The Society for Financial Studies

Is Default Event Risk Priced in Corporate Bonds? Author(s): Joost Driessen Source: The Review of Financial Studies, Vol. 18, No. 1 (Spring, 2005), pp. 165-195 Published by: Oxford University Press. Sponsor: The Society for Financial Studies. Stable URL: . Accessed: 09/10/2011 18:21
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact

Oxford University Press and The Society for Financial Studies are collaborating with JSTOR to digitize, preserve and extend access to The Review of Financial Studies.

Is Default EventRisk Pricedin CorporateBonds?
Joost Driessen University of Amsterdam
This article provides an empirical decomposition of the default, liquidity, and tax factors that determine expected corporate bond returns. In particular, the risk premium associated with a default event is estimated. The intensity-based model is estimated using bond price data for 104 US firms and historical default rates. Significant risk premia on common intensity factors and important tax and liquidity effects are found. These components go a long way towards explaining the level of expected corporate bond returns. Adding a positive default event risk premium helps to explain the remaining error, although this premium cannot be estimated with high statistical precision.

Compared to the extensive literature on risk premia in equity markets, relatively little is known about risk premia in corporate bond markets. Recent empirical evidence suggests that corporate bonds earn an expected excess return over default-free government bonds, even after correcting for the likelihood of default (see Section 1). In order to explain this excess return (or, equivalently, excess credit spread), existing research has looked at tax and liquidity effects, and risk premia on systematic changes in credit spreads (if no default occurs). However, a complete empirical analysis that incorporates all proposed components is lacking. The main contribution of this article is twofold. First, an empirical decomposition of expected corporate bond returns into the several proposed determinants interest rate and default risk premia, and tax and liquidity factors-is provided. Second, this article is, to my best knowledge, the first to estimate the risk premium associated with a default event, and to assess its importance for expected corporate bond returns. This approach thus explicitly distinguishes the risk of credit spread changes, if no default occurs, from the risk of the default event itself. Estimation of the default event risk premium allows one to test the assumptions underlying the conditional diversification hypothesis of
I thank Frank De Jong, Siem-Jan Koopman, David Lando, Hayne Leland, Pascal Maenhout, Bertrand Melenberg, Theo Nijman, and Ilya Strebulaev for comments and discussions. I also thank Kenneth Singleton (the editor) and two anonymous referees for many helpful comments and suggestions. Finally, I thank participants at the 2001 ESSFM Gerzensee meeting, the EFA 2002 Berlin meeting, the 2002 CREDIT Conference on Credit Risk in Venice, and the Thiele Symposium at the University of Copenhagen, and seminar participants at INSEAD, Tilburg University, the Tinbergen Institute, NIB Capital Management, and ABN-AMRO Bank for their comments. Address correspondence to: Joost Driessen, Finance Group, Faculty of Economics and Econometrics, University of Amsterdam, Roetersstraat 11, 1018 WB, Amsterdam, The Netherlands, or e-mail: j.j.a.g.driessen( The Review of FinancialStudiesVol. 18, No. I ? 2005 The Society for FinancialStudies;all rights reserved. Advance Access publication November 3 2004 doi:10.1093/rfs/hhi009

The Review of Financial Studies / v 18 n 1 2005

Jarrow et al. (2001), who proved that default jump risk cannot be priced if default jumps are conditionally independent across an infinite number of firms. A positive jump risk premium would thus indicate that default jumps are not conditionally independent across firms or that not enough corporate bonds are traded to fully diversify the default jump risk. The Duffie and Singleton (1999) framework is used to specify the model. In case of default a downward jump in the bond price occurs, equal to a (constant) loss rate times the bond price just before default. The product of the risk-neutral jump intensity and the loss rate equals the instantaneous credit spread. Each firm's default intensity is modeled as a function of latent common factors and a latent firm-specific factor. This extends the results of Duffee (1999), who estimated a separate model for each firm. All factors follow square-root diffusion processes. In line with the results of Longstaff and Schwartz (1995) and Duffee (1999), the model also allows for correlation between credit spreads and default-free interest rates, which are modeled using a two-factor "essentially affine" model [Duffee (2002)]. The model also corrects for tax differences between corporate and government bonds, and a liquidity factor that is based on the corporate bond age is included. Finally, the ratio of the risk-neutral and actual jump intensities defines the jump risk premium. In sum, the model allows for six different components of expected excess corporate bond returns. Two components are risk premia on, respectively, market-wide and firm-specific changes in credit spreads. A third component is due to the dependence of credit spreads on default-free interest rates, and a fourth component is due to a risk premium on the default jump. The remaining two components are tax and liquidity effects. The estimation methodology consists of two steps. First, the Kalman filter maximum likelihood is used to estimate the affine factor model for risk-neutral intensities. This step uses weekly US corporate bond price data for 592 bonds of 104 firms, from 1991 to 2000. The results show that the market-wide spread risk, represented by movements in two common factors, is priced, whereas the firm-specific factor risk is not. A negative relationship is found between credit spreads and the default-free term structure. The liquidity factor, which represents the liquidity difference between recently issued and old bonds, is another important determinant of credit spreads. In particular, a downward sloping term structure of liquidity spreads is found. Finally, tax effects also explain part of the level of credit spreads. Next, the implications of this estimated model for actual default probabilities, excluding a default jump risk premium, are studied. In this case the model overestimates observed default rates, and underestimates expected excess corporate bond returns and credit spreads. This "credit spread puzzle" has also been reported by others (see Section 1). 166

a second estimation step is performed. The economic importance of all factors is assessed by decomposing the expected corporate bond return into the different components. The standard error of the estimated premium is obtained by incorporating the sampling error in historical default rates and in the estimates of the factor model parameters. and it leads to a better fit of observed default rates. Section 5 describes the estimation of the default jump risk premium. once common factor risk premia and tax and liquidity effects are included. the jump risk premium mainly matters for BBB-rated bonds. With a t-ratio of 1. The results on the event risk premium obtained in this article differ from the results of Jarrow et al. and common factor risk. Jarrow et al. explaining about 31 basis points of the expected annual return on a 10-year bond. (2001). For AArated and A-rated bonds the most important determinants are tax and liquidity effects. 1. The estimated premium turns out to be positive. the statistical evidence for the existence of a default jump risk premium is inconclusive. In particular. the tax effect and common factor risk are the two other important determinants. 6 and 12 basis points to the expected return of AA-rated and A-rated bonds. the statistical evidence for the existence of a credit spread puzzle is not very strong. besides using corporate bond price data. In other words. Section 1 discusses the related literature. Contributionto Existing Literature This article is most closely related to a set of articles that also evaluate models using both corporate bond price data and historical default rates [Elton et al. in which the default jump risk premium is estimated using data on historical default rates.2. (2001). respectively. (2001) performed a first analysis of conditional diversification. Leland (2002) pointed out the importance of evaluating corporate bond pricing models using observed default rates. The rest of the article is organized as follows. and Section 3 describes the corporate bond data. Section 2 introduces the model. it is shown that the use of the Markov migration model leads to downward biased estimates of the default jump risk premium. They use the estimates of Duffee's (1999) model to compare the model-implied default probabilities (in the absence of a default jump risk premium) with the default probabilities that are implied by an annual Markov migration model. 167 . Section 4 discusses the estimation methodology and the results for the risk-neutral intensity model. This analysis shows that. Jarrow et al. Section 6 concludes. followed by a presentation of the results and robustness checks. The default jump premium contributes. (2001) (see Section 5. but they did not estimate the default event risk premium. and Huang and Huang (2002)]. economically. For this rating category.2).Default Event Risk Next.

there are several differences between the study by Elton et al.The Review of Financial Studies I v 18 n 1 2005 Elton et al. Although these pricedjumps do not necessarily cause a default event. Huang and Huang find that the impact of jumps depends strongly on the parameter values for jump intensity and size. this article uses a latent factor model instead of the Fama-French factor model. and liquidity factors. and this article. An important difference between Huang and Huang and this article is that here the model is estimated using firm-level data. They use a firm value framework to assess the influence of several factors on the level of credit spreads. They find that all models imply too low investment grade credit spreads. or. but that it is unlikely that the incorporation of jumps fully explains the credit spread puzzle. credit spreads. while Huang and Huang calibrate structural models to aggregate data without analyzing the statistical accuracy of their results. in line with the results of Huang and Huang and this article. (2003) propose a theoretical model where a large upward jump in a firm's credit spread causes a moderate market-wide 168 . It is then interesting to see that Huang and Huang explain about 30% of the BBB-credit spread. because Collin-Dufresne et al. I find that tax and liquidity effects (not incorporated by Huang and Huang) are important determinants of credit spreads. (2001) show that observable financial and economic variables (including equity returns) cannot explain the correlation of credit spread changes across firms. Huang and Huang also consider a model that incorporates systematic (and priced) jumps in the firm value. including tax. Collin-Dufresne et al. Although they do not empirically decompose the credit spread into all these factors. Moreover. and compare model-implied and observed credit spreads. amongst others. Most of these models are based on a diffusion process for changes in the firm value. They explain part of this fitting error by relating corporate bond returns to equity market factors. so that it is indeed unlikely that default risk alone can explain credit spread levels. historical default rates and the equity premium. First. and did not use historical default rates. (2001) show that expected default losses and tax effects cannot explain the observed level of credit spreads. The risk premia generated by these models can be interpreted as risk premia on common changes in firm values. which is similar to the explanatory power of the risk premia of the common spread factors in this article. Although this article also finds that common movements in corporate bond prices carry a risk premium. do not incorporate a liquidity factor and a default event risk premium. jump. they argue that priced jumps in firm value cannot solely explain the high level of credit spreads. equivalently. Elton et al. Delianedis and Geske (2001) study only the level of credit spreads. Huang and Huang (2002) fit structural firm value models to. they may capture a similar effect as the priced default jumps in this model.

