Thesis SCM Nivi

EMPIRICAL STUDIES ON CREDIT MARKETS
ISBN 90 5170 713 4 Cover design: Crasborn Graphic Designers bno, Valkenburg a.d. Geul This book is number 320 of the Tinbergen Institute Research Series, established through cooperation between Thela Thesis and the Tinbergen Institute. A list of books which already appeared in the series can be found in the back. This PhD project was partly nanced by the Risk Management & Modelling department of Rabobank International.
Empirical Studies on Credit Markets

(Empirisch onderzoek naar markten voor kredieten)
PROEFSCHRIFT
ter verkrijging van de graad van doctor aan de Erasmus Universiteit Rotterdam op gezag van de Rector Magnicus Prof.dr.ir. J.H. van Bemmel en volgens besluit van het College voor Promoties
De openbare verdediging zal plaatsvinden op vrijdag 3 oktober 2003 om 13.30 uur door
Patrick Houweling
geboren te Leiden
Promotiecommissie Promotor: Overige leden: Prof.dr. A.C.F. Vorst Prof.dr. D. Lando Prof.dr. A.A.J. Pelsser Prof.dr. M.J.C.M. Verbeek
Acknowledgements
Writing my PhD thesis has been the single largest project I ever conducted, so these thank yous may be the most important ones I will ever say. First, and foremost, I would like to thank my supervisor Ton Vorst. Your guidance in nding my way in the world of nancial research has been invaluable. You have taught me a lot about doing research in general and about credit markets in particular. Your comments on early versions of our papers were always relevant and helped me to improve them again and again. I am equally thankful to Albert Mentink, with whom I have worked on several papers, two of which now appear in this thesis. Finding out that you were also doing a PhD on credit markets, and seeing your enthusiasm to start a joint project, has denitely been the turning point in my time as a PhD student. I greatly beneted from your knowledge, refreshing ideas and project management skills. I am also grateful to my other co-authors Jaap Hoek and Frank Kleibergen. Working on our paper, which is now a chapter in this thesis, has been a very instructive and rewarding experience. Over the years, I shared my room at the university with various PhD students: David zuur! Dekker, Jedid-Jah ik zou haast zeggen inteGENdeel Jonker, Erjen deeeze gast van Nierop and Richard had ik dit eerder geweten Kleijn. I sure had a great time with you guys. Although we kept the ball rolling, we also had time to relax with a cup of tea with our other Friends: Dennis het was weer reuze gezellig Fok, Klaas mwuuuh Staal and Bj orn wheee! Vroomen. I also enjoyed the time I spent with the rest of the lunch gang: Wilco van den Heuvel, Joost Loef, Ivo Nobel, Rutger van Oest, Richard Paap, Kevin Pak and Pim van Vliet. Our conversations were sometimes serious, sometimes hilarious, but always entertaining. I much enjoyed occasional chats with Reimer Beneder, Marisa de Brito, Jeab Cumperayot, Anna Gutkovska, Dennis Huisman, Jos van Iwaarden, Daina Konter, Roy Kouwenberg, Bert Menkveld, Robin Nicolai, Bart Oldenkamp, Em oke Oldenkamp, Ioulia Ossokina, Alexander Otgaar, Lennie Pattikawa, Rom Phisalaphong, Raol Pietersz, Lidewey van der Sluis, Jan-Frederik Slijkerman, Mari elle Sonnenberg, Yulia Veld, Ingrid Verheul and Martijn van der Voort. Further, I would like to thank the sta at
vi
the Tinbergen Institute and the Econometric Institute for excellent support. In particular, I am indebted to Carine Horbach, Tineke Kurtz, Carien de Ruijter and especially Elli Hoek van Dijke. During most of my time as a PhD student, I held a part-time position at the Risk Management & Modelling department of Rabobank International. Although I changed rooms and roommates even more often than the department changed its name, I really felt at home. I am especially grateful to Theo Kocken, who hired me at RI. You have given me the unique opportunity to combine my academic work at the university with more practical projects at the bank. I am also thankful to Kees van den Berg. You were my guide in the bank and opened doors that would otherwise have remained closed for me. My roommates made my four years at RI not only a very instructive, but also a very enjoyable period. I especially owe a big thank you to Freddy spa-n-sma van Dijk, Walter copula Foppen, M ac e kan ik hier even een kopietje van maken Mesters, Joeri van der Tonnekreek Potters, Erik okay dan! van Raaij, Marion ik ben even naar een meeting Segeren and Sacha homo! van Weeren. I further thank my other colleagues of the Modelling & Research team: Natalia Borokyvh, Martijn Derix, Elles Jongenelen, Estelle Jonkergouw, Adrian K uckler, Frans Ligtenberg, Roger Lord, Erwin Sandee, Harmen-Jan Sijtsma and Krishna Varu. I derived much pleasure from working on the Specic Risk project with risk managers Gerben Hagedoorn and Micha Schipper. I have learnt a lot from our many discussions in meetings, talks, phone calls and e-mails. From the London branch, I thank Andrew Gates. You have answered more questions on credit markets and credit derivatives than I can ever thank you for. I liked my contacts with Jan van den Bovenkamp, Sander van Geloven, Jan-Willem de Koning and Ren e van der Pol from the IT department. I appreciate our co-operation in developing, maintaining and extending PHsim and Rates. Support from the data base team has been a big timesaver. I particularly give thanks to Marit de Brouwer, Rebecca Groenhuis and Marcel Molenaar. Finally, I thank Marjolijn Benneker-Faber and Annelie Lander for superb assistance and equally superb chats. Besides my research activities, I also conducted educational tasks at Erasmus University and Rabobank International. I especially enjoyed (co-)supervising Masters students Victor Bellido, Georges Beukering, Maaike Duijts, Rob Groot-Zwaaftink, Wing-Hei Chan, Nathalie van der Mheen, Chios Slijkhuis and Michel van der Spek. For the werkcollege bedrijfseconometrie I co-supervised Arjan er staat toch geen punt achter?! van Dijk, Martijn nu zijn alle bugs eruit Krijger and Joost vlookup Kromhout. I hope you learnt as much from me, as I learnt from you.
vii
I am grateful to Kees van den Berg, Albert Mentink and Martijn van der Voort for reading one or more chapters of a preliminary version of this thesis. Your comments really helped me improve the readability of the text. I also thank Dennis Fok, Richard Kleijn
A and Erjen van Nierop for getting me started with L TEX.
Last, but never least, I thank my parents for believing in me, stimulating me to achieve the best and for always being there for me. I love you. Patrick Houweling Rotterdam, May 2003
Contents
Acknowledgements List of Figures List of Tables 1 Introduction 1.1 1.2 1.3 1.4 1.5 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sources of Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Credit Instruments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
v xi xiii 1 1 2 5
Credit Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 17
2 Estimating Spread Curves 2.1 2.2 2.3 2.4 2.5 2.6 2.7
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 Multi-Curve Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Model Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Model Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.A B -Splines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 2.B Variances and Covariances . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3 Measuring Corporate Bond Liquidity 3.1 3.2 45
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Contents
3.3 3.4 3.5 3.6
Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 75
4 Pricing Credit Default Swaps 4.1 4.2 4.3 4.4 4.5 4.6 4.7
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 Default Swaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 111
5 Pricing Step-Up Bonds 5.1 5.2 5.3 5.4 5.5 5.6
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Step-Up Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 137 141 147 151
6 Summary Nederlandse samenvatting (Summary in Dutch) Author Index Bibliography
List of Figures
Chapter 2 2.1 2.2 2.3 2.4 Distribution of the bonds maturity dates by rating category. . . . . . . . . 29 Single-curve estimates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 Single-curve and multi-curve estimates. . . . . . . . . . . . . . . . . . . . . 35 Spread curves. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Chapter 3 3.1 Liquidity premiums for dierent age thresholds. . . . . . . . . . . . . . . . 59
Chapter 4 4.1 4.2 Sensitivity of spreads and default swap premiums to the recovery rate. . . 85 Scatter plots of pricing errors versus default swap premiums per rating. . . 97
Chapter 5 5.1 5.2 5.3 5.4 5.5 Credit ratings history. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 Deutsche Telecom step-up premiums. . . . . . . . . . . . . . . . . . . . . . 127 France Telecom step-up premiums. . . . . . . . . . . . . . . . . . . . . . . 129 KPN step-up premiums. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 Recovery rate sensitivity analysis. . . . . . . . . . . . . . . . . . . . . . . . 132
List of Tables
Chapter 2 2.1 2.2 2.3 2.4 2.5 Distribution of bonds in the data set by rating. . . . . . . . . . . . . . . . 29 Distribution of included bonds in the data set by rating and industry. . . 30 Model specications for single-curve and multi-curve models. . . . . . . . . 31 Summary statistics of single-curve and multi-curve estimates. . . . . . . . . 33 Curve Similarity Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Chapter 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 Overview of liquidity measures from the empirical bond liquidity literature. 56 Overview of liquidity measures, their expected signs and the portfolio order. 62 Results for the entire sample . . . . . . . . . . . . . . . . . . . . . . . . . . 65 Results for the characteristics portfolios. . . . . . . . . . . . . . . . . . . . 67 Portfolio statistics P = 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 Results for model 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 Portfolio statistics P = 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 Results for model 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 Results of the comparison tests. . . . . . . . . . . . . . . . . . . . . . . . . 73
Chapter 4 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 Characteristics of the default swap data set. . . . . . . . . . . . . . . . . . 91 Performance of the direct comparison methods. . . . . . . . . . . . . . . . 96 Paired Z-tests of the direct comparison methods. . . . . . . . . . . . . . . . 98 Goodness of t of the reduced form credit risk models. . . . . . . . . . . . 100 Parameter estimates for the reduced form credit risk models. . . . . . . . . 101 Performance of the reduced form credit risk models. . . . . . . . . . . . . . 104 Paired Z-tests of the reduced form credit risk models. . . . . . . . . . . . . 105 Analysis of absolute pricing errors from reduced form credit risk models. . 108
xiv
List of Tables
Chapter 5 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 Step-up bond types. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 Number of bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 Characteristics of the step-up bonds. . . . . . . . . . . . . . . . . . . . . . 121 Pricing errors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 Paired Z-tests. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 Condence interval coverage percentages. . . . . . . . . . . . . . . . . . . . 125 Volatility analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 Event analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Chapter 1
Introduction
1.1
Motivation
The analysis of credit markets can be traced back to at least Fisher (1959) and the valuation of credit-risky securities made a signicant step forward due to Black and Scholes (1973) and Merton (1974). Nevertheless, this area of research has received relatively little attention, until several developments in the last decade awoke academics and practitioners, and brought about a wave of research on credit markets. The most important developments1 were the following: Companies increasingly raised capital directly from the capital markets by issuing bonds rather than borrowing money from their bank, especially in the United States (US) and Europe. Since loans are privately held and bonds are publicly traded, this changing behavior led to more publicly available data on credit markets. The European Monetary Union (EMU) and the liberalization of the European capital markets eectively integrated the markets of the participating countries into a single European corporate bond market. Liquidity, transparency and competition were greatly improved. New derivatives were developed to take on and lay o credit risk in a exible way. The market for these credit derivatives has grown tremendously over the last decade, both in size and product range. This necessitated the development of new models to price and hedge these new instruments. The prospering economic conditions and the reduction of governments budget decits in the US and EMU in the second half of the 1990s, drove yields and issuance
1
This list draws on Sch onbucher (1999).
Introduction Chapter 1
of government bonds to historically low levels. Investors thus needed other securities to enhance their portfolio yields. Bonds issued by corporations and emerging markets were found as alternatives. Later on, credit derivatives were used to reshape portfolios risk proles. Several well-published derivatives losses (e.g. Barings, MetallGesellschaft and Orange County), nancial turmoil (in Argentina, Asia and Russia), the near-collapse of hedge funds (most notably Long Term Capital Management, LTCM), and the actual defaults of several large companies (e.g. Enron, KPN Qwest and Worldcom) all contributed to a growing awareness among investors and regulators of credit and liquidity risk. These events and trends led to a large growth of credit markets, and at the same time to a growing need to understand them. Numerous models were developed, some extending the classic models of the 1970s, some drawing on the default-free interest rate literature. Especially the valuation of credit derivatives required the development of more sophisticated credit risk models. Empirical studies on credit markets were for a long time hampered by a lack of market data, and often restricted to US corporate bonds. Only since the last few years, research started to appear that analyzed not only US but also European and emerging markets, and not only bonds but also credit derivatives. This thesis adds four studies to the empirical literature on credit markets. The studies deal with credit risk, liquidity risk and credit derivatives. Before their contribution is discussed in Section 1.5, rst Section 1.2 lists the sources of risk to which an investor in credit markets is exposed, Section 1.3 gives an introduction to credit instruments and Section 1.4 describes the main approaches to credit risk modelling.
1.2
Sources of Risk
An investor who holds credit-risky securities is exposed to a number of risks, most importantly market risk, credit risk and liquidity risk. By subtracting the yields of two instruments with equal amounts of market risk, for example a defaultable bond and a default-free, liquid (but otherwise similar) bond, we obtain the (yield) spread. The spread compensates investors for being exposed to credit, liquidity and other risks. The remainder of this section discusses the main sources of risk. Most attention is paid to credit and liquidity risk, since they are the focus of this thesis.
Section 1.2 Sources of Risk
1.2.1
Market Risk
Market risk is the risk of losses resulting from adverse movements in the level or volatility of market prices. In credit markets, changes in interest rates are the most important market factor, although specic instruments may also be sensitive to uctuations in equity prices or exchange rates. To measure the market risk of portfolios, most nancial institutions use the concept of Value-at-Risk (VaR): the potential loss that is associated with a price movement of a given probability over a specied time horizon. For instance, a portfolio with a 1-day VaR of suer a loss in excess of
E10 million in one out of 20 days. Under the internal models
E10 million at a 95% condence level is expected to
approach of the Basel Committee on Banking Supervision (BCBS, 1996) of the Bank for International Settlements (BIS), the amount of regulatory capital banks have to put aside to cover market risks is based on their VaR level.
1.2.2
Credit Risk
Credit risk concerns the losses caused by the possibility that an entity will fail to fully and timely meet its contractual obligations. With traditional debt instruments, such as bonds and loans, the borrower is obliged to pay the coupons and the notional amount in time. With derivatives, such as swaps and options, the amounts due depend on the prevailing market conditions, as specied by the contract. Credit risk can be separated into two components: default risk is the uncertainty about whether or not the entity will fail to meet its obligations; recovery risk is the uncertainty about the amount that will be recovered in case of default. Together they determine the compensation investors receive for bearing credit risk: the credit spread.2 The spread on a borrowers securities can be seen as the markets assessment of its credit risk. Another important indicator of a borrowers credit worthiness is its credit rating : a subjective assessment of its credit or default risk, measured on an alphanumeric scale. The scales of the two major independent rating agencies, Moodys and Standard & Poors (S&P), are {Aaa, Aa, A, Baa, Ba, B, Caa, Ca, C} and {AAA, AA, A, BBB, BB, B, CCC, CC, C}, respectively. Aaa/AAA indicates the highest credit quality, C the lowest. The rating agencies rene these major ratings into minor ratings by adding notches to the letters: Moodys uses postxes 1, 2 and 3 and S&P adds a + sign, no postx, or a sign. The rst four ratings are collectively called investment grade and the remaining ratings speculative grade. Many nancial institutions assign internal credit ratings to their counterparties as well. Sometimes rating migration risk is seen as a separate component
2
Changes in credit spreads are sometimes considered part of market risk instead of credit risk.
of credit risk. Also, some credit derivatives explicitly depend on the credit rating of the underlying issuer. While market risk can typically be measured and hedged on a day-to-day basis, credit risk accumulates over longer-term time horizons, e.g. one year. Moreover, market prices of many credit-risky securities are not daily updated, and defaults (and to a lesser extent rating migrations) occur only infrequently. The data scarcity and the longer time horizon make estimating and backtesting of credit risk models much more dicult than of market risk models. Regulation of credit risk has long followed the Basel Capital Accord (BCBS, 1988), requiring banks to hold regulatory capital equal to at least 8% of a risk-weighted basket of assets to cope with potential losses from these assets. In 1999 and 2001, the BCBS issued consultative documents on the New Basel Capital Accord, proposing more risk-sensitive weights based on external or internal credit ratings (BCBS, 1999, 2001). The new accord is planned to replace the current accord in 2006.
1.2.3
Liquidity Risk
Liquidity risk, also called marketability risk, involves the possibility of not being able to timely buy or sell an instrument in the desired quantity with little impact on prices. Liquidity may dier between instruments (e.g. swaps are generally more liquid than bonds), between instrument types (e.g. plain vanilla instruments are more liquid than exotic instruments), between issuers (e.g. government bonds are typically more liquid than corporate bonds) and between markets (e.g. the euro capital market is more liquid than emerging markets). Liquidity risk tends to aggregate other sources of risk. For example, markets tend to lose liquidity in times of crises and/or high volatility, e.g. the 1987 stock market crash or the nancial turmoil in Asia and Russia in 1998. Also, lower rated bonds are often less liquid than otherwise similar bonds with higher ratings, since many large investors, like pension funds, are not allowed to hold speculative grade securities. At the moment, there are no regulatory guidelines that cover liquidity risk, but the BIS Committee on the Global Financial System has started publishing on the topic; see CGFS (1999, 2001).
1.2.4
Other Risks
This section discusses several other sources of risk, though without being exhaustive. Although institutions and regulators put the most eort in managing and measuring market, credit and liquidity risk, many of the major losses of the 1990s mentioned above (including the collapses of Barings and Enron) were due to operational risk : the risk
Section 1.3 Credit Instruments
of unexpected losses arising from deciencies in a rms management information, procedures, and control systems. Specically, for banks a mistake or fraud in the trading or risk management department can have more harmful eects than a market crash. Operational risk will be covered in the New Basel Capital Accord (BCBS, 2001). Another source of risk that is not limited to nancial institutions is legal risk, which is the risk that a transaction proves unenforceable in law or because it has been inadequately documented. For example, in the early days of credit derivatives, legal risk was the largest concern to participants of the biannual Credit Derivatives Survey by the British Bankers Association (BBA, 1998), because documentation was not yet standardized and dierent counterparties used dierent denitions and legal structures. Documentation disputes led to several lawsuits, for instance after Russias default on its sovereign debt in 1998 and after the restructuring of Consecos debt in 2000. Systemic risk refers to the possibility of disruptions in the functioning of nancial markets that are severe enough to reduce economic activity. Such a systemic event is typically hypothesized to occur after the initial bankruptcy of one large institution, followed by a domino-style contagion that causes the bankruptcy of many more and, as a worst-case outcome, the collapse of the nancial system as a whole. As mentioned above, the studies comprising this thesis focus on credit and liquidity risk. While the studies take into account market risk, the other sources of risk are ignored; not because they are less important, but because they aect all instruments alike and are much harder to quantify.
1.3
Credit Instruments
The empirical studies in this thesis use market data on two types of credit instruments: bonds and credit derivatives. The remainder of this section discusses bonds only briey and credit derivatives in more detail, because the latter are more recently introduced and less well-known.
1.3.1
Bonds
A bond is an obligation on its issuer to its holder with the purpose of raising capital by borrowing. Typically, the holder pays the amount borrowed (notional amount ) to the issuer at issuance and the issuer promises to repay this amount in the future, along with interest payments (coupons ). The frequency at which interest payments are made varies across markets, but is usually annually or semi-annually. Various coupon types
exist: a xed-income bond pays a xed percentage of its notional; the coupon percentage of a oating-rate bond is reset periodically to a specied short-term interest rate plus a specied spread; for a (rating-triggered ) step-up coupon bond, the coupon percentage depends on the credit rating of its issuer; a zero-coupon bond does not make interest payments. More bond avors can be created by varying the redemption method: if the principal is repaid as a whole on the maturity date, the bond is called a bullet bond; with a sinking bond the issuer repays the face value in several terms on a set of pre-specied dates. Uncertainty regarding the bonds maturity is introduced when option-like features are added: a callable bond gives the issuer the right to redeem the principal prematurely; a puttable bond gives the holder the right to sell the bond back to the issuer early; an extendible bond enables the issuer to extend the life of the bond beyond the initially agreed redemption date; nally, with a perpetual bond the issuer never repays the principal. Most chapters in this thesis analyze xed-income or zero-coupon bullet bonds without optionalities (usually called plain vanilla bonds), with the exception of Chapter 5, which studies step-up bonds.
1.3.2
Credit Derivatives
A credit derivative allows the transfer of credit risk3 without transferring the ownership of debt issued by the underlying borrower(s). The pay-out of a credit derivative can depend on: (i) the occurrence of a credit event, e.g. bankruptcy, failure to make an interest payment, debt restructuring or debt acceleration (jointly called default ), or a rating migration; (ii) the payments, price, and/or yield spread of one or more bonds of the underlying borrower(s). Many credit derivatives are insurance-like contracts between two parties, where one party buys protection from the other party to a deterioration of the borrowers credit worthiness. The protection buyer will either make periodic payments or pay an up-front premium to the seller, and the protection seller will, upon the occurrence of the specied credit event, make a payment to the buyer. Hence, the buyer has reduced his credit exposure to the underlying entity in return for a periodic fee. Several examples of credit derivatives are given below. Credit derivatives were rst introduced on the annual meeting of the International Swaps and Derivatives Association (ISDA) of 1992 and some trading began in 1992 as well. Academic papers started to appear around 1995; see e.g. Howard (1995), Smithson (1995) and Das (1996). Several books were published on the subject a few years later; see e.g. Das (1998), Tavakoli (1998) and Francis, Frost and Whittaker (1999). Since 1992,
3
Some credit derivatives, like total return swaps, also transfer market risk.
the global credit derivatives market has experienced impressive growth. Whereas market size (measured in total outstanding notional) amounted to no more than a few billion US dollars in 1995, participants to the latest Credit Derivatives Survey (BBA, 2002) estimated that the market has grown to US$ 2.0 trillion at the end of 2002; participants to the annual Credit Derivatives Survey by Risk Magazine (Patel, 2003) estimated a market size of US$ 2.4 trillion.4 To put these gures into perspective, as of June 2002 the total outstanding notional of interest rate swaps amounted to US$ 68 trillion and of interest rate options to US$ 13 trillion (BIS, 2002, Table 19). Although the market for credit derivatives is still relatively small, it is catching up fast with annual growth rates of at least 50%. The publication of the Credit Derivatives Denitions (ISDA, 1999) was a big move to standardizing the terminology in credit derivatives transactions. The ISDA Denitions were amended in 2001 with the Restructuring Supplement (ISDA, 2001) following disagreements in the market on which obligations can be delivered in physically settled contracts in case of a debt restructuring; see also Tolk (2001). The Denitions established a uniform set of denitions of important terms, such as the range of credit events that could trigger payments or deliveries. In addition to the enhanced enforceability and interpretation of the contracts, the Denitions increased exibility and reduced the complexity of administration and documentation. More than 90% of all credit derivative transactions are being documented by the ISDA conrms (BBA, 2002, page 22). Types According to the BBA (2002), credit default swaps are the most popular type of credit derivative (accounting for 45% of the market), followed by collateralized debt obligations (22%), credit-linked notes (8%), total return swaps (7%), basket products (6%) and credit spread options (5%). These contracts are briey discussed below; see Tavakoli (1998) or OKane (2001) for details and other credit derivatives. Default swap: A (credit) default swap (CDS) is an insurance-like contract that protects the holder of the underlying asset(s) from the losses caused by the occurrence of a specied credit event to the reference entity. The protection buyer makes periodic payments to the protection seller, typically a specied percentage of the notional amount. If the credit event occurs, the protection seller reimburses the loss incurred by the protection buyer, so that the value of the buyers asset(s) is restored to the notional amount. Note that a default swap only pays out if the reference enBoth surveys are based on interviews and estimations, and should therefore be treated as indications rather than hard numbers.
4
tity defaults; reductions in value unaccompanied by default do not compensate the buyer in any way. Credit-linked notes: A credit-linked note is basically a combination of a bond and a default swap. At initiation of the contract, the protection buyer sells a bond to the protection seller and thus receives the notional of the bond. The protection buyer makes (xed or oating) coupon payments on the bond during the contract period. At the maturity date, the bond is to be redeemed at par, unless a credit event has occurred to the reference entity, in which case the buyer only pays the recovered amount. Although much like a default swap, there is an important dierence. With credit-linked notes the protection seller makes his payment in advance and receives it back fully (if no event occurs), or partially (if the event does occur). With default swaps, the protection seller only makes a payment after the event has occurred. Total return swap: In contrast to a default swap, which only transfers credit risk, a total return swap (TRS; also called a total rate of return swap, TRORS) transfers both credit risk and interest rate risk. The buyer makes periodic payments to the seller, usually specied as a spread over interbank interest rates (LIBOR). The seller pays to the buyer the total return of the asset, comprising of interest payments and change-in-values payments. The latter are dened as any appreciation or depreciation in the market value of the reference obligation. Hence, a net depreciation in value results in a payment to the seller. When entering into a total return swap on an asset, the buyer has eectively removed all economic exposure to the underlying asset. The seller on the other hand has gained exposure to the underlying without the initial outlay required to purchase the reference obligation. Credit spread option: A credit spread option is similar to a standard stock option, except that the underlying is a credit spread rather than a stock price. With a credit spread option, one party pays an up-front premium to the other in return for a payment at the maturity date that is linked to the dierence between the actual spread and the specied strike. Credit spread options thus allow users to bet on or hedge against future spread movements. Collateralized debt obligations: A collateralized debt obligation (CDO) is a structure of xed-income securities, called the tranches, whose cash ows are backed by the payments of an underlying pool of debt instruments, the collateral, through a set of rules, the waterfall structure. When the collateral is a pool of bonds, the structure is called a collateralized bond obligation (CBO); with a pool of loans, it is a collateralized loan obligations (CLO); with a pool of default swaps, it is a synthetic CDO. The tranches have dierent priorities: income from the collateral
is rst paid to the senior tranches, than to the mezzanine tranches and nally to the equity tranches. A CDO allows the redistribution of the credit risk of a pool of assets to create securities with a variety of risk proles. Basket products: An nth -to-default swap is similar to a regular default swap, but now the credit event that triggers the payment to the protection buyer is the nth default in a specied basket of borrowers. For instance, in a rst-to-default swap the rst borrower to default, triggers the contract. Likewise, a basket total return swap is just like a regular TRS, but instead of single underlying bond, the cash ows and price changes of a portfolio of bonds are passed through.
Applications Like any derivative, a credit derivative can be used to take on or lay o risk. Because a derivative does not transfer the ownership of the underlying assets and often does not require an initial investment, risk can be transferred more eciently than in the cash market: buyers can reduce credit exposure without physically removing assets from their balance sheets, and sellers get the opportunity to run credit risk without actually buying the reference asset. Specically, credit derivatives have the following applications (ranked in order of importance according to the BBA, 2002): Managing credit lines: Banks are limited in the amount of business they can do with a particular borrower. Yet even if a credit line is full, a bank may want to lend additional funds to a borrower to prevent a deterioration in its relationship with the client. Credit derivatives oer a solution to this dilemma: the bank can give the client a new loan and, without having to notify the client, simultaneously buy default protection on the client in the credit derivatives market. Now the bank has both fostered its relationship with the client and kept its credit risk within the specied limits. Regulatory arbitrage: In the 1988 Capital Accord, corporations have a risk weight of 100% in calculating the amount of regulatory capital, but banks from OECD countries only 20%. Consequently, by using credit derivatives banks can transfer credit risk on corporate loans to an OECD bank to reduce regulatory capital. These risk transfers are protable to banks, because the freed capital can be put to other uses again. Banks regulators are dissatised with such transactions, because regulatory capital is not in line with actual risk taking behavior. The New Capital Accord addresses this issue by proposing rating-based risk weights.
10
Product structuring: Credit derivatives can be used to split the credit risk of one or more assets and redistribute it into more risky or less risky forms that suit the risk appetites of dierent investors. This also allows the creation of tailor-made investment products. Portfolio management: Credit derivatives allow investors to change their portfolio characteristics by reducing or taking on exposure to particular companies, sectors, regions and/or maturities, which would be much harder or even impossible just using bonds and loans. For example, the lack of a market for repurchase agreements (repos ) for most corporates makes shorting corporate bonds infeasible. So, credit derivatives are the only viable way to go short corporate credit risk.5
1.4
Credit Models
In the literature, there are two classes of credit risk models: portfolio models and pricing models. Portfolio models are primarily used for risk measurement and management purposes, like calculating VaR-like risk measures and marginal risk contributions. In this thesis, only instances of the class of pricing models are used, so we restrict ourselves to mentioning some examples of the class of portfolio models: CreditMetrics (J.P. Morgan: Gupton, Finger and Bhatia, 1997), CreditPortfolioView (McKinsey: Wilson, 1997a,b), CreditRisk+ (Credit Suisse Financial Products, 1997) and Portfolio Manager (Moodys KMV: Kealhofer, 2001); see Crouhy, Galai and Mark (2000) for a comparison of this class of models or the books by Caouette, Altman and Narayanan (1998) and Saunders (1999) for in-depth treatments. Pricing models are mainly used for investment-related purposes, including pricing of bonds and credit derivatives, calculating hedge ratios, and seeking favorable investment opportunities. There are two types of pricing models: structural form models and reduced form models. The distinction between these two types of models is blurring though, since hybrid models have also started to appear (e.g. Madan and Unal, 2000). In fact, Due and Lando (2001) showed that under asymmetric information, reduced form models can be seen as the reduced form of a particular type of structural model. The remainder of this section outlines structural and reduced form models; see also Nandi (1998), Jeanblanc and Rutkowski (1999) and OKane and Schl ogl (2001).
Even if a bond can be shorted on repo, investors can only do so for short periods of time (one day to one year), exposing them to changes in the repo rate, next to changes in credit spreads. On the other hand, credit derivatives allow investors to go short credit risk at a known cost for long time spans.
5
Section 1.4 Credit Models
11
1.4.1
Structural Models
In the class of structural models, also called rm value models or Merton models, a rm defaults when the value of the rms assets drops below a certain threshold. If this happens, bond holders get the residual value of the rm and share holders receive nothing. If the rm survives, bond holders are paid o, and share holders receive the remaining value of the assets. In this framework, both bonds and stocks are contingent claims on the value of the rms underlying assets, so that option pricing theory can be used to calculate theoretical debt and equity values. The advantage of structural models is that they describe how default actually occurs and that recovery is determined endogenously. However, structural models have diculties incorporating complex debt structures. Moreover, their parameters are hard to estimate, because the assets market value and volatility are dicult to observe. Finally, structural models may be better suited for corporate than for sovereign issuers, because for countries the asset value concept is not applicable, and even though a country is able to pay, it may not be willing to do so. Structural models were rst developed by Black and Scholes (1973) and Merton (1974). In their model, a rm has a simple debt structure consisting of one zero-coupon bond. Default can only occur at the bonds maturity date. The classic model has been extended by several authors. Black and Cox (1976) introduced the possibility of intermediate default into the model, as well as indenture clauses, safety covenants and subordination arrangements. Geske (1977) and Geske and Johnson (1984) allowed a more general capital structure and considered coupon-bearing bonds instead of zero-coupon bonds. Shimko, Tejima and van Deventer (1993) and Longsta and Schwartz (1995a) relaxed the assumption of deterministic interest rates and used a Vasicek (1977) model to describe the evolution of the default-free term structure. Leland (1994), Leland and Toft (1996) and Mella-Barral and Perraudin (1997) combined the model with strategic behavior models from corporate nance. Sch onbucher (1996) and Zhou (1997) generalized the continuous asset value process to a jump-diusion process, so that defaults can also come as a surprise. Moodys KMV commercially applies the structural approach (Crosbie, 2002). Fitting a Merton-type model to balance sheet and equity price data, KMV calculates an Expected Default Frequency measure, which is a rms probability of default for the next year.
1.4.2
Reduced Form Models
In reduced form models, also called intensity-based models, the direct reference to the rms asset value process is abandoned. Instead, default is modelled as an exogenous
12
event. In particular, default is linked to a counting process that literally counts the number of defaults. Typically, we are only interested in the rst default, so the default time is dened as the time of the rst jump of the counting process. A popular example is the Poisson process, whose stochastic behavior is driven by a hazard process, also called an intensity process. The hazard rate is the arrival rate of the default event and can be interpreted as a conditional, instantaneous default probability. Theoretical prices for bonds and derivatives are computed using equivalent martingale measures; see Bielecki and Rutkowski (2001) for a detailed account of the mathematics of reduced form models. The advantage of reduced form models is that their parameters are easy to estimate. Also, they can be calibrated to the market prices of liquid instruments and subsequently used for the pricing of credit derivatives; this is very similar to the calibration and pricing of default-free interest rate derivatives. Their drawback is that it is dicult to realistically model the recovery process (see below). The rst reduced form model was developed by Litterman and Iben (1991), who considered a simple setup without recovery in case of default. Jarrow and Turnbull (1995) formalized their model using risk-neutral valuation and assumed a xed recovery rate at maturity and a Poisson process with a xed hazard rate. Lando (1998) further generalized the framework by making the hazard rate stochastic; this is called a Cox process or doubly-stochastic Poisson process. Typically, a factor model is employed to drive both the default-free term structure and the hazard rate, so that default-free rates and the default time are correlated. With Cox processes, hazard rates can also be made dependent on equity prices, hence bringing in new information; see Jarrow and Turnbull (2000), Jarrow (2001) and Pan (2001). In a somewhat dierent approach, due to Due and Singleton (1999), there is no need to separately model the hazard and recovery components of credit risk, but it suces to model the spread process. Other implementations of this approach include Longsta and Schwartz (1995b) and Das and Sundaram (2000). To use the information present in credit ratings and to value securities that explicitly depend on ratings, Lando (1994) and Jarrow, Lando and Turnbull (1997, JLT) developed a rating-based reduced form model. They used a Markov chain with ratings as states. The model was generalized by Das and Tufano (1996) to incorporate stochastic recovery rates, and by Lando (1998) and Arvanitis, Gregory and Laurent (1999) to make transition intensities stochastic and possibly dependent on state variables. Sch onbucher (1999), Bielecki and Rutkowski (2000) and Acharya, Das and Sundaram (2002) embedded the Markov chain in a Heath, Jarrow and Morton (1992) framework. A dicult issue in reduced form models is the recovery assumption. Whereas in structural models the amount recovered by bond holders in case of default is determined
Section 1.5 Overview
13
endogenously, reduced form models have to specify the recovery process explicitly. Three recovery assumptions are found in the literature: Recovery of Treasury: This approach, rst used by Jarrow and Turnbull (1995), assumes that at the default time, the defaulting bond is replaced by a default-free, but otherwise similar, bond. The main advantage of this approach is that it leads to a closed-form solution of defaultable bonds, so that risk-neutral default probabilities can be easily backed out from default-free and defaultable term structures. Recovery of market value: Under this assumption, introduced by Due and Singleton (1999), a bond loses a constant fraction of its market value. Due and Singleton (1999) showed that this assumption allows credit-risky claims to be valued as if they were default-free, but now discounted by risk-adjusted interest rates. Recovery of face value: This assumption, applied by Jarrow and Turnbull (2000) and Sch onbucher (2000), conforms best to real-world defaults, where investors recover a fraction of the bonds face value (and sometimes accrued interest as well). It also corresponds to the way Moodys and S&P publish their recovery rate estimates. For the management of credit risk on a portfolio basis, and for the pricing of basket credit derivatives and CDOs, models are required that describe the joint credit worthiness of multiple issuers. This has been accomplished in the literature in several ways. Due (1998) and Due and Singleton (1998) introduced correlations between the hazard processes of issuers by making them dependent on common factors. Gupton et al. (1997) and Hull and White (2001) extended the Merton (1974) approach to multiple issuers by correlating the underlying asset processes. Finally, Li (1999, 2000) and Sch onbucher and Schubert (2001) used copula functions to model the dependency structure between the marginal default densities. Since copula functions allow the univariate behavior to be separated from the dependency structure, any correlation structure can be imposed, for example equity correlations or spread correlations.
1.5
Overview
This thesis contributes four studies to the empirical literature on credit markets. Chapters 2 and 3 are concerned with the measurement of yield spreads and corporate bond liquidity, respectively. Chapters 4 and 5 contain two empirical studies on the pricing of credit derivatives. The remainder of this section outlines the four chapters. Chapter 2 presents a robust framework for the estimation of yield spreads. Spreads are an important input for the pricing of defaultable bonds and credit derivatives and for
14
risk management purposes. Inaccuracies or errors in the estimated spreads will result in incorrect prices or risk measures. Traditionally, spread curves are calculated by subtracting independently estimated default-free and defaultable term structures of interest rates. It is illustrated that this results in twisting spread curves that alternately have positively and negatively sloped segments. In Chapter 2, a new framework is presented for the joint estimation of the default-free term structure and corporate spread curves. The model is based on the decomposition of a defaultable term structure into a default-free part and a spread part. The default-free curve is estimated from government bonds, so that the model for the corporate term structure can focus on the spread curve only and can thus be parsimonious. The performance of the new model is compared to the traditional method by estimating them on a data set of German mark-denominated government and corporate bonds. Chapter 3 is concerned with the estimation of liquidity spreads of corporate bonds. For an investor, it is important to know whether a bond is liquid or illiquid, because if he needs to sell an illiquid bond before its maturity, he faces higher transaction costs, due to a larger bid-ask spread and/or order processing costs, than for a comparable, liquid bond. Direct liquidity measures, such as trading volume or trading frequency, are not available for corporate bonds, since most transactions occur on the over-the-counter market. Therefore, the literature has proposed numerous indirect liquidity measures that are based on bond characteristics and/or market prices. In Chapter 3, an empirical comparison is conducted of eight indirect measures of corporate bond liquidity, one of which is new to the literature. Great care is taken to ensure that bond yields are corrected for market and credit risk to properly identify the spread associated with liquidity risk. For each liquidity measure, the signicance of the liquidity eect is determined on a data set of euro-denominated bonds. Moreover, a series of pairwise comparison tests is conducted to establish the eectiveness of the liquidity measures relative to each other. Chapter 4 contains an empirical study on the pricing of default swaps. Since default swaps are the most popular credit derivative, it is important to know how market participants price them. Moreover, default swap data provide an interesting challenge for the credit risk models that have been developed for the pricing of credit derivatives. In Chapter 4, a reduced form model is implemented with a deterministic hazard process and a constant recovery of face value assumption. The model is estimated on market prices of bonds and subsequently used to calculate theoretical default swap premiums. For comparison, a simple spread-based approach is also implemented, which directly compares bonds yield spreads to default swap premiums. The chapter pays attention to the implementation of the approaches, by considering several alternatives for the choice of the
Section 1.5 Overview
15
default-free term structure and by testing the robustness with respect to the assumed recovery rate. Chapter 5 provides an empirical analysis of the pricing of rating-triggered step-up bonds. Step-up bonds are basically xed-income bonds with a built-in credit derivative, whose payo depends on the issuers credit rating. These bonds formed an important source of nancing for European telecom companies in recent years when they had diculty issuing plain vanilla bonds due to nancial distress. Further, step-up bonds allow empirical testing of rating-based credit risk models. In Chapter 5, three methods are compared to value step-up bonds: (i) the Jarrow, Lando and Turnbull (1997) framework, (ii) a similar framework using historical probabilities and (iii) as plain vanilla bonds. It is tested which method provides the best approximation to market prices, and whether step-up bonds oer protection to investors in the form of superior returns or lower price volatility. Chapter 6 concludes the thesis.
Chapter 2
Estimating Spread Curves1
2.1
Introduction
Many credit risk models require an accurate description of the term structures of interest rates of dierent credit risk classes as input data. Measuring a term structure for a particular credit rating class amounts to estimating its credit spread curve relative to the government curve, which proxies the default-free curve. Traditionally, spread curves are calculated by subtracting independently estimated government and corporate term structures. In this chapter, we present a new framework that jointly estimates the government curve and credit spread curves. Unlike the twisting curves one gets from the traditional method, the estimated spread curves are now smooth functions of time to maturity, and are less sensitive to model settings. Moreover, we develop a novel test statistic that allows us to determine the optimal settings of the new model. An important application in which accurately estimated term structures of interest rates are essential inputs is the pricing of defaultable bonds and credit derivatives. The leading frameworks are the Jarrow, Lando and Turnbull (1997) Markov chain model, which extended the work of Litterman and Iben (1991) and Jarrow and Turnbull (1995) to multiple credit ratings, and the Due and Singleton (1999) framework, which can be cast into a defaultable Heath, Jarrow and Morton (1992, HJM) model. Similar to the default-free interest rate models developed in the early 1990s most notably the extended Vasicek (1977) models, such as Hull and White (1990), the lognormal short rate models, like Black, Derman and Toy (1990), and the models in the HJM framework these credit risk models provide an exact t to todays default-free and defaultable term structures
This chapter is a slightly revised version of the article by Houweling, Hoek and Kleibergen (2001), which has been published in the Journal of Empirical Finance.
1
18
Estimating Spread Curves Chapter 2
of interest rates. Any error in the input of such models will be amplied in the prices of interest rate and credit derivatives that are subsequently priced with them. Interest rates and spread curves are also required for risk management purposes, for example in applying the historic simulation method to calculate the Value at Risk for a corporate bond portfolio; see e.g. Saunders (1999, Chapter 11). Future scenarios are generated by adding historical day-to-day movements in interest rates and spread curves to todays curves. Since in each scenario the bond portfolio is revalued to obtain the empirical distribution of the future portfolio value, inaccurate curves may lead to an unnecessarily large Value at Risk and a too large amount of regulatory capital. Other applications of default-free and defaultable interest rates include the pricing of new bond issues and assessing counterparty risk in derivative products; see e.g. Hull and White (1995), Duee (1996) and Caouette et al. (1998). An obstacle in the above mentioned applications is that the term structures are not directly observable in the market and have to be estimated from market prices of traded instruments. Until now the literature has primarily focused on the estimation of the default-free term structure from a data set of government bonds. The standard approach originates from McCulloch (1971, 1975), who modelled the discount curve as a linear combination of polynomial basis functions. Other approaches include the usage of Bernstein polynomials (Schaefer, 1981), exponential splines (Vasicek and Fong, 1982), B-splines (Shea, 1985; Steeley, 1991), exponential forms (Nelson and Siegel, 1987) or a bootstrapping procedure as employed on electronic information systems Bloomberg and Reuters; Anderson, Breedon, Deacon, Derry and Murphy (1996, Chapter 2) provided an extensive overview of these and other term structure estimation methods. After choosing one of these methods, we could independently estimate a separate model for each credit class. We illustrate that these calculations are likely to result in twisting spread curves that alternately have positively and negatively sloped segments. Moreover, the level and shape of the spread curve are shown to be sensitive to model misspecication. Instead, we jointly estimate the default-free and defaultable interest rate curves. Our joint estimation is based on the decomposition of a defaultable term structure into a default-free curve and a credit spread curve. The default-free curve is estimated from government bonds, so that our model for a corporate term structure focuses on the credit spread only. Both the government curve and the corporate spread curve are modelled as B-spline functions and all parameters are jointly estimated from a combined data set of bonds. We apply the model to a data set of liquid, German mark-denominated bonds, whose credit ratings range from Standard and Poors ratings AAA to B. We obtain smooth and reliably estimated spread curves that are relatively robust to model
Section 2.2 Multi-Curve Model
19
misspecication. Moreover, we demonstrate that these results can be attributed to both the joint and the parsimonious modelling. Independently estimating the government curve and a parsimoniously specied corporate curve model does not yield the same results, nor does jointly estimating the government curve and a richly specied corporate spread curve. The remainder of this chapter is structured as follows. Section 2.2 presents the new framework for the joint estimation of the government term structure and corporate credit spread curves. The specication of the model is described in Section 2.3, whereas Section 2.4 goes over several methods to choose between competing models, including a novel statistic that is developed to compare spread curves obtained from alternative model specications. Section 2.5 describes our data set. Section 2.6 applies the new model and compares jointly estimated term structures with independently estimated term structures. Finally, Section 2.7 summarizes the chapter.
2.2
Multi-Curve Model
Ideally, we would like to use a dierent spread curve for each rm, reecting the uniqueness of a rms characteristics that determine its credit risk. Due to data constraints, however, which are discussed in Section 2.5, we have to resort to grouping rms that have similar credit worthiness and face similar operating environments. A disadvantage of grouping bonds is that a particular type of heterogeneity2 may occur; see Helwege and Turner (1999). Suppose we have created C categories of bonds, where category 1 corresponds to government bonds and the other categories are formed by using, e.g., credit rating and industry as criteria. The purpose is to estimate a spread curve for each category. Instead of independently estimating term structures, we propose a joint estimation approach. Since a corporate term structure consists of a default-free curve and a credit spread curve, it seems natural to only model the spread and take the default-free part from the government curve. Several representations of the term structure exist, e.g. as discount factors or spot interest rates, but it is common practice to model the discount curve. We use the following framework for jointly estimating the discount curves D1 (t) = d(t) Dc (t) = d(t) + sc (t), c = 2, 3, . . . , C,
2
(2.1)
Within a data set of bonds of the same rating, the longest maturity bonds usually have been issued by the relatively most credit worthy rms. Therefore, credit spreads may decrease for the longest maturities in such a data set.
20
where Dc () is the discount curve of category c, d() is the model for the government discount curve and sc () is the model for the discount spread curve of category c with respect to the government curve. We impose C constraints Dc (0) = 1, because a payment due today does not need to be discounted. All parameters in the models for the government curve and the discount spread curves are jointly estimated from a combined data set of government and corporate bonds. We refer to this model as the multi-curve model as opposed to a single-curve model that independently estimates a single term structure. To model d() and sc (), we use spline functions, as introduced to the term structure estimation literature by McCulloch (1971). Some commonly used types of splines are exponential splines by Vasicek and Fong (1982), and B-splines, as discussed by Shea (1985) and Steeley (1991). Splines are tailored to approximate a scatter of data points by a continuous and preferably smooth function. Their main advantage is their exibility: there is no need to a priori impose a particular curvature, because the shape of the curve is determined by the data. Bliss (1997) compared several non-parametric term structure estimation models and found that spline models perform at least as good as competing models, and outperform the other considered models if the data contains longer maturity bonds (over 5 years). Jankowitsch and Pichler (2002) estimated our framework on a data set of nine EMU governments using both cubic splines and Svensson (1994) curves. They conrmed our results, and moreover, found that the multi-curve splines model and the multi-curve Svensson model performed similarly. Jankowitsch and Pichler (2003) extended our framework to multiple currencies. Splines are basically piecewise polynomials. The approximation interval3 [a, b] is divided into n subintervals [0 , 1 ], [1 , 2 ], . . . , [n1 , n ], where the knots i are chosen such that a = 0 < 1 < . . . < n = b. The data points in each subinterval are modelled as a k th degree polynomial. The n polynomials are constrained by the condition that the spline has to be k 1 times continuously dierentiable. This condition imposes k constraints on the coecients of two adjacent polynomials at the knots 1 , 2 , . . . , n1 . In sum, we have n(k + 1) coecients minus (n 1)k constraints, leaving n + k degrees of freedom. A more parsimonious way of representing splines is by means of basis functions; see e.g. Powell (1981, page 228). Any k th degree spline function S () with knots = (0 , . . . , n ) can be expressed as a linear combination of n + k basis functions f () = {f1 (), f2 (), . . . , fn+k ()}
n+k
S (t) =
s=1
3
s fs (t) = f (t) .
With term structure estimation the approximation interval runs from 0 to the longest bond maturity in the sample.
Section 2.2 Multi-Curve Model
21
Once the basis is chosen and the degree k and the knots are set, the basis functions are fully specied. The spline weights , however, are unknown and have to be estimated from the data. Powell (1981) recommended the use of a basis of B-spline functions, because of their eciency and numerical stability. Steeley (1991) applied B-splines to term structure estimation. Appendix 2.A contains a concise description of constructing a basis of B-splines; see Powell (1981) for more details. We use B-splines to model the government discount curve and corporate discount spread curves in Equation (2.1). We set d(t) = g1 (t) 1 and sc (t) = gc (t) c , where gi ()
th contains ni + ki B-spline basis functions that span a spline of ki degree with knots i .
Section 2.3 discusses the specication of the degrees and knots. Using B-spline basis functions, we rewrite the multi-curve model in Equation (2.1) as D1 (t) = g1 (t) 1 Dc (t) = g1 (t) 1 + gc (t) c , c = 2, 3, . . . , C. (2.2a) (2.2b)
To estimate the unknown spline weights 1 , 2 , . . . , C , we construct a data set, consisting of Bc bonds of category c, and use the discounted cash ow (DCF) principle to link the bond prices to a discount curve. According to the DCF principle, the price that an investor is willing to pay for the bth bond of category c equals the sum of the present values of the cash ows
Ncb DCF Pcb
=
i=1
CFcbi Dc (tcbi ),
(2.3)
DCF where Pcb is the DCF bond price, Ncb is the number of remaining cash ows and
CFcbi is the ith cash ow that is paid at time tcbi . By using the DCF equation as the theoretical bond price model, we have to conne our data set to xed-income bonds with known redemptions and exclude any bonds with optional elements such as callable and puttable bonds and bonds with oating or index-linked coupons. The DCF method is valid if we assume a perfectly competitive capital market, i.e. if all relevant information is widely and freely available and no barriers, frictions and taxes exist. Brealey and Meyers (1991, page 20) stated that even though these conditions are not fully satised, there is considerable evidence that security prices behave almost as if they were.
22
For category 1, i.e. for government bonds, we substitute Equation (2.2a) into Equation (2.3), yielding
N1b DCF P1 b n1 +k1 n1 +k1 N1b
CF1bi
i=1 n1 +k1 s=1 s=1
1s g1s (t1bi )
=
s=1
1s
i=1
CF1bi g1s (t1bi ) (2.4a)
x1bs 1s = x1b 1 ,
Ncb i=1
where xcbs =
CFcbi g1s (tcbi ). For categories 2, 3, . . . , C substitution of Equation (2.2b)
into Equation (2.3) results in

