Available at www.defaultrisk.

com

Analytical Pricing of Basket Default Swaps

in a Dynamic Hull & White Framework
Frédéric D. Vrins∗
This Version : January 2010

Abstract
In this paper, some analytical results related to the Hull & White dynamic model of credit portfolio
of N obligors in the case of constant jump size are provided. For instance, this specific assumption
combined with the moment generating function of the Poisson process lead to analytical calibration for
the model with respect to the underlying CDSs. Further, extremely simple analytical expressions are
obtained for first-to-default swaps; the more general case of quantities related to nth-to-default swaps
also have a closed form and remain tractable for small n. Similarly, pairwise correlation between default
indicators also proves to be simple. Although the purpose of this note is not to compare models, we
compare the shape of pairwise default correlations of the Hull & White, the Gaussian copula and the Mai
& Scherer model with compound Poisson process as Lévy subordinator. It is shown that only the models
including jumps can lead to non-vanishing default correlation for short-term maturities. Further, these
models can generate higher default correlation levels compared to the Gaussian one. When calibrated on
default probability of first default time, Jump-based models also lead to much higher default probability
for the last obligor to default. Finally, we tackle the problem of simultaneous jumps, which prevent the
above class of models to be usable when recoveries are name-specific. To that end, we propose a tractable
compromise to deal with baskets being non-homogeneous recovery-wise under the Hull & White model
by splitting isolated and non-isolated default events.

1 Introduction
As form the Summer 2007 credit crisis, a lot of researchers both from top universities and professional
institutions try to develop a dynamic alternative to the (static) Gaussian copula model for the pricing of
multi-underlying credit derivatives. The challenge lies in the fact that the investigated models either do
not allow to reach sufficiently high default correlation levels (as it is the case when, among others, stochas-
tic hazard rates are modeled as correlated geometric Brownian motions, see [Duffie and Garleanu (2001),
Chapovsky et al. (2006)]) or become very quickly untractable from a practical point of view. Among the
alternatives proposed in the literature is the Hull & White model [Hull and White (2008)]. In that paper,
satisfactory correlation levels are got by introducing jumps in the hazard rates. More precisely, the name-
specific hazard rates are decomposed as a sum of a deterministic continuous part and a pure-jump random
process, so that a calibration of the continuous part must be performed in order for the model to be in-line
with single-name probabilities. These are typically obtained from single-name market of Credit Default
Swaps (CDS).
Defining the random default time of entity i by τi , the key point here is that the single-name survival
.
probabilities Si (t) = Pr[τi > t] are modeled in two different ways depending if we are considering them as
∗ Dr F. Vrins is with the Financial Markets/Credit Derivatives Modeling Team of ING SWE. Contact: +32 2 557 13 13,

[email protected], Aue Marnix 24, M2+4, 1050 Brussels, Belgium. The author is grateful to John Hull for discussions on
the present model.

1
 Available at www.defaultrisk.com

isolated entities or as a part of a portfolio. From the CDS market, we infer the risk-neutral (deterministic,
piece-wise constant) hazard rate curves (λi (s))s>0 , so that :
R
t
Si (t) = e− s=0
λi (s)ds
. (1)

When focusing on the portfolio, the hazard becomes stochastic :

Z t Jt
. . X
Ŝi (t) = e−Xi (t) , Xi (t) = µi (s)ds + H(j) ,
s=0 j=1

where (µi (s))s>0 is a continuous, deterministic function, H(j) > 0 is again a determinisitic function defined
on the set of natural numbers and (Js )s>0 is a non-homogeneous Poisson process with time-t probability
mass function (see e.g. [Schoutens (2003)]):
t
e−Λ(t) (Λ(t))k
Z
.
Pr[Jt = k] = , Λ(t) = λ(s)ds .
k! s=0

The quantity λ(t) is the time-t jump intensity of the process (Js )s>0 (hence, λ(s) > 0 for all s). Recall that
dJs is binary : either 1 or 0. Because the jump process is the same for all variables, this is the “common”
variable creating the dependency. Conditionally on Jt = j, the defaults are independent1 . As a corollary,
the cases λ(s) = 0 ∀s 6 t or H = 0 correspond to independence. In the sequel, we shall assume that
PJt
H(j) = H > 0 is constant, so that j=1 H(j) = Jt H. Following [Hull and White (2008)], we shall use the
. R t
short-hand notation J = Jt = s=0 dJs when no ambiguity is possible about the time-value at which the
Poisson process is evaluated.
The purpose if this paper is to complete the Hull & White model with analytical expressions for key
quantities. This will avoid iterations for calibration of default probabilities. Further, the numerical integra-
tion of J-conditional expected number of defaults could be avoided in some specific cases with the help of
the closed form expression for the survival distribution being derived below. These results could also be used
to extend the application of the model to baskets with non-homogeneous recoveries as well. Throughout this
paper, special emphasis is put on nth-to-default swaps.
The remaining of this note is organized as follows. First we show that there is a closed-form expression
for the µi such that single-name calibration is ensured. It turns out that the survival probability up to time
t is given by a simple rescaling of that obtained assuming independence. This scaling function, called the
jointure function, is proved to play a key role in the Hull & White model; this is shown in Section 2. In
Section 3, pairwise distributions and conditional distributions are worked out, and the probability to have
simultaneous defaults is emphasized. This will be extended to the N -dimensional case. The distribution
of the nth default time is then computed in Section 4. The first-to-default (FtD) case proves to have an
extremely simple form. Finally, the explicit formula of correlation between default indicators, which we call
default correlation, is given. Its is used to compare the evolution of the default correlation under the static
Gaussian copula with respect to the dynamic models of Hull & White and Mai & Scherer; this is the purpose
of Section 5. Finally we tackle the major problem of this class of pure-jump models, which is the possibility
of facing simultaneous defaults at a time, preventing a reliable handling of cases where baskets are built of
underlyings having different recovery rates. To that end, we propose in Section 6 a “first-order” approach
towards proper dealing of the non-homogeneous case by splitting events where simultaneous defaults happen
or not.
1 One could question this as there is a remaining component in all hazard rate curves, Λ(t). However, this component is

deterministic, so that each conditional survival probability is determined by deterministic hazard rate curve, and hence they
are indeed independent. See for example [Papageorgiou and Sircar (2008)] for more details.

