copulas from infinitely divisible distributions-applications to credit value at risk
TRANSCRIPT
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
1/28
Copulas from Infinitely Divisible Distributions:
Applications to Credit Value at Risk
Thomas Moosbrucker
Graduate School of Risk ManagementUniversity of Cologne
Albertus-Magnus-Platz50923 Koln, Germany
Phone: +49 (0)221 470 3967Fax: +49 (0)221 470 7700
First Draft: March, 2006This Draft: June, 2006
JEL classification: G13
Keywords: Credit Value at Risk, Credit Portfolio Models, LHP approximation,
Infinitely Divisible Distributions, Levy Processes
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
2/28
Copulas from Infinitely Divisible Distributions:Applications to Credit Value at Risk
Abstract
This article proposes infinitely divisible distributions for credit portfolio modelling. First,
we show that these distributions are well suited to fit into a one factor copula approach. We
compare four different specifications (Normal Inverse Gaussian, Variance Gamma, Merton
Jump Diffusion and Kou Jump Diffusion), where the last two distributions are new to credit
portfolio modelling. The comparison in this article is done with respect to Value at Risk(VaR) measures. We find that the more extreme tail behaviour of the infinitely divisible
distributions leads to significantly different VaR measures. Therefore, we conclude that the
identification of the correct dependence structure is of major importance for the estimation
of VaR.
JEL classification: G13
Keywords: Credit Value at Risk, Credit Portfolio Models, LHP approximation,
Infinitely Divisible Distributions, Levy Processes
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
3/28
1 Introduction
Besides single-name default probabilities, the dependence structure of defaults is of major
importance for credit portfolio risk management. The most important industry models
such as KMV and CreditMetrics assume that a Gaussian copula can describe these de-
pendencies. Furthermore, the Internal Ratings based approach of the latest Basel accord
relies on this assumption. The Gaussian model traces back to the work of Merton (1974),
who assumes that asset value processes are driven by Brownian motions.
One of the main advantages of the Gaussian assumption is that these models are
easy to implement and that they lead to fast computations. Vasicek (1987) derived an
analytical formula for the loss distribution of an infinitely large portfolio called the largehomogeneous portfolio (LHP) approximation. This allows practitioners to implement an
oversimplified but fast version before setting up a more sophisticated credit risk model.
Furthermore, a single parameter (correlation) determines the entire dependence structure
between two defaults. This number is easy to communicate.
In practice, there are several ways to estimate default correlations. One method is based
on historical default time series. However, more often than not, very little data is available.
Another method to estimate default correlations consists of using asset and equity returncorrelations as proxying variables. In this case, usually, more data is available. However,
log returns are not normally distributed. They are leptokurtic and often skewed. Thus,
the estimation of a Gaussian correlation parameter is inappropriate.
In the market of equity options, it is well-known that the assumption of normally
distributed log returns in the Black-Scholes model leads to a volatility smile. The implied
volatility of a low strike price option is higher than that of a high strike price option.
Since the work of Black and Scholes, researchers have developed several models to cover
the stylized facts of log returns for the valuation of options. In particular, models based
on Levy processes have proven to be successful. These models have a significantly better
fit to equity options than the approach via Brownian motions.
The volatility smile in equity options has its counterpart in credit derivatives. In liquid
CDS index tranches, there is a pronounced correlation smile when a Gaussian copula model
is assumed. This shows that market participants do not fully believe in the Gaussian copula
model. Instead, the risk-neutral distribution exhibits stronger extremal dependence than
1
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
4/28
that captured by a Gaussian copula.
In this paper, we consider factor copulas resulting from infinitely divisible distributions
for credit risk management. These distributions arise from the before mentioned Levy
processes. Namely, we use Variance Gamma (VG) and Normal Inverse Gaussian (NIG)
distributions and distributions of the jump-diffusion models by Merton (1976) and Kou
(2002) (MJD and KJD). While NIG and VG factor copulas have been used for the valuation
of Collateralized Debt Obligations (CDOs) in a risk-neutral setting in Kalemanova et al.
(2005) and Moosbrucker (2006), the other distributions have not been used in a factor
copula approach for credit portfolios yet.
In our setting, infinitely divisible distributions have the following advantages: First,
their defining property of infinite divisibility makes them ideally suited for a factor ap-
proach. As a consequence, the resulting formulas for the distribution of the number of
defaults and the LHP approximation are simple. The original LHP approximation for the
Gaussian copula by Vasicek (1987) is a special case. Second, all copulas proposed in this
article contain the Gaussian copula as a special or limiting case. They may therefore be
seen as natural extensions. Third, these distributions are not limited to fitting to stylized
facts of log returns and lead to a flattening of the volatility smile in equity options. Fur-
thermore the VG and NIG copulas have proven to explain the correlation smile in CDS
index tranches (see Kalemanova et al. (2005) and Moosbrucker (2006)).
