Bayesian Estimation of Time-Changed Default Intensity Models

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Bayesian Estimation of Time-Changed Default Intensity Models Michael B. Gordy and Pawel J. Szerszen 2015-002 Please cite this paper as: Gordy, Michael B. and Pawel J. Szerszen (2015). “Bayesian Estimation of Time-Changed DefaultIntensityModels,”FinanceandEconomicDiscussionSeries2015-002. BoardofGovernors of the Federal Reserve System (U.S.). http://dx.doi.org/10.17016/FEDS.2015.002 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Bayesian Estimation of Time-Changed Default Intensity Models∗ Michael B. Gordy and Pawel J. Szerszen Federal Reserve Board, Washington, DC, USA January 6, 2015 Abstract Weestimateareduced-formmodelofcreditriskthatincorporatesstochasticvolatilityin default intensity via stochastic time-change. Our Bayesian MCMC estimation method overcomes nonlinearity in the measurement equation and state-dependent volatility in the state equation. We implement on firm-level time-series of CDS spreads, and find strongin-sampleevidenceofstochasticvolatilityinthismarket. Relativetothewidelyused CIR model for the default intensity, we find that stochastic time-change offers modest benefit in fitting the cross-section of CDS spreads at each point in time, but very large improvements in fitting the time-series, i.e., in bringing agreement between themomentsofthedefaultintensityandthemodel-impliedmoments. Finally,weobtain model-impliedout-of-sampledensityforecasts viaauxiliaryparticlefilter, andfind that the time-changed model strongly outperforms the baseline CIR model. JEL classification: C11, C15, C58, G12, G17 Keywords: Bayesian estimation; MCMC; Particle filter; CDS; Credit derivatives; CIR process; Stochastic time change ∗WehavebenefittedfromdiscussionwithDobrislavDobrev,andfromtheexcellentresearchassistanceof HenryFingerhut,DannyKannell,JimMarrone,DannyMartsandBobakMoallemi. Theopinionsexpressed hereareourown,anddonotreflecttheviewsoftheBoardofGovernorsoritsstaff. Addresscorrespondence to Pawel Szerszen, Federal Reserve Board, Washington, DC 20551; email hpawel.j.szerszen@frb.govi.

1 Introduction Credit spreads widened dramatically during the financial crisis of 2007–09. The CDX index of five year North American investment grade corporate credit default swaps (CDS) rose from under 30 basis points (bp) in February 2007 to 280bp in November 2008. Perhaps less widely appreciated, the financial crisis also witnessed bursts of extreme volatility in spreads. This can be seen in Figure 1, where we plot daily changes in the log of the par spread on the on-the-run CDX index. The figure suggests that stochastic volatility is not a phenomenon peculiar to the financial crisis, as bursts of volatility can be seen, for example, around March 2005 and September 2011. More formally, Gordy and Willemann (2012) find thatahypothesisofconstantvolatilityinaVasicekmodeloflog-spreadsisstronglyrejected even in the pre-crisis data, and Alexander and Kaeck (2008) provide evidence of regimedependent volatility in a Markov switching model of CDS spreads. Arguably, financial institutions might have been better prepared for the financial crisis had risk management and rating models incorporated stochastic volatility. As a case study in model risk, Gordy and Willemann (2012) show that the high investment grade ratings assigned pre-crisis to constant proportion debt obligations depended crucially on the rating agencies’ assumption of constant volatility in spreads. In practical application and in the empirical literature, the most widely-used pricing modelsforCDStaketheso-calledreducedformapproach,pioneeredbyJarrowandTurnbull (1995) and Duffie and Singleton (1999), in which a firm’s default occurs at the first event of a non-explosive counting process with stochastic intensity. Broadly speaking, there are three ways in which to accommodate the patterns of Figure 1 in this class of models. First, if we begin with the assumption that the log of the intensity follows a Vasicek process, as in Pan and Singleton (2008), then we can simply augment the process with positive and negative jumps. The diffusion component would drive the low volatility periods, and the jumps would accommodate the high volatility periods. Observe here that the log transform oftheintensityiswhatallowsustoaccommodatenegativejumpswithoutviolatingthezero lowerboundontheintensityprocess, butitcarriesthepriceofcomputationalintractability. A second approach builds on the widely-used and analytically tractable single-factor CIR specification for the intensity. Jacobs and Li (2008) employ a two-factor specification in which a second CIR process controls the volatility of the intensity process. They estimate the model on time-series of corporate bond spreads and find strong evidence of positive volatility of volatility. The model retains tractability in pricing, but at the expense of the zero lower bound. Except when volatility of volatility is zero, there is no region of the parameterspaceforwhichthedefaultintensityisboundednonnegative. Thethirdapproach is to induce stochastic volatility via stochastic time-change. The constant volatility model is assumed to apply in a latent “business time.” The speed of business time with respect to 1

calendartimeisstochastic, andcapturestheintuitionoftime-variationintherateofarrival of news to the market. When applied to default intensity models, analytic tractability is sacrified to a degree, but Mendoza-Arriaga and Linetsky (2014) and Costin, Gordy, Huang, and Szerszen (forthcoming) derive computationally efficient series solutions to the timechanged model. An advantage to this approach is that the calendar-time default intensity is bounded nonnegative so long as the business-time intensity is bounded nonnegative. Inthispaper,wetakethethirdapproach. AlthoughCDSpricingistractable,estimation of the model presents significant challenges. The model has two latent state variables (one for the default intensity in business time, and one for the time-change process). Due to nonlinearities in the pricing function and state-dependent variance in the state evolution, the model is not well-suited to maximum likelihood estimation by Kalman filter or its extensions. Bayesian Markov chain Monte Carlo (MCMC) estimation is viable, but the conventional MCMC algorithm will be prone to very poor convergence rates in our setting primarilybecausethelatentdefaultintensityhashighserialdependence. Variablesthatare highlycorrelatedarebestsampledinblocks,butthemodelnonlinearityandstate-dependent variance make this infeasible. To overcome these obstacles, we build on the approach introduced by Stroud, Müller, and Polson (2003) in which an auxiliary linearized model facilitates sampling of the latent state variables in blocks. As a by-product of the MCMC estimation, we obtain smoothed estimates of the path of the default intensity and timechange increments. Moreover, missing data is easily addressed in the MCMC algorithm. Weestimatethemodelonfirm-leveltime-seriesofCDSspreadsandfindstrongevidence of material stochastic volatility. Relative to the model without time-change, we find that stochastic time-change offers modest benefit in fitting the cross-section of CDS spreads at each point in time. However, we find that stochastic time-change offers very large improvements in fitting the time-series, i.e., in bringing agreement between the in-sample moments of the default intensity and the model-implied moments. Finally, we assess model performance in out-of-sample density forecasts. We hold parameters fixed at the mean of the in-sample posterior distribution, and filter state variables throughtheout-of-sampleperiod. Althoughnon-adaptedparticlefiltersarestraightforward toimplementinourmodelsetting, theyarepronetoparticledegenerationwheneverabrupt changesofCDSspreadsoccur. Wedevelopapartially-adaptedauxiliaryparticlefilterwhich effectively mitigates particle degeneration while preserving tractability. For nine of the ten firms in our sample, we reject the baseline CIR model in favor of the model with stochastic time-change. Our model specification and CDS pricing are set forth in Section 2. Our econometric methodology is described in Section 3. Data and estimation results are presented in Section 4. Out-of-sample performance is assessed in Section 5. Section 6 concludes. 2

2 Model specification and pricing Mendoza-Arriaga and Linetsky (2014) and Costin et al. (forthcoming, hereafter CGHS) introduce stochastic time change to default intensity models for pricing credit default swaps and other credit sensitive instruments. A firm’s default occurs at the first event of a nonexplosive counting process. Under the business-time clock, the intensity of the counting process is λ . The intuition driving default intensity models is that λ dt is the probability t t of default before business time t+dt, conditional on survival to business time t. We assume complete markets, no arbitrage, and the existence of an equivalent martingale measure Q. Unless stated otherwise, all specifications of stochastic processes and expectations are taken with respect to this measure. We assume that λ is a CIR process t under the business clock, i.e., that it follows the stochastic differential equation p dλ = (µ−κλ )dt+σ λ dW (1) t t t t where W is a Brownian motion. We assume that µ > 0 and initial condition λ ≥ 0, but t 0 do not restrict κ. Our approach can easily be generalized to allow for independent positive Poisson jumps in the default intensity process, but we do not pursue this extension here. Define Λ (s) as the time-integral (or “compensator”) of the default intensity from time t t to time t+s, i.e., Λ (s) = Rt+sλ du. Conditional on the survival to time t and the state t t u λ , the probability of survival to date t+s is t S (s;‘) = Pr(τ > t+s|τ > t,λ = ‘) = E[exp(−Λ (s))|λ = ‘]. (2) t t t t where τ denotes the default time of the firm. This function has an exponential affine form as detailed in Duffie and Singleton (2003, Appendix A.5). We now introduce stochastic time-change. Let T be the stochastic business time assot ciated with calendar time t. For a given firm, let τ˜ denote the calendar default time, so that τ = T is the corresponding time under the business clock. We assume that processes τ˜ λ and T are independent, i.e., that there is no leverage effect. In the empirical literature on stochastic volatility in stock returns, there is strong evidence for dependence between the volatility factor and stock returns (e.g., Andersen et al., 2002; Jones, 2003; Jacquier et al., 2004). In the credit risk literature, the evidence is rather less compelling. Across the firms in their sample, Jacobs and Li (2008) find a median correlation of around 1% between the default intensity diffusion and the volatility factor. We assume that the time-change T is an inverse Gaussian (IG) process with mean t parameter t and shape parameter αt2. This is an almost surely increasing Lévy process such that E[exp(u(T −T ))] = exp(sΨ(u)) where the Laplace exponent is Ψ(u) = t+s t (cid:16) p (cid:17) α 1− 1−2u/α . As E[T −T ] = s for all s and t, we say that the business clock is t+s t 3

unbiased, i.e., business time moves at the same speed on average as calendar time. The parameter α can be interpreted as a precision parameter. As α → ∞, business time converges in probability to calendar time (hence, no time-change). As a Lévy subordinator, the process has increments T −T ,T −T ,T −T ,... that are independent t+s t t+2s t+s t+3s t+2s and identically distributed, so the specification rules out the possibility of volatility clustering in time. The methodology of Mendoza-Arriaga and Linetsky (2014) allows T to be t any Lévy subordinator, and CGHS allow for a still broader class of time-change processes. Nonetheless, we have found the IG specification to be flexible and numerically well-behaved in our econometric application, so impose this specification throughout this paper. Independence of the default intensity and business clock implies that time-changing the defaulttimeisequivalenttotime-changingΛ ;thatis,thatthecompensatorincalendartime t ˜ isΛ (s) = Λ (T −T ). Sinceλ andT areMarkovprocesses,thetriplet(T ,λ(T ),τ˜ > t) t T(t) t+s t t t t t is a sufficient statistic for the information at time t. The conditional calendar-time survival probability function is S˜(s;‘) = Pr(τ˜> t+s|τ˜ > t,T ,λ(T ) = ‘) = Pr(τ > T |τ > T ,T ,λ(T ) = ‘) t t t t+s t t t (cid:20) h i (cid:12) (cid:12) (cid:21) = E E exp(−Λ T(t) (T t+s −T t ))|T t+s ,T t ,λ(T t ) = ‘ (cid:12)T t ,λ(T t ) = ‘ (cid:12) h i = E S (T −T ;‘) . t T(t) t+s t Mendoza-Arriaga and Linetsky (2014) provide a series solution to S˜(s;‘) via Bochner t subordination of the eigenfunction expansion. The expansion is uniformly convergent when the intensity process is stationary (µ > 0 and κ > 0), but cannot be employed in the case of non-stationarity (κ ≤ 0). Both Duffee (1999) and Jacobs and Li (2008) find that the default intensity process is indeed non-stationary under the risk-neutral measure for the typical firm. In our empirical results below, we find that the mean of the posterior distribution for κ under the risk-neutral measure is negative for every firm in our sample. Therefore, we employ the “expansion in derivatives” method of CGHS, which does not impose any restriction on κ. Calendar-time survival probabilities are given by M m S˜(s;‘) ≈ X α−mX c sjDm+jS (s;‘) (3) t m,j s T(t) m=0 j=0 where D is the differential operator d , M is the order of approximation, and s ds ! 1 1 1 2m c = m,j 2mm(j −1)! m−j for m ≥ j ≥ 1. For j = 0, c = 1 and c = 0 for m > 0. Derivatives of the business-time 0,0 m,0 4