They also incorporatea reduced-form liquidity correction. where Xt=(X1. (2003). The model generates risk premiaand correlationbetweenthe factors._ - "H Kll O 1-X1 t+ -A 0^dt - XA / t 0 0 0 dW. in which the instantaneous default-freeshort rate r. is differentfrom the relationshipbetweenthe intensitiesand these variables.1 Modelsetup I For US Treasurybonds (assumedto be default-free) use a two-factor "essentiallyaffine" model [Duffee (2002)]. Yu (2002) also providesa decompositionof corporatebond returns of framework Duffie and Singleton(1999).Default Event Risk jump in the credit spreads of other firms (a contagion effect).However.t V/ +. 169 . assuming that the relationshipbetween liquidity and variablessuch as the spot rate. A recent example of work in this direction is Bierens et al. 2. X2.which is similarto the default event risk premiumin this article.whichis comparedto the resultsin Section5. rt=o + ' Xt.but using the intensity-based he did not estimatethe size of the components. Bond Wtis a two-dimensional risk premiaare modeledfollowingthe specification proposedby Duffee (2002). the A1(2) model is used.t-dX1t dX2. from the contagion risk premium.t. Janosiet al. which is time-varying to fit interestratedata [Dai and Singleton(2000)]. They distinguishthe directjump risk premium.Theirfocus is on the bond pricingperformance the model.t)and 5= (81.they do not incorporate effectsor a defaulteventrisk tax of premium.21 X.1 tion to obtain an indicationof the size of the jump and contagionrisk premia.t vectorof independent Brownianmotions.The factors important follow an affineprocessunderthe actualprobability measureP -dX. 82)'.2.The Brownianmotion vector Wtundera risk-neutral probability ' It may be econometrically hard to distinguish between the effects of a market-wide jump process and common diffusion factors using discretely observed credit spread data. (2001)estimatean intensity-based modelin whichcommon factors influencethe intensities. -X2. Modelfor Defaultable BondPrices 2. in line with the resultsof this article.but empirically Collin-Dufresne al. performa calibraet theymay capturesimilareffects. et The market-wide jumps of Collin-Dufresne al.and volatility. is a function of two factors. and theircommonfactormodelgivesa good fit of corporatebond prices.. equitymarketreturn. Following the notation of Dai and Singleton (2000). (1) _K21 k22. are differentfrom the commondiffusionfactorsfor creditspreadsin this article.

Assuming the absence of arbitrage opportunities guarantees the existence of an equivalent martingale measure Q. the assumption of a constant loss rate is innocuous. j= 1..u)du (4) with instantaneous spread sj. it is shown that this assumption is only used for the estimation of the default event risk premium. As in Duffee (1999) and Elton et al. Madan and Unal (1998). Although this risk premium parameter may vary assumed to be constant for simplicity.thQL and where EQ denotes the Qexpectation conditional upon the information set at time t. As noted by Jarrow et al. Duffie and Singleton (1999) show that the time t price Vj(t. N. This shows that it suffices to model the instantaneous spread sj. there is a downward jump in the bond price equal to Lj.ttimes the price of the bond just before the default event. it is exceed 1. (3) will If the risk associated with default events is priced. The stochastic intensity of this jump process at time t under P is denoted by hft. this loss rate is assumed to be constant at 56%. where -A/4X 0 0 2/1 +/821Xl. (2001). T) = EQ exp - (r + s)du) (ru s. is given by t Vj(t. and Jarrow and Turnbull (1995).The Review of Financial Studies / v 18 n 1 2005 measure Q satisfies dWt = Atdt + dWt. and not for the estimation of the risk premia of credit spread changes. this also implies that for the modeling of credit spreads and credit spread risk premia. As mentioned above. the intensity hQ under this measure is related to the P-intensity through the risk premium parameter /i hQ = ht. issued by firmj and maturing at time T. a default event is modeled as the first jump of a conditional Poisson process.. As in Duffie and Singleton (1999).T) of a defaultable zero-coupon bond (that has not yet defaulted). (2001).. the parameter Au over time. to price defaultable bonds. In case of a default event at time t. 170 ..t At = .2) Duffee (2002) shows that this model leads to the well-known exponentialaffine pricing formula for default-free bonds. Below..t)/2 0 A2(21) 0 A2(22)_ X t (2) X2.All A12 + 0 0 0 (1 +321X. for firm j. Duffie and Singleton (1999) call this the recovery of market value assumption.

Given the affine processes for the factors under Q.2 Both for the common and firmspecific factors.).2 we discuss the interpretation and estimation of this liquidity component.. which motivates the inclusion of a common liquidity component (13o + P3jlt) in the instantaneous credit + spread. i= 1. where W^F is a Brownian motion i I i.)dt+ o dFi G-dWj.. For most results. 2 Not all parameters in the processes in Equation (6) can be identified. N. and the ar-parameterscan be interpreted as volatility parameters. The K common factors and N firm-specific factors follow independent square-root processes under P dFi.. Equation (4) implies that corporate bond prices are exponential-affine functions of the factors [see Duffie and Kan (1996)].t . plus two terms that allow for correlation between spreads and default-free rates K hQtL= . the 0-parameters represent the unconditional factor means. K.t 1.t dGj.K... Gj. Given the evidence for systematic changes in credit spreads across firms [see Collin-Dufresne et al. Elton et al. corporate bonds are less liquid than government bonds (as discussed in Section 1)..t -O) + aj(X. (5) A latent factor model is used since Collin-Dufresne et al. it is assumed that the market price of factor risk depends on the factor level. (6) Here. In Appendix B it is shown that the identification problem can be solved by normalizing the means of the factors Of. The liquidity factor It is assumed to follow a standard square-root diffusion process under P and Q. i= 1 . (2001)]. coupons on corporate bonds are subject to state taxes.t KF(FKG(O Fi.875%.. For example.K j 1.t. and a firm-specific factor Gjt.. (2001) and Elton et al... 171 . this pricing formula is modified in two ways.. (2001) show that financial and economic variables cannot explain the correlation structure of credit spreads across firms. for each common factor it is assumed W that dWF = dWf + (AF/-F)F)/.. For the liquidity factor It a similar assumption for the liquidity risk premium Al is made. .t i.. this tax rate is used. while government bond coupons are not [Elton et al. In Appendix A it is shown how this tax effect can be incorporated in the intensity-based pricing model. First. the risk-neutral intensity of firm j is modeled as a function of K common factors Fi.t)dt + CF.t under Q.. the K-parametersare mean-reversion parameters. . Second..dt.t-0 ) +a2jX2. so that s = hQtL (Ilj + 13j. (2001) estimated the effective tax rate to be 4..OF) + (Gj.oj+ i=l yj(Fi. (2001)]. In Section 4. i= 1.t.Default Event Risk In order to apply formula (4) to the data on coupon-paying corporate bond prices.

1). Xt. through the dependence of credit spreads on default-free factors. This notation will be useful below.t) + (u . incorporates explicitly the dependence of the functions AI(. Finally. a2i)' and where the functions Bi{.) depend on the model parameters [see.t)Cj(T .t)dt (7) A(8.1)hj.. + (ay+ 5.) dt.) and A2(..t)Atl. 82)'. because one has to incorporate the influence of a default event on the expected return. Using the results of Yu (2002). In total. L + (10j + /3 .) on the parameter vector 8 =(8. Fourth. T) denote the time t default-free bond price for maturity date T.(T. Applying Ito's lemma to the well-known exponential-affine bond price expression of Duffie and Kan (1996) gives E with P(tT) t_P(t. the tax difference between corporate and government bonds also explains part of the pre-tax excess corporate bond return. Appendix C derives the following expression for the instantaneous expected return on a corporate discount bond. The notation used in this article.t . for coupon-paying bonds.T. via a risk premium on firm-specific credit spread changes. because the risk of common changes in credit spreads is priced. . Cj(.T) = rtdt + A(S. Second.) can easily be solved for numerically.) and A2(.t)_-(T- 0 (0 x t)[Al(6. Let P(t.A2(21) A2(22) X The functions Al(. Xt.t)AG.(T . (8) A..The Review of Financial Studies I v 18 n 1 2005 2.t)B(T - t)AFFi.. the expression is slightly more complicated. Pearson and Sun (1994) for explicit expressions for these loading functions in square-root models].. in excess of a government bond with the same maturity K -Z(T i=l .. Fifth.t)A2(8. 172 .. (9) where ay= (aly. and Dj(.T.t)] + . Third. ' 0 - 0 x+ - 0 X . T . the model generates expected excess corporate bond returns in six ways.).. X.2 Expected bond returns and conditional diversification Let us start with the default-free expected bond returns. For corporate bond returns. First. expected bond returns are positively related to the jump risk premium (/. T . T . due to the liquidity difference between corporate and government bonds.t) -A(. Xt.12 A1212 0- ..t)Dj(T .T. e.g..).