In this article, we investigate the consequences of the use of the alternative copulas for
Value at Risk (VaR) measures. We apply three different methods. First, we calculate VaR
measures for the different copulas using the same correlation parameters for the default
triggering variables. This method is similar to the one used by Frey and McNeil (2001)
in order to compare the Gaussian and the student t copula. Since identical correlations
of default triggering variables in different models result in different correlations of default
events, we also calibrate all copulas such that the default correlations are matched. This
second method leads to a more realistic comparison of resulting VaR measures. The
approach was proposed by Schonbucher (2002) and is also used by Schloegl and OKane
(2005). For a comparison between the Gaussian and the student t copula, Schloegl and
OKane (2005) found that there are only minor differences in VaR measures. Finally, we
apply the method used by Hamerle and Rosch (2005). In this case, we simulate default time
series based on one of the infinitely divisible copulas and estimate correlation parameters
2
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
5/28
and default probabilities under the assumptions of the Gaussian copula model. Again,
there do not seem to be large differences in VaR measures between the Gaussian and the
student t copula. However, the results are different for the other one factor copulas.
The structure of this article is as follows. The next section presents the distributions
and the corresponding factor copulas. We derive the number of defaults distribution and
the LHP approximation. Section 3 gives the numerical results. Section 4 concludes the
article.
2 Infinitely divisible distributions and factor copulas
2.1 Infinitely divisible distributions
A probability distribution F is called infinitely divisible if for every n N there are n i.i.d.random variables such that their sum has distribution F.1 Infinitely divisible distributions
are in a one-one correspondence to Levy processes: If (Xt)t0 is a Levy process, then for
every t, Xt has an infinitely divisible distribution. Conversely, if F is an infinitely divisible
distribution then there is a Levy process (Xt)t0 such that the distribution of X1 is given
by F.An example of an infinitely divisible distribution is the normal distribution which
corresponds to Brownian motion. Since the distributions we use in this article are better
known through their process counterparts in the literature of option pricing, we introduce
the corresponding processes here. The distributional properties are given in the Appendix.
The first class of Levy processes we consider consists of time-changed Brownian mo-
tions. In this case, a Brownian motion (Wt)t0 is subordinated by an increasing Levy
process (Gt)t0, that isXt = t + Gt + WGt.
Throughout this article, we set = = 0 and we restrict the distributions to unit
variances.
If (Gt)t0 is an inverse Gaussian process, then (Xt)t0 is called a Normal Inverse Gaus-
sian process. This process was proposed by Barndorff-Nielssen (1997). Madan, Carr
and Chang (1998) model (Gt)t0 as a Gamma process. (Xt)t
0 is then called a Variance
1See for example Cont and Tankov (2003).
3
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
6/28
Gamma process. We denote the variance of the subordinating process by . For 0,the limiting process is a Brownian motion.
The jump diffusion processes considered in the papers of Merton (1976) and Kou (2002)
can be written as
Xt = t + Wt +Nti=1
Yt.
In this equation, (Nt)t0 is a Poisson process with intensity and Yi is the distribution
of jumps. In the Merton (MJD) case, Yt is normally distributed, while Kou (KJD) uses
exponential distributions for the upward and downward jumps. In this article, we restrict
both the jump distributions Y1 and the distribution ofX1 to zero mean and unit variance.
We set = 0 as before. Thus, the jump intensity is the only additional parameter in
these models. For = 0, we see that the processes contain Brownian motion as a special
case.
2.2 Factor copulas
Let n N be the number of entities in a portfolio and for 0 < < 1 let X = (X1, . . . , X n)be a random vector where Xi =
M +
1 Zi for every i {1, . . . , n} and M and
all Zi are mutually independent random variables with zero mean and unit variance. We
then call the copula function of X a one factor copula.
The most important one factor copula is the Gaussian copula, where the factors M
and Zi (and thus Xi) are standard normally distributed. The Gaussian copula underlies
the approaches of KMV, CreditMetrics and the latest Basel accords. A good reference for
this model is Li (2000).
If M = M kX
and Zi = Zi kX
, where M and Zi = Zi are standard normallydistributed and X 2(k), then the resulting copula is a student t copula with k degreesof freedom.
In this article, we propose infinitely divisible distributions for factor copulas. The prop-
erty of infinite divisibility makes these distributions ideally suited for a factor approach:
if the parameters for the factors M and Zi are chosen suitably, then the distribution of
the default triggering variable Xi is of the same class. This property is shown in the
Appendix for VG, NIG, MJD and KJD distributions. Since the density functions of these
distributions are available (at least approximately) in closed form, this leads to simple and
4
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
7/28
fast computations. Furthermore, the parameters have a clear economic interpretation.
2.3 Number of Defaults Distribution
As for the Gaussian copula, the number of defaults distribution is obtained by conditioning
on the common factor M. In the following, we assume that the portfolio is homogeneous
with respect to default probabilities. If the unconditional default probability of entity i is
given by p = P(Xi < C), then the default probability conditional on M = m is
p(m) = P(Xi < C|M = m) = P
Zi z) .
From this equation, the VaR of the loss distributions can be calculated in closed form.
It is given by
VaRz(L) = FZ
F1X .(p)
q1z(M)
1
,
where q1z(M) is the (1 z)-quantile of the systematic factor M.
3 Numerical Results
3.1 Identical correlation for default triggering variables
In this subsection, we compare the factor copulas with respect to VaR measures when the
marginal default probabilities and the correlations of the default triggering variables are
identical. In the Gaussian case, these inputs determine VaR measures. The other copulas
have additional parameters. We use the third and fourth moments (skewness and kurtosis)
to appoint these additional parameters.
