The Perfect Copula
John Hull and Alan White1
Joseph L. Rotman School of Management
University of Toronto
First Draft: June 2005
This Draft: June 2006
1
We are grateful to Bill Bobey for research assistance and to Leif Andersen, Jon Gregory, Harald Skarke,
and anonymous referees for comments that have improved this paper.
The Perfect Copula
John Hull and Alan White
As the CDO market has grown researchers have increased their efforts to develop models
for losses on portfolios of exposures. The standard market model is the Gaussian copula
model. The structure for this model was suggested by Vasicek (1987) and it was first
applied to credit derivatives by Li (2000) and Gregory and Laurent (2003). Market
participants now imply correlations from the model in much the same way that they
imply volatilities from the Black-Scholes model.
The evolution of research for valuing credit derivatives in some ways mirrors the
evolution of research for valuing options. Like the Black-Scholes model, the Gaussian
copula model does not provide a good description of market prices. Implied correlations,
both tranche (compound) correlations and base correlations, vary significantly from
tranche to tranche. Just as researchers have developed stochastic volatility models and
jump models to explain volatility smiles they have developed alternatives to the Gaussian
copula model to explain correlation smiles. Among the copulas that have been considered
are the Student-t, double-t, Clayton, Archimedian, and Marshall Ohkin.
An important development in the pricing of stock options was the work of Breeden and
Litzenberger (1978) that showed that it is possible to imply a stock price distribution for a
future time from the prices of options with different strike prices that mature at that time.
Jackwerth and Rubinstein (1996) showed how the procedure can be operationalized. The
result is a model that perfectly fits the prices of options with the chosen maturity and is
particularly useful for valuing derivatives that provide a non-standard payoff at that
maturity. In this paper, analogously to Jackwerth and Rubinstein, we show how to
develop a copula that provides a perfect fit to iTraxx or CDX quotes. This is particularly
useful for valuing CDO-squareds and other instruments with non-standard structures.
2
One-Factor Copula Models
Suppose that we are interested in modeling the joint defaults of n different obligors. In a
one-factor copula model we first define variables xj (1≤ j ≤ n) by
x j = a j M + 1 − a 2j Z j (1)
where M and the Zj’s have independent probability distributions. The variable xj can be
thought of as a default indicator variable for the jth obligor: the lower the value of the
variable, the earlier a default is likely to occur. Each xj has two stochastic components.
The first, M, is the same for all xj while the second, Zj, is an idiosyncratic component
affecting only xj.
Suppose that tj is the time to default of the jth obligor, Qj is the cumulative probability
distribution of tj, and Fj is the cumulative probability distribution for xj. The copula model
maps xj to tj on a “percentile to percentile” basis. The 5% point on the xj distribution is
mapped to the 5% point on the tj distribution; the 10% point on the xj distribution is
mapped to the 10% point on the tj distribution; and so on. In general, the point tj = t is
mapped to xj = x where
[
x = F j−1 Q j (t ) ] (2)
or equivalently
[
t = Q −j 1 F j ( x) ]
The copula model defines a correlation structure between the tj’s while maintaining their
marginal distributions. The essence of a copula model is that we do not define the
correlation structure between the variables of interest directly. We map the variables of
interest into other more manageable variables and define a correlation structure between
those variables.
3
From equation (1)
⎡x−a M ⎤
Prob ( x j < x M ) = H j ⎢ ⎥
j
⎢ 1 − a 2j ⎥
⎣ ⎦
where Hj is the cumulative probability distribution of Zj. It follows from equation (2) that
⎧ F −1 ⎡Q ( t ) ⎤ − a M ⎫
⎪ j ⎣ j ⎦ ⎪
Q j ( t M ) = Prob ( t j < t M ) = H j ⎨
j
⎬ (3)
⎪⎩ 1− a j 2
⎭⎪
Conditional on M, defaults are independent. When using the model to value a CDO
tranche we set up a procedure to calculate expected cash flows on the tranche conditional
on M and then integrate over M to obtain the unconditional expected cash flows.
In the Gaussian copula model both M and the Zj have standard normal distributions. In
this case xj also has a standard normal distribution so that Hj = Fj = N for all j where N is
the cumulative normal distribution function. As already mentioned the one-factor
Gaussian copula model does not fit market data well.
