Dynamic CDO Term Structure Modelling∗

Damir Filipović† Ludger Overbeck‡ Thorsten Schmidt§

first version: November 2007, this version: 19 June 2008

Abstract
This paper provides a unifying approach for valuing contingent claims
on a portfolio of credits, such as collateralized debt obligations (CDOs).
We introduce the defaultable (T, x)-bonds, which pay one if the aggre-
gated loss process in the underlying pool of the CDO has not exceeded
x at maturity T , and zero else. Necessary and sufficient conditions on
the stochastic term structure movements for the absence of arbitrage are
given. Moreover, we show that any exogenous specification of the forward
rate volatility curve actually yields a unique consistent loss process and
thus an arbitrage-free family of (T, x)-bond prices. For the sake of ana-
lytical and computational efficiency we then develop a tractable class of
affine term structure models.

Key words: affine term structure, collateralized debt obligations, loss

process, single tranche CDO, term structure of forward spreads

1 Introduction
Collateralized debt obligations (CDOs) are securities backed by a pool of refer-
ence entities such as bonds, loans or credit default swaps. The reference entities
form the asset side of a CDO-structure. Traded products are notes on the CDO
tranches. They have different seniorities and build the liability side of the CDO.
The most liquidly traded CDOs are those based on so-called indices. In
2004 the CDX in North-America and the Itraxx in Europe have been created
and are nowadays very liquid. Both indices consist of the most liquidly traded
and quoted credit default swaps in the given market, for example the corporate
∗ Filipović is supported by WWTF (Vienna Science and Technology Fund). Thanks to

René Carmona, Philipp Schönbucher, and Gennady Samorodnitsky for helpful discussion.
† Vienna Institute of Finance, University of Vienna, and Vienna University of Economics

and Business Administration, Heiligenstädter Strasse 46-48, A-1190 Wien, Austria. Email:
[email protected]
‡ Department of Mathematics, University of Giessen, Arndtstr. 2, D-35392 Giessen, Ger-

many. Email: [email protected]

§ Department of Mathematics, University of Leipzig, D-04081 Leipzig, Germany. Email:

[email protected]

1
investment grade iTraxx in Europe consists of the 125 most liquid investment
grades corporate credit default swaps. In addition to contingent claims based on
the indices, there are also many options on the market spread of those indices
and single tranche CDOs (STCDOs) (a STCDO is a credit default swap on a
tranche, see Section 6 below for a definition); typically calls and puts. However,
the corresponding products are less liquid and only quoted by a few market
makers. For more background and references we refer to the respective chapters
in [17].
Concerning the valuation of the basic STCDOs, the one-factor Gaussian
copula approach [16] has emerged as the industry standard. It is basically a
static description of the default times on the asset side of the CDO. However,
it is well acknowledged that this approach has a number of deficiencies. First
of all, there are only two parameters: the average default probability and a
correlation parameter, which is a stylized version of Mertons asset correlation.
This model is not able to capture all market quotes on the liability side with
these two parameters. Therefore each tranche can only be priced with a different
correlation, the so-called implied correlation. Moreover, the latest credit crises
has clearly illustrated the limited capability of [16] to appropriately capture the
dependence structure between the single names.
Recently, there have emerged several new attempts on CDO valuation based
on the aggregate loss function (“top-down”), as opposed to the above men-
tioned (“bottom-up”) single tranche default models. Giesecke and Goldberg
[12] decompose the portfolio loss process into single name loss processes. Ben-
nani [1] models, for a fixed maturity T and loss variable LT , the T -Forward
Loss process L(t, T ) = E[LT | Ft ] and the T -Forward Outstanding Notional
ON (t, T ) = 1 − L(t, T ). From there the excess loss E[(L(T, T ) − K)+ ] for strike
K and maturity T can be computed numerically by Monte Carlo methods. How-
ever, this approach focuses on one maturity date T only, and neither market
interest rate and nor spread risk is considered. Schönbucher [18] introduces the
forward loss distributions and finds a Markov chain with the same marginal
distribution as the loss process. Ehlers and Schönbucher [8] extend [18] by
considering non-constant interest rates for pricing. They introduce conditional
forward interest-rates fn (t, T ) and forward protection rates (spreads) Fn (t, T )
given the realization of the loss process L(t) = n. An HJM-type specification of
the loss-contingent forward interest and loss rates fn and Fn is then proposed
and no-arbitrage conditions are given. Ehlers and Schönbucher [9] analyze the
interplay of the background (forward interest and protection rates, say) and
loss process conditional on an increasing sequence of filtrations. However, the
technical analysis in [18, 8, 9] relies on the assumption that the loss process lives
on a finite grid, and their extension to multi-step increments (loss given default
risk) becomes notationally demanding. The paper of Sidenius et al. (SPA) [20]
is closest to our framework. However, SPA assumes zero risk-free rates. More-
over, some crucial problems, e.g. regarding the construction of a consistent loss
process, have remained open in [20] and will be completed from a global point of
view in our paper. A stable non-parametric calibration algorithm has recently
been developed in Cont and Minca [6].

