On Pricing CDOs with Meixner Distribution

Kridsda Nimmanunta, Anant Chiarawongse, Sunti Tirapat

Department of Banking and Finance Faculty of Commerce and Accountancy Chulalongkorn University, Thailand

Abstract
This paper provides a framework to price CDOs using an asymmetric dependence structure based on the Meixner distribution/process since it possesses desirable properties such as fat-tail, skewness, and jump component and is relatively simple to implement comparing to other Lvy processes. It is shown that the Meixner distribution can be applied to both copula and structural form approaches. Using the historical prices of CDOs on the CDX NA IG, the performances of the proposed models are examined and compared to those of standard models such as Gaussian copula model, double-t copula model, and correlated Brownian motion structural model. It is found that the Meixner-based models have the edge over the standard models in all cases in terms of the mean absolute pricing errors (MAPEs). Using the paired Z-test also conrms that the proposed models seem to outperform the standard ones. The risk measures of the models are also examined. It is shown that the risk measures from Meixner-based models are more intuitive than those of the standard models.
Key words: CDOs, Meixner, Lvy, Copula Models, Structural Models JEL classication: C40, C63

Introduction

CDOs have been growing rapidly in recent years. The market size for CDOs was estimated to be $2 trillion by the end of 2006 which covers around 40%

Kridsda Nimmanunta gratefully acknowledges the support by Chulalongkorn University Graduate Scholarship to Commemorate the 72nd Anniversary of His Majesty King Bhumibol Adulyadej. Email addresses: [email protected] (Kridsda Nimmanunta), [email protected] (Anant Chiarawongse), [email protected] (Sunti Tirapat).
Preprint submitted to MF A 2008 January 13, 2008

of the total credit derivative market, making pricing and risk management for CDOs essential for market participants. In general, there are two major approaches to model CDOs: copula-based and correlated structural models. In most copula-based models (see for example Li; 2000; Hull and White; 2004; Brunlid; 2006), the marginal default probabilities are estimated via a reducedform model and then are combined using a copula to derive the portfolio loss distributions. In correlated structural models (see for example Hull et al.; 2005; Luciano and Schoutens; 2005; Baxter; 2006), usually Monte Carlo methods are employed to generate paths of rm value under certain dependence structure. A rm defaults when its rm value falls below some thresholds. Basic credit-risk models are based on Gaussian distribution. Recent studies attempt to improve the models by applying alternative distributions such as Student-t and double-t; both are symmetric and have fat tails (excess kurtosis). A recent trend in Finance is to move beyond Gaussian distribution to more generalized distribution such as Lvy distribution. Lvy distribution can have skewness and excess kurtosis. When considering its associated process, the Lvy process has three components i.e. deterministic, Brownian, and Lvy measure which can be interpreted as drift, diusion, and jump parts, respectively; while a Brownian process has only drift and diusion parts. See Schoutens (2003) for Lvy distributions/processes and its applications in nance. There are many special cases of Lvy distributions, e.g., Compound Poisson, Normal Inverse Gaussian, Variance Gamma, CGMY, Generalized Hyperbolic, but most of them do not have analytic expressions for their density functions. Usually, only analytical characteristic functions are available . The Meixner distribution is an attractive choice in credit risk modeling since it has analytical density function as well as nice properties for nancial assets 1 . This study is an exploratory investigation of the pricing models of CDOs. The objective is to provide a framework for incorporating the Meixner distributions into the copula-based and structural approaches in modeling credit risk. The paper also compares the performance of the Meixner-based models to the more traditional ones in pricing the CDOs. It also investigates the characteristics of the risk measures based on expected discounted loss (EDL), Value-at-Risk (VaR), and expected shortfall (ES) for each model. Using the series 5 and 6 of CDX NA IG from September 20, 2005 to October 13, 2006 and from March 23, 2006 to October 16, 2006, we compare the pricing errors of the proposed models with the traditional ones. It is found that the

The skewness and kurtosis of the Meixner distribution are adjustable which is suitable for modeling the return distribution, while the Meixner process is a pure jump process which is suitable for modeling the movement of rm value. See Schoutens (2002); Albrecher et al. (2006).
1

Meixner-based models outperforms the standard models in all cases. Moreover, under Meixner distribution, the movements at lower tail are more dependent than those at the upper tail. Regarding the risk measures for the proposed models, the results show that the risk measures derived from Meixner copula model, Meixner structural model, and Double-t copula model are more sensible than those of Gaussian copula and Brownian structural models which are extremely low for the senior tranches of the CDOs. This is not surprising since both Gaussian copula model and Brownian structural model do not have the fat-tail property. The paper proceeds as follows. Section 2 reviews the previous works in CDOs pricing. The data and methodology are discussed in Section 3. In Section 4, we report the accuracy of the proposed models and compare them with the traditional models. Section 5 concludes and discusses the main ndings.

A Review of Previous Studies

The asset correlation is essential in pricing the CDOs. The pioneer work is done by Gupton et al. (1997) for JPMorgan's CreditMetrics where default probabilities are derived from the transition matrix. Since it is assumed that the asset value follows the Gaussian distribution, the probability of credit quality migration is interpreted as the likelihood that the asset value falls below a threshold. Hence, the probability of joint default can be computed by simulation based on the equity correlations. In this section we briey review two strands of developments based on the copula and the structural form models.

2.1

Copula Models

The default (survival) time copula model was proposed by Li (2000) which has now become the industry's standard model. Under this framework, the default probabilities are implied from the defaultable instruments such as bonds, and those marginal default probabilities are linked together using Gaussian copula. Similar to the CreditMetrics, the equity correlation matrix is used as the Gaussian copula's parameter. The framework is later extended to other copulas such as Student-t and Archimedean, (see Schnbucher; 2000; Laurent and Gregory; 2003; Galiani; 2003; Elizalde; 2005). Since the credit risk tends to move together in the extreme events, Hull and White (2004) apply the onefactor double-t copula model to price CDOs and nth -to-default CDSs. They show that the double-t model ts the prices reasonably well since it takes into account the tail dependence in a pool of credit risk. 3

Burtschell et al. (2005) compares the following copula models: Gaussian, stochastic correlation, Student-t, double-t, Clayton and Marshall-Olkin. It is found that the Student-t and Clayton copula models give results similar to the Gaussian copula whereas Marshall-Olkin copula model results in overly fattening of the tail of the loss distributions. The double-t model lies in between and gives a better t to market quotes while the stochastic correlation copula model can produce a reasonable skew as observed in the market. In recent years, Lvy-based models have become widely used due to their exibility. Kalemanova et al. (2005) apply a Lvy distribution called Normal Inverse Gaussian (NIG) with the assumption of large homogeneous portfolio to extend the standard copula model. The model tted the market quotes reasonably well. Moosbrucker (2006a) and Moosbrucker (2006b) further extend this approach for Variance Gamma copula model and other distributions. Brunlid (2006) compares a number of Hyperbolic Lvy copula models i.e. Normal Inverse Gaussian, Variance Gamma, Skewed Student t copula models. The results suggest that the distribution of credit risk should have negative skewness and high kurtosis. In general, Albrecher et al. (2006) suggest that one can use any Lvy processes and their associated distributions such as Shifted Gamma, Shifted Inverse Gaussian, Variance Gamma, Normal Inverse Gaussian, Meixner, and many more for the one-factor copula model.