Equation (3) defines the cumulative probability of default by time t conditional on M.
The variable M defines the default environment for the whole life of the model. Once M
has been determined the cumulative probability of default Qj is a known function of time.
A one-factor copula model can be thought of as a model where there are a many possible
outcomes for the Qj and the realization of M that occurs at time zero defines which will
be selected for each j. There is no stochastic evolution for hazard rates or CDS spreads in
the model. As M increases Qj(t) decreases for all obligors and all t. There is perfect
dependence between the outcomes in that when the outcome for one obligor has been
specified the outcomes for all other obligors are determined. In the homogeneous case
where the aj and the Qj are the same for all j, the Q-outcome is the same for all obligors.
4
Implied Hazard Rate Paths
Instead of formulating the model in terms of conditional Q’s we can instead use
conditional hazard rates. Define λ j ( t M ) as the hazard rate at time t conditional on M
for company j. The relationship between λ j ( t M ) and Q j ( t M ) is
⎡ t ⎤
Q j ( t M ) = 1 − exp ⎢ − ∫ λ j ( τ M ) d τ ⎥
⎣ 0 ⎦
or equivalently
dQ j ( t M ) dt
λ j (t M ) =
1 − Q j (t M )
This equation can be used in conjunction with equation (3) to calculate the hazard rate as
a function of time for alternative values of M. We shall refer to this conditional function
of time as a hazard rate path.
Figure 1 shows hazard rate paths for the Gaussian copula model (M and the Zj both
normal) for the situation where the unconditional hazard rate is 1% per year for all
obligors and the copula correlation, aj2, is 0.15. It shows that the model has some
unrealistic properties. Uncertainty about the future hazard rate decreases with the time
(except for a short initial period). This seems to be a general property of the Gaussian
copula model, true for virtually all assumptions about the unconditional hazard rate and
the copula correlation. Figure 2 shows results for the double t copula model where M and
the Zj have four degrees of freedom. Interestingly, this model has the more realistic
property that uncertainty about the future hazard rate increases with the time.
The Implied Copula Approach
As we have just seen the homogeneous one-factor copula model implies a set of paths for
the hazard rates of obligors. The probability of each path occurring is determined by the
probability distribution for M. Each hazard rate path represents a future credit
environment (FCE). We will refer to the set of all possible hazard rate paths and their
5
probabilities as the distribution of future credit environments (DFCE). It is easy to see
that the DFCE is the only information we need about the underlying copula in order to
value a CDO tranche or similar instrument. We calculate expected cash inflows and
expected cash outflows on the CDO tranche for a particular FCE. To value a CDO
tranche we integrate these expected cash flows over the DFCE.
This suggests a new approach to modeling default correlation. Instead of specifying a
copula we specify the DFCE directly. There is a copula corresponding to any particular
DCFE. (All the one-factor copulas that have been proposed are particular cases of our
approach.) However, from the perspective of model implementation there is no need to
determine what this copula is. The default correlation in the model depends on the
dispersion of the DFCE. If there were only one FCE the correlation would be zero. As the
dispersion of the DFCEs increases default correlation increases.
At this stage, instead of defining the most general model possible we focus on developing
practical implementation procedures. For ease of exposition we start by assuming that a)
all obligors have the same set of hazard rate paths2 and b) the hazard rate is constant
along each hazard rate path. Later we will explain how these assumptions can be relaxed.
We first choose a number of different values for the five-year default rate. The number of
values chosen, which we will denote by n, must be sufficiently large to provide enough
degrees of freedom to fit the market data and the values chosen must span a plausible
range of default rates. Good results are obtained when n ≥ 50. It turns out that to fit the
market data it is necessary to include some very high default rates and very low default
rates in the set. The steps are then as follows:
1. Calculate hazard rate paths corresponding to each value of the five-year default
rate. When the hazard rate is constant it equals –ln[1–Q(5)]/5.
2. Conditional on each hazard rate path, calculate the present value of the cash
inflows and the present value of the cash outflows for each CDO tranche and for
the credit default swap (CDS) represented by the index.
2
The corresponding assumption in the one-factor copula model is that aj and Qj are the same for all j.