All distributions in this article have zero skewness. For kurtosis, we consider two cases:
an excess kurtosis of 13
(T1, VG1, NIG1, MJD1, KJD1) and an excess kurtosis of 1 (T2,
VG2, NIG2, MJD2, KJD2). The higher excess kurtosis corresponds to a more extreme
scenario which differs more from the Gaussian model and exhibits larger tail dependence.
Table 1 summarizes the parameters used in this article.2Schloegl and OKane (2005) derive the cdf of this mixing variable.
6
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
9/28
Insert Table 1 about here.
Throughout this study, we consider four different credit portofolios, two medium-risk
portfolios with marginal default probabilities of 1% and two high-risk portfolios with
marginal default probabilities of 7.5%. In both cases, we consider a case with low and a
case with high correlation (20% and 40%). We calculate all VaR measures by the LHP
formula.
Insert Table 2 and 3 and Figure 1 and 2 about here.
Table 2 and 3 and Figure 1 and 2 show the ratio of the VaR given by the alternative
copulas to that given by the Gaussian copula. Therefore, the specified model leads to a
higher VaR measure than the Gaussian model if and only if this ratio is larger than 1.
Typically, the student t copula results in higher VaR measures than the Gaussian copula
the only exceptions are low marginal distributions and VaR quantiles below 95%. The
other distributions show a different pattern: while VaR measures of the specified models
are lower for quantiles below 99%, the ratio to the Gaussian VaR increases dramatically
above 99%. In general, the use of copulas from infinitly divsible distributions therefore puts
more weight on extreme scenarios. Medium-size losses are less likely, whereas very large
losses are more likely to occur. This effect is larger for low marginal default probabilities
and low correlations. As Schloegl and OKane (2005) point out, the tail dependence
does manifest itself more in this case. Furthermore, this effect is more pronounced if the
distributions exhibit heavier tails. If we compare the different copulas, we find that the
NIG and VG copulas behave quite similarly. The difference to the Gaussian copula is
more pronounced for the MJD and especially the KJD copula.
3.2 Identical correlation of default events
Instead of using identical correlations for the default triggering variables, it may seem
more appropriate to compare the different models with identical default correlations. This
correlation is defined as the correlation of the Bernoulli distributed variables 1X1
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
10/28
p12 = P(X1 < C X2 < C), the default correlation is
de =p12 p1p2
p1(1 p1)p2(1 p2)
.
We calculate default correlations resulting from the Gaussian copula approach and
determine the correlations of the default triggering variables for the alternative factor
copulas. Schloegl and OKane (2005) use this method for the student t copula. The
approach was originally proposed by Schonbucher (2002). The results of this calibration
are given in Table 4.
Insert Table 4 about here.
In almost all cases, the correlation of default triggering variables is smaller than in
the Gaussian model. The only exception is the MJD copula for lower correlations. Since
the VaR is increasing in correlation, the ratio of Gaussian and alternative VaR should
therefore decrease in comparison to section 3.1. We calculate the VaR ratios as in the
previous subsection, this time based on the calibration results of Table 4. The results are
given in Table 5 and 6 and Figure 3 and 4.
Insert Table 5 and 6 and Figure 3 and 4 about here.
In line with Schloegl and OKane (2005), we find that the resulting VaR measures
are quite similar for the Gaussian and student t copulas. The ratio is between 0.982 and
1.100 in all cases. In contrast, the difference to the Gaussian copula is not wiped our forthe other copulas we have considered. Compared to the study in section 3.1, VaR ratios
decrease, but there are still significant differences between the Gaussian and the other
copulas. In particular, at the 99.9% level, all but two VaR ratios are larger than 1.
3.3 Model Risk in the dependence structure
In this subsection, we simulate a default time series based on the different copulas and es-
timate marginal default probabilities and correlation based on the incorrect assumption of
8
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
11/28
a Gaussian copula. This is the method used in Hamerle and Rosch (2005) for a comparison
between the Gaussian and the student t copula.
We use the same sample portfolios and parameter sets as in the previous subsections.
As in subsection 3.1, we use correlations of default triggering variables of 20% and 40%.
We now assume that the portfolio consists of 1000 entities. For each portfolio, we simulate
1000 default time series for each copula. As in Hamerle and Rosch (2005), we set the time
series to length T = 10.
For estimation, we employ the maximum likelihood framework as in Hamerle and Rosch
(2005). In all cases except the student t copula, the log-likelihood function is
l(, ) =
T
t=1 ln
ndtp(m)dt(1 p(m))ndtdFM ,while for the student t copula it is
l(, ) =Tt=1
ln
0
n
dt
p(m)dt(1 p(m))ndtdd2(k)
.
For each model and each portfolio, we simulate a default time series and estimate the
parameters and under the (false) assumption of a Gaussian copula. Table 7 gives the
average parameters of this estimation. We find that default probabilities can be estimated
very well: results are between 0.990% and 1.142% for a true default probability of 1% and
between 7.404% and 7.631% for a true default probability of 7.5%. This is not a surprise
since the estimation of the default probability is based on the relative number of defaults.
This number does not depend on the correlation structure.