2
 The aim of our paper is to provide a unifying approach for valuing contingent
claims on CDOs, which encompasses the above mentioned and puts them on a
sound mathematical basis. We therefore introduce the defaultable (T, x)-bonds,
which pay one if the aggregated CDO loss process has not exceeded x at maturity
T , and zero else. It turns out that essentially all contingent claims on the
CDO-pool, such as STCDOs, can be written—and thus hedged and priced—as
linear combinations of (T, x)-bonds. We then model the corresponding (T, x)-
forward rates as semimartingales driven by some Brownian motion, reflecting
market information, and the jump measure associated to the loss process. This
setup is most general, and allows for feedback, or contagion effects, from the
loss process on the forward curve. As a first result, we provide necessary and
sufficient conditions for the absence of arbitrage in terms of a drift condition
and a relation between the short end of the spread curve and the prevailing
loss intensity. Most important from a modelling point of view, we then provide
mathematical evidence that arbitrage-free (T, x)-bond models uniquely exist
under general assumptions. This is very much in the spirit of the Heath–Jarrow–
Morton [15] approach to the modelling of the term structure of risk free interest
rates. Risk-neutral pricing has to be done numerically in general. However,
under omittance of the aforementioned contagion effects—which corresponds
to the doubly stochastic setup for single name models—we find an efficient
CDO derivatives pricing formula. For the sake of analytical and computational
efficiency we then develop a tractable class of affine term structure models,
which lead to closed form solutions for STCDO values and swaption prices.
The significance of our approach is its focus on the (T, x)-bonds and their
exogenous stochastic specification. This perspective facilitates the mathemat-
ical analysis and it should also facilitate the empirical estimation for dynamic
CDO term structure modelling, as it is the case for Heath–Jarrow–Morton [15]
type forward rate models. Moreover, to our knowledge, the integrated affine
specification of the (T, x)-term structure developed below is new in the litera-
ture.
The structure of the paper is as follows. In Section 2, we formally introduce
the (T, x)-bonds. In Section 3, we provide necessary and sufficient conditions
for the absence of arbitrage. In Section 4, we give sufficient conditions on
the stochastic basis such that arbitrage-free (T, x)-bond models uniquely exist
under general assumptions. This is then further improved in the more restrictive
doubly stochastic framework in Section 5. In Section 6, we derive STCDO and
swaption price formulas. In Section 7, we provide an affine specification for the
doubly stochastic framework.

2 (T, x)-Bonds
As stochastic basis, we fix a filtered probability space (Ω, F, (Ft ), Q). We assume
that Q is a risk-neutral pricing measure. An equivalent measure change will be
discussed below in Remark 3.5.
Consider a pool of credits (the CDO-pool) with an overall nominal normal-

3
ized to 1, and let I ⊂ [0, 1] denote the set of all attainable loss fractions, i.e.
x ∈ I represents the state where 100x% of the overall nominal has defaulted.
Throughout, we assume that
• either I = {i/n | i = 0, . . . , n} is finite,
• or I = [0, 1],

endowed with the usual point and Borel topology and σ-algebra B(I), respec-
tively. The Lebesgue measure on [0, 1] is denoted dx. To facilitate notation, we
shall write dx also for the uniform counting distribution if I is finite. That is,
for any set A ⊂ [0, 1],
Z X
1
f (x)dx = f (x).
A n+1
i/n∈A

We denote by L the I-valued increasing aggregate CDO-loss process with

L0 = 0. That is, Lt represents the ratio of CDO-losses occurred by time t.
The basic instrument that we consider is a (T, x)-bond which pays 1{LT ≤x}
at maturity T , for x ∈ I. Its price at time t ≤ T is denoted by P (t, T, x).
Obviously, P (t, T, x) is increasing in x and decreasing in T . Since Lt ≤ 1 for all
t, the risk free T -bond price P (t, T ) at time t ≤ T equals

P (t, T ) = P (t, T, 1). (1)

(T, x)-bonds are the fundamental components for the hedging and pricing
of CDO-derivatives. Indeed, any European type contingent claim on the loss
process with (regular enough) payoff function F (LT ) at maturity T can be
decomposed into a linear combination of (T, x)-bonds
Z
F (LT ) = F (1) − F 0 (y)1{LT ≤x} dx.
I

Hence the static portfolio

Z
F (1)P (t, T ) − F 0 (x)P (t, T, x) dx
I

replicates, and thus prices the claim at any time t ≤ T , model independently.
For example, the basic components of the payment legR of the STCDO below in
Section 6 are put options with payoff (K − LT )+ = (0,K] 1{LT ≤x} dx.

3 Arbitrage-free Term Structure Movements

Our aim is to describe the (T, x)-bond price term structure movements explicitly
in the form RT
P (t, T, x) = 1{Lt ≤x} e− t f (t,u,x) du (2)

4
where f (t, T, x) denotes the (T, x)-forward rate prevailing at date t. That is,
f (t, T, x) is the rate that one can contract for at time t, given that Lt ≤ x,
on a defaultable forward investment of one euro that begins at date T and is
returned an instant dT later conditional on LT +dT ≤ x. The following forward
rate agreement replicates the corresponding cash flows:
P (t,T,x)
• at t: sell one (T, x)-bond and buy P (t,T +dT,x) (T + dT, x)-bonds.

This zero net investment at t yields the following future cash flows:
• at T : pay (“invest”) 1{LT ≤x} euro.
P (t,T,x)
• at T + dT : receive P (t,T +dT,x) 1{LT +dT ≤x} euros.

The corresponding continuously compounded forward rate f (t, T, x) is given, in

first order in dT , by
P (t, T, x)
ef (t,T,x) dT ≈ .
P (t, T + dT, x)
In the limit dT → 0, we obtain (2).
Note that, in line with (1),

f (t, T ) = f (t, T, 1) and φ(t, T, x) = f (t, T, x) − f (t, T ) (3)

are the corresponding risk free T -forward rate and (T, x)-forward spread, respec-
tively.
The (T, x)-bond specification (2) combines default risk, 1{Lt ≤x} , and market
risk, f (t, T, x), in one instrument. This is a slight, but crucial difference to the
SPA [20] framework where the market and default risks are specified by a two
layer-process. In what follows we analyze the consistency of the loss process L
and market term structure f (t, T, x) movements in order that the (T, x)-bond
market be free of arbitrage.
As for the loss process, we make the general assumption
P
(A1) Lt = s≤t ∆Ls is an increasing marked point process1 which admits an
absolutely continuous compensator ν(t, dx) dt.
This setup implies totally inaccessible default times of the (T, x)-bonds and
a fundamental relation between their intensity processes and the compensator
ν(t, dx):
Lemma 3.1. Assume that (A1) holds. Then, for any x ∈ I, the indicator
process 1{Lt ≤x} is càdlàg with intensity process

λ(t, x) = ν(t, (x − Lt , 1] ∩ I). (4)

That is, Z t
Mtx = 1{Lt ≤x} + 1{Ls ≤x} λ(s, x) ds (5)
0
1 Also called multivariate point process. For a definition see e.g. [13] or [2].