6
3. Select probabilities to assign to the hazard rate paths so that the unconditional
expected value of each CDO tranche and the unconditional expected value of the
CDS are zero.
The procedure we have outlined is open to the criticism that the solution is not unique
because many probabilities are being implied from six data points. This is true. There are
potentially many different distributions of future credit events that are consistent with the
observed market data. As a result there is a degree of uncertainty about the value of any
non-standard contract. An alternative way of stating this is that there are an infinite
number of one-factor copulas that can exactly fit observed market prices. It is not clear
how one would choose amongst all these copulas to select the ‘correct’ model.
We have not found the multiple solutions to be a serious problem. While it is possible to
construct contracts that are very sensitive to the exact distribution of hazard rates, for
most contracts that are encountered the value does not vary much when we move from
one distribution that exactly fits market prices to another. We discuss the multiple
solutions issue in more detail in Hull and White (2006). We show that, contrary to the
perception of some researchers, the perfect copula model is more stable than the
Gaussian copula base correlation model that is used by many market participants. This is
because the latter is highly sensitive to the interpolation procedure used for the base
correlations. Some interpolation procedures result in prices for non-standard tranches that
are not arbitrage free. By contrast, the prices given by the implied copula approach are
always arbitrage free.
To avoid any ambiguity about the results from using the implied copula approach we
have standardized our implementation in a number of ways. First we avoid any
arbitrariness in the way the five-year default rates are chosen. Suppose we wish to choose
n default rates. We choose the spacing between the default rates to reflect the sensitivity
of the prices of the six instruments (five CDO tranches and one CDS) to the default rate.
The idea here is that when this sensitivity is low we can afford to space the default rates
more widely whereas when there is a significant price change as a result of a small
change in default rate the default rates need to be spaced more closely. Suppose that L is
the sum of the values of the six instruments when the five-year default rate is zero and U
7
is the sum of the values when it is 100%. We divide the range between L and U into n–1
equal intervals and set the ith five-year default rate, di, to be the one that corresponds to
the beginning of the ith interval. The nth default rate is set equal to U.
Next we choose a single solution from the set of possible solutions by adding to the
objective function a penalty for non-smoothness of the distribution. The problem that is
solved is shown in Table A. This is a standard quadratic programming problem with a
unique solution. The parameter c determines the relative importance of smoothness and
the fit to market prices. As is usual in financial engineering appropriate values for this
parameter are determined by trial and error.3
Recovery Rate Assumption
Research by Altman et al (2002) and Cantor et al (2002) and Hamilton et al (2005)
provides empirical support for the existence of a negative correlation between default
rates and recovery rates. The best fit relationship reported by Hamilton et al (2005) is
RR=0.52 – 6.9×DR (4)
where RR is the recovery and DR is the annual default rate. Our implied copula
procedure can incorporate this relationship. In step 2 above we use a different recovery
rate for each hazard rate path.
When testing the model using data from the second half of 2004 we found that a model
with a constant recovery rate of 40% can fit CDX and iTraxx data exactly. However, in
tests using more recent data we found it necessary to incorporate the relationship in
equation (4) to get an exact fit. The results we will present in the next section use
equation (4). In implementation RR is set equal to zero when equation (4) gives a value
less than zero.
3
An alternative approach is to minimize the second term in the objective function subject to the first term
being less than some small amount. Yet another approach involving linear programming is in Hull and
White (2006)
8
Results
Table B shows iTraxx data and CDX data for December 1, 2005 obtained from Reuters.
Figure 3 shows the probability density functions obtained for iTraxx and CDX using the
5-year data using 50 points on the distribution. The probability density for iTraxx peaks
at a five-year default rate of about 2.8% while that for CDX peaks at a five-year default
rate of about 3.5%. Figure 4 shows the five-year CDX distribution when different choices
are made for the number of points on the distribution. It shows that when a smoothing
condition is applied the implied probability density is well behaved as the number of
points is increased.