2
 Available at www.defaultrisk.com

2 Calibration of the µi
In order for our model survival probabilities to be in line with the values got from CDS market, it is needed
.
that the expected value of the conditional survival probability Si (t|J) = Ŝi (t|J) matches Si (t):

Si (t) = E[Si (t|J)] . (2)

. Rt . −M (t)−
P J
H(j)
Hence, with Mi (t) = s=0 µi (s)ds, and Si (t|J) = e i j=1 , we have

e−Mi (t) E e−JH

 
Si (t) =
= e−Mi (t) φt (−H) , (3)
.
where φt (u) = E[euJ ] is the moment generating function of J = Jt . This function will be very helpful in the
sequel. It admits the handy closed-form expression
u
−1)
φt (u) = eΛ(t)(e .

With this in hands, equating R.H.S. of Eq. 1 with that of Eq. 3 yields

Mi (t) log(φt (−H)) − log(Si (t))

=
dMi (t)
µi (t) =
dt
−H dΛ(t) d log(Si (t))
= (e −1) − .
dt dt
dΛ(t)
Using dt = λ(t) and λi (t) = fi (t)/Si (t) with fi (t) = −dSi (t)/dt, one gets the calibration equation

µi (t) = λ(t)(e−H −1) + λi (t) ⇔ λi (t) = µi (t) − λ(t)(e−H −1) . (4)

. Rt
The above expression leads to “auto-calibrated” equations. Indeed, with Λi (t) = s=0 λi (s)ds, replacing
Mi (t) by Λ(t)(e−H −1) + Λi (t) will automatically ensure that Eq. 4 holds.
As a corollary, we have that
" n
Y
#
h n P i n
Y
E Si (t|J) = E e− i=1 (Mi (t)+HJ) = ψ(n, H, Λ(t)) Si (t) , (5)
i=1 i=1

where we have used

. −H
ψ(n, H, Λ(t)) = φt (−nH) e−nΛ(t)(e −1)
 
Λ(t) (e−nH −1)−n(e−H −1)
= e . (6)

Note that the L.H.S. of Eq. 5 is nothing but the probability thatQthe names 1, . . . , n all survive up to
. n
time t, as given in the Hull & White model. Therefore, with S ⊥ (t) = i=1 Si (t) to denote the joint survival
probability up to time t under independence :

S(t) = Pr[τi > t . . . , τn > t]

Yn
= ψ(n, H, Λ(t)) Pr[τi > t]
i=1
⊥
= ψ(n, H, Λ(t))S (t) . (7)

3
 Available at www.defaultrisk.com

.
In the following, we shall make use of short-hand notation ψ(n) = ψ(n, H, Λ(t)) and call ψ the jointure
function for obvious reasons.2
Lemma 2.1 [Porperties of the jointure function]
The jointure function satisfies3
ψ(n, 0, Λ(t)) = ψ(n, H, 0) = 1
and
dψ(n, H, Λ(t))  
= ψ(n, H, Λ(t)) log ψ(n, H, λ(t)) .
dt
Further, if H > 0, Λ(t) > 0, then ψ(n, H, Λ(t)) > 1 for all n > 1.
The proof of this lemma is skipped due to its simplicity (note : last result can be proven by induction).

3 Some results for the N = 2 case

It is informative to work out some results in the two-dimensional case; they will be useful for analyzing
pairwise default correlation for instance. We first compute the joint distribution
Fij (x, y) = Pr[τi 6 x, τj 6 y] .
This will allow us to check that simultaneous defaults are possible in that model, that is that Pr[τi = t|τj =
t] > 0. Further, this is useful to compute correlation between the default indicators Ai (t), Aj (t) where
.
Ai (t) = 1I{τi 6t}
is the random variable indicating whether name i defaulted by time t or not.

3.1 Distribution
We suppose first that y > x and that calibration of the (µi (s))06s6t is achieved via Eq. 4.
h x R  y R i
Fij (x, y) = E 1 − e− 0 µi (s)ds−Jx H 1 − e− 0 µj (s)ds−Jy H
Jy =Jx +(Jy −Jx )
h i
= 1 − Si (x) − Sj (y) + e−Mi (x)−Mj (y) E e−2Jx H−(Jy −Jx )H .

Under the additional assumption that the jump intensity is constant over time, λ(t) = λ (hence Λ(t) = λt),
Jt becomes a homogeneous Poisson process and it comes that
Jy − Jx ∼ Jy−x , Jy−x ⊥ Jx ,
where X ⊥ Y has to be understood as (X, Y ) being independent. Hence,
h i
E e−2Jx H−(Jy −Jx )H = E e−2HJx E e−HJy−x
   

= φx (−2H)φy−x (−H) .
Consequently, extending this result to the case x 6 y one gets
−H
−1)((x∧y)(e−H +1)+|y−x|)
Fij (x, y) = 1 − Si (x) − Sj (y) + e−Mi (x)−Mj (y) eλ(e .
2 The above structure is close to that given by Sibuya [Sibuya (1960)], where the joint distribution H(x, y) is modeled as
the product of the marginals F (G) of X (Y ) (evaluated at x and y respectively) times a function Ω(F (x), G(y)) usually called
Sibuya function. It is different, though, as i) here we are working with survival distributions instead of cumulative ones and
ii) our jointure function is of the form Ω? (x, y) instead of Ω(F (x), G(y)). Our jointure function precisely corresponds to the
local dependence index as defined by Anderson et al., see e.g. [Anderson et al. (1992)], [Drouet Mari (1999)]. We are grateful
to Prof. Jaap Spreeuw from Cass Business School (CityQn University, London)
Qn
for pointing that out.
3 The H = 0 or Λ(t) = 0 cases corresponds to E
i=1 Si (t|J) = i=1 E[Si (t|J)], that is independence, as argued above

4
 Available at www.defaultrisk.com

3.2 Conditional probability Pr[τi 6 x|τj = y]

. ∂Fij (x,y)
The conditional probability Fi|j (x|y) = Pr[τi 6 x|τj = y] is obtained by computing ∂y /fj (y) where
dSj (y)
fj (y) = − dy is the marginal density function of τj . Straightforward algebraic manipulations yield,
together with the assumption λ(t) = λ,

Fi|j (x|y) = Sj (y)λj (y)

−H
−1)((x∧y)(e−H +1)+|y−x|)
+(−µj (y) + (e−H −1)λχ(y)) e−Mi (x)−Mj (y)+λ(e

with
e−H

. (y < x)
χ(y) = .
1 (y > x)
Hence, there is a discontinuity in this CDF at x = y provided that both H > 0, λ > 0, showing that there
is a non-zero probability that τi = x given τj = x when there is a positive probability to have downwards
jumps in the survival distribution. In other words, simultaneous defaults are indeed possible in the case τi
and τj are not independent.