5
is a martingale. Moreover, λ(t, x) is progressive, decreasing and càdlàg in x ∈ I
with λ(t, 1) = 0.
Conversely, λ(t, x) uniquely determines ν(t, dx) via

ν(t, (0, x]) = λ(t, Lt ) − λ(t, Lt + x), x ∈ I, (6)

where we denote λ(t, x) = 0 for x ≥ 1.

Note that SPA [20] postulate that λ(t, x) exist with the above properties.
With Lemma 3.1 we now put their assumption on a sound mathematical basis.
Proof. Right-continuity of 1{Lt ≤x} follows from the structure (A1) of Lt . By
the very definition of ν(t, dx),
Z tZ
F (Lt ) − (F (Ls + y) − F (Ls ))ν(s, dy) ds (7)
0 I

is a martingale, for any bounded measurable function F . In particular, for

F (Lt ) = 1{Lt ≤x} we have

F (Ls + y) − F (Ls ) = −1{Ls +y>x} 1{Ls ≤x} . (8)

This proves (5). The other properties of λ(t, x) hold by inspection.

Remark 3.2. As can be seen from the proof, Lemma 3.1 actually holds true for
any, not necessarily increasing, marked point process.
We next assume that any (T, x)-forward rate process follows a semimartin-
gale of the form
Z t Z t
f (t, T, x) = f (0, T, x) + a(s, T, x)ds + b(s, T, x)> · dWs
0 0
Z tZ
+ c(s, T, x; y) µ(ds, dy) (9)
0 I

where W is some d-dimensional Brownian motion (market noise) and

X
µ(ω; dt, dx) = 1{∆Ls (ω)6=0} δ(s,∆Ls (ω)) (dt, dx)
s>0

denotes the integer-valued random measure associated to the jumps of L, where

we write δa for the Dirac measure at point a.
This specification is very general2 , and it allows for two kinds of feedback,
or contagion, of the loss process on the forward rates:
(C1) direct, via simultaneous jumps driven by L: ∆f (t, T, x) = c(t, T, x; ∆Lt ).
2 This framework can obviously be further generalized by adding Lévy and/or Poisson

random measure driven market noise to W .

6
(C2) indirect, via letting the model parameters a, b, and c be explicit functions
of the prevailing loss path L (“regime switching”).
We will provide, in Section 4 below, mathematical evidence that CDO term
structure models (9) with properties (C1) and (C2) exist under very general
assumptions.
To assert that the subsequent analysis and formal manipulations be mean-
ingful, we make the following technical assumptions, where O and P denote the
optional and predictable σ-algebra on Ω × R+ , respectively:
(A2) the initial forward curve f (0, T, x) is B(R+ ) ⊗ B(I)-measurable, and lo-
cally integrable:
Z T
|f (0, u, x)| du < ∞ for all (T, x),
0

(A3) the drift parameter a(t, T, x) is R-valued O ⊗ B(R+ ) ⊗ B(I)-measurable,

and locally integrable:
Z T Z T
|a(t, u, x)| dt du < ∞ for all (T, x),
0 0

(A4) the volatility parameter b(t, T, x) is Rd -valued O⊗B(R+ )⊗B(I)-measurable,

and locally bounded:

sup kb(t, u, x)k < ∞ for all (T, x),

t≤u≤T

(A5) the contagion parameter c(t, T, x; y) is R-valued P ⊗B(R+ )⊗B(I)⊗B(I)-

measurable, and locally bounded:

sup |c(t, u, x; y)| < ∞ for all (T, x).

t≤u≤T, y∈I

Conditions (A2)–(A5) assert that the risk free short rate rt = f (t, t) has a
RT
progressive version and satisfies 0 |rt | dt < ∞ for all T , see e.g. [11]. Hence the
Rt
savings account e 0 rs ds is well defined.
It is well known that there exists no admissible3 arbitrage strategy in the
(T, x)-bond market if the discounted price processes
Rt
e− 0
rs ds
P (t, T, x) are local martingales for all (T, x). (10)

We now give necessary and sufficient conditions for (10) to hold.

3 A self-financing trading strategy is admissible if its discounted value process is a super-

martingale, e.g. bounded from below or a martingale.

7
Theorem 3.3. Assume (A1)–(A5) hold. Then the no-arbitrage condition (10)
is equivalent to
Z T °Z °2 Z
1°
°
T °
°
³ RT ´
a(t, u, x) du = ° b(t, u, x) du° + e− t c(t,u,x;y) du − 1 ν(t, dy),
t 2° t ° I
(11)
rt + λ(t, x) = f (t, t, x) (12)
on {Lt ≤ x}, dt ⊗ dQ-a.s. for all (T, x).
Note that (12) is part of the assumptions in the SPA [20] framework, while
here we demonstrate that this is in fact a necessary consequence of the no-
arbitrage condition (10).
Proof. We denote RT
p(t, T, x) = e− t
f (t,u,x) du
(13)
so that P (t, T, x) = 1{Lt ≤x} p(t, T, x). Using a stochastic Fubini argument pro-
posed by Heath et al. [15], see also [11], we transform
Z TZ t Z tZ T Z tZ T Z tZ u
· · · ds du = · · · du ds = · · · du ds − · · · ds du,
t 0 0 t 0 s 0 0