2.2

Correlated-Structural Models

Hull, Predescu and White (2005) propose a CDOs pricing methodology based on the rst passage model by Black and Cox (1976). Under this framework, the rm's asset value is assumed to follow the correlated geometric Brownian motion. The model is also extended to incorporate the empirical evidences that default correlations are positively dependent on the default rates and the recovery rates are negatively dependent on the default rates. Their study nds that the model with stochastic correlation has done a good job in improving the accuracy of the model. The model with stochastic recovery rate also has some improvements but less than the case of stochastic correlation. Similar to the copula framework, Lvy-based models under the structural framework have been recently investigated. Luciano and Schoutens (2005) provide a model based on assumptions that the market factor component follows Gamma process whereas the idiosyncratic component follows Wiener process. Their assumptions lead to the Variance Gamma process. Unlike the work of Luciano and Schoutens (2005) that only investigates the case of identical Gamma process with independent Brownian motions, Moosbrucker (2006b) further extends the framework to various cases: correlated Gamma processes with independent Brownian motions, identical Gamma processes with de4

pendent Brownian motions, and sum of two independent Variance Gamma processes. The study also compares these structural models with the copula based models. It is found that in general the Variance Gamma structural model is better in tting the market quotes than the Gaussian and double-t copula models. In addition, Baxter (2006) studies a number of Lvy structural models such as Catastrophe Gamma, Variance Gamma, Gamma, Brownian Gamma structural models, and many combinations, etc. toward single-name and multiname credits and conclude that the simple Gamma is eective in pricing the credit derivatives. Although Lvy based models seem to outperform the traditional models, they are quite dicult to implement in practice. Recently the Meixner process is introduced by Schoutens and Teugels; 1998. Later, Schoutens (2002) is the rst to apply Meixner distribution to t daily stock returns. The study also extends Black and Scholes (1973)'s model by using Meixner process and derive the closed-form solutions for option valueand concludes that the Meixner distribution is more realistic than Gaussian distribution. In this study we propose to use the Meixner distribution as an alternative due to its ability in explaining nancial assets returns and its implementation's simplicity. The framework in extending the traditional credit risk models with the Meixner distribution are presented in the next section.

3
3.1

Data and Methodology

Data

The study tests the performance of each model by pricing the CDOs written on CDX NA IG. The CDX NA IG is a credit default swap index. It consists of 125 most actively-traded CDSs of investment-graded rms in North America, equally weighted. It is divided into ve standard tranches: 0-3%, 3-7%, 7-10%, 10-15% and 15-30%. One can buy or sell each tranche separately, called the single tranche trade. We use daily mid bid-ask price quotes of the CDX NA IG series 5 and 6 from September 20, 2005 to October 13, 2006, and from March 23, 2006 to October 16, 2006, respectively. The CDOs with maturity of 5-, 7-, and 10-year are investigated. Only the days where all tranches have both bid and ask quotes are included. The CDO data are from Reuters (GFI) while CDS index and USD Zero Curves are downloaded from DataStream. We also use the linear interpolation for the interest rates that fall between the available maturities. 5

3.2

Methodology

This section reviews the models used in this study. However only the Meixnerbased models are discussed in detail. The rst model is Gaussian copula model proposed by Li (2000), which has become the industry's standard model. The second model is double-t copula by Hull and White (2004), which has the capability to capture the premium structure via the tail dependence. The third model relax the normality assumption of Gaussian copula model by replacing Gaussian distribution with Meixner distribution. The fourth model is based on the structural model proposed by Hull, Predescu and White (2005). The last model utilizes Meixner process to relax the assumption of path continuity in the standard structural model. We call the last model correlatedMeixner structural model, or more generally correlated-Lvy structural model. For more details on Meixner distribution and process, see Appendix A.

3.2.1

CDO Pricing Framework

Credit derivatives have a similar structure to interest rate swaps: both have a oating leg and a xed leg. The oating leg for credit derivatives, also known as the default leg, represents the amount of losses the investors have to pay when defaults occur; while the xed leg, also known as the premium leg, represents the amount of premium investors receive periodically as a compensation for the protection they provide. 2 The key quantity in the pricing of CDOs is a spread s. Let ti denote the time at the end of an arbitrary period i, and P (ti ) the corresponding notional principal of one of the tranches. For the premium leg, the accrued premium paid at time ti is an annualized xed proportion s to the sum of the expected outstanding notional principal at that particular default time. The spread is determined by equating the value of both legs at initiation. Suppose as in Hull and White (2006b) that the defaults, if any, occur at the middle of each period . The value of the premium leg is the present value of all the cash ows:
n

premium leg = s
i=1 n

(ti ti1 )E [P (ti )]erti (

i=1
ti +ti1 ti ti1 ){E [P (ti1 )] E [P (ti )]}er( 2 ) 2

+s
2

CDOs have xed payment dates which are on 20th of March, June, September and December of each year.
6

The value of the default leg at time ti is the sum of discounted change of the expected notional principal of the tranche from the previous period:
n

default leg =
i=1

{E [P (ti1 )] E [P (ti )]}erti

Then, s is chosen to equate the value of the two legs:

s=

n i=1 (ti

n rti i=1 {E [P (ti1 )] E [P (ti )]}e ti +ti1 ti1 ti1 )E [P (ti )]erti + ( ti 2 )(E [P (ti1 )] E [P (ti )])er( 2 )

(1)

Now we explain how E [P (ti )] can be computed. Let KL and KU denote the lower and upper loss limits of the tranche, respectively. Let Zti stands for the portfolio loss rate at time ti , representing the percentage of the cumulative loss in the portfolio value at time ti . Given Zti , the value of the outstanding notional principal of the tranche is:

P (ti ; Zti ) = P (0) min 1, max 0,

(KU Zti ) (KU KL )

(2)

where P (0) is the initial notional principal of the tranche. The expected outstanding notional principal of the tranche at the future time ti is then:
z =1

E [P (ti )] =

z =0

P (ti ; z )dFti (z )
z

P (ti ; z ) P r(Z = z )

(3)

where Fti is the cumulative distribution function of the portfolio loss rate Zti . See Elizalde (2005); Hull and White (2006b); Laurent and Gregory (2003) for more details. It is apparent that the component that drives the value of CDOs is the distribution of portfolio loss rate. All models follow the above framework, but they dier on how the portfolio loss distributions are modeled. In this study, we will rely on the following assumptions which are common in practice: (1) The underlying portfolio is homogeneous, so their recovery rates, marginal default probabilities, and their default correlations are the same among all rms. (2) The recovery rates R of each rm are all constant at 40%, conforming to the research of Varma and Cantor (2005). (3) The default correlations are assumed to be constant over time. (4) The weights of rms in the portfolio are equal and are constant over time. 7

As a consequence, the portfolio loss rate Zti at time ti can be expressed as follow X (1 R) (4) Zti = N where X is the number of default rms by time ti , and N is the number of rms in the portfolio. The portfolio loss distribution function Fti can be derived from the distribution function of the number of defaults or the default rate. The next two subsections discuss how the default rate and hence the portfolio loss distribution can be derived under the Meixner distribution.