3.3 Positive quadrant dependency

It can be shown (see Lemma 6.2 below) that

S(t1 , t2 ) = S1 (t1 )S2 (t2 )ψ(2, H, t1 ∧ t2 ) .

Because ψ > 1, the model exhibits positive quadrant dependency (see e.g. [Drouet Mari (1999)] or
[Cherubini et al. (2006)]), and hence (Pearson’s) correlation is always positive. Also, Bayes’ rule implies

Pr[τ1 > t1 |τ2 > t2 ] > Pr[τ1 > t1 ] .

In other words, the information about the default time of name 2 arriving after some time t2 < ∞ tends
increase the probability of name 1 defaulting after t1 . This is an interesting feature as it mimics the impact
of good news propagation : knowing that a company survives in the near future tends to increase the
probability that another firm could also survive in a near future as well.

4 Distribution of n-th to default time

.
Let τ (1) = argminj τj be the random time of the first default. Equation Eq. 7 directly yields

N
Y
Pr[τ (1) > t] = S(t) = ψ(N ) Si (t) . (8)
i=1

This result is extremely appealing : under this model the joint survival distribution of N names S(t, . . . , t)
up to time t is just the one obtained by independence (product of individual survival distributions) adjusted
by a correction factor, only depending of the jump size H and of time via the deterministic function λ(t)
(observe that there is no assumption about the “time-independency” of λ(s) here) .
This is useful to price FtD swaps when all the names have the same loss-given-default (i.e. notionals and
recoveries are not name-specific) as in this case this information combined to payment calendar and discount
rates is enough to value the swap; see e.g. [Laurent and Gregory (2003)].
The more general n-th to default case (NtD, 1 6 n 6 N , where N is the number of constituents in the
basket) is also available in closed form and tractable for small N (as it is typically the case for NtD). First,

5
 Available at www.defaultrisk.com

we observe that if τ (1) 6 τ (2) 6 . . . 6 τ (N ) is the ordered list of the default times, it holds
n−1
"N # n−1
X X X
(n)
Pr[τ > t] = Pr Ai (t) = j = Pr [N (t) = j]
j=0 i=1 j=0

. PN
with N (t) = i=1 Ai (t) the number of defaults up to time t.
Taking into account all the possibilities that exactly n out of N names defaulted prior to t and exploiting
conditional independence :
X
Pr [N (t) = n|J] = Pr[τi1 6 t, ·, 6 τin 6 t, τj1 > t, ·, τjN −n > t|J]
16i1 <·<in 6N
{i1 ,·,in ,j1 ,·,jN −n }={1,·,N }

X n 
Y −n
 NY
= 1 − Sik (t|J) Sjk0 (t|J) .
16i1 <·<in 6N k=1 k0 =1
{i1 ,·,in ,j1 ,·,jN −n }={1,·,N }

The interpretation of the sum subscript above is that we partition the set {1, . . . , n} in two subsets of size
n and N − n, respectively. The elements of the first set ({i1 , . . . , in }) are assumed to have defaulted prior
to t, not the others. The choice of the members of the first set uniquely determined those of the other set
({j1 , . . . , jN −n }). Finally, the condition {1 6 i1 < · · · < in 6 N } avoids counting twice the same events in
terms of default indicators, as the permutation of elements in a same subset leads to the same partitioning.
Note that we need to define
0
Y .
Sjk0 (t|J) = 1 (9)
k0 =1

to obtain the correct result in the case n = N . The expansion of the above product of default probabilities
can be dealt with by applying the result
n
Y n
X X P l
(1 − exi ) = 1 + (−1)l e− z=1 xjz
.
i=1 l=1 16j1 <...<jl 6n

with xi ← −Λi (t) − JH (see [Mai and Scherer (2009)]). Hence,

"N −n #
X Y
Pr [N (t) = n] = E Sjk0 (t|J)
16i1 <·<in 6N k0 =1
{i1 ,·,in ,j1 ,·,jN −n }={1,·,N }
−n
n
" l N
#
X X Y Y
l
+ (−1) E Simz (t|J) Sjk0 (t|J)
l=1 16m1 <·<ml 6n z=1 k0 =1
N −n n
Eq. 5 X Y
= Sjk0 (t) × ψ(N − n)
16i1 <·<in 6N k0 =1
{i1 ,·,in ,j1 ,·,jN −n }={1,·,N }
n
X X l
Y o
+ (−1)l ψ(N + l − n) Simz (t) . (10)
l=1 ˙ l 6n
16m1 <<m z=1

Observe that calibration to CDS probabilities is already included in these equations.

6
 Available at www.defaultrisk.com

5 Correlation of default indicators

In this section, we first derive the pairwise default correlations, defined as Corr(Ai (t), Aj (t)), the correlation
between default indicators, in the Hull & White model. We then compare the shapes under the three various
models mentioned in the introduction.

5.1 Default correlations under the Hull & White, the Gaussian copula and the
Mai-Scherer models
The covariance between indicator functions is :


Cov 1I{τi 6t} , 1I{τj 6t} = E[1I{τi 6t} 1I{τj 6t} ] − E[1I{τi 6t} ] E[1I{τi 6t} ]
= Fij (t, t) − Fi (t)Fj (t)
 
= 1 − Si (t) − Sj (t) + Sij (t, t) − (1 − Si (t))(1 − Sj (t))
= Sij (t, t) − Si (t)Sj (t) .

The variance of 1I{τi 6t} is given by

E[1I2{τi 6t} ] − E2 [1I{τi 6t} ] = Fi (t) − Fi2 (t) = Fi (t)Si (t) .

So that,
Fij (t, t) − Fi (t)Fj (t) Sij (t, t) − Si (t)Sj (t)
Corr(Ai (t), Aj (t)) = p p =p p . (11)
Fi (t)Fj (t) Si (t)Sj (t) Fi (t)Fj (t) Si (t)Sj (t)

5.1.1 Hull & White

For the Hull & White model, Eq. 7 gives Sij (t, t) = ψ(2)Si (t)Sj (t) and the correlation is given by the simple
expression
s
Si (t)Sj (t)
Corr(Ai (t), Aj (t)) = (ψ(2) − 1) . (12)
Fi (t)Fj (t)
We observe that at first sight, the Hull & White model can potentially leads to correlation higher than
1 when the default probability are taken sufficiently small and for high jump size H occurring with high
likelihood (λ(s) high). However, this cannot happen if we constraint Mi (t) > 0. This is proven in the
appendix.

