Swiss Finance Institute

Research Paper Series N06 39

Pricing Interest Rate-Sensitive
Credit Portfolio Derivatives

Philippe EHLERS
ETH Zurich, D-Math
Philipp J. SCHONBUCHER
ETH Zurich, D-Math





























Established at the initiative of the Swiss Bankers' Association, the Swiss
Finance Institute is a private foundation funded by the Swiss banks and
SWX. It merges 3 existing foundations: the International Center FAME, the
Swiss Banking School and the Stiftung "Banking and Finance" in Zurich.
With its university partners, the Swiss Finance Institute pursues the
objective of forming a competence center in banking and finance
commensurate to the importance of the Swiss financial center. It will be
active in research, doctoral training and executive education while also
proposing activities fostering interactions between academia and the
industry. The Swiss Finance Institute supports and promotes promising
research projects in selected subject areas. It develops its activity in
complete symbiosis with the NCCR FinRisk.

The National Centre of Competence in Research Financial Valuation and
Risk Management (FinRisk) was launched in 2001 by the Swiss National
Science Foundation (SNSF). FinRisk constitutes an academic forum that
fosters cutting-edge finance research, education of highly qualified finance
specialists at the doctoral level and knowledge transfer between finance
academics and practitioners. It is managed from the University of Zurich and
includes various academic institutions from Geneva, Lausanne, Lugano,
St.Gallen and Zurich. For more information see www.nccr-finrisk.ch .


This paper can be downloaded without charge from the Swiss Finance
Institute Research Paper Series hosted on the Social Science Research
Network electronic library at:

http://ssrn.com/abstract=957744



Pricing Interest Rate-Sensitive Credit Portfolio
Derivatives
Philippe Ehlers and Philipp J. Sch onbucher

D-MATH, ETH Zurich

July 2006, this version December 2006

Abstract
In this paper we present a modelling framework for portfolio credit risk which
incorporates the dependence between risk-free interest-rates and the default loss
process. The contribution in this approach is that besides the traditional diusion-
based covariation between loss intensities and interest-rates a direct dependence
between interest-rates and the loss process is allowed, in particular default-free
interest-rates can also depend on the loss history of the credit portfolio. Amongst
other things this enables us to capture the eect that economy-wide default events
are likely to have on government bond markets and/or central banks interest-rate
policies. Similar to Sch onbucher (2005), the model is set up using a set of loss-
contingent forward interest-rates f
n
(t, T) and loss-contingent forward credit pro-
tection rates F
n
(t, T) to parameterize the market prices of default-free bonds and
credit-sensitive assets such as CDOs. We show that (up to weak regularity condi-
tions), existence of such a parametrization is necessary and sucient for the absence
of static arbitrage opportunities in the underlying assets. We also give necessary
conditions and sucient conditions on the dynamics of the parametrization which
ensure absence of dynamic arbitrage opportunities in the model. Similar to the
HJM drift restrictions for default-free interest-rates, these conditions take the form
of restrictions on the drifts of f
n
(t, T) and F
n
(t, T), together with a set of regularity
conditions.

Authors Address: ETH Zurich, D-MATH, R amistr. 101, CH-8092 Zurich, Switzerland.
Ehlers@math.ethz.ch, www.math.ethz.ch/ehlers, P@Schonbucher.de, www.schonbucher.de.
This paper was presented at: Bachelier Finance Society, Fourth World Congress, 2006, Tokyo and RiskDay
2006, Zurich.
Financial support by the National Centre of Competence in Research Financial Valuation and Risk
Management (NCCR FINRISK) is gratefully acknowledged. NCCR FINRISK is a research program
supported by the Swiss National Science Foundation (SNSF).
All errors are our own. Comments and suggestions are very welcome.

The rst draft of the paper was entitled Dynamic Credit Portfolio Derivatives Pricing.
1


Keywords: AP, MI; Credit Portfolio Risk, Top-Down, Forward Model, Contagion, Collateralized Debt Obligations
JEL: G13
1 Introduction
In this paper we present a modelling framework for portfolio credit risk which incorpo-
rates a new methodology to model the dependence between risk-free interest-rates and
the default loss process, allowing direct dependence between interest-rates and the loss
process.
We provide a stochastic and arbitrage-free framework for the evolution of the prices of a
set of contingent-claims on the credit portfolios loss distribution. This set of contingent
claims is complete in the sense that it spans all European contingent claims on the loss pro-
cess L(t).
1
In particular, the prices of index credit default swaps (CDS) and single-tranche
collateralized debt obligations (STCDO) of all maturities and attachment points can be
easily constructed from these contingent claims. This allows a straightforward calibra-
tion of the model to the so-called correlation smile. In contrast to Sch onbucher (2005),
though, these prices can not be interpreted as probabilities under the spot martingale
measure as we allow dependence between the loss process and risk-free interest-rates.
The prices of the basic contingent claims are parameterized using a set of loss-contingent
forward interest-rates f
n
(t, T) and loss-contingent forward credit protection rates F
n
(t, T).
The forward interest-rates f
n
(t, T) must be loss contingent in order to allow us to capture
its credit dependence. These rates can be viewed as the interest-rates of forward-rate
agreements that are contingent on a certain number of losses L(T) = n. Clearly, if
there is dependence between the loss process and the default-free interest-rates, the loss-
contingent forward rates f
n
(t, T) must dier over dierent values of n. We show that
(up to weak regularity conditions), existence of such a parametrization is necessary and
sucient for the absence of static arbitrage opportunities in the underlying assets, i.e.
the parameterization fully describes the set of arbitrage-free price systems in this model.
Next, we analyze the possible dynamics of the thus dened market for portfolio credit
derivatives and interest-rate derivatives. We give necessary conditions and sucient con-
ditions on the dynamics of the parametrization which ensure absence of dynamic arbitrage
opportunities in the model. Similar to the HJM drift restrictions for default-free interest-
rates, these conditions take the form of restrictions on the drifts of f
n
(t, T) and F
n
(t, T),
together with a set of regularity conditions on the stochastic characteristics of the param-
eterization.
The modelling framework presented in this paper can be applied very eciently to the
pricing of exotic portfolio credit-derivatives such as options on index CDS, leveraged super-
senior tranches, options on STCDOs and in particular also hybrid derivatives on credit
portfolios and interest-rates. Given the increased liquidity of the markets for index CDS
and STCDOs, a particular advantage of this model over most competing models is that
it can be calibrated very easily to a given set of prices for index CDS and STCDOs. Such
a calibration only requires changing the initial conditions of the loss-contingent forward
interest- and protection-rates f
n
(t, T) and F
n
(t, T) but it does not require any changes to
the dynamics of the model.
Empirical studies of the dependence between default rates and risk-free interest-rates
1
I.e. any contingent claim with payo X at time T, where X is (L(T))-measurable can be replicated
using a static portfolio of these contingent claims.
2
usually nd a negative dependence between interest-rates and defaults. Such results are
found investigating the interest-rate correlations of credit spreads of corporate bonds
(e.g. Duee (1998), Duee (1995), Collin-Dufresne et al. (2001)), and also when actual
default event arrivals are investigated (see e.g. Due et al. (2006)). The most common
explanation of this negative dependence is that high default rates and low interest-rates
tend to coincide in recessions, while booms are usually characterized by low default rates
and high interest-rates.
Traditionally, interest-rate risk is incorporated in intensity-based models of credit risk by
specifying diusion-based dynamics for the interest-rates and correlating them with the
dynamics of the default intensities of the obligor(s) in question. Most importantly, a direct
inuence from the credit event itself to the level or dynamics of the default-free interest-
rates is usually (implicitly) ruled out. This approach is followed most frequently in models
of single-obligor default risk (e.g. Due and Singleton (1997),Madan and Unal (1998),
Jarrow et al. (1997) and the literature following it) where indeed it may be argued that
the individual obligor under consideration is small compared to the macroeconomic
forces determining the level of default-free interest-rates and where the inuence of rm-
specic factors determining the default-risk dominates. Thus it is plausible that the
dependence between the default of any given individual rm and interest-rates should be
indirect and weak. Nevertheless, even in this situation some diculties arise: e.g. negative
covariation between interest-rates and intensities is essentially impossible to achieve in an
exponentially ane framework with positive interest-rates and positive intensities (the
most common specication of intensity-based default risk).
The assumption of weak and indirect dependence cannot be sustained when default risk
is considered on a portfolio level. Almost by denition, if a representative portfolio suers
a large number of defaults, a recession or a similar macroeconomic crisis must be present:
if not as the cause, then certainly as the eect of the defaults. This crisis in turn will
directly inuence the level of default-free interest-rates through the interest-rate policies
of central banks and through eects on investor preferences and ination expectations in
xed-income markets. Thus, for economic reasons, there should be a stronger and more
direct link between portfolio defaults and interest-rates.
Moreover, diversication eects also shift the focus towards macro-variables when mov-
ing from a single-obligor point of view to portfolio credit risk modelling as idiosyncratic
inuences average out in a portfolio context. Thus, we nd it very plausible that the
dependence between interest-rates and large losses in credit portfolios should be stronger
and more direct than the dependence between interest-rates and the default of one specic
individual obligor.
Also from a mathematical point of view it is natural to allow non-defaultable quantities
in the model to be driven by the portfolio loss process: If the models setup admits a mar-
tingale representation theorem, the resulting representation must include integrals with
respect to the compensated loss process