and similarly for dWs du and µ(ds, dy) du. We thus derive by Itô’s formula
½ Z T °Z °2
dp(t, T, x) 1°°
T °
°
= f (t, t, x) − a(t, u, x) du + ° b(t, u, x) du°
p(t, T, x) t 2° t °
Z ³ R ´ ¾
T
+ e− t c(t,u,x;y) du − 1 ν(t, dy) dt
I
Z T
− b(t, u, x)> du · dWt
t
Z ³ RT ´
+ e− t
c(t,u,x;y) du
− 1 (µ(dt, dy) − ν(t, dy) dt) . (14)
I
Rt
Denote Z(t, T, x) = e− 0
rs ds
P (t, T, x). Integrating by parts and using (5) yields
µ ¶
dp(t, T, x)
dZ(t, T, x) = Z(t, T, x) −rt dt + dMtx − λ(t, x)dt + . (15)
p(t, T, x)
Combining (14) and (15) shows that (10) holds if and only if
Z T °Z °2
1°°
T °
°
− rt − λ(t, x) + f (t, t, x) − a(t, u, x) du + ° b(t, u, x) du°
t 2 ° t °
Z ³ R ´
T
+ e− t c(t,u,x;y) du − 1 ν(t, dy) = 0 (16)
I

on {Lt ≤ x}, dt ⊗ dQ-a.s. for all (T, x).

Setting T = t, we obtain that (16) is equivalent to (11)–(12).

8
 Static arbitrage strategies, such as holding long a (T, x)-bond and short a
(T, y)-bond if P (0, T, x) > P (0, T, y) for some x < y, may not be admissible
and hence not excluded by condition (10). Hence (10) does not necessarily
assert monotonicity of P (t, T, x) in T and x. Note, however, that (12) implies
monotonicity of f (t, t, x) in x in any case. The following corollary is obvious:
Corollary 3.4. If, in addition to (11)–(12), the discounted prices processes in
(10) are true martingales, then P (t, T, x) is decreasing in T and increasing in
x. This holds in particular if forward rates are positive: f (t, T, x) ≥ 0.
Remark 3.5. We present our approach under the assumption that Q is a risk-
neutral measure, i.e. the no-arbitrage condition (10) is supposed to hold under
Q. It is of course possible to consider the above dynamic equations with respect
to some objective probability measure P ∼ Q. The measure change from P to Q
will have the following impact:
a(t, T, x) Ã a(t, T, x) + b(t, T, x)> · Φ(t)
ν(t, dx) Ã Ψ(t, x)ν(t, dx)

for some appropriate stochastic processes Φ(t) and Ψ(t, x) with values in Rd and
(0, ∞), respectively. We do not intend to provide further general results on this,
as it is rather standard and regularity conditions have to be checked from case
to case. For a general reference see Theorem III.3.24 in [14], for Markovian
models see also [5].

4 A Martingale Problem
Theorem 3.3 states that, under the no-arbitrage condition (10), the drift param-
eter a(t, T, x) is determined by the volatility and contagion parameters b(t, T, x)
and c(t, T, x), respectively. However, there is still an implicit relation between
the loss process L and the short end of the forward curve f (t, t, x) in (12) which
cannot be expressed directly in terms of the volatility and contagion parameters.
This circumstance has been addressed in the previous works [1, 18, 8, 9, 20]
by ad-hoc methods, such as the construction of conditional Markov loss pro-
cesses given the market information. This special case will be further discussed
in Section 5 below.
In this section, we provide mathematical evidence that arbitrage-free CDO
term structure models (9) with properties (C1) and (C2), in fact, uniquely exist
under general assumptions. Our framework contains and unifies the approaches
in [1, 18, 8, 9, 20] as particular cases.
Without loss of generality we henceforth assume that the stochastic basis
satisfies:
(A6) Ω = Ω1 × Ω2 , F = G ⊗ H, Q(dω) = Q1 (dω1 )Q2 (ω1 , dω2 ), where ω =
(ω1 , ω2 ) ∈ Ω, and Ft = Gt ⊗ Ht , where
(i) (Ω1 , G, (Gt ), Q1 ) is some filtered probability space carrying the market
information, in particular the Brownian motion W (ω) = W (ω1 ),

9
 (ii) (Ω2 , H) is the canonical space of paths for marked point processes
endowed with the minimal filtration (Ht ): the generic ω2 ∈ Ω2 is a
càdlàg, increasing, piecewise constant function from R+ to I. Hence-
forth, we let the loss process

Lt (ω) = ω2 (t)

be the coordinate process. The filtration (Ht ) is thus Ht = σ(Lt (s) |

s ≤ t), and H = H∞ ,
(iii) Q2 is a probability kernel from (Ω1 , G) to H to be determined below.
This setup implies that the volatility and contagion parameters

b(ω; t, T, x) = b(ω1 , ω2 ; t, T, x), c(ω; t, T, x) = c(ω1 , ω2 ; t, T, x)

in (A4)–(A5) actually are functions of the loss path ω2 . Hence the indirect
contagion property (C2) is satisfied. The evolution equation (9) can thus be
solved on the stochastic basis (Ω1 , G, (Gt ), Q1 ) along any genuine loss path ω2 ∈
Ω2 . Indeed, the integral with respect to µ in (9) is path-wise in ω2 . However,
in view of (6), condition (12) is equivalent to

ν(ω; t, dx) = −f (ω; t, t, ω2 (t) + dx), (set f (t, t, x) ≡ rt for x ≥ 1). (17)

Hence, unless the contagion parameter c is zero,

a(t, T, x) = a(t, T, x, f (t, t, ·))

in (11) becomes an explicit linear functional of the (short end of the) prevailing
spread curve. But since this dependence on f (t, t, ·) is linear, for any given loss
path ω2 ∈ Ω2 , equation (9) will generically be uniquely solvable.
It thus remains to find a probability kernel Q2 such that ν in (17) becomes
the compensator of L. This is a martingale problem for marked point processes,
which has completely been solved by Jacod [13]. It turns out that Q2 exists and
is unique.
Theorem 4.1. Assume (A6) holds. Let f (0, T, x), b(t, T, x) and c(t, T, x) sat-
isfy (A2), (A4) and (A5), respectively. Define ν(t, dx) by (17) and a(t, T, x)
by (11) for all (t, T, x).
Suppose, for any loss path ω2 ∈ Ω2 , there exists a solution f (t, T, x) of (9)
such that f (t, t, x) is progressive, decreasing and càdlàg in x ∈ I. Then

(i) (A3) is satisfied.