5.1.2 Gaussian Copula

The default probability is here given by the bivariate normal distribution [Li (1999)] :
 √ 
Fij (t, t) = Φ2 Φ[−1] ◦ Fi (t), Φ[−1] ◦ Fj (t), ρ ,

so that,  √ 
Φ2 Φ[−1] ◦ Fi (t), Φ[−1] ◦ Fj (t), ρ − Fi (t)Fj (t)
Corr(Ai (t), Aj (t)) = p p . (13)
Fi (t)Fj (t) Si (t)Sj (t)

7
 Available at www.defaultrisk.com

5.1.3 Mai and Scherer

This model assumes that i defaulted prior to time t if a stochastic process ζλi (t) reaches a random threshold
Ei [Mai and Scherer (2009)] :
Ai (t) = 1I{τi 6t} = 1I{ζλi (t) >Ei } .
More precisely, Ei ∼ exp(1), i ∈ {1, . . . , n} are i.i.d. random default thresholds, and ζt is a Lévy subordina-
tor4 with Laplace exponent Ψ independent of the Ei ’s : E[esζt ] = etΨ(s) (Recall that the Laplace exponent
Ψ(x) is nothing but the Characteristic exponent of the process (Xs )s>0 evaluated at −ix : Ψ(x) = Ψ̃(−ix),
. . PNt
Ψ̃(x) = log(E[eiuX1 ]), see e.g. [Cont and Tankov (2004)])5 . When ζt = µt + l=1 Gl where Nt is a Poisson
process with intensity β and Gl are i.i.d. exp(η) random variables, ζt is a compound Poisson process. In
this case, we have
Ψ(s) = µs + βs/(η − s) . (14)
The calibration with respect to default probabilities is achieved via the constraint Ψ(−1) = −1, which
uniquely determines µ as being equal to 1 + β/(1 + η) (this is the equivalent of Eq. 4 in the Hull & White
model). In the sequel, we fix β = 1, which allows η to be chosen arbitrarily in R+ . Finally, the default
correlation is given by
  
1 − (Fi (t) ∧ Fj (t))−(Ψ(−2)+1) 1 − (Fi (t) ∨ Fj (t))
Corr(Ai (t), Aj (t)) = p
Fi (t)Fj (t)Si (t)Sj (t)
Fi (t) + Fj (t) − 1 − Fi (t)Fj (t)
+ p (15)
Fi (t)Fj (t)Si (t)Sj (t)

and the survival probability by

Pr[τi > t, τj > t] = Sij (t, t) = (Si (t) ∧ Sj (t))(Si (t) ∨ Sj (t))Ψ(−1)−Ψ(−2) .

5.2 Default correlation comparison

Very often, models are compared based on market-implied quantities. For instance, we compare time-stability
of market-implied parameters or how good is the fit of models when applied to multiple tranches. We believe
this equation is also useful for model comparison when we are lacking market data, as it is e.g. the case
regarding NtD swaps. For instance, suppose that we don’t know the price of the FtD because of a lack of
liquidity. But suppose that analysts are confident in the survival probability by time T . In such a case,
it makes sense to compare models under the constraint that they have the same survival probability up
to T . For instance, we could compare correlation of default indicators coming from different models under
constraint that they have the same time-T survival probability. In the sequel, we shall do this exercise to
compare the Gaussian Copula, the Hull & White and the Mai-Scherer model with compound Poisson process
chosen as Lévy subordinator, that is with Laplace exponent given by Eq. 14.

5.2.1 Two-dimensional case

In the simulation below, we fix H, λ(t) = λ, λi (t) = λi and λj (t) = λj with T = 5. We compute the survival
probability by time T , Pr[τi > t, τj > t], as given by the Hull & White model using ψ(2) e−t(λi +λj ) , and
calibrate the loading factor ρ of the Gaussian copula and the rate parameter η of the Mai & Scherer model
to obtain the same probability S(T ), that is we look for ρ and η such that the S(T ) are equal for all three
models.6
4 Recall that a Lévy subordinator is nothing but a non-decreasing Lévy process.
5 This .
directly results from Lévy processes theory : E[esζt ] = (Φζ1 (−is))t = (eΨ(s) )t = etΨ(s) since Ψ(s) = log(Φζ1 (−is)) is
the log of the ζ1 -characteristic function evaluated at (−is).
6 In the two-dimensional case, calibrating on F (t, t) or S (t, t) is equivalent.
ij ij

8
 Available at www.defaultrisk.com

Model Margins H&W GC M&S

Case λi (t) λj (t) H λ(t) ρ β η
1 0.05 0.05 5 0.01 57% 1 1.68
2 0.05 0.02 0.01 0.82 16 % 1 9
3 0.001 0.001 10 0.001 99.5 % 1 0.05%
.
Table 1: Testing cases for model comparison leading to a same survival probability Sij (t) = Pr[τi > t, τj > t]
at maturity t = T together with the implied loading factor of the Gaussian copula ρ. We also give the implied
rate η for the Mai-Scherer model when a compound Poisson process with rate 1 and threshold distributed
as exp(η) is chosen as the Lévy subordinator.

Equipped with this calibration procedure, we can now have a look at the evolution of the survival
probability as a funciton of t and at the correlation Corr(Ai (t), Aj (t)) with respect to time for the three
models. This exercise is performed on the three various test cases described in Table 1 and is illustrated in
Figure 1.
We note S HW (t), S GC (t) and S M S (t) the survival probabilities up to time t, that is the generated Sij (t, t)
for each model and define Err(t) = S HW (t) − S GC (t) and Err(t) = S HW (t) − S M S (t). The first interesting
phenomenon to observe is that the survival probability curves are very similar. This means that calibration
of the curves at maturity gives a pretty good fit for all t 6 T as well. This tends to indicate that all models
behave similarly with respect to time. Another quantity of interest is the default correlation. One can see
that contrarily to the Gaussian copula, both jump-based models tend to produce a default correlation being
much more constant over time. Further, in the Gaussian case, the short-term default correlation tends to
zero when so does t; for very short maturities, the correlation between default indicators tends to zero. This
is not the case for the Hull & White model with the proposed calibration scheme. Same holds true for the
Mai & Scherer model. Finally, another key difference is that the Hull & White model allows to reach much
higher levels of default correlation; this is emphasized in Case 3, where a value of ρ = 99.5% only leads to a
default correlation less than 85%. By contrast, the Hull & White model with the same survival probability
at maturity allows to have default correlations up to 100%. Regarding the Mai & Scherer model, even if
the absolute levels of default correlation do not exactly match those of the Hull & White models, they are
shown to be rather close, and exhibit the same slowly decreasing behaviour as time passes.