L(t) (or another martingale which generates the
jumps of L(t)). After all, the loss process itself is adapted to the models ltration. Thus,
if one were to specify the generic dynamics of any other martingale (e.g. a discounted
bond price process) in the model, it would exhibit joint jumps with L and if it does
not have such joint jumps this amounts to a modelling assumption which needs to be
justied.
3
As argued above, we believe that a credit portfolios loss process can convey important
macroeconomic information which should aect the prices of many dierent nancial
instruments. With the same justication we also argue to allow inuences between L
and other variables such as equity indices, exchange rates etc. Here we focus on the
dependence between losses and default-free interest-rates. For practical implementations
this is the most important case because interest-rates enter the pricing formulae of all
securities, in particular also the pricing formulae of elementary calibration securities such
as defaultable bonds, CDS or STCDOs.
Unfortunately, almost invariably the literature on portfolio credit risk assumes indepen-
dence between defaults and interest-rates. For multivariate intensity models some depen-
dence can be introduced by correlating default intensities and interest-rates in the same
way as for single-obligor intensity models, yet as mentioned above this approach is not
without diculties and problems. In other types of portfolio credit risk models the incor-
poration of interest-rate dependence is not so straightforward, this applies in particular
to portfolio credit risk models based upon static latent factors to drive the dependence
like the very widely-used Gauss copula models (Li (2000)).
The model presented in this paper follows the top-down approach to credit risk modelling.
Models of this class do not focus on the defaults or survival of individual obligors but rather
attempt to capture the behavior of the aggregate portfolio loss process. Representatives
of this approach are e.g. Davis and Lo (2001), Giesecke and Goldberg (2005) or Frey and
Backhaus (2004). More concretely, the parametrization of the prices of the basic securities
and the idea of modelling the price dynamics of a whole term- and strike structure of
STCDOs in this paper goes back to the forward loss surface modelling approach proposed
in Sch onbucher (2005), which in turn is closely related to the model proposed by Sidenius
et al. (2005). We show later that the model by Bennani (2005) can be viewed as a
projection of the model presented in this paper, removing the loss-dependence from the
state variables. Furthermore, the model used in Brigo et al. (2006) and the empirical
study of CDO prices in Longsta and Rajan (2006) can be viewed as simplied versions
of these loss surface models.
The rest of the paper is structured as follows: In the following part we introduce the
basic securities for portfolio loss modelling: loss contingent zero coupon bonds P
n
(t, T)
and up-front prices U
n
(t, T) for protection against the n-th default. After showing how
liquid portfolio credit derivatives like index CDS and STCDOs can be priced using solely
these instruments, we analyze the conditions under which a system of these prices can
be represented using loss-contingent forward interest-rates f
n
(t, T) and loss-contingent
forward protection rates F
n
(t, T). We discuss the economic interpretation of these rates
and show that existence of such a representation is essentially equivalent to absence of
static arbitrage opportunities in the basic securities.
The next part of the paper addresses the problem which dynamics of the parametrization
{f
n
(t, T), F
n
(t, T)} are consistent with absence of dynamic arbitrage opportunities. We
characterize the dynamics that {f
n
(t, T), F
n
(t, T)}, the loss process L(t) and the short
rate process r(t) have to follow under any spot martingale measure and also give sucient
conditions for the dynamics to be arbitrage-free. As these conditions entail very specic
forms for the drift parameters of the f
n
(t, T) and F
n
(t, T), we also have to address the
problem of existence of solutions to these restricted stochastic dierential equations.
4
2 Basic Portfolio Credit Derivatives
In this section we show that index CDS and STCDOs on a portfolio of N obligors can be
priced by setting up simple static portfolios of the building block securities P
0
, P
1
, . . . , P
N
and U
1
, . . . , U
N
. Index CDS and STCDOs are by far the most liquid portfolio credit
derivatives in the market and the most popular assets used to calibrate portfolio credit
risk models. The calibration of our model to these instruments is facilitated by the
relatively simple form of the pricing formulae.
Furthermore, we show how the payo functions of many more exotic portfolio credit
derivatives can be represented in our setup. Thus, these derivatives can be priced by
simulation in our model.
2.1 Setup
We are on a ltered probability space (, F, F, P). T

< is a nite investment horizon,

F = (F
t
)
t[0,T

]
satises the usual conditions, all processes are F-adapted and c` adl` ag and
P is understood as the physical probability measure if nothing else is specied. P denotes
the predictable and O the optional -eld, B([0, T

]) is the Borel -eld on [0, T

] and
L denotes the Lebesgue measure, and M(P) is the space of uniformly integrable P-
martingales. When we make a statement about a term structure, that is a family of
processes X(t, T)
t[0,T]
indexed by T [0, T

], we will omit the phrase for all 0 t

T T

if no ambiguities can arise, and we always assume X(t, , T) to be OB([0, T

])-
measurable.
We consider a given credit portfolio consisting of N N obligors whose cumulative credit
losses are described by the loss process L
t
. We assume L
t
is a piecewise constant,
bounded process with jumps of size one and L
0
= 0. The credit exposures of the obligors
are normalized to unity so that we can identify the number of losses with the amount
lost in the portfolio. The time of the n
th
default (loss) in the portfolio is denoted by

n
:= inf{t > 0; L
t
n}, n = 1, . . . , N.
2
We assume the following nancial instruments
are traded at all times t [0, T

] for all T t and n = 0, . . . , N:

P
n
(t, T), is the price at time t of an asset paying 1$ at time T if and only if L
T
= n.
U
n
(t, T), is the up-front price for protection against the n
th
loss,
i.e. the price at time t of an asset paying 1$ at
n
if and only if
n
(t, T].
It is natural to set P
m1
(t, T) = U
m
(t, T) = 0 for all m L
t
. Further, money can always
be invested in the money market account b
t
, which is nondecreasing and satises b
0
= 1
a.s.
2
We use the convention inf = . Because L
t
only has one-step increments, no two jump times will
coincide if one of them is nite, i.e.
n
<
n+1
for all n N with
n
< .
5
2.2 Index CDS and Single-Tranche CDOs
In an index credit default swap (index CDS) with maturity T, the protection seller
promises to pay all credit losses that occur in the portfolio up to time T, i.e. he promises
to pay 1 at
n
if
n
T for all n N. Expressed in terms of the building blocks, the
value of such a protection at time t is
Prot
0,N
(t, T) :=
N

n=1
U
n
(t, T). (1)
The protection buyer on the other hand promises to pay at the payment dates T
k
, k =
1, . . . K a protection fee s times the remaining notional of the index and times a daycount
fraction. It is important to note that at a default, the notional on which the fee is paid
is reduced: At any time t, the remaining notional of the index is N L
t
=

N
n=1
1
{n>t}
.
If we approximate the discrete payments with a continuous payment stream the value of
the fee stream at time t is
s Fee
0,N
(t, T) := s
_
T
t
N

n=0
(N n)P
n
(t, u)du. (2)
The dierence between a STCDO and an index CDS is that the protection of the STCDO
only covers a certain range [a, b] of the possible losses. In a STCDO with attachment
point a and detachment point b > a and maturity T, the protection seller promises to pay
1 at
n
if
n
T for all n [a + 1, b]. (An index CDS is reached by setting a = 0 and
b = N.) The value of this protection is
Prot
a,b
(t, T) :=
b

n=a+1
U
n
(t, T).
Similar to the case of the index CDS, the protection fee on the STCDO is also paid on a
variable notional amount which is reduced by all losses that fall in the range [a, b] between
attachment point and detachment point. Overall the value of the fee stream is
s Fee
a,b
(t, T) := s
_
T
t
(b a)
a

n=0
P
n
(t, u)du +s
_
T
t
b

n=a+1
(b n)P
n
(t, u)du,
which is composed of the value of the fee payment on the full notional which is paid as
long as L
t
a (the rst term), and the value of the fees paid on a notional amount which
is reduced by the losses already incurred.
Given these formulae, the fair fee quoted in the market is the fee that makes the value of
the fee leg equal to the value of the protection leg:
s
a,b
(t, T) :=
Prot
a,b
(t, T)
Fee
a,b
(t, T)
.
Our assumption of unit losses in default allowed us to slightly simplify the payos by
identifying the losses with the default count. In a real-world STCDO, payos are based
solely upon the cumulative loss process, i.e. the protection seller pays the increments in
6
the tranche loss process (L
t
a)
+
(L
t
b)
+
, where a, b R. For the fee payment, the
notional of a STCDO is reduced by the tranches loss process, while the notional of a
real-world index CDS is reduced by the total notional of the defaulted obligor (and not
just the loss in default). If a constant recovery rate is assumed, it is straightforward to
adapt the pricing formulae given above to these modications.
Recently, tranchelets, i.e. very thin tranches have gained popularity in the market. In our
setup, the thinnest possible tranche would cover one default, i.e. b = a+1. Its protection
leg has simply the value U
b
(t, T), while the fee leg has the value s
_
T
t