n1 +k1 DCF Pcb Ncb nc +kc Ncb
=
s=1 n1 +k1
1s
i=1
CFcbi g1s (tcbi )

nc +kc
+
s=1
cs
i=1
CFcbi gcs (tcbi ) (2.4b)
s=1
xcbs 1s +
s=1 Ncb i=1
ycbs cs = xcb 1 + ycb c ,
where ycbs =
CFcbi gcs (tcbi ). Note that these equations for the theoretical bond price
are linear in the unknown parameters, because all terms in x and y are either known from the characteristics of the bond or the specication of the spline models. Also, the constraints Dc (0) = 1 on the discount functions are linear restrictions on the spline weights; see Equation (2.2).
DCF In order to estimate the spline weights, we substitute the theoretical prices Pcb by
observed market prices Pcb and add an error term cb to the equations. The error term is necessary, because due to market imperfections the DCF method is not able to perfectly explain bond prices.4 Using matrix notation, we obtain the following linear regression model
P2 X2 Y2 0 P3 = X3 0 Y 3 . . . . . . . . . . . . XC 0 0 PC
P1
X1
...
0 ... 0 ... . . . . . . YC ...
2 3 + . . . C
2 2 3 ), , c i.i.d.(0, c . . . C
(2.5)
where Xc is a (Bc (n1 + k1 )) matrix with rows {xc1 , xc2 , . . . , xcBc } and Yc is a (Bc (nc + kc )) matrix with rows {yc1 , yc2 , . . . , ycBc }. We allow the disturbances to have dierent
It is possible to obtain an arbitrary high goodness of t by increasing the number of parameters. However, the resulting term structures are likely to have twisting shapes and wide condence intervals.
4
Section 2.3 Model Settings
23
variances for each category, because prices of lower rated bonds are generally more noisy due to lower liquidity and a higher uncertainty about their perceived credit worthiness. Also, the residuals of independently estimated single-curve models can be shown to have 1 , 2 , . . . , C of signicantly dierent variances using a heteroscedasticity test. Estimates the spline weights are readily obtained by applying Restricted Feasible Generalized Least Squares estimation to (2.5); see e.g. Greene (2000, page 473). Once we have estimated c () for any desired the spline weights, we can evaluate the category-c discount curve D maturity. It is important to emphasize, however, that discount factors for maturities beyond the maximum maturity bond become unreliable.
2.3
Model Settings
Before we are able to estimate the regression model (2.5), we have to specify the exact form of the basis functions. The functional form of the basis functions follows by choosing the degree of the splines and the number and location of the knots. These choices reect the familiar trade-o between exibility and smoothness. The degree of the splines should not be chosen too high, to preclude the problems of higher order polynomials.5 If the order is too low however, the estimated curve will not t the data very well and thus will not reect the interest rates that are prevalent in the market. Similarly, if the number of knots is chosen too low, the model will not be able to t term structures with dicult shapes. On the other hand, if it is too high, the estimated curve is sensitive to outliers. For the spline model for the government discount curve, we can use results from the term structure estimation literature. Almost all studies that employ spline functions to model the discount curve use third degree splines. Only McCulloch (1971) used quadratic splines in his pioneering study, resulting in knuckles in the forward curve, which made him switch to cubic splines in his follow-up paper (McCulloch, 1975). Beim (1992) conducted a simulation study and concluded that cubic splines are preferable. Poirier (1976, page 49) demonstrated that tting a cubic spline minimizes the integral of the square of the second derivative, which is an approximation of a functions smoothness; see also Adams and van Deventer (1994). Consequently, cubic splines are a convenient compromise between high goodness of t and smooth curves. With regard to the specication of the knots, there is less agreement. McCulloch (1971) proposed to set the number of knots equal to the integer nearest to the square root of the
Using higher order polynomials often results in spurious curvature between the data points; see e.g. Shea (1984, page 255). This is especially true if such polynomials are tted to data that are not uniformly distributed over the approximation interval, as is the case with term structure estimation; see Figure 2.1.
5
24
number of bonds in the sample. The knots are then located such that an approximately equal number of bonds is placed in each segment. Litzenberger and Rolfo (1984) stated that the McCulloch scheme is likely to result in a poor t for longer maturities due to the larger number of shorter maturity bonds. As an alternative, they suggested to exogenously place the knots at 1, 5 and 10 years, roughly corresponding to an economic segmentation into short, medium and long maturities. Langetieg and Smoot (1989) tested the McCulloch knot placement scheme against the economic scheme and found that the latter typically performed better. Steeley (1991) experimented with the specication of the knots and recommended placing knots at 5 and 10 years as a starting point for future research. The simulation study by Beim (1992) revealed that cubic splines with two knots minimized the standard error of t between the estimated and the simulated true discount curves. We also use spline functions to model discount spread curves, but there is no prior evidence available on the specication of the degree and the knots. Given the disagreement in the literature on the specication of the default-free discount curve, our task of specifying the splines of the spread curve is not an easy one. Our choices are guided by the observation that a spread curve generally has a less complicated shape than a term structure. Therefore, we reduce the exibility of the spline model for the spread curve by reducing the degrees of freedom. Compared to the spline for the discount curve of category c in a single-curve model, we specify the discount spread curve sc () in the multi-curve model as a lower degree spline with a smaller number of knots. That is, we use a quadratic spline function and the knots are chosen to be a subset of the knots of the single-curve spline model. This still leaves us with several competing degree-knot combinations. To choose the optimal combination we use a newly developed test statistic that allows us to compare spot spread curves that are obtained from competing multi-curve models. We describe this curve similarity test in the next section and apply it in Section 2.6.
2.4
Model Comparison
A problem in comparing dierent single- and multi-curve models amongst and against one another is that there is no general estimable model that encompasses all other models. Therefore, we cannot use standard econometric testing procedures. Moreover, most econometric tests only focus on goodness of t, i.e. the ability of a model to t the data. In term structure estimation, however, practitioners are additionally interested in other features of the models, such as smoothness and statistical reliability. For these reasons, we compare single- and multi-curve models in three ways:
Section 2.4 Model Comparison
25
Usage of statistics that reect the goodness of t, smoothness and reliability, such that models can be compared by confronting these statistics, though without being able to determine the statistical signicance of possible dierences. Usage of a newly developed test statistic that allows two curves from two dierent multi-curve models to be compared to one another. We focus here on spot spread curves, because of their importance as inputs for pricing and risk management models, but the statistic may also be employed to compare other curves that can be calculated from the multi-curve models. Visual inspection of the estimated term structures, most notably the spot curves and the spot spread curves. Desirable features are smoothness and monotonicity. The remainder of this section discusses the rst two ways in more detail.
2.4.1
Statistics
Since interest rates are the main determinants of bond prices, any term structure model should be able to explain market prices fairly accurately. Therefore, goodness of t is a useful criterion to compare models. We measure the t as the Root Mean Squared Error of the residuals RM SEc = 1 Bc
Bc
e2 cb
b=1
where RM SEc denotes the Root of Mean Squared Error for category c, Bc the number of category c bonds and ecb the residual of the bth category c-bond, which is calculated as the market price of the bond minus its theoretical DCF price (2.4a) or (2.4b). Although a low value of the RMSE statistic is desirable, we run the risk of ending up with a twisting curve. Therefore, we also measure the smoothness of estimated term structures. Following Poirier (1976, page 49) and Powell (1981), the smoothness of a function over an interval [t1 , t2 ] is computed as the integral of the square of its second derivative
t2
s(, t1 , t2 ) =
t1
(t)2 dt.
We evaluate this statistic for both spot curves and spot spread curves. Finally, we want to judge to what degree deviations between theoretical and market prices is transformed into uncertainty about estimated interest rates and spreads. The reliability of a point (maturity) on an estimated curve is indicated by its standard error.
26
The reliability can be evaluated in a number of maturities to compare dierent segments of the curve. Appendix 2.B derives the standard errors for a number of curves: discount curve, discount spread curve, spot curve and spot spread curve.
2.4.2
Curve Similarity Test
The Curve Similarity Test described below helps in choosing between two curves that are estimated with two dierent multi-curve models. The test especially guides in striking a balance between goodness of t and smoothness; see Section 2.3. Given our focus on credit risk models, we describe the construction of the test for spot spread curves, but the test is suitable for any other curve for which standard errors and covariances can be computed. Suppose we estimate a richly specied multi-curve model and compose a vector s1 r,c (t) of spot spread rates of category c evaluated in a q -vector of maturities t = (t1 , . . . , tq ). We would like to know to what extent we can reduce this model to a more parsimonious model, i.e. a smoother curve, without loosing too much on the goodness of t criterion. Consider therefore the vector s0 r,c (t) that contains spot spread rates evaluated in the same vector of maturities that result from a more parsimonious multi-curve model. This alternative model contains less parameters due to a lower degree and/or less spline knots. The Curve Similarity Test (CST) aims to test whether s0 s1 r,c (t) lies in the realm of r,c (t). To compute the CST statistic we weigh dierences between the spot spread vectors with the covariance matrix 1 s1 c (t) of r,c (t)
1 1 1 CST = ( s1 s0 sr,c (t) s0 r,c (t) r,c (t)) (c (t)) ( r,c (t)).
The covariance matrix, which is constructed in Appendix 2.B, measures the uncertainty in the spread estimates and by using it as weight matrix we put more emphasis on the reliable maturities of the spread curve, and vice versa. We compare the CST value to critical values from a 2 distribution with q degrees of freedom to determine whether s0 r,c (t) is approximately equal to s0 1,c (t) at the selected maturities. The testing procedure can only be applied to spot spread vectors from multi-curve models. Spread curves from singlecurve models are obtained by subtracting independently estimated term structures, so that we are unable to construct the covariance matrix of a spread vector. As s1 r,c (t) curve we choose the multi-curve model with the same degree-knot settings as the single-curve model, because its spread curve resembles the spread curves obtained from single-curve models the most; see Section 2.6. As s0 1,c (t) curves we consider several more parsimonious multi-curve models, i.e. with a lower degree and/or less knots. These dierent parsimo-
Section 2.5 Data
27
nious models are all compared to the most richly specied model. The results that stem from such a model comparison should be interpreted with care as the testing procedure is conceptually dierent from standard econometric testing procedures. For example, the test statistic may prefer a model with low order splines that has appropriately selected knots to a high order model with badly located knots. Therefore, the test may reject a model that has a larger number of parameters than a competing model that is not rejected; this outcome is not possible with traditional econometric tests that compare nested models. To make the test operational we have to specify the maturity vector t. Since the covariance matrix of the spot spread vector s1 r,c (t) is derived from the covariance matrix 1 1 1 , c ) of the richly specied model, we cannot construct the covariof the estimators ( ance matrix for an arbitrarily chosen maturity vector t. For example, if the number of maturities q exceeds the number of parameters in the underlying regression model, the covariance matrix of s1 r,c (t) becomes singular. Another issue is the location of the maturities. Because of the smoothness of the curve, spreads for two adjacent maturities cannot be very dierent from each other. Therefore, the grid points should not be chosen too close to each other to preclude a near-singular covariance matrix. A nal consideration is the location of the maturities relative to the spline knots. Since each spline interval corresponds to an extra parameter, we cannot place too much grid points of the test in one spline interval. Again, doing so would lead to a near-singular matrix. In practice, the above mentioned conditions on the maturity vector t imply that we can only conduct a joint comparison of the spot spreads at a limited number of maturities, which lie reasonably far apart. To determine the robustness of the results from the testing procedure we can vary the maturity vector t while satisfying the conditions.
2.5
Data
To appraise the performance of the proposed multi-curve model and compare it to independently estimated single-curve models, we use a data set of German mark-denominated bonds. Their characteristics, like maturity dates, coupon percentages and credit ratings, are obtained from Bloomberg, whereas bond quotes are retrieved daily at 4.00pm from Reuters Treasury and Eurobond pages. These Reuters pages are connected to broker pages, and each time a broker updates a quote for a bond, that quote is also refreshed on the Reuters page. Therefore, the Treasury and Eurobond pages provide a good representation of the market for German mark-denominated bonds.
28
For illustration, we present the results for the trading days of June 1998, yielding a total of 1291 quoted bonds. To estimate the term structures, we construct a sample of xed-coupon, bullet bonds and to ensure their liquidity, we only use bonds that are quoted on at least 18 of the 20 trading days of June 1998; see also the liquidity criterion missing prices in Chapter 3. There are 624 bonds that satisfy these conditions. In estimating the term structure on a particular trading day, we also consistently exclude all bonds with a remaining maturity of less than 1 year. Unlike the US and UK Treasury Bills, short-term German government bonds and short term discount bonds of other credit ratings typically have low liquidity. In our data set, such bonds showed constant prices or they were not quoted at all for several consecutive days Since we cannot use short-term bonds, we add four synthetic zero-coupon bonds to the sample, whose prices correspond to 1-, 3-, 6- and 12-month money-market rates, respectively; see also B uhler, Uhrig-Homburg, Walter and Weber (1999). We use Frankfurt Interbank Oered Rates (FIBORs) as money-market rates, but since most commercial banks have an AA rating these rates are not straightly applicable to other rating classes. Therefore, we correct the FIBORs by adding or subtracting a category- and maturitydependent spread. The price Pct of a synthetic bond for category c of maturity t is thus computed as Pct = 1 . (1 + FIBORt + correctionct )t
Note that we do not constrain the curves to pass exactly through the corrected FIBORs; the synthetic bonds are just additional data points to support the curve in a sparse data segment. Due to the subjectivity of corrections, however, care must be taken in using rates from the short end of the estimated term structures. The number of suitable bonds of a single corporate debtor is too small to reliably estimate a separate term structure for each debtor. Our data set comprises 168 unique issuers; only 14 of them have 10 or more suitable bonds outstanding, and only 2 out of these 14 issued more than 20 suitable bonds. Therefore, we resort to grouping rms by rating and industry. First we show a division of the included bonds by rating; see Table 2.1. We use ratings published by Bloomberg, which compounds the major ratings of rating agencies Moodys and Standard & Poors. We consider AAA-rated government bonds as a separate category, indicated by rating symbol GOVT. From the table, it is clear that the number of bonds per rating decreases with credit quality. Therefore, the reliable estimation of term structures of lower rating categories may be hampered by an insucient number of bonds if we were to use a single-curve method. The proposed multi-
Section 2.5 Data
29
Table 2.1: Distribution of bonds in the data set by rating.

GOVTa AAA AA A BBB BB B NRb Total
Allc Includedd
a b c d
112 93
411 228
329 145
119 53
30 13
53 16
38 16
199 60
1291 624
AAA-rated government bonds. Not rated. All German mark Eurobonds that are quoted at least once in June 1998. Fixed-income, bullet bonds that are quoted on at least 18 of the 20 trading days of June 1998.
Figure 2.1: Distribution of the bonds maturity dates by rating category.
GOVT
AAA
AA
BBB
BB
B 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020 2022 2024 2026 2028
curve model oers a solution, because it only focuses on the spread curve and involves less parameters. The distribution of the bonds maturity dates is shown in Figure 2.1. For government bonds and AAA bonds, large gaps in the maturity distribution are observed beyond a maturity of 10 years. This is caused by the very infrequent issuance of long term bonds (e.g. 30-year bonds), as opposed to the regular issuance of 5-year and 10-year bonds. Therefore, a term structure estimation procedure has few data points beyond 10 years and identication of that part of the term structure is more dicult.
30
Table 2.2: Distribution of included bonds in the data set by rating and industry.
Financialsa Government Agenciesb Supra-Nationalsc Industrialsd Total
AAA AA
a b c d
139 99
23 33
59 0
7 13
228 145
Consists of Reuters classications Banks and Financials. Government National, Government Regional and Government Agency. Worldbank, Inter-American Development Bank and Supra-National. Gas & Transmission, Utility & Electricity, Transport-Nonrail, Telephone and Industrial.
For credit classes AAA and AA there are enough bonds to allow a further classication by industry, but for the other ratings this is not feasible due to the limited number of bonds. The included bonds from ratings AAA and AA are assigned to one of four compounded Reuters industry classications. According to Table 2.2, the majority of the bonds is issued by nancial institutions.
2.6
2.6.1
Results
Single-Curve Results
Before we discuss the results of the multi-curve (MC) model, we rst present some estimation results of single-curve (SC) models. We only describe the results for the rating classes in Table 2.1. The results for the industry classications of Table 2.2 are similar and do not provide additional insights into the performance of the models. Table 2.3 summarizes the specications of the SC model. We use third degree B-splines for all bond categories, but the number and placement of the knot points dier. For the ratings GOVT, AAA, AA and A, we set the knots at 3 and 9 years, approximately corresponding to a segmentation into short-, medium- and long-term maturities. For the lower rating classes, BBB, BB and B, we use only one knot at 5 years, because the number of bonds in these classes is relatively small. The table also mentions the corrections to be applied to the AA-rated Frankfurt Interbank Oered Rates (FIBORs) in the calculation of the synthetic bonds prices to take credit risk into account. The corrections are taken from Bloomberg and are inevitably approximations. For simplicity, we apply the same correction to all four synthetic bonds of a rating. The estimated curves are rather insensitive to small changes in these corrections, though completely omitting them is ill advised.
Section 2.6 Results
31
Table 2.3: Model specications for single-curve and multi-curve models.

Single-curve Degreea Knotsb Multi-curve Degreea Knotsb FIBOR correctionc
GOVT AAA AA A BBB BB B

a, b c
3 3 3 3 3 3 3
3,9 3,9 3,9 3,9 5 5 5
3 2 2 2 2 2 2
3,9 9 9 9 5 5 5
-20 -10 0 +10 +20 +30 +40
Degree and knots refer to the specication of the B-splines. Spread (in basis points) applied to Frankfurt Interbank Oered Rates (FIBORs) in calculating the prices of synthetic bonds.
Table 2.4 summarizes the term structure estimations for all 20 trading days of June 1998 by averaging our evaluation statistics for goodness of t, smoothness and reliability. Figure 2.2 graphically illustrates the SC estimation results for the rst day of our data set, 2 June 1998, by depicting (a) spot interest rate curves with their 95% condence intervals and spot spread curves, (b) residuals and (c) standard errors for the estimated discount curves. Note that we estimate discount curves and discount spread curves and that the corresponding spot curves and spot spread curves are subsequently calculated from these estimates. The residual scatter plots in Figure 2.2b show that goodness of t decreases with the credit quality of the bonds. The market prices of government bonds are reasonably approximated by the theoretical DCF prices, since all absolute deviations, except one, are less than 25 basis points (bps). The scatters for the bond categories AAA, AA, A and BBB are more dispersed, but the absolute errors are still smaller than 100 bps. Fitted prices for BB- and B-rated bonds are the least accurate with the largest residuals being about 1000 bps. Similar conclusions are drawn from the RMSE statistic in Table 2.4, which has the lowest value for GOVT bonds and generally increases for lower rated bonds. Apparently, the bonds in lower rating categories are increasingly more heterogeneous, which can be attributed to relative dierences in perceived default probabilities and recovery rates of the issuers. Liquidity dierences between the bonds may also contribute to the larger dispersion of pricing errors.
32
Figure 2.2: Single-curve (SC) estimates for 2 June 1998.

6% SC GOVT spot 5% 0.5 0.0 1998 -0.5 -1.0 0.3% 1.0 SC AAA residuals 0.2% 0.5 0.0 1998 -0.5 3% 0 6% 5% 0.2% 4% 3% 0 5 10 15 0.1% 0.0 1998 -0.5 -1.0 1.0 0.5 0.5% 0.3% 0.1% 0 6% 5 SC BBB spot SC BBB spread 10 15 1.0% 0.8% 0.6% 4% 0.4% 0.2% 0 5 SC BB spot SC BB spread 10 15 -1.0 10 10% 8% 6% 4% 2% 0% 0 5 SC B spot SC B spread 10 15 -10 10 10% 8% 6% 4% 2% 0% 0 5 10 15 -10 5 0 1998 -5 SC B residuals 5 0 1998 -5 0 1998 40 30 20 2003 2008 2013 2018 2023 2028 10 0 1998 2003 2008 2013 2018 2023 2028 SC BB residuals 0.0 1998 -0.5 -1.0 SC BBB residuals 2003 2008 2013 2018 2023 2028 SC A residuals 0 1998 3 2 1 0 1998 2003 2008 2013 2018 2023 2028 2003 2008 2013 2018 2023 2028 1 5 10 15 20 25 30 0.3% 0.0% -1.0 1.0 0.5 SC AA residuals 0 1998 3 2003 2008 2013 2018 2023 2028 2 1.0 SC GOVT residuals 3 GOVT bonds SC GOVT s.e.
4%
2003
2008
2013
2018
2023
2028
3% 0 6% 10 SC AAA spot SC AAA spread 20 30
0 1998 3
2003
2008
2013
2018
2023
2028
AAA bonds SC AAA s.e.
5%
4%
0.1%
2003
2008
2013
2018
2023
2028
SC AA spot SC AA spread
AA bonds SC AA s.e.
6% 5% 4% 3%
SC A spot SC A spread
0.7%
A bonds SC A s.e.
2003
2008
2013
2018
2023
2028
1.0 0.5 0.0 1998 -0.5
BBB bonds SC BBB s.e.
5%
2003
2008
2013
2018
2023
2028
3%
0 1998 40 30 20 2003 2008 2013 2018 2023 2028 10
2003
2008
2013
2018
2023
2028
12% 10% 8% 6% 4% 2% 0%
BB bonds SC BB s.e.
12% 10% 8% 6% 4% 2% 0%
B bonds SC B s.e.
2003
2008
2013
2018
2023
2028
(a)
(b)
(c)
Graphs (a) show estimated spot curves (with 95% condence intervals) and spot spread curves; graphs (b) contain residuals; graphs (c) display (1000 times) the standard errors (s.e) of estimated discount curves.
Section 2.6 Results
33
Table 2.4: Summary statistics of single-curve and multi-curve estimates.

Goodness of ta Smoothnessb Spot Spread Reliabilityc 2y 5y 10y
Single-Curve GOVT AAA AA A BBB BB B Multi-Curve AAA AA A BBB BB B

a b c
0.09 0.32 0.28 0.32 0.47 3 .5 5.3 0.33 0.29 0.32 0.52 3 .8 5.3
4.5 2.4 9.1 23 80 1049 689 2.2 4.5 4.6 6.8 31 829
2.5 3.6 7 .5 61 1076 594 0.1 0.7 0.9 1.5 131 788
1 .5 3 .2 3 .5 6 .2 23 177 209 2 .0 2 .2 3 .7 17 36 87
1 .7 3 .3 3 .9 6 .5 25 143 383 2 .6 2 .9 4 .7 19 30 110
4 .6 6 .2 9 .3 19 186
5 .4 7 .8 14 39
Root Mean Squared Error (RMSE) of the residuals. (108 times) the integral of the square of the second derivative (104 times) the standard error of the estimated spot curve, evaluated at maturities 2, 5 and 10 years.
The extent to which market prices can be tted accurately, has consequences for the reliability of the estimated curves. For the GOVT, AAA, AA and A curves, the condence bounds in Figure 2.2a almost coincide with the spot curve, whereas for the BBB, BB and B curves the upper and lower bounds are distinctly observable in the graphs. Therefore, the reliability decreases with credit worthiness. The plotted standard errors in Figure 2.2c and the reliability statistics in Table 2.4 also show that the estimation error increases in maturity segments that contain a small number of bonds. Most interesting for our purposes are the credit spread curves that can be obtained by subtracting the estimated corporate spot curve from the estimated government spot curve. All spot spread curves of 2 June 1998 in Figure 2.2a have an unrealistically twisting shape, because they have alternately positively and negatively sloped segments. Especially the spread curves for investment-grade bonds display twisty behavior. This can be explained from the relatively small magnitude of these spreads compared to interest rate levels: spot interest rates for classes AAA, AA, A and BBB are approximately 3.5% to 5.5%, whereas spot spreads range from 0.1% to 0.6%. Therefore, small deviations in
34
either the government or the corporate spot curve can result in substantial irregularities in the spread; in other words, any dissimilarity in the curves curvatures implies twists in the spread curve. These uctuations are indeed observed in the graphs. The twisting spread curves are not in line with the smooth spread curves that are predicted by the theoretical bond price models of Merton (1974) and Longsta and Schwartz (1995a). The curves also contradict the empirical research by Helwege and Turner (1999) who found statistically signicant evidence for increasing credit spreads for both investmentgrade and speculative-grade issuers. Usage of these spread curves for, e.g., Value at Risk calculations or credit derivatives pricing is likely to result in erroneous outcomes.
2.6.2
Multi-Curve Results
In the MC model we explicitly and parsimoniously model the spread curve, so that we hypothesize to obtain smooth spread curves. The model for a corporate bond category now only has to focus on the discount spread relative to the default-free government discount curve. Together with the use of a combined data set of government and corporate bonds, we therefore expect the parameters to be estimated more reliably. A consequence of the joint parameter estimation is that the parameters of the government spline model change somewhat compared to the SC estimates. Therefore, the government curve obtained from the MC model diers somewhat from the SC government curve. The changes are very small, though, and here we are only interested in the corporate curves. Table 2.3 shows the specication of the degree and knots for the MC model. For the government discount curve, we use exactly the same settings as in the SC model, but for all corporate discount spread curves we lower the degree and for some curves also the number of knots. Rating categories AAA, AA and A use quadratic splines with one knot at 9 years, while the BBB, BB and B discount spread curves are modelled as quadratic splines with one knot at 5 years. Later we show how to nd these specications, but rst we show that they indeed yield the hypothesized favorable properties of the MC model over its SC competitor. Note that we use the same corrections to FIBOR to value the synthetic zero-coupon bonds as in the SC estimations, so that any bias that may be caused by using incorrect values is equal for the SC and MC models. Figure 2.3 provides a graphical representation of the results of 2 June 1998 and Table 2.4 again summarizes the main characteristics by averaging the goodness of t, smoothness and reliability statistics over all 20 trading days of June 1998. The graphs in Figure 2.3a contain the estimated spot spread curves from the MC model and, for comparison, also the SC spot spreads from Figure 2.2a. For all ratings, there is a
Section 2.6 Results
35
Figure 2.3: Single-curve (SC) and multi-curve (MC) estimates for 2 June 1998.
0.3% SC AAA spread: degree 3, knots {3,9} MC AAA spread: degree 2, knots {9} 0.2% 0.5 0.0 1998 -0.5 0 0.4% 0.3% 0.2% 0.1% 0 5 10 15 -1.0 1.0 0.5 0.0 1998 -0.5 0 0.8% 0.7% 0.6% 0.5% 0.4% 0.3% 0 5 10 15 10 5 0 1998 -5 0 7% 6% 5% 4% 3% 2% 1% 0% -1% 0 5 10 15 -10 MC B residuals SC B residuals MC BB residuals SC BB residuals 0.0 1998 -0.5 -1.0 2003 2008 2013 2018 2023 2028 5 10 15 -1.0 MC BBB residuals SC BBB residuals MC A residuals SC A residuals 5 10 15 20 25 -1.0 1.0 0.5 0.0 1998 -0.5 MC AA residuals SC AA residuals 1.0 MC AAA residuals SC AAA residuals 3 2 1 0 1998 3 AAA bonds MC AAA s.e. SC AAA s.e.
0.1%
2003
2008
2013
2018
2023
2028
0.0%
2003
2008
2013
2018
2023
2028
SC AA spread: degree 3, knots {3,9} MC AA spread: degree 2, knots {9}
AA bonds MC AA s.e. SC AA s.e.
2003
2008
2013
2018
2023
2028
0 1998 3
2003
2008
2013
2018
2023
2028
0.7% 0.6% 0.5% 0.4% 0.3% 0.2% 0.1%
SC A spread: degree 3, knots {3,9} MC A spread: degree 2, knots {9}
A bonds MC A s.e. SC A s.e.
2003
2008
2013
2018
2023
2028
0 1998 3
2003
2008
2013
2018
2023
2028
SC BBB spread: degree 3, knots {5} MC BBB spread: degree 2, knots {5}
1.0 0.5
BBB bonds MC BBB s.e. SC BBB s.e.
0 1998 40 30 20 2003 2008 2013 2018 2023 2028 10 0 1998 40 30 20 2003 2008 2013 2018 2023 2028 10 0 1998
2003
2008
2013
2018
2023
2028
5% 4% 3% 2% 1% 0%
SC BB spread: degree 3, knots {5} MC BB spread: degree 2, knots {5}
BB bonds MC BB s.e. SC BB s.e.
2003
2008
2013
2018
2023
2028
SC B spread: degree 3, knots {5} MC B spread: degree 2, knots {5}
10 5 0 1998 -5
B bonds MC B s.e. SC B s.e.
10
15
-10
2003
2008
2013
2018
2023
2028
(a)
(b)
(c)
Graphs (a) show spot spread curves (with 95% condence intervals for MC spreads); graphs (b) contain residuals; graphs (c) display (1000 times) the standard errors (s.e) of estimated discount curves.
large improvement in the smoothness of the spread curves. Compared to the uctuating SC spread curves, the MC spread curves are smooth, mostly increasing functions of time to maturity. For the longest maturities however, all spread curves, except for ratings AA and A, have a negatively sloped segment; this behavior may be caused by a sampling bias
36
as discussed in Helwege and Turner (1999). Table 2.4 adds quantitative evidence to these graphical insights: the smoothness statistic is decimated for all ratings, except B. The improvements range from a factor 5 for A to 40 for BBB-rated bonds. Corporate spot curves, calculated as the sum of the estimated government and spread curves, also become smoother, again with rating B as the only exception. Note that the smoothness statistics for the spot spread curves are now much smaller than for the spot curves, conrming our prior belief in Section 2.3 that spreads have a less complicated shape. The second advantage of the MC model over independently estimated SC models resides in the increased reliability of the estimated corporate term structures.6 Figure 2.3c and Table 2.4 show that for all ratings and all maturities the standard errors of MC curves are lower than those of SC curves. The reliability for the BB and B categories improves most. For segments with a small number of bonds, such as the interval 2010-2020 for category AAA and 2004-2006 for AA, the standard errors become much smaller. This can be attributed to the combination of data sets, which improves the density of the distribution of maturity dates. Since the total number of parameters in the MC model is smaller than in the corresponding SC models, the goodness of t may be expected to decrease. This is also predicted by the familiar trade-o between exibility and smoothness. However, the price that we have to pay for the improvements in smoothness and reliability is rather modest, because the RMSE statistics in Table 2.4 are approximately equal for the SC and MC model for all rating categories. The scatter plots in Figure 2.2b also show about the same dispersion of the residuals and do not reveal any signicant biases. Apparently, the focus on the credit spread and the joint estimation with the default-free government curve oset the negative eects of the imposed parsimonious structure.
2.6.3
Robustness
In this section we show that the shape of the spread curve in the MC model is relatively robust to the precise specication of the model, whereas the SC model is more sensitive to model misspecication. Along the way, we illustrate that the favorable properties of the MC model cannot be obtained from a parsimoniously specied corporate SC model or from a richly specied MC model. We show these results for the AAA-, AA- and A-curves, since for these curves we can consider more alternative degree and knot settings.
Note that we cannot judge the change in reliability of the spread curves, because we are unable to calculate standard errors of estimated spreads for the SC case: the parameters of the government and corporate term structure models are estimated independently, so we do not get the required covariances.
Section 2.6 Results
37
As starting point for the specication of the discount spread curves of AAA, AA and A in the MC model, we choose the specication of their discount curves in the SC model, i.e. a cubic spline with knot scheme {3, 9}. Subsequently, we reduce the exibility of the spline model by lowering the degree or the number of knots (or both). Including the initial combination, this yields 6 alternative specications: degree 3 and knots {3,9}; 3,{3}; 3,{9}; 2,{3,9}; 2,{3} and 2,{9}. The specication for the government curve is unchanged as a cubic spline with knots {3, 9}. Figure 2.4 shows estimated spot spread curves for the three ratings using MC models with the 6 degree/knots combinations. For comparison, the gure also contains spot spread curves that are calculated from independently estimated SC models, where the specication of the GOVT discount curve is unchanged and the AAA-, AA- and A- discount curves are specied according to one of the 6 degree-knots combinations. The bold lines in Figure 2.4 depict the estimated SC and MC spreads for the rst combination. The shapes of the MC spreads are just as twisty as those of the SC models. It is therefore clear that an MC model with the same number of parameters as an SC model, does not yield the described favorable results of a parsimoniously specied MC model. We can smooth the spread curve by decreasing the spline degree and reducing the set of knots. We apply this idea for both the corporate discount curves in the SC model and the corporate discount spread curves in the MC model. The graphs in Figure 2.4 reveal that the shapes of the SC spot spread curves calculated with these new settings vary considerably. This shows that estimating a parsimoniously specied SC model does not yield smooth and intuitively shaped spread curves. The spread curves obtained from the MC model, however, stien and are not as sensitive to dierent degree-knots settings. Particularly for AA and A, there is much less variation in shape and curvature of the curves. This robustness to model specication is a distinct advantage of the MC model over its SC competitor.
2.6.4
Spread Curve Specication
To choose between the competing parsimonious MC models of the previous section, we apply the Curve Similarity Test (CST) of Section 2.4. The most richly specied MC models, i.e. with the highest degree and the largest number of knots, are used as base case. These models are restricted by lowering the degree and/or reducing the number of knots. The null hypothesis in each CST claims that the spot spread curve of a restricted model is equal to the curve of the largest model if evaluated in a prespecied maturity vector. Section 2.4 discussed several criteria on the dimension and the spacing of this
38
Figure 2.4: Spread curves.