3.2.2

Meixner Copula Model

First, we explain the general framework for copula approaches. We introduce the model where rm defaults are independent and then extend to the more realistic one. Meixner copula model are formulated afterward. Let 1k,ti ,{def ault} be the indicator function of the default event of rm k by time ti . Then the number of default rms is X = N k=1 1k,ti ,{def ault} . If we assume that the individual default probabilities at time ti are all pti , and the defaults occur independently among each other, the probability that m out of N rms default will follow a binomial distribution.

P r(X = m) =

N m

pti (1 pti )
m N m

(5)

P r(X m) =
k=1

N k N k pti (1 pti )

(6)

As a result, the portfolio loss distribution becomes

Fti (z ) = P r X

zN (1 R)

zN/(1R)

=
k=1

N k N k pti (1 pti )

(7)

We follow Li (2000) who applies a survival time model to the marginal default probability pti : pti = 1 eti where is a hazard rate.

(8)

In the real world, the default is dependent. We model the default dependence via a single-factor model. Let be the degree of dependency. The proxy for the default of rm j , Xj , can be decomposed into the common factor M and 8

the nonsystematic part Yj : 3

Xj =

M +

1 Yj

If the market factor M is given, the probability of default pti |m is conditional independent: 1 FM m +Y (pti ) pti |m = FY (9) 1 Hence, the portfolio loss distribution under all possible factor values based on our assumptions is

k=1

m=+ zN/(1R)

Fti (z ) =

pti |m (1 pti |m )
k N k

m=

dFM (m)

(10)

where FM , FY and FM +Y are the cumulative probability distributions of M, Y and X , respectively. For our Meixner copula model, we substitute FM , FY and FM +Y with , (.), a Meixner distribution with zero-mean and unit-variance as described in Appendix A. We will nd the Meixner's parameters and that minimize the pricing error among all of the market quotes. Consequently, the portfolio loss distribution function (9) and (10) become:
m=+ zN/(1R)

Fti (z ) =

m=

N k N k d, (m) pti |m (1 pti |m )

k=1

(11)

1 ti , 1 e p(ti |M = m) = , 1

(12)

3.2.3

Correlated-Meixner Structural Model

This section explains the generalized correlated structural model that relies on Monte Carlo simulation, derived from the frameworks in Hull et al. (2005); Schoutens (2002, 2003); Luciano and Schoutens (2005); Black and Cox (1976); Hull and White (2001). We starts with a structural model in an independent
3

return, etc.

Xj may be the survival time, the rst passage time, the rm value, or the asset

world rst, then extending it to the correlated structural case. After that, the correlated-Meixner structural model is presented. Assuming that the rm's asset value follows an exponential of a Lvy process, under the risk-neutral setting, the asset value follows

Vj (t) = Vj (0)exp (j j )t + Zj (t)

where Vj (t) is the value of rm j at time t

(L)

(13)

j is the adjusted risk-neutral rate of return j is the mean-correcting parameter that makes the rm value process becomes a martingale Zj is a Lvy process
Rescaling the Lvy process can reduce the number of parameters. Hence,
(L)

Vj (t) = Vj (0)exp (j j )t + Xj (t)

where is the scaling factor of the process and Xj (t), a scaled Lvy process (L) of Zj , corresponds to the proxy of asset returns. To construct the distribution of portfolio default rate, we need a default condition We follow Black and Cox (1976) who assume growing default barrier: K (t) = Kexp(t) where is the growth parameter. The default condition is then: Vj (t) K (t) Since rms are assumed to be homogeneous, Vj (0), j , and j are the same for all j . Thus, we can drop the subscript j . The default barriers are also the (L) same for all rms. The default condition rewritten in term of Xj is:
(L)

(L)

Xj (t) K (t) K (t) a + bt is the default barrier: a= lnK lnV (0) ( ) b=

(L)

Whenever Xj (t) falls below K (t), the rm j is considered default. Thus, the default time of rm j is

(L)

j = inf t 0; Xj (t) K (t)

10

(L)

When the rm values are dependent, the model need modications. The proxy of asset return is assumed to follow a correlated Lvy process with the degree of dependency . Based on our assumptions, the formulas become

Xj (t + dt) = Xj (t) + dXj (t) (L) dXj (t) = dM (t) + 1 dYj (t)

(L)

(L)

(L)

(14)

where M and Yj are independent Lvy processes with M (0) = Yj (0) = 0. The distribution of default rate is translated to the portfolio loss distribution via (4). To construct the default rate distribution, we sample the paths of (L) (L) M and Yj and calculate Xj ; whenever the Xj falls below K , the rm j is considered default after that time. The default rate is counted to create a frequency table representing the default rate distribution. Now, we specialize the correlated Lvy process Xj to a correlated Meixner (M eixner ) process Xj by assuming that the rm's asset value follows Meixner process with parameters of , , ; moreover, we replace the drift parameter with rf , and specify the mean-correcting parameter :
(L)

Vj (t) = V (0)exp (rf )t + Xj

(M eixner )

(15)

cos(/2) = 2 ln cos(( adjusts the rm's asset value so that the discounted + )/2) asset value process becomes a martingale. However, since the markets are incomplete model, the market participants may use diernt formula for valuation. This need not concern us because we imply the parameters from CDS quotes.The default barrier for Meixner process is K (t) = a + btwhere (r f ) lnV (0) and b = . When all parameters are known, we can price a = lnK (M eixner ) CDOs by sampling the paths of each rm j , Xj by sampling the paths of market factor F and the rm specic Yj :

Xj (t + dt) = Xj (t) + dXj (t) dXj (t) = dF (t) + 1 dYj (t)

where F and Yj are independent Meixner processes which F (0) = Yj (0) = 0 and their increments distribute as Meixner(, ,dt,dt tan(/2)).