a
n=0
P
n
(t, u)du.
2.3 Forward-Starting STCDOs and Tranche Options
A T
1
-into-(T
2
T
1
) forward-starting STCDO with attachment point a and detachment
point b > a is a contract which at time T
1
turns into a STCDO over [T
1
, T
2
], where the
protection fee has already been xed at the initial time t < T
1
at the level s
f
. The twist
is that the attachment and detachment points are shifted by any loss amounts that occur
before T
1
, i.e. it turns into a STCDO with attachment point (a + L
T
1
) and detachment
point (b +L
T
1
), and not simply into a STCDO with attachment point a and detachment
point b. (The latter case would have been straightforward to price.)
While it is not possible to directly give the price of a forward-starting STCDO, we can at
least give its payo at T
1
. At T
1
, the value of the position value to the protection seller is
_
s
f
s
(a+L
T
1
),(b+L
T
1
)
(T
1
, T
2
)
_
Fee
(a+L
T
1
),(b+L
T
1
)
(T
1
, T
2
).
A Put (Payer) is usually dened as the right to by protection (i.e. sell risk) and a Call
(Receiver) as the right to sell protection (i.e. buy risk). Tranche options come in various
shapes and forms and we can only give the simplest variations: The payo at time T
1
of
a European Put on a forward-starting STCDO is
_
s
f
s
(a+L
T
1
),(b+L
T
1
)
(T
1
, T
2
)
_
+
Fee
(a+L
T
1
),(b+L
T
1
)
(T
1
, T
2
),
while a European Put on a plain STCDO without adjustment of attachment- and detach-
ment points has the payo
_
s
f
s
a,b
(T
1
, T
2
)
_
+
Fee
a,b
(T
1
, T
2
).
Leveraged Super-Senior tranches are a form of barrier option on a STCDO, usually a
super-senior tranche, i.e. a tranche with a high attachment point. In these transactions,
a trigger event is specied, usually in terms of a hitting time of the loss process L
t
or a
hitting time of the portfolio spread s
0,N
(t, T) or a combination of these. Upon the trigger
event, the protection seller makes a payment to the protection buyer which is linked to
the value of a given STCDO-tranche. The precise specication of such transactions varies
but it is clear that the trigger events and the payos of these structures can be expressed
in terms of our basic securities.
7
3 Forward Model and Absence of Arbitrage
3.1 Static No-Arbitrage Conditions
In our setup, the default-free zero coupon can be expressed as B(t, T) :=

N
n=0
P
n
(t, T),
and it is well-known that
B(t, T) (0, 1] and nonincreasing in T and B(t, t) = 1. (B)
is a necessary condition in order to avoid static arbitrage opportunities if cash can be
stored at zero cost. In particular, if L
t
= N, then P
N
(t, T) = B(t, T) will be the only
remaining traded asset apart from the money market account b
t
. When L
t
< N, the
situation more involved. Obviously, absence of (static) arbitrage requires
P
n
(t, T) 0 and P
n
(t, t) = 1
{Lt=n}
, n = 0, . . . , N. (P)
U
n
(t, T) nondecreasing in T and U
n
(t, t) = 0, n = 1, . . . , N. (U)
because P
n
and U
n
yield nonnegative payos, and the larger T, the more protection
provides U
n
. Additionaly we dene the portfolios
3

n
(t, T) := P
n
(t, t) +U
n
(t, T) U
n+1
(t, T) P
n
(t, T).
By the denition of P
n
and U
n
, holding
n
(t, T) pays o as follows:
+1$ at t and 1$ at
n+1
T for n = L
t
+1$ at
n
T and 1$ at
n+1
T for n > L
t
Consequently, for any > 0,
n
(t, T +)
n
(t, T) pays nothing if the loss level n is not
reached in [T, T +) (i.e. L
u
= n for all u [T, T +)), and otherwise +1$ at max{T,
n
}
and 1$ at min{(T + ),
n+1
} (i.e. later). Hence in an arbitrage-free market,
n
(t, T)
must at least satisfy

n
(t, T) nondecreasing in T, n = 0, . . . , N. ()
We call (B), (P), (U) and () the static no-arbitrage conditions. If they hold, then
it is easily checked that also the following statements are true: P
n
(t, T) 1 and for
each T there is always at least one n L
t
with P
n
(t, T) > 0. U
n
(t, T) [0, 1] and
is nonincreasing in n.
n
(t, T) [0, 1] and
n
(t, t) = 0. And the defaultable zero bond
S
n
(t, T) :=

n
j=Lt
P
j
(t, T), which pays 1$ at T if at most n losses have occurred by time T,
satises S
n
(t, T) [0, 1] and is nonincreasing in T and S
n
(t, t) = 1 for all n = L
t
, . . . , N.
Remark 1 (
n
as Loss Contingent Forward Lending). If the law of each
n
is diuse,
then a full term structure of
n
(t, T) allows to approximate the n
th
loss-contingent for-
ward interest rate arbitrarily well in the following sense. Suppose n L
t
. Then for
every T t and > 0, there exists a.s. an F
t
-measurable = (t, ) > 0 such that
P[
n
,
n+1
/ (T, T + ) | F
t
] > 1 (on the set {L
t
n}). And hence with probability
greater than 1 ,
n
(t, T + )
n
(t, T) will pay
4
+1$ at T and 1$ at T + if L
T
= n,
nothing otherwise,
3
If L
t
= n, then P
j
(t, T) and U
j+1
(t, T), j = 0, . . . , n 1 are not traded anymore, and we are free to
set them equal to zero. Further we make the convention U
N+1
(t, T) = 0.
4
And with probability less than it pays +1$ and then 1$ within [T, T + ] as described above.
8
i.e. holding
n
(t, T + )
n
(t, T) approximately replicates forward lending of 1$ over
[T, T +], contingent on L
T
= n.
We denote the space of rightcontinuous functions in T with C
0,r
(T), and C
1,r
(T) is the
space of continuous functions that admit a right derivative which is in C
0,r
(T). We write

T
for the right derivative operator. Suppose now (B), (P), (U) and () hold and
P
n
(t, T) and U
n
(t, T) are in C
1,r
(T). Then
F
n
(t, T) := 1
{Pn(t,T)=0}
1
Pn(t,T)

T
U
n+1
(t, T) and
f
n
(t, T) := 1
{Pn(t,T)=0}
1
Pn(t,T)

n
(t, T)
(3)
are well-dened and nonnegative. Assuming additionally
_
T

t
f
n
(t, u) + F
n
(t, u)du <
5
for all n = 0, . . . , N and dening P
1
:= 0 ensures that we have by construction
P
n
(t, T) = e
t,T
n
_
1
{Lt=n}
+
_
T
t
(e
t,u
n
)
1
P
n1
(t, u)F
n1
(t, u)du
_
and
U
n
(t, T) =
_
T
t
P
n1
(t, u)F
n1
(t, u)du, (4)
where e
t,T
n
:= e

R
T
t
fn(t,u)+Fn(t,u)du
. This way, P
0
=, . . . , = P
Lt1
= U
1
=, . . . , = U
Lt
= 0 is
automatically satised, and

T
P
n
(t, T) = P
n
(t, T)
_
f
n
(t, T) +F
n
(t, T)
_
+P
n1
(t, T)F
n1
(t, T) (5)
holds with initial condition P
n
(t, t) = 1
{Lt=n}
.
The classical forward rate curve f(t, T) C
0,r
for default-free interest-rates is character-
ized by f(t, T) :=
T
log B(t, T), which is equivalent to
_
T
t
B(t, u)f(t, u)du = 1B(t, T)
for all T [t, T

]. Similarly, we can uniquely dene a curve F(t, T) C

0,r
(T) such that
Prot
0,N
(t, T) =
N

n=1
U
n
(t, T) =
_
T
t
B(t, u)F(t, u)du
for all T [t, T

]. The integral
_
T
t
B(t, u)F(t, u)du is the value of a continuous non-
constant payment stream of F(t, u)du over the interval u [t, T] which by construction
is equal to the the value Prot
0,N
(t, T) of protection against all losses over [t, T] (see also
equation (1)). Hence, F(t, T) can be interpreted as a type of non-standard protection fee
which is not subject to the usual reduction of a standard index CDS (cp. the discussion
preceding (2)). Using this notation, we obtain
Proposition 1. f(t, T) and F(t, T) decompose into
f(t, T) =
N

n=0
P
n
(t, T)
B(t, T)
f
n
(t, T) and F(t, T) =
N1

n=0
P
n
(t, T)
B(t, T)
F
n
(t, T),
i.e. f(t, T) and F(t, T) are weighted averages over f
n
(t, T)and F
n
(t, T), respectively.
5
with the convention F
N
(t, T) = 0.
9
Proof. Note that

N
n=0

n
(t, T) = 1 B(t, T) and hence B(t, T)f(t, T) =
T
B(t, T) =

N
n=0

n
(t, T) =

N
n=0
P
n
(t, T)f
n
(t, T). Analogously, B(t, T)F(t, T) =
T
Prot
0,N
(t, T) =

N
n=1

T
U
n
(t, T) =

N
n=1
P
n
(t, T)F
n
(t, T).
Thus, our model renes the notion of forward interest-rates and forward protection
rates by conditioning these rates on the realization of L(t) = n. The corresponding
unconditional rates are reached by projection. A model which only uses unconditional
forward loss rates F(t, T) and does not allow a dependence on L
t
is the forward loss
model of Bennani (2005). This modelling approach amounts to an implicit assumption
of F(t, T) = F
n
(t, T) (and also f(t, T) = f
n
(t, T)) for all n. This has implausible conse-
quences, e.g. the price of the next-to-default zero bond (which defaults at the next loss
event) P
Lt
(t, T) = e