" "

Spot spread curves estimates for 2 June 1998 for ratings AAA, AA and A estimated with (a) single-curve (SC) models and (b) multi-curve (MC) models, with dierent degree and knot settings.
maturity vector; one choice that satises these criteria is to set the maturities at the spline knots and at maturities exactly between these knots. For rating AAA, for example, the knots of the largest MC model are set at {0, 3, 9, 25} (including the end points), so that we specify the maturities of the CST as {1.5, 3, 6, 9, 17}. Table 2.5 shows the maturity vectors for the other rating classes7 , as well as the settings of the most richly specied MC model and of the restricted MC models.
7
We also conducted the test with other maturity vectors, where all maturities were either decreased or increased by 0.25 years. Since the results were about the same, the CST statistic is fairly robust to the precise choice of the maturity vector.
Section 2.6 Results
39
Table 2.5: Results of Curve Similarity Tests applied to spot spread curves obtained from multi-curve models with dierent degree and knot settings.
H0 c 2,{9} 3,{3}
Maturities
H1
2,{3,9}
3,{9}
2,{3}
2,{5}
AAA AA A BBB BB B
a b c
{1.5, 3, 6, 9, 17} {1.5, 3, 6, 9, 10} {1.5, 3, 6, 9, 10.5} {2.5, 5, 6.5} {2.5, 5, 8} {2.5, 5, 7.5}
3,{3,9} 3,{3,9} 3,{3,9} 3,{5} 3,{5} 3,{5}
0.00 0.92 0.37
0.16 1.00 0.87
0.00 0.09 0.07
0.00 0.76 0.85
0.00 0.93 0.49 0.07 0.04 0.50
Maturities in which the test statistic is evaluated. Degree and knots of the most richly specied multi-curve model used as alternative hypothesis Average p-values of the test for all restricted models over all trading days of June 1998.
For each trading day of June 1998, we calculate the CST statistic and confront the test statistics with critical values from a 2 -distribution to obtain 20 p-values for each rating and model specication. The averages of these p-values are reported in Table 2.5. We rst consider the results of AA, a case where the CST is indeed helpful in the specication process. All alternative degree-knot combinations, except degree 2 with knot {9}, yield average p-values well in excess of any reasonable condence level. This means that the spot spread curves corresponding to these combinations are - on average - not signicantly dierent from the most richly specied MC model. In other words, these combinations result in spreads that are just as twisting as those from the SC model of Section 2.6.1. The model with degree 2 and one knot at 9, on the other hand, delivers spread curves that are on the edge of the rejection/non-rejection interval with an average p-value of 0.09. In the sense of Section 2.4, this model strikes the optimal balance between goodness of t and smoothness. A similar conclusion may be drawn for ratings A, BBB and B, whose p-values range from 0.04 to 0.07. For AAA-rated bonds, all models, except 3,{9}, are signicantly dierent from the largest MC model. Taking into account the very twisting nature of the spread curve in Figure 2.4b this is hardly surprising. In this case, we feel that visual inspection of the estimated spread curves has to take prevalence over the use of the CST statistic. Thus, any of the four rejected degree-knots combinations may be used and for consistency with AA and A we choose the model with degree 2 and one knot at 9. Likewise, for rating class B we use the same settings as for BBB and BB.
40
2.7
Summary
In this chapter, we presented a new framework for the joint estimation of term structures and credit spreads. By decomposing a corporate term structure in a default-free curve and a credit spread curve, we could use a parsimoniously specied model for the spread curve and take the default-free part from the government curve. Both the government and the spread curve were modelled as B-splines and their parameters were jointly estimated from a combined data set. We used a data set of liquid, German mark-denominated bonds, with Standard and Poors ratings ranging from AAA to B. For comparison, we rst independently estimated separate single-curve term structure models for each of the rating classes. Spread curves that were obtained by subtracting the estimated government and corporate curves, had unrealistically twisting shapes and we argued that this could be attributed to the relatively small magnitude of spreads compared to interest rates. Therefore, any dissimilarity in the curvatures of the government and corporate curves implies twists in the spread curve. Next, we applied the proposed multi-curve model that explicitly and parsimoniously models the spread curve and jointly estimated it with the government curve. We illustrated that the model yielded smooth spread curves that were more in line with the theoretical bond price models. Because the parameters were estimated from a combined data set of government and corporate bonds, the reliability of estimated term structures improved considerably. This eect was strongest for those maturity segments of a term structure that contained only a small number of bonds for that particular rating class. In spite of the smaller number of parameters, the ability of the multi-curve model to accurately t market prices of bonds was hardly aected. We found that the favorable results of the model could be attributed to both the joint and the parsimonious modelling of the spread curves. To determine the optimal settings of the spline model for the spread curve, we used a newly developed test statistic that allowed us to compare spot spread curves that were calculated from competing multi-curve models.
Section 2.A B -Splines
41
Appendix
2.A B -Splines
Below we give a brief description of the construction of a basis of B-splines; see Powell (1981) for a more elaborate discussion and Steeley (1991) for an application of B-splines to term structure estimation. Given n + 1 knots 0 < 1 < ... < n , the k th degree B -spline basis function is dened as
s+k+1 k Bs (t) = l=s
1 max (t l , 0)k , h l h=s,h=l
s+k+1
where subscript s denotes that this B-spline is only non-zero if t is in [s , s+k+1 ]. Powell (1981, page 234) provided an ecient recurrence relation to evaluate the spline functions
k Bs k1 k1 (t s ) Bs (t) + (s+k+1 t) Bs +1 (t) (t) = , s+k+1 s
with start conditions =

i+1 t (i+1 i1 )(i+1 i ) ti (i+1 i )(i+2 i )
if j = i 1 if j = i if j = i 1, j = i.
1 Bj (t)
To construct a basis we need n + k linearly independent B-splines. Because a k th order B-spline is only non-zero in k + 1 subintervals, within the interval [0 , n ] only n k Bsplines are dened. To construct a basis of n + k functions, another (n + k ) (n k ) = 2k splines are required. A convenient way of choosing them so that they are also B-splines is to introduce extra knots {i ; i = k, k + 1, . . . , 1} and {i ; i = n + 1, n + 2, . . . , n + k } outside the interval 0 , n ]. Commonly, these auxiliary knots are set as {i = 0 + i(1 0 ); i = k, k +1, . . . , 1} and {i = n +(i n)(n n1 ); i = n +1, n +2, . . . , n + k }. Then
k ; s = k, k + 1, . . . , n 1 . we construct a basis of n + k B-splines consisting of Bs
42
2.B
Variances and Covariances
With spline estimation it is straightforward to calculate the variance of an estimated c (t) of the discount factor discount factor for a specic maturity t. Since the estimate D for maturity t is a linear combination of the parameter estimates, we can apply the result that the variance of a linear combination a of a vector of random variables with covariance matrix cov ( ) is given by the quadratic form var (a ) = a cov ( ) a. Hence, 1 (t) = g1 (t) cov 1 g1 (t), and var D c (t) = (g1 (t), gc (t)) cov 1 , c (g1 (t), gc (t)), c = 2, 3, . . . , C. var D c (t) D 1 (t) equals Similarly, the variance of an estimated discount spread s c (t) = D c gc (t) , c = 2, 3, . . . , C. var ( sc (t)) = gc (t) cov Because the spot curves and spot spread curves depend non-linearly on the estimated parameters, their variances are not straightforwardly computed. A useful approximation is given by the delta method; see e.g. Greene (2000, page 118): the variance of a function () of is approximated by the quadratic form var ( ( )) ( ) cov ( ) ( ) . (2.B.1)
c (t) The estimated spot rate r c (t) is a function of the estimated discount factor D c (t) = ln Dc (t) , r c (t) = 1 D t so that its variance is approximately 1 c (t) tD c (t) var D 1 c (t) tD c (t) var D c (t) tD
2
var ( rc (t))
Section 2.B Variances and Covariances
43
To construct the variance for an estimated spot spread curve (t) + s ln d c (t) t (t) ln d , t
s r,c (t) = r c (t) r 1 (t) =
(2.B.2)
1 and c we write it explicitly as a function of the estimated parameters 1 , c = s r,c (t) = 2 1 + gc (t) c ln g1 (t) t 1 ln g1 (t) t
Using (2.B.1), the variance is then approximated as var ( sr,c (t)) where g1 (t) s r,c (t) g1 (t) + = and 1 1 1 + gc (t) c tg1 (t) t g1 (t) s r,c (t) gc (t) . = 1 + gc (t) c c t g1 (t) (2.B.4a) (2.B.4b) s r,c (t) s r,c (t) , 1 c 1 , c cov s r,c (t) s r,c (t) , 1 c , (2.B.3)
The idea of a variance of a spot spread at a single maturity can be extended to a full covariance matrix of a vector containing spreads for several maturities. Let sr,c (t) be a vector function of the maturity vector t = (t1 , . . . , tq ) , where the ith element is given by the spot spread function (2.B.2) evaluated in maturity ti . Then we can apply a vector version of (2.B.1) to obtain a result similar to (2.B.3) 1 sr,c (t)) c (t) = cov ( sr,c (t) sr,c (t) 1 , c cov , 1 , c 1 , c
1 , c where sr,c (t) /
is a q (n1 + k1 + nc + kc ) matrix with the ith row equal to
(2.B.4a) and (2.B.4b) evaluated in ti .
Chapter 3
Measuring Corporate Bond Liquidity1
3.1
Introduction
The eect of liquidity on bond yields has been frequently studied in the recent nance literature. Since liquidity is a rather subjective concept, a lot of measures have been proposed to approximate the extent to which a bond is liquid or illiquid. For corporate bonds, where most transactions occur on the over-the-counter market, direct liquidity measures (based on transaction data) are often not reliable and dicult to obtain. Therefore, researchers resorted to indirect measures that are based on bond characteristics and/or end-of-day prices. This chapter makes a number of contributions to this literature on measuring corporate bond liquidity. First, we pay great attention to control for other sources of risk to properly identify the premium associated with liquidity risk. As far as we know, this is the rst study in this strand of literature to use the well-known Fama and French (1993) two-factor bond-market model to control for interest rate and credit risk and to augment it with individual bond characteristics, as recommended by Gebhardt, Hvidkjaer and Swaminathan (2001). Second, we do not make a subjective choice of which indirect liquidity measures to work with, but implement as much of the measures proposed in the literature as possible on our data set. We evaluate the relative performance of all measures, employing a method recently applied by Goldreich, Hanke and Nath (2002) on Treasury bonds. Third, the vast majority of empirical papers on sovereign and corporate bond liquidity studied data from the United States and relatively little is known about the extent to which these results apply to the euro market. Although euro corporate bond data were also studied by other authors, including Annaert and De
1
This chapter is a slightly revised version of the paper by Houweling, Mentink and Vorst (2003a).
46
Measuring Corporate Bond Liquidity Chapter 3
Ceuster (1999), Dimson and Hanke (2001), McGinty (2001) and D az and Navarro (2002), none of them analyzed the euro corporate bond market using data on individual bonds over a substantial time period. To properly identify a securitys premium for the liquidity risk it imposes on its bearer, researchers have to control for other sources of risk that aect the securitys expected return too. Fortunately, theory (like the standard discounted cash ow equation for defaultfree bonds and the reduced form credit risk models following Jarrow and Turnbull (1995) for defaultable bonds) nominates two risk factors: (i) interest rate risk and (ii) credit risk. Gebhardt et al. (2001) found that both the Fama and French (1993) term and default factors and the individual bond characteristics duration and rating are important to properly capture the impact of interest rate and credit risk on bond yields. We augment these four variables with a fth variable, a bonds denomination currency, to get our nal set of variables. We test the presence of liquidity eects after correcting for the eects of these ve variables. In this study, we use the Brennan and Subrahmanyam (1996) methodology to test whether liquidity is priced in the euro-denominated corporate bond market. We use eight indirect measures of bond liquidity: issued amount, coupon, listed, age, missing prices, price volatility, number of contributors and yield dispersion; see Section 3.3.5 for a detailed description. For each liquidity measure, we construct P , mutually exclusive portfolios by sorting all bonds on their value of the liquidity measure and assigning the rst 100/P % of the bonds to portfolio 1, the next 100/P % to portfolio 2, and so on, until the last 100/P % of the bonds are assigned to portfolio P . The P time series of portfolio yields are subsequently used in two regression models. In the rst model, the regression equation for each portfolio has its own intercept and under the null hypothesis that liquidity does not aect bond yields, these intercepts should be jointly zero. In the second model, all portfolios share a common intercept, but a portfolio-specic liquidity variable is added to the regression equation. Here, the null hypothesis is that the intercept and the coecient of the liquidity variable are jointly equal to zero. To determine the eectiveness of the dierent liquidity measures relative to each other, we run a series of regressions with pairwise combinations of the liquidity measures, as proposed by Goldreich et al. (2002). By running the regressions for all possible combinations, we can count the number of times a measure adds power to another measure, and vice versa, the number of times a measure subsumes another measure. This allows us to rank the dierent liquidity measures we consider. We use a detailed data set consisting of daily yields of individual corporate bonds denominated in euros or in one of the currencies of the euro-participating countries (legacy
Section 3.2 Literature
47
currencies). The results for the rst regression model indicate that the ve-variable model should be rejected for seven out of eight liquidity measures. The premium between liquid and illiquid portfolios ranges from 9 to 24 basis points and is the largest for the measures age and yield dispersion. For the second model the null hypothesis of no liquidity eects is even always rejected. Finally, the pairwise comparison tests point out that the measures price volatility and number of contributors add power to most other measures, and that most other measures are subsumed by them. The remainder of this chapter is structured as follows. Section 3.2 gives an overview of the theoretical and empirical liquidity literature. In Section 3.3, the Brennan and Subrahmanyam (1996) approach of testing the compensation for liquidity in equity returns and the Fama and French (1993) model are explained and our modications to both approaches are given. This section also describes the portfolio construction and the eight liquidity measures that are used in this construction. Next, in Section 3.4, the data that are used to test the hypotheses of corporate bond liquidity are described. In Section 3.5, the results from the model implementation are shown. Finally, Section 3.6 summarizes the chapter.
3.2
Literature
Both theoretical and empirical evidence demonstrate that liquidity risk is priced in security markets. The market microstructure models of Amihud and Mendelson (1986), Boudoukh and Whitelaw (1993) and Vayanos (1998) show that transaction costs cause liquidity dierences between securities, and that illiquid securities have higher expected rates of return than liquid securities. Empirical evidence on priced liquidity risk in equity markets is provided by, e.g., Amihud and Mendelson (1986), Brennan and Subrahmanyam (1996), Haugen and Baker (1996), Brennan, Chordia and Subrahmanyam (1998), Chordia, Roll and Subrahmanyam (2001) and Chordia, Subrahmanyam and Anshuman (2001). These studies on equity markets had to cope with an important drawback: they resorted to approximating expected returns by realized returns, which are, by denition, realizations of a stochastic process instead of expectations. For bonds, on the other hand, the yield-to-maturity can be used as expected return measure. It may be for this reason, that bond liquidity has been the topic of numerous papers. A substantial part of these studies analyzed data from US Treasury markets, including Sarig and Warga (1989), Amihud and Mendelson (1991), Warga (1992), Daves and Ehrhardt (1993), Kamara (1994), Elton and Green (1998), Fleming (2001), Strebulaev (2001), Fleming (2002), Goldreich et al. (2002) and Krishnamurthy (2002). Markets for other countries government bonds were stud-
48
ied by Boudoukh and Whitelaw (1991, 1993, Japan), Kempf and Uhrig-Homburg (2000, Germany) and Jankowitsch, M osenbacher and Pichler (2002, six EMU countries). Research on corporate bonds has also been predominantly conducted on US data; references include Cornell (1992, high yield mutual funds), Gehr and Martell (1992, investment grade bonds), Shulman, Bayless and Price (1993, high yield bonds), Crabbe and Turner (1995, new issues), Fridson and J onsson (1995, high yield indices), Chakravarty and Sarkar (1999, corporate, municipal and Treasury bonds), Alexander, Edwards and Ferri (2000, high yield bonds), Hong and Warga (2000), Collin-Dufresne, Goldstein and Martin (2001, corporate bonds), Ericsson and Renault (2001, zero-coupon bonds), Elton, Gruber, Agrawal and Mann (2002, corporate bonds) and Mullineaux and Roten (2002, corporate bonds). Non-US corporate bond data were used by Annaert and De Ceuster (1999, euro bond indices), Dimson and Hanke (2001, UK equity index-linked bonds), McGinty (2001, euro-denominated corporate bonds; data for only one month) and D az and Navarro (2002, Spanish corporate bonds). So, although there are numerous empirical papers on bond liquidity, none of them studied the euro corporate bond market using individual bond data over a substantial time period.
3.3
Methodology
We use the Brennan and Subrahmanyam (1996) approach to test whether liquidity risk is priced in the euro-denominated corporate bond market. Therefore, we rst describe their framework, followed by our modications and our implementation. Next, the Goldreich et al. (2002) method to compare dierent liquidity measures is presented. Finally, our liquidity measures are discussed.
3.3.1
Brennan-Subrahmanyam Approach
Brennan and Subrahmanyam (1996) used two variables to construct their portfolios. First, the stocks were sorted on their market capitalization and divided into size quintiles. Next, within each size quintile, the stocks were sorted on their estimated Kyle (1985) liquidity measure of market depth and assigned to one of ve portfolios. In this way, Brennan and Subrahmanyam (1996) evenly divided the total sample of stocks into 25 portfolios depending on the rms size and liquidity. Finally, they tested whether the constructed portfolios had signicantly dierent returns. To control for other sources of risk, Brennan and Subrahmanyam (1996) used the Fama and French (1993) equity-market model.
Section 3.3 Methodology
49
Fama and French (1993) found three common factors that explained the return of an equity portfolio: an overall market factor, a size factor and a book-to-market factor. To test their model, Fama and French (1993) subdivided the stocks in their sample in portfolios based on several other criteria, and regressed the portfolios excess returns over the default-free interest rate on the three factors. The intercept coecients from these regressions were almost never statistically signicant, indicating the validity of their model on their data set. Similarly, Brennan and Subrahmanyam (1996) regressed the realized excess returns of their portfolios on the Fama-French equity market factors. For each portfolio, they ran a Fama-French regression model augmented with an intercept term. Under the null hypotheses that liquidity has no additional power in explaining excess equity returns, the intercepts of these regressions should not be jointly signicantly dierent from zero. This null hypothesis was also tested by an alternative regression model. Again the portfolios excess returns were used as dependent variables. However, the portfolio-specic intercepts were replaced by a common intercept for all portfolios, and a portfolio-specic liquidity variable was added to the regression equation. For both regression models, the null hypothesis had to be rejected, implying that liquidity was an additional source of risk, beyond the Fama-French risk factors, that was priced by nancial market participants.
3.3.2
Modications
We use the Brennan and Subrahmanyam (1996) methodology to test whether liquidity is priced in the euro-denominated corporate bond market. We make three modications to their method. First, we use a bonds yield-to-maturity instead of its realized return to proxy for its expected return. Amihud and Mendelson (1991) provided the following argument why the yield of a bond should contain a compensation for liquidity. Suppose an investor wants to buy an illiquid bond. Since the bond is illiquid, he faces higher transaction costs when he wants to unwind his position before the bonds maturity, due to a larger bid-ask spread and/or order processing costs, than for a comparable, liquid bond. In order to persuade him to buy the illiquid bond, the investor should be compensated by a lower price. The yield of the illiquid bond should thus be higher than that of the liquid bond. In other words, the investor is willing to accept a lower yield on a liquid bond, because of the option to liquidate his position before maturity at lower costs. The advantage of yields over realized returns is twofold. First, yields really represent the markets expectation of a bonds expected return to maturity; realized returns, on the other hand, are, by denition, realizations of a stochastic process rather than expectations.
50
The second advantage of yields over realized returns can be understood by considering the hypothetical case where the prices of a liquid and an illiquid bond always have a xed ratio to each other; the realized returns of these bonds will be exactly equal, but their yields will dier. So, to the extent that there is a xed percentage price discount for illiquidity, realized returns are unsuited to determine whether liquidity is a determinant of security prices. The second modication to the Brennan-Subrahmanyam framework is that we replace the three-factor Fama and French (1993) equity-market model by their two-factor bondmarket model, augmented with bond characteristics as recommended by Gebhardt et al. (2001). Fama and French (1993) found two common risk factors in the returns on bonds. These two bond-market factors explained the excess returns of seven bond portfolios: two government bond portfolios with average maturities of 1 to 5 years and 6 to 10 years, and ve corporate bond portfolios with average, Moodys ratings of Aaa, Aa, A, Baa and below Baa. The excess return was dened as the portfolio return minus the one-month Treasury rate. Fama and French (1993) dened the rst bond-market factor as the dierence between the long-term government bond return and the one-month Treasury rate at the end of the previous month. Thus, this slope factor should explain variations in excess bond returns by changes in the slope of the Treasury yield curve. The second factor was dened as the dierence between the long-term corporate bond return and the long-term Treasury return. This credit factor was therefore related to the likelihood of credit events in the corporate bond portfolio. These two factors did a very good job in explaining the excess returns of the seven bond portfolios (Fama and French, 1993, Table 3): all estimated parameters were statistically signicant with t-values ranging from 8 to 140 and the R2 -values were also high, with values ranging from 79% to 98%. Moreover, the intercepts of all ve corporate bond portfolios were statistically not signicant. In their analysis, Fama and French (1993) used the government curve to calculate the excess returns and the two bond-market factors. In contrast, we use the swap curve. Recently, xed-income investors have moved away from using government securities to extract default-free interest rates and started using interest rate swap rates instead. Golub and Tilman (2000) and Koci c, Quintos and Yared (2000) mentioned the diminishing amounts of US and European government debts, the credit and liquidity crises of 1998, and the introduction of the euro in 1999 as primary catalyzing factors for this development. The results in Chapter 4, which studies the same market over the same time period as the current chapter, indicate that the swap curve is a better proxy for the default-free curve than its government counterpart.
51
Gebhardt et al. (2001) looked at the validity of the Fama-French bond market model by analyzing whether individual bond characteristics could rival the two Fama-French factors. Three characteristics were considered: rating, duration and Altman (1968) Z-scores. Using bivariately sorted portfolios and multivariate regressions, Gebhardt et al. (2001) concluded that both factor loadings and characteristics were important in explaining bond yields and recommended a model containing four variables: the Fama-French slope and credit factors, rating and duration. In Section 3.5.2, we show that for our data set three characteristics are relevant: rating, maturity and an indicator variable that equals 1 if a bond is euro-denominated, and 0 otherwise. Therefore, our null model consists of ve variables: two Fama-French factors and three characteristics. Clearly, all our conclusions about the relation between liquidity and bond yields are based on the assumption that our ve-variable pricing model is well-specied; see also Dimson and Hanke (2001). The third dierence with the Brennan-Subrahmanyam framework is that we do not use the Kyle (1985) direct liquidity measure. Instead, we consider eight indirect liquidity measures, which are detailed in Section 3.3.5.
3.3.3
Implementation
For each liquidity measure, we construct P time series of portfolio yields as follows (the choice for P will be discussed shortly). Every two weeks, we order all bonds in the sample by the value of that liquidity measure; only bonds that have already been issued and have not yet matured or been redeemed on that date are used in the ordering. Then, 100/P % of the bonds gets assigned to each of the P portfolios. We do this in such a way that for each liquidity measure portfolio 1 contains the most liquid bonds and portfolio P the most illiquid bonds. Every day we calculate the portfolio yield as the unweighted average of the yields of the bonds that make up the portfolio. The bond yield is calculated as follows: if a bond is not quoted, we disregard it for that day; if it is quoted by one pricing source, we use that yield; if it is quoted by more than one pricing source, we use the average quote. Like Brennan and Subrahmanyam (1996), we consider two regression models. For measure i, model 1 is
2 i Ypt 3 i jp Fjt j =1
i p
+
j =1
i i j Cjpt + i pt ,
E [i pt ] = 0
i i E [i pt qs ] = pq , if t = s, and 0 otherwise,
(3.1)
52
i is the excess yield of the pth measure-i portfolio on day t, and F1t and F2t are the where Ypt i i i two Fama-French factors and C1 pt , C2pt and C3pt are the three portfolio characteristics.
The disturbance terms are allowed to be heteroscedastically distributed and to be crosssectionally correlated, but we do assume that they are uncorrelated across time. For meai i i i i i i i i sure i, we estimate all 3P + 3 coecients (1 , . . . , P , 11 , . . . , 1 P , 21 , . . . , 2P , 1 , 2 , 3 )
for all P portfolios simultaneously with Feasible Generalized Least Squares (FGLS) as a system of seemingly unrelated regressions (SUR); see e.g. Greene (2000, Chapter 15). To correct for possible autocorrelations in the disturbances, we apply the Newey and West (1987) estimator for the covariance matrix. To test the null hypothesis that liquidity has no additional power in explaining bond yields beyond the two Fama-French factors and the three portfolio characteristics, we use a Wald test to determine the joint signii i = 0 . . . P = 0. The test statistic is asymptotically cance of the intercepts: H0 : 1
2 -distributed with P degrees of freedom. Further, to test the null hypothesis that all
i i intercepts are equal, we use a Wald test for H0 : 1 = . . . = P . This test statistic is
asymptotically 2 -distributed with P 1 degrees of freedom. Regression model 2 reads

= +
+
j =1
i i i j Cjpt + i Li pt + pt ,
(3.2)
where the assumptions on the disturbances are equal to those in (3.1) and Li pt is the value of the liquidity measure of the pth measure-i portfolio on day t in deviation from its daily average; so, if li denotes the value of the liquidity measure, and li is its daily average, i.e.
pt t
1 i lt = P then Li pt is calculated as
P i lpt , p=1
i i Li pt = lpt lt .
In Equation (3.2), the portfolio-specic intercepts of Equation (3.1) have been replaced by a single intercept and an additional regressor has been introduced that contains a proxy for
i in model 1 has been portfolio ps liquidity. Therefore, the constant liquidity premium of p i replaced by a time-varying premium i + i Li pt , where gives the eect of the deviation
of a liquidity measure from its mean. Here, the null hypothesis of no liquidity premiums
53
is tested with a Wald test on the joint signicance of i and i , H0 : i = 0 i = 0. The test statistic is asymptotically 2 -distributed with 2 degrees of freedom. For both model 1 and 2, there is the problem that if we want to test whether a particular measure i is a good liquidity measure that we are actually testing a joint hypothesis: illiquidity leads to yield increases and i is a proxy for liquidity; see also Kempf and Uhrig-Homburg (2000) and Jankowitsch et al. (2002). If we reject this joint hypothesis, then either i is a not a good liquidity measure or illiquidity does not lead to yield increases (or both). Given the strong empirical evidence mentioned in Section 3.2, we feel condent that a rejection of the joint hypothesis can in fact be traced to i being an inadequate liquidity measure. We now discuss the choice for the number of portfolios P for both models. For model 1, we create two portfolios for each liquidity measure. This gives an intuitive interpretation of portfolio 1 as the liquid portfolio and portfolio 2 as the illiquid portfolio. Moreover,
i i the dierence 2 1 between the two intercepts can be interpreted as the yield premium
investors get for bearing liquidity risk caused by measure i. In model 2, we have to estimate the slope coecient i , i.e. the relation between a portfolios value for liquidity measure i and its excess yield. Clearly, two portfolios would be insucient to estimate a slope. However, using too much portfolios diminishes the power of the Wald test; see Lys and Sabino (1992). From their Figure 1, it follows that if the portfolios contain approximately 25% of the bonds the power of the test of no relation between the liquidity measure and the excess yield is maximized. Therefore, we use 4 portfolios for model 2. Lo and MacKinlay (1990) showed that serious biases can arise in the test statistics when portfolios are used to test the ecient market hypothesis. The portfolios are constructed by sorting securities on empirical characteristics, which typically follow either from own research on a data base or from results of other papers that have analyzed the same data base. If the statistical tests are carried out on exactly the same data base, signicant biases in the tests can occur. Lo and MacKinlay (1990) called this data-snooping. In our analysis, data-snooping is probably limited, since our liquidity measures originate from two sources. First, some of our measures follow from theoretical models. Second, the remaining measures have been taken from empirical research on other bond data bases, so that no data mining has been applied to our data base.
3.3.4
Comparison
Given the large number of liquidity measures that have been proposed in the literature, a natural question to ask is if all measures are equally suited to proxy bond liquidity or if
54
some measures work better than others. One way to test this is to extend model 2 from the previous section with additional regressors for all other measures
2 i Ypt 3 i jp Fjt j =1 8 i i j Cjpt j =1
= +
Li pt
+
j =1,j =i
i ij Lij pt + pt ,
th where Lij measure-i portfolio in deviation of its pt is value of liquidity measure j for the p
daily average; the assumptions on the disturbances are equal to those in Equation (3.1). The problem with such an approach, however, is that the liquidity measures are strongly correlated. This may lead to multicollinearity problems between the regressors. Instead we follow Goldreich et al. (2002) by running a series of regressions with pairwise combinations of the liquidity measures. For each combination (i, k ) of measures, we estimate a regression like Equation (3.2) for measure i, augmented with measure k
= +
+
j =1
i i ik ik ik j Cjpt + i Li pt + Lpt + pt ,
(3.3)
where the disturbances behave as in Equation (3.1). In this regression equation2 , we test for the signicance of ik . If it is signicant, we say that k adds explanatory power to i, and otherwise we say that k is subsumed by i (this follows the terminology in Goldreich et al. (2002)). By repeating this procedure for all possible combinations, we can count the number of times a measure adds power to another measure, and the number of times a measure subsumes another measure. This allows us to rank the dierent liquidity measures we consider.
3.3.5
Liquidity Measures
Empirical papers that examined liquidity in bond or equity markets, used both direct and indirect measures of liquidity. Examples of direct liquidity measures are quoted bidask spreads, eective bid-ask spreads, quote sizes, trade sizes, quote frequencies, trade frequencies and trading volume. For corporate bonds, where most transactions occur on the over-the-counter market, these direct measures are often not reliable and dicult to obtain. Therefore, we use eight indirect liquidity measures from the empirical liquidity literature.
Goldreich et al. (2002) rst orthogonalized the values of measure k relative to measure i and used the orthogonalized values in Equation (3.3) instead of Lik pt . This is not necessary, since, by the Frisch-Waugh theorem (see e.g. Greene, 2000, Section 6.4.3)), the regression already automatically does this for us.
2
55
The measures issued amount, coupon and listed are static, as they are xed characteristics of a bond or its issuer. The age measure changes gradually over time. The other measures are dynamic and depend on market information; the measures missing prices and price volatility use daily price information, whilst the measures number of contributors and yield dispersion also consider quote composition information. Table 3.1 shows which papers used which measures and the eects they found; missing prices and yield dispersion are not mentioned in this table, because they were not used in previous studies. We will now discuss each measure in more detail. Issued Amount The issued amount of a bond is often assumed to give an indication of its liquidity. Most investment banks use it as liquidity criterion in building their bond indices; for example, Lehman Brothers uses this criterion for their Euro-Aggregate Corporate Bond index. Its use was rst proposed by Fisher (1959), who claimed that large issues should trade more often, so that the indirect measure issued amount is actually a proxy for the direct liquidity measure trading volume. Since Fisher, several alternative hypotheses have been put forward that also predict a positive eect of issued amount on liquidity (and thus on bond prices). In market microstructure models, like Smidt (1971) and Garman (1976), transaction costs arise, because dealers hold inventories. Further, dealers inventory costs are higher if it is more dicult to obtain information about a security and if the expected holding time is longer. Crabbe and Turner (1995) subsequently reasoned that large issues may have lower information costs, since more investors own them or have analyzed its features; similarly, information about small issues may be less broadly disseminated among investors. Therefore, small issues will have a higher yield due to an illiquidity premium. Another frequently heard argument, for instance in Sarig and Warga (1989) and Amihud and Mendelson (1991), is that bonds with smaller issued amounts tend to get locked in buy-and-hold portfolios more easily, reducing the tradeable amount and thus their liquidity. To summarize the above, we hypothesize a negative eect of issued amount on yields. Table 3.1 shows that many empirical papers considered issued amount as liquidity measure. The papers on Treasury bonds found negative and mostly signicant eects, so that larger Treasury issues have lower yields, as expected. Research on corporate bonds is inconclusive, though: both negative and positive coecients are observed. McGinty (2001) conrmed this by showing that even though most large issues in his corporate bond sample were liquid, some large issues were illiquid and some small issues were liquid.
56
Table 3.1: Overview of liquidity measures from the empirical bond liquidity literature.
Liquidity measures price coupon listed age volatility
Authorsa Data
issued amount
number of contributors
Corporate bonds AEF00 US CT95 US EGAM02 US ER01 US GM92 US + + HW00 US M01 EMU MR02 US + S01 US SBP93 US Treasury bonds AM91 US EG98 US F02 US JMP02 EMUb K02 US KU00c Germany SW89 US W92 US + Corporate & Treasury bonds DN02 Spain + Corporate, municipal & Treasury bonds CS99 US
a
+ + +
+ +
+ +
+ + + +
b c
Legend: negative; + positive; signicant; insignicant. AEF00=Alexander, Edwards and Ferri (2000), AM91=Amihud and Mendelson (1991), CS99=Chakravarty and Sarkar (1999), CT95=Crabbe and Turner (1995), DN02=D az and Navarro (2002), EG98=Elton and Green (1998), EGAM02=Elton, Gruber, Agrawal and Mann (2002), ER01=Ericsson and Renault (2001), F02=Fleming (2002), GM92=Gehr and Martell (1992), HW00=Hong and Warga (2000), JMP02=Jankowitsch, M osenbacher and Pichler (2002), K02=Krishnamurthy (2002), KU00=Kempf and Uhrig-Homburg (2000), M01=McGinty (2001), MR02=Mullineaux and Roten (2002), S01=Schultz (2001), SBP93=Shulman, Bayless and Price (1993), SW89=Sarig and Warga (1989), W92=Warga (1992). JMP02 considered 6 countries: Austria, France, Germany, Italy, Spain and The Netherlands. We used the price discounts in KU00s Table 2 to calculate the impact of maturity on yields.
57
Coupon Amihud and Mendelson (1991) argued that nancial institutions that are constrained to distribute only coupon income on their investments may prefer bonds with higher coupon percentages. This higher demand for high-coupon bonds implies lower yields. On the other hand, coupon is also frequently seen as a proxy for tax eects; see e.g. Shiller and Modigliani (1979): if coupon income is taxed, then bonds with higher coupons will have higher before-tax yields. In addition, lower-rated companies will typically issue higher-coupon bonds, so that higher coupons are again associated with higher yields. The predicted sign of the measure coupon is thus ambiguous. Few empirical papers employed the measure coupon ; see Table 3.1. Two papers found an insignicant, positive coecient, whereas one paper found a signicant negative eect. Listed Alexander et al. (2000) reasoned that companies whose equity is listed on a stock exchange must disclose more information than privately held companies. According to the market microstructure models mentioned above, the costs of making a market in bonds of listed rms should thus be smaller. Therefore, we hypothesize that the measure listed is associated with higher liquidity and lower yields. Since Alexander et al. (2000) were the only authors to use the liquidity measure listed, the empirical evidence is limited to their results. Contrary to their expectations, they found that issues of private rms trade more actively and thus are more liquid than issues of listed rms. Their explanation of this result was that for private rms debt is the only investment vehicle, while for public rms both debt and equity are traded; therefore, debt of private rms might trade more and have higher liquidity. Age The age of a bond is a popular measure of its liquidity. Sarig and Warga (1989) observed that as a bond gets older, an increasing percentage of its issued amount is absorbed in investors buy-and-hold portfolios. Thus, the older a bond gets, the less trading takes place, and the less liquid it becomes. Moreover, once a bond becomes illiquid, its stays illiquid until it matures. McGinty (2001) and Schultz (2001) also noted that new issues trade more than old issues. McGinty mentioned lead managers commitment to making market in the newly issued bond. Schultz pointed out that new issues are typically underpriced, so that traders buy bonds after the oering and sell them shortly thereafter. Following these arguments, we hypothesize a positive relation between age and yield.
58
Empirical research strongly conrms the positive eect of age on yields; see Table 3.1. This nding holds for corporate and sovereign bonds and for US and European data sets. Moreover, Schultz (2001) found evidence for the argument by Sarig and Warga (1989), since in his sample most bonds were bought and not sold; in other words, the bonds were put in buy-and-hold portfolios. Market practitioners often use a threshold value to determine if a bond is old or young: for some T , they mark all bonds with an age smaller than T as young and an age larger than T as old. Some academic papers also use such a dichotomous approach for the liquidity measure age. For instance, Alexander et al. (2000) set T = 2 years, Ericsson and Renault (2001) used T = 3 months, and Elton et al. (2002) employed a threshold value of 1 year. To determine which threshold values give useful divisions of bonds, we estimate model 1 from two portfolios, where portfolio 1 contains all bonds younger than T months and portfolio 2 older than T months, for T = 2, 4, . . . , 30. The dierence 2 1 between the portfolio intercepts, i.e. the liquidity premium between old and young bonds, and the signicance of the Wald test on H0 : 2 1 = 0 are displayed in Figure 3.1. Thresholds from 4 to 24 months give rise to a signicant liquidity premium, while the 2-month threshold and thresholds larger than 24 months do not. All thresholds, except for the smallest of 2 months, yield a signicant premium of at least 10 bps. The division between young and old bonds seems to be the strongest for a threshold of 14 months, where the premium equals 36 bps. For the remainder of this study, we arbitrarily use a threshold of 1 year for the measure age, although any other value between 4 months and 2 years could also be used. Missing Prices The occurrence of price runs and missing values is our rst liquidity measure that uses market information. Sarig and Warga (1989) argued that if the liquidity of a bond is suciently low, it may happen that on some business days there is virtually no trading in that bond. In their data set, this was recorded as a price run: two consecutive prices for a bond were identical. We extend their notion of illiquidity by considering not only the occurrence of a price run, but also the occurrence of a missing value, since in both cases there is no activity in that bond on that day. We will jointly refer to these events as the measure missing prices. We hypothesize a positive relation between missing prices and yield.
59
Figure 3.1: Liquidity premiums for dierent age thresholds.