When analyzing correlation, the results are different. For the student t copula, the
Gaussian estimates are higher than the true correlation. The only exception is a default
probability of 7.5% and a correlation of 40%, where the real correlation and the Gaussianestimate are nearly identical. This finding is in line with both the result of Hamerle
and Rosch (2005) and the results of section 3.2. One might be tempted to conclude
that the estimation procedure in this subsection chooses parameters to match the default
correlation. However, when we compare the other models, we see that this is not true.
Instead of overestimating correlation (as one might expect form the results of section 3.2),
we find lower correlation under the misspecified Gaussian model. This behaviour may be
explained as follows: typically, the extreme tail of the distribution (quantiles above 99%)does not manifest itself in a simulated default time series of length T = 10. The Gaussian
9
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
12/28
correlation estimation is therefore based on lower quantiles of the distribution. By the
results of the previous subsections, this Gaussian estimation is therefore lower than the
true parameter value.
Insert Table 7 about here.
As in the previous subsections, we calculate VaR ratios of the true and misspecified
model. The results are given in Table 8 and 9 and Figures 5 and 6. As in the previous
subsection, we see that the differences between the student t and the Gaussian copula are
wiped out. From this analysis, Hamerle and Rosch (2005) conclude that there is not muchmodel risk in the correlation structure. However, our findings for the alternative copulas
show that this is not true. The different copula assumptions lead to significantly different
VaR measures.
Insert Table 8 and 9 and Figure 5 and 6 about here.
4 Conclusion
In this article, we have compared six different one factor copulas with respect to Value at
Risk. These copulas arise, if the normal, student t, and infinitely divisible distributions
are implemented into a one factor copula. From our analysis, we draw the following con-
clusions: First, there is substantial model risk in the dependence structure. While the
Gaussian and the student t copula lead to very similar VaR measures (if default corre-lations are identical), these figures are very different for the NIG, VG, MJD and KJD
copulas. Overall, these copulas result in higher risk measures for quantiles above 99%.
Second, when marginal default probabilities and correlation are estimated based on his-
torical default series, correlation is underestimated under the assumption of a Gaussian
copula. Therefore, the ratios of true and estimated VaR are even higher than their theo-
retical values. This behaviour is in sharp contrast to the student t copula, where the true
VaR and its Gaussian estimation are almost identical. Contrary to Hamerle and Rosch
10
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
13/28
(2005), we therefore find that there is substantial model risk in the dependence structure
of a credit portfolio.
There is a variety of further research questions, both theoretically and empirically.
Obviously, one could try other infinitely divisible distributions in a one factor approach.
However, if one does not want to rely on Monte Carlo simulations, their density or distri-
bution functions should be known.
Empirically, it might be interesting to apply the copulas proposed in this article to
liquid CDS index tranches. While the NIG and VG distributions have shown to fit to the
correlation smile in liquid CDS index tranches, it might be interesting to check whether
the other distributions lead to a similar fit.
Appendix
This Appendix provides the properties of the infinitely divisible distributions used in this
article.
A.1 NIG Distributions
Definition 1. A random variable X is said to be Normal inverse Gaussian (NIG) dis-
tributed with parameters ,,, with 0, 2 2, > 0, R if its characteristicfunction is given by
X(z) = E[eizX ] = eiz+(0iz), z R,
where iz =
2 (+ iz)2. We write X N IG(,,,).
Proposition 1. LetX N IG(,,,). Then:
1. The density fX of X is
fX(x) =
e
K1(
2 + (x )2)2 + (x )2 e
(x).
K is the modified Bessel function of the third kind,
K
(x) =1
2
0
y1 exp12x(y + y1) dy.
11
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
14/28
2. X is infinitely divisible. The corresponding Levy process is
Xt = t + Gt + WGt,
where = 2 , = and (Gt)t0 is an inverse Gaussian process with parameters
((0)1, 20). (Wt)t0 is a standard Brownian motion.
3. The first four centralized moments (mean, variance, skewness and kurtosis) are
E[X] = +
,
Var[X] =2
3,
S[X] = 32
5
,
K[X] = 3 + 32(2 + 42)7.
with = 0.
The transformation properties are given by the following proposition:
Proposition 2. 1. If X N IG(,,,) and c > 0, then
cX N IGc
, c
,c,c .2. If X1 N IG( , , 1, 1) and X2 N IG( , , 2, 2) then
X1 + X2 N IG ( , , 1 + 2, 1 + 2) .
Proof. This can be seen directly from the definition and the properties cX(z) = X(cz)
and X+Y(z) = X(z) Y(z) for c R and independent X and Y.
Corollary 1. If the factors distributions are
M N IG(, 0, , 0)
and
Zi N IG(
1 , 0,
1 , 0),
then the distribution of Xi =
M +
1 Zi is
Xi N IG(, 0, , 0).
12
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
15/28
A.2 VG Distributions
Definition 2. A random variable X is said to be Variance Gamma (VG) distributed with
parameters ,
R and , > 0 if its characteristic function is given by
X(z) = E[eizX ] = eiz
1 iz + 1
22z2
1
, z R.
We write X V G(,,,).
Proposition 3. LetX V G(,,,). Then:
1. The density fX of X is given by
fX(x) =2exp
x21 2( 1)
(x )22
2
+ 21
2 14
K1
1
2 12(x )222
+ 2 .