3.2.4

Risk Measures

Expected Discount Loss The expected discount loss is equal to the value of the default leg in CDO 11

pricing framework i.e.

n

EDL = def ault leg =

i=1

{E [P (ti1 )] E [P (ti )]}erti

where E [P (ti )] is the expected outstanding principal of a CDO tranche at time ti . At initiation the default leg is equal to premium leg. Value-at-Risk Moosbrucker (2006a) computes the VaR of the portfolio under the assumption of innite large homogeneous portfolio; this study relies on the assumption of nite homogeneous portfolio. Suppose Lp is a random variable of the portfolio loss and P (0) is the total value of the portfolio. The maximum potential loss of the portfolio at a condence level (1 )%, VaRp can be calculated as follows

P r(Lp > V aRp ) = V aRp )=1 Fp ( P (0) 1 V aRp = P (0)Fp (1 )

where Fp is the distribution of the portfolio loss rate. Let Pj (0) is the total value of the single tranche j . The credit VaR of the tranche j is

V aRj = Pj (0) min 1, max 0,

KUj

V aRp P (0)

KUj KLj

1 (1 ) KUj Fp V aRj (1 ) = Pj (0) min 1, max 0, KUj KLj

(16)

Expected Shortfall Mashal et al. (2003); Antonov et al. (2005) suggests that the expected shortfall of CDOs can be calculated as:

ESj =

1 1

V aRj (x)dx

where ESj is the expected shortfall or conditional VaR of a single tranche j and is the signicant level.

Results

In this section, we discuss the model calibration and present the parameters estimated from the samples. Then the performances of the Meixner-based mod12

els are examined and compared with those of the traditional models. Finally, the risk measures of the CDOs from each model are presented and discussed.

4.1

Model Calibration

4.1.1

Calibration for Copula Models

There are three elements to be calibrated from the market data: a hazard rate of each rm, the correlation parameter, and copula's parameters. Using the homogeneity assumption, the hazard rate of a rm can be backed out from the CDS index. In this study, we assume that the recovery rate is 40%. Specically, the hazard rate is implied by setting the market quote (s) in the following equation:

s=

n i=1 (ti

n rti i=1 {E [P (ti1 )] E [P (ti )]}e ti +ti1 ti ti1 ti1 )E [P (ti )]erti + n ){E [P (ti1 )] E [P (ti )]}er( 2 ) i=1 ( 2

(1 R)

where E [P (ti )] = e

ti

(17)

The correlation parameter is implied from the equity tranche since, as suggested by Hull, Predescu and White (2005), this tranche is the most sensitive to the correlation parameter. First, we x the distribution parameters; for the double-t copula, and for the Meixner copula. For each day, we solve for in the equation (12) that makes the model price of equity tranche(1), (3), (11), and (12)equal to the market price. We do this for each series and then calculate the mean absolute pricing error. After that, we change the copula parameters and solve for again until the overall errors are minimized.

4.1.2

Calibration for Correlated Structural Models

For the structural models, instead of implying the hazard rate, we back out the default barriers. Then the correlation parameter and the distribution's parameters can be estimated similar to the case of the copula models.

Similar to the copula models, we view the CDS index as a CDS of a representative company underlying the portfolio. However, we must assume the process parameters rst, then imply the default barriers. For correlated-BS model, the parameter a = and b = of the default bar rier K will be estimated by using 5- and 10- year CDS spread and the 13
lnK lnV (0) rf 2
2

equation (17) where E [P (ti )] = 1 p(0, ti ) in which p(0, ti ) is the probability of default from the rst passage time's equation (Harrison; 1990) i.e. +bti bti 4 p(0, ti ) = a . + exp (2ab) a ti ti
lnV (0) f For MS model, the parameters a = lnK and b = of the default barrier K will be estimated similar to correlated-BS model. Unfortunately, we have to estimate two more parameters which are the and of the Meixner process. Moreover, the rst passage time distribution of the Meixner process has no closed-form as in the case before; this makes calibration harder. To estimate a and b, we have to assume the value of and rst and then simulate until we gure out the barrier parameters that make the CDS prices equate the market quotes. N + 1 sets of sample paths are needed where N is the number of underlying rms. Note that in all cases we use time step t = 0.25. In Appendix B, we propose an approximation method to speed up the calibration. (r )

For illustration, considering a day on series 6 with maturity of 5 and 10 years, we sample the paths that follow Meixner process(1, , ) over time horizon of 10 years and then we nd the probability of rst hitting the barrier a + bti by time ti i.e. p(0, ti ). After that the 5Y and 10Y CDS premium are computed using the equation (17); we adjust the barrier parameters (a, b) until these CDS premiums are matched with the market quotes. However for 7YS6, we use the data from 5YS6 and 7YS6 to nd the suitable barrier. As in the case of the copula models, the set of correlation parameter and the process' parameters and are chosen such that the overall mean absolute pricing error is minimized. The parameters estimated are presented in Table 1. The degree of freedom of double-t copula model that minimizes MAPE is 3 for all time series excepts 10YS5 and 10YS6 which are 6. For Meixner copula model, the optimal parameters (, ) vary across the time series: (2.6753,-0.7004 for 5YS6, (1.3724,-1.8162) for 7YS6, and (0.0132,-3.0949) for 10YS6. The parameters for the correlated Meixner structural model, (, ) are (-1.0472, 0.0500) for 5YS6, (-0.8261, 0.0280) for 7YS6, and (-1.3694, 0.0400) for 10YS6. It should be noted that the 's of MC and MS are all negative. Note that the correlated Meixner structural's parameters of 10YS5 and 10YS6 may not be global optimum. Even so, it is sucient for the MS model to be outperform other models.
4

If we need to know the asset volatility and the relative dierence between the default barrier and the asset value K/V (0), we can calculate them given the USD zero curves and the decay parameter . However, in this study it is not necessary to nd these values; only the values of a and b are sucient to price CDOs.
14

Table 1 Summary of parameters calibrated from the market quotes of CDOs on the CDX NA IG index CDX NA IG 5YS5 GC DC MC BS MS

5YS6 NA 3 2.6753 NA 0.0500

7YS5 NA 3 1.6657 -1.9726 NA 0.0268

7YS6 NA 3 1.3724 -1.8162 NA 0.0280

10YS5 NA 6 0.0180 -3.0644 NA 0.0400

10YS6 NA 6 0.0132 -3.0949 NA 0.0400

NA df

3 3.2225 NA 0.0400

-0.6911 -0.7004

-0.6435 -1.0472 -0.7499* -0.8261* -1.3694* -1.3694*

* This results from the approximation approach.