40
35
30 liquidity premium (bps)
25
20
15
10
0 0 6 12 18 age threshold (months)
24
30
Solid ( ) and empty ( ) squares denote signicance and insignicance, respectively, of the Wald test on the joint signicance of the two intercepts (p-value < 0.05).
Price Volatility The measure price volatility is a measure of price uncertainty. In the market microstructure models discussed above, dealers inventory costs are higher if information uncertainty is higher. An important source of uncertainty is related to the predictability of future price movements. Therefore, we hypothesize that a higher price volatility leads to higher bid-ask spreads, and thus to lower liquidity and higher yields. The empirical evidence for price uncertainty as liquidity measure is mixed; see Table 3.1. Shulman et al. (1993) used price volatility as proxy for price uncertainty and found a signicantly positive eect on bond spreads. Hong and Warga (2000) proxied uncertainty with squared price return and estimated a positive and signicant coecient in a regression using bid-ask spread as dependent variable; this also implies a positive eect of uncertainty on bond yields. Alexander et al. (2000) approximated uncertainty as the average of absolute price returns; in their regressions, they found a signicant, positive eect on trading volume, implying a negative relation between uncertainty and yields.
60
Number of Contributors The number of contributors is our following measure of a bonds liquidity, and the rst that uses quote composition information. In Ericsson and Renault (2001), a larger number of active traders competing for the same bond leads to a smaller price discount for illiquidity and thus a smaller yield premium. Alternatively, Gehr and Martell (1992) and Jankowitsch et al. (2002) argued that a larger number of market participants makes it easier to trade a bond, because it is easier to nd a counterparty for a transaction and large orders can be split up into smaller parts without aecting the market price. Either way, we hypothesize a positive relation between the measure number of contributors and liquidity and therefore expect a negative eect of this measure on bond yields. Direct empirical evidence on the number of contributors liquidity measure is limited. Jankowitsch et al. (2002) found that bonds with more contributors have lower yields for all but one of the six European countries they analyzed. Indirect evidence is provided by Schultz (2001), who showed that there was a positive relation between the number of trades in a bond and the number of dealers as counterparties. Further, the results of Gehr and Martell (1992) showed a negative, though insignicant eect of the number of dealers on the bid-ask spread. Yield Dispersion Our nal liquidity measure, yield dispersion, reects the extent to which market participants agree on the value of a bond. Tychon and Vannetelbosch (2002) derived a model that predicts that if investors have more heterogeneous beliefs, the liquidity premium is larger. The inventory costs argument, mentioned above, applies here as well, since dealers face more uncertainty if prices show a larger diusion among contributors. Either way, we hypothesize a positive relation between scatter and bond yields. We proxy this notion of liquidity with a yield dispersion statistic, which has not been used before in the literature, as far as we know. We dene the yield dispersion of bond i on day t as the standard deviation of percentage yield dierences relative to the mean 1 nit 1
nit j =1
Dispersionit =
yitj y it y it
(3.4)
where yitj is the yield quoted by pricing source j , y it is the average yield and nit is the number of contributors. This measure can only be calculated if we have at least two quotes for a bond on a particular day, i.e. if nit > 1.
Section 3.4 Data
61
Application Table 3.2 gives details on the calculation of each liquidity measure. It also shows the
i expected sign of the measure. To get the lpt variable of Section 3.3.3, we multiply measures
with a negative expected sign by 1. This makes sure all i coecients in model 2 are expected to be positive. Finally, the table shows the order in which bonds are put in the portfolios: the rst portfolio always contains the bonds that are hypothesized to be most liquid, the last portfolio contains bonds that we expect to be most illiquid. As described in Section 3.3.3, every two weeks the portfolios for each liquidity measure are rebalanced according to each bonds value for that measure. For the measures issued amount, coupon, listed and age, we use the value of the liquidity measure on the rebalancing date. For the measures missing prices, number of contributors and yield dispersion, we use the average value over the two weeks prior the rebalancing date. For the measure price volatility, we calculate the standard deviation of the observed prices over the two weeks prior to the rebalancing date. If for a particular bond it is not possible to calculate the value of a liquidity measure on the rebalancing date, that bond is ignored for that measure until the next rebalancing date.
3.4
Data
The data are downloaded from three dierent sources. Lehman Brothers provides the International Securities Identication Numbers (ISINs) of the members of their EuroAggregate Corporate Bond index. The required characteristics of these corporate bonds are downloaded from Bloomberg. Reuters 3000 EXtra provides daily bid yields of each bond quoted by dierent pricing sources. The download period starts on 1 January 1999 and ends on 31 May 2001. The ISINs are obtained for 31 May 2000. The total number of bonds on this date equalled 1190. All bonds that are issued in euros directly after the currencys introduction are included in this analysis. Moreover, the yield time series of each corporate bond has at least twelve months history.
3.4.1
Lehman Brothers
Lehman Brothers provides the ISINs of the corporate bonds in their Euro-Aggregate Corporate Bond index. This index serves as a proxy of the investment-grade eurodenominated, corporate bond market. Lehman Brothers imposes a number of criteria before the corporate bonds can enter its index. All bonds must be denominated in euros or in one of the legacy currencies. Further, all bonds are investment grade, have a xed-
62
Table 3.2: Overview of liquidity measures, their expected signs and the portfolio order.
Liquidity measure
Details
Signa
Portfoliob rst last
Issued amount Coupon Listed Age Missing prices Price volatility Number of contributors Yield dispersion
total notional in billions of euros coupon rate 1 if a rms equity is publicly traded, 0 otherwise time between issue date and quote date in years 1 if price is missing or equal to previous price, 0 otherwise standard deviation of prices since previous rebalancing number of market participants quoting the bond see Equation (3.4)
? + + + +
largest smallest yes young least lowest largest smallest
smallest largest no old most highest smallest largest
Expected signs of the relationship between the measures and bond yields. Order in which the ranked bonds are assigned to the rst (most liquid) portfolio and the last (most illiquid) portfolio.
Section 3.4 Data
63
rate coupon, at least one-year to maturity and an issued amount of at least 150 million euro. The country of issuance and the country of the issuer are no index criteria. The credit ratings of all corporate bonds are also provided by Lehman Brothers. All ratings are downloaded for 31 May 2000. Due to data limitations, we have kept these ratings unchanged during the whole sample period. Finally, their Euro-Aggregate Corporate Bond BBB sub index is used to construct the Fama-French credit factor.
3.4.2
Bloomberg
Bloomberg provides the required bond characteristics. Using the ISINs that are given by Lehman Brothers these characteristics are downloaded. In case an ISIN code is not recognized by Bloomberg, the bond data are obtained from Lehman Brothers PC Product system. From the initial 1190 ISINs, three are not available in the Bloomberg data base. The downloaded corporate bond characteristics are: issued amount, issue date, maturity date, currency, call dates, put dates and sinking fund dates. Euro-denominated par swap data, which are used to calculate the two Fama-French factors and the portfolio excess yields, are also downloaded from Bloomberg.
3.4.3
Reuters
Reuters 3000 EXtra provides the bid yields of the selected corporate bonds. Most corporate bond yields in the Lehman Brothers Euro-Aggregate index are bid yields; only newly issued corporate bonds have ask yields during their rst month in the index; see Lehman Brothers (1999). Therefore, we download bid yields from Reuters. For each corporate bond, all pricing sources (also called contributors) are downloaded. We exclude two Reuters pricing sources, the clearing agency ISMA and two anonymous pricing sources from the list of contributors, since they are averages of other pricing sources. The total number of dierent pricing sources thus obtained equals 74. From the original 1190 ISINs in the Lehman Brothers Euro-Aggregate Corporate Bond index, 191 bonds cannot be analyzed, because they either have no Reuters Identication Code (RIC) that matches their ISIN or they do have a RIC but no contributor. For the remaining 999 bonds, all bid yields from all pricing sources are downloaded. This means that a number of time series, equal to the number of pricing sources, shows the yield development of each bond. Most bonds are quoted by more than one pricing source.
64
3.5
Results
We rst present the results of applying the Fama-French bond-market model to the entire sample and show the extension of this model with portfolio characteristics. Next, the regression results for models 1 and 2 are given. Finally, the performance of the liquidity measures is compared.
3.5.1
Entire Sample
To test whether the euro-dominated corporate bond market can, on average, be described by the two-factor Fama-French model, we rst run their model on a portfolio consisting of all bonds in our sample, i.e.
2
Yt = +
j =1
j Fjt + t , t i.i.d.(0, 2 ),
(3.5)
where the excess yield Yt is the average bond yield, calculated over all bonds in the sample, minus the one-year euro swap rate, the slope factor F1t is dened as the ten-year swap rate minus the one-year swap rate of the previous day and the credit factor F2t is calculated as the Lehman Brothers Euro-Aggregate Corporate Bond BBB sub index minus the ten-year euro swap rate. The rst row of Table 3.3 shows the R2 and the estimated coecients along with their t-values. The R2 value is high and comparable to the values reported by Fama and French (1993). The estimated slope and credit coecients have the expected positive sign and are strongly statistically signicant. The intercept is not statistically signicant, so that the Fama-French model cannot be rejected for the entire sample. To test our choice for approximating default-free interest rates with swap rates, regression model (3.5) is estimated again, but with the swap rates replaced by the government rates. So, the excess yields and the slope and credit factors are now calculated with government yields. Our proxy for euro government rates is the Lehman Brothers EuroAggregate Treasury index. The second row of Table 3.3 shows the regression results. Both the R2 and the t-values of the slope and credit factors have decreased compared to the model with swap rates. Moreover, the intercept is now signicantly dierent from zero. Therefore, the Fama-French model should be rejected in case government rates are used as default-free rates. This empirically conrms our choice for using swap rates as proxy for default-free interest rates instead of Treasury rates.
Section 3.5 Results
65
Table 3.3: Results for the entire sample

Intercept Slope Credit R2
Swap rates Government rates
0.0371 (1.01) 0.419 (12.4)
0.785 (36.5) 0.540 (31.4)
0.173 (6.66) 0.273 (5.46)
97.9% 95.0%
Regression results for the Fama-French model estimated from the entire sample with either swap rates or government rates as default-free interest rates (t-values between parentheses).
3.5.2
Characteristics
As recommended by Gebhardt et al. (2001), we analyze the added value of incorporating characteristics into the model. We consider three characteristics: Rating: rating of the bonds issuer at 31 May 2000: AAA, AA, A or BBB. Maturity: the remaining time to maturity of a bond, measured in years. Euro: whether a bond is denominated in euros or in one of the legacy currencies. This variable was not used in Gebhardt et al. (2001), who studied US data, but we nevertheless consider it to be relevant for our data set. The motivation is that the market generally sees legacy bonds as less liquid, because these bonds are relatively old, not well known to investors and more dicult to trade due to the legacy currency. To determine whether a characteristic is important for explaining excess bond yields, we follow the same procedure as for our liquidity measures, as described in Section 3.3.3, except that the null model is now the Fama-French model of the previous section. For each characteristic i, we create portfolios and estimate the following regression model
2 i Ypt
i p
+
j =1
i Fjt + i jp pt ,
where the assumptions on the disturbances are equal to those in Equation (3.1). For the characteristic rating, we create four portfolios: portfolio 1 contains the AAA-rated bonds, portfolio 2 the AAs, portfolio 3 the As, and portfolio 4 the BBBs. For the characteristic maturity, two portfolios are constructed: portfolio 1 consists of the 50% shortest bonds,
66
and portfolio 2 of the 50% longest bonds.3 Finally, for the characteristic euro, two portfolios are created as well: portfolio 1 contains the euro-denominated bonds and portfolio 2 the legacy bonds. The regression results are reported in Table 3.4. For rating, we nd that the intercepts are larger for lower ratings, although the step from AA to A is very small. The FamaFrench factor loadings are all signicant. The Wald test indicates that the four intercepts are jointly highly signicant. For maturity, the intercepts of the portfolios reveal that short-maturity bonds have smaller yields than long-maturity bonds. The null hypothesis that the two intercepts are jointly equal to zero is easily rejected. Finally, for euro, the results imply that euro-denominated bonds have smaller yields than legacy bonds, with an average spread of 21 bps between them. Again, the Wald statistic is signicantly dierent from zero. From these results, we conclude that the rating, maturity and euro characteristics are important determinants of excess yield in the euro corporate bond market. To make the characteristics operational, we have to transform them to a numerical scale: Rating: the letters are mapped as follows: AAA=1, AA=2, A=3 and BBB=4. Although this linearity assumption is somewhat crude, it is not uncommon in the literature. Moreover, since our bonds are all investment grade, and the non-linearities in S&Ps and Moodys rating scales are especially apparent for speculative grade ratings, we believe that the linear scale is a reasonable approximation; see also panel I of Table 4.1 in Chapter 4. Maturity: this is already a continuous variable, and thus needs no transformation. Euro: this characteristic is represented by an indicator variable that equals 1 if the bond is denominated in euros, and 0 if it is denominated in a legacy currency.
i The value of characteristic j for the pth measure-i portfolio on day t, denoted Cjpt in
Section 3.3.3, is calculated analogously to the liquidity variable Li pt below Equation (3.2). For instance, for the characteristic maturity, it is the average maturity of all quoted bonds in that portfolio on that day, in deviation from the average maturity of all quoted bonds on that day.
3.5.3
Model 1
For the rst regression model, Equation (3.1), we create two portfolios for all eight liquidity measures. Table 3.5 contains some summary statistics for these 16 portfolios, averaged
3
The portfolios are updated every two weeks, just like in Section 3.3.3
Section 3.5 Results
67
Table 3.4: Results for the characteristics portfolios.

Intercept Slope Credit Wald R2
Rating AAA AA A BBB Maturity short long Euro euro legacy
-0.220 (8.30) 0.120 (5.60) 0.122 (4.95) 0.453 (8.28) -0.135 (3.09) 0.247 (13.6) -0.066 (1.85) 0.139 (5.53)
0.736 (46.8) 0.732 (55.8) 0.856 (59.4) 0.824 (24.4) 0.635 (25.0) 0.944 (82.0) 0.888 (42.1) 0.682 (45.7)
0.0838 (4.40) 0.0310 (2.04) 0.295 (15.8) 0.435 (10.6) 0.165 (5.36) 0.138 (10.6) 0.291 (11.2) 0.0242 (1.35)
946 (0.00)
97.2%
474 (0.00)
98.8%
112 (0.00)
97.8%
Regression results for the Fama-French model estimated from portfolios based on the rating, maturity and euro characteristics (t-values between parentheses). The Wald column shows the test on the joint signicance of the intercepts (p-value between parentheses).
over the full sample period of 602 trading days. We observe that the average yields of portfolio 1 (containing the hypothesized liquid bonds) and portfolio 2 (illiquid bonds) are quite dierent. The deviations range from -28 bps (for age ) to 59 bps (for price volatility ). We also see that the average values of the liquidity measures dier substantially. Except for the measures price volatility and number of contributors, we could prematurely conclude that the liquidity premium is negative, since portfolio 1 has a higher average yield than portfolio 2. However, it is not correct to fully attribute the yield dierences to dierences in liquidity, since the average maturity and the average credit worthiness also strongly vary. Therefore, this table illustrates the necessity of using the Fama-French factors and the portfolio characteristics to correct for dierences in maturity and credit rating. Table 3.6 displays the results of estimating model 1 for all liquidity measures. All Fama-French factor loadings are statistically signicant and have the expected positive
68
Table 3.5: Portfolio statistics P = 2.

Yielda 1 2 Maturityb 1 2 Ratingc 1 2 Liquidityd 1 2
Issued amount Coupon Listed Age Missing prices Price volatility Number of contributors Yield dispersion
5.33 5.21 5.26 5.44 5.28 4.91 5.19 5.40
5.09 5.21 5.01 5.16 5.07 5.50 5.27 5.14
6.47 6.05 5.68 6.91 6.09 3.81 5.58 7.42
4.64 5.04 5.11 5.31 4.57 7.24 5.56 4.91
0.22 0.25 0.27 0.17 0.19 0.31 0.22 0.09
0.35 0.32 0.33 0.30 0.46 0.26 0.42 0.14
0.65 4.29 1.00 0.64 0.19 0.21 2.31 0.47
0.20 6.91 0.00 3.80 0.46 0.45 0.76 1.50
a b c
Summary statistics of the two constructed portfolios using the eight liquidity indicators. Portfolio 1 (respectively 2) contains the bonds that are hypothesized to be most liquid (respectively most illiquid). Average portfolio yield. Average time to maturity in years. Average credit worthiness, measured on the following scale: AAA=1, AA=2, A=3, BBB=4. Average value of the liquidity measure.
sign. The same holds for the coecients of the rating and maturity characteristics (with one exception: the rating coecient for issued amount is insignicant). The coecient for the euro characteristic is mostly insignicant, though it does have the expected negative sign for seven out of eight cases. All R2 -values are around 98%. Except for the liquidity measure listed, all intercept pairs are jointly statistically different from zero at a 95% signicance level, as evidenced by the p-values of the Wald statistics. This indicates that the remaining seven measures are indeed able to separate the bonds in our data set into two mutually exclusive portfolios that have statistically dierent yields, after controlling for dierences in maturity, rating and currency. Next we look at the portfolio intercepts themselves. If our hypotheses on the sign of the liquidity eects are correct, the intercept of portfolio 1 should be smaller than that of portfolio 2 for all liquidity measures. We see that this holds for six out of eight cases; for listed and price volatility the order is reversed. For listed, this poses no problem, since the Wald test already indicated that for this measure the null model cannot be rejected. For price volatility, the intercepts are inconsistent with our expectations, since the low-volatility portfolio has a higher yield than the high-volatility portfolio. Further, for the measure coupon, the results show that low coupon bonds have lower yields than high coupon bonds;
Section 3.5 Results
69
this contradicts the liquidity hypothesis of Amihud and Mendelson (1991), but instead supports the alternative interpretation of coupon as tax and/or credit risk proxy.
i i Another way of looking at the intercepts, is to calculate their dierences 2 1 ,
which we interpret as the liquidity premium for measure i. The second to last column of Table 3.6 shows that the premiums for measures yield dispersion and age are the largest with 24 bps and 20 bps, respectively, while the premiums for the other measures are between 9 and 13 bps. All premiums are statistically signicant at the 95% condence level.
3.5.4
Model 2
For model 2, we create four portfolios since it maximizes the power of the test for the presence of liquidity eects; see Section 3.3.3. Unfortunately, this means we cannot conduct the test for measures listed and age, since they are both binary variables (listed versus not listed, and young versus old). The summary statistics for the other six measures are shown in Table 3.7. Clearly, the dierences between the portfolios are now larger than in Table 3.5, since we have assigned the bonds to four size percentiles instead of two. The regression results are displayed in Table 3.8.4 The Wald statistic that tests for the joint signicance of the intercept and the coecient of the liquidity measure is statistically signicant for all six measures. So, also using model 2, we nd statistical evidence of the presence of liquidity eects in our data set, after controlling for maturity, rating and currency dierences between the portfolios. The signs of all liquidity coecients are positive, as hypothesized.
3.5.5
Comparison
Table 3.9 summarizes the results of conducting the pairwise comparisons between the liquidity measures, as described in Section 3.3.4. For each measure i, we count the number of times it adds power to a model that already contains measure j . We also count the number of times a measure j is subsumed if it is added to the model of measure i. Looking at the sum of both counts, we see that measures number of contributors and especially price volatility outperform the other four measures.
The Fama-French factor loadings and the coecients for the portfolio characteristic are omitted from Table 3.8 for space considerations.
70
Table 3.6: Results for model 1.
Intercept
Factors Slope Credit Walda Premiumb R2
Characteristics Rating Maturity Euro
0.866 (51.5) 0.712 (36.9) 0.786 (51.9) 0.783 (37.2) 0.783 (45.0) 0.748 (42.4) 0.912 (53.6) 0.741 (36.4) 0.252 (11.8) 0.180 (7.40) 0.224 (5.51) 0.115 (11.7) -0.0520 (0.909) 0.165 (7.95) 0.143 (6.71) 0.143 (2.37) 0.109 (7.64) -0.0581 (0.667) 1.72 (0.42) 0.174 (9.54) 0.179 (7.13) 0.304 (6.62) 0.121 (8.04) -0.0348 (0.515) 21.7 (0.00) 11.0 (0.00)
0.211 (10.4) 0.130 (5.53)
0.0720 (1.56)
0.121 (8.24)
-0.263 (2.84)
6.90 (0.03)
12.4 (0.01)
98.0%
97.7%
-4.46 (0.19)
98.0%
Issued amount large -0.0320 (0.880) small 0.0924 (2.30) Coupon small -0.0187 (0.671) large 0.0918 (2.67) Listed yes 0.0153 (0.476) no -0.0293 (0.842) Age <1y -0.0365 (1.35) >1y 0.159 (5.25) 54.9 (0.00) 19.6 (0.00) 98.1%
continued on next page
Table 3.6: continued
Section 3.5 Results
Intercept
Factors Slope Credit Walda Premiumb R2
Characteristics Rating Maturity Euro
0.148 (6.95) 0.177 (6.78) 0.168 (6.34) 0.157 (8.45) 0.164 (7.38) 0.108 (4.82) 0.169 (8.18) 0.113 (5.18) 0.323 (8.32) 0.0925 (16.7) -0.107 (1.52) 0.205 (6.89) 0.140 (26.9) 0.0169 (0.608) 52.5 (0.00) 0.314 (6.42) 0.0965 (22.2) -0.169 (2.22) 24.8 (0.00) -13.3 (0.00) 98.5%
0.189 (5.47)
0.164 (23.1)
-0.0351 (0.890)
27.8 (0.00)
12.8 (0.00)
97.6%
9.15 (0.00)
97.6%
Missing prices few -0.0315 0.798 (1.01) (44.9) many 0.0969 0.732 (2.85) (35.6) Price volatility small 0.121 0.630 (3.32) (28.6) large -0.0123 0.903 (0.454) (54.9) Number of contributors large 0.0165 0.786 (0.544) (44.2) small 0.108 0.747 (5.02) (40.1) Yield dispersion small -0.0331 0.893 (1.13) (50.7) large 0.208 0.698 (7.03) (36.9) 64.4 (0.00) 24.1 (0.00) 98.4%
Regression results for the Fama-French model augmented with portfolio characteristics (see Equation (3.1)) estimated from two portfolios based on one of the eight liquidity measures (t-values between parentheses). Test on the joint signicance of the intercepts (p-value between parentheses). Dierence between the portfolio intercepts in basis points (p-value between parentheses). 71
72
Table 3.7: Portfolio statistics P = 4.
Yielda 2 3 4 1 4 1 4 1
Maturityb 2 3
Ratingc 2 3
Liquidityd 2 3
Issued amount Coupon Missing prices Price volatility Number of contributors Yield dispersion
5.38 5.16 5.35 4.77 5.27 5.43
5.28 5.27 5.21 5.02 5.10 5.37
5.08 5.31 5.07 5.38 5.25 5.19
5.10 5.11 5.12 5.63 5.32 5.07
6.93 6.31 6.53 2.92 6.24 7.94
5.93 5.78 5.63 4.55 4.92 6.88
4.93 6.03 4.57 6.65 5.59 5.44
4.34 4.05 4.67 7.88 5.41 4.30
2.17 2.07 2.21 1.95 2.18 2.16
2.23 2.22 2.14 2.11 2.08 2.20
1.98 2.11 2.08 2.23 2.15 2.14
2.21 2.20 2.22 2.28 2.42 2.12
0.88 3.48 0.05 0.15 3.48 0.35
0.38 5.17 0.34 0.25 1.13 0.60
0.24 5.99 0.45 0.37 0.80 0.92
0.16 7.85 0.53 0.54 0.57 2.15
Summary statistics of the four constructed portfolios using six liquidity indicators. Portfolio 1 (respectively 4) contains the bonds that are hypothesized to be most liquid (respectively most illiquid). Average portfolio yield. Average time to maturity in years. Average credit worthiness, measured on the following scale: AAA=1, AA=2, A=3, BBB=4. Average value of the liquidity measure.
Section 3.5 Results
73
Table 3.8: Results for model 2.

Intercept Liquidity Walda R2
0.0316 (1.32) 0.0405 (1.71) 0.0894 (3.23) 0.0637 (2.75) 0.102 (3.99) 0.0855 (4.26)
0.338 (11.9) 0.0569 (8.07) 0.357 (7.92) 0.0295 (0.922) 0.00100 (0.173) 0.0335 (3.15)
147 (0.00) 65.3 (0.00) 66.2 (0.00) 8.22 (0.02) 15.9 (0.00) 28.9 (0.00)
97.9% 97.6% 96.3% 98.3% 96.2% 97.8%
Regression results for the Fama-French model augmented with portfolio characteristics and a liquidity variable (see Equation (3.2)) estimated from four portfolios based on one of six liquidity measures (t-values between parentheses). The coecients and t-values of the Fama-French factors and the characteristics are omitted for space considerations. Test on the joint signicance of the intercept and the coecient of the liquidity variable (p-value between parentheses).
Table 3.9: Results of the comparison tests.

Adds power Subsumes Total
3 3 2 4 4 3
1 1 1 4 2 1
4 4 3 8 6 4
Results of the pairwise comparisons of six liquidity measures (see Equation (3.3)). The table displays the number of times a measure adds explanatory power to another measure and the number of times a measure subsumes another measure.
74
3.6
Summary
In this chapter, we used the Brennan and Subrahmanyam (1996) methodology to test which measures can be used to approximate liquidity in the euro-denominated corporate bond market. Eight indirect measures of liquidity were implemented: issued amount, coupon, listed, age, missing prices, price volatility, number of contributors and yield dispersion. For each liquidity measure, we constructed mutually exclusive portfolios. The time series of portfolio yields were subsequently used in two Fama and French (1993) regression models, augmented with portfolio characteristics as recommended by Gebhardt et al. (2001), to control for dierences in maturity, credit worthiness and currency. We also conducted pairwise comparisons between the liquidity measures, as proposed by Goldreich et al. (2002). The results indicated that the augmented Fama-French model should be rejected for seven out of eight liquidity measures. The premium between liquid and illiquid portfolios depended on the liquidity measure and ranged from 9 to 24 basis points. The highest premiums were found for the measures age and yield dispersion. The comparison tests showed that the measures price volatility and number of contributors most often had added value over other measures or that other measures were subsumed by them.
Chapter 4
Pricing Credit Default Swaps1
4.1
Introduction
During the last decade, credit derivatives have become important instruments to lay o or take on credit risk. Until today only very limited empirical research has been devoted to these new instruments, although several reduced form models have been developed to price them. Most empirical papers on credit risk modelling have focussed on defaultable bonds. In this chapter, we estimate reduced form models and compare model-implied credit default swap premiums to market data. We show that using a reduced form model outperforms directly using bonds yield spreads to approximate default swap premiums. Moreover, we shed light on the choice of the default-free term structure of interest rates. We nd that swap and repo curves signicantly outperform the government curve as proxy for default-free interest rates for investment grade issuers, but that their performance is similar for speculative grade issuers. As such, this is one of the rst studies to empirically conrm that nancial markets no longer see Treasury bonds as the default-free benchmark. A default swap protects its buyer from losses caused by the occurrence of a default event to a corporate or sovereign debt issuer. In exchange for this default protection, the buyer pays a periodic premium to the protection seller. The no-arbitrage value of the default swap premium can be derived by applying a reduced form credit risk model. In these models prices of default-sensitive instruments are jointly determined by the risk-neutral default probability and the recovered amount at default. Default is often represented by a random stopping time with a stochastic or deterministic arrival intensity (hazard rate), while the recovery rate is oftentimes assumed to be constant. We have a large data set of
1
This chapter is a somewhat revised version of the paper by Houweling and Vorst (2002).
76
Pricing Credit Default Swaps Chapter 4
market quotes on credit default swaps at our disposal, allowing us to conduct empirical testing of reduced form models, which the literature has lacked so far. To the best of our knowledge, the only other study that analyses credit default swap data is Aunon-Nerin, Cossin, Hricko and Huang (2002). Their primary analyses are regressions of default swap premiums on proxies for credit risk, whereas we estimate and apply a reduced form credit risk model. Aunon-Nerin et al. (2002) also implement the Das and Sundaram (2000) tree model, but on just 75 observations. As a rst indication, the default swap premium is often estimated by the yield spread of a bond with a similar maturity issued by the same borrower. We show analytically that this relationship only holds approximately. Moreover, we show empirically that the approximation results in fairly large deviations between calculated and market default swap premiums. By deriving the risk-neutral pricing formula for a defaultable coupon-bearing bond, we can explicitly express its dependence on risk-neutral processes for default-free interest rates, hazard rates and and the recovery rate. Since we focus on the estimation and application of credit risk models, we use a priori estimated default-free term structures. The choice for default-free interest rates has received little attention in the literature. Virtually all empirical papers on credit risk modelling used zero-coupon rates extracted from government bonds. However, since 1998, nancial markets have moved away from estimating default-free interest rates from government securities, and started using swap and repo contracts instead. We nd that using the government curve results in statistically signicant overestimation of credit risk for investment grade issuers, that swap curves result in a small but signicant bias, and that repo curves estimate credit risk unbiasedly. We also nd that the government curve is signicantly outperformed by both the swap and repo curves, and that repo curve models perform signicantly better than, or not signicantly dierent from swap curve models. For speculative grade issuers, the choice for the default-free curve is less important, as the performance dierences between the three curves are smaller. We also pay attention to the choice of the recovery rate. Since it is not possible to extract both the hazard rate and the recovery rate from prices of bonds of a single seniority class, we x the recovery rate to identify the model. We show that not only bond spreads, but also default swap premiums are relatively insensitive to changes in the recovery rate as long as the hazard function is scaled accordingly. Therefore, there is no need to determine the recovery rate very accurately, as long as it takes a reasonable value. We model the hazard function as a constant, linear or quadratic function of time to maturity. The parameters of the hazard function are estimated using non-linear least squares from market prices of bonds of a single issuer. The estimated credit model is
Section 4.2 Default Swaps
77
subsequently applied to the pricing of credit default swaps written on the same issuer. We observe that both the in-sample t to bonds and the out-of-sample t to default swaps declines with an issuers credit quality. We also nd that using the various hazard rate functions yield more accurate estimates of default swap premiums than directly using the yield spread of a similar bond. An analysis of the deviations between calculated and market premiums reveals that the deviations for all models are related to the maturity of the default swap and the rating of the underlying issuer. The remainder of this chapter is structured as follows. Section 4.2 discusses the characteristics of credit default swaps. The literature on reduced form credit risk modelling is reviewed in Section 4.3. In Section 4.4, we derive reduced form valuation models for bonds and default swaps, and present our estimation framework. The construction of our data set is outlined in Section 4.5. In Section 4.6 we present the results of applying the direct comparison methods and the reduced form models. Finally, Section 4.7 concludes the chapter.
4.2
Default Swaps
Default swaps are the most popular type of credit derivatives: according to the latest Credit Derivatives Survey by the British Bankers Association (BBA, 2002) they account for 45% of the global credit derivatives market and according to the most recent Credit Derivatives Survey by Risk Magazine (Patel, 2003) for even 73%. A default swap is a contract that protects the holder of an underlying obligation from the losses caused by the occurrence of a credit event to the obligations issuer, referred to as the reference entity. Credit events that trigger a default swap can include one or more of the following: bankruptcy, failure to make a principal or interest payment, obligation acceleration, obligation default, repudiation/moratorium (for sovereign borrowers) and restructuring; these events are jointly referred to as default. A default swap only pays out if the reference entity defaults; reductions in value unaccompanied by default do not compensate the buyer in any way. Also, the event of default must be veriable by publicly available information or an independent auditor. The buyer2 either pays an upfront amount or makes periodic payments to the seller, typically a specied percentage of the notional amount. In the latter case, the percentage that gives the contract zero value at initiation is called the spread, premium or xed rate. If default occurs, the default swap can be settled in one of two ways. With a cash settlement, the buyer keeps the underlying
The party purchasing credit protection is called the buyer ; similarly, the seller refers to the party providing (selling) protection.
2
78
asset(s), but is compensated by the seller for the loss incurred by the credit event. In a physical settlement procedure, the buyer delivers the reference obligation(s) to the seller, and in return, he receives the full notional amount. Either way, the value of the buyers portfolio is restored to the initial notional amount.3 Several features of default swaps are worth mentioning. If the contract species periodic payments, and default occurs, the buyer is typically required to pay the part of the premium payment that has accrued since the last payment date; this is called the accrual payment. The credit event may apply to a single reference obligation, but more commonly the event refers to any one of a much broader class of debt securities, including bonds and loans. Similarly, the delivery of obligations in case of physical settlement can be restricted to a specic instrument, though usually the buyer may choose from a list of qualifying obligations, irrespective of currency and maturity as long as they rank pari passu with (have the same seniority as) the reference obligation. This latter feature is commonly referred to as the delivery option. Theoretically, all deliverable obligations should have the same price at default and the delivery option would be worthless. However, in some credit events, for instance a debt restructuring, not all obligations become immediately due and payable, so that after such an event bonds with dierent characteristics will trade at dierent prices. This is favorable to the buyer, since he can choose the cheapest bonds for delivery to the seller. Counterparties can limit the value of the delivery option by restricting the range of deliverable obligations, e.g. to non-contingent, interest-paying bonds. Counterparty risk is generally not taken into account in determining deal prices; if a party is unwilling to take on credit risk to its counterparty, it either decides to cancel the trade or to alleviate the exposure, for instance by demanding that a collateral is provided or that the premium is paid up-front instead of periodically (Culp and Neves (1998) and OKane and McAdie (2001)). An important application of credit default swaps is shorting credit risk. The lack of a market for repurchase agreements (repos ) for most corporates makes shorting bonds unfeasible. So, credit derivatives are the only viable way to go short corporate credit risk. Even if a bond can be shorted on repo, investors can only do so for relative short periods of time (one day to one year), exposing them to changes in the repo rate. On the other hand, default swaps allow investors to go short credit risk at a known cost for long time spans: default swaps with maturities of up to 10 years can be easily contracted, but liquidity rapidly decreases for even longer terms.
In a so-called binary default swap the seller pays the buyer a pre-specied amount, independent of the realized loss in case of default.
3
Section 4.3 Literature
79
4.3
Literature
In Chapter 1, the literature on credit risk modelling was reviewed, including dierent types of reduced form models. Here, the empirical literature on reduced form models is discussed. That literature has focused on estimating the parameters of one of three processes: the hazard process, the spread process or the risky short rate process. The rst approach seems to be most popular. Cumby and Evans (1997) considered both cross-sectional estimation of a constant hazard rate model and time-series estimation of several stochastic specications. Madan and Unal (1998) estimated recovery and hazard processes in a two-step procedure using Maximum Likelihood (ML) and Generalized Methods of Methods (GMM). Duee (1998), Keswani (2000) and Driessen (2001) applied ML with Kalman ltering to obtain parameter estimates of Cox, Ingersoll and Ross (1985, CIR) processes from time-series data. Bakshi, Madan and Zhang (2001), Fr uhwirth and S ogner (2001) and Janosi, Jarrow and Yildirim (2002) used non-linear least squares to estimate the model parameters from cross-sectional data. Janosi et al. specied a stochastic hazard rate that depends on the default-free short rate and an equity market index; Bakshi et al. estimated a model with correlated interest rates, hazard rates and recovery rates; Fr uhwirth and S ogner used a constant hazard rate. The second approach encountered in the empirical literature refrains from modelling the default and/or recovery components of credit risk and directly estimates the spread process instead. Nielsen and Ronn (1998) estimated a log-normal spread model using nonlinear least squares from cross-sectional data. Taur en (1999) utilized GMM to estimate the spread dynamics as a Chan, Karolyi, Longsta and Sanders (1992) process. D ulmann and Windfuhr (2000) and Geyer, Kossmeier and Pichler (2001) implemented a ML procedure with Kalman ltering to obtain parameter estimates of Vasicek (1977) and/or CIR models for the instantaneous spread. Due, Pedersen and Singleton (2003) used an approximate Maximum Likelihood method to estimate a multi-factor model with Vasicek and CIR processes. The third approach is to consider the sum of the default-free rate and the spread and estimate a model for the total risky rate. Due and Singleton (1997) utilized this approach to estimate the swap rate as a 2-factor CIR process using Maximum Likelihood. All discussed papers gauged the quality of the implemented models on their ability to t spreads and/or bond prices. Since credit derivatives allow credit risk to be traded separately from other sources of risk, they provide a clean way of putting a price on credit risk. Therefore, we may obtain better insights in the performance of credit risk models by applying them to the pricing of credit derivatives.
80
4.4
Methodology
In this section, we rst discuss the valuation of bonds and credit default swaps in our reduced form credit risk model. Then, we elaborate on the specication and estimation of the model.
4.4.1
Valuing Bonds
Following Jarrow and Turnbull (1995), we assume a perfect and arbitrage-free capital market, in which default-free and defaultable zero-coupon bonds, a default-free moneymarket account and defaultable coupon bonds are traded. The uncertainty is represented by a ltered probability space (, F , Q), where denotes the state space, F is a -algebra of measurable events in and Q is the actual probability measure. The information structure is represented by the ltration F (t). We take as given some non-negative, bounded and predictable default-free short-rate process r(t), which drives the default-free denote the equivalent martingale measure that is money-market account B (t). Let Q is the associated with the numeraire B (t); see Harrison and Pliska (1981). That is, Q risk-neutral measure. Let p(t, T ) and v (t, T ) denote the time-t values of a default-free and a defaultable zero-coupon bond with maturity T and face value 1, respectively. Default .4 Let P (t, T ) denote the riskoccurs at a random time , independent of r(t) under Q (t, T ) = E t [1{ >T } ] with E t [X ] = EQ neutral survival probability, i.e. P [X |Ft ] and 1{A} the indicator function of event A. We assume the existence of a non-negative, bounded and predictable process (t), which represents the default intensity or hazard rate for . Intuitively, (t)t is the risk-neutral probability of default between times t and under Q t + t as seen at time t, conditional on no earlier defaults. Then, (t, T ) = E t exp P
t T
(s)ds
t [exp((t, T ))] , =E
T t
(4.1)
where (t, T ) denotes the integrated hazard function : (t, T ) =
(s)ds.
Now, consider a defaultable coupon bond with coupon payment dates t = (t1 , . . . , tn ), coupon payment c, maturity tn and notional 1. The price v (t, t, c) of this bond at time t equals the sum of the expected values of its coupons and face value and a potential recovery payment. The ith coupon payment is only made if the bonds issuer has not gone
In our empirical application of the model we use readily available default-free term structures instead of specifying a risk-neutral process for r(t) and estimating its parameters. Therefore, we cannot estimate the correlation between default-free interest rates and the default time, so there is no use in allowing for a correlation parameter in our model.
4
81
bankrupt yet at time ti . Similarly, the face value is only fully paid if the bond is still alive at time tn . If the bond does default before it matures, a constant5 recovery fraction of the notional (and not of the remaining coupons too, see Jarrow and Turnbull (2000) and Sch onbucher (2000)) is made at the random default time . Applying the risk-neutral valuation principle to all cash ows, yields (cf. Due and Singleton, 1997, Equation (26))
n
v (t, t, c) =
i=1 n
t c1{ >t } + p(t, tn )E t 1{ >tn } + E t p(t, ) 1{ tn } p(t, ti )E i (t, ti ) + p(t, tn )P (t, tn ) + p(t, ti )cP
i=1 t tn
(4.2) p(t, s)f (s)ds,
where f (t) denotes the probability density function associated with the intensity process (t). In our empirical application, we replace the integral expression in Equation (4.2) by a numerical approximation.6 To do so, we dene a grid of maturities s0 , . . . , sm , where s0 = t and sm = tn and set
tn m
p(t, s)f (s)ds

t i=1
(t, si1 ) P (t, si ) . p(t, si ) P
In our implementation, we work with a monthly grid.
4.4.2
Valuing Default Swaps
Similar to a plain vanilla interest rate swap, a default swap contract may be viewed as consisting of two legs: a xed leg and a oating leg. The former contains the payments by the buyer to the seller and is called the xed leg, because its payments are known at initiation of the contract. The oating leg comprises the potential payment by the seller to the buyer and at the start date it is unknown how much the seller has to pay if he has to pay at all. Consider a default swap contract with payment dates T = (T1 , . . . , TN ), maturity TN , (t, T, P ) premium percentage P and notional 1. Denoting the value of the xed leg by V (t), the value of the default swap to the buyer equals and the value of the oating leg by V
If we assume a stochastic recovery rate that is risk-neutrally independent from the default-free shortrate process and the default time, all formulas remain valid, except that should be interpreted as the expected recovery rate under the risk-neutral measure. Moreover, the results of Bakshi et al. (2001) indicated that a model with a stochastic recovery rate performs equally well as a model with a constant recovery rate. 6 This approximation is necessary, because we do not have an analytical expression for p(t, ), but use the markets default-free term structure instead.
5
82
(t) V (t, T, P ) and to the seller V (t, T, P ) V (t). At initiation, the premium P is V chosen in such a way that the value of the default swap is equal to zero, because only then the value to the buyer equals the value to the seller. Since the value of the xed leg is homogeneous of degree one in P , the premium percentage should be chosen as (t) V P = . V (t, T, 1) We rst determine the value of the xed leg.7 At each payment date Ti , the buyer has to pay (Ti1 , Ti )P to the seller, where (Ti1 , Ti ) is the year fraction between Ti1 and Ti (T0 is equal to t), taking into account the day count convention specied in the contract. If the reference entity does not default during the life of the contract, the buyer makes all payments. However, if default occurs at time s TN , the buyer has made only I (s) payments, where I (s) = max(i = 0, . . . , N : Ti < s) and the remaining payments I (s) + 1, . . . , N are no longer relevant; in addition, he has to make an accrual payment of (TI (s) , s)P at time s.8 The value of the xed leg at time t is thus equal to
N
(t, T, P ) = V
i=1 N
t (Ti1 , Ti )P 1{ >T } + E t p(t, )(TI ( ) , )P 1{ T } p(t, Ti )E i N (t, Ti ) + p(t, Ti )(Ti1 , Ti )P P

i=1 t TN
(4.3) p(t, s)(TI (s) , s)P f (s)ds.
To calculate the value of the oating leg, we have to distinguish between cash settlement and physical settlement. If default occurs, and the contract species cash settlement, the buyer keeps the reference obligation and the seller pays the buyer an amount equal to the dierence between the reference price and the nal price of the reference obligation. The reference price is specied in the contract and is typically equal to 100%. The nal price is the market value of the reference obligation on the default date as computed by the specied calculation agent by the specied valuation method; commonly, he has to poll one or more dealers for quotes on the reference obligation, disregard the highest
In calculating the values of the xed and oating legs, we do not take into account counterparty risk for two reasons. First, as noted in Section 4.2, market participants do not consider counterparty risk when determining default swap premiums. Second, Lando (2000) showed that allowing for counterparty default has virtually no eect on default swap premiums, unless the correlation between the default processes of the counterparty and the reference entity is very high. 8 We assume that if the default time exactly coincides with a payment date Ti , the buyer does not make the regular payment, but makes an accrual payment, i.e. I (Ti ) = i 1. Since the regular payment and the accrual payment are equal on a payment date, this assumption does not aect the value of the default swap.
7
83
and lowest quotes and calculate the arithmetic mean of the remaining quotes. Under our recovery assumption, the nal price of the reference obligation is equal to , so that the value of the oating leg under cash settlement equals t p(t, )(1 )1{ T } = (t) = E V N
TN
p(t, s)(1 )f (s)ds.