survival function are easily computed using a recurrence rule in §4 of CGHS. CGHS show that expansion (3) diverges as M → ∞ but yields an accurate approximation when the series is truncated close to the numerically least term. In our empirical application, we fix M = 2, which CGHS find generally accurate to within the bid-ask spread. Let r denote the riskfree short rate, which we assume to be independent of the default t intensity and time-change. Our pricing and estimation methods each can be generalized to allow for dependence, but at significant cost in complexity and computational resources.1 Duffee (1999) finds that the correlation between riskfree rates and default intensity is typically negative but has second-order importance, so we do not pursue the extension here. The discount function is given by (cid:20) (cid:18) Z t+s (cid:19)(cid:21) P (s) = E exp − r du t t u t CDS pricing follows the treatment of Leeming et al. (2010, §2). By independence of the short rate and default risk, we can write the value of the remaining quarterly premium leg payments at t+s ,...,t+s at date t as 1 n n PremiumLeg = c X P (s )S˜(s ;λ )B(s ) t i t i T(t) i i=1 whereB(s) = min{0.25,s}isthecoupontimeinterval. Thevalueofprotection leg payments is Z s(n) ProtectLeg = (1−ρ) P (s)q˜(s;λ )ds t t T(t) 0 where ρ is the expected recovery rate as a fraction of face value, and q˜(s;‘) is the density t of the default time under calendar time conditional on business-time intensity λ = ‘. T(t) This density is simply q˜(s;‘) = −S˜0(s;‘). The model-implied par CDS spread equates the t t value of the premium leg to that of the protection leg.2 In Section 4, we will compare the smoothed estimates of the default intensity in a model without time-change to the calendar-time intensity λ˜ in the full model with timet change. The default intensity coincides with the instantaneous forward default rate, so λ˜ = q˜(0;λ ) is readily computed.3 t t T(t) As emphasized by Jarrow et al. (2005), market prices can reveal only the Q-intensity, 1Onpricinginamultifactorframeworkwithdependence,seeRemark4.1inMendoza-ArriagaandLinetsky (2014) and §7 in CGHS. 2SinceApril2009,CDShavebeentradedonan“upfront”basiswithannualcouponsfixedto1%or5%. Forobservationssincethischangeinmarketconvention,weconvertupfronttoparspreadusingthestandard formula published by ISDA. Leeming et al. (2010, §4). 3Formaltreatmentoftheexistenceofthecalendar-timedefaultintensitycanbefoundinMendoza-Arriaga and Linetsky (2014, Theorems 3.2(iii), 3.3(iii)). 5

which may or may not contain a risk-premium for the event risk associated with default. While the intensity under the physical measure P is not identified in our setting, our smoothed estimate of λ reveals the P-dynamics of the Q-intensity.4 We specify the dynamt ics of λ under the two measures jointly in the standard fashion. The drift and volatility t parametersin(1)arerestrictedtobethesameacrossthetwomeasures, buttherisk-neutral Q P mean-reversion parameter κ differs from the physical κ by an unrestricted risk-premium. This change of measure is called “drift change in the intensity” by Jarrow et al. (2005), and has been adopted by Duffee (1999) and Jacobs and Li (2008), among others. In the P empirical implementation below, we will restrict κ > 0 to guarantee stationarity under the Q physical measure, but will not restrict κ . In principle, we can introduce a second risk-premium on the uncertainty due to timechange. Say we assume that T − T ∼ IG(s,αs2) under the physical measure P. For t+s t the time-change process T to remain within the IG family under Q, the increments to t the business clock must be distributed T −T ∼ IG(sψ,αs2) under Q where the riskt+s t premium parameter ψ is nonnegative.5 The issue is whether ψ can be estimated. For large α, T − T converges in probability to sψ. Since λ follows a scale-invariant CIR t+s t t process, the parameters of the time-changed default intensity could not be identified in the asymptotic case of α → ∞. In numerical experiments based on simulated time-series of lengthcomparabletoourdataandvaluesofαfromtherangeofestimatesreportedbelowin Section 4, we find that ψ is only very weakly identified. Therefore, we make the simplifying assumptionthatuncertaintyduetotime-changeisunpriced,i.e.,thatT −T ∼ IG(s,αs2) t+s t holds under both P and Q. Stochastic time-change can have a dramatic effect on the kurtosis of changes in the time-series of CDS spreads. In Figure 2, we plot kurtosis of the stationary distribution of changes in the CDS spread as a function of α on a log-log scale. Let CDS(λ;α) be the five-year par spread as a function of the current business-time intensity. For each value of α (i) and horizon δ, we first draw a sample λ , i = 1,...,I, from the stationary distribution for 0 λ . Next, we draw the elapsed business-time T (i) ; and then draw λ(i)(T (i) ) from the CIR t δ δ transitiondensity. Fromthissample, wecalculatethekurtosisofthestationarydistribution of CDS(λ(i)(T (i) );α)−CDS(λ (i) ;α). We plot separate curves for a one day horizon (δ = δ 0 1/250, assuming 250 trading days per year), a one month horizon (δ = 1/12), and an P annual horizon (δ = 1). Parameters for the business-time CIR process are fixed to κ = Q κ = 0.2,µ = 0.004,σ = .1, and recovery is fixed to ρ = 0.4. As we expect, kurtosis at all 0 horizons tends to its asymptotic CIR limit as α → ∞. For fixed α, kurtosis also tends to 4This limitation applies as well to Duffee (1999), Jacobs and Li (2008), and Pan and Singleton (2008). Driessen (2005) brings additional ratings-based information to bear to identify λP and λQ separately. t t 5ThisistheonlyformthatsatisfiesthenecessaryintegrabilityconditionsontheRadon-NikodymderivativeoftheLévymeasuresofjumpsunderPandQsetforthinTheorem7.3ofBarndorff-NielsenandShiryaev (2010). 6

its CIR limit as δ → ∞. This is because an unbiased trend stationary time-change has no effect on the distribution of a stationary process far into the future. For intermediate values of α (say, between 1 and 10), we see that time-change has a modest impact on kurtosis beyond one year, but a material impact at a one month horizon, and a very large impact at a daily horizon. 3 Estimation method Our choice of estimation method is guided by two characteristics of the model of Section 2. First, the mapping from default intensity to CDS spreads is nonlinear and sensitive to parameters. Second, the latent default intensity is persistent under the physical measure andhasstate-dependentvariance. BayesianMarkovchainMonteCarlo(MCMC)estimation is a natural choice in this setting, but the conventional MCMC algorithm will suffer from very poor convergence rates in our setting primarily because the latent default intensity has high serial dependence. Variables that are highly correlated are best sampled in blocks, but the model nonlinearity and state-dependent variance make this infeasible. To overcome this obstacle, we build on the approach introduced by Stroud, Müller, and Polson (2003, hereafter SMP) in which an auxiliary linear model facilitates sampling of the latent state variables in blocks. MCMC estimation has additional advantages. As a by-product of the estimation, we obtain smoothed estimates of default intensity and time-change increments. Moreover, missingdataiseasilyaddressedintheMCMCalgorithmbyaugmentingthestates. Relative toparticlefiltering,alimitationofMCMCisthatonlineestimationofstatesandparameters is computationally infeasible. We turn to particle filtering for out-of-sample forecasting exercises in Section 5. Section 3.1 offers an introduction to Bayesian MCMC methods in general and the SMP algorithm in particular. The presentation is in a simplified setting and may be skipped over by readers familiar with the SMP method. In Section 3.2, we adapt the SMP method to the model of Section 2. 3.1 MCMC estimation with a linear auxiliary model Our introduction to Bayesian MCMC methods is modeled after Chib and Greenberg (1996) and Jones (2003). Let Y = (y ,...,y ) denote the vector of observations, X = (x ,...,x ) 1 T 1 T bethevectoroflatentstatevariablesandΘbethevectorofmodelparameters.6 InBayesian inference the joint posterior of states and parameters is of interest and can be derived given 6Throughout the paper, we employ caligraphic font to distinguish dataset dimensions from model quantities, e.g., T as a count of time-series observations vs. T as the stochastic time-change process. t 7

the prior distribution on the parameters: p(Θ,X|Y) ∝ p(Y|X,Θ) · p(X|Θ) · p(Θ), where p(Y|X,Θ) is the likelihood function of the model, p(X|Θ) is the probability distribution of state variables conditional on the parameters and p(Θ) is the prior probability distribution on the parameters of the model. The joint posterior distribution p(Θ,X|Y) is in general analytically and computationally intractable. MCMC methods overcome this problembybreakingthehigh-dimensionalvectorsX andΘintolow-dimensionalsubvectors with complete conditional posterior distributions that are more easily sampled. The joint posterior is, by construction, the invariant distribution of the chain. The Gibbs sampler of Geman and Geman (1984) is typically the preferred method when the conditional posterior can be sampled directly. We partition X and Θ into, respectively, JX and JΘ subvectors X(1),X(2),...,X(JX) and Θ(1),Θ(2),...,Θ(JΘ). The Markov chain is initialized at chosen values X and Θ . The chain is formed by drawing iteratively 0 0 from transition densities p(X (i) |X (−i) ,Θ ,Y), i = 1,2,...,JX, and p(Θ (j) |Θ (−j) ,X ,Y) n n n−1 n n n j = 1,2,...,JΘ, where X (−i) ≡ (X (k) ) ∪(X (k) ) and Θ (−j) ≡ (Θ (k) ) ∪(Θ (k) ) . n n k<i n−1 k>i n n k<j n−1 k>j Under mild regularity conditions, the chain (X ,Θ ) converges to its invariant distribution n n p(Θ,X|Y). When it is difficult or impossible to draw directly from a complete conditional posterior forasubvectorX(i)orΘ(j),thatstepisreplacedbytheMetropolis-Hastings(MH)algorithm (Metropolis et al., 1953). Say X(i) is the variable to be sampled in iteration n. We sample from a tractable proposal density p˘(X˘;X (i) |X (−i) ,Θ ,Y). The draw X˘ is accepted with n−1 n n−1 probability   min  1, p(X˘|X n (−i) ,Θ n−1 ,Y)·p˘(X n (i − ) 1 ;X˘|X n (−i) ,Θ n−1 ,Y)   p(X n (i − ) 1 |X n (−i) ,Θ n−1 ,Y)·p˘(X˘;X n (i − ) 1 |X n (−i) ,Θ n−1 ,Y) If accepted, we set X (i) = X˘. If rejected, then the value X (i) is retained, i.e., we set n n−1 X (i) = X (i) . A similar procedure is used when Θ(j) is the variable to be sampled. n n−1 Observe that if the true conditional density is taken as the proposal density, the draw is accepted with probability one, and the MH step is simply a Gibbs sampler. In application, it is often the case that the true conditional density is too complicated to approximate with any reliability, in which case a standard approach is to sample as a normal random walk, i.e., draw X˘ as a normal random variable centered at X (i) . While simple to implement, n−1 this approach may suffer from slow convergence. Correlation across parameters or state variables poses a computational challenge in MCMC estimation. Joint sampling of correlated variables yields faster convergence rates, butisoftenmuchmoredifficulttoimplement. Inourapplication, thisdilemmaisespecially 8