(ii) there exists a unique probability kernel Q2 from (Ω1 , G) to H, such that the
loss process Lt (ω) = ω2 (t) satisfies (A1) and the no-arbitrage condition
(10) holds.

10
(iii) ν(t, dx) dt is the compensator of L with respect to (G ⊗ Ht ). Moreover,

Q [τn+1 − τn > t | G ⊗ Hτn ]

R τn +t
= Q2 [τn+1 − τn > t | Hτn ] = e− τn
ν(ω1 ,ω2 (τn );s,I) ds
, t≥0 (18)

and
ν(τn , A)
Q [∆Lτn ∈ A | G ⊗ Hτn − ] = Q2 [∆Lτn ∈ A | Hτn − ] = (19)
ν(τn , I)
for all A ∈ B(I) on {τn < ∞}, where 0 < τ1 < τ2 < · · · denote the
successive jump times of L.
(iv) Q2 (·, A) is Gt -measurable for all A ∈ Ht and t. Consequently, every (Gt )-
martingale is a (Ft )-martingale.
Remark 4.2. Property (iv) is known as “(H)-hypothesis”, see [3, 9]. Property
(iii) has been explored in [9] as “successive (H)-property” for finite I. The
formulas (18)–(19) will be most useful for numerical implementation as sketched
below.
Proof. That (A3) holds follows from the local boundedness assumptions (A4)–
RT
(A5), which in particular imply that 0 |f (t, t, 0)| dt < ∞ for all T , see [11].
Whence (i).
By Lemma 3.1 and (12), if the loss process L satisfies (A1), its compensator
is necessarily of the form (17). Theorem 3.6 in [13] now implies that there exists
a unique probability kernel Q2 from Ω1 to H, such that ν is the compensator
of L. Indeed, Jacod’s [13] notation (in quotation marks) corresponds to ours as
follows: “Ω = Ω0 × Ω00 ”↔ Ω = Ω1 × Ω2 , “F0 ”↔ G ⊗ {∅, Ω2 }, “Ft ”↔ G ⊗ Ht ,
“P0 ”↔ Q1 (its trivial extension to G ⊗{∅, Ω2 }, respectively). In particular, what
Jacod [13] refers to as “past” (“F0 ”) corresponds to our market information G.
In [13, eqn (10)], Jacod then recursively defines a unique probability measure
“P ” on “(Ω, F∞ )”, starting with a transition probability from “(F0 , P0 )”. This
yields our probability kernel Q2 . By construction, in [13], ν becomes the pre-
dictable projection of µ on “(Ω, F, (Ft ), P )”, which is (G ⊗ Ht ). This proves (ii)
and the first part of (iii). In view of [2, Section VIII.1], ν(t, I) is the (G ⊗ Ht )-
intensity of the point process {τn }. Hence, on {τn < ∞},

Φ(t) = Q[τn+1 − τn > t | G ⊗ Hτn ]

·Z τn +t ¸
=1+E ν(s, I)1{τn+1 −τn >s} ds | G ⊗ Hτn
τn
Z τn +t
=1+ ν(ω1 , ω2 (τn ); s, I)Φ(s) ds,
τn

which implies (18). Formula (19) follows from [2, Theorem 6, Chapter VIII].
Now fix T and replace “F0 ” above by GT ⊗{∅, Ω2 } and “Ft ” by GT ⊗Ht . Then
ν(t, dx) dt, t ≤ T , is “(Ft )”-predictable and Jacod’s [13] argument above yields

11
 (T )
a unique probability kernel Q2 from (Ω1 , GT ) to HT such that ν becomes
the predictable projection of µ. But then, by uniqueness again, we conclude
(T )
that Q2 (·, A) = Q2 (·, A) is GT -measurable for all A ∈ HT . From this, (iv)
follows.
Remark 4.3. The SPA [20] paper leaves open the question whether a “condi-
tional Markov” loss process consistent with the loss distributions for all x ∈ I
can be constructed in their way. Theorem 4.1 now resolves and clarifies this
in full generality, as it implies that the law of the loss process L is uniquely
determined by f (0, T, x), b(t, T, x) and c(t, T, x).
The pricing of loss path-dependent CDO derivatives, such as the default leg
of a STCDO in Section 6 below, becomes a computational issue. In general,
one has to resolve to Monte-Carlo methods such as the following. Fix a finite
time horizon τ , and denote by f (0, T, x) a given initial forward curve, for T ≤ τ
and x ∈ I. We now sketch an algorithm to simulate N trajectories for the joint
process (f (t, T, x), Lt ), t ≤ T ≤ τ .
(i) Simulate N independent samples of ω1 , i.e. standard Brownian paths
(1) (N )
ω1 (t), . . . , ω1 (t), t ∈ [0, τ ].
(ii) Initialize: set T (i,0) := 0, f (i,0) (0, T, x) := f (0, T, x), j := 1, and the initial
(i,j)
placeholder loss process ω2 (t) := 0, t ∈ [0, τ ], for i = 1, . . . , N .
(i) (i,j)
(iii) Solve (9) along (ω1 , ω2 ), e.g. via Euler scheme. This gives f (i,j) (t, T, x)
for all t ≤ T ≤ τ and x ∈ I. In fact, you can set f (i,j) (t, T, x) :=
f (i,j−1) (t, T, x) for t < T (i,j−1) .
(iv) Simulate an independent standard exponential random variable ²(j) , set
λ(i,j) (t, x) := f (i,j) (t, t, x) − f (i,j) (t, t, 1) (:= 0 for x ≥ 1), and determine
the jth jump time via
½ Z t ¾
(i,j)
T (i,j) = inf t ≥ T (i,j−1) | λ(i,j) (s, ω2 (T (i,j−1) )) ds ≥ ²(j) .
T (i,j−1)

This is justified by (18).

