pan, J., & Singleton, K. J. (2008). Default and recovery implicit in the termstructure of sovereign CDS spreads. Journal of Finance, 63 (5), 2345-2384.

Park, Y. C. (1998). The financial crisis in Korea and its lessons for reformof the international financial systems. In J. J. Teunissen & A. Akkerman(Eds.), Diversity in development - Reconsidering the Washington consensus(p. 25-64). The Hague: FONDAD.

Pedersen, L. H., & Brunnermeier, M. K. (2007). Market liquidity andfunding liquidity (Working Paper No. 12939). National Bureau of EconomicResearch (NBER).

References 124

Pericoli, M., & Sbracia, M. (2003). A primer on financial contagion. Journalof Economic Surveys, 17 (4), 571-608.

Perron, P. (2005). Dealing with structural breaks. In T. C. Mills & K. Pat-terson (Eds.), Palgrave handbook of econometrics, volume 1: Econometrictheory (chap. IV). Palgrave Macmillan.

Pritsker, M. (2001). The channels for financial contagion. In S. Claessens& K. J. Forbes (Eds.), International financial contagion (chap. 4). KluwerAcademic Publishers.

Quandt, R. E. (1960). Tests of the hypothesis that a linear regressionsystem obeys two separate regimes. Journal of the American StatisticalAssociation, 55 (290), 324-330.

Radelet, S., & Sachs, J. (1998a). The East Asian financial crisis: Diagnosis,remedies, prospects. Brookings Papers on Economic Activity , 1998 (1), 1-90.

Radelet, S., & Sachs, J. (1998b). The onset of the East Asian financialcrisis (Working Paper No. 6680). National Bureau of Economic Research(NBER).

Rigobon, R. (2002). Contagion: How to measure it? In S. Edwards& J. A. Frankel (Eds.), Preventing currency crises in emerging markets(p. 269-334). University of Chicago Press.

Scharfstein, D. S., & Stein, J. C. (1990). Herd behavior and investment.The American Economic Review , 80 (3), 465-479.

Schinasi, G. J., & Smith, R. T. (2000). Portfolio diversification, leverage,and financial contagion. IMF Staff Papers, 47 (2), 159-176.

Sgherri, S., & Zoli, E. (2009). Euro area sovereign risk during the crisis(Working Paper No. 09/222). International Monetary Fund (IMF).

Shiller, R. J. (1995). Conversation, information, and herd behavior. TheAmerican Economic Review , 85 (2), 181-185.

Stock, J. H., & Watson, M. W. (2007). Introduction to econometrics (2nded.). Addison Wesley.

References 125

Stulz, R. M. (2010). Credit default swaps and the credit crisis. Journal ofEconomic Perspectives, 24 (1), 73-92.

Van Rijckeghem, C., & Weder, B. (2001). Sources of contagion: Is itfinance or trade? Journal of International Economics, 54 (2), 293-308.

Verdier, P.-H. (2004). Credit derivatives and the sovereign debt restructur-ing process (LL.M. Paper). Harvard Law School.

Wermers, R. (1999). Mutual fund herding and the impact on stock prices.The Journal of Finance, 54 (2), 581-622.

Whitt, J. A. (1996). The Mexican peso crisis. Economic Review , 81 (1),1-20.

Zhu, H. (2006). An empirical comparison of credit spreads between thebond market and the credit default swap market. Journal of FinancialServices Research, 29 (3), 211-235.

Documents

Sovereign Credit Risk and Contagion