acuteforsamplingofthedefaultintensity. Strongpersistenceinthedefaultintensityimplies that we ought to sample jointly the default intensity states over extended blocks of time. Were the model linear and Gaussian, we could sample jointly via the forward filtering backward sampling (FFBS) method of Carter and Kohn (1994) and Frühwirth-Schnatter (1994). Inourmodelsetting,nonlinearityandstate-dependentvariancemakethisinfeasible. The single-move MCMC sampler (Carlin et al., 1992) can be applied, but we have found the cost in rate of convergence to be insurmountable. SMP propose an MCMC algorithm that is well-suited to our problem. Their approach is quite general, but for illustration we adopt a simplified setting with nonlinearity in the measurement equation, constant variance in the state equation, and both y and x unit t dimensional. The essential idea is to extend the state vector to include auxiliary mixing variables ~k = (k ,...,k ) with k having discrete support {1,...,K}. Conditional on the 1 T t mixing variable, the measurement equation is approximated as Gaussian and linear in x , t which allows FFBS to be applied to the proposal distribution. Finally, this FFBS draw is accepted or rejected, as in MH, to account for the gap between the true measurement equation and the linearized approximation that provides the proposal density. There is a trade-offinchoosingthenumberofnodesintheapproximation: ThelargerisK, thesmaller is the gap between true and linearized models, so the higher the acceptance rate. However, the computational cost per MCMC iteration increases with K. SMP specify the auxiliary model for y as Gaussian with pa(y |x ,k = k,Θ) ∼ N(γ + t t t t t,k β x ,ω2 ) for (possibly) time-varying coefficients γ ,β ,ω . The mixture weights are t,k t t,k t,k t,k t,k standardized Gaussian weights φ(x ;ν ,ξ ) pa(k = k|x ,Θ) = t k k t t w(x ) t where φ(x;ν,ξ) denotes the normal density with mean ν and variance ξ2 and the scaling factorw(x) = PK φ(x;ν ,ξ )guaranteesthattheweightssumtoone. Fortractability,the k=1 k k mixture variables are assumed to be serially independent, conditional on x . The auxiliary t mixture model is then K pa(y |x ,Θ) = X pa(y |x ,k = k,Θ)·pa(k = k|x ,Θ) (4) t t t t t t t k=1 Thenodeconstants(ν ,ξ )andidentifyingrestrictionsfor(γ ,β ,ω )arechosensothat k k t,k t,k t,k the auxiliary density pa(y |x ,Θ) is close to the target density p(y |x ,Θ). t t t t In each iteration of the MCMC procedure, we draw Θ in the usual way, and draw the state variables as follows. For notational compactness, we drop the subscript n for indexing MCMC iterations. 9

1. Generatemixtureindicatorsk fromthecompleteconditionalposteriorpa(k |y ,x ,Θ), t t t t which has a multinomial distribution. By Bayes’ rule, pa(k |y ,x ,Θ) ∝ pa(y |x ,k ,Θ)·pa(k |x ,Θ) t t t t t t t t This is a Gibbs step. 2. Conditional on the mixture indicators, draw states jointly via FFBS from the MH proposal density for the mixture model. The proposal density is given by: T p˘(X|~k,Y,Θ) = p(x ) Y p(x |x ,Θ)·pa(y |x ,k ,Θ)·φ(x ;ν ,ξ ) 0 t t−1 t t t t k(t) k(t) t=1 This is similar to the MH proposal density in a linear model without auxiliary variables, but incorporates an extra Gaussian kernel φ(x ;ν ,ξ ) to capture the adt k(t) k(t) ditional conditioning information from the auxiliary variables. Denote the proposal draw X˘ = (x˘ ,...,x˘ ). 1 T 3. Accept or reject X˘. The acceptance probability is ( Y T p(y t |x˘ t ,Θ) w(x t )pa(y t |x t ,Θ) ) min 1, (5) w(x˘ )pa(y |x˘ ,Θ) p(y |x ,Θ) t=1 t t t t t where the {x } denote the previous draw of X in the Markov chain. t Observe that the acceptance probability is maximized when the auxiliary model likelihood pa(y |x ,Θ) is close to the true likelihood p(y |x ,Θ). When the acceptance probability is t t t t low, performance may be improved by dividing the time-series into blocks, b = 1,...,B, and tuning the node constants (νb,ξb) on a block-by-block basis. k k 3.2 Application to estimation of the time-changed default intensity model In this section, we cast the model of Section 2 as a discrete-time state-space model, and then adapt the SMP algorithm to accommodate a multidimensional measurement equation and state-dependent variance in the state equation for the default intensity. Ournotationneedstotrackthreetime-scales,oneforobservations,oneforcalendartime, and one for business time. Henceforth, we let t = 1,...,T index a set of daily observations. The associated (deterministic) calendar time is denoted T˜. For simplicity, we take trading t days as equally spaced in calendar time at an interval of ∆ = T˜ −T˜ = 1/250 of a year. t+1 t The stochastic business time associated with observation time t is T(T˜). For notational t convenience, let h = λ(T(T˜)) be the business-time default intensity at the business time t t 10

associated with observation date t.7 Fixing a given reference entity, let y be the log of the CDS spread observed on date t,m Q t for maturity m ∈ M = {1,2,3,5,7,10} years. Let F (h ;µ,κ ,σ,α,ρ ,P ) map h to a m t t t t model-implied log-spread on a CDS of maturity m. For notational compactness, we drop explicitreferencetothedate-triskfreediscountfunctionandrecoveryrate,andsimplywrite Q F (h ;µ,κ ,σ,α). t,m t We assume that CDS log-spreads are observed with noise. Without measurement noise, the spread would be a deterministic function of underlying parameters and the state, which would lead to stochastic singularity in filtering and smoothing.8 In the option pricing literature, this approach has been used by Eraker (2004) and Johannes et al. (2009), among others. Our measurement equation is Q (y) y = F (h ;µ,κ ,σ,α)+ζ ε (6) t,m t,m t t,m (y) Measurement errors ε are assumed to be i.i.d. standard normal random variables. The t,m new parameter ζ scales the pricing errors. For parsimony, we impose homoscedasticity across both time and contract maturity. The model has two state evolution equations. The first governs the distribution of increments to the business time clock. Let χ = T(T˜ )−T(T˜) i ∼ id IG(∆,α∆2) (7) t+1 t+1 t be the inverse Gaussian increment to the business clock between observations t and t+1. The second governs the evolution of the default intensity under the physical measure. We apply a first order Euler discretization scheme to SDE (1) with stochastic time-increments to obtain q P (λ) h = h +(µ−κ h )χ +σ h χ ε (8) t+1 t t t+1 t t+1 t+1 (λ) (y) The {ε } are i.i.d. standard normal random variables, independent of {ε }. We assume t t,m that both κ P and µ are strictly positive, so that h is bounded nonnegative and stationary.9 t We refer to equations (6)–(8) as the state-space representation. Our framework extends the basic SMP algorithm of Section 3.1 along two dimensions. First, our measurement equation is of dimension #M, as we have one observation for each contract in maturity set M. As in the standard SMP algorithm, we introduce a 7This should not be confused with the calendar time default intensity λ˜(T˜). t 8Inoursetting,measurementerrorcouldalternativelybejustifiedbymisspecificationorestimationerror inthefittedriskfreeterm-structureP orbythepresenceoftime-varyingliquiditypremiaintheCDSmarket. t 9Discretization introduces the possibility of negative realizations of the default intensity. We avoid this in the MCMC simulation by simple truncation at zero. 11

(y) linear auxiliary model for the measurement equation with mixing variable k . Second, t and of greater consequence, the state equation (8) for h has state-dependent volatility, t as V[h |h ] = σ2h χ . As FFBS requires state-independent volatility, we introduce t+1 t t t+1 (λ) a linear auxiliary model for the state equation (8) with a new mixing variable k . For t parsimony, the two mixing variables share the same support {1,...,K}, as well as mean and volatility constants (ν ,ξ ).10 k k (λ) (y) We assume that log-spread y is conditionally independent of k given k . The t,m t t auxiliary model for the measurement equation is conditionally Gaussian with pa(y |h ,k (y) ,Θ) = Y pa(y |h ,k (y) ,Θ) t t t t,m t t m∈M Hence, the auxiliary model is given by the following mixture of Gaussian densities: K pa(y |h ,Θ) = X pa(k (y) = k|h ,Θ) Y pa(y |h ,k (y) = k,Θ) t t t t t,m t t k=1 m∈M where pa(k (y) = k|h ,Θ) = φ(h t ;ν k ,ξ k ) for w(x) = X K φ(x;ν ,ξ ) t t w(h ) k k t k=1 AsinthestandardSMPalgorithm, weassumethattheauxiliarymodelforthemeasurement equation is linear and Gaussian conditional on the mixture indicator. We linearize the measurement equation (6) locally at node k, and define pa(y |h ,k (y) = k,Θ) ∼ N(F (ν ;Θ)+F0 (ν ;Θ)·(h −ν ),ζ2) t,m t t t,m k t,m k t k Observe that the coefficients in the mean of the auxiliary distribution vary over time and across maturity, as well as across nodes. We now introduce the auxiliary model for the state evolution equation (8). Conditional (λ) (y) on k and h , h is independent of the measurement equation mixing variable k . The t t−1 t t auxiliary model is conditionally linear and Gaussian. We linearize state equation (8) at node k, and define pa(h |h ,k (λ) = k,χ ,Θ) ∼ N(µχ +(1−κ P χ )h ,σ2χ ν ). t t−1 t t t t t−1 t k As the drift of the true evolution equation is affine in h , the mean of the auxiliary model t−1 can be matched to the true mean without conditioning on the node. Only the variance of 10In our implementation, we achieve higher acceptance rates if the time-series is divided into contiguous blocks of, say, 40–50 observations each. The node constants are tuned on a block-by-block basis. 12

the auxiliary model depends on k(λ). The auxiliary model for state evolution equation is decomposed in the usual way: K pa(h |h ,χ ,Θ) = X pa(h |h ,k (λ) = k,χ ,Θ)·pa(k (λ) = k|h ,Θ) (9) t t−1 t t t−1 t t t t−1 k=1 (λ) (y) The state k is assumed to be conditionally independent of k with t t φ(h ;ν ,ξ ) pa(k (λ) = k|h ,Θ) = t−1 k k (10) t t−1 w(h ) t−1 We introduce additional notation so that the algorithm can be presented compactly. The data vector is Y = (y ,...,y ) where y = (y ) is the vector of CDS log-spreads 1 T t t,m m∈M P Q observedattimetforallmaturitiesm ∈ M. TheparametervectorisΘ = (µ,κ ,κ ,σ,α,ζ). ThestatevectorX containsfourcomponents: defaultintensitiesX(λ) = (h ,...,h ); time- 1 T change increments X(χ) = (χ ,...,χ ); mixture variables for the observation equation 1 T X(y) = ~k(y) = (k (y) ,...,k (y) ) with k (y) ∈ {1,...,K}; and mixture variables for the state 1 T t evolution equation X(x) = ~k(λ) = (k (λ) ,...,k (λ) ) with k (λ) ∈ {1,...,K}. For any subset 1 T t V ⊂ X, let X(−V) denote X\V. In each iteration of the MCMC procedure, we follow these steps: Draw the mixture indicator variables. By Bayes’ law and the assumption of conditional independence, pa(k (y) |X(−~k(y)),Y,Θ) ∝ pa(y |h ,k (y) ,Θ)·pa(k (y) |h ,Θ) t t t t t t pa(k (λ) |X(−~k(λ)),Y,Θ) ∝ pa(h |h ,k (λ) ,χ ,Θ)·pa(k (λ) |h ,Θ) t t t−1 t t t t−1 The complete conditional posterior distribution for each mixture indicator is multinomial and is easily sampled. Generate proposal default intensity states. WeapplyBayes’lawtotheauxiliarymodel for the state equation: pa(X(λ)|X(−λ),Y,Θ) ∝ pa(Y|X(λ),~k(y),~k(λ),Θ)·pa(X(λ)|~k(y),~k(λ),X(χ),Θ) ∝ pa(Y|X(λ),~k(y),Θ)·pa(~k(y)|X(λ),Θ)·pa(X(λ)|~k(λ),X(χ),Θ) 13