R
T
t
f
L
t
(t,u)+F
L
t
(t,u)du
does not depend on the current loss level L
t
.
We have seen that if the no-arbitrage requirements (B), (P), (U) and () hold, then
f
n
and F
n
as dened in (3) are nonnegative. Next, we turn to the converse case.
Denition 1. If there exists nonnegative term structures f
n
(t, T) and F
n
(t, T), n =
0, . . . , N, with
_
T

t
f
n
(t, u) + F
n
(t, u)du < for all t T

and n = L
t
, . . . , N, such that
P
n
and U
n
satisfy (4) a.s., then we say P
n
and U
n
have the representation property with
respect to f
n
and F
n
. (For short, they have the representation property.)
In denition 1, we do not require that P
n
and U
n
are in C
1,r
(T), and f
n
and F
n
need not
be in C
0,r
(T). Also note that f
0
, . . . , f
Lt1
and F
0
, . . . , F
Lt1
can take arbitrary values
without aecting P
n
and U
n
in (4). Therefore, f
n
F
n
in denition 1 are not unique.
The results below are understood t-wise a.s.
Theorem 2. Let P
n
and U
n
have the representation property. Then (B), (P), (U)
and () hold.
The proof can be found in appendix A. With a slight abuse of notation we obtain
6
Corollary 3. Let f
n
and F
n
be in C
0,r
(T) with
_
T

t
|f
n
(t, u)| + |F
n
(t, u)| du < for
n = 0, . . . , N. Then the following are equivalent.
(i) P
n
and U
n
are in C
1,r
(T) and satisfy (B), (P), (U) and ().
(ii) P
n
and U
n
have the representation property with respect to f
n
and F
n
, in particular
f
Lt
, . . . , f
N
and F
Lt
, . . . , F
N1
are nonnegative.
Informally, we can state that under weak regularity, for absence of static arbitrage in
the market consisting of P
n
and U
n
(and b
t
) it is necessary that P
n
and U
n
have the
representation (4) with F
n
and f
n
nonnegative.
Remark 4. In practical applications, it would be possible to t all traded tranches
with piecewise constant (in T) term structures f
n
(t, T) and F
n
(t, T). Therefore we think
that requiring P
n
(t, T) and U
n
(t, T) to be continuously dierentiable in T (or almost
equivalently f
n
(t, T) and F
n
(t, T) be continuous in T) might be too restrictive.
6
By abuse of notation we mean that f
n
and F
n
can be dened likewise by (3) or denition 1. But as
we showed, under sucient regularity the two notions coincide.
10
3.1.1 Strictly Positive Interest Rates, Default- and Survival Probabilities
At the beginning of this section, we required only minimal properties for the evolution of
the bank account b
t
and the distribution of the loss times
n
, n = 1, . . . , N under P. It
is realistic and removes implausible investment opportunities
7
to assume in addition that
b
t
is a.s. increasing and that t-wise a.s.
P[
n
> T | F
t
] (0, 1) for all T > t and is decreasing in T, n = L
t
+ 1, . . . , N.
Intuitively, there always remains a minimum of uncertainty about the occurrence of losses
in the future (unless the entire portfolio has already defaulted) and the money market
account pays positive interest. In this modied setup,
U
n
(t, T) increasing in T and U
n
(t, t) = 0, n = L
t
+ 1, . . . , N. (U

n
(t, T) increasing in T, n = L
t
, . . . , N. (

)
are necessary conditions for absence of arbitrage. In case (B), (P), (U

) and (

)
hold, f
n
and F
n
as dened in (3) will be positive, and it holds in addition: B(t, T) is
decreasing in T. P
n
(t, T) (0, 1) for all T > t. U
n
(t, T) (0, 1) and is decreasing in n.

n
(t, T) (0, 1) for all T > t and
n
(t, t) = 0. And the defaultable zero bond S
n
(t, T)
is decreasing in T and satises S
n
(t, t) = 1 for all n = L
t
, . . . , N. With a little more
work along the proof of theorem 2, we then obtain the counterpart of theorem 2 and its
corollary.
Theorem 5. Let P
n
and U
n
have the representation property with respect to f
n
and F
n
,
and let f
n
, n = L
t
, . . . , N and F
n
, n = L
t
, . . . , N 1 be positive a.s. Then (B), (P),
(U

) and (

) hold.
Corollary 6. Let f
n
and F
n
be in C
0,r
(T) with
_
T

t
|f
n
(t, u)| + |F
n
(t, u)| du < for
n = 0, . . . , N. Then the following are equivalent.
(i) P
n
and U
n
are in C
1,r
(T) and satisfy (B), (P), (U

) and (

).
(ii) P
n
and U
n
have the representation property with respect to f
n
and F
n
, and
f
n
, n = L
t
, . . . , N and F
n
, n = L
t
, . . . , N 1 are positive.
3.2 Representation with Auxiliary Markov Chain
If P
n
and U
n
have the representation property and moreover f
n
and F
n
are in C
0,r
(T),
then the building blocks P
n
(t, T) and U
n
(t, T) can also be interpreted as follows. For
xed t (and ), consider the auxiliary {L
t
, . . . , N}-valued Markov chain

L = (

L
T
)
T[t,T

]
with law

P, dened by

P[

L
t
= L
t
] = 1 and the time-inhomogeneous generator matrix
F(t, T)
T[t,T

]
with (omitting the (t, T)-argument)
F :=
_
_
_
_
_
F
0
F
0
0 . . .
0 F
1
F
1
0
.
.
. 0
.
.
.
.
.
.
0 . . . 0 F
N
_
_
_
_
_
7
E.g. that P
n
(t, T) = 1 for T > t or that U
n
(, T) trades at zero even though L
t
< n.
11
This way the law

P is uniquely characterized. Further, we dene an auxiliary short rate
process r = ( r
T
)
T[t,T

]
by
r
T
:= f
e
L
T
(t, T)
and auxiliary stopping times
Lt
:= t and
n
:= inf{T > t;

L
T
n} for n = L
t
+1, . . . , N.

E[ ] denotes expectation wrt.

P. Then we have
Proposition 2. For every n = L
t
, . . . , N and T t it holds
P
n
(t, T) =

E
_
1
{
e
L
T
=n}
e

R
T
t
e rsds
_
and U
n
(t, T) =

E
_
1
{t<e nT}
e

R
e n
t
e rsds
_
.
I.e. under weak regularity, there always (i.e. for every t) exists a unique auxiliary Markov
chain model (

L,

P) together with an auxiliary short rate r = r(

L) that allow to represent

P
n
(t, T) and U
n
(t, T) as their expected discounted nal payos. The proof of proposition
2 can be found at the end of this subsection.
Remark 7. This statement also informally conrms that there exist no static arbitrage
opportunities if f
n
and F
n
in (4) are nonnegative (cf. corollary 3 where we show necessity).
Indeed, if the auxiliary Markov chain model were the true model, then the prices U
n
and
P
n
as dened in proposition 2 would be arbitrage-free, there would not exist any admissible
trading strategy oering arbitrage opportunities. In particular, this holds for the static
trading strategies
8
. Whether a static trading strategy admits arbitrage depends only on
the initial asset prices and the nal asset payos (and whether these payos occur with
positive probability), which are identical in the true and the auxiliary model. Formally,
to complete the argument, we would need to check here, that asset payos indeed have
the same null sets in both the true and the auxiliary model, e.g. that P[
n
> T | F
t
] > 0
if and only if

P[
n
> T ] > 0. We will not go further into this but refer to corollary 9
and the remark following it, which show that under weak regularity the auxiliary model
is a certain projection of the true model. Intuitively, this makes clear that asset payos
have the same null sets in both models.
Omitting the argument (t, T), we additionally dene the matrices P := (P
0
, . . . , P
n
),
f = diag(f
0
, . . . , f
N
) and the unit vectors e
0
:= (1, 0, . . . , 0)

, . . . , e
N
:= (0, . . . , 0, 1)

.
Then the term structure ODE (5) reads more compactly

T
P(t, T) = P(t, T)
_
f(t, T) +F(t, T)
_
, P(t, t) = e

Lt
.
This ODE looks similar to the Kolmogorov equation obtained in Sch onbucher (2005).
However, as we have noted in proposition 2, in general P(t, T) is not a probability dis-
tribution (its elements do not sum up to one), and the matrix f(t, T) + F(t, T) is not
a generator matrix (since F(t, T) is a generator matrix, but f(t, T) is not). Consider
D(t, T) :=
1
B(t,T)
P(t, T) instead, which is a probability distribution for all T t. By
dierentiation we get

T
D(t, T) = D(t, T)
_
f
ex
(t, T) +F(t, T)
_
, D(t, t) = e

Lt
,
8
To be precise, with a static trading strategy, we mean that one invests at t in a number of traded
assets (i.e. P
n
(t, T), U
n
(t, T), b
t
) and holds each of these positions until its respective maturity.
12
where f
ex
:= diag(f
0
f, . . . , f
N
f).
9
This distinguishes our results from those in
Sch onbucher (2005) as follows. For every t, D(t, T) is a Markov chain in T [t, T

], if
and only if
f
n
(t, T) = f(t, T) (6)
for all T t and n = L
t
, . . . , N, i.e. the term structure of interest rates coincides with
the term structure of loss-contingent interest rates for all possible loss levels n. See also
remark 11 below for further comments on this.
Proof of proposition 2. Let (