5.2.2 N -dimensional case

As previously stated, the calibration of the model parameters in order to obtain a same survival probability
at maturity Pr[τ (1) > T ] for each model will lead to the same prices for FtD swaps with basket built of two
constituents. This is because of the joint effect that i) FtD prices are more dependent upon the presence of
a default or not before maturity than to the exact defaulting time, and ii) the implied survival probability
curves are very similar for 0 < t < T . It is instructive, however, to have a look at “last to default”-related
quantities coming out of the three models under the constraint that they all lead to a same value of the
survival probability S(T ) = Pr[τ (1) > T ] for larger baskets. Note that this is just the survival probability
evaluated at (t1 , . . . , tn ) = (T, . . . , T ). This quantity could not be easily obtained for the Gaussian copula
as the last model rather provides the distribution function. However, the above no-default probability can
be readily obtained via the integral
Z ∞ Y N
!!
GC Φ[−1] (Fi (t)) − ρZ
S (T ) = 1−Φ p dΦ(z)
z=−∞ i=1 1 − ρ2

which could be efficiently evaluated with the help of some quadrature.

These probability curves can be obtained as follows, for the One factor Gaussian Copula, Hull & White

9
 Available at www.defaultrisk.com

and Mai & Scherer models :

N
!
∞
Φ[−1] (Fi (t)) − ρz
Z Y
GC
Pr[N (t) = N ] = Φ p dΦ(z)
z=−∞ i=1 1 − ρ2
N l
Eq. 9 & 10 X X Y
Pr[N HW (t) = N ] = 1+ (−1)l ψ(l) Sjz (t)
l=1 16j1 <...<jl 6N z=1
N
X X
Pr[N M S (t) = N ] = 1+ (−1)l Clν (Sj1 (t), . . . , Sjl (t))
l=1 16j1 <...<jl 6N

where we have used

l
Ψ(−(i−1))−Ψ(−i)
Y
(x1 , . . . , xl ) 7→ Clν (x1 , . . . , xl ) = x(1) x(i)
i=2

with x(1) 6 x(2) 6 . . . 6 x(N ) is the ordered list of {x1 , . . . , xN }.

One can see on Figure 2 that the impact for that “tranche” may be far from being negligible. We
assumed a 5-names basket (N =5) with default intensity vector (5.17%, 8.2%, 6.87%, 5.4%, 9.7%). We have
chosen (H, λ) = (10, 1%) yielding Pr[N (T ) = 0] = 20.9%. The implied values for the two other models are
44.5% (Gaussian correlation) and (β, η) = (1, 2.86). Survival probability curves are shown in Panel 2(a)
and converges to the same value at maturity, which shows that proper calibration is achieved. Panel 2(b)
shows the difference between Hull & White curve compared to the others, Err(t) = S HW (t)−S · (t). Pairwise
default correlations are shown in Panel 2(c). Finally, one can see in Panel 2(d) that the model-implied curves
Pr[τ (N ) < T ] are, in spite of the calibration, very different. Both jump-based models rend extreme scenario
much more likely than the Gaussian copula one. The Hull & White model gives the maximum probability
for this “catastrophe” scenario to happen.

6 A route for pricing non-homogeneous recovery FtD baskets

The major drawback of jump-based models like that of Mai & Scherer and of Hull & White for instance,
is that they do not allow to price baskets which have name-specific recoveries. Indeed, in this case, the
contingent leg of a n-th to default basket is given in the literature by

T N P
Pr[τi ∈ [t, t + ds), Aj (t) = n − 1]
Z
j6=i
X
δ(t) (1 − Ri ) lim dt (16)
t=0 i=1
ds→0 ds

where δ(t) is the discount factor, N is the number of constituents in the basket, Ri is the recovery rate of
entity i (i ∈ {1, . . . , N }), T is the maturity of the deal and the above probability is that of the i-th name
causing the n-th default in the interval [t, t + ds). There is a major assumption encapsulated in the above
formula : it only allows one name to default at a time. This requirement is met for continuous copula models,
like the Gaussian copula (see for example [Laurent and Gregory (2003)]). This is not true in the present
case, however, as simultaneous are proven to be possible (see Section 3.2). This becomes non-negligible as
the size of the jumps or the rate of them become moderately high.
In order to see this, consider the following lemma (for which a proof is given in the Appendix) :
Lemma 6.1 (Densities of first default and conditional density of τi given J) The density of the first
default time τ (1) is given by
N
! N
!
X X
(1) ⊥
f (t) = λi (t) − log(ψ(N, H, λ(t))) ψ(N, H, Λ(t))S (t) = λi (t) − log(ψ(N, H, λ(t))) S(t)
i=1 i=1

10
 Available at www.defaultrisk.com

If further the intensity of the jumps is constant (λ(t) = λ), the density of τi conditional on J = j is given by
−H
fi (t|j) = λi (t)Si (t) e−Λ(t)(e −1)−jH

Equation Eq. 16 would hold if the following is true

N P
(1) ?
X Pr[τi ∈ [t, t + ds), j6=i Aj (t) = 0]
f (t) = lim (17)
i=1
ds→0 ds

Indeed, also in the specific case n = 1, we need Eq. 16 to be in-line with the homogeneous expression when
Ri = R, which is
Z T
CL(R) = (1 − R) δ(t)f (1) (t)dt (18)
t=0

We now compare the L.H.S. of Eq. 17 with that of the R.H.S. with the help of Lemma 6.1. It comes that
using conditional independence, the R.H.S is

P
 h P i
N
X Pr τi ∈ [t, t + ds), j6=i Aj (t) = 0|J ⊥
XN h i
E  lim  = E fi (t|J) e− j6=i Mj (t)−(N −1)JH

i=1
ds→0 ds i=1
N
X h P i
= E fi (t|J) e− j6=i Mj (t)−(N −1)JH

i=1
N
X −H
= S ⊥ (t) λi (t) e−N Λ(t)(e −1)
E e−N JH
 
i=1
N
X −H
= S ⊥ (t) λi (t) e−N Λ(t)(e −1)
φt (−N H)
i=1
N
!
X
= λi (t) S ⊥ (t)ψ(N )
i=1
N
!
X
= λi (t) S(t)
i=1