(2001). were collected. until February 18. Second. semi-annual coupon payments that do not contain any put or call options. if (1) default processes are independent. The model can however also allow for simultaneous default. mention two reasons why default jump risk could be priced. there may be a (small) possibility that some firms default simultaneously. One interpretation of the model is that the common factors drive intensities. only the most recently issued bond is retained. 1991. of data on bond prices of firms that did not default during the sample period. This analysis only uses bonds with constant. Jarrow et al. Jarrow et al. By estimating the default jump risk premium. it is possible to test the assumptions underlying this "conditional diversification" hypothesis. if a common intensity factor F. which leaves 104 of the 161 firms. refer to this as a case of"diversifiable default risk". More than 80% of the remaining bonds is senior unsecured. observations on bond prices with less than one year to maturity were dropped. This study is restricted to the set of 161 firms analyzed by Duffee (1999).t describes the arrival rate of joint default events. given these two assumptions.1). Other bonds were included only if a bond has the same rating as the senior unsecured bonds. they show that the P-intensity is approximately equal to the Q-intensity (tu. Weekly data from February 22. As in Duffee (1999). At the end of the sample period. A firm is included in this analysis only if there are data on at least two corporate bonds for at least 100 weeks. default jumps can be (approximately) diversified away and are therefore not priced. Description of the Dataset Data on mid-quotes for US-dollar corporate bond prices and bond characteristics are taken from the Bloomberg Corporate Bonds Database. if the maturity difference of two bonds is smaller than six months. Intuitively. with independent defaults conditional upon this intensity process. Also. Because it is not necessary to specify whether default events are conditionally independent or not in order to price corporate bonds that have not yet defaulted [Equation (4)]. First. both explanations for a positive default jump risk premium are in line with the model for default events used in this study. because of the use. and (2) there is a countable infinite number of firms in the economy.. 2000.Default Event Risk At this stage it is appropriate to discuss the "asymptotic equivalence" results of Jarrow et al. 3. In an intensity-based framework. conditional on the path of default intensities. or sinking fund provisions. below. This also implies that it is not possible to distinguish empirically between these two interpretations of the model. all 104 firms are rated 173 . in an economy with a finite number of bonds a positive default jump risk premium may exist even if default jumps are conditionally independent.

8 Median 378 3. This study uses the estimation results for the twofactor affine model in Equation (1). coupon spreads are calculated.1 The table reports summary statistics on weekly observations for corporate bond prices from February 22.0 22. investment-grade: 2 AAA-rated firms. conditional upon two bond prices observed during this week.0 10. "Mean number of fitted bonds" contains the mean number of bonds fitted per week.0 4. Observations for which the coupon spread is above 400 basis points or below -50 basis points. Unreported results show that. 5. 58 A-rated firms. The "middle" observation of observations for which the coupon spread moves more than 50 basis points in one week. Besides corporate bond price data. which will be discussed in Section 4. Figure 1 also shows that there is considerable correlation between the spreads of the different rating categories.6 1. on average three different bonds are used to estimate the model. To further analyze the corporate bond price data.9 11. 3 Some bond price observations are very likely incorrect. 3. and have remained high since.0 2.8 50. For the median firm. and again more than 50 basis points in the opposite direction in the next week are also deleted. for the maturities at 2. spreads increased dramatically.0 7. until February 18. Due to the Russia/LTCM crisis in the fall of 1998. this study also uses Bloomberg data on the six-month US Treasury bill. whereas low-rated firms have higher and more steeply increasing spread term structures. The row "Weeks of data" contains the number of weeks for which at least two bond prices are observed for a given firm.0 15. and 30 years.0 7. and 31 BBB-rated firms. 1991. for 592 bonds of 104 firms. on average.The Review of Financial Studies I v 18 n 1 2005 Table 1 Summary statistics for corporate bond data Across 104 firms Firm-level statistic Weeks of data Mean number of fitted bonds per week Mean years to maturity of fitted bonds Minimum years to maturity of fitted bonds Maximum years to maturity of fitted bonds Mean coupon of fitted bonds Minimum 100 2. Table 1 contains information on the bond data. 2000. and observations that relate to a coupon spread movement of more than 100 basis points in one week are eliminated. averaged within each rating category. 7. 10.7 5. and at 378 of the 470 weeks in the data at least two bond prices are observed for this median firm. to calculate the associated yield-to-maturities of default-free coupon bonds.6 Maximum 470 8.389 bond price observations are eliminated. in line with the summary statistics in Duffee (1999). 616 of the 140. The coupon spread is defined as the difference between a corporate bond yield and the yield of a default-free bond with the same coupon and maturity.3 Figure 1 plots the time series of the coupon spreads.7 1. 13 AA-rated firms. 174 . The graph shows that from 1991 to 1998 spreads have declined. which motivates the common factor model. high-rated firms have a low and slowly increasing spread term structure. Thus. and the most recently issued US Treasury bonds.

However. 1991 1992 1993 1994 1995 1996 1997 WeeklyData Feb-1991 . This avoids using an ad-hoc smoothing method to calculate zero-coupon interest rates.Feb-2000 1998 1999 2000 Figure 1 Time series of the coupon spreads For each week..Default Event Risk 180 L 160 - . I. Kalman filter QML is used to estimate the two-factor affine model in Equation (1) for the default-free term 175 . Appendix D provides details on the Kalman filter QML. The graph depicts the resulting time series. 4. coupon spreads of 592 bonds of 104 firms are averaged within each rating category. The estimation of the default jump risk premium is uL described in Section 5. since the number of parameters is large. this study uses quasi maximum likelihood (QML) based on the Kalman filter to estimate the model for risk-neutral intensities in Equations (5) and (6).. Estimation of the Factor Model for Risk-Neutral Intensities The setup of the model is such that it can be estimated in two steps. First.i~^ Ou 120 60 ' 2 ' AAA/ BBB-rated firms iI '1 ' V ' 'uui:'lr VMe' c~~~~~~~n~~~ ' : c 100 - a 80- 80 -A-rated firms o 6040s AAA/AA-rated firm 20 . this section describes the estimation methodology for the model for the risk-neutral intensity hJ. the estimation is performed in four steps. .. As in Duffee (1999) this estimation directly uses the yield-to-maturities of coupon-paying Treasury and corporate bonds to estimate the parameters. In principle. First. 4.1 Estimation methodology Similar to Duffee (1999). a joint estimation of all parameters is most efficient.

. These spreads are used in step three. This restriction turns out to be not binding for the common factors. the third and fourth step of this estimation procedure are repeated. On analyzing whether applying another iteration leads to important changes in the parameter estimates... but this is excessively time-consuming. using data on Treasury-bill rates and Treasury bond yields. ?? . these two firms are treated as AA-rated firms.d. assuming that the parameters that are estimated in previous steps are estimated without error. which involves the estimation of the common factors parameters and the liquidity parameters/3P and iyj. Finally. Thus. the restriction is binding for 7 out of the 104 firms. This step is discussed in Section 4. j1. For the firm-specific factors.i.t.... To reduce the number of parameters to be estimated in the third step. the coupon spreads are calculated.2 Estimation of the liquidity factor This subsection describes the construction of the liquidity factor it and the estimation of the 13-parameters in the liquidity spread specification 4 Since there are only two AAA-rated firms. i = 1. 176 . For all square-root processes in the model. using Kalman filter QML. In each estimation step standard errors and t-ratios for the parameter estimates are calculated [correcting for heteroscedasticity using the White (1982) covariance matrix]. for the common factors. In principle. 104..2 and also provides a time series of liquidity factor values 11. measurement error assumption is replaced by the structure implied by the estimated firm-specific factor processes and the measurement error structure that was assumed in this fourth step. this method explicitly incorporates the presence of firm-specific factors when estimating the common factor processes. it is possible to calculate the standard errors by taking into account the previous steps.. the restriction 2KFOF > (oF)2. it was found that this is not the case.The Review of Financial Studies I v 18 n 1 2005 structure. by means of bootstrapping. . for instance. as described in Appendix D. the firm-specific parameters a i... 2j. The second step of the estimation procedure involves estimation of the parameters in the square-root process for the liquidity factor lt.4 In step four. These parameters are estimated for each firm separately.. where in the third step the i. j= 1. K is imposed. the parameters are estimated given the Feller condition. For example. IT Using the model for default-free Treasury rates.again using Kalman filter QML. YKjare restricted to be constant across firms that have the same credit rating at the end of the sample period. The constant terms yoj in Equation (5) are also estimated in this step. 4. the residual fitting errors for the coupon spreads from step three are used to estimate the parameters of the firm-specific factors Gj. 8jy.