F
T
)
T[t,T

]
be the natural ltration of

L. We will rst show
that P
n
(t, T) as dened above satises (5). Clearly

E
_
1
{
e
Lt=n}
_
= 1
{Lt=n}
= P
n
(t, t).
We use Landaus symbol o() for g with g()/ 0 as 0. By construction, we
have

E
_
1
{
e
L
T+

e
L
T
>1}
_
= o() and on the set {

L
T
=

L
T+
= n} it holds e

R
T+
T
e rsds
=
1 f
n
(t, T) +o()

P-a.s. Thus
P
n
(t, T +) =
nLt

k=0

E
_
e

R
T
t
e rudu
1
{
e
L
T
=nk}

E
_
1
{
e
L
T+

e
L
T
=k}
e

R
T+
T
e rudu

F
T
_ _
= P
n
(t, T)
_
1 F
n
(t, T)
__
1 f
n
(t, T)
_
+o()
+

E
_
e

R
T
t
e rudu
1
{
e
L
T
=n1}

E
_
1
{
e
L
T+

e
L
T
=1}

F
T
_ _

E
_
e

R
T
t
e rudu
1
{
e
L
T
=n1}

E
_
1
{
e
L
T+

e
L
T
=1}
_
1 e

R
T+
T
e rudu
_

F
T
_ _
.
We notice that 1
{
e
L
T
=n1}

E
_
1
{
e
L
T+

e
L
T
=1}

F
T
_
= 1
{
e
L
T
=n1}
F
n1
(t, T) + o()

P-a.s.
Using this and r
u
f
n1
(t, u) +f
n
(t, u) on {n 1

L
u
n}, we deduce
1

E
_
e

R
T
t
e rudu
1
{
e
L
T
=n1}

E
_
1
{
e
L
T+

e
L
T
=1}
_
1 e

R
T+
T
e rudu
_

F
T
_ _

_
1 e

R
T+
T
f
n1
(t,u)+fn(t,u)du
_
1

E
_
1
{
e
L
T
=n1}

E
_
1
{
e
L
T+

e
L
T
=1}

F
T
_ _
0
0.
Combining the steps above we obtain that

E
_
1
{
e
L
T
=n}
e

R
T
t
e rsds
_
solves the forward ODE
(5). Second, we show that U
n
(t, T) as dened above, satises (4). Since
_
T
t
F
e
Lu
(t, u)du is
the predictable compensator of

L under

P,

E
_
1
{t<e nT}
e

R
e n
t
e rudu
_
=

E
_ _
T
t
1
{
e
L
u
=n1}
e

R
u
t
e rvdv
d

L
u
_
=
_
T
t

E
_
1
{
e
Lu=n1}
e

R
u
t
e rvdv
_
F
n1
(t, u)du
=
_
T
t
P
n1
(t, u)F
n1
(t, u)du = U
n
(t, T).
9
f
ex
can be termed loss-contingent excess forward rate matrix.
13
3.3 Dynamic No-Arbitrage Conditions
So far we have only excluded static arbitrage opportunities by restricting the term struc-
tures f
n
and F
n
in the representation (4) separately for every t. In this section, we are
concerned with the dynamic evolution of f
n
and F
n
. For derivatives pricing, we mostly
like to set up a model directly under an equivalent martingale measure (EMM) and the
link to the real-world probability measure is of less importance. We therefore give neces-
sary and sucient martingale/drift restrictions for the dynamics of f
n
and F
n
under an
EMM, but we do not question the existence of an EMM. Once the model is set up under
an EMM, every candidate for a real-world probability measure can be constructed using
appropriate measure transformations. Consequently, P does not necessarily denote the
real-world probability measure here.
Assumption 1.
(i) P
n
and U
n
have the representation property with respect to f
n
and F
n
.
(ii) The money market account is absolutely continuous.
(iii) A oating rate note with a continuous dividend rate r
fl
t
, ensuring that the value of
the oating rate note is constant, is traded.
By (ii) there exists a short rate process r
t
with b
t
= e
R
t
0
rsds
and hence the discount factor

t
:= 1/b
t
satises
t
= 1
_
t
0

s
r
s
ds. Note that P
n
(t, T) is an asset without dividends and
terminal value 1
{L
T
=n}
and U
n
(t, T) is a dividend-paying asset with terminal value zero
(U
n
(T, T) = 0) and dividend 1
{L
t
=n1}
dL
t
. It is well-known that absence of arbitrage is
equivalent to the existence of an equivalent (local) martingale measure.
Denition 2 (EMM). We say Q P is an equivalent martingale measure (EMM), if
every asset (i.e. b
t
, P
n
, U
n
and the oating rate note) with price process S
t
and accumulated
dividend process D
t
(i.e. with dividend dD
t
), satises

t
S
t
+
_
t
0

s
dD
s
M
loc
(Q). (7)
For the oating rate note, this means
t
+
_
t
0

s
r
fl
s
ds must be a Q-local martingale. But
since it is continuous and of bounded variation, it must be constant, which is equivalent
to r
fl
t
= r
t
P L(dt)-a.e. This allows to characterize the short rate (if it exists) in an
arbitrage-free market as the unique
10
process r
t
such that
t
+
_
t
0

s
r
s
ds is a P-martingale
(cf. Hughston and Rafailidis (2005)). Similarly, the predictable compensator of L under
P is the unique predictable nondecreasing process A
t
with A
0
= 0 such that L
t
A
t
is a
P-martingale. If A is absolutely continuous, then we may write A
t
=
_
t
0

s
ds, and we say
L admits the loss intensity under P.
Clearly, P
n
and U
n
are bounded with bounded dividends. Hence if P itself is an EMM,
then it holds P
n
(t, T) = E
_
1
{L
T
=n}

T
/
t

F
t

and
U
n
(t, T) = E
_ _
T
t
1
{L
s
=n1}

s
/
t
dL
s

F
t
_
. (8)
Now we are ready to formulate the main result of this subsection.
10
Unique in L
1
(PL).
14
Theorem 8. Let
t
P
n
(t, T)f
n
(t, T) and
t
P
n
(t, T)F
n
(t, T) M(P) for all n = 0, . . . , N
and T T

, let L admit the loss intensity

t
= F
Lt
(t, t) under P and let the short rate
satisfy r
t
= f
Lt
(t, t). Then P is an EMM.
Conversely, let P be an EMM, and assume that f
n
(t, T
0
) and F
n
(t, T
0
) are integrable and
lim
TT
0
E[ |x
n
(t, T) x
n
(t, T
0
)| ] = 0 for all t T
0
T

, n = 0, . . . , N and x = f, F.
Then
t
P
n
(t, T)f
n
(t, T) and
t
P
n
(t, T)F
n
(t, T) M(P) for all n = 0, . . . , N and T T

,
L admits an intensity and
r
t
= f
Lt
(t, t) and
t
= F
Lt
(t, t)
are right-continuous versions of the short rate and the loss intensity, respectively.
We immediately derive the following corollary, which illustrates best the relation between
r, and the forward rates f
n
, F
n
.
Corollary 9. Let L admit an intensity under P, and ssume that f
n
(t, T
0
) and F
n
(t, T
0
)
are integrable and lim
TT
0
E[ |x
n
(t, T) x
n
(t, T
0
)| ] = 0 for all t T
0
T

, n = 0, . . . , N
and x = f, F. Then P is an EMM if and only if
f
n
(t, T) = 1
{Pn(t,T)=0}
E
_

T
1
{L
T
=n}
r
T

F
t

t
P
n
(t, T)
and
F
n
(t, T) = 1
{Pn(t,T)=0}
E
_

T
1
{L
T
=n}

F
t

t
P
n
(t, T)
. (9)
Proof. If
t
= F
Lt
(t, t), then P
n
(T, T)F
n
(T, T) = 1
{L
T
=n}
F
n
(T, T) = 1
{L
T
=n}

T
. Anal-
ogously, P
n
(T, T)f
n
(T, T) = 1
{L
T
=n}
r
T
if r
t
= f
Lt
(t, t). Then the claim follows from
theorem 8.
Remark 10. The possibility that the loss intensity
t
of a credit portfolio jumps (up)
at the occurence of defaults is usually referred to as contagion. Due to theorem 8,
t
=
F
Lt
(t, t) must hold up to regularity under every EMM. This implies in particular that
modelling F
n
(t, T) as continuous processes (for every T) does not rule out contagion
because by construction,
t
= F
Lt
(t, t) automatically admits discontinuities at each loss
event. Contagion is a natural feature of our model. We will present a diusion-based
Heath-Jarrow-Morton-type (HJM) setup of the model in section 4.
For the rest of this section we suppose the conditions of corollary 9 are satised and P is an
EMM. We denote with P
T
the T-forward measure
11
as introduced in Jamshidian (1987).
Remark 11. If P[ L
T
= n ] > 0, then P
n
(0, T) > 0 and similar to the T-forward
measure we may dene the the n
th
loss-contingent T-forward measure P
T
n
P by
dP
T
n
dP

Ft
:=

t
P
n
(t, T)
P
n
(0, T)
Then (9) simplies to f
n
(t, T) = E
P
T
n
[ r
T
| F
t
] and F
n
(t, T) = E
P
T
n
[
T
| F
t
]. It is well-
known that f(t, T) = E
P
T
[ r
T
| F
t
] and using proposition 1 we also obtain F(t, T) =
11
P
T
is induced by the P-martingale
t
B(t, T) via
dP
T
dP