Recall that Gaussian copula has no tail dependence while the double-t copula has symmetric tail dependence, and Meixner copula has an asymmetric dependence structure. Figure 1 illustrates their dependence structures. They are the plots of two 3000-sample variates with dierent degree of dependency = 0.2, 0.5, 0.7, and 0.9. For Gaussian copula, the shape of the bi-variate dependent structure is ellipse. For the double-t copula, it is more scattered in the middle and is also more clustered at both tails. Meixner copula has the shape that is even more scattered in the middle and even more clustered at both tails. Due to the negative , the PDF of Meixner has fatter left tail than the right, and the lower tail of Meixner copula is, in turn, more clustered than the upper tail. For the correlated structural models, both BS and MS are based on the rst passage time. However, MS is more exible than BS because it can adjust for heavier/lighter tails while BS has no fat-tailedness. Unfortunately, unlike the Brownian motion, there is an analytical solution for the probability of rst passage time in in Meixner process. In multi-variate settings, correlated asset returns follows Brownian motion in BS and Meixner process in MS. Figure 2 shows the sample paths of the asset return simulated from the correlated Brownian motion and the correlated Meixner process (-1.0472, 0.0500) with the degree of dependency of 0.2, 0.5, 0.7, and 0.9. In BS, both market factor and the idiosyncratic factor diuse without jumps; whereas those factors of MS are pure jump processes. 15

Figure 1. The dependence structure of Gaussian copula, double-t copula with df=3, and Meixner copula with parameters = 2.6753, = 0.70036.

(a) = 0.2

(b) = 0.5

(c) = 0.9

4.2

Comparison of Models

The Mean Absolute Pricing Error (MAPE) is examined to compare the performance of the proposed models. Specically the MAPE is calculated as:
n i=1

M AP E =

AP Eti n

16

Figure 2. The sample paths of the cumulative asset log return that follows correlated Brownian motion and correlated Meixner process (-1.0472, 0.0500) with dierent degree of dependency.

(a) = 0.2

(b) = 0.5

(c) = 0.9
where

AP Eti =

tranche5 tranche2

|M odelSpreadti M arketSpreadti |,

n is the number of period until maturity

Table 2 reports the MAPEs of each model for all series, from which we can 17

Table 2 The Mean Absolute Pricing Errors (MAPEs) and their standard deviations . CDX NA IG 5YS5 GC DC MC BS MS 102.65 25.53 (17.32) 21.93 (13.76) 84.68 22.12 (10.64) 5YS6 99.35 24.36 (9.91) 13.74 (8.02) 78.84 10.55 (6.34) 7YS5 168.31 59.98 43.76 156.72 53.86* 7YS6 152.61 39.98 25.94 146.60 25.81* 10YS5 200.15 158.47 154.66 255.14 98.53* 10YS6 185.40 158.15 151.11 235.55 72.97*

(33.70) (19.18) (54.17) (26.00) (48.29) (24.01) (31.19) (16.13) (33.22) (43.64) (24.66) (13.00) (31.69) (42.81)

(40.73) (37.36) (47.89) (73.78) (73.07) (44.50) (34.05) (17.13) (58.65) (32.97)

Note: All numbers are in basis point except the t-stats that is in unit of one. * This results from the approximation approach.
draw the following conclusion. For every series, the models based on Meixner distribution (MC, MS) have the lowest MAPE. Model based on double-t distribution provides the next best performance, while the Gaussian based models yield the worst performance. To conrm the performances of the models, we employ a paired Z-test as Houweling and Vorst (2005) 5 . The results reported in Table 3 conrms that for copula models, Meixner copula yields the best tting performance, followed by double-t copula and the Gaussian copula. This suggests that the model taking account of the tail dependence of each asset in the baskets will improve the capability to capture the capital structure of CDOs. A more exible model that can handle the asymmetric tail dependence will further increase the performance. The parameter of MC that minimize the MAPE are negativefor all data samples, conrming that the assets in the portfolio are more correlated in lower tail than in upper tail. We conclude that, for CDO valuation, consid5
A,B Specically, Z-test is expressed as ZA,B = n d sA,B where dA,B,ti = AP EA,ti A,B is the sample mean of dA,B,t , sA,B is the sample standard deviation AP EB,ti , d i of dA,B,ti , and n is the sample size.

18

Table 3 The results of the paired Z-tests. CDX NA IG 5YS5 GC vs MC Mean Di. t-stats DC vs MC Mean Di. t-stats GC vs DC Mean Di. t-stats BS vs MS Mean Di. t-stats 80.72 26.36 3.60 2.51 77.12 41.68 62.57 16.86 5YS6 85.61 37.73 10.63 8.02 74.98 65.01 65.10 36.44 7YS5 124.55 22.62 16.22 6.61 108.33 32.16 94.82 18.91 7YS6 126.68 51.26 14.04 12.45 112.63 74.60 103.80 13.30 10YS5 45.49 (37.24) 13.10 3.81 (9.22) 4.43 41.68 (38.29) 11.67 140.70 14.98 10YS6 34.29 (42.76) 8.06 7.05 (3.17) 22.35 27.25 (43.00) 6.37 128.29 14.89

S.D. of Di. (34.65) (24.12) (59.81) (25.92)

S.D. of Di. (16.22) (14.09) (26.65) (11.83)

S.D. of Di. (20.93) (12.26) (36.59) (15.84)

S.D. of Di. (41.99) (18.99) (54.48) (81.85) (100.72) (86.61)

Note: The number of observations of 5YS5 is 128 while 5YS6's is 113 observations; 7YS5's, 7YS6's, 10YS5's, and 10YS6's are 118, 110, 115, and 101, respectively. All numbers are in basis point except the t-stats that is in unit of one.
ering only the symmetric tail dependence is not sucient since rms tends to default together during the downturns but are more independent during the upturns. For the structural models, the correlated Meixner structural model signicantly outperforms the standard structural model in all cases. For 5YS6, the t-stat is 36.44 while 7YS6's is 13.30, and 10YS6's is 14.89. The t-stats in series 5 for 5Y, 7Y, and 10Y are 16.86, 18.91, and 14.98, respectively. This suggest that the assumption that rm value follows the Brownian motion may not be entirely appropriate. The rm value can have rare large moves governed by jumps in addition to the frequent small moves governed by diusion. This study assumes that the rm value is driven wholly by jumps of innite activity and innite variation which can capture both the rare events and the extremely frequent small moves. The result shows that modeling the asset return as Meixner process is more realistic. Furthermore, the optimal parameters having negative 's for all cases, implying that a rm is more likely to 19

suer from large down-sided jumps than the large up-sided jumps. The large down-sided jumps can be from market factor or rm specic. When it is from market factor, the default probabilities of all rms increase tremendously.
4.3 Risk Measures