K is the modified Bessel function of the third kind,
K(x) =1
2
0
y1 exp
12
x(y + y1)
dy.
2. X is infinitely divisible. The corresponding Levy process is given by
Xt = t + Gt + WGt,
where (Gt)t0 is a Gamma process with parameters (1
, ) and(Wt)t0 is a standardBrownian motion.
3. The first four centralized moments are
E[X] = + ,
Var[X] = 2 + 2,
S[X] = 32 + 22
(2 + 2)3/2,
K[X] = 3(1 + 2 4(2 + 2)2).
Proposition 4. 1. If X V G(,,,) and c > 0, then
cX V G(c,,c,c).
2. IfX1 V G(1, 1, 1, 1) and X2 V G(2, 2, 2, 2) are independent and such that211 =
222 and
121
= 222
then
X1 + X2 V G1 + 2, 121 + 2
,21 + 22, 1 + 2 .13
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
16/28
Proof. This can be seen directly from the definition and the properties cX(z) = X(cz)
and X+Y(z) = X(z) Y(z) for c R and independent X and Y.
Corollary 2. If the factors distributions are
M V G(0,
, 1, 0)
and
Zi V G(0, 1 , 1, 0),
then the distribution of Xi =
M +
1 Zi is
Xi V G(0, , 1, 0).
A.3 MJD Distributions
Definition 3. We call a random variable X Merton-jump-diffusion (MJD) distributed
with parameters R, , , > 0 if its characteristic function is given by
X(z) = E[eizX ] = exp
2z2
2+
e
2z2/2+iz 1
.
We write X MJD(,,,).
Proposition 5. LetX MJD(,,,). Then:1. The density fX of X is
fX(x) = e
k=0
k exp (xk)2
2(2+k2)
k!
2(2 + k2).
2. X is infinitely divisible. The corresponding Levy process is given by
Xt = Wt +Nt
i=1Yi,
where Yi N(, 2) is normally distributed, (Nt)t0 is a Poisson process with inten-sity and (Wt)t0 is a standard Brownian motion.
3. For = 0, the first four centralized moments are
E[X] = 0,
Var[X] = 2 + 2,
S[X] = 0,
K[X] = 3(1 + 3).
14
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
17/28
Proposition 6. 1. If X MJD(,,,) and c > 0 then
cX MJD(c,,c,c) .
2. If X1 MJD(1, 1, , ) and X2 MJD(2, 2, , ) are independent then
X1 + X2 MJD
21 + 22, 1 + 1, ,
Proof. This can be seen directly from the definition and the properties cX(z) = X(cz)
and X+Y(z) = X(z) Y(z) for c R and independent X and Y.
Corollary 3. If the factors distributions are
M MJD1 , , 0, 1
and
Zi MJD
1 , (1 ), 0, 11
,
then the distribution of Xi =
M +
1 Zi is
Xi M J D
1 ,, 0, 1
.
A.4 KJD Distributions
Definition 4. We call a random variable X Kou-jump-diffusion (KJD) distributed with
parameters (,,+, , p) if its characteristic function is given by
X(z) = E[eizX ] = exp
2z2
2+ iz
p
+ iz 1 p
+ iz
.
We write X KJD(,,+, , p).Proposition 7. LetX KJD(,,+, , p). Then:
1. The density fX of X is approximatively given by
fX(x) =1
x
+
p+e
(22+)/2e(x)+
x 2+
+ (1 p)e(22)/2e(x)x + 2+
.
15
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
18/28
2. X is infinitely divisible. The corresponding Levy process is
Xt = Wt +Nt
i=1Yi,
where Yi N(, 2) is two-sided exponentially distributed with parameters + and, (Nt)t0 is a Poisson process with intensity and(Wt)t0 is a standard Brownian
motion.
3. The first four centralized moments are
E[X] =
p
+ 1 p
,
Var[X] = 2 + p2+
+ 1 p2
,S[X] =
p
3+ 1 p
3
,
K[X] =
p
4++
1 p4
.
Proposition 8. 1. If X KJD(,,+, , p) and 0 < c < 1, then
cX
KJDc, , +c , c , p2. If X1 KJD(1, 1, +, , p) and X2 KJD(2, 2, +, , p) are independent,
then
X1 + X2 KJD
21 + 22, 1 + 2, +, , p
.
Corollary 4. If the factors distributions are
M KJD1 2+
,, +, +, pand
Zi KJD
1 2+
, (1 ),
1 +,
1 +, p
,
then the distribution of Xi =
M +
1 Zi is
Xi
KJD1
2+
, , +, +, p .
16
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
19/28
References
[1] Barndorff-Nielssen, O.E. (1997). Normal inverse gaussian distributions and
stochastic volatility modelling. Scandinavian Journal of Statistics 24, 1-13.
[2] Cont, R., and Tankov, P. (2003). Financial Modelling with Jump Processes. Chap-
man & Hall, London.
[3] Hamerle, B., and Rosch, D. (2005). Misspecified Copulas in Credit Risk Models:
How good is Gaussian?. Journal of Risk, 8, Fall, 41-58.
[4] Li, D.X. (2000). On default correlation: A copula function approach. Journal of
Fixed Income 9, March, 43-54.