Ft
=
tB(t,T)
B(0,T)
.
15
E
P
T
[
T
| F
t
]. This shows that both standard and loss-contingent forward interest rates
and forward loss rates are projections of the true future interest rates and default rates,
with respect to an appropriate probability measure. It also shows that f
n
(t, T) 0 and
F
n
(t, T) 0 must be the case if P is an EMM.
We mentioned in section 3.2 that the representation (4) discussed in this paper coincides
with that obtained in Sch onbucher (2005) if and only if (6) holds, i.e. f
n
(t, T) = f(t, T)
for all n. This is in turn is equivalent to
E
P
T
_
1
{L
T
=n}
r
T

F
t

= E
P
T
_
1
{L
T
=n}

F
t

E
P
T
[ r
T
| F
t
] ,
i.e. conditional on F
t
, L
T
and r
T
are uncorrelated wrt. the T-forward measure P
T
. (To
see this, note L
T
=

N
n=0
n1
{L
T
=n}
.) Empirical studies as e.g. Due et al. (2006) indicate
that correlation between interest rates and defaults is typically negative.
Proof of theorem 8. To show the rst result, we only need to check that all traded assets
satisfy (7). The oating rate note satises (7) by construction.
_
t
0
F
Ls
(s, s)ds being the
predictable compensator of L under P, we have
E
_ _
T
t

s
1
{L
s
=n1}
dL
s

F
t
_
= E
_ _
T
t

s
1
{L
s
=n1}
F
Ls
(s, s)ds

F
t
_
= E
_ _
T
t

s
P
n1
(s, s)F
n1
(s, s)ds

F
t
_
=
_
T
t

t
P
n1
(t, s)F
n1
(t, s)ds =
t
U
n
(t, T) (10)
for all n = 1, . . . , N and t T T

(in the second equality we used that L

s
= L
s
P L(ds)-a.e. and P
n1
(s, s) = 1
{Ls=n1}
.). Hence U
n
satises (7). As concerns P
n
, we
note that, since r
t
= f
Lt
(t, t),

T
1
{L
T
=n}
=
t
1
{Lt=n}

_
T
t

s
1
{L
s
=n}
f
Ls
(s, s)ds +
_
T
t

s
(1
{L
s
=n1}
1
{L
s
=n}
)dL
s
.
Then, the latter together with (10) and the denition of
n
and P
n
(T, T) = 1
{L
T
=n}
yields
E[
T
P
n
(T, T) | F
t
] =
t

n
(t, T) +
t
P
n
(t, T)
_
T
t
E
_

s
1
{Ls=n}
f
Ls
(s, s)

F
t

ds
=
t

n
(t, T) +
t
P
n
(t, T)
_
T
t
E[
s
P
n
(s, s)f
n
(s, s) | F
t
] ds
=
t

n
(t, T) +
t
P
n
(t, T)
_
T
t

t
P
n
(t, s)f
n
(t, s)ds
=
t
P
n
(t, T)
for all n = 0, . . . , N and t T T

, i.e. P
n
satises the EMM condition (7).
For the converse direction, we will rst prove that the martingale property holds and
then show that both L
t

_
t
0
F
Ls
(s, s)ds and
t
+
_
t
0

s
f
Ls
(s, s)ds are in M(P). Let
16
0 t
1
< t
2
T

and F F
t
1
. First of all we note that U
n
(t
1
, T) = U
n
(t
1
, t
2
) +
_
T
t
2
P
n1
(t
1
, u)F
n1
(t
1
, u)du for T t
2
. Hence, if P is an EMM, then
0 = E
_
1
F
_

t
2
U
n
(t
2
, T)
t
1
U
n
(t
1
, T) +
_
t
2
t
1

s
1
{L
s
=n1}
dL
s
_ _
= E
_
1
F
_
T
t
2

t
2
P
n1
(t
2
, u)F
n1
(t
2
, u)
t
1
P
n1
(t
1
, u)F
n1
(t
1
, u) du
_
E
_
1
F
_

t
1
U
n
(t
1
, t
2
)
_
t
2
t
1

s
1
{L
s
=n1}
dL
s
_ _
=
_
T
t
2
E
_
1
F
_

t
2
P
n1
(t
2
, u)F
n1
(t
2
, u)
t
1
P
n1
(t
1
, u)F
n1
(t
1
, u)
_
du (11)
for all T [t
2
, T

] and n = 1, . . . , N. In the last equality we used (8). This implies that

the integrand in (11) vanishes for L-a.e. u [t
2
, T]. E[ 1
F

t
P
n1
(t, u)F
n1
(t, u) ] exists
for all t u and n. It is right-continuous in u for all t and n because
E[ 1
F

t
|P
n
F
n
(t, u) P
n
F
n
(t, u
0
)| ] E[ |P
n
F
n
(t, u) P
n
F
n
(t, u
0
)| ]
E[ P
n
(t, u) |F
n
(t, u) F
n
(t, u
0
)| ]
+E[ F
n
(t, u
0
) |P
n
(t, u) P
n
(t, u
0
)| ]
E[ |F
n
(t, u) F
n
(t, u
0
)| ]
+E[ F
n
(t, u
0
) |P
n
(t, u) P
n
(t, u
0
)| ]
uu
0
0
(lim
uu
0
E[ F
n
(t, u
0
) |P
n
(t, u) P
n
(t, u
0
)| ] = 0 by dominated convergence.) Hence the
integrand in (11) vanishes for every u [t
2
, T] and it follows
t
P
n
(t, T)F
n
(t, T) M(P)
for all n = 1, . . . , N and T T

n
(t, T) = 1
{Lt=n}
P
n
(t, T) + E
_
_
T
t

s
/
t
(1
{L
s
=n1}
1
{L
s
=n}
)dL
s

F
t
_
is a conse-
quence of (8) and the denition of
n
(t, T). We proceed similar as above.
n
(t
1
, T) =

n
(t
1
, t
2
) +
_
T
t
2
P
n
(t
1
, u)f
n
(t
1
, u)du for T t
2
. Hence, if P is an EMM, then
0 = E
_
1
F
_

t
2

n
(t
2
, T)
t
1

n
(t
1
, T)
t
2
1
{Lt
2
=n}
+
t
1
1
{Lt
1
=n}
_
+E
_ _
t
2
t
1

s
_
1
{L
s
=n1}
1
{L
s
=n}
_
dL
s
_
=
_
T
t
2
E
_
1
F
_

t
2
P
n
(t
2
, u)f
n
(t
2
, u)
t
1
P
n
(t
1
, u)f
n
(t
1
, u)
_
du
for all n = 0, . . . , N and T T

. With the same arguments as above this shows that

t
P
n
(t, T)f
n
(t, T) M(P) for all n = 0, . . . , N and T T

.
By the martingale property,

N
n=1

t
P
n1
(t, s)F
n1
(t, s) = E[
s
F
Ls
(s, s) | F
t
] for every
s t. Note that dL
s
= 1
{L
s
<N}
dL
s
. Using (8), we observe
E
_ _
T
t

s
dL
s

F
t
_
=
N

n=1

t
U
n
(t, T) = E
_ _
T
t

s
F
Ls
(s, s)ds

F
t
_
17
for every t T T

. This proves that

_
T
t

s
dL
s
admits the intensity
t
F
Lt
(t, t) and,
since
t
is positive a.s., L admits the intensity
t
= F
Lt
(t, t), which is right-continuous.
Analogously,

N
n=0

t
P
n
(t, s)f
n
(t, s) = E[
s
f
Ls
(s, s) | F
t
] for every s t. We note that
1 B(t, T) =

N
n=0

n
(t, T) and therefore
E[
t

T
| F
t
] =
t
_
1 B(t, T)
_
=
N

n=0

n
(t, T) = E
_ _
T
t

s
f
Ls
(s, s)ds

F
t
_
for every t T T

. As pointed out above, this uniquely characterises f

Lt
(t, t) as the
short rate r
t
, and f
Lt
(t, t) is right-continuous by construction.
Remark 12. Recall 1 B(t, T) =

N
n=0

n
(t, T). If P
n
and U
n
are in C
1,r
(T), then so
are
n
, and the marginal (short-maturity) bond return rate can be dened as usual as
lim
0
1

_
1
B(t, t +)
1
_
= f
Lt
(t, t). (12)
If moreover P is an EMM and f
n
and F
n
are suciently regular (cf. the converse part of
theorem 8), then the marginal bond return rate is also equal to the short rate. However,
this would not yet prove that f
Lt
(t, t) satises
t
+
_
t
0

s
f
Ls
(s, s)ds = const. a.s. because
we did not use (12) as the denition of the short rate.
4 A HJM-Type Forward Model
In this section, we consider a HJM-type specication of the loss-contingent forward in-
terest and loss rates f
n
and F
n
. As we mentioned in remark 10, assuming that f
n
and F
n
are continuous processes seems not too restrictive. We maintain the general setup and
notation of section 3. P is not necessarily the physical probability measure, and W
t
is a
d-dimensional Brownian motion under P (d N).
Assumption 2 (HJM-type Forward Rates). L
t
admits the intensity
t
under P. For
every T T