This shows that Eq. 17 does not hold when ψ(N, H, λ) 6= 1, which precisely the case for which independence
is not achieved provided that N > 1 (this condition is assumed to be met as we are dealing with baskets
including more than one single entity). The impact of multiple defaults can be evaluated by checking how
far is the following ratio to 1:
PN
i=1 λi (t) − log(ψ(N, H, λ))
PN .
i=1 λi (t)
Properties of jointure function show that this impact decreases as size of the jump decreases or as rate of
jump arrivals decreases. This is very intuitive as then the impact of jumps decreased, and we converge to
the independent case, given by the product copula.
It is possible, however, to try to go one step further in the context of the Hull & White model. In order
to cope with that, we propose to split the contingent leg in two parts. The first one deals with the events
where only one default is allowed at a time, and the second part deals with cases where multiple simultaneous
defaults are allowed. In the first part, there will be no ambiguity about which recovery applies, and a formula

11
 Available at www.defaultrisk.com

similar to Eq. 16 could be used to value this part. In the second case, however, there will be an ambiguity,
and some assumptions need to be made. In that case, one could decide to take the average of recovery rates,
or to adopt a more conservative approach, for example taking the smallest or largest recovery (depending
on the position we have); we note the chosen value for this recovery rate upon multiple defaults as R̂.
The remaining is based on the two following lemmas, proven in the Appendix:

Lemma 6.2 (Survival distribution in n dimensions) Under the assumption of homogeneous jumps, that
is λ(t) = λ, the survival probability Pr[τ1 > t1 , . . . , τN > tN ] under the Hull & White model with constant
jump size is given by
N
Y  
S(t1 , . . . , tN ) = Si (ti )ψ (N − i + 1), H, λ(t(i) − t(i−1) )
i=1

.
where 0 = t(0) 6 t(1) 6 t(2) 6 . . . 6 t(N ) is the ordered version of {t1 , . . . ,tN }.
In the case t1 = . . . = tN = t, the jointure function is ψ N, H, λt (if i = 1) or 1 (otherwise) and
therefore:
 N
Y
S(t, . . . , t) = ψ N, H, λt Si (t)
i=1

where one recognizes Eq. 5.

Lemma 6.3 (Densities of i-th default exclusively) Under the assumption of homogeneous jumps, the
probability that name i defaults in [t, t + dt) and that all the remaining names survive up to t + dt is given by
Pn o
t , t+dt, . . . , t+dt)−S(t+dt) = S(t) e− j6=i λj (t)dt ψ(N − 1, H, λdt) − ψ(N, H, λdt) e−λi (t)dt
S(t+dt, . . . , t+dt, |{z}
i

so that the associated density obtained by taking the limit of the ratio of the above difference by dt is given
by
  
? . S(t + dt, . . . , t + dt, t, t + dt, . . . , t + dt) − S(t + dt) ψ(N − 1, H, λ)
fi (t) = lim = S(t) λi (t) + log .
dt↓0 dt ψ(N, H, λ)

The requirement for not to have multiple names defaulting at the same of the first one to default mathe-
matically translates to dN (τ (1) ) = 1 : at the time of the first default, the total number of defaults increases
by one. As a corollary,
N Z T
X
Pr[τ (1) 6 T, dN (τ (1) ) = 1] = fi? (t)dt .
i=1 t=0

Consequently, one could evaluate the probability to have one default before time T which happened
simultaneousley that at least one other obligor. Indeed, noting that
h i h i h i
Pr τ (1) 6 T = Pr τ (1) 6 T, dN (τ (1) ) = 1 + Pr τ (1) 6 T, dN (τ (1) ) > 1 ,

and combining this with Lemma 6.1 yields

N
!
h i Z T X
Pr τ (1) (1)
6 T, dN (τ ) > 1 = f (1) (t) − fi? (t) dt .
t=0 i=1

12
 Available at www.defaultrisk.com

P 
N
(H, λ) Pr[τ (1) 6 T ] Pr[τ (1) 6 T, Ω] Pr[τ (1) 6 T, Ω̄] log(ψ(N, H, λ))/ i=1 λi (t)
(0,1%) 22.12 22.12 0 0
(10,0.1%) 20.55 20.1 0.45 8
(10,1%) 4.878 0.001 4.877 80

Table 2: Basket of N = 5 names with idiosynchratic rates λi = 1%; values are expressed in [%]. The
. .
shorthand notations Ω = {dN (τ (1) ) = 1} and Ω̄ = {dN (τ (1) ) > 1]} have been used.

Based on these results, and supposing λi (t) = λi 7 , we can turn the “homogeneous” contingent leg given in
Eq.18 into
N
X Z T N Z
X T  
CL(Ri ) = (1 − Ri ) δ(t)fi? (t)dt + (1 − R̂) δ(t) f (1) (t) − fi? (t) dt
i=1 t=0 i=1 t=0
N    Z T
X ψ(N − 1, H, λ)
= (1 − Ri ) λi + log δ(t)S(t)dt
i=1
ψ(N, H, λ) t=0
 Z T
+(1 − R̂) (N − 1) log(ψ(N, H, λ)) − N log(ψ(N − 1, H, λ)) δ(t)S(t)dt
t=0
N   
nX ψ(N − 1, H, λ)
= (1 − Ri ) λi + log
i=1
ψ(N, H, λ)
 o Z T
+(1 − R̂) (N − 1) log(ψ(N, H, λ)) − N log(ψ(N − 1, H, λ)) × δ(t)S(t)dt (19)
t=0

Obviously, the premium leg of the first-to-default does not depend on the recoveries, so it is exclusively
dependent upon the rates (or discount factors) and the density of τ (1) , i.e. f (1) (t).
Observe that in the case of homogeneous recoveries, Ri = R, Eq. 19 will be in line with Eq. 18 if and
only if R̂ = R, which is obviously the case if R̂ is defined using percentile-based approach on the vector
(R1 , . . . , RN ).
The above development shows that one could not use formula given by Eq. 16 to evaluate the default
probability of a first to default swaps. The error made by using this expression vanishes when H = 0 or
Λ(t) = 0, but increases as we deviate from this “independent” situation. This is illustrated in Table 2.