where T pa(Y|X(λ),~k(y),Θ) = Y Y pa(y |h ,k (y) ,Θ) t,m t t t=1m∈M pa(~k(y)|X(λ),Θ) = Y T φ(h t ;ν kt ,ξ kt ) w(h ) t=1 t T pa(X(λ)|~k(λ),X(χ),Θ) = pa(h |k (λ) ,Θ) Y pa(h |h ,k (λ) ,χ ,Θ) 0 1 t t−1 t t t=2 The proposal distribution is T p˘(X(λ)|X(−λ),Y,Θ) ∝ pa(X(λ)|X(−λ),Y,Θ)· Y w(h ) t t=1 The standard FFBS algorithm can be used to sample jointly default intensity states from the proposal distribution p˘. Accept or reject proposal draw. By repeated application of Bayes’ law: p(X(λ)|X(−λ),Y,Θ) p(Y|X(λ),Θ)·p(X(λ)|X(χ),Θ)·pa(~k(y)|X(λ),Θ)·pa(~k(λ)|X(λ),Θ) ∝ p˘(X(λ)|X(−λ),Y,Θ) pa(Y|X(λ),~k(y),Θ)·pa(X(λ)|~k(λ),Θ)·pa(~k(y)|X(λ),Θ)·QT w(h ) t=1 t p(Y|X(λ),Θ)·p(X(λ)|X(χ),Θ) ∝ pa(Y|X(λ),Θ)·pa(X(λ)|X(χ),Θ)·QT w(h ) t=1 t Hence, the MH acceptance probability is given by: ( p(Y|X˘(λ),Θ)·p(X˘(λ)|X(χ),Θ)·pa(Y|X˘(λ),Θ)·pa(X˘(λ)|X(χ),Θ)·QT w(X˘(λ) ) ) min 1, t=1 t p(Y|X(λ),Θ)·p(X(λ)|X(χ),Θ)·pa(Y|X(λ),Θ)·pa(X(λ)|X(χ),Θ)·QT w(h ) t=1 t (50) whereX˘(λ) = (h˘ ,...,h˘ )istheproposaldrawandX(λ) = (h ,...,h )isthecurrent 1 T 1 T value. Relative to the standard SMP acceptance probability in (5), expression (50) introduces the ratio of kernels pa(X˘(λ)|X(χ),Θ) and p(X˘(λ)|X(χ),Θ) to account for the second auxiliary model for the state equation. Calculation of the acceptance probabilities is straightforward, as the kernels for true model p involve only Gaussian densities, and the auxiliary kernels pa are mixtures of Gaussian densities. Draw the parameters. The conditional posteriors for the parameters κ P and ζ2 are conjugate under the choice of normal prior and inverse gamma prior, respectively. Thus, the Gibbs sampler can be used for these parameters. The remaining parameters Q (µ,κ ,σ,α) have complete conditional posteriors that depend intractably both on the 14

time-series of default intensity states (h ) and the observed data Y. We sample t t=1,...,T these parameters via the MH normal random walk algorithm. For all parameters we P assume flat uninformative priors. We restrict κ and α to be strictly positive. Draw the time-change increments. By Bayes’ law and the conditional independence of the increments, p(χ |X(−χ),Y,Θ) ∝ p(h |h ,χ ,Θ)·p(χ |Θ) t t t−1 t t It is straightforward to show that these kernels give a generalized inverse Gaussian (GIG) density from which we can sample directly in a Gibbs sampler step. 4 Empirical results We estimate the CIR and time-changed CIR models on daily data for CDS spreads on ten single-name reference entities for the period from the beginning of 2002 to the end of 2009. The sample period spans the financial crisis of 2007–09. We retain the period from the beginning of 2010 to the end of 2011 for out-of-sample forecasting exercises in Section 5. The sample of reference entities includes U.S. and European nonfinancial corporations of varying credit quality: Alcoa, Anadarko, CenturyLink, Clear Channel, Ford, Lennar, Limited Brands, RadioShack, Sprint Nextel, and Tyson Foods. All have liquid trading in CDS. For each name, spreads on CDS of maturities 1, 2, 3, 5, 7 and 10 years are taken from the Markit database.11 In the upper panels of Figures 3 and 4 we plot time-series of 1, 5 and 10 year CDS spreads for Alcoa and Ford, which are among the most heavily-traded names in our sample. Even for these liquid names, it is not unusual to find missing observations in CDS data. In Table 1 we report the number of missing observations in our eight year sample by reference entity and maturity. As a practical matter, it would be infeasible to estimate the model without a convenient and rigorous means to accommodate missing observations. In MCMC estimation, we can simply treat missing observations as latent variables and sample from their complete conditional posteriors. This approach avoids additional assumptions on the missing data points and does not interfere with the law of motion imposed by the observed data points.12 Model-impliedCDSspreadsdependontheriskfreediscountfunctionP andonrecovery t rates. We parameterize P using the Svensson yield curve, for which we rely on daily t 11The time-series for Sprint Nextel Corp. begins on August 24, 2005. Data for the other nine firms span the entire sample period. 12See Tsay (2010, §12.6) on the handling of missing data via data augmentation. 15

parameter estimates provided by the Federal Reserve Board.13 Recovery rates are taken from the Markit database. Weestimatemodelspecificationsbothwithandwithouttime-changeindefaultintensity using the algorithm of Section 3.2. Recall that the model without time-change is nested in the time-change model as the limiting case α = ∞. After discarding the burn-in period, we simulate chains with 100,000 draws from the invariant distribution for each model. The parameter estimates and the smoothed state estimates are calculated as means of draws from the joint posterior distribution of parameters and states. The parameter estimates of CDS pricing models are reported in Tables 2.a–2.j. The tables report for each firm the mean, standard deviation, 5% quantile and 95% quantile of the posterior distribution for the state-space model in equations (6)–(8). All parameters are estimated with high precision except for the speed of mean reversion under the physical measure. This exception is not unexpected. In maximum likelihood estimation of a CIR specification for interest rates, Chapman and Pearson (2000) and Phillips and Yu (2009) P show that κ has a large standard error in reasonable sample sizes. The speed of mean Q reversion under the risk-neutral measure (κ ) is for all firms negative even at the 95% quantiles. This finding is qualitatively consistent with Duffee (1999), Pan and Singleton (2008) and Jacobs and Li (2008), and underscores the necessity of a pricing methodology that remains valid under non-stationarity. The estimates of the parameter α provide strong evidence for stochastic time-change in default intensities. For two of the firms (Lennar and Sprint), the mean of the posterior distribution of α is near 20. At such values, stochastic time-change has essentially no effect on the term-structure of credit spreads. Nonetheless, as illustrated in Figure 2, the kurtosis of daily changes in CDS spreads is nonetheless an order of magnitude larger than the kurtosis in the CIR model without time-change. For RadioShack, the mean and even the 95%quantileoftheposteriorforαisbelow2.0. Atsuchvalues,time-changehasanoticeable effect on credit spreads. For the remaining seven firms in our sample, the estimated α lie between these extremes. Alcoa and Ford are typical in this regard, with estimated α between 7 and 8. In the middle panels of Figures 3 and 4, we plot smoothed estimates of the calendar-time default intensity (λ˜ ) for these two firms for the models with and without time-change. The t smoothed time-series track quite closely across the two models, except when the estimated CIR default intensity dips below 1bp. Intuitively, the close alignment is a consequence of the small effect of time-change on the model-implied term-structure of spreads and the sufficiency of the observed term-structure to pin down the latent default intensity. In the lower panels, we plot smoothed estimates of the time-change increments (χ ) t 13Daily Svensson curves are publicly available for download at the Federal Reserve Board website. For details on data sources and methodology, see Gürkaynak et al. (2007). 16

for the time-changed model. We scale by ∆ so that the increments have unit mean. Not surprisingly, we find very large outliers during the financial crisis of 2007–09 with business time ticking 15 times faster than calendar time for the most active days for Alcoa and up to about 25 times faster for Ford. Counter to our model assumption of Lévy time change, the smoothed time-series suggest serial correlation in time-change increments. We leave this as an avenue for future research. The moments of the estimated residuals in the measurement equations and state equation for the default intensity provide metrics of in-sample fit for the CDS pricing models with and without time-change. Table 3 reports the posterior mean, standard deviation, skewness and kurtosis of the smoothed distribution of the stacked measurement equation (y) residuals, ε . Acorrectspecificationimpliesthattheseresidualsarestandardnormalwith t,m mean zero, standard deviation one, skewness zero and kurtosis three. Comparison of the moments suggests that stochastic time-change offers modest benefit in fitting CDS spreads. The most notable improvement is a reduction in skewness for nine of the ten reference entities. It is in the time-series dimension where we see a clear distinction between the models with and without time-change. Table 4 reports the moments of the smoothed distribution (λ) of the innovations (ε ) in the default intensity state equation. Across all firms in our t sample, introducing stochastic time-change brings the mean, standard deviation, skewness andkurtosisofthestateequationinnovationsclosertothemomentsofthestandardnormal distribution. InthecaseofAlcoa,forexample,themeanfallsfrom0.05fortheCIRmodelto 0.01 for the time-changed model, variance increases from 0.67 to 0.82, skewness decreases from 0.34 to 0.03, and kurtosis falls from 4.27 to 3.22. Indeed, introducing time-change essentially eliminates skewness and reduces kurtosis to below 3.25 for every firm in the sample. We use the estimated parameters for Alcoa to illustrate how stochastic time-change alters the implications of the baseline CIR model. In Figure 5, we plot the term structure of CDS spreads on three dates, one during the credit boom (18 March 2005), one in the early phase of the crisis (19 June 2008), and one in the late phase of the crisis (19 March 2009).14 The five-year CDS spreads on these dates for Alcoa were 19bp, 93bp and 851bp, respectively. For each model and date, model parameters are fixed to the mean of the posterior distribution for the model, as reported in Table 2.a. The business time default intensity is taken from the estimated smoothed distribution of h for the date and model. t Despite the two-orders-of-magnitude difference across dates in the initial condition (h ), the t two models produce very similar term structures, which suggests that the two models will 14Dates are chosen to be representative of the range of observed spreads over the sample period. Each of these dates immediately precedes a CDS settlement date, so the nominal maturity of a CDS on these dates equals the exact actual maturity. 17

yield similar fit to the cross-section of observed spreads on each sample date. The fitted models differ, however, in the forecast distribution for spreads at a future horizon. We again fix parameters for each model to the mean of the posterior distribution for Alcoa. For simplicity, we fix a riskless rate of 3% and recovery of 40%. For the CIR model, we simulate the quantiles of the forecast distribution of CDS(λ ) for δ = 1/250 (i.e., δ a horizon of one trading day) given initial condition λ . Similarly, for the time-changed 0 model we simulate the quantiles of the forecast distribution of CDS(λ(T );α). We do this δ for two values of λ , one below the long-run mean (5 bp) and one above the long-run mean 0 (50 bp), and report results in Table 5. Consistent with the findings on kurtosis of spread changes (Figure 2), we find that the time-change model assigns higher probability to very large changes in the spread over the short horizon. We return briefly to consider whether empirical results are consistent with two orthogonality assumptions imposed in Section 2: (λ) • There is no leverage effect, i.e., the innovation ε in the default intensity state t equation is uncorrelated with time-change increments χ . t (λ) • Thedefaultintensityprocessisindependentoftheriskfreeshortrateprocess, i.e., ε t is uncorrelated with the change in the federal funds rate from t−1 to t. Totesttheseassumptions,wereportinTable6theposteriormeansandstandarddeviations (λ) of the respective time-series correlations. We find the correlation between ε and changes t in the short rate is economically small and statistically insignificant for all ten names in our (λ) sample. The correlation between ε and χ is positive for all ten names and in some cases t t statistically significant, but generally of second-order economic magnitude. For all firms except Sprint, the posterior mean correlation is under 3%. 5 Out-of-sample performance In this section, we assess out-of-sample model performance. For each observation date in the out-of-sample period from January 2010 through December 2011, we calculate the day-ahead forecast density implied by the models with and without time-change in default intensity. For each model and each date, the forecast density is evaluated at the realized next-day observation. The test statistic, due to Diebold and Mariano (1995) and Amisano and Giacomini (2007), allows for a two-sided test of the null hypothesis that the two models perform equally well. Construction of forecast densities presents new challenges. The MCMC algorithm of Section 3.2 delivers a smoothed estimator of the default intensity, whereas for this application online filtering is required. As in our MCMC estimator, the filtering algorithm must accommodate nonlinearity in the measurement equation and non-Gaussian state evolution. 18