t
(4.4)
If the default swap contract species physical settlement things work somewhat dierently. At the default time, the buyer delivers one or more of the deliverable obligations with a total notional of 1 to the seller and the seller pays 1 to the buyer. Assuming one deliverable, the net value of these transfers is equal to 1 , so that the value of the oating leg under physical settlement is equal to its value under cash settlement. However, a default swap contract generally allows the buyer to choose from a list of qualifying obligations, irrespective of currency and maturity as long as they have the same seniority as the reference obligation. This feature is called the delivery option ; see also Section 4.2. We refrain from valuing the delivery option, and use the value of the oating leg under cash settlement. To numerically approximate the integrals in Equations (4.3) and (4.4), we use the same method as for the defaultable coupon bond price. Again, a monthly grid is chosen. Our default swap pricing formula is very similar to other models encountered in the literature. The models by Aonuma and Nakagawa (1998), Brooks and Yan (1998), Scott (1998), Jarrow and Turnbull (1998) and Due (1999) are equal to our model, except that they only allow defaults on premium payment dates. Consequently, they lack the accrual payment in the xed leg valuation. Cheng (2001) formally showed that the last four models, as well as a special version of Das and Sundaram (2000), are all mathematically equivalent. Nakagawa (1999) and Hull and White (2000), like us, also allowed defaults to occur on other dates than payment dates. However, Nakagawa (1999) did not incorporate the accrual payment and Hull and White (2000) assumed that the protection buyer makes a continuous stream of premium payments, rather than a set of discrete payments.
4.4.3
Specication
Since we focus on the estimation and application of credit risk models, we refrain from estimating a model for the default-free short-rate. Instead we use a priori estimated zerocoupon curves and use them to calculate the prices of default-free zero-coupon bonds. To completely specify the model, we have to (i) select a model for the risk-neutral haz-
84
ard process, (ii) pick a recovery rate and (iii) choose a proxy for the default-free term structure. Hazard Process In the existing empirical literature, as discussed in Section 4.3, all studies that use time series estimation model the hazard rate stochastically, typically as a Vasicek or CIR process. Papers that apply cross-sectional estimation techniques consider either constant or stochastic hazard rates; in the latter case, the stochastic process is chosen in such a way that the survival probability curve in Equation (4.1) is known analytically. We follow an intermediate approach by using a deterministic function of time to maturity. This specication facilitates parameter estimation, while still allowing for timedependency. We model the integrated hazard function as a polynomial function of time to maturity
d
(t, T ) =
i=1
i (T t)i ,
where d is the degree of the polynomial and 1 , . . . , d are unknown parameters. Note that we have imposed the required restriction (t, t) = 0 by omitting the constant term. This specication implies that the hazard rate itself is a polynomial of degree d1. The survival probabilities follow directly from Equation (4.1) as exp((t, T )). To the extent that the survival probability curve from a stochastic hazard specication can be approximated by our exponential-polynomial function, deterministic and stochastic models will yield similar results. Recovery Rate There are two approaches for the estimation of the recovery rate. The rst is to consider it as just another parameter, and estimate it from the data along with the other parameters. The second method is to a priori x a value. Although the rst method seems preferable, it turns out that it is hard to identify the recovery rate from the data. Figure 4.1a illustrates this for a constant hazard rate model estimated from a data set of Deutsche Bank bonds on 4 May 1999 (the rst day in our sample) using the swap curve as proxy for the default-free curve. We vary the recovery rate from 10% to 90% in steps of 10% and for each value we estimate the hazard rate. It is clear from the gure that the tted zero-coupon curves are virtually identical, except for the one estimated with a recovery rate of 90%. To get some intuition for this outcome, consider the price of
85
Figure 4.1: Sensitivity of spreads and default swap premiums to the recovery rate.
4.5%
zero-coupon rate
4.0%
10%
3.5%
30% 50%
3.0%
70%
2.5% 0 1 2 3 4 5 maturity 6 7 8
90%
9 10
(a) 1.2 1recovery rate 1.0 0.15 0.8 % 0.6 0.4 0.05 0.2 0.0 0 0.2 0.4 0.6 recovery rate (b) 15 default swap premium (bps) 0.8 1 0.00 0.10 % hazard rate (1recovery) hazard 0.20
14
13
12
11
10 0 0.2 0.4 0.6 recovery rate (c)
0.8
A reduced form model with a constant hazard rate is tted to market bid quotes of Deutsche Bank bonds on 4 May 1999 for various recovery rates. The default-free term structure is approximated by the swap curve. Graph (a) shows the tted zero-coupon curves; (b) shows 1 minus the recovery rate and the hazard rate on the left axis, and their product on the right axis; (c) shows the calculated premiums for a 5 year default swap.
86
a defaultable zero-coupon bond (cf. Jarrow and Turnbull, 1995, Equation (49)) (t, T ))]. v (t, T ) = p(t, T )[1 (1 )(1 P So, given a default-free curve p(t, T ), the price only depends on the product of 1 and (t, T ). Using Equation (4.1) and a rst order Taylor expansion, the bond price can 1P be approximated as v (t, T ) p(t, T )[1 (1 )(t, T )]. For the constant hazard rate model, (t, T ) = 1 (T t), so that the zero-coupon spread s(t, T ) with respect to the default-free rate is approximately equal to s(t, T ) (1 )1 ; see also Due and Singleton (1999, below Equation (5)). Decreasing 1 and simultaneously increasing 1 by the same ratio will result in approximately the same spread. Figure 4.1b shows that this indeed happens when we estimate the hazard rate for dierent values of recovery rate. As long as the recovery rate is chosen between roughly 10% and 80%, the product of 1 and 1 is approximately constant. It is clear that it is hard to identify both the hazard and recovery processes from bond data; see also Duee (1998, page 203), Due (1999, page 80), Due and Singleton (1999, page 705) and Fr uhwirth and S ogner (2003). This may pose a problem for some applications, but for our purpose of pricing default swaps it fortunately does not. It turns out that the default swap premium is also relatively insensitive to the assumed recovery rate. Figure 4.1c shows the premiums for a 5 year default swap written on Deutsche Bank for varying recovery rates (and thus varying hazard rates). As long as the recovery rate is chosen between roughly 10% and 80%, the estimated default swap premium is approximately between 13 and 15 basis points (bps). A smaller range of 14 to 15 bps is obtained, if the recovery is chosen between 10% and 60%. In conclusion, both the bond spread and the default swap premium primarily depend on the product of one minus the recovery rate and the integrated hazard function. Therefore, it is not very important to choose the recovery rate very accurately, as long as it takes a reasonable value. In our implementation, we set = 50%. Default-Free Interest Rates Our bond and default swap valuation models require a term structure of default-free interest rates as input data. Since a few years, xed-income investors have moved away from using government securities to extract default-free interest rates and started using plain
87
vanilla interest rate swap rates instead. Golub and Tilman (2000) and Koci c, Quintos and Yared (2000) mentioned the diminishing amounts of US and European government debts, the credit and liquidity crises of 1998, and the introduction of the euro in 1999 as primary catalyzing factors for this development. Nowadays, government securities are considered to be unsuitable for pricing and hedging other xed-income securities, because in addition to interest rate risk they have become sensitive to liquidity risk. Swaps, on the other hand, being synthetic instruments, are available in unlimited quantities, allowing investors to go long or short any desired amount. A disadvantage of swap rates is that they contain a credit risk premium due to two sources. First, being a bilateral agreement between two parties, an investor is exposed to the potential default of its counterparty. Due and Huang (1996) showed that this premium is quite small however: only one or two basis points for typical dierences in counterparties credit qualities. Second, the swaps oating leg payments are indexed on a short-term LIBOR rate, which is a defaultrisky rate. Therefore, the swap rate will be higher than the default-free rate even though the swap contract is virtually default-free; see Collin-Dufresne and Solnik (2001). An instrument that is less sensitive to the risk of counterparty default and is not linked to a risky rate is a repurchase agreement (repo for short; see e.g. Due, 1996). A repo is basically a collateralized loan, typically between two banks for a relatively short time period (1 day to, at most, 1 year). Each instrument has its own repo rate, and the highest repo rate is referred to as the general collateral (GC) rate.9 GC rates have historically been close to swap rates, but they were typically several basis points lower. The usage of repo rates as default-free interest rates was recommended by Due (1999, page 75). Repo rates were also used by Longsta (2000). Even though the above seems to be well-known to practitioners, the academic literature has paid little attention to the choice of the default-free curve. This is demonstrated by the fact that almost all empirical papers that estimate reduced form credit risk models used the government curve as the default-free curve; Due et al. (2003) are the only exception by using the swap curve. We estimate our models for all three proxies government, swap and repo curves and see which curve gives the best t to bond prices and default swap premiums.
Instruments whose repo rates are at or near the GC rate, are called general collateral. Instruments whose repo rates are signicantly below the GC rate are referred to as special. Since data on repo specialness is hard to obtain, we assume that all considered bonds are general collateral.
88
4.4.4
Estimation
We use cross-sectional estimation to estimate the parameters of our model. So, if we are given a default-free zero-coupon curve at time t, and the market prices P1 (t), . . . , Pb(t) (t) of b(t) defaultable bonds issued by a single entity, where the ith bond has payment dates ti and coupon percentage ci , we may estimate the parameters of that entitys integrated hazard function. We minimize the criterion function,
b(t)
(Pi (t) v (t, ti , ci ))2 ,

i=1
using the Gauss-Newton algorithm; see e.g. Greene (2000, Chapter 10). We repeatedly estimate the model until all residuals are smaller than 2.5 standard deviations, removing the bond with the largest residual (in absolute sense) each time this condition is not met. This procedure prevents strongly mispriced bonds from unreasonably aecting the estimated curves; see also Perraudin and Taylor (1999, Section 2.2). To estimate a curve on a particular trading day, we also consistently exclude all bonds with a remaining maturity of less than 3 months. In our data set, such bonds showed constant prices or they were not quoted at all for several consecutive days. Moreover, we require that on each day, quotes should be available for at least 5 bonds. This ensures some degree of statistical reliability of the estimated parameters. Strictly speaking, our approach of cross-sectional estimation, where we re-estimate the model parameters on a daily basis, is inconsistent with our assumption of a deterministic hazard function. Nevertheless, it is used in order to t the model to observed bond prices as closely as possible. This approach could be compared to the practice of estimating the implied volatility of the Black-Scholes model using one or more observed option prices, rather than estimating the (assumed constant) volatility using a time-series of returns of the underlying asset.
4.5
Data
The bond data set consists of corporate and sovereign bonds and is obtained from two sources. From Bloomberg, we obtain bond characteristics, like maturity dates, coupon percentages and seniorities; a time series of credit ratings for each issuer is also downloaded from Bloomberg. Clean bid and ask price quotes are retrieved daily at 4.00pm from Reuters Treasury and Eurobond pages. The data covers the period from 1 January 1999 to 10 January 2001 and contains prices of almost 10800 bonds issued by over 1600
Section 4.5 Data
89
dierent entities. The total number of price quotes is close to 2.5 million. To estimate the credit risk models, we construct a sample of xed-coupon, bullet, senior unsecured bonds that are denominated in euros or in one of the currencies of the participating countries. This reduces the number of bonds to 3920, the number of unique issuers to 704 and the number of quotes to approximately 1.1 million. The default swap data set is constructed by combining quotes from two sources. Firstly, it contains indicative bid and ask quotes from daily sheets posted by commercial and investment banks, such as J.P. Morgan Chase, Salomon Brothers, Deutsche Bank and Credit Suisse, and by brokers, such as Prebon, Tradition and ICAP. Secondly, it comprises bid and ask quotes from internet trading services creditex and CreditTrade, whose participants, in addition to banks and brokers, also include other nancial institutions and corporates. The data period ranges from 1 May 1999 to 10 January 2001. In this period, we observed 48098 quotes on default swaps on 837 distinct reference entities. Contracts denominated in US dollars make up 82% of the quotes, euro-denominated contracts account for 17% and the remaining 1% is comprised of British pounds, Japanese yens and Australian dollars. Quotes on dollar contracts are observed in the entire data period, whereas quotes on euro-denominated default swaps are only observed from March 2000 to January 2001. All contracts specify quarterly payments by the protection buyer. Virtually all quotes (99.7%) are for contracts with a notional amount of 10 million (denominated in one of the above mentioned currencies); the remaining contracts have face values of 1, 5, 15, 20, 25 or 50 million. The maturity of the default swaps ranges from 1 month to 20 years, with multiples of 6 months up to 10 years being most common; 5-year deals are most popular, making up 53% of the observations, followed by 3-year (10%), 10-year (7%) and 1-year (4%) contracts. For our subsequent analyses, we constrain ourselves to default swaps that are euro- or dollar-denominated, have a maturity of at most 10 years and a notional amount of 10 million. Imposing these constraints reduces the number of observations by 2.7%, but creates a more uniform data set by removing the least liquid contracts. Whenever we refer to the default swap data set we mean this restricted version of the original data set. For our research, we need reference entities for which both bond and default swap prices are available. Restricting the data sets to this subset of entities, leaves us with 225 reference entities, 1131 bonds, about 258000 bond prices and about 23000 default swap prices. As proxy for default-free interest rates, we consider three alternatives: government rates, swap rates, and general collateral (GC) repo rates. The zero-coupon euro government curve is estimated on a daily basis from a subset of the bond data set, consisting of liquid German government bonds. We model the discount function as a linear com-
90
bination of third degree B-splines basis functions with knots at 2, 5 and 10 years; see Chapter 2 for more details. Euro swap rates are downloaded from Bloomberg. We apply a standard bootstrapping procedure to extract zero-coupon rates and interpolate linearly between the available maturities to get a curve for all required maturities. Finally, we download euro repo benchmark rates from the website of the British Bankers Association (BBA, 2001). Unfortunately, the longest maturity for which GC rates are available is 1 year, which is too short for our purposes. Therefore, we use the following method to calculate approximate GC rates for all required maturities: on each day, we determine the 1-year swap-GC spread and assume that this spread may be subtracted from the swap rates of all other maturities to get the GC rates. Analogously to the swap curve, we use bootstrapping and linear interpolation to obtain a zero-coupon curve.
4.6
Results
In this section we rst discuss the properties of the default swap data set and implement an approximate default swap pricing method often applied by nancial market participants. Then, we present the results of applying our reduced form credit risk model to our data set. We conclude by analyzing the pricing errors of the model.
4.6.1
Analyzing Default Swap Premiums
Since the empirical literature on credit default swaps is restricted to just one other study (Aunon-Nerin et al., 2002), it is interesting to look at the properties of the data rst. Table 4.1 summarizes the default swap quotes on several characteristics. Panel I subdivides the 46820 observations by the reference entitys credit rating at the quote date. As may be expected, the rating is a very important determinant of default swap premiums as average premiums decrease monotonously with credit quality. Aunon-Nerin et al. (2002) found a similar pattern of premiums with credit rating; their average premiums per rating dier from ours, but this may be due to the dierent sample period. Within the class of investment grade ratings, a linear relation seems to exist: average premiums roughly double each time the credit quality decreases by one rating. For speculative grade issuers, average default swap quotes rapidly increase. In panel II of Table 4.1 the sample is further subdivided by deal type. The number of bid quotes is roughly equal to the number of ask quotes. As usual in nancial markets, a spread exists between them. The average bid-ask spread is 8 bps, but an increasing pattern with ratings may be observed. For AAA-rated entities the average bid-ask spread
Section 4.6 Results
91
Table 4.1: Characteristics of the default swap data set.

AAA AA A BBB BB B CCC NR All
I: Rating 12.0 24.5 43.4 (1794) (8321) (18613) II: Rating and deal type bid 9.8 20.7 38.4 (882) (3724) (8672) ask 14.2 27.6 47.8 (912) (4597) (9941) III: Rating and currency dollar 13.7 25.0 47.1 (1325) (5193) (14910) euro 7.5 23.7 28.6 (469) (3128) (3703) IV: Rating and maturity (0,1] 12.1 18.4 31.9 (55) (208) (790) (1,2] 9.4 22.3 30.3 (24) (80) (721) (2,3] 9.6 26.4 37.2 (387) (562) (2142) (3,4] 11.8 27.3 45.3 (24) (453) (1242) (4,5] 11.8 24.1 44.6 (675) (6072) (11743) (5,-) 14.0 26.3 51.3 (629) (946) (1975) continued on next page 87.9 (13187) 81.5 (6484) 94.1 (6703) 89.8 (12543) 51.7 (644) 100.8 (1191) 108.1 (788) 74.5 (1750) 87.0 (686) 82.3 (6710) 103.0 (2062) 269.5 (1595) 236.0 (868) 309.5 (727) 269.7 (1589) 223.0 (6) 199.1 (306) 290.2 (295) 242.4 (306) 330.9 (32) 281.6 (555) 419.0 (101) 483.4 (1118) 431.9 (592) 541.3 (526) 482.1 (1106) 598.3 (12) 407.6 (359) 432.8 (183) 501.0 (222) 567.9 (47) 561.2 (230) 622.1 (77) 1957.5 (10) 1985.0 (5) 1930.0 (5) 1957.5 (10) 55.7 (2182) 55.7 (880) 55.8 (1302) 59.9 (1775) 37.5 (407) 61.9 (124) 52.0 (71) 45.7 (279) 50.6 (335) 60.4 (1217) 45.2 (156) 70.6 (46820 ) 66.4 (22107 ) 74.4 (24713 ) 79.7 (38451 ) 28.7 (8369 ) 120.2 (3033 ) 133.5 (2166 ) 75.6 (5648 ) 65.8 (2822 ) 58.4 (27203 ) 75.5 (5948 )
2900.0 (4)
1016.7 (3) 425.0 (1) 2250.0 (2)
amounts to 4.4 bps, rising gradually within high grade ratings to 12.6 bps for rating BBB. A much larger spread exists between bid and ask quotes for speculative grade issuers: about 75 bps for BB and 110 for B. The bid-ask spread for CCC-rated default swaps is negative, but this is most likely caused by the small number of observations. Note that the bid-ask spreads are relatively large compared to the quote size. For instance, for AA the average bid-ask spread of 6.8 bps amounts to 28% of the average quote of 43.4 bps. Contracts in our data base are denominated in one of two currencies : either US dollars or euros. Panel III shows that dollar-denominated default swaps prevail, but recall that euro-denominated default swaps are only observed during the second half of the data period. For all ratings, dollar quotes are on average larger than euro quotes, except for
92
Table 4.1: continued

AAA AA A BBB BB B CCC NR All
V: Rating and date Q2-1999 13.4 19.7 (265) (839) Q3-1999 13.4 24.2 (211) (642) Q4-1999 12.2 25.5 (77) (622) Q1-2000 10.4 23.5 (32) (355) Q2-2000 11.9 27.8 (109) (837) Q3-2000 9.5 25.2 (393) (2489) Q4-2000 11.9 23.6 (619) (2271) Q1-2001 17.6 30.6 (88) (266)
56.0 (2374) 58.8 (2106) 42.4 (1835) 34.9 (1367) 37.8 (2990) 35.9 (4321) 41.4 (3166) 58.9 (454)
103.5 (2529) 104.1 (2258) 100.1 (1052) 63.1 (953) 73.1 (1636) 67.9 (2893) 93.7 (1566) 114.4 (300)
259.6 (345) 300.0 (253) 301.7 (280) 172.8 (118) 271.5 (99) 299.5 (304) 211.9 (186) 220.5 (10)
768.1 (105) 698.4 (217) 545.3 (104) 356.3 (99) 481.8 (56) 353.8 (415) 342.1 (118) 518.8 (4)
2585.7 (7)
491.7 (3)
98.2 (110) 134.7 (110) 46.6 (90) 50.6 (185) 47.1 (570) 46.5 (618) 53.3 (464) 61.4 (35)
90.7 (6567 ) 106.8 (5797 ) 89.4 (4067 ) 58.4 (3109 ) 53.7 (6297 ) 60.0 (11436 ) 52.8 (8390 ) 66.7 (1157 )
The table shows average default swap premiums by rating (Panel I), rating and deal type (II), rating and currency (III), rating and maturity (IV) and rating and quote date (V). The number of observations per cell is shown in parentheses.
rating B where the number of euro observations is rather small. This nding still holds, if we also control for quote date, though to a lesser extent (not shown here). The relation between premium level and contract maturity is assessed in panel IV. Notice that more than 50% of the observations resides in the 4- to 5- year maturity range; Aunon-Nerin et al. (2002) also reported this asymmetric distribution of the quotes over the maturity range. Aggregated over all ratings, there does not seem to be a clear relation between the average default swap premium and maturity. Only for A- and Brated reference entities an increasing pattern may be detected, but for the other ratings premiums are approximately constant as a function of time to maturity. Our ndings are in line with those of Aunon-Nerin et al. (2002), who tested several specications for the maturity eect, but none of them appeared to be signicant. Finally, panel V looks at the behavior of average premiums over time by grouping the default swaps by quote date into three-month periods. Except for AA and BB, the quotes for all ratings roughly follow a U-shape pattern over time: in the middle of the sample period, the average premium is lower than at the start and at the end.
Section 4.6 Results
93
4.6.2
Comparing Bond Spreads and Default Swap Premiums
To admit a direct comparison between bonds and default swaps, we make the following intuitive argument. Suppose an investor in a coupon-bearing defaultable bond buys protection by entering into a credit default swap. The package consisting of the bond and the default swap is free of default risk, so defaultable bond + default swap = defaultfree bond. Therefore, we also have default swap = default-free bond defaultable bond. Consequently, the default swap premium should be equal to the spread between the defaultable and the default-free bond. To formalize our argument, consider a defaultable bond with coupon payment dates t = (t1 , . . . , tn ), coupon c, maturity tn and notional 1. Further, consider a default swap with the same maturity, premium percentage P and notional 1. For simplicity, assume that the default swaps and bonds payment dates coincide. (t, t, P ) + V (t), The value V (t) of the package to the investor is given by v (t, t, c) V whose formulas are given by Equations (4.2) to (4.4). We replace the integrals in the pricing formulas by the approximations from Section 4.4; the grids for both the bond and default swap are chosen to be equal to the payment dates t. Under these assumptions, the loss of 1 on the bond on a potential default date is exactly oset by the reception of 1 from the default swap. Moreover, the accrual payments of the default swaps xed leg vanish, because for the current grid choice default can only occur on a payment date. Therefore, the value of the package can be written as
n
V (t) =
i=1 n
(t, ti ) + p(t, ti )(c P (ti1 , ti ))P (4.5) (t, ti1 ) P (t, ti ) + p(t, tn )P (t, tn ). p(t, ti ) P
i=1
The rst line indicates that the coupon payments c on the bond are reduced by the insurance premium P (ti1 , ti ) on the default swap. The second line shows that the notional of 1 will be paid eventually, but that the timing depends on the occurrence of the credit event. The investor is thus protected against default risk, but is now exposed to the risk of prematurely receiving the notional and thus missing some of the promised coupons (prepayment risk ). If we additionally assume that both the default-free and defaultable bond are priced at par, then their yields are equal to their coupon rates (ignoring the prepayment risk). Let y and Y denote the yield of the defaultable and the default-free bond, respectively, then y = c and Y = c P , so that y Y = P . This conrms our intuition that the bond
94
spread should be equal to the default swap premium. Note however that we had to make several assumptions to get this result, so that it is only approximately valid. Obviously, its validity is further impaired by the implicit assumptions that the bond and default swap pricing models are correct, and that bonds and default swaps are priced o the same survival curve by nancial market participants. Nevertheless, bond spreads and default swap premiums should be comparable. This relation was also presented by Aunon-Nerin et al. (2002), though without proof. Due (1999) showed that this relation holds exactly for par oating rate notes instead of xed income coupon bonds. OKane and McAdie (2001) discussed a large number of factors that may aect the dierence between bond spreads and default swap premiums. We will now determine to what extent this relation holds for our data set. For each quoted default swap written on an entity, we have to nd a quoted bond issued by that same entity with the same maturity. Unfortunately, a bond with exactly the same maturity as the default swap is rarely available. Therefore, we examine two alternative methods: 1. Find a quoted bond whose maturity diers at most 10% from the default swaps maturity; 2. Find two quoted bonds, one whose maturity is smaller than, but at most twice as small as, the default swaps maturity, and one whose maturity is larger than, but at most twice as large as, the default swaps maturity, and linearly interpolate their spreads. We will refer to method 1 as the matching method and to method 2 as the interpolation method. The performance of each method is evaluated for all three proxies for the defaultfree term structure. That is, a bonds spread is calculated by subtracting either the swap, repo or government rate from its yield-to-maturity. This gives a total of six method-proxy combinations. Each time a pair can be formed of a default swap premium and a (matched or interpolated) bond spread, we calculate two pricing errors. One by subtracting bond bid spreads from default swap ask quotes, and the other by subtracting bond ask spreads from default swap bid quotes.10 The pricing errors are summarized in two ways. First, as the average, denoted by the Mean Pricing Error (MPE), and second as the average of the absolute values, called the Mean Absolute Pricing Error (MAPE). A negative (positive) sign of the MPE statistic indicates that bond spreads are, on average, too large (small) and thus that
In this way, we are contrasting similar sides of the market. For instance, an investor can create an exposure to an issuers default by buying a bond for which he pays the ask price or by writing default swap protection for which he receives the bid premium.
10
Section 4.6 Results
95
the bond markets estimate of the issuers credit risk is larger (smaller) than the default swaps market estimate. To test if this under- or overestimation is signicant, we create a time series MPEi1 , . . . ,MPEiS , where MPEit denotes the mean pricing error for method i on date t. We then apply a one-sample Z-test (see e.g. Arnold, 1990, Chapter 11) Zi = MPEi , S si
where MPEi and si are the sample mean and sample standard deviation of the MPEit series, respectively, and S is the sample size. Asymptotically, Zi has a standard normal distribution. Similarly, to determine if signicant performance dierences exist between our methods, we create a time series MAPEi1 , . . . ,MAPEiS of mean absolute pricing errors for each method i. Then a paired Z-test (Arnold, 1990, Chapter 11) can be used to determine if method is pricing performance is signicantly dierent from method j s. The test lets us nd out whether two time series have the same mean, while allowing for non-zero correlation and unequal variances. The test statistic is dened as Zij = ij d S , sij
ij and sij are the sample mean and sample standard deviation, respectively, of where d dijt = MAPEit MAPEjt , t = 1, . . . , S . Asymptotically, Zij has a standard normal distribution. Figure 4.2 depicts scatter plots of pricing errors versus default swap premiums per rating for the interpolation method that uses the swap curve as default-free curve; the plots for the other methods are similar. If interpolated bond spreads over the swap curve are good estimates of default swap premiums, all (default swap premium, bond spread) pairs should lie on the horizontal axis. For ratings AAA, AA and A, this seems indeed to be the case, so that on average the method does a good job. This is conrmed by the MPE values in Table 4.2, which are approximately zero, though only for AAA insignicant. For rating BBB, the scatter plot and MPE statistic indicate that bond spreads are on average smaller than default swap premiums; for ratings BB and B this is almost always the case. Moreover, the Z-test indicates that the MPEs are signicantly dierent from zero, so that for BBB, BB and B bond spreads are biased estimates of default swap premiums. For all ratings, the dispersion around the horizontal axis is fairly large. In fact, the MAPE statistics in Table 4.2, together with the average default swap premiums in Table 4.1, imply that the calculated premiums deviate on average 19% (for BBB) to 68% (for AAA) in absolute value from the market values.
96
Table 4.2: Performance of the direct comparison methods.

AA A BBB IG a BB B NR SG a All a
AAA
Matching Obs.b 1058 2168 1951 1188 6365 441 297 738 40 7144
Swap
Repo
Government
5.9 (9.1) 1.7 (8.3) -31.1 (31.4) 292 1839 2260 1067 5458 316 387 703 61
-1.4 (14.6) -5.8 (15.9) -32.6 (34.4)
-4.9 (11.4) -9.7 (13.5) -37.1 (37.8)
9.4 (34.3) 4.4 (34.3) -15.8 (41.0)
0.7 (16.4 ) -3.8 (17.3 ) -30.6 (36.1 )
129.7 (137.0) 124.9 (133.6) 106.6 (118.5)
174.9 (187.2) 170.3 (183.0) 151.3 (165.6)
148.0 (157.3 ) 143.3 (153.5 ) 124.7 (137.5 )
0.2 16.0 (27.0) (31.0 ) -5.5 11.4 (29.1) (31.5 ) -29.1 -14.5 (41.8) (46.7 ) 6222
Interpolation Obs.b
Swap
Repo
Government
-1.6 0.8 (8.2) (11.1) -3.4 -6.0 (8.2) (12.0) -33.9 -33.4 (34.3) (34.2) -3.7 (10.6) -8.7 (12.8) -31.7 (32.5) 16.6 (29.5) 11.6 (29.2) -7.5 (34.8) 1.2 (14.3 ) -3.5 (15.5 ) -27.6 (33.6 ) 154.1 (156.0) 149.4 (151.8) 177.2 (178.7)
200.5 (201.5) 196.0 (197.2) 133.6 (137.2)
179.6 (181.1 ) 175.1 (176.8 ) 157.6 (160.0 )
-4.3 21.3 (28.6) (33.3 ) -9.9 16.6 (32.0) (33.9 ) -32.6 -6.8 (49.5) (48.1 )
The table shows mean pricing errors (MPE) and mean absolute pricing errors (between parentheses) by rating, pricing method and proxy for the default-free curve. All MPEs are signicant at condence levels above 99%, except for those marked with and , which are insignicant at condence levels up to 95% and 99%, respectively. IG=investment grade subsample; SG=speculative grade subsample; All=entire sample. Number of (bond spread, default swap premium) pairs that could be formed.
Section 4.6 Results
97
Figure 4.2: Scatter plots of pricing errors versus default swap premiums per rating.
40 30 20 10 0 0 -10 -20 0 10 20 AAA 175 150 125 100 75 50 25 0 -25 -50 0 50 100 A 500 400 300 200 100 0 -100 0 250 BB 500 750 1000 700 600 500 400 300 200 100 0 -100 -200 0 200 400 B 600 800 1000 1200 150 200 250
-100 0 100 200 300 400 500 600 700 0 100 200 300 400
100 75 50 25
-25 -50 -75 30 40 50

0 25 50 75 100
AA
BBB
The graphs depict scatter plots of pricing errors (on the vertical axis) versus default swap premiums (horizontal axis). A pricing error is dened as a default swap quote minus a bond spread calculated with the interpolation method and using swap rates as proxy for default-free interest rates.
Tables 4.2 and 4.3 also shed light on the performance of the other methods. A striking feature of these results is the abominable performance of the government-based methods for AAA- to A-graded issuers. Their MAPE values are up to four times higher than for methods that proxy the default-free curve with the swap or repo curve. As the paired Ztests in Table 4.3 point out, the underperformance of the government curve for investment grade issuers (so including BBB too) is highly signicant. Moreover, since the MPEs are negative, almost identical in size to the MAPEs and statistically signicant, bond spreads relative to the government curve are virtually always larger than default swap premiums for high grade issuers. The MPE values for the swap and repo curve methods are much closer to but still signicantly dierent from zero. Looking at the MAPE statistics as
98
Table 4.3: Paired Z-tests of the direct comparison methods.

Matching Repo Government Interpolation Repo Government
Swap
Swap
Investment grade Matching Swap Repo Government Interpolation Swap Repo Government Speculative grade Matching Swap Repo Government Interpolation Swap Repo Government
-12.60
-28.08 -27.41
6.87 9.56 29.27
4.41 7.32 29.28 -9.54
-14.38 -12.55 7.63 -24.71 -25.73
14.30
13.89 12.21
-4.13 -4.61 -6.36
-3.45 -3.93 -5.68 31.28
-1.46 -1.94 -3.73 22.72 17.26
The table shows t-values of paired Z-tests for all combinations of pricing methods and proxies for the default-free curve.
well, the swap curve methods perform signicantly better than the repo curve methods for investment grade, but signicantly worse for speculative grade entities. For speculative grade issuers, the government curve methods signicantly outperform their swap- and repo-based counterparts. For all methods the MPE statistics take large, signicantly positive values though, indicating that they all result in bond spreads being smaller than default swap premiums. Our overall impression is that bond and default swap markets deviate considerably, with absolute deviations increasing for lower credit ratings. For AAA- to A-rated issuers, bond spreads over swap rates give reasonable estimates of default swap premiums. For ratings BBB to B, default swap premiums are substantially larger than bond spreads for all considered methods. Recall that the applied direct comparison method is only an approximation to the actual relationship between bonds and default swaps; specically, a part of the dierence between bond spreads and default swap premiums is due to the presence of prepayment risk; see the discussion below Equation (4.5). An application
Section 4.6 Results
99
of our reduced form credit risk model, which species the exact dependence of both instruments on interest rates, hazard rates and recovery rates, may yield better results. This is the topic of the remainder of the chapter.
4.6.3
Estimating Hazard Functions
We now turn to the estimation of credit risk models as described in Section 4.4.4. We consider three proxies for the default-free curve government, swap and repo curves and three degrees for the integrated hazard function linear, quadratic and cubic. This yields a total of nine models. For each issuer, we estimate all models for each day that we have at least one default swap quote and at least ve bond quotes. The rst row of Table 4.4 shows the number of issuer-days on which this was the case. Most observations are for AAA- and AA-rated issuers even though rating classes A and BBB contain the largest number of default swap quotes; see Table 4.1. This is caused by the composition of the bond data base, which contains (a) a small number of BBB-rated bonds and (b) a small number of bonds per A-rated issuer. The remainder of Table 4.4 shows the average t of the model to the bond prices. A models quality at time t may be assessed by looking at the root mean squared error (RMSE) of the deviations between the market prices and the model prices 1 b(t)
b(t)
RMSE(t) =
ei (t)2
i=1
where ei (t) = Pi (t) v (t, ti , ci ). A models RMSE is calculated as the average of its RMSE(t) values over all days in the sample. Looking at the dierences between ratings rst, we observe that the goodness of t deteriorates as the rating declines. AAA- and AA-rated bond can be tted quite accurately with RMSE values of 15 to 20 bps. The RMSEs for other investment grade bonds are well below 1% and usually below 50 bps. For speculative grade bonds, RMSE statistics range from 1% to 2%. Next we try to determine which model performs best. We nd that using more parameters obviously yields a better in-sample t. This holds for all three proxies for the default-free curve, but especially for the government curve if we move from a linear to a quadratic model. Table 4.4 also indicates that using a cubic model has little advantage over a quadratic model for high grade bonds, but does improve the t for low grade bonds. As to the choice of the default-free curve, we see that for the linear model, the government curve clearly underperforms the swap and repo curves. For quadratic and cubic models, on the other hand, choosing a proxy for the default-free curve is less
100
Table 4.4: Goodness of t of the reduced form credit risk models.

AAA AA A BBB BB B NR All
Obs.a
933
955
166
182
146
255
2639
Swap 1 0.17 2 0.15 3 0.15 Repo 1 0.17 2 0.15 3 0.15 Government 1 0.37 2 0.18 3 0.16
0.22 0.17 0.15 0.22 0.17 0.15 0.38 0.17 0.15
0.39 0.26 0.23 0.39 0.27 0.23 0.51 0.27 0.24
0.74 0.54 0.46 0.74 0.54 0.46 0.87 0.55 0.46
2.15 1.26 1.13 2.14 1.26 1.13 2.29 1.25 1.12
1.95 1.21 0.97 1.95 1.20 0.96 2.10 1.21 0.95
0.53 0.30 0.15 0.54 0.30 0.16 0.61 0.31 0.17
0.88 0.55 0.46 0.88 0.55 0.46 1.02 0.56 0.46
The table shows the average t measured by the root mean squared error (RMSE) of the bond residuals. Each model is characterized by the proxy for the default-free curve and the degree of the integrated hazard function. Number of issuer-days on which we have at least one default swap quote and at least ve bond quotes.
important. Overall it seems sucient for AAA and AA to use a linear model with a swap or repo curve, for A and BBB a quadratic model with a swap or repo curve, and for BB and B a cubic model with any proxy for the default-free curve. Now we turn to the discussion of the estimated coecients of the integrated hazard function; see Table 4.5. Let us rst look at the case where the integrated hazard function is modelled as a rst degree polynomial. Since the hazard function is a constant in this model, the 1 parameter may be interpreted as the issuers average risk-neutral default intensity. The estimation results for all three default-free curve proxies indicate that the average default intensity increases with the issuers credit rating, except that Bs is somewhat below BBs. Therefore, on average, credit ratings do a good job in ranking rms by credit worthiness. However, the level of the default intensity diers considerably between the models: the default rate for the government curve model is about 50 to 60 bps higher than the default rate in the repo curve model, which in turn is about 10 bps higher than in the swap curve model. Obviously, these dierences reect the average spreads between these three curves, but especially for investment grade bonds they lead to substantially dierent default rates. For instance, if we use the swap curve we would
Section 4.6 Results
101
Table 4.5: Parameter estimates for the reduced form credit risk models.
AAA AA A BBB BB B NR
Swap 1 1 0.07 2 1 0.10 2 -0.02 3 1 0.10 2 -0.02 3 0.00 Repo 1 1 0.15 0.18 2 1 2 -0.02 3 1 0.18 2 -0.02 3 0.00 Government 1 1 0.75 2 1 0.39 2 0.07 3 1 0.26 2 0.12 3 -0.01
0.36 0.27 0.02 0.27 0.02 0.02 0.45 0.36 0.02 0.36 0.01 0.02 1.04 0.50 0.10 0.45 0.13 0.02
0.72 0.46 0.05 0.42 0.07 -0.01 0.83 0.55 0.05 0.52 0.07 -0.01 1.36 0.71 0.11 0.61 0.17 -0.01
1.42 1.05 0.05 1.16 0.00 0.01 1.52 1.14 0.05 1.23 0.01 0.01 1.98 1.32 0.10 1.29 0.10 0.00
7.18 4.83 0.38 5.46 0.18 0.02 7.28 4.93 0.38 5.55 0.18 0.02 7.66 5.06 0.43 5.61 0.24 0.02
6.65 4.52 0.39 2.98 1.02 -0.06 6.75 4.61 0.39 3.08 1.02 -0.06 7.23 4.78 0.45 3.15 1.11 -0.06
1.95 0.43 0.43 -2.77 2.47 -0.31 2.05 0.53 0.43 -2.69 2.48 -0.31 2.59 0.73 0.53 -2.53 2.61 -0.31
The table shows (100 times) the average parameter estimates i , i = 1, 2, 3 (0 is restricted to zero) per rating and model. Each model is characterized by the proxy for the default-free curve and the degree of the integrated hazard function.
conclude that AAAs default rate is only 7 bps, but if we use the government curve that number would be about ten times as large. In the quadratic integrated hazard model, the hazard function is linear, so that 1 and 2 are the level and slope coecient of the hazard function, respectively. The estimates for both coecients rise as the rating declines (with two exceptions: the level of B for all default-free curve proxies is smaller than that of BB and BBBs slope in the government curve model is slightly smaller than As). These ndings imply that if we compare the estimated spread curves of two ratings, the worst ratings spread curve both starts at a higher level and is steeper. Again, noticeable dierences exist between the three defaultfree curve types. Using the government curve results in higher and steeper spread curves
102
than using the swap curve. Comparing swap and repo curves, we nd that the levels dier by about 10 bps, just like in the linear model, but that the slopes are exactly equal. This simply reects the construction of our repo curve as a parallel shift of the swap curve. Finally, we look at the results of the cubic model for the integrated hazard function. Unlike the nicely ordered level and slope coecients, there does not seem to be relation between 3 and the credit rating. Both the size and sign of this parameter appear to be uncorrelated with credit worthiness. Considering the results for the government curve, we observe that, compared to the quadratic model, the level coecient has decreased and the slope coecient has increased for all ratings (except BB). Interestingly, the 3 parameters of the government curve models are almost equal to those of the swap and repo curve models. Further, in the latter models, the level coecient is virtually unchanged for AAA, AA and A, has increased for BBB and BB and has decreased for B. Similarly, the level coecients is virtually unchanged for AAA, AA and A, has decreased for BBB and BB and has increased for B. In conclusion, the choice of a proxy for the default-free curve has a signicant impact on the estimated credit risk model. Both the level and shape of the hazard function are substantially eected by this decision. Moreover, the t of the model to investment grade bonds is better if we use the swap or repo curve instead of the government curve. For speculative grade bonds, choosing a proxy is less important. Which combination of a default-free curve proxy and polynomial degree is optimal, can be judged by applying the models to the pricing of credit default swaps.
4.6.4
Comparing Model and Market Premiums
Having estimated a credit risk model for an issuer allows us to calculate model premiums of credit default swaps written on that issuer.11 Like above, we dene the pricing error as the market premium minus the model premium and summarize it by the Mean Pricing Error (MPE), and Mean Absolute Pricing Error (MAPE). A negative (positive) sign of the MPE statistic implies that the model premiums are, on average, too large (small) and thus that the model overestimates (underestimates) the issuers credit risk. Table 4.6 contains the MPE and MAPE gures for all nine models subdivided by rating, as well as one-sample Z-tests for the MPE statistics. Table 4.7 shows the outcomes of the paired Z-tests for the investment grade and speculative grade subsamples. If we
All premiums are calculated using euro proxies for the default-free term structure. Ideally, we should use dollar curves for dollar-denominated contracts, but we are unable to obtain dollar repo rates. However, repeating the analysis in this section for government and swap curves with dollar curves, gives very similar results: dollar-based and euro-based premiums have a correlation of over 99%.
11
Section 4.6 Results
103
compare these gures to the ones in Tables 4.2 and 4.3, we see very similar patterns. First, mean absolute pricing errors increase with credit ratings for all models, except for high grade entities in the models that use the government curve. Second, governmentbased models perform very badly for investment grade issuers: their MAPE statistics take signicantly larger values than for swap- and repo-curve models, and their signicantly positive MPE statistics indicate a large overestimation of default swap premiums. Third, the MPE values for the swap and repo curve models are close to zero, and for rst degree swap and second and third degree repo models mostly statistically insignicant. Fourth, for speculative grade bonds, the government curve models outperform the swapand repo-based models by two to 20 basis points; for quadratic and cubic models this outperformance is signicant. Finally, for all models the MPE statistics take signicantly positive values, indicating that they all underestimate the credit risk of low grade entities. Estimating a hazard rate model gives a clear improvement over the direct methods of Section 4.6.2. Comparing the best direct method to the best model for each default-free curve proxy, we nd that MAPE statistics are reduced by 35% to 55% for investment grade issuers. For A-rated entities the reduction is less spectacular with a decrease of about 15%. Also, for government-based approaches the best method and best model perform similarly poor. For speculative grade issuers hazard rate models outperform the direct methods by 15% to 20%. Even though using a model works better than directly comparing bond spreads and default swap premiums, the model premiums of the bestperforming model still deviate on average 20% to 50% in absolute value from the market values. So the models t rather well to bonds, but their out-of-sample performance on default swaps is somewhat poor. We now try to identify the best model. The results show that models that use the swap or the repo curve on average do a reasonable job for investment grade issuers. The swap-based models somewhat underestimate the true default swap premiums, and except for the linear model, this bias is signicant; the repo-based model, on the other hand, slightly overestimates the market premiums, but these dierences are mostly insignicant (except for the linear model). Looking at the MAPE statistics as well, the linear and quadratic swap curve models signicantly outperform their repo- and government-based counterparts for investment grade entities; the dierences between swap and repo models are economically limited though. For speculative grade issuers, the quadratic and cubic government curve models signicantly work better than swap and repo models with equal degrees. As to the choice of the optimal degree of the integrated hazard function, the paired Z-tests indicate that the quadratic model works signicantly better than, or not signicantly dierent from, the linear and cubic models. This result holds for both
104
Table 4.6: Performance of the reduced form credit risk models.