An implied 10-year default rate distribution can be determined in a similar way to an
implied five-year distribution. Figure 5 shows the cumulative probability distribution for
the 5-year and 10-year default rate for CDX. At this point it is natural to suggest that the
methodology be extended to find a 10-year hazard rate path distribution that
simultaneously matches the market data for both five and ten year maturities. It is
possible to find such a hazard rate path distribution, but it is important to avoid over
fitting the model. The model is a description of the average hazard rate environment
between time zero and some time T as seen at time zero. The model does not say
anything about the dynamics of hazard rates. A different type of model is needed to
answer a question such as “if the hazard rate between now and year 5 is h, what is the
probability distribution for hazard rates between years 5 and 10?”
When valuing CDO tranches and other similar instruments that have maturities between
five and ten years, it makes sense to interpolate between default rates. For example, in the
case of CDX the probabilities that the five- and ten-year default rates will be less than or
equal to a threshold level of 10% are 0.930 and 0.652, respectively. From this we can
estimate the chance of the seven-year default rate being less than or equal to 10% to be
0.819. Similar calculations can be carried out for other threshold levels and a complete
probability distribution for seven-year default rate calculated.
9
The Non-Homogeneous Case
We now discuss how the model we have presented can be extended so that it matches the
term structures of credit spreads for different companies.
Suppose that the life of the model is tN years. As a first step we show how the model can
be extended so that it matches the tN year CDS spread for each company. We then show
how the term structure of CDS spreads for each company can be matched
We define variables as follows:
di,index (1 ≤ i ≤ n): the ith cumulative default rate that is used in the homogeneous constant
hazard rate model. (In Table 1, this is simply di.)
dij (1 ≤ i ≤ n): the ith cumulative default rate for the jth company (to be determined)
Uj(d): the value of a CDS on the jth company when the cumulative default rate is d and
the hazard rate is constant
Uindex(d): the value of a CDS on the index when the cumulative default rate is d and the
hazard rate is constant
An equation for Uj(d) is
U j (d ) = X (d ) − Y (d ) s j
where X(d) is the value of the payoff for a principal of $1, Y(d) is the value of payments
at the rate of $1 per year, and sj is the CDS spread for the jth company.
We choose the dij so that
U j (d ij )
= Cj
U index (d i ,index )
where Cj is a constant independent of i. The Cj can be chosen one company at a time
using a fast iterative search. The advantage of setting things up so that this equation is
satisfied is that, when the probabilities pj are chosen so that the index CDS spread is
matched, the CDS spread for individual names are also matched. Specifically when
n
∑ pU
i =1
i index (d i ,index ) = 0
10
then
n
∑ pU
i =1
i j (d ij ) = 0
We now explain how a similar idea can be used to extend the model so that it matches the
term structure of CDS spreads. Suppose that the CDS spread is known for company j for
maturities t1, t2, t3,…, tN. We assume that the ith hazard rate for the company is constant
between time tk-1 and time tk (t0=0). We choose the hazard rates so that the ratio of the
value of the tk year CDS to the value of the tN year CDS is independent of i while
ensuring that the hazard rates on the ith path are consistent with the cumulative default
rate dij. This is not difficult to do. The ratio of the value of the tk year CDS to the tN year
CDS can be chosen as (wksk)/(wNsN) where wk is the present value of payment at the rate
on $1 per year on a tk year CDS when the hazard rate is zero and sk is the tk year CDS
spread.
There is perfect correlation between obligors in the generalized model we have outlined
in that when we know the hazard rate path for one obligor we know the hazard rate path
for all other obligors. Searching for the optimal set of probabilities in the model is no
more difficult than in the homogeneous model because the model is set up so that when
we match the CDS spread for the index we automatically match the term structure of
CDS spreads for all companies underlying the index. An algorithm such as that in
Andersen et al (2003) or Hull and White (2004) must be used to calculate loss
distributions on the portfolio for each i, but such an algorithm must also be used for each
value of the factor that is considered in the nonhomogeneous Gaussian copula model.
Default correlation in the generalized perfect copula model we have just presented is
determined by the dispersion of the hazard rate paths which in turn depends on the
tranche quotes. If the dispersion is small, the default correlation is low because regardless
of the default environment all companies tend to have similar hazard rate paths. If it is
large, the default correlation is high because a high (low) hazard rate path for one
company occurs at the same time as a high (low) hazard rate path for all other companies.