Particle filtering is well-suited to such challenges, but the most tractable filters, such as the SIR filter of Gordon et al. (1993), are ill-suited to our setting. Since the latent timechange process has serially independent increments, propagation of state variables from t to t+1 using only time-t information leads to particle degeneration, because the filter cannot “anticipate” abrupt changes in CDS spreads.15 Contrasted to such non-adapted filters are fully-adapted filters, in which propagation is conditioned on the observation at t+1. Here the challenge is tractability. To navigate between these obstacles, we develop a partiallyadapted particle filter using an auxiliary linearized model in the spirit of our approach in Section 3.2. To reduce the dimensionality of the problem, we fix all model parameters to the in-sample posterior means estimated in Section 4. 5.1 Partially-adapted particle filter on a linearized model The particle filter is a recursive algorithm that constructs a discrete approximation to the filtered distribution of the state variables at each observation date: x (i) i ∼ id p(x |Yt) for all t = T +1,...,T∗ t t where Yt = (y ,...,y ) are observations available up to time t, x is the state vector, I is 1 t t the number of particles i = 1,...,I, and T∗−T is the number of trading days in the outof-sample period. To reduce notation, explicit dependence on model parameters is omitted throughout this section. In the auxiliary particle filter (APF) of Pitt and Shephard (1999), the sample of particles at observation time t is constructed as the marginal distribution for x in the joint t distribution for (x ,x ) given Yt, for which the kernel can be decomposed as t t−1 p(x ,x |Yt) ∝ p(x |x ,y )·p(y |x )·p(x |Yt−1). (11) t t−1 t t−1 t t t−1 t−1 Fully-adapted sampling from this distribution is conducted in two steps. First, the particles {x (i) } that serve as a discrete approximation to p(x |Yt−1) are resampled with t−1 i=1,...,I t−1 (i) weights proportional to p(y |x ). This step requires pointwise evaluation of the required t t−1 (i) density and induces consistency of the “old” particles x with the “new” observation t−1 y . Second, the resampled time-t−1 particles are propagated to time t by sampling from t p(x |x ,y ). As is often the case in application, in our model setting we cannot easily t t−1 t evaluate p(y |x ) for the resampling step, nor can we sample directly from p(x |x ,y ). t t−1 t t−1 t 15Particledegenerationariseswhenonlyasmallnumberofparticleshavesignificantweightintheresampling stage. See Lopes and Tsay (2011) for a review of particle filter methods with discussion of particle degeneration. 19

We substitute x = (h ,χ ) in equation (11) and re-write the first two kernels as t t t p(h ,χ |h ,χ ,y )·p(y |h ,χ ) t t t−1 t−1 t t t−1 t−1 = p(h |χ ,h ,χ ,y )·p(y |χ ,h ,χ )·p(χ |h ,χ ,Yt−1). t t t−1 t−1 t t t t−1 t−1 t t−1 t−1 Inourmodel,theχ areiid,sop(χ |h ,χ ,Yt−1) = p(χ ). Conditionalonh ,theCDS t t t−1 t−1 t t−1 spread at t does not depend on χ , so p(y |χ ,h ,χ ) = p(y |χ ,h ). Similarly, cont−1 t t t−1 t−1 t t t−1 ditional on h , h does not depend on χ , so p(h |χ ,h ,χ ,y ) = p(h |χ ,h ,y ). t−1 t t−1 t t t−1 t−1 t t t t−1 t Combining these terms, we re-write equation (11) as p(h ,χ ,h ,χ |Yt) ∝ p(h |χ ,h ,y )·p(y |χ ,h )p(χ )·p(h ,χ |Yt−1) (12) t t t−1 t−1 t t t−1 t t t t−1 t t−1 t−1 As the propagation kernel p(h |χ ,h ,y ) is intractable, we introduce an auxiliary t t t−1 t linear model.16 Begin by writing p(h |χ ,h ,y )·p(y |χ ,h ) = p(y |h ,χ ,h )·p(h |χ ,h ) (13) t t t−1 t t t t−1 t t t t−1 t t t−1 In our setting, this is simplified by noting that Y p(y |h ,χ ,h ) = p(y |h ) = p(y |h ). t t t t−1 t t t,m t m∈M In the auxiliary model, the measurement equation is linearized as y = γ +β h +ζ ε (y) (60) t,m t,m t,m t t,m so that pa(y |h ) ∼ N(γ +β h ,ζ2). The choice of constants γ and β determines t,m t t,m t,m t t,m t,m how well the auxiliary linear model approximates the true nonlinear model. The auxiliary modelstateequationsareassumedtobethesameasthoseforthetruemodelin(7)and(8). In the appendix, we derive a proposal propagation density p˘(h |χ ,h ,y ) and auxiliary t t t−1 t evaluation weights pa(y |χ ,h ) such that equation (13) can be replaced by t t t−1 p˘(h |χ ,h ,y )·pa(y |χ ,h ) = pa(y |h )·p(h |χ ,h ) (130) t t t−1 t t t t−1 t t t t t−1 The proposal propagation density is Gaussian (so sampling is trivial), and the auxiliary evaluation weights are similarly convenient to compute. We arrive at the following algorithm for sampling {h ,χ } conditional on particles t t {h (i) } and on Yt: t−1 i=1,...,I 16Thepossibilityofconstructingaproposaldistributiononalinearizationofthetruemodelwassuggested by Pitt and Shephard (1999, §3.3). 20

(i) 1. Draw particles {χ } blindly from the unconditional IG distribution for χ . t i=1,...,I t 2. Draw particles {h˙(i) ,χ˙ (i) } by resampling {h (i) ,χ (i) } with weights prot−1 t i=1,...,I t−1 t i=1,...,I portional to pa(y |χ (i) ,h (i) ). t t t−1 3. Propagate {h˘(i) } by sampling from p˘(h |χ˙ (i) ,h˙(i) ,y ). t i=1,...,I t t t−1 t 4. To restore consistency with the law of the true model, generate{h (i) ,h¨(i) ,χ (i) } t t−1 t i=1,...,I by resampling from {h˘(i) ,h˙(i) ,χ˙ (i) } with weights proportional to17 t t−1 t i=1,...,I p(h˘(i) |χ˙ (i) ,h˙(i) ,y )·p(y |χ˙ (i) ,h˙(i) ) p(y |h˘(i) ) ω (i) = t t t−1 t t t t−1 = t t (14) t p˘(h˘(i) |χ˙ (i) ,h˙(i) ,y )·pa(y |χ˙ (i) ,h˙(i) ) p˘(y |h˘(i) ) t t t−1 t t t t−1 t t 5. Discard the sample of χ , and regenerate by sampling χ (i) from p(χ |h (i) ,h¨(i) ). As t t t t t−1 noted in Section 3.2, the conditional distributed is GIG. This algorithm is fully-adapted for h with respect to the auxiliary model. Even though t the auxiliary model diverges from the true model, with reasonable choices of γ and β t,m t,m we can still expect that the realization of y will inform the propagation of h to h . We t t−1 t impose β t,m = F t 0 ,m (hbt ) γ t,m = F t,m (hbt )−β t,m hbt where hbt is given by the inverse of model-implied 5-year CDS spread observed at time t: hbt = F t − ,5 1(y t,5 ). Sampling of χ in Step 1 is non-adapted, but the regenerated sample in Step 5 is fully t adapted, i.e., it utilizes all the information content of y under the true model. The regent eration step is permissible only because χ does not enter the observation equation (i.e., t because y is independent of χ when conditioned on h ). It is straightforward to show t t t that the regenerated {χ } has the same conditional distribution as the sample after t i=1,...,I Step 4, but regeneration improves the efficiency of sampling. We include the final step to facilitate examination of the time-series of filtered states in Section 5.2. Finally, it is trivial to modify the algorithm for the baseline model without time-change, because the pricing model is a special case of the model with time-change. We can simply (i) fix χ = ∆ in Step 1 and drop Step 5. t 17We denote the resampled h as {h¨ } in order to avoid confusion with the original cloud of t−1 t−1 i=1,...,I particles {h } , which are conditioned only on Yt−1. t−1 i=1,...,I 21

5.2 Density forecast analysis In this section we present the out-of-sample model selection analysis based on the Diebold and Mariano (1995) and Amisano and Giacomini (2007) test statistic. We begin with a brief description of the test. The average predictive likelihood for the out-of-sample points t = T +1,...,T∗ is given by: L = 1 X T∗ log (cid:16) p(y |Yt−1,Θ ) (cid:17) M T∗−T t M t=T+1 where M ∈ {TC,CIR} denotes a model specification with “CIR” for the baseline model without time-change and “TC” for the model with time-change, p denotes the predictive likelihoodevaluatedatthefuturerealizedCDSlog-spreadattimet,givenmodelparameters Θ , which are fixed at the in-sample MCMC estimates. The predictive likelihood can be M computed from the filtering distribution as follows: Z p(y |Yt−1,Θ ) = p(y |x ,Θ ) p(x |Yt−1,Θ ) dx t M t t M t M t Z p(x |Yt−1,Θ ) = p(x |x ,Θ ) p(x |Yt−1,Θ ) dx t M t t−1 M t−1 M t−1 Theparticlefilterprovidesasample{x (i) } fromthefilteringdistributionp(x |Yt−1,Θ ). t−1 i=1,...,I t−1 M We propagate this sample forward via the state evolution equations by drawing from the (i) (i) transition density p(x |x ,Θ ) to obtain a sample {xˆ } from the distribution t t−1 M t i=1,...,I p(x |Yt−1,Θ ). The predictive likelihood for observation t is approximated as the Monte t M Carlo integral 1 X I (i) ! LbM,t = log I p(y t |xˆ t ,Θ M ) . i=1 Finally, we average across time to obtain 1 X T∗ LbM = T∗−T LbM,t . t=T+1 Intuitively, a model that performs better out-of-sample has a higher value of average predictive likelihood. Under the null hypothesis of equal performance the Amisano and Giacomini (2007) statistic U = √ T∗−T (LbTC −LbCIR ) (15) r (cid:16) (cid:17)2 (cid:16) (cid:17)2 T∗− 1 T−1 PT t= ∗ T+1 LbTC,t −LbCIR,t − LbTC −LbCIR 22

has an asymptotic standard normal distribution. The partially-adapted APF of Section 5.1 is implemented for each model and each firm with I = 100,000 particles. We confirm that our particle filter avoids significant particle degeneration. In the case of the time-changed model for Ford, for example, the range of the efficient sample size (as a fraction of the number of particles) for the partially-adapted APF is (68.3%,99.8%) with the 1%-quantile of 80.6%. We have implemented the standard SIR filter as well, and in this case find a range of (1.7%,88.4%) along with the 1%-quantile of 15.6%. The test statistic U is estimated for each of the ten names in our sample and reported in Table 7. For nine of the names (all firms but Lennar), the coefficient is positive and significant at the 1% level. In Figures 6 and 7, we contrast the typical case of Ford with the exceptional case of Lennar. We plot CDS spreads (top graph), the filtered estimates of default intensity (middle graph), and the filtered estimates of time-change increments (bottom graph) for the out-of-sample period covering 2010 and 2011.18 The filtered timeseries of χ for Ford exhibits significant volatility with multiple spikes, whereas for Lennar t this time-series is roughly constant with only two minor instances of higher activity. In each case, the spikes in χ coincide with relatively large movements in the one-year CDS t spreadandthefiltereddefaultintensity. Thus,addingstochastictime-changetothebaseline CIR model improves the forecast performance when the CDS spread exhibits volatility in volatility, but may lead to overfit when volatility is relatively constant. 6 Conclusion Stochastic volatility has been studied extensively in markets for equities, interest rates and commodities, but there has been relatively little empirical work to date on stochastic volatilityincreditmarkets. Weappeartobethefirsttoestimateamodelinwhichstochastic volatility in default intensity is induced by stochastic time change. Nonlinearity in the CDS pricing function and state-dependent variance in the evolution of the default intensity present significant challenges for estimation. We overcome these difficulties using a new Bayesian MCMC estimation method that builds on earlier work by Stroud, Müller, and Polson (2003). For out-of-sample density forecasts, we also develop an auxiliary particle filter. Although computationally intensive, both methodologies performed well on our data, and may have application to related problems in modeling interest rate or volatility swaps. We estimate our model on firm-level time-series of single-name credit default swap spreads. In-sample, stochastic time-change is found to be a statistically and economically significant feature of the data for all firms in our sample. In out-of-sample density forecasts 18To be precise, the plotted value at date t for the filtered estimate is the mean of the sample of {x(i)} from the partially-adapted APF. t i=1,...,I 23