AA A BBB IG a BB B NR SG a All a
AAA
Swap 1 12.5 (12.5) 74.4 (74.4)
1.8 (4.5) 3.0 (4.1) 3.7 (4.9) -3.1 -4.7 (9.6) (11.8) -1.4 -2.3 (7.8) (10.4) 0.0 -1.4 (8.1) (9.4) -36.6 (36.7) -28.7 (28.8) -28.4 (28.5) -35.7 (36.3) -25.3 (27.0) -22.9 (25.0) -29.9 (36.0) -17.5 (24.7) -13.0 (22.3) -34.8 (36.1 ) -25.4 (27.2 ) -23.5 (26.0 ) 34.6 (103.6) 104.8 (114.8) 96.6 (123.0) 56.8 (149.6) 123.8 (144.5) 148.1 (151.0) 47.4 (130.2 ) 115.8 (131.9 ) 126.1 (139.0 ) -8.5 -4.5 (24.8) (12.6 ) 0.3 -1.4 (18.8) (10.4 ) 2.6 0.0 (19.6) (10.4 ) 55.6 (109.7) 118.7 (126.7) 109.4 (134.0) 83.7 (151.1) 141.6 (156.4) 164.5 (166.8) 71.8 (133.6 ) 131.9 (143.8 ) 140.9 (152.7 ) 7.4 (8.2) 69.3 (69.3) 9.7 (35.1 ) 24.9 (36.7 ) 30.3 (41.1 ) -18.6 (18.6) 55.1 (55.1) -19.5 (53.7 ) 0.5 (46.4 ) 6.4 (48.6 )
0.6 (8.2) 2.5 (6.9) 4.0 (7.9)
-0.4 -3.6 (11.7) (24.7) 2.1 5.1 (10.0) (19.3) 2.8 7.4 (9.3) (20.6) 60.3 (110.8) 123.4 (130.5) 114.0 (137.2) 88.0 (152.4) 145.9 (159.5) 168.7 (170.7) 76.3 (134.8 ) 136.3 (147.2 ) 145.3 (156.3 ) 14.3 (35.6 ) 30.2 (38.2 ) 36.6 (44.2 )
-0.4 (12.0 ) 2.8 (10.1 ) 4.3 (10.8 )
Repo 1
-2.0 (4.7) 2 -0.3 (3.7) 3 0.4 (3.9) Government 1 -33.4 (33.5) 2 -25.3 (25.3) 3 -25.9 (25.9)
The table shows mean pricing errors (MPE) and mean absolute pricing errors (between parentheses) by rating and model. Each model is characterized by the proxy for the default-free curve and the degree of the integrated hazard function. All MPEs are signicant at condence levels above 99%, except for those marked with and , which are insignicant at condence levels up to 95% and 99%, respectively. IG=investment grade; SG=speculative grade; All=entire sample.
Section 4.6 Results
105
Table 4.7: Paired Z-tests of the reduced form credit risk models.
Swap 2 Repo 2 Government 1 2
Investment grade Swap 1 4.50 0.70 2 -2.79 3 Repo 1 2 3 Government 1 2 3 Speculative grade Swap 1 -1.43 -4.02 2 -5.04 3 Repo 1 2 3 Government 1 2 3
-7.87 -7.11 -3.06
1.85 -4.71 0.43 5.95
0.19 -3.48 -1.30 2.80 -1.28
-30.40 -27.47 -24.48 -29.94 -27.42 -24.27
-22.44 -24.04 -18.50 -20.57 -25.22 -19.14 16.39
-18.78 -21.57 -22.26 -16.82 -21.63 -23.73 14.06 1.07
1.06 1.44 3.97
-0.29 19.22 6.61 -0.37
-3.03 -3.28 23.75 -3.01 -4.88
-0.87 0.71 2.96 -1.27 -0.13 2.15
3.03 15.54 10.23 2.71 12.82 8.61 2.49
-0.18 1.51 22.14 -0.27 0.01 17.54 0.14 -3.87
The table shows t-values of paired Z-tests for all combinations of models and proxies for the default-free curve.
investment grade and speculative grade issuers. All in all, a quadratic model that uses the repo curve seems to be the best choice for investment grade issuers: it gives unbiased estimates, and has the second lowest MAPE values. For speculative grade entities, none of the considered models can be recommended, since they all signicantly underestimate credit risk. The underestimation of default swap premiums for speculative grade issuers is substantial. If an investor would like to exploit this dierence between bond and default swap markets, he has to write default swap protection (receiving the high default swap
106
premium) and short a bond (paying the low bond spread). However, as noted earlier in Section 4.2, shorting corporate bonds is typically unfeasible, so that positive pricing errors can persist in the market. An explanation for these pricing errors may be found in missing features in the bond and/or default swap pricing models. OKane and McAdie (2001) and Hjort, McLeish, Dulake and Engineer (2002) discussed a large number of factors that may aect the dierence between default swap and bond markets. They mentioned that for speculative grade issuers, the delivery option in physically settled default swaps (see Section 4.2) is particularly important: for higher default probabilities, it is more likely that the default swap will be triggered, and the protection buyer actually has to deliver obligations to the seller. The delivery option will thus be more valuable for low grade issuers.12 Since the delivery option is not taken into account in our default swap pricing model in Section 4.4.2, the pricing error can be used as a rough estimate of the value of the delivery option. Hence, if the delivery option is truly an important missing factor, pricing errors should be positively correlated with hazard rates. If we calculate the correlation 1 , we get values between of pricing errors from the linear integrated hazard model with 35% and 47%, depending on the proxy for the default-free curve.
4.6.5
Analyzing Pricing Errors
In the previous section, we showed that the mean absolute pricing error increases if the credit rating deteriorates. Moreover, speculative grade default swaps are grossly underpriced, but investment grade contracts can be priced more or less unbiasedly. It is also interesting to see if dierences between market and model premiums can be related to other characteristics than the issuers credit rating. We try to accomplish this by regressing absolute pricing errors on dummy variables for the following characteristics: deal type (bid or ask), currency (euro or dollar), rating (AAA, AA, A, BBB, BB), maturity (1-year intervals up to 5 years, and an interval from 5 to 10 years) and quote date (6month periods). Since each set of dummies contains mutually exclusive categories, their values sum to 1 for each observation. Usually, the coecient of one of the dummies is set to 0 as identifying restriction. We take a dierent approach here, and set a linear combination of the coecients to 0, where the weight of a coecient is equal to the sample mean of the corresponding dummy variable. For instance, if and denote the coecients of the bid and ask dummies, we could have set one of them to 0. Instead, we impose b + a = 0, where b is the percentage of bid observations and a the percentage
OKane and McAdie (2001) derived a back-of-the-envelope estimate of the value of the delivery option: for a default swap written on an issuer with an annual default probability of and a potential gain of G of switching from an expensive to a cheaper deliverable, the value is approximately G.
12
Section 4.6 Results
107
of ask observations. The advantage of imposing these restrictions is that the constant of the regression equals the sample mean of the dependent variable. Furthermore, each coecient can be interpreted as the change in the absolute pricing error for that dummy for an otherwise representative observation. Table 4.8 shows the regression results for the quadratic specication of the integrated hazard function for all three proxies for the default-free curve; the results for the linear and cubic model are similar. We observe that most parameters are statistically dierent from zero, the signs of the parameters are largely consistent between the models and the R2 values are between 46% and 58%. Mispricings strongly dier between deal types as errors on bid quotes are larger than on ask quotes. Since we use hazard functions estimated from bond bid quotes to calculate default swap ask premiums, and vice versa, this implies that the bid-side of the bond market is somewhat more in line with the ask-side of the default swap market than vice versa. The maturity of the default swap contract is also predictive of the pricing error, since the coecients of the maturity dummies are signicant (except for the interval from 3 to 4 years) and monotonously increasing. Likewise, the coecients of the rating dummies are highly signicant and monotonously increasing. The currency dummies are only signicant for the government model, indicating that the swap and repo models price dollar- and euro-denominated default swaps similarly. Finally, the parameters for the quote date dummies show that for most models pricing errors in 2000 were smaller than in 1999 or 2001, although not all coecients are statistically signicant.
108
Table 4.8: Analysis of absolute pricing errors from reduced form credit risk models.
Swap Repo Government
R2 constant Deal type bid ask Maturity (0,1] (1,2] (2,3] (3,4] (4,5] (5,-) Rating AAA AA A BBB BB B Currency dollar euro Date Q2 1999 Q3,Q4 1999 Q1,Q2 2000 Q3,Q4 2000 Q1,Q2 2001
38.2 5.4 -7.0 -48.1 -26.6 -8.7 -2.7 7.7 19.7 -36.8 -36.5 -30.6 -22.2 102.6 138.3 -0.9 2.1 6.0 0.4 -6.4 -4.3 11.1
58% (52.6) (8.4) (8.4) (26.6) (8.3) (4.1) (0.8) (9.6) (6.7) (10.5) (26.5) (27.5) (11.0) (38.9) (60.3) (1.3) (1.3) (2.7) (0.2) (2.4) (1.7) (1.9)
36.7 5.7 -6.9 -45.2 -24.5 -7.9 -2.1 7.0 16.9 -33.9 -33.0 -28.5 -21.4 98.9 135.1 -0.6 1.3 6.1 1.0 -6.5 -5.1 8.2
57% (52.9) (8.9) (8.9) (24.5) (7.8) (3.9) (0.7) (9.2) (6.1) (11.0) (25.6) (26.9) (11.0) (38.4) (60.3) (0.9) (0.9) (2.8) (0.4) (2.6) (2.1) (1.5)
46.4 5.6 -6.4 -37.7 -18.2 -7.5 -1.4 5.8 7.8 -25.7 -23.3 -21.1 -19.8 79.2 113.2 -3.2 6.0 -1.0 7.0 -4.9 -3.3 -0.9
46% (72.3) (9.2) (9.2) (18.2) (6.1) (4.0) (0.5) (8.7) (3.3) (10.7) (20.4) (21.0) (10.7) (31.9) (52.3) (4.7) (4.7) (0.5) (2.7) (2.0) (1.4) (0.2)
The table shows estimated coecients and t-values (between parentheses) of regressions of absolute pricing errors on dummy variables for deal type (bid or ask), currency (euro or dollar), rating (AAA, AA, A, BBB or BB), maturity (1-year intervals up to 5 years, and an interval from 5 to 10 years) and quote date (6-month periods). For each set of dummies, we set the weighted average of the coecients to 0, where the weight of a coecient equals the sample mean of the corresponding dummy variable.
Section 4.7 Summary
109
4.7
Summary
In this chapter, we empirically investigated the market prices of credit default swaps. We showed that a simple reduced form model priced credit default swaps better than directly comparing bonds yield spreads to default swap premiums. The model worked reasonably well for investment grade issuers, but quite poor in the high yield environment; so, there is denitely room for more empirical research and further model development. Further, we found evidence that government bonds are no longer considered by the markets to be the reference default-free instrument. Swap and/or repo rates have taken over this position. We also showed that bond spreads and default swap premiums are relatively insensitive to changes in the assumed constant recovery rate as long as the hazard function is scaled accordingly.
Chapter 5
Pricing Step-Up Bonds1
5.1
Introduction
European telecom companies have issued rating-triggered step-up coupon bonds in order to compensate bond investors for losses in the event of rating downgrades. Several investment banks (McAdie, Martin and OKane (2000), Fumagalli and Taur en (2001) and Sirinathsingh (2001)) analyzed step-up bonds. Most of the time, they used historical and subjective rating transition probabilities in their valuation and showed the results of their model for only one day. In contrast, we apply the risk-neutral valuation framework of Jarrow, Lando and Turnbull (1997, hereafter JLT) to value the step-up bonds and demonstrate the models results over a long time period. For comparison purposes, we also value step-up bonds using historical transition probabilities and as equivalent plain vanilla bonds, i.e. bonds similar to step-up bonds except for the step-up feature. Moreover, we analyze the protection that the step-up feature oers to investors in two ways new to the literature. First, we test the volatility of a step-up bond versus its equivalent plain vanilla bond. Second, we determine whether the step-up bond oers superior returns over its equivalent plain vanilla bond in case of rating downgrades and negative outlooks. Our results indicate that the market seems to use the JLT model to value step-up bonds that make a single step-up after the rating trigger. On the other hand, step-up bonds that make multiple step-ups seem to be treated as plain vanilla bonds. Also, the JLT model approximates the step-up feature premium always better than the historical method. We further nd that the step-up feature reduces bond price volatility for most of the considered step-up bonds. Finally, for all bonds in our sample, the step-up feature
1
This chapter is a slightly revised version of the paper by Houweling, Mentink and Vorst (2003b).
112
Pricing Step-Up Bonds Chapter 5
does not oer investors positive excess returns in case of a rating downgrade or a change to negative outlook. The coupon of a step-up bond depends on its issuers rating or the rating of the issuers long term debt. If the rating deteriorates and hits a predened level, the step-up feature is triggered and the coupon rises with a predened amount. Depending on the exact specication of the step-up bond, the coupon can rise even more if the rating further deteriorates. For most step-up bonds, the reverse also applies: the coupon is reduced if the rating increases again. Step-up bonds are a relatively new phenomenon in the eurodenominated corporate bond market. Both issuers and investors benet from this step-up feature; see McAdie et al. (2000). Issuers have placed a larger volume of debt at lower prices than would otherwise have prevailed. Investors have proted too, because they will be compensated in case of rating downgrades and because issuers should be more committed to their rating as a downgrade will penalize them directly by higher coupons. We analyze ve step-up bonds that have been issued by three companies: Deutsche Telecom, France Telecom and KPN. These are the only three companies in our data set that have issued both step-up bonds and enough euro-denominated plain vanilla bonds to reliably estimate issuer-specic interest rate curves. Even though we can analyze only a limited number of step-up bonds, this study is worthwhile, because it provides one of the rst empirical tests of the JLT model on rating-sensitive instruments and because the behavior of step-up bonds has not yet been documented in the academic literature. Various investment banks have published about the characteristics and valuation of step-up bonds, for example Lehman Brothers (McAdie et al., 2000), J.P. Morgan (Sirinathsingh, 2001), Schroder Salomon Smith Barney (Fumagalli and Taur en, 2001) and Soci et e General (Turc, 2001). J.P. Morgan used historical transition and default probabilities in their valuation; Lehman Brothers estimated these probabilities subjectively by using their analysts opinions; Schroder Salomon Smith Barney applied both subjective and historical transition probabilities; Soci et e General implemented a JLT-type of model and used risk-neutral probabilities. These studies all showed the results of their analysis for only one day. In contrast, we price the step-up bonds for a longer period (March 2001 February 2002), and implement three pricing methods. Despite the large interest from practitioners, only one academic paper (Conroy, 2000) has been published about the valuation of euro step-up bonds, as far as we know. Like most investment banks, Conroy also valued step-up bonds using historical rating transition probabilities. Risk-neutral valuation models that can be used to price rating-triggered instruments have appeared in the academic literature though. The basis for most of these models is the JLT Markov chain model, which uses a rms credit ratings as an
Section 5.2 Step-Up Bonds
113
indicator of the likelihood of default. Kijima and Komoribayashi (1998) adjusted the JLT model to make it numerically more stable by replacing default probabilities with survival probabilities in the calculation of risk premia. The JLT model has been generalized by Das and Tufano (1996) to incorporate stochastic recovery rates, and by Lando (1998) and Arvanitis, Gregory and Laurent (1999) to make transition and default intensities stochastic and possibly dependent on state variables. Sch onbucher (1999), Bielecki and Rutkowski (2000) and Acharya, Das and Sundaram (2002) embedded the Markov chain in a Heath, Jarrow and Morton (1992) framework with stochastic forward rates. Acharya et al. (2002) also provided an illustration of their model on a stylized step-up bond. In independent work, Lando and Mortensen (2003) also analyzed step-up bonds in the JLT framework. Although their work is similar to our study, they focused on rening the JLT model, while we compare three dierent valuation methods. Further, we analyze the protection that the step-up feature oers to investors. The remainder of this chapter is structured as follows. First, we describe the characteristics of step-up bonds in Section 5.2. Next, in Section 5.3, the JLT model is briey summarized and the risk-neutral valuation of step-up bonds is explained. Further, we describe the valuation methods using historical probabilities and as equivalent plain vanilla bonds. Section 5.4 describes our data set. The results of applying the three valuation methods to the data are given in Section 5.5. We also test whether including step-up features oers the investor sucient protection using both volatility tests and we conduct an event study on rating and outlook changes. Finally, Section 5.6 summarizes the chapter.
5.2
Step-Up Bonds
The coupon of a step-up bond depends on the issuers rating or the rating of its long term debt. If the rating deteriorates and hits a predened level, the step-up coupon is triggered and the coupon rises with a predened number of basis points. Depending on the step-up coupon type, the coupon can rise even more if the rating further deteriorates. For most step-up bonds, the reverse also applies: the coupon is reduced if the rating improves again. This is called the step-down feature. The coupon can never go below the original level at issuance though. Step-up conditions can dier between bonds. The most important discriminating conditions of the step-up bonds are the following: whether the coupon can step up and down or only steps up, whether both Moodys and S&P have to downgrade the issuer or only one of them before the step-up trigger is hit, the timing of the coupon adjustment, the rating-trigger level and the number of basis points of the step-up. Based on McAdie
114
Table 5.1: Step-up bond types.

Type Step-upa Step-downb One-o c And/ord Accruale
A B C
a b c
or and or
next next immediately
Coupon increases if the rating decreases and hits the rating-trigger. Coupon decreases if the rating increases again. Coupon increases only once, even if the rating falls further below the rating-trigger; for bonds that are not one-o, each further decrease in the rating, causes a further increase in the coupon. Determines whether the coupon is adjusted if Moodys and S&P adjust their ratings, or if Moodys or S&P adjusts its rating. Determines the date at which the adjusted coupon starts accruing: immediately following a rating action, or on the next coupon date after the rating action.
et al. (2000) and Marchakitus, Soderberg and Bramley (2001), Table 5.1 denes three types of step-up bonds.
5.3
5.3.1
Model
Rating Transitions
The value of a bond equals the sum of the discounted, expected cash ows. Unlike a plain vanilla bond, a step-up bonds coupons are a function of the issuers rating and therefore we have to model the issuers rating transition process under the equivalent martingale measure. The JLT model provides a suitable framework for this purpose, since it uses a companys credit rating as an indicator of that companys creditworthiness. We follow the setup of the JLT model. We assume that a unique equivalent martingale exists that makes all default-free and defaultable bond prices martingales, after measure Q normalization by the default-free money-market account. Finally, the recovery rate is constant; we follow Jarrow and Turnbull (2000) and Sch onbucher (2000) by assuming that applies to the principal only, not to the coupons. The companys credit rating Rt at day t is modelled as a Markov chain on a nite , defaultstate space S = {1, . . . , K } under the historical probability measure. Under Q free interest rates and ratings are assumed to be independent. The state space S contains all possible ratings, including the minor ratings: state 1 represents the Aaa rating, state 2
Section 5.3 Model
115
Aa1, state 3 Aa2, . . . , state K 1 Caa, and the last state, K , default. It is assumed that default is an absorbing state. Under the Markov property, it holds for all i, j K, t s 0 and ru S, 0 u < s that qij (s, t) = P(Rt = j |Rs = i, Ru = ru , 0 u < s) = P(Rt = j |Rs = i), i.e. the probability of going from rating i to rating j in the period from s to t only depends on the rating Rs at time s and not on the history Ru , 0 u < s, of reaching that rating. We are aware of the limitations of this assumption, as past rating movements do seem to aect future rating transitions, as shown by Nagpal and Bahar (2000); however, it is an assumption common to most theoretical models. In order to value step-up bonds, we need the rating transition process under the risk . JLT started with the observed, historical transition probabilities, for neutral measure Q example from a rating transition matrix from Moodys or S&P, and applied risk premia to transform these into risk-neutral probabilities. We choose to adjust the (T t)-year (t, T ) as follows transition matrix2 Q(t, T ) to get the risk-neutral transition matrix Q (t, T )q (t, T ) for j = K ij q ij (t, T ) = 1 (t, T )(1 q (t, T )) for j = K iK
(5.1)
for some risk premium (t, T ). It follows that the risk premium can be calculated as (t, T ) = 1q Rt ,K (t, T ) . 1 qRt ,K (t, T ) (5.2)
The risk premium for time T is the ratio of the risk-neutral survival probability to the historical survival probability, so that we retain the numerical stability of Kijima and Komoribayashi (1998, hereafter KK). Note that we use cumulative probabilities q ij (t, T ) instead of JLTs and KKs forward probabilities q ij (t, t + 1), i.e. q ij (t, t + 1) = i (t, t + 1)qij (t, t + 1). The advantage of our risk premia is that their calculation is easier, because they do not require matrix inversion, cf. JLTs Equation (16) and KKs Equation (19). Both
Usually, Moodys and S&P report (multiples of) 1-year transition matrices. However, the time between the valuation day and coupon and redemption dates equals almost never exactly (multiples of) one year. Therefore, we adjust the historical transition matrix and make it maturity-dependent using a generator matrix, as described by Israel, Rosenthal and Wei (2001).
2
116
approaches, forward and cumulative, generate the same results though. Also, note that we only use one risk premium for all rating categories. Ideally, we would like to use a separate risk premium for all ratings. However, each euro-denominated telecom issuer does not cover the full rating spectrum, so that we have to resort to one risk premium, derived from the issuers current rating, and apply it to all ratings. For each day, we estimate the issuer-specic survival probability curve of each telecom company. Following Chapter 4, we specify a linear hazard function, assume a recovery rate of 50%, and use the euro zero-coupon swap curve as proxy for the default-free term structure. The parameters of the hazard function are estimated from the market prices of that issuers plain vanilla bonds using non-linear least squares. Given the estimated survival probability curve for a company on a particular day, we calculate its risk premia (5.2) and risk-neutral transition matrices (5.1) for all required maturities. These risk-neutral matrices are used to calculate theoretical values for step-up bonds using the methods of the next section. We also calculate 95% condence bounds for the survival probability curve, and repeat this series of calculations for the upper and lower bound, hence obtaining upper and lower bounds for the step-up bond values as well.
5.3.2
Step-Up Valuation
To determine the theoretical value of a defaultable step-up and step-down bond, we add the following assumption to the set of JLT assumptions mentioned in the previous section: both Moodys and S&P alter their rating of an issuer at the same time.3 In Section 5.4, we show that this is a plausible assumption to make. In our analysis, we apply the Moodys rating actions. The remainder of this section discusses our three methods to value step-up bonds, starting with the JLT model. Consider a step-up and step-down bond with n remaining coupon payments and a face value of 1. The bond issuer makes the j th coupon payment at day tj , j = 1, 2, . . . , n, but only if he has not gone into default before tj . If the rating at tj 1 is equal to r, the coupon payment at tj is equal to cr , r = 1, . . . , K . The coupon payment at t1 depends on the rating at t0 , which we dene as the previous coupon date, or, if there is no previous coupon, the issue date. The step-up bonds principal amount is paid at maturity tn , again only if the issuer has not defaulted before tn . In case the issuer does default before the bond matures, the constant recovery rate of the notional is paid at the default time. Applying the risk-neutral valuation principle to these coupon, principal and recovery cash
Fumagalli and Taur en (2001) assume that a rating action of one agency is followed by the other agency in the next six months.
3
Section 5.3 Model
117
ows, yields
n
JLT
(t, t, c) =
j =1
t 1{ >t } cRt p(t, tj )E j j 1
(5.3)
t 1{ >tn } + E t p(t, )1{ tn } , p(t, tn )E where t is an n-vector with the coupon payment dates, c is a K -vector with the coupon percentages per rating category, p(t, T ) denotes the time-t default-free discount factor for t [X ] denotes the Q -expected value of X given the information at day t and 1{A} time T , E is the indicator function of event A. The rst line of Equation (5.3) contains the coupon payments, the second line the principal payment and the potential recovery payment. For j > 1, the rst risk-neutral expectation in Equation (5.3) can be evaluated as4
K
t 1{ >t } cRt E = j j 1
k=1 K
t ( > tj Rt = k )ck P j 1 t ( > tj |Rt = k )P t (Rt = k )ck P j 1 j 1

k=1 K
= =
k=1
q Rt ,k (t, tj 1 )(1 q k,K (tj 1 , tj ))ck ,
t (A) denotes the risk-neutral probability of event A given the information at day where P t; for j = 1, the coupon amount is already known (because Rt0 is already known), so t 1{ >t } cRt = P t ( > t1 )cRt = (1 q E Rt ,K (t, t1 ))cRt0 . 1 0 0 After evaluating the second and third risk-neutral expectations in Equation (5.3) too, the JLT value of our rating-triggered step-up and step-down bond equals B JLT (t, t, c) = p(t, t1 )(1 q Rt ,K (t, t1 ))cRt0 +
n K
p(t, tj )
j =2 k=1
q Rt ,k (t, tj 1 )(1 q k,K (tj 1 , tj ))ck
+ (5.4)
p(t, tn )(1 q Rt ,K (t, tn )) +

n
p(t, tj )( qRt ,K (t, tj ) q Rt ,K (t, tj 1 )),

j =1
Note that the summation may be reduced from K terms to K 1 terms, since the K th term is zero: t ( > tj Rt P j 1 = K ) = 0.
118
where we follow Chapter 4 by replacing the integral that results from the recovery payment with a numerical approximation with potential default dates equal to the coupon payment dates. Note that the only dierence between Equation (5.4) and the value of a plain vanilla (PV) bond,
n
B P V (t, t,c) =
j =1
p(t, tj )(1 q Rt ,K (t, tj ))c + p(t, tn )(1 q Rt ,K (t, tn )) +

n
(5.5)
p(t, tj )( qRt ,K (t, tj ) q Rt ,K (t, tj 1 )),

j =1
is the part relating to the specic coupon structure of the step-up bond. If we set ck = c for all k , the expected value of the j th coupon payment reduces to that of a plain vanilla bond
K K K
q Rt ,k (t, tj 1 )(1 q k,K (tj 1 , tj ))c = c

k=1 k=1
q Rt ,k (t, tj 1 )
k=1
q Rt ,k (t, tj 1 ) qk,K (tj 1 , tj )
t (Rt = K )) = c(1 P j = c(1 q Rt ,K (t, tj )). As a second valuation method, we treat the step-up bond as a bond that is identical except for the step-up feature. We refer to this bond as the equivalent plain vanilla bond. Just like the step-up bond, the rst coupon of this bond also depends on the rating at t0 , but the remaining coupons are assumed to follow from the current rating, i.e. they are all equal to cRt . The value of the equivalent plain vanilla (EPV) bond thus equals B EP V (t, t, c) = p(t, t1 )(1 q Rt ,K (t, t1 ))cRt0 +
n
p(t, tj )(1 q Rt ,K (t, tj ))cRt +

j =2
p(t, tn )(1 q Rt ,K (t, tn )) +

n
p(t, tj )( qRt ,K (t, tj ) q Rt ,K (t, tj 1 )).

j =1
The dierence with the JLT formula (5.4) is the second line, which no longer includes a summation over possible future ratings, but instead assumes the current rating Rt will
Section 5.4 Data
119
prevail. The EPV formula also strongly resembles the PV formula (5.5), except that the rst coupon may dier from the other coupons. The third method we consider to value step-up bonds, the historical valuation method, is based on the methods investment banks often applied, e.g. McAdie et al. (2000) and Fumagalli and Taur en (2001). This method uses the telecom companys zero-coupon curve to discount expected coupons, where the expectation is calculated with historical transition probabilities rather than risk-neutral probabilities; again, the rst coupon is known. This gives
B (t, t, c) = v (t, t1 )cRt0 +

j =2
v (t, tj )
k=1
qRt ,k (t, tj 1 )ck + v (t, tn ),
where v (t, T ) denotes the issuers time-t discount factor for time T .
5.4
Data
Most euro-denominated step-up bonds have been issued by telecom companies. Analyzing these rating-triggered step-up corporate bonds thus automatically means focusing on the telecom sector. Step-up bonds are an important source of nancing for telecom companies, because at the end of March 2001, step-up bonds accounted for 42% of the market capitalization of the telecom bond market; see Fumagalli and Taur en (2001). We analyze the set of euro-denominated step-up coupon telecom bonds as listed by both Lehman Brothers (McAdie et al., 2000) and J.P. Morgan (Marchakitus et al., 2001). For these European telecom companies, we download the main characteristics and the price time series for all their bonds from Bloomberg on a daily basis for the period from 4 January 1999 to 13 February 2002. We use the Bloomberg Generic (BGN) price. The BGN price is an average of prices that are quoted by a long list of banks and brokers and reects the bid side of London closing. BGN prices are also used to price the Bloomberg/EFFAS government bond indices; see Brown (1994). We make sure that the bonds used in the curve estimation are plain vanilla. We classify a bond as a plain vanilla bond if the bond has no step-up language, no embedded options, is not oating and not convertible. From the downloaded prices, we remove quotes that equal the quote of the preceding day(s) and quote spikes. We seldom have to remove quotes though. Table 5.2 shows the number of all euro-(re)denominated bonds, plain vanilla bonds and euro-denominated step-up coupon bonds of the telecom companies that have
120
Table 5.2: Total number of euro-(re)denominated, quoted plain vanilla and step-up bonds for all telecom companies.
All bonds Quoted, plain vanilla bonds Step-up bonds
British Telecom Group Deutsche Telecom France Telecom KPN Olivetti SPA/Tecnost Telecom Italia
5 29 37 10 12 7
0 14 7 5 0 2
3 2 2 2 4 2
issued step-up bonds. Three issuers, Deutsche Telecom, France Telecom and KPN, have a large number of quoted, plain vanilla bonds compared to the other telecom companies, British Telecom, Olivetti/Tecnost and Telecom Italia. In our analyses, we focus on the step-up bonds of the rst three issuers only, because we need to estimate issuer-specic survival probability curves as described in Section 5.3 and this curve estimation requires a certain number of plain vanilla bonds. These three companies are large corporate bond issuers as they together represent 6.5% of the Lehman Brothers Euro-Aggregate Corporate Bond Index on 31 May 2002. Table 5.3 displays the characteristics of the step-up bonds we will use in our analysis: the step-up type, as dened in Table 5.1, the number of basis points step-up and the rating-trigger level. We restrict ourselves to step-up bonds with step-up and step-down coupons, because from our three telecom issuers just one KPN bond has step-up only language. Valuing a step-up only bond, type C in Table 5.1, diers from a bond with step-up and step-down features, types A and B in Table 5.1, because the coupon of a stepup only bond is path-dependent. This means that if the rating at any time before the coupon date has been below the trigger-level the coupon payment includes the step-up, even after the issuer is upgraded to or even above the pre step-up rating again; this pathdependency characteristic necessitates an other valuation procedure than Equation (5.4). In short, once triggered, a step-up only bond becomes a plain vanilla bond. This is exactly what happened to the KPN step-up only bond seven months after its issuance. Therefore, from the rst coupon date after this rating event, we treat it as a plain vanilla bond and use it in our estimation of the KPN curve. We download the Moodys and S&P rating and outlook history of Deutsche Telecom, France Telecom and KPN from Bloomberg as well. From Moodys, we use the issuers
Section 5.5 Results
121
Table 5.3: Characteristics of the step-up bonds.

Typea Step-upb Triggerc
Deutsche Telecom 2005 B 2010 B France Telecom 2004 A 2008 A KPN 2006 A
a b c
50 50 25 25 37.5
Baa1/BBB+ Baa1/BBB+ Baa1/BBB+ Baa1/BBB+ Baa3/BBB
Type of step-up bond; see Table 5.1. Number of basis points the coupon steps up. Rating-trigger level (Moodys/S&P).
rating and from S&P the rating of long term debt in local currency. Figure 5.1 shows the rating and outlook migrations of Deutsche Telecom, France Telecom and KPN by both Moodys and S&Ps for the sample period. A positive (negative) outlook is denoted by a + () sign. The gures show that the rating dynamics of both rating agencies are very similar. At the end of our period the two agencies assign identical ratings to each of our three telecom companies. We believe that these gures justify our additional modelling assumption in Section 5.3.2. As historical transition matrix, we use Moodys average one-year senior rating transition matrix for corporate bond issuers, estimated over the period from 1983 to 2001; see Cantor, Hamilton and Ou (2002). Finally, euro swap rates are downloaded from Bloomberg. We apply a standard bootstrapping procedure to extract zero-coupon rates and interpolate linearly between the available maturities to get a curve for all required maturities.
5.5
5.5.1
Results
Step-Up Bond Values and Step-Up Premiums
In this section, we analyze the results from the implementation of the framework from Section 5.3 on the ve step-up bonds in Table 5.3. We dene the pricing error of a step-up bond as its market price minus its theoretical value. Specically, PEi bt is the pricing error of bond b at day t, if method i is used to calculate the theoretical value,
122
A3
A2
A1
Ba1
Aa3
Aa2
Aa1
Aaa
A3 A3 A2 A1
A2
A1
Baa3
Baa2
Baa1
Ba1 Ba1 Aa3 Aa2 Aa1 Aaa Baa3 Baa2 Baa1
Aa3
Aa2
Aa1
Aaa
Baa3
Baa2
Baa1
Jan-1999 Feb-1999 Apr-1999 Apr-1999 Feb-1999
Jan-1999
Jan-1999
Feb-1999
Apr-1999
--
May-1999
May-1999
May-1999
Jul-1999 Aug-1999 Sep-1999 Nov-1999 Nov-1999 Dec-1999 Feb-2000 Mar-2000 May-2000 Sep-1999 Aug-1999
Jul-1999
Jul-1999
Aug-1999
Sep-1999
Nov-1999
Dec-1999 Feb-2000 Mar-2000
Dec-1999
Feb-2000
Mar-2000
May-2000 Jun-2000 Aug-2000 Sep-2000 Nov-2000 Dec-2000 Feb-2001 Mar-2001 May-2001 Jun-2001 Aug-2001 Aug-2000 Sep-2000 Nov-2000 Dec-2000 Jun-2000
May-2000
Deutsche Telecom
France Telecom
Figure 5.1: Credit ratings history.
KPN
Feb-2001 Mar-2001 May-2001 Jun-2001 Aug-2001
Jun-2000
Aug-2000
Sep-2000
Nov-2000
Dec-2000
Feb-2001
Mar-2001
May-2001
- -
Jun-2001
Aug-2001
Sep-2001 Oct-2001 Dec-2001 Jan-2002 S&P
Sep-2001
Sep-2001 Oct-2001 S&P S&P Moody's Moody's Dec-2001 Jan-2002
Oct-2001 Moody's
Dec-2001
Jan-2002
Section 5.5 Results
123
i {EP V, JLT, H }. Further, APEi bt denotes the absolute pricing error for the three
i methods. The sample means of these six statistics are denoted by MPEi b and MAPEb .
A negative (positive) sign of an MPE statistic indicates that theoretical values are, on average, too large (small). To test if this over- or underestimation of step-up bond prices is signicant for method i, we apply a one-sample Z-test (see e.g. Arnold, 1990, Chapter 11)
i Zb
MPEi b Nb , i Sb
i where Sb is the sample standard deviation of the PEi bt series and Nb is the sample size i for bond b. Asymptotically, Zb has a standard normal distribution. Similarly, in order to
determine if signicant performance dierences exist between the three methods, we also use a paired Z-test ; see Arnold (1990, Chapter 11). This test tells us whether two time series have the same mean, while allowing for non-zero correlation and unequal variances. The test statistic to compare measures i and j is dened as
ij Zb
Mbij Nb ij , Sb
ij where Mbij and Sb are the sample mean and sample standard deviation, respectively, of ij j ij Dbt := APEi bt APEbt , i, j {EP V, JLT, H }, i = j , t = 1, . . . , Nb . Asymptotically, Zb
also has a standard normal distribution. Table 5.4 shows the MPE and MAPE statistics for the telecom step-up bond values generated by the three valuation methods and, for comparison, also the MAPE values for the plain vanilla bonds that are used in the estimation of the issuer-specic survival probability curves.5 The paired Z-tests are presented in Table 5.5 for all combinations of the three valuation methods. Further, as an alternative test of the performance of our three methods, we calculate in Table 5.6 the percentage of days that the market price of a step-up bond lies between the 95% condence bounds of the calculated theoretical values. This tells us how uncertainty in the estimated survival probability curves translates into uncertainty in the calculated theoretical step-up bond values. We also zoom in on the value of the step-up feature. We dene the market premium of the step-up feature as the step-up bonds market price minus the value of the equivalent plain vanilla bond. Similarly, we calculate the JLT premium (respectively the historical premium ) of the step-up feature as the JLT (respectively the historical) value of the stepup bond minus the value of the equivalent plain vanilla bond. By subtracting the value of
Note that per denition MPE for the plain vanilla bonds is zero, since the pricing error for a plain vanilla bond is simply its residual from the least squares estimation of the survival probability curve.
5
124
Table 5.4: Pricing errors.

Plain vanillas MAPE Step-ups JLT MPE MAPE
EPV MPE MAPE
Historical MPE MAPE
Deutsche Telecom plain vanilla 0.29 2005 2010 France Telecom plain vanilla 0.64 2004 2008 KPN plain vanilla 1.09 2006
0.15 1.05
0.24 1.15
-0.09 0.20
0.25 0.82
-0.29 -0.34
0.36 0.89
-0.26 -0.87
0.33 0.94
-0.41 -1.99
0.46 1.99
-0.56 -2.57
0.57 2.57
-0.44
1.13
-0.95
1.42
-2.28
2.29
Mean absolute pricing errors (MAPE) of the plain vanilla bonds and both mean pricing errors (MPE) and MAPEs of the step-up bonds for the equivalent plain vanilla (EPV), Jarrow, Lando and Turnbull (1997, JLT) and historical valuation methods. Indicates signicance of the one-sample Z-test at a 95% condence level.
the equivalent plain vanilla bond, we correct the market, JLT and historical values for all bond characteristics except for the step-up feature. So, we assume that the step-up feature fully determines the remainder and no other factors are of importance. In addition to the premium, the 95% condence interval around the premium is calculated using the condence bounds of the equivalent plain vanilla bond. Similar to the pricing errors introduced for the step-up bond prices, we look at the dierences between the market step-up premium and the JLT and historical premiums. Since all three gures contain the value of the equivalent plain vanilla bond as correction term, the last four columns in Table 5.4 and the last column in Table 5.5 also apply to the step-up premium. Yet Table 5.6 does not pertain to the step-up premium; the coverage percentages for step-up premiums are very similar to those for the step-up bond prices, are therefore omitted. Deutsche Telecom We now discuss the results for each telecom company, beginning with Deutsche Telecom. The MPE values in Table 5.4 tell that the EPV method typically undervalues and the JLT and historical methods usually overvalue the 2005-bond. The JLT model has the smallest bias to the market price, since both MPEEP V and MPEH dier more from zero
Section 5.5 Results
125
Table 5.5: Paired Z-tests.