[5] Kalemanova, A., Schmid, B. and Werner, R. (2005). The Normal inverse Gaus-
sian distribution for synthetic CDO pricing. Working paper, RiskLab Germany, Mu-
nich.
[6] Kou, S.G. (2002). A Jump-Diffusion Model for Option Pricing. Management Science
48, August, 1086-1101.
[7] Madan, D.B., Carr, P.P. and Chang, E.C. (1998). The Variance Gamma process
and Option Pricing. European Finance Review 2, 79-105.
[8] Merton, R.C. (1974). On the Pricing of Corporate Debt - The Risk Structure of
Interest Rates. Journal of Finance 3,449-470.
[9] Merton, R.C. (1976). Option pricing when underlying stock returns are discontin-
uous. Journal of Financial Economics3
, 115-144.
[10] Moosbrucker, T. (2006). Explaining the Correlation Smile Using Variance Gamma
Distibutions. Journal of Fixed Income 15, ..-.. .
[11] Schloegl, L., and OKane, D. (2005). A note on the large homogeneous portfolio
approximation with the Student-t copula. Finance and Stochastics, 577-584.
[12] Schonbucher, P. (2002). Taken to the Limit: Simple and Not-so-Simple Loan Loss
Distributions. Working Paper, University of Bonn.
17
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
20/28
[13] Vasicek, O. (1987). Probability of Loss on Loan Portfolio. Memo, KMV Corporation.
18
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
21/28
Distribution mean variance skewness kurtosis jump copmponent
Gauss = 0 = 1
T1 = 0 = 1 df = 22
T2 = 0 = 1 df = 10
NIG1 = 0 = = 0 = 3
NIG2 = 0 = = 0 =
3
VG1 = 0 = 1 = 0 = 19
VG2 = 0 = 1 = 0 = 13
MJD1 = 0 =
1 2 = 19
= 1
MJD2 = 0 = 1 2 = 13 = 1KJD1 = 0 =
1
2+
p = 12
, + = = 112 + =12
KJD2 = 0 =
1 2+
p = 12
, + = = 14 + =12
Table 1: Parameters sets used in the numerical examples in section 3. Parameters of T1, NIG1, VG1,
MJD1 and KJD1 are chosen to lead to an excess kurtosis of 13
. T2, NIG2, VG2, MJD2 and KJD2 have
an excess kurtosis of 1. In MJD and KJD the additional parameters are chosen such that the distribution
of jumps has unit variance.
Correlation = 20% Correlation = 40%
Confidence Level Confidence Level
90% 95% 99% 99.5% 99.9% 90% 95% 99% 99.5% 99.9%
T1 1.052 1.168 1.348 1.268 1.468 0.933 1.032 1.165 1.193 1.217
NIG1 0.814 0.837 1.054 1.151 1.648 0.817 0.809 0.992 1.117 1.411
VG1 0.816 0.854 1.080 1.155 1.596 0.822 0.820 0.999 1.116 1.381
MJD1 0.854 0.836 0.889 1.023 1.876 0.823 0.807 0.927 1.048 1.568
KJD1 0.421 0.309 0.245 0.601 3.143 0.471 0.401 0.534 0.915 3.023
T2 1.045 1.304 1.720 1.563 1.970 0.829 1.042 1.345 1.408 1.447
NIG2 0.654 0.662 0.984 1.236 2.386 0.650 0.616 0.916 1.209 1.979
VG2 0.654 0.697 1.050 1.240 2.248 0.663 0.650 0.936 1.188 1.921
MJD2 0.655 0.591 0.773 1.336 2.870 0.614 0.527 0.867 1.335 2.119
KJD2 0.512 0.445 0.501 0.601 3.699 0.350 0.216 0.237 1.044 3.044
Table 2: VaR ratios of specified model over Gaussian model for identical correlations of default triggering
variables. Marginal default probabilities are 1% and VaR is calculated via LHP approximation.
19
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
22/28
Correlation = 20% Correlation = 40%
Confidence Level Confidence Level
90% 95% 99% 99.5% 99.9% 90% 95% 99% 99.5% 99.9%
T1 1.054 1.077 1.104 1.108 1.111 1.024 1.040 1.051 1.050 1.042NIG1 0.916 0.965 1.152 1.240 1.405 0.920 0.977 1.123 1.161 1.176
VG1 0.913 0.975 1.169 1.251 1.391 0.914 0.979 1.133 1.167 1.172
MJD1 0.962 0.967 1.019 1.077 1.447 0.960 0.978 1.050 1.095 1.195
KJD1 0.919 0.924 0.989 1.080 1.955 0.910 0.940 1.092 1.262 1.332
T2 1.114 1.162 1.215 1.222 1.221 1.052 1.087 1.107 1.103 1.083
NIG2 0.815 0.904 1.319 1.519 1.762 0.812 0.937 1.297 1.352 1.281
VG2 0.795 0.915 1.378 1.576 1.756 0.787 0.932 1.344 1.372 1.275
MJD2 0.865 0.882 1.175 1.530 1.839 0.855 0.922 1.259 1.370 1.296
KJD2 0.691 0.691 1.291 2.225 2.062 0.631 0.728 1.781 1.638 1.331
Table 3: VaR ratios of specified model over Gaussian model for identical correlations of default triggering
variables. Marginal default probabilities are 7.5% and VaR is calculated via LHP approximation.