442 0.and the age of the on bond.and Houwelinget al. dummiesfor each ratingcategory. smallerand largerthanfive years.218 0.84 bp 7.Theestimation in constanttermandslopecoefficient theliquidity strategy described Sections 4.obtained usingKalman low-ageandhigh-age in is model.Elton et al.2). are and ratingdummycoefficients not and maturity.the amountissued.638 0.Thus.380 0. all bonds with bond age less than three years.451 0. For the ratingcategories. smallerand largerthan five years.BBB 7.059 oa1 11.1 and4.Default Event Risk is regressed two liquidityproxies. 177 .The other two portfolios respectively. the panel dataset of coupon spreads of all firms Table2 Estimates the liquidity of premium Estimate on of proxies (a) Regression the couponspreads the liquidity -0.023 age x maturity modelfor liquidity resultsof the one-factor spread square-root (b) Estimation K Standard error 0.Table 2 shows that bond age is a significant of determinant creditspreads. again with maturities.48 bp 6.264 issued) log(amount -0. (2002). are not equallyweighted. with maturities.38bp 20.AA 31. and Chakravarty Sarkar(1999).282 of model(Section factorin thebenchmark for estimates theliquidity Thetablecontains 4.Instead. areincludedas controlvariables. four portfoliosof all corporatebonds are formedin this dataset.the weightsare chosen to match the of durationand the ratingdistribution the first two portfolios.419 0.028 0.273 P I.14 bp 0. (2003)also find that bond age influences Based on these regressionresults.Theselattertwo portfolios respectively.043 0. To start with. contain bonds with bond age above three years.By calcuthe differencebetweenthe averageyield of the low-age portfolios lating (j3o +13ljlt).031 issued)x maturity log(amount 1.while the amountissuedis not statistically significant.the bond maturityand dummyvariables To effectsof liquidity.017 0. Estimates a one-factor square-root (b) usingthe NeweyandWest(1987)covariance for filterQML.200 0.A P I.Foreachbond. allowfor maturity both the amount issued and the bond age with maturityare allowed to interactwith maturity.The maturity autocorrelation and for are errors corrected heteroscedasticity first-order shown.standard modelfor of matrix.432 1.AA P1oA /3O.the bond age is used as a liquidity indicator. creditspreads. for a given week.429 Age -0. The first two (equallyweighted)portfolioscontain.009 Al 0.BBB 0.2. The age-effect decreases as bond maturity increases.58bp 01 constanttermand slope for estimates the liquidity (c) Factormodelfor creditspreads: 3 O.361 1.(c) Estimates the between the spread bonds.(a) Estimates bond of a regression all couponspreads[in basispoints(bp)]on the four listedexplanatory variables.

The graph shows a positive relationship 5 To simplify the estimation procedure.6 basis points. which implies that long-maturity bonds have a lower liquidity spread than short-maturity bonds. and the time series of the average corporategovernment coupon spread. versus the high-age portfolios. it is possible to obtain. Estimation is again performed using Kalman filter QML.T/e Review of Financial Studies I v 18 n 1 2005 Oa100- l 7n0i C ? o 50- A PVIu 11 Liquidityspread 1991 0 1992 1993 1994 1995 1996 1997 1998 Weekly data: Feb . for each week.s Results are given in Table 2b.2). The downward sloping liquidity term structure is also in line with the results of Janosi et al. (2002). The graph depicts the filtered factor values for this liquidity spread and the average coupon spread of all bonds for each week. the liquidity spreads of the portfolios are treated as spreads on zero-coupon bonds with maturities equal to the average durations of the portfolios (3. The unconditional expectation of the instantaneous liquidity spread equals 20. both a short-maturity and long-maturity liquidity spread. Next. respectively). This is consistent with the regression results in Table 2a. a one-factor square-root model is estimated for the spread between high-age and low-age bonds (see Section 4. a one-factor square-root term structure model [similar to Equation (6)] is estimated for the liquidity factor Itusing the time series of these two liquidity spreads.1991 untilFeb 2000 1999 2000 Figure 2 Time series of the liquidity spreads Using Kalman filter QML.2 and 8. Figure 2 plots the time series of the liquidity spread It (obtained using the Kalman filter). 178 . The risk premium parameter is actually positive.3 years.

At the average maturity of 8. Using a likelihood ratio test6 results in a model with two common factors for the intensity model in Equation (5). Also. there seems to be a downward trend in the liquidity spread.7 basis points between these two bonds. which has also been documented by Chen and Wei (2002) and Janosi et al. reported in Table 4. 179 . It is assumed that the liquidity of a bond that has just been issued with a large amount issued is similar to the liquidity of government bonds. for which the amount issued is about one fifth of the 95th percentile and for which the bond age is 9.7 basis points [for a bond with the coupon rate equal to the average coupon rate for the median firm (7. The estimates for the interest rate risk premia are such that government bonds earn. This is compared with the average bond. thus allowing credit spreads to load on It through the parameter /3j. Based on these results.Default Event Risk between liquidity and credit spreads. The average liquidity spread is estimated using the regression results in Table 2.Effectively. the mean absolute yield error is 8.1. liquidity and credit spreads are both high in the beginning of the sample period and during the Russia/LTCM crisis.3 Estimation results for the risk-neutral intensity model Table 3 presents the estimation results for the default-free part of this model.6%)]. The liquidity model also contains a constant liquidity term /3j. the parameter /3 is chosen such that the unconditional expectation of the total liquidity spread at 8. on average. show that the market prices of risk of both factors are negative and jointly statistically significant. since the model in Equation (5) already contains a constant term Yoj.6 years. part of this constant term is interpreted as a liquidity component. indicating that corporate bond 6For this likelihood ratio test.2 years maturity equals 12. Across all maturities. In particular. it seems safe to assume that the liquidity factor is related to the time variation in market-wide liquidity. the regression results in Table 2a imply a yield difference of 12. This slope parameter is estimated in step three of the procedure described in Section 4.05 basis points. which allows one to fit the average liquidity spread. a positive return in excess of the risk-free short rate [Equation (8)]. (2002). The estimates. For each rating category and given the estimates for all other parameters in the risk-neutral intensity model.2 years. For this bond. The fit on the Treasury yields is reasonable. (2001). the age is set equal to zero and the amount issued equal to the 95th percentile of the empirical distribution of the amount issued. This is in line with the liquidity time series of Liu et al. This term can be estimated separately. the non-normality of the factor changes is neglected and a normal approximation is used for these changes. 4.

t).026 (0. 10.061 (0.111) A rating 26. low-rated firms are more sensitive to the common factors than high-rated firms.006) -0.022 (0. whereas the loading function B2j{T.R Y2.The Review of Financial Studies I v 18 n 1 2005 Table 3 Kalman filter QML estimates of the two-factor essentially affine model for default-free rates Parameter Kll K21 Estimate 0.152 (0.054 (0.010) -0. First.28 bp (15.182 K22 All 01 So (1 A12 A2(21) A2(22) Using Kalman filter QML (see Appendix D).049 (0.619 (0.590 (0.151) AA rating 12.359 (0. Table 4 Kalman filter QML estimates of the common factor model for intensities Factor i KiF AiF oiF Factor 1 Factor 2 YO. there is a negative relationship between spreads 180 . For both factors.871 (0. 3.t) in Equation (9) that is increasing with maturity (T. The parameter 0/ is normalized to 50 basis points for both factors. 5. It is assumed that all interest rates are observed with i.041) -3.052 (0. so that the second factor mostly influences short-maturity spreads.177) 0. and 30 years.235) 0.007) 0.642) -0.39 bp (14.007) (0. the benchmark model with two common factors in Equations (5) and (6) is estimated (see Appendix D).292) 0.002) 0. For both factors.006 (0. both factors influence credit spreads in the same direction for all firms. This evidence for common factors in risk-neutral intensities is in line with the results of Das et al. spread term structures are more steeply increasing for lower ratings.001 (0. the estimates for the factor loadings yij are positive for all three rating categories (AA.i. (2002).R 0.017) -0.125) -0. Indeed.018) Using Kalman filter QML.R YI. second.038 (0.017) 0. measurement errors. A.002 (0.d.067) 0.204) 2.012 (1.31 bp (21.152) 0. in line with results of Longstaff and Schwartz (1995) and Duffee (1999). investors demand a positive return in excess of default-free bond returns.057 (0.012) BBB rating 54.853 (0. spreads are more volatile for low-rated firms.007) 0. 7.629 (0. Standard errors (in brackets) are corrected for heteroscedasticity using the White (1982) covariance matrix.521 (0.191 (0.62 bp) 0.060) -0.481 (0.162 (0.001) Parameter 52 P21 Estimate 0. a higher value for yij implies both steeper spread term structures and more volatile spreads. the two-factor essentially affine model in Equation (1) is estimated using weekly data on the six-month Treasury-bill rate and Treasury bond yields with maturities of 2. and.R 61a lR S622. The explanation for this result is twofold.014 (0. Thus. who provide evidence for common factors in actual default probabilities across firms.200) 0.981) -0. to be compensated for the risk associated with common spread movements.326 (0.075) -0. Standard errors are calculated using the White (1982) heteroscedasticityconsistent covariance matrix. Finally.012 (0.682) 1.092) -0.004 (0.18 bp) 0.66 bp) 1.t) is decreasing with maturity. The first factor has a loading function Bj(T.058 (0.024) -0.444 (1.003) 0. and BBB).

147 Median 0.015 39. a model without common factors is also estimated [K=0 in Equation (5)].027 0.Default Event Risk Table5 factorparameter estimates Firm-specific Firstquartile Estimate KG Estimate Ak oG Estimate 05 Estimate K=0: Estimate AG 0. which is much lower than 0. Estimatesare obtainedusing the QML based on Kalman filter (see D). The firm with the worst fit has a mean absolute yield error of 17.004 0.63 basis points for the median firm. The average of these cross-firm correlations is 0. is negative for almost all firms.049. reported in Table 5.7 basis points. the cross-firm correlations of weekly changes in these firmspecific factors are calculated. The results show a positive dependence of credit spreads on the liquidity factor I. which is in the same order of magnitude as the fit on the Treasury bonds.029 -0.348. the constant terms are also positive. This model.053 of The table reportsquartiles parameter estimatesfor the firm-specific factorsin the modelwith two for commonfactors(K= 2) in Equations and (6).092 0. in line with the graph in Figure 2. In this case. which still seems reasonable.064 0. the average cross-firm correlation that is found for the model without two common factors (K= 0). The last row reports estimates the market (5) priceof risk in a model with K=0. The explained variation is however quite small.59bp -0. For comparison. the estimate for the market risk price.008 22. the fit of the model on corporate bond yields is assessed.004 11.108 Thirdquartile 0.7 Thus. firm-specific 181 . by 8A three-factor of For correlation the modelhas also beenestimated.8 The liquidity parameters /30j and /3lj are reported in Table 2c. To compare with the in similarto the estimates 7Estimates for the other parameters the model with K= 0 are qualitatively reported Duffee(1999). Finally.85bp -0.32bp -0. To verify that the firm-specific factors are really firm specific. this modelthe averagecross-firm factorsequals0. Appendix and default-free rates for firms in all three rating categories. the market price of the risk associated with movements in the firm-specific factors is close to zero for the median firm. with firm-specific factors only. after correcting for market-wide spread risk by including two common factors. and much larger in absolute size relative to the model with K= 2. the remaining firm-specific movements in spreads are hardly priced. Most strikingly. The time-series R2 for the changes in credit spreads is also calculated. In order to match the average liquidity spread of 12. Table 5 presents the estimation results for the parameters in the firmspecific factor processes.076. The model with K= 2 has a mean absolute yield error of 8.41 basis points. is similar to the model of Duffee (1999).049 -0.