EPV JLT EPV Historical JLT Historical
Deutsche Telecom 2005 0.00 2010 0.33 France Telecom 2004 -0.13 2008 -1.04 KPN 2006 -0.30
-0.12 0.26 -0.24 -1.63 -1.16
-0.12 -0.07 -0.11 -0.58 -0.86
Pairwise dierences between mean absolute pricing errors of the equivalent plain vanilla (EPV), Jarrow, Lando and Turnbull (1997, JLT) and historical values for the step-up bonds. Indicates signicance of the paired-sample Z-test at a 95% condence level.
Table 5.6: Condence interval coverage percentages.

EPV JLT Historical
Deutsche Telecom 2005 73% 2010 33% France Telecom 2004 90% 2008 48% KPN 2006 82%
85% 68% 85% 0% 76%
62% 50% 72% 6% 33%
Percentages of market step-up bond prices that lie between the 95% upper and lower bounds of the equivalent plain vanilla (EPV), Jarrow, Lando and Turnbull (1997, JLT), and historical values.
126
than MPEJLT , despite that all three MPEs are statistically dierent from zero. The percentages in Table 5.6 also show that the JLT model better approximates the stepup market price than the EPV and historical methods. Although the bond values vary between methods, the paired Z-test in Table 5.5 shows that these dierences are small and in case of the dierence between the EPV and JLT methods not even statistically signicant. Figure 5.2a shows the market, JLT and historical step-up premiums and the 95% condence bounds around the market premium for the 2005-Deutsche Telecom bond.6 Both the JLT and historical premiums are positive and more or less constant, except for a small increase caused by Moodys downgrade of Deutsche Telecom on 13 June 2001; the historical premiums are larger than the JLT premiums. In contrast, the market premiums are much more volatile and sometimes even negative.7 A negative market premium for the step-up feature is counterintuitive: as long as the step-up has not been triggered yet, the coupon cannot decrease and hence the step-up feature must have a positive value. From the condence bounds, however, if follows that the negative market premiums are not statistically signicant. The market premium uctuates mostly between 0 and 1, except in August 2001, where it is above 1 and in November and December 2001, where it is below 0. For the 2010-bond, the MPE values in Table 5.4 for the JLT and historical methods are relatively small, whereas for the EPV method the MPE value is much larger. Further, all three valuation methods give statistically signicant dierent values at a condence level of 95% according to Table 5.5. Table 5.6 also displays that the highest coverage percentages are generated by the JLT model. The market step-up premium of the 2010-bond displays an even more volatile pattern than of the 2005-bond; see Figure 5.2b. This makes sense, because due to its longer maturity, more coupons are aected by the step-up language. During the rst half of the analyzed period, until October 2001, the market premium is larger than the JLT and historical premiums, later this pattern is reversed. The market step-up premium becomes negative in December 2001 and January 2002, but like above, the negative market premium is not signicant. Both JLT and historical premiums remain roughly constant during the entire period.
Due to a lack of plain vanilla Deutsche Telecom bonds prior to 14 May 2001, its gures start well after the step-up bonds issue date 6 July 2000. 7 Lehman Brothers (McAdie et al., 2000) also found a negative value in their analysis of euro step-up bonds on 13 July 2000.
Section 5.5 Results
127
Figure 5.2: Deutsche Telecom step-up premiums.

4
historical JLT
3
market 95% confidence bounds market value
-1
-2 May-2001
Jun-2001
Jul-2001
Aug-2001
Sep-2001
Oct-2001
Nov-2001
Dec-2001
Jan-2002
(a)
5
historical JLT
4
market 95% confidence bounds market value
-1
-2 May-2001
Jun-2001
Jul-2001
Aug-2001
Sep-2001
Oct-2001
Nov-2001
Dec-2001
Jan-2002
(b)
The market, JLT and historical step-up premium of the (a) 2005- and (b) 2010-Deutsche Telecom step-up bonds. The dotted lines are the 95% condence bounds of the market premium. The vertical line indicates a rating downgrade by Moodys.
128
France Telecom Next, we look at the two France Telecom step-up bonds. For the 2004-bond, the MPE and MAPE values for the EPV method in Table 5.4 approximate the market prices better than the JLT and historical methods. Apparently, the market values this bond as a plain vanilla bond. Also, from Table 5.6, we observe that the 95% condence bounds of the EPV values enclose the market prices in 90% of the days, and the JLT and historical methods only 85% and 72%, respectively. Again, Table 5.5 exhibits that the approximations produced by the three valuation methods dier signicantly. Figure 5.3a displays the market step-up premium with its 95% condence bounds and the JLT and historical premiums. The same pattern as with Deutsche Telecom emerges, with volatile, though insignicant, market premiums and steady JLT and historical premiums. Also, there is a large drop in both theoretical premiums on 26 September 2001 after the downgrade by Moodys from A3 to Baa1. This downgrade triggered the stepup feature (see Table 5.3), so that the remaining number of step-ups is lower than the number of step-ups at issuance and the coupon can also step down as the rating improves again. The consequence is a decrease in the theoretical step-up premiums, as observed in Figure 5.3a. For the 2008-France Telecom bond, the MPE value is (in absolute value) equal to the MAPE value for both the JLT and historical methods; see Table 5.4. This means that both methods always overestimate the market price. The MPE value for the EPV method is the smallest, but there is still a large bias. The condence bounds statistics in Table 5.6 reveal that in 48% of the days the market price and EPV value are statistically indistinguishable, while the condence bounds of the JLT and historical methods almost never include the market price. The paired Z-tests also indicate that the JLT and historical values do not vary much, but that they both dier to a great extent from the EPV values. As Figure 5.3b shows, the market premium of the step-up feature of the 2008-bond is again volatile and is even signicantly negative during the rst few months and also after the downgrade. In contrast, both the JLT and historical premiums remain positive during the full period, also after the downgrade, where their values drop as well. As above, we observe that the JLT and historical premiums are typically larger than the market premium and this dierence is quite large, because they virtually never lie within the market premiums condence bounds.
Section 5.5 Results
129
Figure 5.3: France Telecom step-up premiums.

2
historical JLT market

1
95% confidence bounds market value
-1
-2
-3 Mar-2001
Apr-2001
May-2001
Jun-2001
Jul-2001
Aug-2001
Sep-2001
Oct-2001
Nov-2001
Dec-2001
Jan-2002
(a)
4
historical
3
JLT market
-1
-2
-3
-4 Mar-2001
Apr-2001
May-2001
Jun-2001
Jul-2001
Aug-2001
Sep-2001
Oct-2001
Nov-2001
Dec-2001
Jan-2002
(b)
The market, JLT and historical step-up premium of the (a) 2004- and (b) 2008-France Telecom step-up bonds. The dotted lines are the 95% condence bounds of the market premium. The vertical line indicates a rating downgrade by Moodys.
130
KPN Finally, we analyze the 2006-KPN step-up bond. Again the MPE value of the EPV method is closest to zero, though it is signicantly smaller than zero; see Table 5.4. The JLT model also typically overvalues the market price, and the historical method even more, as evidenced by their negative MPE values. The same ranking of the three methods emerges from Table 5.6, since in 82% of the cases the market price lies within the EPV condence bounds, while for the JLT and historical methods this is 76% and 33%. Therefore, the market seems to value this bond as a plain vanilla bond. Finally, we calculate the market, JLT and historical step-up premiums; see Figure 5.4. Similar to the two France Telecom bonds above, the JLT and historical premiums are less volatile and higher than the market premium, but the premiums of the JLT model often lie within the 95% condence bounds. The drops in all premiums on 6 September 2001 follow from Moodys rating downgrade from Baa2 to Baa3. The market premium rst shows a dramatic fall and an impressive recovery in the subsequent months after this downgrade and becomes signicantly positive and then falls for the second time and becomes negative. Also the JLT premium drops after the downgrade and becomes virtually equal to zero. A similar pattern can be observed for the historical method, since it drops following the downgrade too; however, unlike the JLT premium, it stays positive. Again, the historical premium is always larger than the market premium, something we also noticed for the Deutsche Telecom and France Telecom step-up bonds. Recovery Rate The step-up bond values constructed by the EPV, JLT and historical methods all use a recovery rate of 50%. However, the true recovery rate for these three Telecom issuers is not known exactly, in fact this is the case for all corporate issuers that have not experienced default. So, we analyze here how sensitive the mean absolute pricing errors (MAPEs) are to variations in the recovery rate in the range from 30% to 70%. Figure 5.5 shows that for the 2005-Deutsche Telecom bond, the MAPE values produced by the historical and EPV valuation methods always increase whereas the MAPEs associated with the JLT valuation method always decrease as the recovery rate increases. In contrast, for the 2010-Deutsche Telecom bond, both MAPEs and recovery rate always move in the same direction, the same applies for the 2004-France Telecom bond. The MAPEs generated by the historical valuation method rise as the recovery rate becomes larger for both the 2008-France Telecom and the 2006-KPN bond. The opposite pattern can be observed for
Section 5.5 Results
131
Figure 5.4: KPN step-up premiums.

2
historical JLT market

1
-1
-2
-3 Mar-2001
Apr-2001
May-2001
Jun-2001
Jul-2001
Aug-2001
Sep-2001
Oct-2001
Nov-2001
Dec-2001
Jan-2002
The market, JLT and historical step-up premium of the 2006-KPN step-up bonds. The dotted lines are the 95% condence bounds of the market premium. The vertical line indicates a rating downgrade by Moodys.
the MAPEs associated with the JLT and EPV valuation methods, with the exception of the MAPEs of the 2006-KPN bond that increases again at the 70% recovery rate. We can observe that for three of the ve step-up bonds, i.e. the 2010-Deutsche Telecom and both France Telecom bonds, the order of the valuation methods is always the same, irrespective of the value of the recovery rate. For the other two bonds, the 2005-Deutsche Telecom bond and the 2006-KPN bond, the historical method always performs worst, but the JLT and EPV methods cross at a recovery rate of 50% and 70%, respectively; for smaller values, the EPV method is better, whereas for larger values the JLT method yields smaller errors.
5.5.2
Step-Up Protection
Volatility Analysis We now analyze the protection a step-up provision oers to an investor. We hypothesize that the price volatility of the step-up bond is lower than the volatility of its equivalent plain vanilla bond, because the step-up feature compensates a lower rating with a higher
132
Figure 5.5: Recovery rate sensitivity analysis.

0.45 0.40 0.35 0.30 0.25 0.20 0.15 30 40 50
(a)
1.5 1.4 1.3 1.2 1.1 1.0 0.9 0.8 0.7

60 70
30
40
50
(b)
60
70
0.8 0.7 0.6
3.5 3.0 2.5 2.0
0.5 1.5 0.4 0.3 30 40 50

(c)
1.0 0.5 60 70 30 40 50
(d)
60
70
3.0
2.5
2.0
1.5
1.0 30 40 50
(e)
60
70
The graphs depict the mean absolute pricing errors for various recovery rates, for the equivalent plain vanilla bond (thick grey line), JLT (thick black line) and historical (thin black line) method respectively, for all step-up bonds: (a) 2005-Deutsche Telecom, (b) 2010-Deutsche Telecom, (c) 2004France Telecom, (d) 2008-France Telecom, and (e) 2006-KPN.
Section 5.5 Results
133
Table 5.7: Volatility analysis.

Test statistic p-value
Deutsche Telecom 2005 0.72 2010 0.56 France Telecom 2004 1.04 2008 0.75 KPN 2006 1.00
0.025 0.000 0.768 0.031 1.000
F-statistics and p-values for the test that the step-up bonds and their equivalent plain vanilla bonds have equal variances.
coupon, and vice versa; see also McAdie et al. (2000).8 So, the step-up feature should work as a cushion against rating migrations.
2 We compare the variance of a step-up bonds market prices, step up , with the variance 2 of its equivalent plain vanilla bonds values, EP V . The null hypothesis 2 step up = 1, 2 EP V
H0 : is tested using the test statistic
V =
2 Sstep up , 2 SEP V
2 2 where Sstep up and SEP V are the sample variances of the step-up bond market prices
and equivalent plain vanilla bond values, respectively. V follows an F -distribution with nstepup 1 and nEP V 1 degrees of freedom, where ni equals the number of i-observations, with i {step-up, EPV}; see e.g. Madsen and Moeschenberger (1986). Table 5.7 shows that the outcome of this test statistic is equivocal. The null hypothesis is easily rejected at the 95% condence level for both the 2005- and 2010-Deutsche
Also, Olivetti/Tecnosts Chief Financial Ocer, after linking the coupons of its bonds to its credit rating, stated: we think having these sort of volatility protection measures associated with our bonds should result in a lower capital cost (Bloomberg Equity News, June 16, 2000, quoted in Acharya et al. (2002, footnote 9)).
8
134
Telecom bonds and the 2008-France Telecom bond. So, for these three bonds the step-up bond price variance is lower than the variance of the equivalent plain vanilla bond.9 For the 2004-France Telecom and the 2006-KPN bonds, the two variances are statistically indistinguishable. Event Analysis Next, we carry out an event analysis to test the protection a step-up bond oers in case of rating (outlook) downgrades. As long as the timing and/or size of a rating downgrade or negative outlook change are not fully anticipated by the market, plain vanilla bonds have a negative return on the event date; see e.g. Hand, Holthausen and Leftwich (1992). Since a step-up bond compensates investors for the decreased creditworthiness via a higher coupon, we hypothesize that it has a higher return than its equivalent plain vanilla bond. To test this hypothesis, we dene the excess step-up return as the step-up bond market return minus the return of its equivalent plain vanilla bond; so, the step-up bond return is fully corrected for all bond characteristics except for the step-up language. We rst calculate the excess return ERit for bond i for the event day, t = 0, and the three succeeding trading days, t = 1, 2, 3. Then, we calculate the average ERt of these excess returns for each day over the ve step-up bonds. We also consider cumulative excess returns, dened as
t
CERit =
s=0
ERis ,
and their averages CERt . We test the signicance of the average excess returns and the cumulative average excess returns using simple t-tests as described in Ritter (1991). If both Moodys and S&Ps change their rating (outlook) of the same company at the same time, we treat this as a single event. Furthermore, if a rating action occurs within three trading days of an earlier rating action, we delete this rating action from the analysis. Table 5.8 shows the results of this event analysis. On the event date, the return on the EPV bond is indeed negative, as expected. However, the return on the step-up bond is even more negative, so that the excess step-up return is negative. Although this negative excess return is not signicant, it does not conform with our expectations. During the post event period, the average excess return becomes positive, but stays statistically
If the step-up bonds were less liquid than the plain vanilla bonds, then this outcome could be explained by the occurrence of stale prices. However, the step-up bonds are probably more liquid than the plain vanilla bonds, because they are younger and have larger notionals than the latter; see Chapter 3 and the references therein.
9
Section 5.6 Summary
135
Table 5.8: Event analysis.

Returns EPV Excess Cumulative returns EPV Excess
Day
Step-up
t-value
Step-up
t-value
0 1 2 3
-1.89% 0.36% -0.11% 0.16%
-0.77% -0.14% -0.61% 0.18%
-1.12% 0.50% 0.50% -0.02%
-1.14 0.65 1.72 -0.15
-1.53% -1.64% -1.47%
-0.90% -1.51% -1.32%
-0.63% -0.13% -0.15%
-0.97 -0.20 -0.23
Average (cumulative) returns for the step-up bonds and their equivalent plain vanilla (EPV) bonds, as well as the excess returns of the former over the latter and their t-values, after rating downgrades and negative outlook changes, for the event date and the three succeeding days.
insignicant. The cumulative excess returns are always negative and insignicant for the whole post event period. Therefore, contrary to our expectations, the step-up feature does not oer investors positive excess returns in case of a rating downgrade or a negative outlook, but instead oers statistically identical returns as the EPV bond.
5.6
Summary
In this chapter, we have empirically compared several pricing methods for rating-triggered step-up coupon bonds. European telecom companies have issued these bonds in order to compensate bond investors for losses in the event of rating downgrades. The coupon of such a step-up bond depends on its issuer rating or the rating of the issuers long term debt. If this rating deteriorates and hits a predened level, the step-up coupon is triggered and the coupon rises with a predened number of basis points. We applied risk-neutral transition probabilities using the Jarrow, Lando and Turnbull (1997, JLT) framework to value these rating-triggered step-up bonds. For comparison purposes, we also valued stepup bonds using historical probabilities and as plain vanilla bonds comparable to step-up bonds except for the step-up feature. Next, we tested the volatility of a step-up bond versus the equivalent plain vanilla bond and we performed a rating and outlook change event analysis of excess step-up bond returns. We found that the market seemed to value the Deutsche Telecom step-up bonds, whose coupons make a single step-up after the rating hits the trigger level, according to the JLT model. On the other hand, for the France Telecom and KPN step-up bonds,
136
whose coupons step up every time a rating hits a trigger level, the market seemed to resort to valuation as plain vanilla bonds. Further, we found that the market premiums of the step-up feature were much more volatile than the JLT and historical premiums. These theoretical premiums were stable, except for changes caused by downgrades. Further, the JLT model approximated the market premiums always better than the historical valuation method. From the ve step-up bonds, the two Deutsche Telecom bonds and the 2008-France Telecom bond had a signicantly lower price volatility than their equivalent plain vanilla bonds at a 95% condence level. So, these three step-up bonds oered protection in terms of a lower price volatility. The volatilities of the 2004-France Telecom and the KPN step-up bond were statistically indistinguishable from their equivalent plain vanilla bonds. Therefore, these step-up bonds did not oer a cushion against rating migrations in the form of dampened price movements. The results from the event analysis, another way of looking at the step-up protection, demonstrated that step-up bonds did not oer superior returns to an investor in case of rating downgrades or negative outlooks.
Chapter 6
Summary
This thesis contains four empirical studies on credit markets. Chapter 1 listed several developments in the 1990s that stimulated research on credit markets, where such research had been virtually absent in the earlier decades. Both academics and practitioners required a better understanding of the risks that drive credit markets. They also needed more sophisticated models to price and hedge the newly developed credit derivatives. While the literature initially brought about primarily theoretical studies, in recent years empirical papers started to appear too. This thesis contributes to this growing empirical literature by presenting four studies on the issues of estimating spread curves, measuring corporate bond liquidity, pricing default swaps and pricing step-up bonds. In this concluding chapter, the main ndings of these studies are summarized. Chapter 2 presented a new framework for the joint estimation of the default-free term structure and corporate spread curves. Accurately estimated spread curves are important inputs for various applications, like pricing bonds and credit derivatives and calculating risk measures such as Value-at-Risk. In Chapter 2, it was illustrated that the traditional way of calculating spread curves, i.e. by subtracting independently estimated default-free and defaultable term structures, results in unrealistically twisting spread curves. This can have important consequences for subsequent calculations with these curves. The proposed model was designed to overcome these problems. It decomposed a defaultable term structure into a default-free part and a spread part. The default-free curve was estimated from government bonds, so that the model for the corporate term structure could focus on the spread curve only and could thus be parsimonious. Both default-free and spread curves were modelled using B-splines, a exible way of tting a continuous and smooth function to the data without a priori imposing a particular curvature. The results indicated that the model indeed yielded more realistic and smoother spread curves.
138
Summary Chapter 6
Moreover, the reliability of the estimated curves was greatly improved. In spite of the smaller number of parameters, the new model could still accurately t market data. Chapter 3 empirically compared several measures of corporate bond liquidity. Investors need to know which bonds are illiquid and which ones are liquid, because selling an illiquid bond before its maturity is more costly, due to higher bid-ask spreads and order processing costs, than for its liquid counterpart. Since not only liquidity, but also other characteristics aect bond yields, it is important to accurately control for these characteristics in identifying liquidity premia. Chapter 3 found ve signicantly priced characteristics for a data set of euro-denominated corporate bonds. After controlling for these ve variables, seven out of eight considered liquidity measures were found to generate statistically signicant liquidity spreads. The liquidity premium ranged from 9 to 24 basis points and was the largest for the measures age and yield dispersion (see Table 3.2 on page 62 for a description of the liquidity measures). By conducting a series of pairwise comparison tests, it was found that the liquidity measures price volatility and number of contributors ranked favorably compared to the other measures. Chapter 4 conducted an empirical analysis on the pricing of credit default swaps. Credit derivatives are relatively new instruments on nancial markets, and have experienced tremendous growth over the last years. They can be used to take on or lay o credit risk in a exible way. The most popular credit derivatives are default swaps, insurancelike instruments that oer protection against the default of the underlying borrower. So, understanding how they work and how they are priced is important. Chapter 4 implemented a basic reduced form model with a deterministic hazard process and a constant recovery of face value assumption. The model was estimated to market data of bonds and subsequently used to calculate theoretical default swap premia. For comparison, it also applied a more simple method, namely using a bonds yield spread as a direct estimate of the default swap premium. Attention was paid to the practical implementation of the model by considering several alternatives for the choice of the default-free term structure and by testing the robustness with respect to the assumed recovery rate. The results indicated that the reduced form model outperformed the direct comparison method. The model worked well for investment grade issuers, but only if swap or repo rates were used as proxy for default-free interest rates. For speculative grade issuers, considerable deviations were found between market and model premiums. Moreover, both bond prices and default swap premiums were shown to be relatively insensitive to the choice of the recovery rate, as long as it takes a reasonable value. Chapter 5 studied the pricing of rating-triggered step-up bonds, a xed-income bond with a built-in credit derivative. The coupon of a step-up bond depends on the credit
139
rating of its issuer: if the rating deteriorates and hits a predened level, the step-up feature is triggered and the coupon rises with a predened number of basis points. In recent years, European telecom companies extensively used step-up bonds to raise capital, because due to nancial distress they had diculty borrowing money without special provisions. The pricing of step-up bonds allows an interesting test of rating-based reduced form credit risk models. This task was undertaken in Chapter 5, which implemented an important representative of such models, the Jarrow-Lando-Turnbull (JLT) model. For comparison, it also valued step-up bonds using a similar framework with historical transition probabilities and as plain vanilla bonds. The results indicated that the market seemed to value single step-up bonds according to the JLT model, while it valued multiple step-up bonds as plain vanilla bonds. Further, most considered step-up bonds oered a cushion against rating migrations via dampened price movements, but did not oer better returns than plain vanilla bonds. The results of the empirical studies conducted in this thesis provide useful guidance for future research on credit markets. The rst two studies oer a robust framework for the estimation of spread curves, and identify readily implementable measures of bond liquidity. The last two studies provide an empirical basis for the application of reduced form credit risk models to the pricing of credit derivatives.
Nederlandse samenvatting (Summary in Dutch)
Inleiding
In dit proefschrift staan obligaties en kredietderivaten centraal. Gedurende de laatste tien jaar vonden diverse ontwikkelingen en gebeurtenissen plaats waardoor onderzoek naar deze instrumenten volop in de belangstelling kwam te staan. Voorbeelden van dergelijke ontwikkelingen zijn de vorming van de Europese Monetaire Unie, de liberalisering van Europese kapitaalmarkten, de toenemende uitgifte van bedrijfsobligaties en de opkomst van kredietderivaten; enkele dramatische gebeurtenissen waren bijvoorbeeld de nanci ele crises in Azi e, Zuid-Amerika en Rusland, de enorme verliezen van Long Term Capital Management en de faillissementen van grote bedrijven als Enron, KPN Qwest en Worldcom. Hierdoor ontstond zowel in de academische als in de nanci ele wereld de behoefte om de risicos van het beleggen in kredieten beter te begrijpen. Bovendien waren nieuwe modellen nodig voor de waardering van kredietderivaten. In eerste instantie werden vooral theoretische artikelen naar deze onderwerpen gepubliceerd, maar na verloop van tijd verschenen ook empirische studies. Dit proefschrift levert met vier studies een bijdrage aan de groeiende empirische literatuur op het gebied van kredietrisico, liquiditeitsrisico en kredietderivaten. De belangrijkste bevindingen van deze studies worden in de rest van deze samenvatting beschreven. Eerst worden echter enkele belangrijke begrippen ge ntroduceerd. Een belegger in kredieten loopt meer risico dan een belegger in staatsobligaties. De vergoeding die een belegger voor dit extra risico krijgt wordt de (yield) spread genoemd. De belangrijkste componenten van de spread zijn vergoedingen voor kredietrisico en liquiditeitsrisico. Kredietrisico heeft betrekking op de mogelijkheid dat een entiteit niet in staat zal zijn haar contractuele verplichtingen na te komen. Kredietrisico wordt vaak
142
Nederlandse samenvatting
opgesplitst in twee componenten: faillissementsrisico is de onzekerheid over het wel of niet in staat zijn om aan de verplichtingen te voldoen; verliesrisico is de onzekerheid over de omvang van het verlies dat geleden zal worden in geval van faillissement. De mate van kredietrisico van een bedrijf of land wordt weerspiegeld in haar rating : een subjectief oordeel van een ratingbureau over de kredietwaardigheid van een debiteur. Liquiditeitsrisico is de mogelijkheid dat een belegger niet in staat zal zijn binnen de gewenste tijd en in de gewenste hoeveelheid een instrument te kopen of te verkopen met een beperkte invloed op de prijzen. Uit onderzoek is gebleken dat een illiquide instrument een lagere prijs, hogere transactiekosten en een groter verschil tussen bied- en laatprijzen heeft dan een vergelijkbaar, liquide instrument. De hoofdstukken in dit proefschrift analyseren marktprijzen van obligaties en kredietderivaten. Een obligatie is feitelijk een lening, waarbij de emittent bij uitgifte het geleende bedrag van de belegger krijgt en belooft dit in de toekomst terug te betalen, vaak met tussentijdse rentebetalingen (coupons ). In tegenstelling tot een lening, die vast staat op de balans van een bank, kan een obligatie dagelijks worden verhandeld op de kapitaalmarkt. Dit levert waardevolle datasets van marktprijzen op, die gebruikt kunnen worden voor onderzoek. Steeds meer bedrijven gaan ertoe over geld te lenen door het uitgeven van obligaties in plaats van door het aangaan van een lening bij een bank. Een kredietderivaat is een instrument dat kredietrisico kan overdragen zonder dat de onderliggende kredieten van eigenaar verwisselen. De uitbetaling van een kredietderivaat kan afhangen van het optreden van een faillissement of ratingverandering of van de uitbetalingen of prijsveranderingen van een krediet. Kredietderivaten kunnen worden gebruikt om meer of juist minder kredietrisico te lopen. Doordat de onderliggende schuld niet van eigenaar hoeft te verwisselen, vindt deze overbrenging van kredietrisico bij kredietderivaten eci enter plaats dan bij obligaties of leningen. De markt voor kredietderivaten is de laatste jaren enorm gegroeid met jaarlijkse groeipercentages van minstens 50%.
Het schatten van spread curves

Hoofdstuk 2 beschrijft en implementeert een nieuwe methode voor het schatten van yield spreads. Spreads zijn van belang bij de waardering van obligaties en kredietderivaten en bij de berekening van risicomaatstaven. Onnauwkeurigheden of fouten in de geschatte spreads kunnen leiden tot incorrecte prijzen of risicomaatstaven. De traditionele manier om spread curves te berekenen is om individueel geschatte risicovrije en risicovolle termijnstructuren van elkaar af te trekken. In hoofdstuk 2 wordt ge llustreerd dat dit leidt tot onrealistisch slingerende spread curves.
Summary in Dutch
143
Als alternatief voor deze oude manier wordt een nieuwe methode voorgesteld die risicovrije en risicovolle termijnstructuren gezamenlijk schat. De nieuwe methode is gebaseerd op het idee dat een risicovolle termijnstructuur bestaat uit een risicovrij gedeelte en een spread gedeelte. De risicovrije curve wordt geschat op basis van marktprijzen van staatsobligaties, zodat het model voor de risicovolle curve slechts de spread hoeft te beschrijven. Dit laatste model kan daarom volstaan met een beperkt aantal parameters. De oude en de nieuwe methode worden beide toegepast op een dataset van staats- en bedrijfsobligaties. De resultaten laten zien dat het nieuwe model inderdaad realistischere en gladdere spread curves oplevert. Bovendien is de statistische betrouwbaarheid van de schattingen sterk toegenomen. Ook geeft het model, ondanks het kleinere aantal parameters, nog steeds een goede beschrijving van de obligatieprijzen.
Het meten van de liquiditeit van bedrijfsobligaties

Hoofdstuk 3 gaat over de liquiditeit van bedrijfsobligaties. Als een belegger een obligatie voor het einde van de looptijd wil verkopen, is het belangrijk om te weten of deze liquide of illiquide is, omdat de verkoop van een illiquide obligatie kostbaarder is dan van een vergelijkbare, liquide obligatie. Bedrijfsobligaties worden veelal bilateraal verhandeld, en niet op beurzen, zodat directe liquiditeitsmaatstaven als de verhandelde hoeveelheid en het aantal transacties niet beschikbaar zijn. In de literatuur vinden we daarom veel indirecte liquiditeitsmaatstaven die gebaseerd zijn op de kenmerken of marktprijzen van een obligatie. In hoofdstuk 3 worden acht indirecte maatstaven ge mplementeerd voor een dataset van in euros uitgegeven bedrijfsobligaties. Yields van bedrijfsobligaties worden echter niet alleen bepaald door liquiditeit, maar ook door andere kenmerken zoals de kredietwaardigheid van de emittent en de resterende looptijd van de obligatie. Uit de resultaten van het onderzoek blijkt dat vijf kenmerken van belang zijn. Na correctie van de yields voor deze vijf kenmerken, kan de liquiditeitspremie worden vastgesteld. Voor zeven van de acht beschouwde liquiditeitsmaatstaven wordt een statistisch signicante liquiditeitspremie gevonden. De grootte van de premie verschilt per maatstaf en varieert tussen de 9 en 24 basispunten. De grootste premies worden gevonden voor de volgende twee maatstaven: de leeftijd van een obligatie en de mate waarin door verschillende partijen afgegeven prijzen van elkaar afwijken. Vervolgens wordt een paarsgewijze vergelijking uitgevoerd om de relatieve kracht van de maatstaven te bepalen. Hierbij komen twee maatstaven gunstig naar voren: de prijsvolatiliteit en het aantal partijen dat een prijs afgeeft.
144
Nederlandse samenvatting
Het waarderen van default swaps

Hoofdstuk 4 onderzoekt de waardering van default swaps. Een default swap is een kredietderivaat dat lijkt op een verzekering: de koper betaalt periodiek een premie, en de verkoper vergoedt de verliezen op de onderliggende kredieten indien de emittent failliet gaat tijdens de looptijd van het contract. Default swaps zijn het meest verhandelde instrument onder de kredietderivaten. Daarom is het belangrijk te weten hoe ze worden geprijsd in de markt. Verder kunnen met behulp van marktprijzen van default swaps kredietrisicomodellen worden getoetst. In hoofdstuk 4 wordt gebruik gemaakt van een zogenoemd gereduceerd kredietrisicomodel om default swaps te waarderen. De parameters van dit model worden dagelijks geschat op de marktprijzen van obligaties van individuele debiteuren. Vervolgens worden met het geschatte model theoretische default swap premies berekend. Door deze te vergelijken met marktpremies kan een inschatting worden gemaakt van de kwaliteit van het model. Er wordt bijzondere aandacht besteed aan de praktische implementatie van het model door verschillende keuzes te vergelijken voor de risicovrije termijnstructuur en voor het verliespercentage in geval van faillissement. Ter vergelijking wordt ook een eenvoudigere methode bekeken, waarbij de yield spread van een obligatie gebruikt wordt als schatting van de default swap premie. De resultaten van het onderzoek laten zien dat het kredietrisicomodel nauwkeurigere schattingen voor default swap premies oplevert dan de spread-methode. Het model blijkt goed te werken voor emittenten met een hoge kredietwaardigheid, maar alleen als de swap of repo curve wordt gebruikt als risicovrije termijnstructuur. Voor emittenten met een hoog kredietrisico worden grote verschillen gevonden tussen de markten modelpremies.
Het waarderen van step-up obligaties