This section discusses the risk measures derived from each model: EDL, VaR, ESs of CDX NA IG by assuming that investors hold the CDOs until the expiry. First, the portfolio loss distribution of 5YS6, 7YS6, and 10YS6 implied from each model are investigated. Then, we discuss the risk measures of each CDO tranche at condence level of 95%. It is important to bear in mind that we imply the risk measures under risk-neutral world; the results may be dierent in the real world. Figure 3 reveals the portfolio loss distribution of CDX NA IG 5YS6, 7YS6, and 10YS6 implied from the market quotes on the issued date (23 Mar 2006) for each model. Table 4 represents the risk measures for all data samples. Recall that these risk measures have dierent interpretations. Expected Discount Loss (EDL) is the expected loss derived from each model while VaR is the potential loss in extreme case at a condence level of 95%. In contrast, ES is the average of the potential losses given that an extreme event occurs at the same condence level. In addition, the undiscounted total premium (UTP) is an approximate indicating actually how much investors get compensated totally. Therefore, VaR/EDL indicates how much the loss in extreme case is relative to the expected loss. For ES/EDL, it indicates how much the average potential loss given an extreme case is, compared to the expected loss. Similarly, VaR/UTP and ES/UTP indicate how much the loss in extreme case and the average loss given extreme case are relative to what market views. The VaR and ES are relatively high in some tranches when comparing with EDL and the market premiums (Undiscounted Total Premium). This is interesting because EDL is what we expect in average and it indicates directly how much investors should fairly get in compensation, but in extreme case, the investors can end up with a huge loss than what is expected. We will discuss the risk measures from 5YS6; however, the other series have only slightly dierent interpretations. For the equity tranches (0-3%), it is quite certain that in extreme cases the investors holding this tranche loss all of their investments. Both VaR and ES are equal to 1 in almost all of the cases. Moreover, the risk measures computed from dierent models provide approximately the same in this tranche. The investors holding 5YS6 equity tranches can loss, in extreme cases, about twice 20

of the expected loss while it is around 1.8 times of what market views. For junior mezzanine tranches (3-7%), investors are exposure to large losses in extreme cases comparing to what they expect to get in average. For 5YS6 as an example, investors can loss, in extreme cases, around 7-8 times of the expected loss while about 5-14 times of what market views. Given that the extreme cases occur, the loss is approximately 12-18 times of the expected loss, whereas it is about 14-20 times of what market views. For this tranche, each of the models provides slightly dierent risk measures except GC; it is the most pessimistic. It is interesting that although tranche 0-3% can worthless in extreme case, the premium is quite high. Even though the loss of this tranche is less than that of tranche 0-3% in extreme cases, the premium is even less. For tranche 7-10%, GC and DC provide similar higher risk measures; whereas, MC's and MS's are around the same; BS has much less risk measures. Though, the losses in extreme case are zeros at condence level of 95%, the average potential loss given the extreme events is about 24-26 times of the expected loss. Moreover, it is about 11 times of what market views for BS but 20-28 times for the other models. Notice that BS's risk measures start to fall sharply because the model neglects extreme cases. For tranche 10-15% and 15-30% of 5YS6, DC provides the highest risk measures, while GC has less risk measures and BS has much less ones. Given extreme events occur, the average potential loss of holding the two tranches can be 24-26 times of the expected loss. For BS and GC, it is around 1.7-9 times of what market views for the rst tranche and only 0.0330-0.2317 times for the later tranche. Meanwhile, it is 18-35 times for MC, MS, and DC for tranche 10-15% and it is around 12-14 times for MC and MS and 33 times for DC in tranche 15-30%. Notice that GC and BS fall sharply; the probable reason is they ignore the extreme cases. Risk measures derived from 5YS5 is very similar to the case of 5YS6. Moreover, the risk measures for 7YS5, 7YS6, 10YS5, and 10YS6 reveal consistency with the interpretation of 5YS6. Firstly, BS and GC neglect the extreme cases so that their risk measures are higher in earlier tranches and then plummet later. Secondly, DC provides relatively high risk measures in most cases. Finally, both MC's and MS's risk measures tend to be less than DC's in most cases. However, the noticeable dierences among dierent series are: in case of 5Y, only tranche 0-3% is wiped out in extreme cases; for 7Y, the rst two tranches are wiped out much likely in rare cases; lastly, the rst three tranches of 10Y are worthless in extreme events quite surely, and rather surely for the fourth tranche. 21

Figure 3. The portfolio loss distribution of CDX NA IG Series 6

(a)

(b)

(c)
22

Table 4: The risk measures for various Tranches of CDX NA IG 5YS6

3%-7% RM 0.0090 0.0452 0.0727 0.0402 0.0369 0.0618 0.0355 0.6182 0.2905 0.2604 0.4915 0.2497 0.8985 0.7200 0.6733 0.7837 0.6477 12.681 18.247 18.231 17.913 15.946 14.912 17.356 14.343 12.355 19.899 0.2692 0.2819 0.2099 0.1081 0.1908 7.0348 5.5297 0.0000 7.9527 10.884 0.0000 0.0000 0.0000 24.706 23.869 24.370 25.791 24.895 7.0514 5.7675 0.0000 0.0000 7.2272 6.4334 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 26.877 28.136 20.949 10.787 19.051 8.5004 13.690 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0439 0.1659 0.0993 0.0084 0.0855 NA 0.7860 0.0077 NA 0.7653 0.0035 NA 1.3686 0.0042 NA 0.4182 0.0003 NA NA 0.0000 0.0000 0.0000 0.0000 0.0000 24.945 23.685 24.396 25.799 24.692 NA 0.8179 0.0086 NA 0.8596 0.0041 NA NA 0.8902 0.0118 NA 1.1788 0.0070 NA 1.4933 0.8675 0.0697 0.7383 0.0000 0.0000 0.0000 0.0000 0.0000 9.3570 35.369 21.165 1.7982 18.229 NA 1.6106 0.0109 NA 1.0879 0.0018 NA 0.3751 0.0001 0.0036 0.0015 0.0000 0.0013 0.0000 0.0000 0.0000 0.0000 0.0000 0.0015 0.0843 0.0363 0.0002 0.0321 NA NA 0.0100 NA NA 0.0047 NA NA 0.0025 NA NA 0.0020 NA NA 0.0009 NA NA 0.0005 NA NA NA NA NA NA NA 0.0000 0.0000 0.0000 0.0000 0.0000 25.178 23.499 24.513 25.988 24.257 RM/EDL RM/UTP RM RM/EDL RM/UTP RM RM/EDL RM/UTP RM RM/EDL 7%-10% 10%-15% 15%-30% RM/UTP NA NA 0.0234 1.4175 0.5846 0.0038 0.5220 0.0000 0.0000 0.0000 0.0000 0.0000 0.5883 33.310 14.330 0.0975 12.663

0%-3%

RM NA NA 0.8259 0.8285 0.8326 0.8672 0.8470 1.8282 1.8282 1.8282 1.8282 1.8282 1.8282 1.8282 1.8282 1.8282 1.8282

RM/EDL RM/UTP

Mean Mkt. Quotes

0.2970

NA

UTP

0.5470

NA

GC

0.4518

NA

DC

0.4532

NA

EDL

MC

0.4554

NA

BS

0.4744

NA

MS

0.4633

NA

23

GC

1.0000

2.2135

DC

1.0000

2.2065

VaR

MC

1.0000

2.1959

BS

1.0000

2.1081

MS

1.0000

2.1583

GC

1.0000

2.2135

DC

1.0000

2.2065

ES

MC

1.0000

2.1959

BS

1.0000

2.1081

MS

1.0000

2.1583

Note: RM is referred risk measures, EDL is expected discount loss, and UTP means undiscount total premium.