Estimation of the default jump risk premium A/is based on moment conditions for the conditional default rate. R = AA. and 4 is a parameter vector that contains all other parameters of the factor model.4) denotes the model-implied conditional default rate under the actual probability measure..R(U. The average R2 across all portfolios equals 38. Thus.. (10) where qn. conditional upon no default between time t and t + n and the rating at time t. This illustrates that the latent factor model is capable of providing a good fit of credit spread data.BBB. the assumption of constant loss rates is not likely to have a large impact on the estimates for the default jump risk premium.1 Estimation methodology To set notation. the performance of the model in volatile market circumstances is tested.t + Zj. this default rate is a decreasing function of p/ and a function of the factor model parameters 4.t as a variable that is equal to 1 if firmj defaults in the annual time interval [t.t+lRj.. Given that the mean of hQ is the most important determinant of conditional default rates. 14. 182 . (2001). (2001). This model-implied default rate can be calculated from the process for the actual default intensity hfp. t + 1] and 0 otherwise. Zj. Since only the product h L enters the model. the fit on corporate bond yields at the end of these weeks is calculated. the moment conditions are EP(Zj.90 basis points for the median firm. In formulas.t+n-1 = 0] = qn.7%. Let Rjt denote the rating of firmj at time t. which is defined as the probability of default in year t + n.R(u.t = R. For a final evaluation of the model fit. the mean absolute yield error equals 9.The Review of Financial Studies / v 18 n 1 2005 common factor model of Elton et al. For these 25 weeks. Since h = h t/!-.R(4L. the 25 weeks that have the largest absolute movements in the coupon spreads (averaged across all firms) are selected.A.t+l + * + Zj. /Lcan be estimated. n= 0. >)-9 By confronting these model-implied conditional default rates with the observed default rates. which is much larger than the average R2 (about 17%) of the factor model of Elton et al. including a time-varying loss rate would most likely lower the estimated variation of hQ but hardly change its unconditional expectation. Then. this model excludes the firmspecific factors for this calculation and focuses on the credit spread changes of the equally weighted portfolios across ratings and one-year maturity intervals. Estimation of the Default Jump Risk Premium 5. Both Moody's and Standard & 9 A constant loss rate L is assumed.4). define Zj. 5. Appendix E derives the expression for qn.

one needs a relatively long period to reliably estimate default probabilities. However. Next. and. The firms in a cohort are followed from this date onwards. credit ratings are used at the end of the sample period (February 2000). while the long-maturity bonds provide information on these long-term default rates. if one would use the default rate data for the 1991-2000 period. so that using the 1991-2000 period would lead to very high but imprecise estimates for the risk premium . This has the disadvantage that the credit spread data and the default rate data are not entirely based on the same sample period. Moreover. one would ideally use default rates on the cohort that starts in 1991 up to the cohort starting in 2000. These empirical conditional default rates provide consistent estimates of the conditional expectation in the left-hand side of Equation (10). From an econometric viewpoint the use of two partly overlapping data periods is not a problem. since this estimation is based on averages of conditional default rates and the assumption of 1 stationarity.10 Given that the data on credit spreads start in 1991 and end in 2000. since defaults do not occur frequently. which are based on the 1970-2000 period [Moody's Special Comment (2001)]. Therefore. S&P data [Standard & Poor's Special Report (2001)] show that for 1991-2000 the total number of downgrades was only slightly larger than the number of upgrades. For example. In this case it is assumed that the Moody's and S&P ratings are the same. The S&P default rate data are based on cohorts starting in 1981 up to 2000 [Standard & Poor's Special Report (2001)]. I also perform an estimation based on Moody's data. This study uses both S&P and Moody's data. S&P does not provide default rate data for longer horizons. in the 1991-2000 period. we use a longer data period to obtain historical default rates. Table 1 shows that for the median firm the maximum maturity of the fitted bonds is 15. l For the firms in this sample. cumulative default rates for more than a 10-year period cannot be used. the first step of the generalized method of moments (GMM) [Hansen (1982)] is applied to Equation (10).0 years. The cumulative default rates from 1 year up to 15 years are used. second. longer horizons are not used for two reasons. For both datasets the cumulative default rates are converted into yearly conditional default rates qData (Appendix E discusses this transformation). The average default rates across cohorts are used to average out potential time or cohort effects. However. For comparison. and default rates can be calculated for each cohort. First. This may lead to conservative estimates for A/if the credit ratings of firms have deteriorated over the sample period. given that these rates are fully characterized by the initial rating and the conditioning period n. default rates are low relative to the 1970s and 1980s. 183 .t. which are averages of cumulative default rates of cohorts of firms that are formed each year. and the sum of squared differences between 10Each cohort corresponds to a given initial date and contains all firms with the same credit rating at this date. so that there is not necessarily a systematic rating drift over time.Default Event Risk Poor's (S&P) provide average historical cumulative default rates per rating category.

Given that the model for risk-neutral intensities fully specifies the behavior of government yields and corporate bond yields for all relevant rating categories. it is only necessary to re-estimate the default event risk premium for this model.R (11) over . the estimate for the default event risk premium may be interpreted as a lower bound.2 Estimation results This subsection presents the results for the three models. Thus. Appendix E describes the calculation of the standard error of the estimate for /x. as long as there is no systematic liquidity difference between AA-rated bonds and A/BBB-rated bonds. Estimation of this model is performed as before. Since in reality AA-rated firms contain a small amount of credit risk. the difference between the AA term structure and the government bond term structure is fully attributed to time-varying liquidity and tax effects. Given the complete exclusion of tax and liquidity effects. Figure 3 shows the influence of the risk premium ut on default probabilities. The line "Risk-neutral: BM" depicts the risk-neutral default probabilities 184 . inserting the estimates for the factor model 4. 5. leaving out the tax correction and step two of the estimation procedure described in Section 4. This figure plots yearly conditional default probabilities for the BBB rating category. This approach is also used by Jarrow et al. which incorporates both the estimation error in the parameters of the factor model and the estimation error in the historical default rates. Besides the model in Equations (1)-(6) that has been discussed so far (the "benchmark model"). to calculate credit spreads for A-rated and BBB-rated firms. For brevity. This model is included to see how much of corporate bond excess returns can be explained using default risk factors only [as in Huang and Huang (2002)]. (2001). For the lower bound model. However.The Review of Financial Studies I v 18 n 1 2005 the model-implied and observed conditional default rates is minimized 14 min E(( R=AA.B n=O ))2qn. an "upper bound" model and a "lower bound" model are also estimated.i. but qualitatively the parameter estimates are similar to the estimates reported in Tables 4 and 5.1 (the liquidity correction). instead of the government bond term structure. the tax and liquidity effects are again excluded from the benchmark model. We start with the results for the benchmark model. the estimate for /u can be regarded as an upper bound for the true value. The upper bound model is obtained by excluding the tax and liquidity corrections from the benchmark model. the results for the intermediate estimation steps are not presented.A. the term structure of AA-rated firms is used.

the actual probabilities for the upper bound and lower bound models (UBM and LBM) are given./--"'^ ^ ^ BM 1. obtained from S&P data.=1." :-' '- BM -.51:UBM 0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Conditioningperiod (n) in years Figure 3 Conditional default probabilities The graph contains yearly conditional default probabilities for BBB-rated firms. * " = 2.15 based on the Moody's 185 . These are calculated by setting tt = 1 and using the risk-neutral measure Q instead of the actual probability measure P for calculation of the default probabilities. The yearly conditional default probability is defined as the probability of default in the next year. * "= 1: BM" refers to the actual probabilities in case iu= 1 for the BM. both for the case where /L= 1 and at the estimated values (5. Furthermore.._31: Z. under the risk-neutral probability measure for the benchmark model.5 - 2 lVl g=1:UBM _. the risk premia on the intensity factors and the tax and liquidity effects cannot fully explain the level of observed default probabilities. Next. The difference between these actual and risk-neutral = probabilities is completely caused by the risk premia on the factors that drive the intensities.31 in case of the S&P data and 2.51 and 1. These default rates are all below the default probabilities implied by the model with . the risk premium tuis estimated. respectively).31) for the BM.80.5- H=5. Thus. since the risk premia on these factors imply that the expectation of the path of the intensity under Q differs from the expectation under P. LBM ---BM =2. assuming that there is no risk premium on the default jump (AL 1).l = 1 (this is also the case for AA and A ratings). Table 6a shows that this gives an estimate of 2.'-. The line " L= 1: BM" shows what the model implies for actual default probabilities.5- Ri-^msk-neutral: 10o __-=1*: gj_ _ . The line "S&P Data" gives the historically estimated default probabilities.31: BM" refers to the actual probabilities at the estimated value for At (2. Figure 3 also contains the empirical yearly conditional default rates based on S&P data from 1981-2000.. The other lines all are model-implied default probabilities: * "Risk-neutral: BM" refers to the risk-neutral probabilities for the benchmark model (BM). LBM S&P Data 0. given that no default has occurred before.80: .Default Event Risk 2.