Marginal default probability
1% 7.5%
Gaussian 20% 40% 20% 40%
default de 2.413% 7.736% 7.024% 16.823%
T1 10.42% 32.89% 16.77% 37.75%
NIG1 14.60% 33.27% 18.12% 37.55%
VG1 15.07% 33.78% 18.01% 37.40%
MJD1 21.66% 36.35% 20.93% 38.77%
KJD1 12.54% 20.33% 17.03% 34.23%
T2 0.23% 23.34% 12.08% 34.18%
NIG2 9.62% 25.44% 16.02% 34.44%
VG2 8.98% 26.58% 15.23% 33.73%
MJD2 23.73% 35.05% 22.15% 36.33%
KJD2 17.77% 24.96% 13.71% 23.55%
Table 4: Correlations of default triggering variables of the specified models that lead to the the same
default correlations as the Gaussian copula.
20
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
23/28
Gaussian Correlation = 20% Gaussian Correlation = 40%
Confidence Level Confidence Levelde 90% 95% 99% 99.5% 99.9% de 90% 95% 99% 99.5% 99.9%
T1 10.42% 1.017 1.020 1.016 1.012 1.001 32.89% 1.000 1.008 1.013 1.012 1.010
NIG1 14.60% 0.764 0.742 0.861 0.966 1.306 33.27% 0.841 0.775 0.853 0.938 1.177
VG1 15.07% 0.771 0.769 0.906 1.004 1.292 33.78% 0.844 0.790 0.873 0.953 1.164
MJD1 21.66% 0.860 0.860 0.947 1.058 2.005 36.35% 0.847 0.791 0.843 0.940 1.438
KJD1 12.54% 0.599 0.455 0.360 0.360 1.489 20.33% 0.559 0.387 0.275 0.297 1.715
T2 0.23% 1.078 1.100 1.081 1.058 0.994 23.34% 0.982 1.041 1.095 1.100 1.091
NIG2 9.62% 0.694 0.634 0.676 0.744 0.992 25.44% 0.694 0.634 0.676 0.744 0.992
VG2 8.98% 0.549 0.522 0.659 0.788 1.264 26.58% 0.673 0.593 0.704 0.838 1.312
MJD2 23.73% 0.659 0.618 0.916 1.535 3.215 35.05% 0.636 0.522 0.757 1.163 1.933KJD2 17.77% 0.426 0.310 0.225 0.340 2.358 24.96% 0.405 0.237 0.162 0.292 2.194
Table 5: VaR ratios of specified model over Gaussian model for identical default correlations. Marginal
default probabilities are 1% and VaR is calculated via LHP approximation.
Gaussian Correlation = 20% Gaussian Correlation = 40%
Confidence Level Confidence Level
de 90% 95% 99% 99.5% 99.9% de 90% 95% 99% 99.5% 99.9%
T1 16.77% 1.010 1.012 1.014 1.014 1.013 37.75% 1.010 1.009 1.011 1.010 1.008
NIG1 18.12% 0.888 0.925 1.093 1.177 1.348 37.55% 0.888 0.941 1.074 1.117 1.150
VG1 18.01% 0.884 0.933 1.107 1.186 1.331 37.40% 0.894 0.942 1.081 1.120 1.144
MJD1 20.93% 0.974 0.986 1.049 1.110 1.054 38.77% 0.950 0.960 1.025 1.071 1.183
KJD1 17.03% 0.880 0.865 0.890 0.950 1.840 34.23% 0.880 0.857 0.951 1.100 1.328T2 12.08% 1.016 1.014 1.005 1.001 0.992 34.18% 1.009 1.010 1.008 1.007 1.003
NIG2 16.02% 0.856 0.879 1.026 1.106 1.281 34.44% 0.856 0.896 1.013 1.059 1.114
VG2 15.23% 0.732 0.813 1.186 1.376 1.636 33.73% 0.732 0.844 1.206 1.274 1.243
MJD2 22.15% 0.890 0.925 1.273 1.622 1.875 36.33% 0.828 0.866 1.174 1.311 1.283
KJD2 13.71% 0.651 0.607 0.712 1.519 2.017 23.55% 0.651 0.527 1.084 1.512 1.317
Table 6: VaR ratios of specified model over Gaussian model for identical default correlations. Marginal
default probabilities are 7.5% and VaR is calculated via LHP approximation.
21
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
24/28
Marginal default probability
1% 7.5%
Correlation Correlation
20% 40% 20% 40% T1 1.100 26.930 1.142 41.654 7.497 21.280 7.565 39.081
NIG1 1.010 16.483 1.030 29.754 7.508 17.152 7.501 33.149
VG1 0.984 12.401 1.004 26.255 7.599 15.754 7.439 30.361
MJD1 0.966 12.058 0.927 22.722 7.476 16.262 7.557 31.741
KJD1 0.950 6.850 0.916 12.479 7.463 14.708 7.397 27.941
T2 1.076 35.945 1.089 46.627 7.404 25.197 7.631 41.991
NIG2 0.977 13.744 0.953 22.934 7.408 16.156 7.546 29.567
VG2 1.010 9.086 0.945 17.557 7.530 12.949 7.578 24.847
MJD2 0.927 7.309 0.909 15.016 7.454 13.386 7.430 25.840
KJD2 0.941 3.493 0.980 7.563 7.618 11.555 7.489 21.618
Table 7: Estimated parameters under the assumption of a Gaussian copula.