Conclusion

This paper provides a framework to price CDOs based on Meixner processes. Meixner processes have properties such as fat-tail, skewness, and jump component andare relatively simpler to implement than other Lvy processes. It is shown that the Meixner processes can be applied to both copula and structural form approaches. Meixner distribution can be used to construct the Meixner copula which has asymmetric tail dependence, whereas the asset returns in the correlated structural model are assumed to follow correlated Meixner process. Using the historical prices of CDOs on the CDX NA IG, the performance of the proposed models are examined and compared to those of standard models such as Gaussian copula model, double-t copula model, and correlated Brownian motion structural model. The results show that the Meixner-based models outperform the standard models in terms of the mean absolute pricing errors . For example, for series 5YS6, the mean absolute pricing errors of MC is 13.74 bps, which is 10.62 bps less than that of DC and 85.61 bps less than that of GC. Similarly, the pricing error of the MS model is 10.55, which is 68.29 bps less than that of BS. Hence, the asymmetric dependence structure can improve the performance. Based on the calibration results, the skew parameters ( ) of Meixner copula model are negative for all 6 cases which means that the dependence structure is more intense in lower tail than upper tail. Correlated Meixner structural models also have negative in all 6 cases, suggesting that a rm value has more likelihood of large downside jumps than large upside jumps. This is reasonable because the default seems more correlated during the downturns and less correlated in normal situations. We also investigate the risk measurements based on the proposed models. In general it is shown that DC considerably provides higher risk measures in most cases; whereas, GC and BS claim higher likelihood of losses in the earlier tranche and fall sharply afterward. MC's and MS's provide risk measures that tend to be less than DC's in most cases. Thus, including tail dependence and/or fat tail provides more realistic risk measures. In summary, in addition to the fat tail, skewness is also an important property which should also be considered when valuing the credit dependent securities. Imposing a negative skewness in the distribution of market factor provides more dependence in lower tail of its associated copula and also more likelihood of common large downside jumps in its associated process. These extensions are more realistic and improve the performance of CDO pricing models. Moreover, the risk measures have more validity because they are based on the portfolio distribution calibrated from real-world market quotes.

24

References
Albrecher, H., A. Ladoucette, S. and Schoutens, W. (2006). A generic onefactor lvy model for pricing synthetic CDOs. Working Paper. Antonov, A., Mechkov, S. and Misirpashaev, T. (2005). Analytical techniques for synthetic CDOs and credit default risk measures. Research Paper, NumeriX. Baxter, M. (2006). Lvy simple structural models. Nomura Fixed Income Quant Group. Black, F. and Cox, J. (1976). Valuing corporate securities: Some eects of bond indenture provisions, Journal of Finance 31: 351367. Black, F. and Scholes, M. (1973). The pricing of options and corporate liabilities, Journal of Political Economy 81: 637654. Brunlid, H. (2006). A comparative analysis of hyperbolic copulas induced by a one factor lvy model. Working Paper. Burtschell, X., Gregory, J. and Laurent, J.-P. (2005). A comparative analysis of CDO pricing models. Elizalde, A. (2005). Credit risk models IV: Understanding and pricing CDOs, www.abelelizalde.com. Galiani, S. S. (2003). Master's thesis, Department of Mathematics, King's College, London. Gupton, G. M., Finger, C. C. and Bhatia, M. (1997). Creditmetrics - technical document. Harrison, J. M. (1990). Brownian Motion and Stochastic Flow Systems, Krieger Publishing Company. Houweling, P. and Vorst, T. (2005). Pricing default swaps: Empirical evidence, Journal of International Money and Finance 24: 12001225. Hull, J., Predescu, M. and White, A. (2005). The valuation of correlationdependent credit derivatives using a structural model. Working Paper, Joseph L. Rotman School of Management, University of Toronto. Hull, J. and White, A. (2001). Valuing credit default swaps II: Modeling default correlations, Journal of Derivatives 8(3): 1222. Hull, J. and White, A. (2004). Valuation of a CDO and nth to default CDS without Monte Carlo simulation, Journal of Derivatives 12(2): 823. Hull, J. and White, A. (2006b). Valuing credit derivatives using an implied copula approach. Working Paper, Joseph L. Rotman School of Management, University of Toronto. Kalemanova, A., Schmid, B. and Werner, R. (2005). The normal inverse gaussian distribution for synthetic CDO pricing, Technical report. Laurent, J.-P. and Gregory, J. (2003). Basket default swaps, CDO's and factor copulas. Working Paper ISFA Actuarial School, University of Lyon and BNP Paribas. Li, D. X. (2000). On default correlation: a copula approach, Journal of Fixed Income 9: 4354. 25

Luciano, E. and Schoutens, W. (2005). A multivariate jump-driven nancial asset model. UCS Report 2005-02, K.U. Leuven. Mashal, R., Naldi, M. and Zeevi, A. (2003). Extreme events and multi-name credit derivatives. Working Papers, Lehman Brothers Inc. and Columbia University. Moosbrucker, T. (2006a). Copulas from intnitely divisible distributions: Applications to credit value at risk. Working Paper, Graduate School of Risk Management, University of Cologne. Moosbrucker, T. (2006b). Pricing CDOs with Correlated Variance Gamma distributions. Working Paper, Department of Banking, University of Cologne. Schnbucher, P. J. (2000). Factor models for portfolio credit risk. Working Paper, Department of Statistics, Bonn University. Schoutens, W. (2002). Meixner processes: Theory and applications in nance, EURANDOM Report 2002-04, EURANDOM, Eindhoven. Schoutens, W. (2003). Lvy Processes in Finance - Pricing Financial Derivatives, Wiley Series in Probability and Statistics, John Wiley & Sons, Ltd. Schoutens, W. and Teugels, J. (1998). Lvy processes, polynomials and martingales, Commun. Statist.- Stochastic Models 14(1 and 2): 335349. Varma, P. and Cantor, R. (2005). Determinants of recovery rates on defaulted bonds and loans for North American corporate issuers: 1983-2003, Journal of Fixed Income 14(4): 2944.