and 3-year conditional default rates are slightly lower than the model-implied probabilities. it is statistically not clear that the observed credit spreads are "too high" or that the average excess corporate bond returns are "too high. Estimation is performed as described in Section 5." Although the statistical precision for /l is not very large.31 (1.03.05) 1. which gives the decomposition of expected excess corporate bond returns proposed in Equation (9). The advantage of the common factor model over Duffee's model is that it provides a decomposition of the total risk premium on the risk of intensity changes into common and firm-specific risk.875%.12Investors thus multiply the actual default probability with a factor of more than 2 for the pricing of corporate bonds.11) S&P data: 1981-2000 2.03 (1.83 (1. 2.77) 2. S&P data: 1981-2000 Moody's data: 1970-2000 2. This implies that once the intensity risk premia. including a default jump risk premium clearly improves the fit on historical default " = 2. due to the estimation error in the model parameters and estimation error in the historical default rates.53) 5.11 and 1. (c) Results for the joint estimation of the tax rate and the event risk premium. as shown by the line Figure 3. the default jump risk premium implies an excess return of 31 basis points 12The default jump risk premium has also been estimated using Duffee's model with only firm-specific factors to model intensities.61 (1. and tax and liquidity effects.37 (1.30 (1.03) Moody's data: 1970-2000 Lower bound model Upper bound model 1. Therefore.83) BBB Category 2. respectively).78 (1. (a) Estimates based on all rating categories together.44) (b) Estimates for the event risk premium uLper rating category for S&P data: 1981-2000 -7. 186 . In this case. For example.39) A Category 1.31: BM" in rates. Calculation of standard errors is described in Appendix E.92) (2.13 The economic contribution of all determinants becomes particularly clear from the plot in Figure 4.47) (c) Benchmark model: Estimates for the event risk premium /J and tax rate Tax rate (%) Risk premium .15 (1. data.67 2. the estimate for the default jump risk premium is of similar size as reported in Table 6.32 (2.60 (0.67 (1.18 AA Category 1. but the individual contributions are quite different. The tax rate in the benchmark model in (a) and (b) is 4. obtained using S&P data and Moody's data. it is impossible to conclude with certainty that tLis larger than 1. (b) Estimates per rating category.51 (1.22) 5. The standard errors are however quite large (1.29) 2.97) 5.The Review of Financial Studies I v 18 n 1 2005 Table 6 Estimates of the default event risk premium Benchmark model (a) Estimates for the event risk premium . 13Results for AA and A ratings are similar.80(1. 2-year. The graph shows that all components contribute to the expected return.46 3% 3% The table contains estimates and standard errors for the default event risk premium. are corrected for. for a 10-year coupon-paying BBB bond.18) 4. For all rating categories the observed 1-year.

. the tax effect explains 33 basis points.' . and 4 basis points. which lowers the expected corporate bond return.2. i . For example. the conditional expected bond return for each week is calculated by applying Equations (9) and (A. ' ::: "DefaultFree Factors Risk ' -0. 187 .. their calibration results in a risk premium of 5 basis points for investment grade firms.:. Since credit spreads are negatively related to default-free rates.:. ':: '. -9. The tax effect is calculated by comparing the tax-corrected bond price [Equation (A.00% ?: :: :" :-i"-' ::':::: :'. and tax and liquidity effects are included according to the benchmark model. while tax and liquidity effects are essentially the same as for the BBB-rated bond.. .20% 1 2 3 4 . 23.::: ' '.00% 0.. (2003). The return is calculated in excess of the return on a government bond with the same maturity and coupon. and the liquidity factor almost 13 basis points. -.20% g ~~~~ I ) 0.. .20% 1.40% i Default Event Risk F irm -specific 'Common Factor Risk Risk :::: Factors :' :: ' 0. For this 10-year bond. :: :: '. For a 10-year A-rated bond the jump premium is 12 basis points. and subsequently averaging over all weeks.. These results can be compared with the informal calibration of Collin-Dufresne et al. the default jump risk premium generates an excess return of about 6 basis points. Using the fact that a coupon bond is a portfolio of discount bonds.80% r _- Tax Effect _ 0. while the interest rate and the common and firm-specific factor risk premia generate. for different maturities.40% 0. they provide a partial hedge against interest rate risk. They effectively set = 2.Default Event Risk Corporate Bond Expected Excess Returns: BBB Rating Benchmark Model 1. for a 10-year AArated bond. 1)] with the uncorrected bond price and annualizing the percentage difference (given the bond maturity). but instead use a lower probability. per year. respectively. the graph provides the model-implied estimate for the annualized expected return on a coupon-paying corporate bond with coupon of 7.. as discussed in Section 2. The contribution of firm-specific and default-free factors is thus rather small.00% 0.. The average return is decomposed into four risk premia. 5 6 8 7 9 in Maturity years 10 11 12 13 14 15 Figure 4 Corporate bond expected excess returns:benchmarkmodel For the median BBB-rated firm.:::: .6% (the average coupon for the median firm). the tax and liquidity factors become more important since they (hardly) vary across rating categories. :::: '::.:. For higher rating categories. which is close to the estimates in this article.60%lm SLiquidityPremium 0. inserting the factor values for that week. Because they do not use historical default rates to estimate the default probability. 1).

It turns out that the historically estimated cumulative default rates. which compares well with the common factor risk premia in this article. In Section 1 it was shown that the common factors may capture similar effects as the contagion jumps of Collin-Dufresne et al. 1. Finally. It also implies that using the Markov migration model leads to a downward bias in the estimate for . is much flatter than the shape of the conditional default probabilities implied by the Markov migration model. Furthermore. . Since /1 is not expected to vary largely across rating categories. The results for the upper bound model show that excluding tax and liquidity components has a dramatic effect on the estimate for tL. and the variation across rating categories is not very large. This is evidence against the assumption of the Markov migration model that rating migrations are independent over time.51 (Table 6a).u is estimated for the lower bound and upper bound model. (2001). with / = 1. and 23 (BBB) basis points for a 10year coupon bond. First. They use the estimates of Duffee's (1999) model to compare model-implied default probabilities. The results in Table 6b provide some support for the framework. 14 (A).80 and 5. Jarrow et al. Second. To check 188 . The estimate is larger than 1 for all models and rating categories. these shape differences disappear to a large extent. To put the benchmark results into perspective. Collin-Dufresne et al. as used in this article. respectively. /A is estimated for each rating category separately. this small difference between benchmark and lower bound estimates supports the correction for tax and liquidity effects in the benchmark model. Figure 3 shows that. (2001). Only for the upper bound model. The results from this study can also be compared with those of Jarrow et al.875%. (2003). The migration model uses a one-year Moody's migration probability matrix. Given that default risk is small for AA-rated firms.The Review of Financial Studies / v 18 n 1 2005 The common factors generate economically and statistically significant risk premia of about 9 (AA). a number of robustness checks are performed. which gives estimates of. The estimate is very large in this case and the fit of observed default rates is less good (Figure 3). reach an upper bound for the contagion risk premium of 39 basis points. the estimates for tu are significantly different from 1. So far the results are based on a tax rate of 4. so that a large part of the AA spread is likely to be caused by tax and liquidity effects. find that the shape of the conditional default probabilities. the estimate for /u is relatively close to the benchmark estimate. with default probabilities that are implied by a Markov rating migration model.L. if one uses the conditional default rates that are directly observed in the data. this can be interpreted as a crosssectional test on the model. the tax effect is analyzed in more detail. as proposed by Elton et al. as implied by the intensity model. For the lower bound model. are much lower than the cumulative default rates implied by this Markov model for rating migrations.