7 Conclusion
In this note, we have worked out the Hull & White dynamic model for pricing structured credit deriva-
tive [Hull and White (2008)]. It is shown that in the constant-jump size case H(j) = H, it turns out that
calibration to underlying single names is automatically achieved thanks to closed-form expressions, just as for
the barrier levels in the Gaussian copula case. Further, it is shown that the distribution of the n-th default
time is also available in closed form, which leads to an extremely easy pricing of homogeneous first-to-default
swaps, and remains tractable for typical N > 1. Correlation between default indicators are available analyti-
cally as well, which allows to derive some procedure in order to properly compare outcomes of various models.
Compared to the Gaussian copula, for instance, the Hull & White model is proven to be able to yield higher
default correlation levels that are non-vanishing as t → 0. Survival distribution is also derived analytically,
which allows model comparison of basket default swaps under calibration of the survival probability of the
first default time at some point in time. This reveals another specificity of the model, which is to lead
7 This is assumption is just made to allow us factorizing λ (t) = λ in front of the integral, but the more general case where
i i
the name-specific hazard rate curves are time-dependent can be dealt with easily by leaving it inside the integral.

13
 Available at www.defaultrisk.com

to a higher likelihood of extreme scenario even when calibration is performed following the aforementioned
procedure. This is shared by the other jump-based model, and is a fundamental difference compared to the
Gaussian copula. Finally, we dropped the restrictive constraint of the homogeneity of the recovery rates
(typically needed for models allowing for simultaneous defaults) by developing a “first-order approach”, in
which simultaneous default cases are treated separately than those where the default is isolated.

8 Appendix
8.1 Default correlation range in Hull & White model
−H 2
First, we observe that ψ(2) = eΛ(t)(e −1) > 1. This shows that the correlation between the default
indicators up to t is positive (the model cannot generate negative correlations). In order to show that
correlation is lower than one, we first work out the case λi (s) = λj (s). Then, correlation ρij < 1 reduces to
ψ(2)Si < 1, and the last inequality is proven when log(ψ(2)) < Λi (t). The last equality is readily seen to
hold :

log(ψ(2)) = Λ(t)(e−H −1) e−H −Λ(t)(e−H −1)

Eq. 4
= Λ(t)(e−H −1) e−H −(Mi (t) − Λi (t))
log(ψ(2)) − Λi (t) = Λ(t) (e−H −1) e|{z}
−H
− Mi (t)
|{z} | {z } | {z }
>0 <0 >0 >0
< 0

As a corollary we have that (ψ(2) − 1) FSii < 1. Extending this to the general case Λi (t), Λj (t), we first
  
note that it is sufficient to show that (ψ(2) − 1)Si /Fi (ψ(2) − 1)Sj /Fj < 1. Because each factor is lower
than one as shown above, this concludes the proof.

8.2 Proof of Lemma 6.1

In order to prove the first result, we note that Pr[τ (1) > t] = S(t) = ψ(N, H, Λ(t))S ⊥ (t), so that the density
is given by
 
d ψ(N, H, Λ(t)) S ⊥ (t)
f (1) (t) = −
dt
dψ(N, H, Λ(t)) ⊥ dS ⊥ (t)
= − S (t) − ψ(N, H, Λ(t))
dt dt
and the result comes by noting that

dψ(N, H, Λ(t))  
= ψ(N, H, Λ(t)) log ψ(N, H, λ(t)) . (20)
dt
The second result is easy to prove too when the Poisson process Jt is homogeneous. In this case, the survival
probability of name i up to time t + dt is, for infinitesimal dt ↓ 0

Si (t + dt|J) = e−Mi (t) e−µi (t)dt e−Jt H E[e−(Jt+dt −Jt )H |Jt ]

= Si (t|J) e−µi (t)dt φdt (−H)
−H
−1) λ(t)dt(e−H −1)
= Si (t|J) e−λi (t)dt e−λ(t)dt(e e
= Si (t|J) e−λi (t)dt ,

14
 Available at www.defaultrisk.com

so that
Pr[τi ∈ [t, t + dt)|J] = Si (t|J) − Si (t + dt|J) = Si (t|J)(1 − e−λi (t)dt ) .
Using the Taylor expansion of the exponential
eu.dt = 1 + udt + u2 /2dt2 + o(dt2 ) , (21)
the conditional density is given by
Pr[τi ∈ [t, t + dt)|J]
fi (t|J) = lim
dt→0 dt
1 − e−λi (t)dt
= Si (t|J) lim
dt→0 dt
= Si (t|J)λi (t) ,
and the claim results from Eq. 4.

8.3 Proof of Lemma 6.2

In order to compute the survival probability function for the Hull & White model, observe first that defining
.
t(0) = 0
N
X N
X N
X
ti = t(i) = N (t(1) − t(0) ) + (N − 1)(t(2) − t(1) ) + . . . + (t(N ) − t(N −1) ) = (N − i + 1)(t(i) − t(i−1) )
i=1 i=1 i=1

and using J0 = 0
N
X N
X N
X
Jti = Jt(i) = (N − i + 1)(Jt(i) − Jt(i−1) ) .
i=1 i=1 i=1
Therefore, using independence of increments of homogeneous Poisson process and calibration equation Eq. 4.,
"N #
Y
−Mi (ti )−Jti H
S(t1 , . . . , tN ) = E e

−
Pi=1
N
h P N
i
= e i=1 Mi (ti )
E e− i=1 Jti H

Λ(t)=λt
N
Y −H P N
h
−H
P N
i=1 (N −i+1)(Jt(i) −Jt(i−1) )
i
= Si (ti ) e−λ(e −1) i=1 ti E e
i=1

⊥
N
Y −H P N
N
Y h
−H(N −i+1)(Jt(i) −Jt(i−1) )
i
= Si (ti ) e−λ(e −1) i=1 ti E e
i=1 i=1
N
Y −H
= Si (ti ) e−λ(e −1)(N −i+1)(t(i) −t(i−1) )
φt(i) −t(i−1) (−(N − i + 1)H)
i=1
N
Y  
= Si (ti )ψ (N − i + 1), H, λ.(t(i) − t(i−1) ) .
i=1

8.4 Proof of Lemma 6.3

We start from the survival probability given in Lemma 6.2, with ti = t and tj6=i = t + dt. In that specific
case, ψ(N − i + 1, H, λ.(t(i) − t(i−1) )) is 1 except for i = 1 and i = 2, where they equal, ψ(N, H, λt) and
ψ(N − 1, H, λdt), respectively. This results from Lemma 2.1. Further, it is trivial to show that
ψ(N, H, λ.(t + dt)) = ψ(N, H, λt)ψ(N, H, λdt)

15
 Available at www.defaultrisk.com

and one gets

t , t + dt, . . . , t + dt) − S(t + dt) =
S(t + dt, . . . , t + dt, |{z}
i
Q  
ψ(N, H, λt) j6=i Sj (t + dt) Si (t)ψ(N − 1, H, λdt) − Si (t + dt)ψ(N, H, λdt)

In order to evaluate the limit of the ratio of the above expression with dt as dt → 0 we use the equivalent
formulation of the numerator given in the lemma, and evaluate the first order expansions of the above
exponentials (see Eq. 21) which directly leads to the claim.