Hoofdstuk 5 analyseert de waardering van step-up obligaties. De coupon van een stepup obligatie is afhankelijk van de rating van de emittent: als de rating verslechtert en beneden een afgesproken niveau daalt, wordt de coupon met een afgesproken aantal basispunten verhoogd. Bij enkelvoudige step-up obligaties wordt de coupon slechts eenmalig aangepast, maar bij meervoudige step-up obligaties wordt de coupon bij elke verdere verslechtering verder verhoogd. Voor beide types wordt de coupon veelal weer verlaagd als de rating weer verbetert. Step-up obligaties zijn feitelijk een combinatie van een gewone obligatie en een rating-afhankelijke kredietderivaat en vormen daarom een interessante uitdaging voor kredietrisicomodellen.
Summary in Dutch
145
In hoofdstuk 5 worden drie verschillende methoden vergeleken voor de waardering van step-up obligaties: ten eerste, het Jarrow-Lando-Turnbull (JLT) model, het bekendste kredietrisicomodel dat gebruik maakt van ratings; ten tweede, een soortgelijk model, maar met gebruikmaking van historische in plaats van risico-neutrale migratiekansen; en tenslotte worden step-up obligaties gewaardeerd alsof het gewone obligaties zijn. De uitkomsten van het onderzoek laten zien dat de markt meervoudige step-up obligaties tijdens de onderzoeksperiode ziet als gewone obligaties, maar dat enkelvoudige step-up obligaties worden gewaardeerd volgens het JLT model. Verder hebben de meeste stepup obligaties een kleinere prijsvolatiliteit dan vergelijkbare, gewone obligaties, maar ze leveren geen hogere rendementen op.
Author Index
Acharya, V. 12, 113, 133 Adams, K. J. 23 Agrawal, D. 48, 56, 58 Alexander, G. J. 48, 5659 Altman, E. I. 10, 18, 51 Amihud, Y. 47, 49, 5557, 69 Anderson, N. 18 Annaert, J. 46, 48 Anshuman, V. R. 47 Aonuma, K. 83 Arnold, S. F. 95, 123 Arvanitis, A. 12, 113 Aunon-Nerin, D. 76, 90, 92, 94 Bahar, R. 115 Baker, N. L. 47 Bakshi, G. 79, 81 Bayless, M. 48, 56, 59 BBA 5, 7, 9, 77, 90 BCBS 35 Beim, D. O. 23, 24 Bhatia, M. 10, 13 Bielecki, T. R. 12, 113 BIS 7 Black, F. 1, 11, 17 Bliss, R. 20 Boudoukh, J. 47, 48 Bramley, A. 114, 119 Brealey, R. A. 21 Breedon, F. 18 Brennan, M. J. 4649, 51, 74 Brooks, R. 83 Brown, P. J. 119 B uhler, W. 28
Cantor, R. 121 Caouette, J. B. 10, 18 CGFS 4 Chakravarty, S. 48, 56 Chan, K. 79 Cheng, W.-Y. 83 Chordia, T. 47 Collin-Dufresne, P. 48, 87 Conroy, P. 112 Cornell, B. 48 Cossin, D. 76, 90, 92, 94 Cox, J. C. 11, 79 Crabbe, L. E. 48, 55, 56 Credit Suisse Financial Products 10 Crosbie, P. 11 Crouhy, M. 10 Culp, C. L. 78 Cumby, R. E. 79 Das, S. R. 6, 12, 76, 83, 113, 133 Daves, P. R. 47 De Ceuster, M. J. K. 46, 48 Deacon, M. 18 Derman, E. 17 Derry, A. 18 D az, A. 46, 48, 56 Dimson, E. 46, 48, 51 Driessen, J. 79 Duee, G. R. 18, 79, 86 Due, D. 10, 12, 13, 17, 79, 81, 83, 86, 87, 94 Dulake, S. 106 D ulmann, K. 79 Edwards, A. K. 48, 5659 Ehrhardt, M. C. 47
148
Author Index
Elton, E. J. 47, 48, 56, 58 Engineer, M. 106 Ericsson, J. 48, 56, 58, 60 Evans, M. D. 79 Fama, E. F. 4550, 64, 74 Ferri, M. G. 48, 5659 Finger, C. C. 10, 13 Fisher, L. 1, 55 Fleming, M. J. 47, 56 Fong, H. G. 18, 20 Francis, J. 6 French, K. R. 4550, 64, 74 Fridson, M. S. 48 Frost, J. 6 Fr uhwirth, M. 79, 86 Fumagalli, R. 111, 112, 116, 119 Galai, D. 10 Garman, M. 55 Gebhardt, W. R. 45, 46, 50, 51, 65, 74 Gehr, A. K. 48, 56, 60 Geske, R. 11 Geyer, A. 79 Goldreich, D. 4548, 54, 74 Goldstein, R. S. 48 Golub, B. 50, 87 Green, T. C. 47, 56 Greene, W. H. 23, 42, 52, 54, 88 Gregory, J. 12, 113 Gruber, M. J. 48, 56, 58 Gupton, G. M. 10, 13 Hamilton, D. 121 Hand, J. R. 134 Hanke, B. 4548, 51, 54, 74 Harrison, M. 80 Haugen, R. A. 47 Heath, D. 12, 17, 113 Helwege, J. 19, 34, 36 Hjort, V. 106 Hoek, J. 17 Holthausen, R. W. 134 Hong, G. 48, 56, 59 Houweling, P. 17, 45, 75, 111
Howard, K. 6 Hricko, T. 76, 90, 92, 94 Huang, M. 87 Huang, Z. 76, 90, 92, 94 Hull, J. 13, 17, 18, 83 Hvidkjaer, S. 45, 46, 50, 51, 65, 74 Iben, T. 12, 17 Ingersoll, J. E. 79 ISDA 7 Israel, R. B. 115 Jankowitsch, R. 20, 48, 53, 56, 60 Janosi, T. 79 Jarrow, R. 79 Jarrow, R. A. 12, 13, 15, 17, 46, 80, 81, 83, 86, 111, 113, 114, 124, 125, 135 Jeanblanc, M. 10 Johnson, H. 11 J onsson, J. G. 48 Kamara, A. 47 Karolyi, G. A. 79 Kealhofer, S. 10 Kempf, A. 48, 53, 56 Keswani, A. 79 Kijima, M. 113, 115 Kleibergen, F. R. 17 Koci c, A. 50, 87 Komoribayashi, K. 113, 115 Kossmeier, S. 79 Krishnamurthy, A. 47, 56 Kyle, A. 48, 51 Lando, D. 10, 12, 15, 17, 82, 111, 113, 124, 125, 135 Langetieg, T. C. 24 Laurent, J.-P. 12, 113 Leftwich, R. W. 134 Lehman Brothers 63 Leland, H. E. 11 Li, D. X. 13 Litterman, R. 12, 17 Litzenberger, R. H. 24 Lo, A. W. 53
Author Index
149
Longsta, F. A. 11, 12, 34, 79, 87 Lys, T. 53 MacKinlay, A. C. 53 Madan, D. 79, 81 Madan, D. B. 10, 79 Madsen, R. W. 133 Mann, C. 48, 56, 58 Marchakitus, S. 114, 119 Mark, R. 10 Martell, T. F. 48, 56, 60 Martin, J. S. 48 Martin, S. 111, 112, 114, 119, 126, 133 McAdie, R. 78, 94, 106, 111, 112, 114, 119, 126, 133 McCulloch, J. H. 18, 20, 23 McGinty, L. 46, 48, 5557 McLeish, N. A. 106 Mella-Barral, P. 11 Mendelson, H. 47, 49, 5557, 69 Mentink, A. 45, 111 Merton, R. C. 1, 11, 13, 34 Meyers, S. C. 21 Modigliani, F. 57 Moeschenberger, M. L. 133 Mortensen, A. 113 Morton, A. 12, 17, 113 M osenbacher, H. 48, 53, 56, 60 Mullineaux, D. J. 48, 56 Murphy, G. 18 Nagpal, K. M. 115 Nakagawa, H. 83 Nandi, S. 10 Narayanan, P. 10, 18 Nath, P. 4548, 54, 74 Navarro, E. 46, 48, 56 Nelson, C. R. 18 Neves, A. M. P. 78 Newey, W. 52 Nielsen, S. S. 79 OKane, D. 7, 10, 78, 94, 106, 111, 112, 114, 119, 126, 133 Ou, S. 121
Pan, G. 12 Patel, N. 7, 77 Pedersen, L. H. 79, 87 Perraudin, W. 11, 88 Pichler, S. 20, 48, 53, 56, 60, 79 Pliska, S. 80 Poirier, D. J. 23, 25 Powell, M. J. D. 20, 21, 25, 41 Price, K. 48, 56, 59 Quintos, C. 50, 87 Renault, O. 48, 56, 58, 60 Ritter, J. R. 134 Rolfo, R. 24 Roll, R. 47 Ronn, E. I. 79 Rosenthal, J. S. 115 Ross, S. A. 79 Roten, I. C. 48, 56 Rutkowski, M. 10, 12, 113 Sabino, J. S. 53 Sanders, A. B. 79 Sarig, O. 47, 5558 Sarkar, A. 48, 56 Saunders, A. 10, 18 Schaefer, S. M. 18 Schl ogl, L. 10 Scholes, M. 1, 11 Sch onbucher, P. J. 1, 1113, 81, 113, 114 Schubert, D. 13 Schultz, P. 5658, 60 Schwartz, E. S. 11, 12, 34 Scott, L. 83 Shea, G. S. 18, 20, 23 Shiller, R. J. 57 Shimko, D. C. 11 Shulman, J. 48, 56, 59 Siegel, A. F. 18 Singleton, K. J. 12, 13, 17, 79, 81, 86, 87 Sirinathsingh, M. 111, 112 Smidt, S. 55 Smithson, C. 6 Smoot, J. S. 24
150
Author Index
Soderberg, M. 114, 119 S ogner, L. 79, 86 Solnik, B. 87 Steeley, J. M. 18, 20, 21, 24, 41 Strebulaev, I. A. 47 Subrahmanyam, A. 4649, 51, 74 Sundaram, R. 12, 113, 133 Sundaram, R. K. 12, 76, 83 Svensson, L. E. O. 20 Swaminathan, B. 45, 46, 50, 51, 65, 74 Taur en, M. 79, 111, 112, 116, 119 Tavakoli, J. 6, 7 Taylor, A. 88 Tejima, N. 11 Tilman, L. 50, 87 Toft, K. B. 11 Tolk, J. S. 7 Toy, W. 17 Tufano, P. 12, 113 Turc, J. 112 Turnbull, S. M. 12, 13, 15, 17, 46, 80, 81, 83, 86, 111, 114, 124, 125, 135 Turner, C. M. 19, 34, 36, 48, 55, 56 Tychon, P. 60 Uhrig-Homburg, M. 28, 48, 53, 56 Unal, H. 10, 79 van Deventer, D. R. 11 Vannetelbosch, V. 60 Vasicek, O. A. 11, 17, 18, 20, 79 Vayanos, D. 47 Vorst, T. 45, 75, 111 Walter, U. 28 Warga, A. 48, 56, 59 Warga, A. D. 47, 5558 Weber, T. 28 Wei, J. Z. 115 West, K. 52 White, A. 13, 17, 18, 83 Whitelaw, R. F. 47, 48 Whittaker, G. 6 Wilson, T. 10
Windfuhr, M. 79 Yan, D. Y. 83 Yared, F. 50, 87 Yildirim, Y. 79 Zhang, F. 79, 81 Zhou, C. 11
Bibliography
Acharya, V., Das, S. R. and Sundaram, R. (2002), A discrete-time approach to no-arbitrage pricing of credit derivatives with rating transitions, Financial Analysts Journal (MayJune), 28 44. Adams, K. J. and van Deventer, D. R. (1994), Fitting yield curves and forward rate curves with maximum smoothness, Journal of Fixed Income 4(1), 5262. Alexander, G. J., Edwards, A. K. and Ferri, M. G. (2000), The determinants of trading volume of high-yield corporate bonds, Journal of Financial Markets 3, 177204. Altman, E. I. (1968), Financial ratios, discriminant analysis and the prediction of corporate bankruptcy, Journal of Finance 23(4), 589609. Amihud, Y. and Mendelson, H. (1986), Asset pricing and the bid-ask spread, Journal of Financial Economics 17, 223249. Amihud, Y. and Mendelson, H. (1991), Liquidity, maturity, and the yields on U.S. Treasury securities, Journal of Finance 46(4), 14111425. Anderson, N., Breedon, F., Deacon, M., Derry, A. and Murphy, G. (1996), Estimating and interpreting the yield curve, John Wiley & Sons, Chicester. Annaert, J. and De Ceuster, M. J. K. (1999), Modelling European credit spreads, Working paper, University of Antwerp-UFSIA. url: http://fetew.rug.ac.be/fineco/jana.html Aonuma, K. and Nakagawa, H. (1998), Valuation of credit default swap and parameter estimation for Vasicek-type hazard rate model, Working paper, University of Tokyo. Arnold, S. F. (1990), Mathematical Statistics, Prentice Hall, New Jersey. Arvanitis, A., Gregory, J. and Laurent, J.-P. (1999), Building models for credit spreads, The Journal of Derivatives (Spring), 2743. Aunon-Nerin, D., Cossin, D., Hricko, T. and Huang, Z. (2002), Exploring for the determinants of credit risk in credit default swap transaction data: Is xed-income markets information sucient to evaluate credit risk?, Working paper, HEC-University of Lausanne and FAME. url: http://www.hec.unil.ch/dcossin/PAPERS.HTM Bakshi, G., Madan, D. and Zhang, F. (2001), Understanding the role of recovery in default risk models: Empirical comparisons and implied recovery rates, Working paper, University of
152
Bibliography
Maryland. url: http://papers.ssrn.com/sol3/papers.cfm?abstract id=285940 BBA (1998), Credit Derivatives Report 1997/1998, British Bankers Association. url: http://www.bba.org.uk/ BBA (2001), BBA Repo Rates, British Bankers Association. url: http://www.bba.org.uk/ BBA (2002), Credit Derivatives Report 2001/2002, British Bankers Association. url: http://www.bba.org.uk/ BCBS (1988), International Convergence of Capital Measurement and Capital Standards, Basel Committee on Banking Supervision, Bank of International Settlements. url: http://www.bis.org/publ/bcbs04a.htm BCBS (1996), Amendment to the Capital Accord to Incorporate Market Risks, Basel Committee on Banking Supervision, Bank of International Settlements. url: http://www.bis.org/publ/bcbs24a.htm BCBS (1999), Consultative Paper on a New Capital Adequacy Framework, Basel Committee on Banking Supervision, Bank of International Settlements. url: http://www.bis.org/publ/bcbs50.htm BCBS (2001), The New Basel Capital Accord, Basel Committee on Banking Supervision, Bank of International Settlements. url: http://www.bis.org/publ/bcbsca.htm Beim, D. O. (1992), Term structure and the non-cash value in bonds, Working Paper 9240, Graduate School of Business, Columbia University. Bielecki, T. R. and Rutkowski, M. (2000), Multiple ratings model of defaultable term structure, Mathematical Finance 10(2), 125139. Bielecki, T. R. and Rutkowski, M. (2001), Credit risk modelling: Intensity based approach, in E. Jouini, J. Cvitanic and M. Musiela, eds, Option Pricing, Interest Rates and Risk Management, Cambridge University Press, Cambridge, pp. 399457. BIS (2002), Regular OTC Derivatives Market Statistics, Bank for International Settlements. url: http://www.bis.org/statistics/derstats.htm Black, F. and Cox, J. C. (1976), Valuing corporate securities: Some eects of bond indenture provisions, Journal of Finance 31(2), 351367. Black, F., Derman, E. and Toy, W. (1990), A one-factor model of interest rates and its application to Treasury bond options, Financial Analysts Journal 46(1), 3339. Black, F. and Scholes, M. (1973), The pricing of options and corporate liabilities, Journal of Political Economy 81, 637654. Bliss, R. (1997), Testing term structure estimation methods, Advances in Futures and Options Research 9(1), 197231.
Bibliography
153
Boudoukh, J. and Whitelaw, R. F. (1991), The benchmark eect in the Japanese government bond market, Journal of Fixed Income 1(2), 5259. Boudoukh, J. and Whitelaw, R. F. (1993), Liquidity as a choice variable: A lesson from the Japanese government bond market, Review of Financial Studies 6(2), 265292. Brealey, R. A. and Meyers, S. C. (1991), Principles of Corporate Finance, 4th edn, McgrawHill, New York. Brennan, M. J., Chordia, T. and Subrahmanyam, A. (1998), Alternative factor specications, security characteristics, and the cross-section of expected stock returns, Journal of Financial Economics 49, 345373. Brennan, M. J. and Subrahmanyam, A. (1996), Market microstructure and asset pricing: On the compensation for illiquidity in stock returns, Journal of Financial Economics 41(3), 441464. Brooks, R. and Yan, D. Y. (1998), Pricing credit default swaps and the implied default probability, Derivatives Quarterly 5(2), 3441. Brown, P. J. (1994), Constructing & calculating bond indices, a guide to the EFFAS standardized rules, Cambridge University Press, Cambridge. B uhler, W., Uhrig-Homburg, M., Walter, U. and Weber, T. (1999), An empirical comparison of forward and spotrate models for valuing interest-rate options, Journal of Finance 54(1), 269305. Cantor, R., Hamilton, D. and Ou, S. (2002), Default & recovery rates of corporate bond issuers, a statistical review of Moodys ratings performance 19702001, Global Credit Research, Moodys Investors Service. Caouette, J. B., Altman, E. I. and Narayanan, P. (1998), Managing credit risk: The next great nancial challenge, John Wiley & Sons, New York. CGFS (1999), A Review of Financial Market Events in Autumn 1998, Committee on the Global Financial System, Bank of International Settlements. url: http://www.bis.org/publ/cgfs12.htm CGFS (2001), Structural Aspects of Market Liquidity from a Financial Stability Perspective, Committee on the Global Financial System, Bank of International Settlements. url: http://www.bis.org/publ/cgfs note01.htm Chakravarty, S. and Sarkar, A. (1999), Liquidity in U.S. xed income markets: A comparison of the bid-ask spread in corporate, government and municipal bond markets, Working paper, Purdue University and Federal Reserve Bank of New York. url: http://papers.ssrn.com/sol3/papers.cfm?abstract id=163139 Chan, K., Karolyi, G. A., Longsta, F. A. and Sanders, A. B. (1992), An empirical comparison of alternative models of the short-term interest rate, Journal of Finance 47(3), 12091227. Cheng, W.-Y. (2001), Recent advances in default swap valuation, The Journal of Derivatives 9(1), 1827.
154
Bibliography
Chordia, T., Roll, R. and Subrahmanyam, A. (2001), Market liquidity and trading activity, Journal of Finance 56(2), 501530. Chordia, T., Subrahmanyam, A. and Anshuman, V. R. (2001), Trading activity and expected stock returns, Journal of Financial Economics 59, 332. Collin-Dufresne, P., Goldstein, R. S. and Martin, J. S. (2001), The determinants of credit spread changes, Journal of Finance 56(6), 21772207. Collin-Dufresne, P. and Solnik, B. (2001), On the term structure of default premia in the swap and LIBOR markets, Journal of Finance 56(3), 10951115. Conroy, P. (2000), Valuing ratings-sensitive coupon language, Credit (July), 6264. Cornell, B. (1992), Liquidity and the pricing of low-grade bonds, Financial Analysts Journal 48(1), 6368. Cox, J. C., Ingersoll, J. E. and Ross, S. A. (1985), A theory of the term structure of interest rates, Econometrica 53(2), 385407. Crabbe, L. E. and Turner, C. M. (1995), Does the liquidity of a debt issue increase with its size? Evidence from the corporate bond and medium-term note markets, Journal of Finance 50(5), 17191734. Credit Suisse Financial Products (1997), CreditRisk+: A credit risk management framework, Technical documentation. Crosbie, P. (2002), Modeling default risk, Working paper, Moodys KMV. url: http://www.moodyskmv.com/research/defaultrisk.html Crouhy, M., Galai, D. and Mark, R. (2000), A comparative analysis of current credit risk models, Journal of Banking & Finance 24(12), 59117. Culp, C. L. and Neves, A. M. P. (1998), Credit and interest rate risk in the business of banking, Derivatives Quarterly 4(4), 1935. Cumby, R. E. and Evans, M. D. (1997), The term structure of credit risk: Estimates and specication tests, Working paper, Georgetown University and National Bureau of Economic Research. url: http://www.georgetown.edu/faculty/evansm1/wpapers.htm Das, S. R. (1996), Pricing credit derivatives total return swaps & credit spread products, Derivatives Laboratory 53(October), 1619. Das, S. R. (1998), Credit Derivatives: Trading & Management of Credit & Default Risk, John Wiley & Sons, New York. Das, S. R. and Sundaram, R. K. (2000), A discrete-time approach to arbitragefree pricing of credit derivatives, Management Science 46(1), 4662. Das, S. R. and Tufano, P. (1996), Pricing credit-sensitive debt when interest rates, credit ratings and credit spreads are stochastic, Journal of Financial Engineering 5, 161198. Daves, P. R. and Ehrhardt, M. C. (1993), Liquidity, reconstitution, and the value of U.S. Treasury strips, Journal of Finance 48(1), 315329.
Bibliography
155
D az, A. and Navarro, E. (2002), Yield spread and term to maturity: Default vs. liquidity, European Financial Management 8(4), 449477. Dimson, E. and Hanke, B. (2001), The expected illiquidity premium: Evidence from equity index-linked bonds, Working paper, London Business School. url: http://phd.london.edu/bhanke/ Driessen, J. (2001), The cross-rm behaviour of credit spreads, Working paper, Tilburg University. url: http://www1.fee.uva.nl/fm/PEOPLE/Driessencv.htm Duee, G. R. (1996), On measuring credit risks of derivative instruments, Journal of Banking & Finance 20, 805833. Duee, G. R. (1998), The relation between Treasury yields and corporate bond yield spreads, Journal of Finance 53(6), 22252242. Due, D. (1996), Special repo rates, Journal of Finance 51(2), 493526. Due, D. (1998), First to default valuation, Working paper, Graduate School of Business, Stanford University. url: http://www.stanford.edu/duffie/working.htm Due, D. (1999), Credit swap valuation, Financial Analysts Journal (JanuaryFebruary), 73 85. Due, D. and Huang, M. (1996), Swap rates and credit quality, Journal of Finance 51(3), 921 949. Due, D. and Lando, D. (2001), Term structures of credit spreads with incomplete accounting information, Econometrica 69(3), 633664. Due, D., Pedersen, L. H. and Singleton, K. J. (2003), Modeling sovereign yield spreads: A case study of Russian debt, Journal of Finance 58(1), 119160. Due, D. and Singleton, K. J. (1997), An econometric model of the term structure of interestrate swap yields, Journal of Finance 52(4), 12871321. Due, D. and Singleton, K. J. (1998), Simulating correlated defaults, Working paper, Graduate School of Business, Stanford University. url: http://www.stanford.edu/duffie/working.htm Due, D. and Singleton, K. J. (1999), Modeling term structures of defaultable bonds, Review of Financial Studies 12(4), 687720. D ulmann, K. and Windfuhr, M. (2000), Credit spreads between German and Italian sovereign bonds: Do one-factor ane models work?, Canadian Journal of Administrative Sciences 17(2). Elton, E. J. and Green, T. C. (1998), Tax and liquidity eects in pricing government bonds, Journal of Finance 53(5), 15331562. Elton, E. J., Gruber, M. J., Agrawal, D. and Mann, C. (2002), Factors aecting the valuation of corporate bonds, Working paper, Stern School of Business, New York University.
156
Bibliography
url: http://pages.stern.nyu.edu/ eelton/Research.htm Ericsson, J. and Renault, O. (2001), Liquidity and credit risk, Working paper, McGill University and Universit e Catholique de Louvain. url: http://papers.ssrn.com/sol3/papers.cfm?abstract id=248368 Fama, E. F. and French, K. R. (1993), Common risk factors in the returns on stocks and bonds, Journal of Financial Economics 33(1), 356. Fisher, L. (1959), Determinants of the risk premiums on corporate bonds, Journal of Political Economy 67, 217237. Fleming, M. J. (2001), Measuring Treasury market liquidity, Working paper, Federal Reserve Bank of New York. url: http://www.ny.frb.org/rmaghome/staff rp/2001/sr133.html Fleming, M. J. (2002), Are larger Treasury issues more liquid? Evidence from bill reopenings, Journal of Money, Credit, and Banking 3(2), 707735. Francis, J., Frost, J. and Whittaker, G. (1999), The Handbook of Credit Derivatives, Mcgraw Hill, New York. Fridson, M. S. and J onsson, J. G. (1995), Spread versus Treasuries and the riskiness of high-yield bonds, Journal of Fixed Income 5(3), 7988. Fr uhwirth, M. and S ogner, L. (2001), The Jarrow/Turnbull default risk model: Evidence from the German market, Working paper, Vienna University of Economics. url: http://www.wu-wien.ac.at/usr/abwldcf/mfrueh/publ.html Fr uhwirth, M. and S ogner, L. (2003), The implicit estimation of default intensities and recovery rates, in Proceedings of the 64th Scientic Conference of the Association of University Professors of Management 2002, Munich. Fumagalli, R. and Taur en, M. (2001), Valuing bonds with step-up coupons, European Credit Research, Credit Strategy, Schroder Salomon Smith Barney. Garman, M. (1976), Market microstructure, Journal of Financial Economics 3, 257275. Gebhardt, W. R., Hvidkjaer, S. and Swaminathan, B. (2001), The cross-section of expected corporate bond returns: Betas or characteristics?, Working paper, Axia Energy Europe, University of Maryland and Cornell University. url: http://papers.ssrn.com/sol3/papers.cfm?abstract id=281209 Gehr, A. K. and Martell, T. F. (1992), Pricing eciency in the secondary market for investmentgrade corporate bonds, Journal of Fixed Income 2(3), 2438. Geske, R. (1977), The valuation of corporate liabilities as compound options, Journal of Financial and Quantitative Analysis 12(4), 541552. Geske, R. and Johnson, H. (1984), The valuation of corporate liabilities as compound options: A correction, Journal of Financial and Quantitative Analysis 19(2), 231232. Geyer, A., Kossmeier, S. and Pichler, S. (2001), Empirical analysis of European government yield spreads, Working paper, Vienna University of Economics.
Bibliography
157
url: http://www.wu-wien.ac.at/usr/or/geyer/publ.htm Goldreich, D., Hanke, B. and Nath, P. (2002), The price of future liquidity: Time-varying liquidity in the U.S. Treasury market, Working paper, London Business School, Institute of Finance and Accounting. url: http://phd.london.edu/bhanke/ Golub, B. and Tilman, L. (2000), No room for nostalgia in xed income, Risk (July), 4448. Greene, W. H. (2000), Econometric Analysis, 4th edn, Prentice Hall, New Jersey. Gupton, G. M., Finger, C. C. and Bhatia, M. (1997), Creditmetricstm , Technical document, J.P. Morgan. url: http://www.riskmetrics.com/creditdocs.html Hand, J. R., Holthausen, R. W. and Leftwich, R. W. (1992), The eect of bond rating agency announcements on bond and stock prices, Journal of Finance 47(2), 733752. Harrison, M. and Pliska, S. (1981), Martingales and stochastic integrals in the theory of continuous trading, Stochastic Processes and Their Applications 11, 215260. Haugen, R. A. and Baker, N. L. (1996), Commonality in the determinants of expected stock returns, Journal of Financial Economics 41, 401439. Heath, D., Jarrow, R. A. and Morton, A. (1992), Bond pricing and the term structure of interest rates: A new methodology, Econometrica 60(1), 77105. Helwege, J. and Turner, C. M. (1999), The slope of the credit yield curve for speculativegrade issuers, Journal of Finance 54(5), 18691884. Hjort, V., McLeish, N. A., Dulake, S. and Engineer, M. (2002), The high beta market - Exploring the default swap basis, Fixed Income Research, MorganStanley, London. Hong, G. and Warga, A. (2000), An empirical study of bond market transactions, Financial Analysts Journal 56(2), 3246. Houweling, P., Hoek, J. and Kleibergen, F. R. (2001), The joint estimation of term structures and credit spreads, Journal of Empirical Finance 8(3), 297323. Houweling, P., Mentink, A. and Vorst, T. (2003a), How to measure corporate bond liquidity?, Working paper, Erasmus University Rotterdam. url: http://www.few.eur.nl/few/people/houweling/ Houweling, P., Mentink, A. and Vorst, T. (2003b), Valuing euro rating-triggered step-up telecom bonds, Working paper, Erasmus University Rotterdam. url: http://www.few.eur.nl/few/people/houweling/ Houweling, P. and Vorst, T. (2002), An empirical comparison of default swap pricing models, Working paper, Erasmus University Rotterdam. url: http://www.few.eur.nl/few/people/houweling/ Howard, K. (1995), An introduction to credit derivatives, Derivatives Quarterly 2(2), 2837. Hull, J. and White, A. (1990), Pricing interest-rate-derivative securities, Review of Financial Studies 3(4), 573592.
158
Bibliography
Hull, J. and White, A. (1995), The impact of default risk on the prices of options and other derivative securities, Journal of Banking & Finance 19, 299322. Hull, J. and White, A. (2000), Valuing credit default swaps I: No counterparty default risk, The Journal of Derivatives 8(1), 2940. Hull, J. and White, A. (2001), Valuing credit default swaps II: Modeling default correlations, The Journal of Derivatives 8(3), 1222. ISDA (1999), Credit Derivatives Denitions, International Swaps and Derivatives Association. url: http://www.isda.org/ ISDA (2001), Restructuring Supplement to the 1999 ISDA Credit Derivatives Denitions, International Swaps and Derivatives Association. url: http://www.isda.org/ Israel, R. B., Rosenthal, J. S. and Wei, J. Z. (2001), Finding generators for Markov chains via empirical transition matrices, with applications to credit ratings, Mathematical Finance 11(2), 245265. Jankowitsch, R., M osenbacher, H. and Pichler, S. (2002), Measuring the liquidity impact on EMU government bond prices, Working paper, Vienna University, CCEFM and Bank Austria. url: http://papers.ssrn.com/sol3/papers.cfm?abstract id=302795 Jankowitsch, R. and Pichler, S. (2002), Parsimonious estimation of credit spreads, Working paper, Vienna University of Technology. url: http://papers.ssrn.com/sol3/papers.cfm?abstract id=306779 Jankowitsch, R. and Pichler, S. (2003), Currency dependence of corporate credit spreads, Working paper, Vienna University of Technology. url: http://papers.ssrn.com/sol3/papers.cfm?abstract id=394521 Janosi, T., Jarrow, R. and Yildirim, Y. (2002), Estimating expected losses and liquidity discounts implicit in debt prices, The Journal of Risk 5(1). Jarrow, R. A. (2001), Default parameter estimation using market prices, Financial Analysts Journal (SeptemberOctober), 7592. Jarrow, R. A., Lando, D. and Turnbull, S. M. (1997), A Markov model for the term structure of credit spreads, Review of Financial Studies 10(2), 481523. Jarrow, R. A. and Turnbull, S. M. (1995), Pricing derivatives with credit risk, Journal of Finance 50(1), 5385. Jarrow, R. A. and Turnbull, S. M. (1998), Credit risk, in C. Alexander, ed., Handbook of Risk Management and Analysis, John Wiley & Sons, pp. 237254. Jarrow, R. A. and Turnbull, S. M. (2000), The intersection of market and credit risk, Journal of Banking & Finance 24, 271299. Jeanblanc, M. and Rutkowski, M. (1999), Models for default risk: An overview, in J. Yong and R. Cont, eds, Mathematical Finance: Theory and Practice, Higher Education Press, Beijing, pp. 171269.
Bibliography
159
Kamara, A. (1994), Liquidity, taxes, and short-term Treasury yields, Journal of Financial and Quantitative Analysis 29(3), 403417. Kealhofer, S. (2001), Portfolio management of default risk, Working paper, Moodys KMV. url: http://www.moodyskmv.com/research/portfoliotheory.html Kempf, A. and Uhrig-Homburg, M. (2000), Liquidity and its impact on bond prices, Schmalenbach Business Review 52, 2644. Keswani, A. (2000), Estimating a risky term structure of Brady bonds, Working paper, Lancaster University. url: http://www.lums2.lancs.ac.uk/ACFIN/Papers/2000.HTM Kijima, M. and Komoribayashi, K. (1998), A Markov chain model for valuing credit risk derivatives, The Journal of Derivatives (Fall), 97108. Koci c, A., Quintos, C. and Yared, F. (2000), Identifying the benchmark security in a multifactor spread environment, Fixed Income Derivatives Research, Lehman Brothers, New York. Krishnamurthy, A. (2002), The bond/old-bond spread, Journal of Financial Economics 66(2 3), 463506. Kyle, A. (1985), Continuous auctions and insider trading, Econometrica 53, 13151335. Lando, D. (1994), Three Essays on Contingent Claims Pricing, PhD thesis, Graduate School of Management, Cornell University. Lando, D. (1998), On Cox processes and credit risky securities, Review of Derivatives Research 2(2/3), 99120. Lando, D. (2000), On correlated defaults in a rating-based model common state variables versus simultaneuous defaults, Working paper, University of Copenhagen. Lando, D. and Mortensen, A. (2003), Mispricing of step-up bonds in the European telecom sector, Working paper, University of Copenhagen and Copenhagen Business School. Langetieg, T. C. and Smoot, J. S. (1989), Estimation of the term structure of interest rates, Research in Financial Services 1, 181222. Lehman Brothers (1999), Introduction to the Pan-European Aggregate Index, Research document, Fixed Income Research. Leland, H. E. (1994), Corporate debt value, bond covenants, and optimal capital structure, Journal of Finance 49(4), 12131252. Leland, H. E. and Toft, K. B. (1996), Optimal capital structure, endogenous bankruptcy, and the term structure of credit spreads, Journal of Finance 51(3), 9871019. Li, D. X. (1999), The valuation of basket credit derivatives, CreditMetrics (April), 3450. Monitor
Li, D. X. (2000), On default correlation: A copula approach, Journal of Fixed Income 9(4), 43 54.
160
Bibliography
Litterman, R. and Iben, T. (1991), Corporate bond valuation and the term structure of credit spreads, Journal of Portfolio Management (Spring), 5264. Litzenberger, R. H. and Rolfo, R. (1984), An international study of tax eects on government bonds, Journal of Finance 39(1), 122. Lo, A. W. and MacKinlay, A. C. (1990), Data-snooping biases in tests of nancial asset pricing models, Review of Financial Studies 3(3), 413467. Longsta, F. A. (2000), The term structure of very short-term rates: New evidence for the expectations hypothesis, Journal of Financial Economics 58(3), 397415. Longsta, F. A. and Schwartz, E. S. (1995a), A simple approach to valuing risky xed and oating rate debt, Journal of Finance 50(3), 789819. Longsta, F. A. and Schwartz, E. S. (1995b), Valuing credit derivatives, Journal of Fixed Income (June), 612. Lys, T. and Sabino, J. S. (1992), Research design issues in grouping-based tests, Journal of Financial Economics 32(3), 355387. Madan, D. B. and Unal, H. (1998), Pricing the risks of default, Review of Derivatives Research 2, 121160. Madan, D. B. and Unal, H. (2000), A two-factor hazard rate model for pricing risky debt and the term structure of credit spreads, Journal of Financial and Quantitative Analysis 35(1), 4365. Madsen, R. W. and Moeschenberger, M. L. (1986), Statistical Concepts with Applications to Business and Economics, 2nd edn, Prentice Hall, New Jersey. Marchakitus, S., Soderberg, M. and Bramley, A. (2001), European telecoms update, watch your step..., Credit Research Sector Report, J.P. Morgan Securities Ltd. McAdie, R., Martin, S. and OKane, D. (2000), Credit-sensitive telecom bonds: A framework for valuation, European Credit Strategy & Research Market Analysis, Lehman Brothers. McCulloch, J. H. (1971), Measuring the term structure of interest rates, Journal of Business 44, 1931. McCulloch, J. H. (1975), The taxadjust yield curve, Journal of Finance 30(3), 811830. McGinty, L. (2001), Issue size versus liquidity in credit, J.P. Morgan Fixed Income Research . Mella-Barral, P. and Perraudin, W. (1997), Strategic debt service, Journal of Finance 52(2), 556556. Merton, R. C. (1974), On the pricing of corporate debt: The risk structure of interest rates, Journal of Finance 29(2), 449469. Mullineaux, D. J. and Roten, I. C. (2002), Liquidity, labels, and medium-term notes, Financial Markets, Institutions and Instruments 11, 401423. Nagpal, K. M. and Bahar, R. (2000), Modelling the dynamics of rating transition, Credit (March).
Bibliography
161
Nakagawa, H. (1999), Valuation of default swap with ane-type hazard rate, Working paper, University of Tokyo. Nandi, S. (1998), Valuation models for default-risky securities: An overview, Federal Reserve Bank Atlanta Ecnomic Review (4th quarter). Nelson, C. R. and Siegel, A. F. (1987), Parsimonious modeling of yield curves, Journal of Business 60, 473489. Newey, W. and West, K. (1987), A simple positive semi-denite, heteroscedasticity and autocorrelation consistent covariance matrix, Econometrica 55, 703708. Nielsen, S. S. and Ronn, E. I. (1998), The valuation of default risk in corporate bonds and interest rate swaps, Working paper, University of Copenhagen and University of Texas at Austin. url: http://www.math.ku.dk/nielsen/ OKane, D. (2001), Credit derivatives explained, Fixed Income Research, Lehman Brothers, London. OKane, D. and McAdie, R. (2001), Explaining the basis: Cash versus default swaps, Fixed Income Research, Lehman Brothers, London. OKane, D. and Schl ogl, L. (2001), Modelling credit: Theory and practice, Analytical research series, Lehman Brothers, Fixed Income Research, London. Pan, G. (2001), Equity to credit pricing, Risk (November), 99102. Patel, N. (2003), Flow business booms, Risk (February), 2023. Perraudin, W. and Taylor, A. (1999), On the consistency of ratings and bond market yields, Working paper, Bank of England and Birkbeck College. url: http://wp.econ.bbk.ac.uk/respap/respap frames.html Poirier, D. J. (1976), The Econometrics of Structural Change, North Holland Publishing Company. Powell, M. J. D. (1981), Approximation Theory and Methods, Cambridge University Press, Cambridge. Ritter, J. R. (1991), The long-run performance of initial public oerings, Journal of Finance 46(1), 327. Sarig, O. and Warga, A. D. (1989), Bond price data and bond market liquidity, Journal of Financial and Quantitative Analysis 24(3), 367378. Saunders, A. (1999), Credit risk management: new approaches to Value at Risk and other paradigms, John Wiley & Sons, New York. Schaefer, S. M. (1981), Measuring a taxspecic term structure of interest rates in the market for British government securities, Economic Journal 91, 415438. Sch onbucher, P. J. (1996), Valuation of securities subject to credit risk, Working paper, University of Bonn. url: http://www.schonbucher.de/
162
Bibliography
Sch onbucher, P. J. (1999), Credit Risk Modelling and Credit Derivatives, PhD thesis, University of Bonn. url: http://www.schonbucher.de/ Sch onbucher, P. J. (2000), A LIBOR market model with default risk, Working paper, University of Bonn. url: http://www.schonbucher.de/ Sch onbucher, P. J. and Schubert, D. (2001), Copula-dependent default risk in intensity models, Working paper, University of Bonn. url: http://www.schonbucher.de/ Schultz, P. (2001), Corporate bond trading costs and practices: A peek behind the curtain, Journal of Finance 56(2), 677698. Scott, L. (1998), A note on the pricing of default swaps, Working paper, Morgan Stanley Dean Witter. Shea, G. S. (1984), Pitfalls in smoothing interest rate term structure data: equilibrium models and spline approximations, Journal of Financial and Quantitative Analysis 19(3), 253269. Shea, G. S. (1985), Interest rate term structure estimation with exponential splines: A note, Journal of Finance 40(1), 319325. Shiller, R. J. and Modigliani, F. (1979), Coupon and tax eects on new and seasoned bond yields and the measurement of the cost of debt capital, Journal of Financial Economics 7(3), 297318. Shimko, D. C., Tejima, N. and van Deventer, D. R. (1993), The pricing of risky debt when interest rates are stochastic, Journal of Fixed Income 3(2), 5865. Shulman, J., Bayless, M. and Price, K. (1993), Marketability and default inuences on the yield premia of speculative-grade debt, Financial Management 22(3), 132141. Sirinathsingh, M. (2001), Valuing rating-triggered step-up bonds, Fixed Income Strategy, J.P. Morgan Securities Inc. Smidt, S. (1971), Which road to an ecient stock market: Free competition or regulated monopoly, Financial Analysts Journal 27(5), 1820, 6469. Smithson, C. (1995), Credit derivatives, Risk (December), 3839. Steeley, J. M. (1991), Estimating the giltedged term structure: basis splines and condence intervals, Journal of Business Finance & Accounting 18(4), 513529. Strebulaev, I. A. (2001), Many faces of liquidity and asset pricing: Evidence from the U.S. Treasury securities market, Working paper, London Business School. url: http://facultyresearch.london.edu/detwp.asp?id=4952 Svensson, L. E. O. (1994), Estimating and interpreting forward interest rates: Sweden 1992-94, Working Paper 4871, National Bureau of Economic Research. url: http://papers.nber.org/papers/W4871
Bibliography
163
Taur en, M. (1999), A comparison of bond pricing models in the pricing of credit risk, Working paper, Indiana University, Bloomington. url: http://papers.ssrn.com/sol3/papers.cfm?abstract id=155688 Tavakoli, J. (1998), Credit Derivatives: A Guide to Instruments and Applications, John Wiley & Sons, New York. Tolk, J. S. (2001), Understanding the risks in credit default swaps, Working paper, Moodys Investors Service. Turc, J. (2001), A rule-of-thumb for valuing telecom coupon step-ups, Quantitative strategy, Euro Credit Monthly, Soci et e General. Tychon, P. and Vannetelbosch, V. (2002), Debt valuation and marketability risk, Working paper, Universit e catholique de Louvain. url: http://www2.econ.ucl.ac.be/Users/v.vannetelbosch/ Vasicek, O. A. (1977), An equilibrium characterization of the term structure, Journal of Financial Economics 5(2), 177188. Vasicek, O. A. and Fong, H. G. (1982), Term structure modeling using exponential splines, Journal of Finance 37(2), 339348. Vayanos, D. (1998), Transaction costs and asset prices: A dynamic equilibrium model, Review of Financial Studies 11(1), 158. Warga, A. D. (1992), Bond returns, liquidity and missing data, Journal of Financial and Quantitative Analysis 27(4), 605617. Wilson, T. (1997a), Portfolio credit risk - part I, Risk (September), 111117. Wilson, T. (1997b), Portfolio credit risk - part II, Risk (October), 5661. Zhou, C. (1997), A jump-diusion approach to modeling credit risk and valuing defaultable securities, Working paper, Federal Reserve Board. url: http://www.federalreserve.gov/pubs/feds/1997/199715/199715abs.html
The Tinbergen Institute is the Institute for Economic Research, which was founded in 1987 by the Faculties of Economics and Econometrics of the Erasmus Universiteit Rotterdam, Universiteit van Amsterdam and Vrije Universiteit Amsterdam. The Institute is named after the late Professor Jan Tinbergen, Dutch Nobel Prize laureate in economics in 1969. The Tinbergen Institute is located in Amsterdam and Rotterdam. The following books recently appeared in the Tinbergen Institute Research Series: 270. N.K. BOOTS, Rare event simulation in models with heavytailed random variables. 271. P.J.M. MEERSMANS, Optimization of container handling systems. 272. J.G. VAN ROOIJEN, Flexibility in nancial accounting; income strategies and earnings management in the Netherlands. 273. D. ARNOLDUS, Family, family rm, and strategy. Six Dutch family rms in the food industry 1880-1970. 274. J.-P.P.E.F. BOSELIE, Human resource management, work systems and performance: A theoretical-empirical approach. 275. V.A. KARAMYCHEV, Essays on adverse selection: A dynamic perspective. 276. A.J. MENKVELD, Fragmented markets: Trading and price discovery. 277. D. ZEROM GODEFAY, Nonparametric prediction: Some selected topics. 278. T. DE GRAAFF, Migration, ethnic minorities and network externalities. 279. A. ZORLU, Absorption of immigrants in European labour markets. The Netherlands, United Kingdom and Norway. 280. B. JACOBS, Public nance and human capital. 281. PH. CUMPERAYOT, International nancial markets: Risk and extremes. 282. E.M. BAZSA-OLDENKAMP, Decision support for inventory systems with complete backlogging. 283. M.A.J. THEEBE, Housing market risks. 284. V. SADIRAJ, Essays on political and experimental economics. 285. J. LOEF, Incongruity between ads and consumer expectations of advertising. 286. J.J.J. JONKER, Target selection and optimal mail strategy in direct marketing. 287. S. CASERTA, Extreme values in auctions and risk analysis. 288. W.H. DAAL, A term structure model of interest rates and forward premia: An alternative approach. 289. H.K. CHAO, Representation and structure. The methodology of econometric models of consumption. 290. J. DALHUISEN, The economics of sustainable water use. Comparisons and lessons from urban areas. 291. P. DE BRUIN, Essays on modeling nonlinear time series. 292. J. ARDTS, All is well that begins well: A longitudinal study of organisational socialisation. 293. J.E.M. VAN NIEROP, Advanced choice models.
294. D.J. VAN VUUREN, The market for passenger transport by train. An empirical analysis. 295. A. FERRER CARBONELL, Quantitative analysis of well-being with economic applications. 296. L.M. VINHAS DE SOUZA, Beyond transition: Essays on the monetary integration of the accession countries in Eastern Europe. 297. J. LEVIN, Essays in the economics of education. 298. E. WIERSMA, Non-nancial performance measures: An empirical analysis of a change in a rms performance measurement system. 299. M. MEKONNEN AKALU, Projects for shareholder value: A capital budgeting perspective. 300. S. ROSSETTO, Optimal timing of strategic nancial decisions. 301. P.W. VAN FOREEST, Essays in nancial economics. 302. A. SIEGMANN, Optimal nancial decision making under loss averse preferences. 303. A. VAN DER HORST, Government interference in a dynamic economy. 304. A.P. RUSSO, The sustainable development of heritage cities and their regions: Analysis, policy, governance. 305. G.A.W. GRIFFIOEN, Technical analysis in nancial markets. 306. I.S. LAMMERS, In conict, een geschiedenis van kennismanagement. 307. O.L. LISTES, Stochastic programming approaches for strategic logistics problems. 308. A.T. DE BLAEIJ, The value of a statistical life in road safety. 309. S.H.K. WUYTS, Partner selection in business markets. A structural embeddedness perspective. 310. H.C. DEKKER, Control of inter-organizational relationships: The eects of appropriation concerns, coordination requirements and social embeddedness. 311. I.V. OSSOKINA, Environmental policy and environment-saving technologies. Economic aspects of policy making under uncertainty. 312. D. BROUNEN, Real estate securitization and corporate strategy: From bricks to bucks. 313. J.M.P. DE KOK, Human resource management within small and medium-sized enterprises. Facts and explanations. 314. T. VERHAGEN, Towards understanding online purchase behavior. 315. R. HOEKSTRA, Structural change of the physical economy. Decomposition analysis of physical and hybrid-units input-output tables. 316. R.K. AIDIS, By law and by custom: Factors aecting small and medium sized enterprises during the transition in Lithuania. 317. S. MANZAN, Essays in nonlinear economic dynamics. 318. K. OLTMER, Agricultural policy, land use and environmental eects: Studies in quantitative research synthesis. 319. H.J. HORBEEK, The elastic workoor. About the implementation of internal exibility arrangements.

Thesis SCM Nivi

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thesis SCM Nivi

Uploaded by

Copyright:

Available Formats

EMPIRICAL STUDIES ON CREDIT MARKETS

Empirical Studies on Credit Markets

Credit Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 17

3.3 3.4 3.5 3.6

5 Pricing Step-Up Bonds 5.1 5.2 5.3 5.4 5.5 5.6

6 Summary Nederlandse samenvatting (Summary in Dutch) Author Index Bibliography

Chapter 3 3.1 Liquidity premiums for dierent age thresholds. . . . . . . . . . . . . . . . 59

This list draws on Sch onbucher (1999).

Section 1.2 Sources of Risk

E10 million in one out of 20 days. Under the internal models

E10 million at a 95% condence level is expected to

Section 1.3 Credit Instruments

Section 1.3 Credit Instruments

Section 1.3 Credit Instruments

Section 1.4 Credit Models

Reduced Form Models

Section 1.5 Overview

Section 1.5 Overview

Estimating Spread Curves1

Estimating Spread Curves Chapter 2

Section 2.2 Multi-Curve Model

Estimating Spread Curves Chapter 2

Section 2.2 Multi-Curve Model

Estimating Spread Curves Chapter 2

CF1bi g1s (t1bi ) (2.4a)

CFcbi g1s (tcbi ). For categories 2, 3, . . . , C substitution of Equation (2.2b)

into Equation (2.3) results in

CFcbi g1s (tcbi )

CFcbi gcs (tcbi ) (2.4b)

ycbs cs = xcb 1 + ycb c ,

0 ... 0 ... . . . . . . YC ...

Section 2.3 Model Settings

Estimating Spread Curves Chapter 2

Section 2.4 Model Comparison

Estimating Spread Curves Chapter 2

Curve Similarity Test

Section 2.5 Data

Estimating Spread Curves Chapter 2

Section 2.5 Data

Table 2.1: Distribution of bonds in the data set by rating.

Figure 2.1: Distribution of the bonds maturity dates by rating category.

Estimating Spread Curves Chapter 2

Section 2.6 Results

Table 2.3: Model specications for single-curve and multi-curve models.

GOVT AAA AA A BBB BB B

3,9 3,9 3,9 3,9 5 5 5

-20 -10 0 +10 +20 +30 +40

Estimating Spread Curves Chapter 2

Figure 2.2: Single-curve (SC) estimates for 2 June 1998.

3% 0 6% 10 SC AAA spot SC AAA spread 20 30

AAA bonds SC AAA s.e.

1.0 0.5 0.0 1998 -0.5

BBB bonds SC BBB s.e.

0 1998 40 30 20 2003 2008 2013 2018 2023 2028 10

Section 2.6 Results

Table 2.4: Summary statistics of single-curve and multi-curve estimates.

Single-Curve GOVT AAA AA A BBB BB B Multi-Curve AAA AA A BBB BB B

1 .7 3 .3 3 .9 6 .5 25 143 383 2 .6 2 .9 4 .7 19 30 110

Estimating Spread Curves Chapter 2

Section 2.6 Results

SC AA spread: degree 3, knots {3,9} MC AA spread: degree 2, knots {9}

AA bonds MC AA s.e. SC AA s.e.

0.7% 0.6% 0.5% 0.4% 0.3% 0.2% 0.1%

" "