(2001). the net coupon received is (1 .tg). there is a 189 . This tax rate changes the corporate bond price in two ways. For eachtax rate.Using 1%intervals. ConcludingRemarks The maincontribution this articleis that it providesa carefulempirical of of decomposition corporatebond returnsinto severalfactors. the differencebetweenthe estimatesfor px acrossratingcategoriesis indeedsmaller.Therefore. which is denote by r.It models can explainthe to would be interesting see whetherutility-based size of the estimatedjump risk premium[see El Karoui and Martellini (2001)]. Given that creditspreadsare lowest for high-rated firms.r)C.Unreportedresultsshow that.decreasing tax rate and increasingLI betterfit of defaultprobabilities acrossratings.Default Event Risk whetherthis value is also appropriatefor this sample.Keswani(1999)and Duffieet al. the effective tax rate is ts(1 . The fact that this estimateis smaller than 4. There are several extensions worth exploring in future research. Let ts denote the state tax rate and tg the federal tax rate. As shown by Elton et al.. at the tax rate of 3%. estimates in creditspread datawith(model-based) intensities on risk-neutral on of actualdefaultprobabilities the firmlevel[Das et al.and a risk credit spreadmovementsare other important premiumon market-wide of of returns investment determinants theexpected gradebonds. 6.875%can be understoodas follows.Liquidityand tax effects.Finally. (2002)]. Second.Table 6b implied shows that the AA rating has the lowest estimate for the jump risk the leads to a premium.includinga associated with thejumpin pricesin case of a defaultevent.but the statisticalevidenceis inconclusiveabout the existenceof a jump risk premium.the relativeimpactof the tax rate on modeldefault probabilitiesis highest for these high ratings.moreresearch neededto obtainpreciseestimates information Thismay be possibleby combining of thejumpriskpremium. (2003)studythe pricing of defaultablesovereigndebt. However. if the coupon size is denoted by C.The article thus providesa first indicationthat defaultjump risk may not be fully is diversifiable. The model in this articlecan be used to analyzethe joint pricingof sovereigndebt spreadsof manycountries. a grid search is to performed see whichvaluefor the tax rategivesthe lowestvalueof the A goal functionin Equation(11). joint estimationis infeasiblesince the factor model has to be re-estimated for it each tax rate.Another extensionwould be to includehigh-yieldbonds in the analysis. was foundthat the optimalrateis 3% for both the S&Pand Moody'sdata.t is re-estimated. riskpremium Thisriskpremium significant of BBBcorpopart explainsan economically rate bond returns. First. Appendix A: Tax Correction on Corporate Bond Prices This appendix estimates the tax correction in the intensity-based framework.

1) and replacing the Q-Brownian motion Wft with a P-Brownian motion WfV. K. i= 1.. (t-. This leads to the following valuation equation for a corporate bond that has n coupon payments and associated payment dates Ti.T)L)ds)] Appendix B: Parameter Identification From Dai and Singleton (2000) and Duffee (2002) it follows that all parametersrelated to the default-free and firm-specific factors can be identified. /OfaF). As discussed in Section 2.T) eau nodefault + (h. The identifiable parameters are the parameters in the processes under P and Q.. These four reduced-form parameters are a function of five structural parameters.r)L. under the risk-neutral measure Q is given by dFi. The derivation is based on Yu (2002). Then.d i = 1. KF +A./ Studies v 18 n 1 2005 TheReview Financial of tax recovery on the default loss if the firm defaults. which changes the loss rate to (1 . and Af. In Equation (5) it is shown that the contribution of the common factor i to the instantaneous spread of firm j is given by Fij. Appendix C: Expected Corporate Bond Returns This appendix derives the expression for the instantaneous expected excess return on a corporate bond.X. (C..n) = ( )CEEQ exp (rL+ .. Equation (C.t) dt.. KF + AF. It then follows that. Ki.. -(T-(t -tj(Bi (T (T-t)B()AFi.. The process of Fij.t A(aj + . OFis normalized to 50 basis points.. it is not possible to recover the remaining Jyo-f). [ VT no default = [rt + s. Tn Vj(t. First. . This appendix analyzes which parameters in the common factor processes can be identified. For another firm k the identifiable parameters are (y ijKF. T)...1) The process under the true probability measure is obtained by removing AF from Equation (B. (B. + t)AjfGG. Applying Ito's lemma to this equation gives the expected return in case of no default E.fAt)(-L) At.. and that normalizing oF. solves this identification problem.1) neglects the loss in case of default. K..2) 190 . (C. t + At) the default probability approximately equals hf At and the loss in case of default equals L V. It is assumed that this loss rate applies to the net coupon payments. the model implies that corporate bond prices are an exponential-affine function of the underlying factors. for a moment. K).t)dt + VyFij Fdwi.T...T.t =(Yi (KF-KF + Af)Fj. the possibility of default is neglected. In a small time interval (t. 1) t)Dj(T-t)A/tl Of course. KFiF. besides K(F structural parameters from the reduced form parameter estimates. (y..f h+ )L)ds ] (A. = yijFi. the expected return over the next time interval approximately equals {(1 -hAt)Et V(T) V j(t. 1) + EQ[exp(-f (rs + 3Ij + 3jls + hQ(1 .

the common factor model for the intensity process is estimated twice. all available observations for that week are used. Appendix D: The Kalman Filter Setup This appendix briefly describesthe setup for the extended Kalman filter estimation of the affine pricing models for default-free and defaultable bonds. Given the estimation results for the firmspecific factors.t.F.t-+h. Discretizing (D. using notation that is unrelated to the notation in the main text.+h= z(Ft+h) + . E. For the default-free model. to construct the weekly likelihood contribution. For each block the conditional covariance matrix of yield residuals V.T. .2). the function z(Ft+h)is nonlinear. 0 and a are K-dimensional vectors. The return in excess of a default-free bond (with the same maturity) is obtained by subtracting the expected return in Equation (7) from the expression in Equation (C..Xt. Therefore. as implied by the estimates for the firm-specific factor process of this firm.1) gives the following transition equation Ft+h = Et[Ft+h Ft] + r1t+h.t)A i .(T ... always some corporate yields are observed in the dataset.(t+h) = V(Ft+hlFt) (D.t) . it is assumed that the covariance matrix of et is a diagonal matrix. Without loss of generality we take the matrix A to be diagonal. Vt(t+h) = Ht. Let Ft be a K-dimensional vector with process under P given by dFt = A( . V.3) the likelihood function can be constructed [see De Jong (2000)].. where each block relates to all corporate bond yields of a single firm. The term (a + B'Ft)1/2 is a diagonal matrix with diagonal elements given by the square root of the elements of the vector (a + B'Ft).L. As in Duffee (1999). Yield measurement errors across firms are assumed to be uncorrelated. a Taylor approximation of this function around the one-period forecast of Ft+h is used to linearize the model.(Et+h) is used.. KK). The measurement equation for Yt+his then given by Y.t)Allt. Since coupon-yields are used.(T.h. (D. Let Yt denote an N-dimensional vector with the observed yields at time t.3).)dt + E(a + B'Ft)/2dWt.2) and (D. (D.1) Here A.3) Here z(Ft+h)is a function that relates the yields to the factors and Et+h is a zero-expectation measurement error that is uncorrelated with Ft+h.t) (T Cj t)\GAj FI (C. which gives the instantaneous expected return K rt + sj.3) + (aj + . In the first estimation. it is assumed that Ht = o-IN.2) Refer to De Jong (2000) for expressions of the conditional expectation and variance in (D. . since one can transform any affine model into a model with diagonal A by rotating the factors. and B are K by K matrices. A = diag(K1. Each week.. Using Equations (D. the common factor model is now re-estimated using a block-diagonal structure for Ht. Thus. For the estimation of the firm-specific factors the treatment of missing 191 . there is a state-space model for these residuals again with a diagonal measurement error covariance matrix. which gives Equation (9)..No).Default Event Risk This expression becomes exact if At -* 0. Ht = diag(o. This dataset consists of yields on coupon bonds.C(T i=1 - t)B. For each firm. An important issue for the estimation of the common factor model is the presence of missing observations. and Wt is a K-dimensional vector of independent Brownian motions.t)Dj(T.(T .. After this first estimation the firm-specific factor parameters are estimated using the residuals from the common factor model.

" Appendix E: Estimation of the Default Jump Risk Premium This appendix first describes the calculation of the model-implied conditional default rate 4). The estimation method in Equatin (11) is the first step of GMM [Hansen (1982)]. 1) can also be used. the results strongly support the choice of the Kalman filter...t.j((r.1) exp -/ The expectation in Equation (E.R(l.l. as ql4.. 4).1) can be calculated explicitly because the process for hJt is affine. 4) that a firm defaults q. qO. To correct the asymptotic variance of the estimator for /u for the estimation error in >.BBB(A.95).Pn. .i is calculated by incorporating both the estimation error in the historically estimated default rates and the estimation error in the factor model parameter estimates. the unconditional expectation of Equation (E. within the next n years. () is defined as (qo.A(/. They document considerable small-sample biases for the EMM/SNP method. similarly. Since the average fitted factor values are mostly lower than the unconditional means. is not normal. As in Duffee (1999). Consistent parameter estimates can be obtained by using the efficient method of moments [EMM. Gourieroux and Monfort (1995)] rF(qData q(~.t}] + hds) -E t exp Jt 'ds) (E.AA(-l). The firm's rating matters since the common factor parameters vary across rating categories. Equation (E. Note that in Equation (E.+ 1-EP _ Zj. . which makes the QML estimator based on the Kalman filter..R(l-. Assuming that all factors follow stationary processes under the true probability measure. Duffee and Stanton (2000) compare EMM/SNP estimation of affine term structure models with the Kalman filter QML estimation. Simulation experiments by Duan and Simonato (1999) and De Jong (2000) show that the induced biases are small. . . O)/lIt are also defined.AA(. Using the following result for the actual probabilitypn. the length of the time interval h is allowed to vary over time in this case. Moreover.The Review of Financial Studies / v 18 n 1 2005 observations is different because there are weeks with no observations at all for a single firm. the unconditional expectation and (co)variances of the factors can be used to initiate the Kalman filter. strictly speaking. Instead of this average. this will most likely lead to higher estimates for AL. 4))' and.4). q14. (E. inconsistent. 0)) are calculated.j(.BBB(/4). ').1) is averaged over all factor values (obtained from the Kalman filter) for the 1991-2000 period. the sample counterpart vector qData well as F = aq(/.t. O))(1 .pn+lj(t.t pnj(ty. q14. and these conditional default probabilities are averaged over all firms in a given rating category to obtain qn. Finally.0.t+l + .. the yearly conditional default rates qn.R.2) 192 .t. .qO. F2.t+n-l Rj . The standard error for the estimate for . Gj. The resulting probabilities are denoted by PnJ{L. it can be noted that this estimation method is quasi maximum likelihood because the conditional distribution of Ft+h conditional on F. conditional upon that no default has occurred yet E[Zj.{FI. The GMM first-order conditions imply that [see. 0). q(.. + Zj. t. Finally. since the model does not allow for rating changes.l.. e. To average out the factor values. and conclude that "for reasonable sample sizes. Refer to De Jong (2000) for all equations in the Kalman filter recursion. the conditional variance in this distribution depends on the unknown values F.t.A. 1) the time-dependence of the firm's rating has been taken out. Gallant and Tauchen (1996)].g. ?)) P 0. ). ) 1 -(1 .A(/. combined with the semi-nonparametric(SNP) method of Gallant and Tauchen (1992).