(i) (i,j)
(v) If T (i,j) > τ , set ω2 (t) := ω2 (t) for all t ∈ [0, τ ]. The ith path is thus
fully simulated and can be omitted in the following. Continue with those
i where T (i,j) ≤ τ . The algorithm terminates if no such i is left.
(vi) Simulate the jth loss jump size ∆LT (i,j) as independent I-valued random
variable with distribution function
(i,j) (i,j)
λ(i,j) (T (i,j) , ω2 (T (i,j) )) − λ(i,j) (T (i,j) , ω2 (T (i,j) ) + x)
(i,j)
, x ∈ I.
λ(i,j) (T (i,j) , ω2 (T (i,j) ))
This is justified by (19). Update the loss path
(
(i,j)
(i,j+1) ω2 (t), t < T (i,j) ,
ω2 (t) = (i,j)
ω2 (t) + ∆LT (i,j) , t ≥ T (i,j) .

12
(vii) Set j := j + 1 and repeat from (iii).
Note that, in every iteration, we have to assume that f (i,j) (t, t, x) is decreas-
ing and càdlàg in x ∈ I.

5 Doubly Stochastic Framework

As a special case of the above general framework we now omit both contagion
effects (C1) and (C2) and assume
(A7) the volatility parameter b(ω; t, T, x) = b(ω1 ; t, T, x) is (Gt )-adapted, and
c(t, T, x; y) ≡ 0.
Note that a(t, T, x) in (11) simplifies considerably. In particular, there is no
explicit dependence on f (t, t, ·) via (17) anymore. Moreover, the forward curve
f (ω; t, T, x) = f (ω1 ; t, T, x) becomes a function of ω1 only.
Theorem 4.1 can now be improved and extended by a very useful formula
for CDO derivatives pricing.
Theorem 5.1. Suppose (A7) and the assumptions in Theorem 4.1 hold.
Then the loss process L becomes G-conditional Markov under Q2 . Moreover,
for any positive G-measurable random variable X and all x ∈ I,
h RT i
E[X1{LT ≤x} | Ft ] = 1{Lt ≤x} E Xe− t λ(s,x)ds | Gt . (20)

Proof. The G-conditional Markov property of L follows since, for given ω1 , the
compensator ν(ω; t, dx) = −f (ω1 ; t, t, ω2 (t) + dx) in (17) is now a function of
the current loss level ω2 (t) only.
Now let X ≥ 0 be G-measurable. Since λ(t, x) is G-measurable, we obtain
£ ¤
Φ(T ) := E X1{LT ≤x} | G ⊗ Ht
" Z T #
x
= E XMT − X1{Ls ≤x} λ(s, x)ds | G ⊗ Ht
0
Z T £ ¤
= XMtx − λ(s, x)E X1{Ls ≤x} | G ⊗ Ht ds
0
Z T
= X1{Lt ≤x} − λ(s, x)Φ(s)ds.
t
We infer that RT
Φ(T ) = X1{Lt ≤x} e− t
λ(s,x)ds
.
Conditioning on Ft yields (20).
Formula (20) states that the market information G is enough to price any,
possibly loss path-dependent, CDO derivative. Indeed, a simple iteration of
(20) yields
" n # " n #
\ Y − R ti λ(s,xi ) ds
Q {Lti ≤ xi } ∩ A | Ft0 = 1{Lt0 ≤mini xi } E e ti−1
1A | G t 0
i=1 i=1

13
for all t0 < · · · < tn and A ∈ G. Hence the Ft0 -conditional law of the loss pro-
cess is given by a Gt0 -conditional expectation of a functional of the G-measurable
compensator λ(t, x). This property generalizes the concept of a doubly stochas-
tic Poisson process to marked point processessee e.g. [2, Section II.1] or [17,
Chapter 9].

6 Single Tranche CDOs (STCDOs)

The standard instrument for investing in a CDO-pool is a single tranche CDO
(STCDO), also called tranche credit default swap.
A STCDO is specified by
• a number of future dates T0 < T1 < · · · < Tn ,
• a tranche with lower and upper detachment points x1 < x2 in I,
• a fixed spread S.
We write Z
H(x) := (x2 − x)+ − (x1 − x)+ = 1{x≤y} dy.
(x1 ,x2 ]

An investor in this STCDO

• receives SH(LTi ) at Ti , i = 1, . . . , n (payment leg),
• pays −dH(Lt ) = H(Lt− ) − H(Lt ) at any time t ∈ (T0 , Tn ] where ∆Lt 6= 0
(default leg).
As in (13), we denote the (Gt )-adapted part of the (T, x)-bond price by
p(t, T, x), so that P (t, T, x) = 1{Lt ≤x} p(t, T, x).
Lemma 6.1. The value of the STCDO at time t ≤ T0 is

Γ(t, S)
Z Ã n
!
X
= 1{Lt ≤y} S p(t, Ti , y) + p(t, Tn , y) − p(t, T0 , y) + γ(t, y) dy
(x1 ,x2 ] i=1
(21)
where Z Tn h Ru i
γ(t, y) = E ru e− t rs ds 1{Lu ≤y} | Ft du.
T0
Moreover,
(i) if the risk free rates f (s, u) and the loss process Ls , for t ≤ s ≤ u, are
Ft -conditionally independent then γ(t, y) in (21) can be replaced by
Z Tn
γ(t, y) = f (t, u)p(t, u, y) du.
T0

14
 (ii) if the doubly stochastic assumptions of Theorem 5.1 are in force then
γ(t, y) in (21) can be replaced by
Z Tn h Ru i
γ(t, y) = E ru e− t f (s,s,y)ds | Gt du.
T0

Proof. The value of the payment leg at time t ≤ T0 is

" n # n Z
X R Ti X
− t rs ds
E e SH(LTi ) | Ft = S P (t, Ti , y)dy.
i=1 i=1 (x1 ,x2 ]

Next we use integration by parts to calculate

Z Tn Ru
e− t
rs ds
dH(Lu )
T0
R Tn R T0
Z Tn Ru
= e− t
rs ds
H(LTn ) − e− t rs ds
H(LT0 ) + ru e− t
rs ds
H(Lu )du.
T0

The (negative) value of the default leg at time t ≤ T0 for the investor is then
given as Ft -conditional expectation. Summing up the two legs, we obtain (21).
Part (i) follows since
h Ru i
E ru e− t rs ds | Ft = f (t, u)P (t, u).