, there exist nonnegative processes f

n
(t, T), n = 0, . . . , N, and F
n
(t, T),
n = 0, . . . , N 1, satisfying
df
n
(t, T) =
f
n
(t, T)dt +
f
n
(t, T)dW
t
,
dF
n
(t, T) =
F
n
(t, T)dt +
F
n
(t, T)dW
t
where
x
n
(t, , u) and
x
n
(t, , u) are P B([0, T

])-measurable for x = f, F and

(i)
_
T

0
f
2
n
(0, u) +F
2
n
(0, u)du < for all n.
(ii)
_
T

0
_
T

s

x
n
(s, u)
2
du ds < a.s. for all n and x = f, F.
(iii)
_
T

0
__
u
0
|
x
n
(s, u)| ds
_
2
du < a.s. for all n and x = f, F.
(iv) P
n
and U
n
have the representation property with respect to f
n
and F
n
.
Assumption 2 (i) to (iii) guarantee that the integrals
_
T
t
f
n
(t, u)du and
_
T
t
F
n
(t, u)du are
nite for all t T T

a.s. (cf. Heath et al. (1992) or corollary 14 below). In the sequel we

18
will use the abbreviations f
n
:= f
n
+F
n
,
n
:=
f
n
+
F
n
,
n
:=
f
n
+
F
n
,
t,T
:=
_
T
t
(t, v)dv
for t T,
t,u,T
:=
_
T
u
(t, v)dv for t u T, where = ,
n
, ,
n
, and
e
t,u,T
n
:= e

R
T
u
f
n
(t,v)dv
.
If a product of functions which share the same argument is considered, we will for brevity
put that argument only at the end of respective product, e.g. P
n
F
n
(t, T) stands for
P
n
(t, T)F
n
(t, T). Also, when integrating functions, we will sometimes completely omit
a term structure argument (t, T), (s, u) etc. if the argument can be guessed/recovered
from the integration bounds, e.g.
_
T
t
e
t,u,T
n
P
n
du is a shortform of
_
T
t
e
t,u,T
n
P
n
(t, u)du. P

n
denotes the left t-limit of P
n
, and L
t
is the compensated process L
t
:= L
t

_
t
0

s
ds.
We dene v
1
:= 0, u
1
:= 0 and
1
:= 0 and then iteratively
v
n
(t, T) := 1
{L
t
=n}
P

n

t,T
n
1
{L
t
<n}
_
T
t
e
t,u,T
n
_
P

n

n
P

n1

F
n1
F
n1
v
n1
_
du
u
n
(t, T) := 1
{L
t
=n}
P

n
_
f
n
(t, t)
t

t,T
n
+
1
2
_
_

t,T
n
_
_
2
_
+1
{L
t
=n1}
e
t,T
n
(
t
F
n1
(t, t))
1
{L
t
<n}
_
T
t
e
t,u,T
n
_
P

n

n
+

n
v
n
P

n1

F
n1
F
n1
u
n1
v

n1

F
n1
_
du

n
(t, T) := 1
{L
t
=n}
P

n
+1
{L
t
=n1}
e
t,T
n
+
_
T
t
e
t,u,T
n
F
n1

n1
du (13)
for n = 0, . . . , N, and
b
n
(t, T) :=
_
T
t
_
P

n1

F
n1
+F
n1
v
n1
_
du
a
n
(t, T) := 1
{Lt=n1}
F
n1
(t, t) +
_
T
t
_
P

n1

F
n1
+F
n1
u
n1
+v

n1

F
n1
_
du

n
(t, T) :=
_
T
t
F
n1

n1
du
for n = 1, . . . , N. The dynamics of P
n
and U
n
are given in
Theorem 13 (Representation). Let assumption 2 be satised. Then the functions v
n
, u
n
,
n
are PB([0, T

])-measurable with
_
T
0
v
n
(s, T)
2
+|u
n
(s, T)|+|
n
(s, T)|
s
ds < a.s. for
every n = 0, . . . , N and T T

and
dP
n
(t, T) = u
n
(t, T)dt +v
n
(t, T)dW
t
+
n
(t, T)dL
t
. (14)
a
n
, b
n
,
n
are P B([0, T

])-measurable,
_
T
0
b
n
(s, T)
2
+|a
n
(s, T)| +|
n
(s, T)|
s
ds <
a.s. for every n = 1, . . . , N and T T

and
dU
n
(t, T) = a
n
(t, T)dt +b
n
(t, T)dW
t
+
n
(t, T)dL
t
. (15)
4.1 Proof of theorem 13
To begin with, we briey review a few properties of the HJM-type term structures f
n
(t, T)
and F
n
(t, T) that we will use in subsequent proofs. Heath et al. (1992) showed that in
19
this setup
_
T

0
f
n
(t, t)dt < a.s. and d
_
T
t
f
n
(t, u)du =
t,T
n
dt +
t,T
n
dW
t
f(t, t)dt for all
n (see also theorem 14 below). This yields
de
t,u,T
n
e
t,u,T
n
=
_
1
{t=u}
f
n
(t, t)
t,u,T
n
+
1
2
_
_

t,u,T
n
_
_
2
_
dt
t,u,T
n
dW
t
(16)
and then the dynamics of 1
{Lt=n}
e
t,T
n
, in particular those of P
0
(t, T) = 1
{Lt=0}
e
t,T
0
, follow
easily with It os lemma. Recalling (4) shows that to obtain the dynamics of the P
n
(t, T)
and U
n
(t, T) for n 1, we need to compute iteratively the dynamics of integrals like
_
T
t
P
n1
(t, u)F
n1
(t, u)du. Therefor we need the following generalization of the integration
rules in Heath et al. (1992).
Theorem 14. Let (t, , u), (t, , T) and (t, , T) be P B([0, T

])-measurable with
_
T

0
_
T

s
|(s, u)| + (s, , u)
2
+ |(s, , u)|
s
du ds < a.s. and let the term structure
X(t, T) satisfy
_
T

0
|X(0, T)| dT < and
dX(t, T) = (t, T)dt + (t, T)dW
t
+ (t, T)dL
t
for every T [0, T

]. Then
t,T
,
t,T
and
t,T
are P-measurable and
d
__
T
t
X(t, u)du
_
=
t,T
dt +
t,T
dW
t
+
t,T
dL
t
X(t, t)dt.
Proof of theorem 14. Heath et al. (1992) showed that for = 0 the result follows from
the stochastic Fubini theorem (see e.g. Protter (2004), chapter IV, theorem 65). We can
thus assume without restriction that = 0, = 0 and X(0, ) = 0. It can be easily
checked that
t,T
is predictable and by the property of the predictable compensator, the
integrals
_
t
0
|
_
T

0
(s, u)du|dL
s
and
_
t
0
|(s, u)| dL
s
, u [t, T

] are nite for all t [0, T

]
a.s. The rest of the proof is due to the ordinary Fubini theorem, which applies separately
to the integrals with respect to dL
s
du and
s
ds du, respectively. Joining them
together (dL
s
= dL
s

s
ds), we conclude
_
T
t
X(t, u)du =
_
T
t
_
t
0
(s, u)dL
s
du =
_
t
0
_
T
s
(s, u)du dL
s

_
t
0
_
t
s
(s, u)du dL
s
=
_
t
0

s,T
dL
s

_
t
0
_
u
0
(s, u)dL
s
du =
_
t
0

s,T
dL
s

_
t
0
X(u, u)du.
The diculty, however, is to show that the integrands as e.g. X(t, T) := P
n1
F
n1
(t, T)
satisfy the conditions of theorem 14. To shortly illustrate this, suppose we have shown
that theorem 13 holds for n 1. This means in particular that
_
T
0
v
n1
(s, T)
2
ds <
a.s. for all T [0, T

]. Then dP
n1
F
n1
(t, T) = F
n1
v
n1
(t, T)dW
t
+ . . . and in order to
apply theorem 14 to X = P
n1
F
n1
we must check that
_
T
0
_
T
s
F
n1
v
n1
(s, u)
2
du ds <
a.s. Obviously, this does not simply follow from H older-type inequalities.
Proof of theorem 13. First, from (16) and It os lemma it follows
d
_
1
{Lt=n}
e
t,T
n
_
= e
t,T
n
_
1
{L
t
=n}
_
f
n
(t, t)
t,T
n
+
1
2
_
_

t,T
n
_
_
2

t
_
+1
{L
t
=n1}

t
_
dt
1
{L
t
=n}
e
t,T
n

t,T
0
dW
t
1
{L
t
=n}
e
t,T
n
dL
t
+1
{L
t
=n1}
e
t,T
n
dL
t
. (17)
20
Also,
_
T
0
f
n
(s, s) +
s
+ |
s,T
n
| +
s,T
n

2
ds < a.s. for n = 0, . . . , N and all T T

.
Using P
0
(t, T) = 1
{Lt=0}
e
t,T
0
, this proves the theorem for n = 0. By induction, we may
continue assuming that the theorem holds for n 1. We rst treat U
n
.
d(P
n1
F
n1
) = y
n1
dt +x
n1
dW
t
+F
n1

n1
dL
t
where we used the abbreviations
x
n
:= P

n

F
n
+F
n
v
n
and y
n
:= P

n

F
n
+F
n
u
n
+v

F
n
for n = 0, . . . , N 1. We will now show that X := P
n1
F
n1
satises the conditions of
theorem 14. Therefor a number of preliminary results are needed. We state them in the
following lemmata and corollaries. Their proofs are postponed to appendix B. A key step
for the proof of these lemmata is given in
Proposition 3. The mapppings t
_
T

t
f
2
n
(t, u)du and t
_
T

t
F
2
n
(t, u)du are a.s. con-
tinuous for all n.
This implies that the process (t, )
_
T

t
f
2
n
(t, , u) + F
2
n
(s, , u)du, and hence also
(t, )
_
t
0
_
T

s
f
2
n
(s, , u) +F
2
n
(s, , u) du ds, are locally bounded for all n.
Proof of proposition 3. It suces to show the result for f
n
assuming f
n
(0, ) = 0. We rst
notice that
_
T

t
f
2
n
(t, u)du
_
T

0
sup
tT
f
2
n
(t u, u)du. By localization we may assume
there exists C < with
_
T