tests, we find that the time-changed model outperforms the baseline CIR model at high levels of significance for nine of our ten firms. Smoothed estimates of the increments of stochastic time show large outliers during the financial crisis of 2007–09 with business time ticking substantially faster than calendar time for the most volatile days in our sample. Intuitively, large jumps in the business clock allow the model to accommodate large daily changes in the default intensity (i.e., relative to volatility). While the time-changed model offers only modest improvement in fitting the term-structure of observable spreads on any given date, it greatly improves the fit in the time-series dimension. The findings have important implications for risk-management applications, because the time-changed model allows for much higher kurtosis in the loss distribution over short horizons, and so higher quantiles in the tail (i.e., higher value-atrisk). For similar reasons, the findings also suggest that the widely-used CIR model will tend to underprice deep out-of-the-money options on CDS. While the model imposes a variety of independence restrictions, only one appears to be materiallyatoddswiththedata. CountertoourmodelassumptionofLévytimechange,the smoothed time-series of state variables exhibit serial correlation in time-change increments. The pricing methodology can easily be extended to allow for such volatility clustering, but our MCMC and particle filter methodologies would be quite challenging without the tractability afforded by serial independence. We leave this as an avenue for future research. Appendix: Proposal and resampling weights in linearized model Application of the proposed adapted filter requires sampling from p˘(h |χ ,h ,y ) and t t t−1 t evaluation of the weights pa(y |χ ,h ). We begin with the kernels on the right-hand-side t t t−1 of equation (130). Since pa(y |h ) ∼ N(γ +β h ,ζ2) and the measurement errors are t,m t t,m t,m t independent across maturity, we have pa(y |h ) = Y φ(y ;γ +β h ,ζ) t t t,m t,m t,m t m∈M = √ 1 !#M Y exp − 1 (cid:18)y t,m −γ t,m −β t,m h t (cid:19)2 ! 2πζ 2 ζ m∈M !#M  !2  = √ 1 Y exp− 1 h t −ν t,m  = Y 1 φ(h t ;ν t,m ,ξ t,m ) 2πζ 2 ξ β t,m t,m m∈M m∈M where ν ≡ (y −γ )/β and ξ ≡ ζ/β . From the state equation (8), we have t,m t,m t,m t,m t,m t,m p(h |χ ,h ) = φ(h ;ν ,ξ ) where ν ≡ h +(µ−κ P h )χ and ξ2 ≡ σ2 h χ , t t t−1 t t,∗ t,∗ t,∗ t−1 t−1 t t,∗ t−1 t where the asterisk indicates parameters of the state equation. For notational convenience in combining these kernels, we define the set N = M∪{∗}. 24

The right-hand-side of equation (130) is   pa(y t |h t )p(h t |χ t ,h t−1 ) =  Y 1 φ(h t ;ν t,m ,ξ t,m )φ(h t ;ν t,∗ ,ξ t,∗ ) β t,m m∈M !#M  !2  = √ 1 √ 1 exp− 1 X h t −ν t,m  (16) 2πζ 2πξ 2 ξ t,∗ m∈N t,m For m ∈ N, let ξ2 ≡ Q ξ2 , and let η be the weight η = ξ2 /P ξ2 . t,−m j∈N\{m} t,j t,m t,m t,−m j∈N t,−j 2 Let ξ be the harmonic sum of the variances, i.e., t  −1 2 Q m∈N ξ t 2 ,m X 1 ξ t ≡ P ξ2 =  ξ2  m∈N t,−m m∈N t,m It is straightforward to sum the fractions within the exp() term of equation (16) as m X ∈N h t ξ − t,m ν t,m !2 = h t −P m∈ ξ N t η t,m ν t,m !2 + ξ 1 2 t    m X ∈N η t,m ν t 2 ,m −   m X ∈N η t,m ν t,m   2    Observe that h appears only in the first term, which we embed in the proposal density as t   X p˘(h t |χ t ,h t−1 ,y t ) = φh t ; η t,m ν t,m ,ξ t m∈N Intuitively, the proposal is the posterior distribution for an unobserved state given a set of independent Gaussian signals with realization {ν } . The mth signal receives weight t,m m∈N η , which is proportional to its precision 1/ξ2 . The precision of the posterior, 1/ξ 2 , is t,m t,m t the sum of signal precisions. To enforce identity (130), remaining terms in equation (16) are captured in the resampling weight 1 !#M ξ 1 P η ν2 −(P η ν )2! pa(y |χ ,h ) = √ exp − m∈N t,m t,m m∈N t,m t,m t t t−1 2πζ ξ t,∗ 2 ξ 2 t Theresamplingweightislowwhenthevariationacrosssignalsν islarge. Thus,aparticle t,m carrying state (χ ,h ) is penalized in the resampling step when it implies large differences t t−1 across the #M+1 equations in the likelihood of y . t 25

References Carol Alexander and Andreas Kaeck. Capital structure and asset prices: Some effects of bankruptcy procedures. Journal of Banking and Finance, 32(6):1008–1021, June 2008. Gianni Amisano and Raffaella Giacomini. Comparing density forecasts via weighted likelihood ratio tests. Journal of Business and Economics Statistics, 25(2):177–199, 2007. Torben G. Andersen, Luca Benzoni, and Jesper Lund. An empirical investigation of continuous-time equity return models. Journal of Finance, 57(3):1239–1284, 2002. Ole E. Barndorff-Nielsen and Albert Shiryaev. Change of Time and Change of Measure, volume 13 of Advanced Series on Statistical Science and Applied Probability. World Scientific, 2010. Bradley P. Carlin, Nicholas G. Polson, and David S. Stoffer. A Monte Carlo approach to nonnormal and nonlinear state-space modeling. Journal of the American Statistical Association, 87(418):493–500, 1992. C.K. Carter and R. Kohn. On Gibbs sampling for state space models. Biometrika, 81(3): 541–553, 1994. David A. Chapman and Neil D. Pearson. Is the short rate drift actually nonlinear? Journal of Finance, 55(1):355–388, February 2000. Siddhartha Chib and Edward Greenberg. Markov chain Monte Carlo simulation methods in econometrics. Econometric Theory, 12(3):409–431, 1996. Ovidiu Costin, Michael B. Gordy, Min Huang, and Pawel J. Szerszen. Expectations of functions of stochastic time with application to credit risk modeling. Mathematical Finance, forthcoming. Francis X. Diebold and Roberto S. Mariano. Comparing predictive accuracy. Journal of Business and Economics Statistics, 13(3):252–263, 1995. JoostDriessen. Isdefaulteventriskpricedincorporatebonds? Review of Financial Studies, 18:165–195, 2005. Gregory R. Duffee. Estimating the price of default risk. Review of Financial Studies, 12 (1):197–226, Spring 1999. Darrell Duffie and Kenneth J. Singleton. Modeling term structures of defaultable bonds. Review of Financial Studies, 12(4):687–720, 1999. 26

Darrell Duffie and Kenneth J. Singleton. Credit Risk: Pricing, Measurement, and Management. Princeton University Press, Princeton, 2003. Bjørn Eraker. Do stock prices and volatility jump? Reconciling evidence from spot and option prices. Journal of Finance, 59(3):1367–1403, 2004. Sylvia Frühwirth-Schnatter. Data augmentation and dynamic linear models. Journal of Time Series Analysis, 15(2):183–202, 1994. D. Geman and S. Geman. Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 6:721–741, 1984. N.J. Gordon, D.J. Salmond, and A.F.M. Smith. Novel approach to nonlinear/non-Gaussian Bayesian state estimation. IEE Proceedings F on Radar and Signal Process, 140(2):107– 113, April 1993. Michael B. Gordy and Søren Willemann. Constant proportion debt obligations: A postmortem analysis of rating models. Management Science, 58(3):476–492, March 2012. Refet S. Gürkaynak, Brian Sack, and Jonathan H. Wright. The U.S. Treasury yield curve: 1961 to the present. Journal of Monetary Economics, 54(8):2291–2304, 2007. KrisJacobsandXiaofei Li. Modelingthedynamicsof credit spreadswithstochasticvolatility. Management Science, 54(6):1176–1188, June 2008. Eric Jacquier, Nicholas G. Polson, and Peter E. Rossi. Bayesian analysis of stochastic volatility models with fat-tails and correlated errors. Journal of Econometrics, 122(1): 185–212, September 2004. Robert Jarrow and Stuart Turnbull. Pricing derivatives on financial securities subject to credit risk. Journal of Finance, 50(1):53–86, March 1995. Robert A. Jarrow, David Lando, and Fan Yu. Default risk and diversification: Theory and empirical implications. Mathematical Finance, 15(1):1–26, January 2005. Michael S. Johannes, Nicholas G. Polson, and Jonathan R. Stroud. Optimal filtering of jump diffusions: Extracting latent states from asset prices. Review of Financial Studies, 22(7):2759–2799, 2009. Christopher S. Jones. The dynamics of stochastic volatility: evidence from underlying and options market. Journal of Econometrics, 116(1–2):181–224, 2003. 27

Matthew Leeming, Søren Willemann, Arup Ghosh, and Rob Hagemans. Standard corporate CDS handbook: Ongoing evolution of the CDS market. Technical report, Barclays Capital, February 2010. Hedibert F. Lopes and Ruey S. Tsay. Particle filters and Bayesian inference in financial econometrics. Journal of Forecasting, 30:168–209, 2011. Rafael Mendoza-Arriaga and Vadim Linetsky. Time-changed CIR default intensities with two-sided mean-reverting jumps. Annals of Applied Probability, 24(2):811–856, 2014. N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, A.H. Teller, and E. Teller. Equation of state calculations by fast computing machines. Journal of Chemical Physics, 21(6): 1087–1092, 1953. Jun Pan and Kenneth J. Singleton. Default and recovery implicit in the term structure of sovereign CDS spreads. Journal of Finance, 63(5), October 2008. Peter C.B. Phillips and Jun Yu. Simulation-based estimation of contingent-claims prices. Review of Financial Studies, 22(9):3669–3705, September 2009. Michael K. Pitt and Neil Shephard. Analytic convergence rates and parameterization issues for the Gibbs sampler applied to state space models. Journal of Time Series Analysis, 20(1):63–85, January 1999. Jonathan R. Stroud, Peter Müller, and Nicholas G. Polson. Nonlinear state-space models with state-dependent variances. Journal of the American Statistical Association, 98(462): 377–386, 2003. Ruey S. Tsay. Analysis of Financial Time Series. John Wiley & Sons, Hoboken, NJ, third edition, 2010. 28

Table 1: Counts of missing observations in CDS data. The table reports the number of missing observations in the estimation sample from the beginning of 2002 to the end of 2009 (with the time series for Sprint Nextel starting on August 24, 2005). Maturity 1y 2y 3y 5y 7y 10y Alcoa 0 0 0 0 0 8 Anadarko 6 32 0 0 0 8 CenturyLink 61 115 12 3 11 107 Clear Channel 0 1 0 0 0 0 Ford 0 0 0 0 0 0 Lennar 120 301 71 2 218 233 Limited Brands 18 42 10 3 0 19 RadioShack 16 17 4 0 1 4 Sprint Nextel 0 0 0 0 0 0 Tyson Foods 39 147 22 30 25 208 29