Part (ii) follows from (20) and (12).

The forward STCDO spread S ∗ (t) prevailing at t ≤ T0 is the spread which

gives Γ(t, S ∗ (t)) = 0. In view of (21) hence
R ¡ ¢
(x1 ,x2 ]
P (t, T0 , y) − P (t, Tn , y) − γ(t, y)1{Lt ≤y} dy
∗
S (t) = Pn R .
i=1 (x1 ,x2 ] P (t, Ti , y) dy

A STCDO swaption with strike spread K gives the holder the right to enter
the above STCDO with spread K at swaption maturity T0 . Its value at T0 is
thus Γ(T0 , K)+ . Note that, since Γ(T0 , S ∗ (T0 )) = 0, this swaption payoff can
also be written as
à n Z !
X +
P (T0 , Ti , y) dy (K − S ∗ (T0 )) . (22)
i=1 (x1 ,x2 ]

As it is the case for single name models, e.g. [7, 19, 4], there is no closed form
solution for swaption prices available in general. See however Remark 7.3 below.

15
7 Affine Term Structure Models
In this section we consider an analytically tractable class of Markov factor mod-
els for the term structure movements (9) in the doubly stochastic framework.
We assume that (A6) holds. Let Z ⊂ Rd be some closed state space with
non-empty interior and Z some Z-valued diffusion process satisfying
dZt = µ(Zt )dt + σ(Zt ) · dWt ,
(23)
Z0 = z
where µ and σ are continuous functions from R+ × Z into Rd and Rd×d , respec-
tively.
In what follows we consider affine term structure models of the form
f (t, T, x) = A0 (t, T, x) + B 0 (t, T, x)> · Zt
that is, in terms of (9),
a(t, T, x) = ∂t A0 (t, T, x) + ∂t B 0 (t, T, x)> · Zt + B 0 (t, T, x)> · µ(Zt )
(24)
b(t, T, x) = B 0 (t, T, x)> · σ(Zt )
for some smooth functions A0 (t, T, x) and B 0 (t, T, x) with values in R and Rd ,
respectively. We denote
Z T Z T
A(t, T, x) = A0 (t, u, x)du, B(t, T, x) = B 0 (t, u, x)du.
t t

The following theorem gives a characterization of those affine term structure

models which satisfy the no-arbitrage condition (10).
Theorem 7.1. Assume that, for all z ∈ Z, there exists a Z-valued continuous
solution Z = Z z of (23) such that the coefficients given in (24) satisfy (11) for
all t ≤ T and x a.s. If the d + d(d+1)
2 functions in (T, x),
Bi (0, T, x), Bk (0, T, x)Bl (0, T, x), k ≤ l, (25)
are linearly independent, then Z is necessarily affine. That is, drift and diffusion
matrix are affine functions of z = (z1 , . . . , zd ) ∈ Z:
d
X Xd
1
µ(z) = µ0 + zi µi , σ · σ > (z) = ν0 + zi νi (26)
i=1
2 i=1

for some vectors µi ∈ Rd and matrices νi ∈ Rd×d . Moreover, A and B solve the
following system of Riccati equations, for t ≤ T ,
−∂t A(t, T, x) = A0 (t, t, x) + µ> >
0 · B(t, T, x) − B(t, T, x) · ν0 · B(t, T, x)
A(T, T, x) = 0
(27)
−∂t Bi (t, T, x) = Bi0 (t, t, x) + µ> >
i · B(t, T, x) − B(t, T, x) · νi · B(t, T, x)
B(T, T, x) = 0
for all (T, x).

16
Proof. Note that
Z T
∂t A0 (t, u, x)du = ∂t A(t, T, x) + A0 (t, t, x),
t

and analogously for B 0 . Hence (11) yields

∂t A(t, T, x) + A0 (t, t, x) + (∂t B(t, T, x) + B 0 (t, t, x))> · Zt + B(t, T, x)> · µ(Zt )

1
= B(t, T, x)> · σ · σ > (Zt ) · B(t, T, x). (28)
2
Letting t ↓ 0, by continuity, we obtain the respective equality for Zt replaced
by z, for all T , x and z. We infer that

σ · σ > (z)
B(0, T, x)> · µ(z) + B(0, T, x)> · · B(0, T, x)
2
is an affine function in z, for all T and x. By assumption (25), we conclude that
µ and σ · σ > /2 must be affine functions of the form (26).
Plugging (26) back in (28) and separating first order terms in zi , we obtain
(27).
The next theorem is the converse to Theorem 7.1 and gives sufficient condi-
tions for the existence of an arbitrage-free affine term structure model.
Theorem 7.2. Assume µ and σσ > are affine of the form (26). Let A0 (t, t, x)
and B 0 (t, t, x) be some given functions with values in R and Rd , respectively,
jointly continuous in (t, x) ∈ R+ × I, and such that A0 (t, t, x) + B 0 (t, t, x)> · z
is decreasing in x ∈ I for all t and z ∈ Z.
Let A and B be given as solutions of the Riccati equations (27), and let Z be
a continuous Z-valued solution of (23), for some z ∈ Z. Then the conclusions
of Theorem 5.1 apply and
>
P (t, T, x) = 1{Lt ≤x} e−A(t,T,x)−B(t,T,x) ·Zt

defines an arbitrage-free, doubly stochastic (T, x)-bond market.