Meixner Distribution/Process

Meixner distribution is a special case of Lvy distribution which is innitely divisible. That means for each positive integer n, the Meixner characteristic function (u) is the n -th power of a Meixner characteristic function i.e. (u; , , , ) = [(u; , , /n, /n)]n . Hence, it is stable under convolution i.e. the distribution function of the sum of independent Meixner random variables is still the Meixner distribution; see Schoutens (2003). Meixner distribution is one of few Lvy distribution that has an analytical probability density function (PDF). The PDF of Meixner distribution is

(x ) (2cos(/2))2 exp fM eixner (x; , , , ) = 2 (2 )

where > 0, < < , > 0 and R.

i(x ) +

, xR

The moments of this distribution is shown in table A.1. The Meixner distribu2 /2) ) and = sin( . tion has zero mean and unit variance when the = 2 cos( 2 Now, we have two parameters left which are that is mainly used to control kurtosis or fat-tail, and mainly used to control skewness. However, changing 26

one parameter aects both kurtosis and skewness as in table A.2.

Table A.1 The moments of the Meixner distribution Meixner(, , ,) mean variance skewness kurtosis
+ tan(/2)
1 2 2 2 (cos (/2))

sin(/2) 2/ 3 + (2 cos( ))/

Since Meixner distribution is innitely divisible, we can nd its associated Lvy process, called Meixner process. Strictly speaking, a Meixner process (M eixner ) (M eixner ) , t 0} is a stochastic process in which X0 = X (M eixner) = {Xt 0, and it has an independent and stationary increments over [t, t + t] i.e. (M eixner ) (M eixner ) Xt+t Xt distributes as Meixner(, , t,t). The distribution (M eixner ) of Xt follows Meixner(, , t,t).

Table A.2 The moments of the zero-mean and unit-variance Meixner distribution
/2) ) Meixner(, , 2 cos( , sin( 2 )
2

mean variance skewness kurtosis

0 1 tan(/2)
2 cos (/2) 3 + ( 3 )2 2 cos2 (/2)
2

Each Lvy process is uniquely characterized by a triplet [, 2 , (dx)]. They correspond to: a deterministic part, a Brownian part, and a jump part. (dx) is called a Lvy measure. A Meixner processhas no Brownian part, thereby being a pure jump process. Moreover, the rst parameter in the Lvy triplet is sinh (x/) = tan (/2) 2 dx sinh (x/) 1 and the Lvy measure is

(dx) =

exp (x/) dx xsinh (x/)

For more details on the Meixner distribution and process, see Schoutens (2003); 27

Albrecher et al. (2006); Schoutens (2002).

An Approximation Approach for Correlated Structural Models

We propose an approximation to speed up the simulation. Note that this is only a conjecture. However, when we compare the CDO pricing performance of the approximation approach to the traditional one, the results in terms of mean absolute pricing error (MAPE) look promising. The verication in terms of proof is in progress. Instead of using the correlated simulation approach as mentioned in the paper, the Bernoulli mixture framework can be applied when some restrictions are imposed. If we assume that the path of the market factor M is given so that the rm values are conditionally independent, the conditional default probability of a rm before time t under rst passage time approach is

p(t|M ) = P r t | inf X ( ) K ( )

1 Z ( ) a + b a + b M ( ) = P r t | inf Z ( ) 1 = P r t | inf

M ( ) +

Therefore, the probability that k out of N rms default before or at time t is as follows

P r(N, k, t|M ) =

p(t|M )
k

1 p(t|M )

N k

Hence, the portfolio loss distribution function is

zN/(1R) k=1

Ft (z ) = EM

N p(t|M )

1 p(t|M )

N k

where EM represents the expectation over all possible paths of the market factor. One can nd the portfolio loss distribution function by rst sampling a number of paths of the market factor and then calculate that expected portfolio loss distribution function under all the possible paths. Alternatively, if the end point of the market factor is given, the conditional portfolio loss distribution can be calculated in the expectation over all possible paths starting from 0 to that given end point. Then, we nd the unconditional one by integrating over all the possible end points. 28

To approximate, we replace the paths starting from 0 to any given points in time t with its expectation (as a line). Specically, for time horizon t, the market factor at the time s before t is M (s) = ms ; s t. Hence, the proxy of asset return becomes

Xj (s) =

ms +

1 Zj (s) ;

st

and the conditional default probability under time t is

b m a + p(t|m) = P r t | inf Z ( ) 1 1

As a consequence, the portfolio loss distribution under time t is

m=+ zN/(1R)

Fti (z ) =

m=

N k N k ti (m) dm pti |m (1 pti |m )

k=1

pti |m

b m a = P r t | inf Z ( ) + 1 1

(B.1)

where the probability distribution of m depends on 1 , i.e. mN (0, 1 ) in case of t t 1 Brownian motion and mM eixner t , , t for Meixner case. Mathematically, one may view these equations as a copula given that there exists an inverse function G1 (pt ) that can reproduce the barrier paramters a and b under time horizon t. By choosing this copula, which is supported by an approximation of correlated structural model, it has more economic sense than by choosing an arbitrary copula. Although, there is no proof yet, the results are acceptable. Particularly for Meixner model, recall that the dynamic of the rm value follows an exponential of Meixner process M (, , ) i.e.

Vi (t) = Vi (0)exp (rf i )t + Xi (t) lnVi (t) lnVi (0) (rf i ) + t Xi (t) =
(M eixner )

(M eixner )

(B.2) (B.3)

where Xj (t) is a Meixner process with parameter (1, , ) and = 2 ( /2) cos(i /2) cos(i /2) i 2i ln cos((i +i )/2) = 4 cos . The term corrects the rm's ln cos(( 2 i +i )/2) i asset value so that the discounted asset value process becomes a martingale. However, since most Lvy models are incomplete model, the market participants may not view as this equivalent measure. Actually, we do not need to know how market participants view since we imply the parameter from CDS quotes which is itself risk-neutralized and reect the view of the market. Under 29

the approximation approach, the Meixner model becomes

m=+ zN/(1R)

Fti (z ) =

m=

N k N k M eixner pti |m (1 pti |m )

m;

k=1

1 , , ti dm ti
(B.4)
a 1

pti |m = P r t | inf Z ( ) a + b
where Z is a Meixner process with parameters of (1, , ), a =
b m . 1

and

b =

Here is the comparison of the traditional structural model (Hull et al.; 2005) and our approximation approach. Note that 20,000 sets of sample paths are used, whereas the approximation approach is semi-analytic.

Table B.1 The comparisons of simulation and approximation approaches in Brownian case 5YS5 Simulation BS MAPE S.D. Approximation BS MAPE S.D. 84.68 63.32 5YS6 78.84 50.62 7YS5 156.72 141.93 7YS6 146.60 121.21 10YS5 255.14 231.50 10YS6 235.55 206.82

(40.73) (37.36) (47.89) (73.78) (73.07) (44.50) (33.80) (13.41) (52.32) (21.39) (79.45) (50.69)

30