11
It is possible to extend the model still further so that the dispersion of hazard rate paths is
different for different companies. However, just as it is difficult to know how to choose
unequal correlations in the Gaussian copula model, so it is difficult to know how to vary
the dispersion of hazard rates from company to company in this model.
The problems in using the implied copula model to value bespoke portfolios are similar
to those in using the Gaussian copula model. If the companies in the portfolio are
considered to be as well diversified as the companies in the iTraxx or CDX index we can
use the hazard rate paths obtained when the model is fitted to quotes for the index. If the
names are not considered to be as well diversified as the index some ad hoc adjustment to
the dispersion of hazard rate paths must be made.
Valuing a CDO Squared
The key advantage of the perfect copula model is in the valuation of non-standard
structure such as a CDO-squared. It is difficult to use the Gaussian copula model for
these instruments because there is no easy way to calculate an appropriate correlation to
input to the model. (This is analogous to the problem in valuing a digital option on a
stock. Although the valuation formula is straightforward there is no easy way to know the
correct volatility to substitute into the formula.)
Consider first the situation where all the names underlying the child portfolios in a CDO
squared are assumed to have the same default risk as the names underlying iTraxx. The
model is first calibrated to iTraxx quotes and the CDO-squared is valued using Monte
Carlo simulation. On each simulation trial a random number between zero and one is
drawn. This random number is used to define the default time and recovery rate for each
name and each hazard rate path. For each hazard rate path the times to default are sorted
and attention is focused on those that occur during the life of the CDO-squared. The latter
are used to calculate the present value of the spread income per basis point and the
present value of the payoff on the CDO-squared. The results for different hazard rate
paths are weighted by the probabilities associated with the hazard rate paths to provide a
present value of spread income and present value of payoff for the simulation trial. An
12
estimate of the expected present value of spread income and expected payoff are
calculated as the arithmetic average of these over many simulation trials.
The advantage of using a copula that exactly matches market prices is that the loss
distributions being assumed are exactly consistent with the market. (This is equivalent to
valuing a one-year digital option on a stock using the one-year stock price distribution
implied from one-year options.) The procedure we have outlined can be modified to
incorporate heterogeneous assumptions in the way described in the previous section.
In Hull and White (2006) we have tested this procedure using CDO-squared examples
similar to those occurring in practice. For example, in one of our tests there were 10
CDOs, all mezzanine tranches with 80 underlying names. We used the breakeven spread
calculated from our model to imply a correlation from the Gaussian copula model. We
found that the implied correlation was quite different for different attachment and
detachment points of the parent CDO and that an analyst would be unlikely to guess the
correct correlation from the correlations implied by CDO quotes.
Conclusions
We have presented a new approach to modeling default correlation. The approach is as
easy to implement as the Gaussian copula model and has good stability properties.
The approach has a number of advantages over the Gaussian copula model and its
extensions. First, the model can be exactly fitted to the market quotes for the actively
traded CDO tranches of standard portfolios. Second, the model is easier to understand
than copula models. Third, the model can easily be used to calculate partial derivatives
with respect to market quotes such as those in Table B. Fourth, the model can be used to
value CDOs on bespoke portfolios and more exotic structures such as CDO-squareds.
Finally, the model is ideally suited for trading. A trader can first calculate the implied
distributions such as those in Figures 3. She can then investigate the effect on market
prices of modifying the distributions to reflect her beliefs. A natural extension is to a two-
factor model that fits iTraxx and CDX tranches simultaneously. This extension can be
13
achieved by assuming a copula correlation structure for the FCE’s for North America and
Europe.
There are limitations of the model. Like the Gaussian copula it does not involve the
dynamic evolution of hazard rates or credit spreads. It is therefore inappropriate for some
instruments. For example, the model is not appropriate for valuing a one-year option on a
five-year CDO because this depends on hazard rates between years one and five
conditional on what we observe happening during the first year.
14
References
Altman, E.I., B. Brady, A. Resti, and A. Sironi, “The link between default rates and
recovery rates: implications for credit risk models and pro-cyclicality,” Working Paper,
Stern School of Business, New York University, April 2002.
Andersen, L., J. Sidenius, and S. Basu, “All your hedges in one basket,” RISK,
November, 2003.
Breeden, D. T. and R. H. Litzenberger, “Prices of state-contingent claims implicit in
option prices,” Journal of Business, 51 (1978), 621-51.