Table 2.a: Parameter estimates of CDS pricing models: Alcoa. The table reports the mean, standard deviation, 5%- and 95%-quantiles of the MCMC posterior distribution for the CDS pricing model in equations (6)-(8). The results are reported for model specifications without time-change (CIR model) and with time-change in default intensity (CIR + IG Time Change Model). Mean Std Dev 5% Quantile 95% Quantile Alcoa: CIR Model P κ 0.4794 0.3698 0.0366 1.1952 σ 0.1877 0.0030 0.1826 0.1925 µ∗100 0.0829 0.0009 0.0814 0.0844 Q κ −0.2526 0.0068 −0.2634 −0.2413 ζ 0.1627 0.0011 0.1609 0.1645 Alcoa: CIR + IG Time Change Model P κ 0.6590 0.5188 0.0483 1.6713 σ 0.2238 0.0019 0.2207 0.2269 µ∗100 0.0688 0.0009 0.0673 0.0702 Q κ −0.3787 0.0068 −0.3903 −0.3679 ζ 0.1583 0.0010 0.1566 0.1600 α 7.1439 0.4384 6.4713 7.9060 Table 2.b: Parameter estimates of CDS pricing models: Anadarko. The table reports the mean, standard deviation, 5%- and 95%-quantiles of the MCMC posterior distribution for the CDS pricing model in equations (6)-(8). The results are reported for model specifications without time-change (CIR model) and with time-change in default intensity (CIR + IG Time Change Model). Mean Std Dev 5% Quantile 95% Quantile Anadarko: CIR Model P κ 0.6569 0.3852 0.0915 1.3452 σ 0.0975 0.0042 0.0910 0.1051 µ∗100 0.1643 0.0012 0.1623 0.1662 Q κ −0.0318 0.0035 −0.0382 −0.0265 ζ 0.1627 0.0011 0.1610 0.1645 Anadarko: CIR + IG Time Change Model P κ 1.7103 1.2431 0.1488 4.0767 σ 0.2596 0.0025 0.2555 0.2636 µ∗100 0.1096 0.0018 0.1067 0.1128 Q κ −0.3398 0.0094 −0.3547 −0.3223 ζ 0.1535 0.0010 0.1519 0.1552 α 5.5933 0.3528 5.0505 6.2100 30

Table 2.c: Parameter estimates of CDS pricing models: CenturyLink. The table reports the mean, standard deviation, 5%- and 95%-quantiles of the MCMC posterior distribution for the CDS pricing model in equations (6)-(8). The results are reported for model specifications without time-change (CIR model) and with time-change in default intensity (CIR + IG Time Change Model). Mean Std Dev 5% Quantile 95% Quantile CenturyLink: CIR Model P κ 1.1616 0.7956 0.1122 2.6471 σ 0.2531 0.0036 0.2471 0.2589 µ∗100 0.2095 0.0031 0.2044 0.2145 Q κ −0.3253 0.0104 −0.3423 −0.3078 ζ 0.2002 0.0013 0.1980 0.2024 CenturyLink: CIR + IG Time Change Model P κ 1.8575 1.3345 0.1640 4.3793 σ 0.2922 0.0023 0.2884 0.2960 µ∗100 0.1558 0.0034 0.1505 0.1618 Q κ −0.5531 0.0132 −0.5730 −0.5302 ζ 0.1944 0.0013 0.1923 0.1965 α 5.9933 0.4425 5.3364 6.7968 Table 2.d: Parameter estimates of CDS pricing models: Clear Channel. The table reports the mean, standard deviation, 5%- and 95%-quantiles of the MCMC posterior distribution for the CDS pricing model in equations (6)-(8). The results are reported for model specifications without time-change (CIR model) and with time-change in default intensity (CIR + IG Time Change Model). Mean Std Dev 5% Quantile 95% Quantile Clear Channel: CIR Model P κ 0.2310 0.1806 0.0171 0.5820 σ 0.2332 0.0040 0.2266 0.2398 µ∗100 0.2594 0.0064 0.2493 0.2702 Q κ −0.3094 0.0115 −0.3275 −0.2904 ζ 0.2689 0.0017 0.2661 0.2719 Clear Channel: CIR + IG Time Change Model P κ 0.1759 0.1373 0.0133 0.4422 σ 0.2662 0.0038 0.2601 0.2725 µ∗100 0.1983 0.0056 0.1889 0.2081 Q κ −0.5013 0.0172 −0.5301 −0.4738 ζ 0.2640 0.0017 0.2613 0.2669 α 3.8124 0.4316 3.1516 4.5369 31

Table 2.e: Parameter estimates of CDS pricing models: Ford. The table reports the mean, standard deviation, 5%- and 95%-quantiles of the MCMC posterior distribution for the CDS pricing model in equations (6)-(8). The results are reported for model specifications without time-change (CIR model) and with time-change in default intensity (CIR + IG Time Change Model). Mean Std Dev 5% Quantile 95% Quantile Ford: CIR Model P κ 0.3466 0.2579 0.0283 0.8446 σ 0.3207 0.0028 0.3162 0.3254 µ∗100 0.4686 0.0094 0.4529 0.4834 Q κ −0.4997 0.0086 −0.5136 −0.4859 ζ 0.1824 0.0012 0.1805 0.1844 Ford: CIR + IG Time Change Model P κ 0.3299 0.2508 0.0263 0.8129 σ 0.3390 0.0028 0.3345 0.3436 µ∗100 0.4147 0.0112 0.3958 0.4318 Q κ −0.6419 0.0127 −0.6626 −0.6213 ζ 0.1801 0.0012 0.1782 0.1821 α 7.5902 0.6288 6.6347 8.6549 Table 2.f: Parameter estimates of CDS pricing models: Lennar. The table reports the mean, standard deviation, 5%- and 95%-quantiles of the MCMC posterior distribution for the CDS pricing model in equations (6)-(8). The results are reported for model specifications without time-change (CIR model) and with time-change in default intensity (CIR + IG Time Change Model). Mean Std Dev 5% Quantile 95% Quantile Lennar: CIR Model P κ 0.4378 0.3252 0.0356 1.0646 σ 0.2576 0.0057 0.2479 0.2670 µ∗100 0.2983 0.0037 0.2924 0.3047 Q κ −0.1736 0.0114 −0.1927 −0.1545 ζ 0.1802 0.0013 0.1782 0.1823 Lennar: CIR + IG Time Change Model P κ 0.6873 0.5193 0.0537 1.6897 σ 0.3197 0.0037 0.3137 0.3258 µ∗100 0.2593 0.0027 0.2549 0.2639 Q κ −0.3330 0.0103 −0.3500 −0.3158 ζ 0.1771 0.0012 0.1751 0.1791 α 18.1031 1.4843 15.8002 20.6781 32

Table 2.g: Parameter estimates of CDS pricing models: Limited Brands. The table reports the mean, standard deviation, 5%- and 95%-quantiles of the MCMC posterior distribution for the CDS pricing model in equations (6)-(8). The results are reported for model specifications without time-change (CIR model) and with time-change in default intensity (CIR + IG Time Change Model). Mean Std Dev 5% Quantile 95% Quantile Limited Brands: CIR Model P κ 0.3738 0.2655 0.0343 0.8747 σ 0.1334 0.0068 0.1218 0.1442 µ∗100 0.2292 0.0041 0.2227 0.2361 Q κ −0.0922 0.0088 −0.1063 −0.0771 ζ 0.2399 0.0016 0.2374 0.2425 Limited Brands: CIR + IG Time Change Model P κ 0.9273 0.7077 0.0722 2.2911 σ 0.2531 0.0034 0.2477 0.2587 µ∗100 0.1518 0.0037 0.1453 0.1576 Q κ −0.3907 0.0148 −0.4178 −0.3680 ζ 0.2334 0.0015 0.2309 0.2358 α 3.8126 0.3055 3.3556 4.3503 Table 2.h: Parameter estimates of CDS pricing models: RadioShack. The table reports the mean, standard deviation, 5%- and 95%-quantiles of the MCMC posterior distribution for the CDS pricing model in equations (6)-(8). The results are reported for model specifications without time-change (CIR model) and with time-change in default intensity (CIR + IG Time Change Model). Mean Std Dev 5% Quantile 95% Quantile RadioShack: CIR Model P κ 0.6419 0.4726 0.0538 1.5468 σ 0.1907 0.0025 0.1866 0.1947 µ∗100 0.0681 0.0022 0.0644 0.0717 Q κ −0.3902 0.0082 −0.4036 −0.3764 ζ 0.2660 0.0018 0.2632 0.2689 RadioShack: CIR + IG Time Change Model P κ 1.3374 1.0143 0.1068 3.2928 σ 0.1968 0.0025 0.1928 0.2008 µ∗100 0.0388 0.0018 0.0358 0.0418 Q κ −0.6591 0.0167 −0.6856 −0.6314 ζ 0.2561 0.0016 0.2534 0.2588 α 1.7946 0.0901 1.6592 1.9488 33

Table 2.i: Parameter estimates of CDS pricing models: Sprint. The table reports the mean, standard deviation, 5%- and 95%-quantiles of the MCMC posterior distribution for the CDS pricing model in equations (6)-(8). The results are reported for model specifications without time-change (CIR model) and with time-change in default intensity (CIR + IG Time Change Model). Mean Std Dev 5% Quantile 95% Quantile Sprint: CIR Model P κ 0.5126 0.4062 0.0367 1.3034 σ 0.2404 0.0032 0.2351 0.2456 µ∗100 0.1407 0.0018 0.1378 0.1437 Q κ −0.3992 0.0101 −0.4154 −0.3820 ζ 0.1836 0.0017 0.1809 0.1864 Sprint: CIR + IG Time Change Model P κ 0.5894 0.4706 0.0422 1.5102 σ 0.2543 0.0028 0.2498 0.2591 µ∗100 0.1320 0.0018 0.1291 0.1348 Q κ −0.4586 0.0106 −0.4769 −0.4424 ζ 0.1822 0.0016 0.1796 0.1848 α 21.3563 2.2668 17.8964 25.3523 Table 2.j: Parameter estimates of CDS pricing models: Tyson Foods. The table reports the mean, standard deviation, 5%- and 95%-quantiles of the MCMC posterior distribution for the CDS pricing model in equations (6)-(8). The results are reported for model specifications without time-change (CIR model) and with time-change in default intensity (CIR + IG Time Change Model). Mean Std Dev 5% Quantile 95% Quantile Tyson Foods: CIR Model P κ 0.7200 0.5119 0.0630 1.6877 σ 0.2469 0.0034 0.2412 0.2524 µ∗100 0.2753 0.0027 0.2708 0.2798 Q κ −0.2400 0.0075 −0.2524 −0.2279 ζ 0.1683 0.0011 0.1665 0.1702 Tyson Foods: CIR + IG Time Change Model P κ 1.1482 0.8454 0.0946 2.7624 σ 0.2916 0.0027 0.2870 0.2960 µ∗100 0.2384 0.0026 0.2339 0.2427 Q κ −0.3805 0.0091 −0.3949 −0.3651 ζ 0.1647 0.0011 0.1630 0.1665 α 12.1615 0.8741 10.8246 13.6870 34

Table 3: Moments of measurement equation residuals. Thetablereportstheposteriormeanofmean,std.deviation,skewnessandkurtosisofmeasurement equation residuals, ε(y), calculated at each iteration of MCMC algorithm for model specifications withouttime-change(CIRmodel)andwithtime-changeindefaultintensity(CIR+IGTimeChange Model). A correct specification implies that the residuals are standard normal with mean zero, standard deviation one, skewness zero and kurtosis three. Mean Std Dev Skewness Kurtosis Alcoa CIR Model −0.0157 0.9998 −0.1532 4.3142 CIR-IG TC Model −0.0051 1.0000 0.0274 4.4102 Anadarko CIR Model −0.0048 1.0000 −0.4178 3.4865 CIR-IG TC Model −0.0044 0.9999 −0.1429 3.4420 CenturyLink CIR Model −0.0169 0.9998 −0.6530 4.0214 CIR-IG TC Model −0.0041 0.9999 −0.4849 4.0898 Clear Channel CIR Model −0.0112 0.9999 −0.2315 3.7204 CIR-IG TC Model −0.0049 1.0000 −0.0709 3.6722 Ford CIR Model −0.0125 0.9999 −0.4358 4.3982 CIR-IG TC Model −0.0048 0.9999 −0.3012 4.4348 Lennar CIR Model −0.0097 0.9999 −0.3456 3.9502 CIR-IG TC Model −0.0040 0.9999 −0.1962 3.8164 Limited Brands CIR Model −0.0068 0.9999 −0.6464 3.5192 CIR-IG TC Model −0.0028 1.0000 −0.4513 3.2429 RadioShack CIR Model −0.0208 0.9998 −0.5183 3.5148 CIR-IG TC Model −0.0022 1.0000 −0.2493 3.3394 Sprint CIR Model −0.0122 0.9999 −0.0525 3.2900 CIR-IG TC Model −0.0047 1.0000 −0.0032 3.3974 Tyson Foods CIR Model −0.0134 0.9999 0.0039 3.6866 CIR-IG TC Model −0.0042 1.0000 0.1442 3.7100 35