Correlation = 20% Correlation = 40%
Confidence Level Confidence Level
90% 95% 99% 99.5% 99.9% 90% 95% 99% 99.5% 99.9%
T1 0.918 0.948 0.988 0.999 1.013 0.812 0.894 1.012 1.042 1.079
NIG1 0.837 0.898 1.206 1.404 1.978 0.764 0.864 1.261 1.482 1.967
VG1 0.917 1.053 1.533 1.803 2.527 0.795 0.941 1.440 1.698 2.247
MJD1 0.984 1.060 1.305 1.501 3.086 0.876 1.057 1.597 1.930 3.132
KJD1 0.756 0.792 1.020 1.288 9.492 0.570 0.678 1.434 2.773 10.870
T2 0.935 0.972 1.001 1.001 0.993 0.801 0.926 1.086 1.117 1.140
NIG2 0.722 0.789 1.312 1.762 3.466 0.672 0.784 1.533 2.164 3.846
VG2 0.773 0.948 1.775 2.382 4.527 0.723 0.928 1.932 2.690 4.923
MJD2 0.885 0.941 1.593 2.889 7.295 0.718 0.832 2.067 3.534 6.478
KJD2 0.663 0.618 0.719 1.301 12.827 0.449 0.416 0.819 4.228 15.668
Table 8: VaR ratios of specified model over Gaussian model; Gaussian parameters estimated by maximum
likelihood. Marginal default probabilities are 1% and VaR is calculated via LHP approximation.
22
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
25/28
Correlation = 20% Correlation = 40%
Confidence Level Confidence Level
90% 95% 99% 99.5% 99.9% 90% 95% 99% 99.5% 99.9%
T1 1.035 1.049 1.064 1.067 1.067 1.024 1.047 1.064 1.064 1.054
NIG1 0.957 1.028 1.255 1.358 1.547 0.981 1.083 1.284 1.329 1.329
VG1 0.967 1.063 1.324 1.429 1.605 1.011 1.143 1.383 1.427 1.408
MJD1 1.024 1.055 1.147 1.220 1.652 1.032 1.103 1.233 1.287 1.386
KJD1 1.009 1.050 1.175 1.297 2.375 1.040 1.148 1.413 1.637 1.695
T2 1.049 1.060 1.065 1.064 1.056 1.018 1.042 1.058 1.057 1.046
NIG2 0.877 0.997 1.499 1.737 2.029 0.895 1.096 1.598 1.670 1.557
VG2 0.896 1.081 1.737 2.017 2.291 0.914 1.180 1.847 1.898 1.735
MJD2 0.974 1.039 1.467 1.937 2.367 0.998 1.165 1.709 1.868 1.734
KJD2 0.793 0.842 1.707 3.001 2.854 0.774 0.991 2.683 2.493 1.994
Table 9: VaR ratios of specified model over Gaussian model; Gaussian parameters estimated by maximum
likelihood. Marginal default probabilities are 7.5% and VaR is calculated via LHP approximation.
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T1NIG1
VG1
MJD1
KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T2NIG2
VG2
MJD2
KJD2
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T1
NIG1VG1
MJD1
KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T2
NIG2VG2
MJD2
KJD2
Figure 1: VaR ratios for identical correlations of default triggering variables, PD= 1%.
23
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
26/28
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T1
NIG1
VG1
MJD1
KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T2
NIG2
VG2
MJD2
KJD2
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T1
NIG1
VG1MJD1
KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T2
NIG2
VG2MJD2
KJD2
Figure 2: VaR ratios for identical correlations of default triggering variables, PD= 7.5%.
24
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
27/28
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T1
NIG1
VG1
MJD1
KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T2
NIG2
VG2
MJD2
KJD2
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T1
NIG1
VG1
MJD1KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T2
NIG2
VG2
MJD2KJD2
Figure 3: VaR ratios for identical default correlations, PD= 1%.
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T1
NIG1
VG1
MJD1
KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T2
NIG2
VG2
MJD2
KJD2
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T1
NIG1VG1
MJD1
KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T2
NIG2VG2
MJD2
KJD2
Figure 4: VaR ratios for identical default correlations, PD= 7.5%.
25
-
8/2/2019 Copulas From Infinitely Divisible Distributions-Applications to Credit Value at Risk
28/28
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T1
NIG1
VG1
MJD1
KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T2
NIG2
VG2
MJD2
KJD2
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T1
NIG1
VG1
MJD1KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T2
NIG2
VG2
MJD2KJD2
Figure 5: VaR ratios between true and misspecified (Gaussian) copula, PD= 1%.
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T1
NIG1
VG1
MJD1
KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=20%
T2
NIG2
VG2
MJD2
KJD2
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T1
NIG1VG1
MJD1
KJD1
0.9 0.92 0.94 0.96 0.98 10
1
2
3
=40%
T2
NIG2VG2
MJD2
KJD2
Figure 6: VaR ratios between true and misspecified (Gaussian) copula, PD= 7.5%.