Cantor, R., D.T. Hamilton, and S. Ou, “Default rates and recovery rates of corporate bond
issuers,” Moody’s Investor’s Services, February, 2002.
Hamilton, D.T., P. Varma, S. Ou, and R. Cantor, “Default and recovery rates of corporate
bond issuers,” Moody’s Investor’s Services, January 2004.
Hull, J. and A. White, “Valuing credit derivatives using an implied copula approach,”
Working Paper, University of Toronto, 2006.
Hull, J. and A. White, “Valuation of a CDO and nth to default CDS without Monte Carlo
simulation,” Journal of Derivatives, 12, 2 (Winter 2004), 8-23.
Jackwerth, J. C., and M. Rubinstein. ``Recovering probability distributions from option
prices,'' Journal of Finance}, 51 (December 1996): 1611--31.
Li, D.X., “On default correlation: A copula approach” Journal of Fixed Income, 9 (March
2000), pp 43-54.
Laurent, J-P and J. Gregory, “Basket default swaps, CDO’s and factor copulas” Working
Paper, ISFA Actuarial School, University of Lyon, 2003
Vasicek, O. ``Probability of loss on a loan portfolio,'' Working Paper, KMV, 1987.
(Published in Risk in December 2002 under the title ``Loan Portfolio Value''.)
15
A: The Optimization
Suppose we have determined n five-year default rates and these, when arranged in
increasing order, are d1, d2,…dn . We have quotes for six instruments (five CDO
tranches and one CDS index). Define viq as the value of the qth instrument for the ith
default rate (calculated using the quote for the qth instrument) and pi as the
probability assigned to the ith default rate. The value of the qth instrument is
n
Vq = ∑ pi viq
i =1
If Vq = 0 for all q the model is consistent with the market. We choose the pi to
minimize
2
6 n −1
⎡ p + pi −1 − 2 pi ⎤
∑
q =1
V + c ∑ ⎢ i +1
q
2
⎥
i = 2 ⎣ 0.5( d i +1 − d i −1 ) ⎦
subject to
n
∑p i =1
i =1
and
pi ≥ 0
16
A: Quotes for CDX IG and iTraxx Tranches on December 1, 2005. Source:
Reuters
CDX IG Tranches
0% to 3% 3% to 7% 7% to 10% 10% to 15% 15% to 30% Index
5-year Quotes 32.5 196 71 26 9 47
10-year Quotes 54.0 507 177 86 33 67
iTraxx Tranches
0% to 3% 3% to 6% 6% to 9% 9% to 12% 12% to 22% Index
5-year Quotes 23.0 131 44 26.5 14 36.0
10-year Quotes 45.5 367 152 87.5 43 53.3
17
Figure 1
Hazard Rate Paths for the Gaussian copula model when the
unconditional hazard rate is 1% per year and the correlation is 15%
0.06
M = -2
M = -1
0.05
M=0
M=1
0.04 M=2
Hazard Rate
0.03
0.02
0.01
0.00
0 1 2 3 4 5
Years
18
Figure 2
Hazard Rate Paths for the double t copula model when the
unconditional hazard rate is 1% per year and the correlation is 15%
0.03
M = -2 M = -1 M=0
M=1 M=2
0.02
Hazard Rate
0.01
0.00
0 1 2 3 4 5
Years
19
Figure 3
Implied Probability Densities for the 5-Year Default
Rate for iTraxx and CDX on December 1, 2005
Implied Density
5Y iTraxx
5Y CDX
0 0.05 0.1 0.15 0.2 0.25
5-Year Default Probability
20
Figure 4
Effect of increasing number of points when estimating the 5-year
CDX default rate distribution on December 1, 2005
30
Implied Density
50
75
100
0 0.05 0.1 0.15 0.2
5-Year Default Probability
21
Figure 5
Implied Cumulative Distribution for 5-year and
10-year CDX default rate on December 1, 2005
1
0.9
0.8
Cumulative Probability
0.7
0.6 5-Yr
0.5
10-Yr
0.4
0.3
0.2
0.1
0
0 0.05 0.1 0.15 0.2 0.25 0.3
5- and 10-Year Default Probability
22