Disclaimer
The views expressed in this work are those of the author, and do not necessarily reflect the position of ING.

References
[Chapovsky et al. (2006)] Chapovsky, A. , Rennie, 1. and Tavares P. (2006) Stochastic Intensity Modeling
for Structured Credit Exotics, Merril Lynch research Report.
[Cont and Tankov (2004)] Cont, R. and Tankov, P. (2004) Financial Modelling With jump Processes, Chap-
man & Hall/CRC.
[Duffie and Garleanu (2001)] Duffie, D. and Garleanu, N. (2001) Risk and Valuation of Collaterized Debt
Obligation, Financial Analyst Journal 57(1).
[Hull and White (2008)] Hull J.C. and White A. (2008) Dynamic Models of Portfolio Credit Risk : A sim-
plified Approach. Journal of Derivatives, 15(4), pp. 9-28.
[Li (1999)] Li, D. (1999) On Default Correlation : A Copula Function Approach. Riskmetrics Working Paper
99-07.
[Mai and Scherer (2009)] Mai, J.F. and Scherer, M. (2009) Pricing k-th to default swaps in a Lévy-time
framework, Journal of Credit Risk 5(3), pp. 55-70.
[Laurent and Gregory (2003)] Laurent, J.-P. and Gregory, J. (2003) Basket Default Swaps, CDO’s and Fac-
tor Copulas, Journal of Risk 7(4), pp.103-122.
[Papageorgiou and Sircar (2008)] Papageorgiou, E. and Sircar, R. (2008) Multiscale Intensity Models and
Name Grouping for Valuation of Multi-Name Credit Derivatives, Research paper, Dpt of Operations
Research & Financial Engineering, Princeton University.
[Schoutens (2003)] Schoutens, W. (2003) Lévy Processes in Finance, Wiley.
[Cherubini et al. (2006)] Cherubini, U., Luciano, E. and Vecchiato, W. (2003) Copula Methods in Finance,
Wiley.
[Sibuya (1960)] Sibuya, M. (1960) Bivariate Extreme Statistics, Annals of the Institute of Statistical Math-
ematics, 11(2) pp. 195-210.
[Anderson et al. (1992)] Anderson, J.E., Louis, T.A., Holm, N.V. and Harvald, B. (1992) Time-dependent
association measures for bivariate survival distributions, Journal of the American Statistical Association,
87(419) pp. 641-650.
[Drouet Mari (1999)] Drouet Mari, D. (1999) La dépendance positive entre deux variables de durées : con-
cepts de dépendance et mesures locales de liason, Revue de Statistique Appliquée, 47(4) pp. 5-24.

16
 Available at www.defaultrisk.com

0.30
0.004

0.25
0.003

Default Indicator Correlation

0.20
● ● ● ● ● ● ● ● ● ●
0.002
Err(t)

0.15
0.001

0.10
0.05
0.000

0.00
0 1 2 3 4 5 0 1 2 3 4 5

t t

(a) Case 1 : Err(t), Sij (T, T ) ≈ 63.8% (b) Case 1 : Default correlation.

0.015
0e+00
−2e−04

Default Indicator Correlation

0.010
−4e−04
Err(t)

−6e−04

0.005
−8e−04

● ● ● ● ● ● ● ● ● ●
−1e−03

0.000

0 1 2 3 4 5 0 1 2 3 4 5

t t

(c) Case 2 : Err(t), Sij (T, T ) ≈ 70.6% (d) Case 2 : Default correlation.
8e−04

1.0

● ● ● ● ● ● ● ● ● ●
0.8
6e−04

Default Indicator Correlation

0.6
4e−04
Err(t)

0.4
2e−04

0.2
0e+00

0.0

0 1 2 3 4 5 0 1 2 3 4 5

t t

(e) Case 3 : Err(t), Sij (T, T ) ≈ 99.5% (f) Case 3 : Default correlation.
17

Figure 1: Err(t) defined as S HW (t) − S GC (t) (green solid) or S HW (t) − S M S (t) (red dashed) and default
correlation Corr(Ai (t), Aj (t)) under various models under constraint that |S HW (T ) − S GC (T )| <  with
T = 5,  = 10E − 3. The legend is Hull & White (blue, dots), Gaussian Copula (green, squares), Mai &
Scherer with compound Poisson process(1, η) as Lévy subordinator (red, solid).
 Available at www.defaultrisk.com

0.00
1.0

●
0.8

−0.01
●
Survival probability

●
0.6

Err(t)

−0.02
●

●
0.4

−0.03
●

●
0.2

0 1 2 3 4 5 0 1 2 3 4 5

t t

(a) Survival probability of first-to-default time (b) Err(t)

0.05

● ● ● ● ● ● ● ● ● ● ●
●
● ●
● ●
● ● ●
● ●
0.04

● ●
0.15

● ● ● ●
● ● ● ● ● ●
● ● ●
● ● ●
● ●
● ● ●
● ● ●
● ● ● ●
● ● ● ●
●
Default Indicator Correlation

● ● ● ● ●
● ● ● ●
● ● ●
● ● ●
● ● ●
● ● ● ●
● ● ●
0.03

● ● ● ●
● ●
● ● ●
●
Prob[N(t)=N]

● ●
● ●
● ●
0.10

● ●
●
● ● ● ● ●
0.02

●
0.05

0.01

●
0.00

0.00

0 1 2 3 4 5 0 1 2 3 4 5

t t

(c) Pairwise default correlations (d) Default probability of last-to-default time

Figure 2: Model impacts on a 5-names basket default swap under calibration of survival probability up to
maturity T = 5. The legend is : Hull & White (blue, dots), Gaussian Copula (green, squares), Mai &
Scherer with compound Poisson process(1, η) as Lévy subordinator (red, solid). On panel 2(a), dashed curve
is survival distribution assuming independence, and we note that the Mai & Scherer curve is almost the
same as that of Hull & White.

18