Proof. It follows as in the proof of Theorem 7.1 that (11) is equivalent to (28),
which again is implied by (27). Moreover,

f (t, t, x) = A0 (t, t, x) + B 0 (t, t, x)> · Zt (29)

satisfies the required properties in Theorem 5.1. Hence the conclusions of The-
orem 5.1 apply.
Remark 7.3. Using the affine toolbox, as developed in e.g. [7, 19, 4], and the
fact that f (t, t, x) in (29) is an affine function of the affine process Z, derivative
prices such as in Lemma 6.1 and (22) can now efficiently be computed.

17
7.1 Example
As simple example, we consider: d = 1, Z = R+ , µ0 ≥ 0, µ1 ∈ R, ν1 = σ 2 /2,
for some σ > 0. That is, Z is a Feller square root process:
p
dZt = (µ0 + µ1 Zt )dt + σ Zt dWt , Z0 = z ∈ R+ .
Moreover, we let A0 (t, t, x) = α(t, x) and B 0 (t, t, x) = β(x), for some R+ -
valued continuous functions α(t, x) and β(x) which are decreasing in x ∈ I with
α(t, 1) ≡ r ≥ 0 and β(1) = 0. That is, we have a constant risk free short rate
rt ≡ r, and λ(t, x) = α(t, x) − r + β(x)Zt .
The Riccati equations (27) become
Z T
A(t, T, x) = (α(s, x) + µ0 B(s, T, x)) ds
t
σ2
−∂t B(t, T, x) = β(x) + µ1 B(t, T, x) − B(t, T, x)2 , B(T, T, x) = 0.
2
The equation for B has the solution
¡ ¢
2β(x) eρ(x)(T −t) − 1
B(t, T, x) ≡ B(T − t, x) = ¡ ¢ ¡ ¢
ρ(x) eρ(x)(T −t) + 1 − µ1 eρ(x)(T −t) − 1
p
where ρ(x) = µ21 + 2σ 2 β(x). Note that
∂T A(t, T, x) = α(T, x) + µ0 B(T − t, x).
Hence, we obtain
f (t, T, x) = α(T, x) + µ0 B(T − t, x) + ∂T B(T − t, x)Zt
f (t, T ) ≡ r.
It is now possible to find a closed form solution for the STCDO value in
Lemma 6.1. We conclude with the remarkable fact that this simple model is
capable of capturing any given initial forward curve f (0, T, x) by an appropriate
choice of the function α(T, x).

References
[1] Bennani, N. (2005), “The forward loss model: A dynamic term structure
approach for the pricing of portfolio credit derivatives,” Working Paper.
[2] Brémaud, P. (1981), Point processes and queues, martingale dynamics,
Springer, New York.
[3] Brémaud, P. and Yor, M. (1978), “Changes of filtration and of probability
measures,” Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebi-
ete, 45, 269–295.

18
 [4] Chen, L. and Filipović, D. (2007), “Credit Derivatives in an Affine Frame-
work,” Asia-Pacific Financial Markets 14, 123-140.

[5] Cheridito, P., Filipović, D. and Yor, M. (2005), “Equivalent and Absolutely
Continuous Measure Changes for Jump-Diffusion Processes,” The Annals
of Applied Probability, 15 1713–1732.
[6] Cont, R. and Minca, A. (2008), “Recovering portfolio default intensi-
ties implied by CDO quotes,” Financial Engineering Report No. 2008-01,
Columbia University Center for Financial Engineering.
[7] Duffie, D., Filipović, D. and Schachermayer, W. (2003), “Affine Processes
and Applications in Finance,” The Annals of Applied Probability 13, 984–
1053.
[8] Ehlers, P. and Schönbucher, P. (2006), “Pricing Interest Rate-Sensitive
Credit Portfolio Derivatives,” Working Paper, ETH Zurich.
[9] Ehlers, P. and Schönbucher, P. (2006), “Background Filtrations and Canon-
ical Loss Processes for Top-Down Models of Portfolio Credit Risk,” Working
Paper, ETH Zurich.

[10] Ethier, S.N. and Kurtz, T.G. (1986), Markov Processes. Characterization
and Convergence, John Wiley&Sons.

[11] Filipović, D. (2008), “Interest Rate Models,” Lecture Notes, Vienna Insti-
tute of Finance.

[12] Giesecke, K.Goldberg, L. (2007), “A Top Down Approach to Multi-Name

Credit”, Working Paper, Stanford University.

[13] Jacod, J. (1975), “Multivariate Point Processes: Predictable Projection,

Radon-Nikodym Derivatives, Representation of Martingales”, Zeitschrift
für Wahrscheinlichkeitstheorie und verwandte Gebiete 31, 235–253.
[14] Jacod, J. and Shiryaev, A.N. (1987), Limit Theorems for Stochastic Pro-
cesses, Springer.
[15] Heath, D., Jarrow, R. A. and Morton, A. J. (1992), “Bond pricing and the
term structure of interest rates,” Econometrica, 60, 77–105.
[16] Laurent, J.P. and Gregory, J. (2005), “Basket default swaps, CDOs and
factor copulas,” Journal of Risk, 7, 103 – 122.
[17] McNeil, A., Frey, R. and Embrechts, P. (2005), Quantitative Risk Manage-
ment: Concepts, Techniques and Tools, Princeton University Press.
[18] Schönbucher, P. (2005), “Portfolio losses and the term structure of loss
transition rates: A new methodology for the pricing of portfolio credit
derivatives,” Working Paper, ETH Zurich.

19
[19] Singleton, K. and Umantsev, L. (2002), “Pricing coupon-bond options and
swaptions in affine term structure models,” Mathematical Finance, 12, 427–
446.
[20] Sidenius, J., Piterbarg, V. and Andersen, L. (2008), “A new framework for
dynamic credit portfolio loss modelling,” International Journal of Theoret-
ical and Applied Finance, 11, 163–197.

20