References withRe-balance.u. The issue is that a large fraction of the firms in the AA-cohort that starts in. while for the 14-year conditional default probability q14. S&P and Moody's report the number of firms per cohort (and per rating category).R. for the default probability in the first year."Specification Analysisof AffineTermStructure Dai.." working Chen. J. and that there is zero correlation between the estimated default rates across rating categories."AnEconometric H. and A. The variance of this default probability estimate can then be determined as nqDRa( qDta) /NR. and over the cohorts starting in 1970 up to 1986 in case of Moody's data. Goldstein. Costsfor Corporate Estimate Transaction of Bonds. Kong. 1943-1978. Changes. where NR is the total number of firms in all yearly cohorts of a given rating category. First." Journal of Finance. andJ.)) - F ( (^ .PennStateUniversity. Huang. Modelof CreditSpreads Bierens. paper. For both terms on the right-hand side of Equation (E. it is assumed that there is perfect correlation between the estimated default rates for different n within each rating category. It is assumed that the estimation error in the historical default rates is uncorrelated with the estimation error in the parameters of the factor model. Journal Models.3). Goldstein.2002.Default Event Risk Using a first-order Taylor expansion for q(.it) P (qD (aaq(. 4) (at the parameter estimates) and rewriting gives (/i .Government workingpaper. 55.. both terms on the right-hand side have asymptotically a normal distribution (if scaled with the square root of the number of observations).3). R. A more difficult issue is the correlation between the observed default rates. the variances of the observed conditional default rates q. it is possible to construct the block-diagonal covariance matrix of all these estimates. Helwege.Purdue and University FederalReserveBankNew York. liquidity factor. To be conservative. it is possible to calculate the full covariance matrix of the default rates qn. will also be present in the AA-cohort that starts in 1983.." workingpaper. For example.3) Given the appropriate regulatory conditions.Carnegie Contagion the Updatingof Beliefs. Q. of A in S. R. 56. Martin." of Collin-Dufresne."AnIndirect StateUniversity.. say. (E.. Singleton. qla. Wei. and finally the covariance matrix of the average conditional default rates of all firms in a given rating category. 2177-2207. Chakravarty. Using the delta-method. MellonUniversity. by assuming that in each year defaults are independently generated by a binomial distribution with probability qDta. Given these assumptions.3) is calculated.and J. ARCHand JumpEffects. J. the covariance matrix of the second part of the right-hand side of Equation (E. S.1999. Sarkar.2003. The standard error for Alcan then be calculated using Equation (E.2001.Dta are calculated. Second. 193 ." of 2000.andW. one can construct the covariance matrix of the model-implied conditional default rates for all firms." Bid-AskSpreadin Corporate.. and K. the calculation of the asymptotic covariance matrix is explained below. S. 1982. the sum is taken over all cohorts. "Is CreditEvent Risk Priced?Modelling P. "The Determinants CreditSpread P. and firm-specific factor parameters) are independent from each other. the sum is taken only over the cohorts starting in 1981 up to 1986 in case of S&P data. via working paper.R.L. and these are summed over all relevant cohorts. It is assumed that the four sets of parameters (default-free. common factor.and J.2003. Finance.Michigan Collin-Dufresne."Liquidity US FixedIncomeMarkets: Comparison the and MunicipalBond Markets. Thus.

2001. A. University of California.. 1996. and K. "The Components of Corporate Credit Spreads: Default. Singleton. R. "How Much of the Corporate-Treasury Yield Spread is Due to Credit Risk?. and L. Tauchen. N. 405-443. 12. Lando. Duffee. G. "Term Premia and Interest Rate Forecasts in Affine Models. 300-318.. pp. De Jong. 657-681. "Large Sample Properties of Generalized Methods of Moments Estimators.. and Market Factors. 379-406. 1029-1054. 1982. 5(1).. "Estimating Expected Losses and Liquidity Discounts Implicit in Debt Prices. Gallant. "A Nonparametric Approach to Nonlinear Time Series Analysis: Estimation and Simulation. Erasmus University Rotterdam. F." Journal of Economics and Business Statistics. 2002. University of Southern California.) New Directions in Time Series Analysis. R. "Estimating the Price of Default Risk. "A Yield-Factor Model of Interest Rates. Parzen. Gallant. Tax. "Explaining the Rate Spread on Corporate Bonds. 687-720. Monfort." Review of Financial Studies. "Pricing Derivatives on Financial Securities Subject to Credit Risk. G. 12. 50. Kan. "Correlated Default Risk. and A. P. Gruber. Duan. and K. D. R. D. and M. 2000. 1992. Simonato." working paper. Liquidity. 2000. "Factors Affecting the Valuation of Corporate Bonds. 1995." Review of QuantitativeFinance and Accounting 13. C. 53-86." Econometrica. Pedersen. M. Huang. A. Cambridge. A.. J. Haas School of Business.. D. Tauchen. M. Turnbull. Brillinger. D.. Cambridge University Press." Journal of Finance." working paper. G. Yu. "Default Risk and Diversification: Theory and Applications." Review of Financial Studies. Gourieroux. Duffee. M. S. 11-135. 2003. "Estimating Exponential-Affine Term Structure Models by Kalman Filter." working paper. Cornell University. 119-159. J. Huang. "Which Moments to Match?. T. and N. 247-277. Yildirim. H. Mann." working paper. M. Duffie." Journal of Finance. and F. Jarrow. L. Part II.P. Taqqu (eds. G." Mathematical Finance... Agrawal and C. Rosenblatt. D. 57. Houweling.. 50. Jarrow. J. 1999." working paper." in D. Recovery. Jarrow. 6. El Karoui. J. 1995." Journal of Finance. E. 197-226. 12. Mann. Freed. and R. "Time-Series and Cross-section Information in Affine Term Structure Models. G. 2001. J. J. and J. S. and G. Singleton. 71-92 Springer-Verlag.. L. 2002. Elton E. Geske. and S. Santa Clara University. "Estimation of Dynamic Term Structure Models. and R. J. UCLA. R. P. E. G. 194 .. Gruber."working paper. Duffie. 1996. Hansen. 1999. 2002. "A Theoretical Inspection of the Market Price for Default Risk. 2001. Jumps. 2001.. Martellini." working paper. L. "How to Measure Corporate Bond Liquidity?. "Modeling Term Structures of Defaultable Bonds. R.. J. Agrawal and C. and T. Geng. A. New York." Journal of Finance. J. Caines. Statistics and Econometric Models... Delianedis. 2003. and Y.... Stanton. Penn State University. and R. E. 56.. 18. Kapadia. Duffie. 1999." working paper.. Duffee." EconometricTheory. Mentink. Janosi. and G. and M. D." The Journal of Risk. "Modeling Sovereign Yield Spreads: A Case Study of Russian Debt. Geweke.The Review of Financial Studies / v 18 n 1 2005 Das. 2002. A.. 58. 2002. Vorst. New York University. Elton E.

"RatingsPerformance 2000. W.Default Event Risk a of Keswani. London Business School. Matrix.H. of in DefaultFrequencies Structural Modelsof Debt. 1279-1304." Standard& Poor's Risk Solutions. Research. 1982. Models."working Leland." Interest workingpaper. S. White. A."TheMarket Analysis Rate SwapSpreads. B. West." GlobalCreditResearch."Pricing Risks of Default.UCLA. 2002. Schwartz.J. Longstaff. the the Pearson. of Standard& Poor's Special Report.N."Predictions Expected paper. 50. 2002. Yu. R. and T. Mandell."Journal Finance.. Debt."Journal FixedIncome. "Defaultand RecoveryRates of CorporateBond Issuers:2000." Journal of Finance. and E. Moody'sInvestors and Heteroskedasticity Newey. 2001. 1998."Reviewof Derivatives 121-160. Priceof CreditRisk:An Empirical of and Liu. 703-708. F. 1999. and H. "A Simple. "Estimating Risky Term Structure BradyBonds..H.. Unal. 1-25.. 2001. Sun. and K."Maximum 195 .D.12. to 1995.. S.A. Consistent Covariance Autocorrelation 55. 49.Positive Semi-Definite. the 2. K. D. Service. the Returnon Corporate Bonds."A SimpleApproach ValuingRiskyand FloatingRate Longstaff. Madan. "Exploiting ConditionalDensity in Estimating Term An of Structure: Application the CIR model. F.F. D.." Econometrica. Likelihood Estimation Misspecified of 50.. Moody's SpecialComment... 1987. 789-819.2001.UCLA."workingpaper. 1994." Econometrica. "Decomposing Expected of 69-81.

Sign up to vote on this title
UsefulNot useful