Table 4: Moments of state equation innovations. Thetablereportstheposteriormeanofmean,std.deviation,skewnessandkurtosisofstateequation residuals, ε(λ), calculated at each iteration of MCMC algorithm for model specifications without time-change(CIRmodel)andwithtime-changeindefaultintensity(CIR+IGTimeChangeModel). A correct specification implies that the residuals are standard normal with mean zero, standard deviation one, skewness zero and kurtosis three. Mean Std Dev Skewness Kurtosis Alcoa CIR Model 0.0463 0.6674 0.3405 4.2667 CIR-IG TC Model 0.0108 0.8230 0.0316 3.2158 Anadarko CIR Model 0.0159 0.9196 0.1544 3.2541 CIR-IG TC Model −0.0018 0.8354 0.0232 3.1892 CenturyLink CIR Model 0.0492 0.7081 0.2956 3.3331 CIR-IG TC Model −0.0024 0.8671 0.0242 3.1424 Clear Channel CIR Model 0.0420 0.8925 0.1341 3.4407 CIR-IG TC Model 0.0015 0.9581 0.0177 3.0508 Ford CIR Model 0.0306 0.8265 0.1025 3.4196 CIR-IG TC Model 0.0001 0.9351 0.0156 3.0696 Lennar CIR Model 0.0436 0.7893 0.2021 3.3338 CIR-IG TC Model 0.0127 0.8083 0.0392 3.2169 Limited Brands CIR Model 0.0308 0.9215 0.1456 3.1110 CIR-IG TC Model −0.0023 0.9208 0.0173 3.0802 RadioShack CIR Model 0.0359 0.7439 0.1330 3.1776 CIR-IG TC Model −0.0040 0.9290 0.0065 3.0734 Sprint CIR Model 0.0728 0.7659 0.4195 4.4997 CIR-IG TC Model 0.0398 0.8156 0.0451 3.2393 Tyson Foods CIR Model 0.0406 0.7290 0.2760 3.4170 CIR-IG TC Model 0.0003 0.8201 0.0351 3.2036 36

Table 5: Forecast distribution for 1-day horizon. For the CIR model, we report the quantiles of the forecast distribution of CDS(λ ) for δ = 1/250 δ (i.e., a horizon of one trading day) given initial condition λ and the posterior mean parameters 0 estimatedforAlcoa. Forthetime-changedmodelwereportthequantilesoftheforecastdistribution of CDS(λ(T );α). Two values of λ are considered, one below the long-run mean (5 bp) and one δ 0 above the long-run mean (50 bp). Parameters for each model are fixed to the mean of the posterior distribution. A riskless rate is fixed at 3% and recovery rate at 40%. λ (bp) Quantiles of 5-year CDS(λ(T )) in basis points 0 δ 0.001 0.01 0.1 0.25 0.5 0.75 0.9 0.99 0.999 CIR Model 5 17.2 17.6 18.9 20.0 21.5 23.3 25.2 28.9 32.2 50 42.7 47.2 54.1 58.5 63.6 69.1 74.3 83.9 90.9 CIR-IG TC Model 5 17.5 17.5 21.2 22.5 23.1 23.7 24.7 32.7 61.8 50 17.5 40.7 68.8 72.5 74.3 76.1 79.3 104.5 177.6 Table 6: Validation of independence assumptions. For the time-changed model, we report the posterior mean and standard deviation of correlations betweeninnovationsε(λ)inthedefaultintensityprocessandtime-changeincrementsχ ,andbetween t t ε(λ) and changes in the federal funds rate (dFFR =FFR −FFR ). t t t t−1 h i h i Corr ε(λ),χ Corr ε(λ),dFFR t Mean Std Dev Mean Std Dev Alcoa 0.0248 0.0125 0.0033 0.0204 Anadarko 0.0209 0.0122 0.0023 0.0218 CenturyLink 0.0212 0.0110 0.0048 0.0209 Clear Channel 0.0284 0.0145 0.0000 0.0216 Ford 0.0251 0.0138 −0.0083 0.0213 Lennar 0.0231 0.0128 0.0109 0.0225 Limited Brands 0.0235 0.0125 0.0010 0.0214 RadioShack 0.0197 0.0109 0.0024 0.0221 Sprint 0.0503 0.0184 −0.0014 0.0279 Tyson Foods 0.0276 0.0122 0.0030 0.0204 37

Table 7: Out-of-sample density forecast performance results. The table reports Amisano and Giacomini (2007) test statistics defined in equation (15). Under the nullhypothesisofequalperformance,thestatisticshaveanasymptoticstandardnormaldistribution. The positive (negative) values support the time-changed (standard CIR) model. U Alcoa 10.98 Anadarko 6.47 CenturyLink 17.65 Clear Channel 47.62 Ford 19.44 Lennar −14.73 Limited Brands 63.07 RadioShack 39.79 Sprint 9.45 Tyson Foods 23.60 38

Figure 1: Daily changes in log-spreads. Daily changes in the log-spread for the on-the-run 5 year CDX.NA.IG index. Index roll dates (marked on the x-axis) are dropped to avoid contamination from changes in index composition. 0.25 0.2 0.15 0.1 0.05 0 −0.05 −0.1 −0.15 −0.2 −0.25 sdaerps−gol ni egnahC 21O ct0 2 3 2M ar0 2 4 0Sep0 2 4 1M ar0 2 5 0Sep0 2 5 0M ar0 2 6 0Sep0 2 6 0M ar0 2 7 0Sep0 2 7 5M ar08 02O ct0 2 8 0M ar09 21Sep0 2 9 2M ar1 2 0 0Sep1 2 0 1M ar1 2 1 0Sep1 2 1 0M ar1 2 2 0Sep1 2 2 0M ar13 39

Figure 2: Kurtosis of change in the CDS spread. Stationary CIR model under business time with parameters κP = κQ = 0.2,µ = 0.004,σ = 0.1, and R = 0.4. Black dotted line marks limit as δ → ∞ of kurtosis of spread change in CIR model. Both axes on log-scale. Moments of the calendar time increments are obtained by simulation with 4 million trials. 104 103 102 101 10−1 100 101 102 103 α sisotruK 1 day 1 month 1 year 40

Figure 3: In-sample CDS spreads and smoothed state estimates: Alcoa. We plot CDS spreads (top panel, log-scale), smoothed estimates of the default intensity (middle panel, log-scale) and smoothed estimates of time-change increments (bottom panel) for Alcoa. Smoothed estimates of the default intensity are obtained as the posterior mean of the MCMC chain forthemodelwithouttime-change(CIR)andwithtime-changeindefaultintensity(CIR+IGTime Change). The smoothed estimates of time-change increments for the model with time-change are scaled by ∆ for ease of comparison against model without time-change. 0 10 −2 10 −4 10 sdaerps SDC CDS spreads 1 year 5 year 10 year 2 2 2 2 3 3 3 3 4 4 4 4 5 5 5 5 6 6 6 6 7 7 7 7 8 8 8 8 9 9 9 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 0/ 2/ 1/ 1/ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 1 1 1 1 1 1 1 1 smoothed default intensity 0 10 CIR CIR + IG time change −2 10 t ˜λ −4 10 −6 10 2 2 2 2 3 3 3 3 4 4 4 4 5 5 5 5 6 6 6 6 7 7 7 7 8 8 8 8 9 9 9 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 0/ 2/ 1/ 1/ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 1 1 1 1 1 1 1 1 smoothed time change increments 15 10 ∆ / t χ 5 0 2 2 2 2 3 3 3 3 4 4 4 4 5 5 5 5 6 6 6 6 7 7 7 7 8 8 8 8 9 9 9 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 0/ 2/ 1/ 1/ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 1 1 1 1 1 1 1 1 41

Figure 4: In-sample CDS spreads and smoothed state estimates: Ford. We plot CDS spreads (top panel, log-scale), smoothed estimates of the default intensity (middle panel, log-scale) and smoothed estimates of time-change increments (bottom panel) for Ford. Smoothed estimates of the default intensity are obtained as the posterior mean of the MCMC chain forthemodelwithouttime-change(CIR)andwithtime-changeindefaultintensity(CIR+IGTime Change). The smoothed estimates of time-change increments for the model with time-change are scaled by ∆ for ease of comparison against model without time-change. 1 10 −1 10 −3 10 sdaerps SDC CDS spreads 1 year 5 year 10 year 2 2 2 2 3 3 3 3 4 4 4 4 5 5 5 5 6 6 6 6 7 7 7 7 8 8 8 8 9 9 9 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 0/ 2/ 1/1/ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/2/ 1 1 1 1 1 1 1 1 smoothed default intensity 0 10 CIR CIR + IG time change −1 10 t ˜λ −2 10 −3 10 2 2 2 2 3 3 3 3 4 4 4 4 5 5 5 5 6 6 6 6 7 7 7 7 8 8 8 8 9 9 9 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 0/ 2/ 1/1/ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/2/ 1 1 1 1 1 1 1 1 smoothed time change increments 25 20 ∆ 15 / t χ 10 5 0 2 2 2 2 3 3 3 3 4 4 4 4 5 5 5 5 6 6 6 6 7 7 7 7 8 8 8 8 9 9 9 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 0/ 2/ 2/ 0/ 2/ 1/1/ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 3/ 6/ 9/2/ 1 1 1 1 1 1 1 1 42

Figure 5: Model-implied term structure of spreads for Alcoa. Term-structure of credit spreads for Alcoa on three dates for fitted model with time-change (solid lines) and without time-change (dashed lines). Spreads plotted in basis points on a log scale. 103 19−Mar−2009 102 19−Jun−2008 18−Mar−2005 101 1 2 3 4 5 6 7 8 9 10 Maturity )pb( daerpS 43

Figure 6: Out-of-sample CDS spreads and filtered state estimates: Ford. WeplotCDSspreads(toppanel, log-scale), filteredestimatesofthedefaultintensity(middlepanel, log-scale) and filtered estimates of time-change increments (bottom panel) for Ford. Filtered estimates of the default intensity are obtained as the mean of the filtered distribution produced by the particle filter for the model without time-change (CIR) and with time-change in default intensity (CIR + IG Time Change). The filtered estimates of time-change increments for the model with time-change are scaled by ∆ for ease of comparison against model without time-change. −1 10 −2 10 −3 10 sdaerps SDC CDS spreads 1 year 5 year 10 year 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/ 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 1 1 filtered default intensity −1 10 t 10 −2 ˜λ CIR CIR + IG time change −3 10 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/ 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 1 1 filtered time change increments 10 ∆ / 5 t χ 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/ 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 1 1 44

Figure 7: Out-of-sample CDS spreads and filtered state estimates: Lennar. WeplotCDSspreads(toppanel, log-scale), filteredestimatesofthedefaultintensity(middlepanel, log-scale) and filtered estimates of time-change increments (bottom panel) for Lennar. Filtered estimates of the default intensity are obtained as the mean of the filtered distribution produced by the particle filter for the model without time-change (CIR) and with time-change in default intensity (CIR + IG Time Change). The filtered estimates of time-change increments for the model with time-change are scaled by ∆ for ease of comparison against model without time-change. −1 10 −2 10 −3 10 sdaerps SDC CDS spreads 1 year 5 year 10 year 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/ 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 1 1 filtered default intensity −1 10 CIR CIR + IG time change t ˜λ −2 10 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/ 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 1 1 filtered time change increments 10 ∆ / 5 t χ 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2/ 1/ 0/ 0/ 1/ 0/ 0/ 0/ 2 2 2 2 2 2 2 2 3/ 6/ 9/ 2/ 3/ 6/ 9/ 2/ 1 1 45