0
__
u
0

f
n
(s, u)

ds
_
2
du C and
_
T

0
_
T

s
_
_

f
n
(s, u)
_
_
2
du ds C.
Then, due to H olders and Doobs inequalities,
E
_ _
T

0
sup
tT

f
2
n
(t u, u)du
_
E
_
_
T

0
sup
tT

__
tu
0

f
n
(s, u)

ds +
_
tu
0

f
n
(s, u)dW
s
_
2
du
_
4E
_
_
T

0
__
u
0

f
n
(s, u)

ds
_
2
+ sup
tT

__
tu
0

f
n
(s, u)dW
s
_
2
du
_
4
_
T

0
E
_
__
u
0

f
n
(s, u)

ds
_
2
+ 4
_
T

u
0
_
_

f
n
(s, u)
_
_
2
ds
_
du
= 4E
_
_
T

0
__
u
0

f
n
(s, u)

ds
_
2
du + 4
_
T

0
_
T

s
_
_

f
n
(s, u)
_
_
2
du ds
_
20C < ,
which shows that
_
T

0
sup
tT
f
2
n
(tu, u)du < a.s. Then the claim follows from rewriting
_
T

t
f
2
n
(t, u)du =
_
T

0
1
{ut}
f
2
n
(t, u)du, dominated convergence and the continuity of f
n
.
Lemma 15.
_
T

0
_
T

s
x
n
(s, u)
2
+ |y
n
(s, u)| du ds < a.s. for n = 0, . . . , N 1.
Lemma 15 guarantees that := x
n1
and := y
n1
satisfy the assumptions of the-
orem 14 and that
_
T
0
b
n
(s, T)
2
+ |a
n
(s, T)| ds < a.s. for all T T

. Further
_
T

0
P
n1
F
n1
(0, T)dT
_
T

0
F
n1
(0, T)dT < . It remains to check the jump part of
P
n1
F
n1
.
21
Lemma 16. P

n

n
1
{L
t
<n}
P

n
a.s. for n = 0, . . . , N and every T T

.
Hence |
n
| 1. By localisation we may assume
_
T

s
F
n1
(s, u)du C < , which implies
_
T
0
|
n
(s, T)|
s
ds
_
T

0
_
T

s
F
n1
|
n1
| (s, u)
s
ds du C
_
T

0

s
ds < a.s. Thus theorem
14 applies to X := P
n1
F
n1
.
dU
n
(t, T) = d
__
T
t
P
n1
F
n1
(t, u)du
_
=
_
T
t
y
n1
du dt +
_
T
t
x
n1
du dW
t
+
_
T
t
F
n1

n1
du dL
t
1
{Lt=n1}
F
n1
(t, t)dt
= a
n
(t, T)dt +b
n
(t, T)dW
t
+
n
(t, T)dL
t
.
In the last equality we also used that 1
{L
t
=n1}
= 1
{Lt=n1}
PL(dt)-a.e. We continue
with P
n
. Its characteristics restricted to the set {L
t
= n} follow already from (17).
d
_
e
t,u,T
n
P
n1
F
n1
_
= e
t,u,T
n
_
P

n1
F
n1
_

t,u,T
n
+
1
2
_
_

t,u,T
n
_
_
2
_

t,u,T
n

x
n1
+y
n1
_
dt
+e
t,u,T
n
_
P

n1
F
n1

t,u,T
n
+x
n1
_
dW
t
+e
t,u,T
n
F
n1

n1
dL
t
.
Lemma 17.
_
T
0
_
T
s
F
2
n1
(s, u)
_
_

s,u,T
n
_
_
2
du ds < a.s. for every n = 1, . . . , N and T T

.
Since
_
_
P

n1
F
n1

t,u,T
n
+x
n1
_
_
2
4(F
n1
_
_

t,u,T
n
_
_
2
+x
n1

2
), lemmata 15 and 17 make
sure that the volatility of e
t,u,T
n
P
n1
F
n1
satises the assumptions of theorem 14.
Lemma 18. For every n = 1, . . . , N and T T

it holds
_
T
0
_
T
s
_
F
n1
_

s,u,T
n

+
1
2
_
_

s,u,T
n
_
_
2
_
+

s,u,T
n

x
n1

_
(s, u) du ds < a.s.
Together with lemma 15 this shows that the drift of e
t,u,T
n
P
n1
F
n1
satises the conditions
of theorem 14. We note that
_
T
0
e
0,u,T
n
P
n1
F
n1
du
_
T
0
F
n1
du < and the jump
part of e
t,u,T
n
P
n1
F
n1
is bounded by that of P
n1
F
n1
. Hence theorem 14 applies to
X := e
t,u,T
n
P
n1
F
n1
. In an intermediate step, we notice that lemma 18 also ensures that
w
n
(t, T) :=
_
T
t
e
t,u,T
n
_
P

n1
F
n1
_

t,u,T
n
+
1
2
_
_

t,u,T
n
_
_
2
_

t,u,T
n

x
n1
_
du
is nite Q L(dt)-a.e. for every t T and in what follows, we need and will not dis-
tinguish between w
n
(t, T) and 1
{|wn(t,T)|<}
w
n
(t, T). Then we have the transformations
rules below.
Lemma 19. For every n = 1, . . . , N and t T T

it holds
_
T
t
e
t,u,T
n
P

n1
F
n1
(t, u)
t,u,T
n
du = 1
{L
t
<n}
_
T
t
e
t,u,T
n
P

n

n
(t, u)du a.s. and
w
n
(t, T) = 1
{L
t
<n}
_
T
t
e
t,u,T
n
_
P

n

n
+

n
v
n
_
(t, u)du a.s.
22
Now, we are ready to carry out the integration rules from theorem 14.
d
__
T
t
e
t,u,T
n
P
n1
F
n1
du
_
=
_
T
t
e
t,u,T
n
_
1
{L
t
<n}
_
P

n

n
+

n
v
n
_
+y
n1
_
du dt
+
_
T
t
e
t,u,T
n
_
1
{L
t
<n}
P

n

n
+x
n1
_
du dW
t
+
_
T
t
e
t,u,T
n
F
n1

n1
du dL
t
e
t,t,T
n
P
n1
(t, t)F
n1
(t, t)dt.
Together with (17), e
t,t,T
n
= e
t,T
n
and 1
{L
t
=n1}
= 1
{Lt=n1}
= P
n1
(t, t) PL(dt)-a.e. this
proves (14).
Lemma 20.
_
T
0
v
n
(s, T)
2
+ |u
n
(s, T)| ds < a.s. for n = 0, . . . , N and every T T

.
Finally
_
T
0
|
n
(s, T)|
s
ds
_
T
0
_
1 +
_
T
s
F
n1
(s, u)du
_

s
ds < a.s. for all T T

, and
the statement of lemma 20 complete the proof.
4.2 EMM Relationship
It is desireable to have a rule telling how to directly set up the model such that P itself is
already an EMM. Such a rule is partly given by theorem 8 and its corollary. Unfortunately,
given the dynamics (14) and (15), we can only restrict
x
n
such that P
n
x
n
(t, T) M
loc
(P)
for all T, n and x = f, F, which is necessary in order for P to be an EMM. In fact, it will
turn out that this is already sucient in the HJM-type model.
As concerns the converse direction of theorem 8, we note that in the HJM-type forward
model, lim
TT
0
E[ |x
n
(t, T) x
n
(t, T
0
)| ] = 0 for all t T
0
T

, n = 0, . . . , N and
x = f, F is in general not satised (nor are f
n
(t, T) and F
n
(t, T) integrable). We will rst
prove a general result and then come back to this point in lemma 24 and corollary 25.
Theorem 21 (EMM HJM model). P is an EMM if and only if
P

n

f
n
+v
n

f
n
= 0 and P

n

F
n
+v
n

F
n
= 0 PL(dt, dT) a.e.
for each n = 0, . . . , N, and f
L
t
(t, t) = r
t
and F
L
t
(t, t) =
t
are predictable versions of
the short rate and the loss intensity, respectively.
Proof of theorem 21. It is easy to see that P is an EMM if and only if it holds for every
T T

u
n
= P

n
r
t
, n = 0, . . . , N
a
n
= U

n
r
t
1
{L
t
=n1}

t
, n = 1, . . . , N.
PL(dt)-a.e. From1
{L
t
=n}
v
n
(t, T) = 1
{L
t
=n}
P

n

t,T
n
and
_
_

t,T
n
_
_
2
= 2
_
T
t

t,u
n

n
(t, u)du,
we deduce
1
{L
t
=n}
u
n
(t, T) = 1
{L
t
=n}
P

n
_
f
n
(t, t)
t

_
T
t
1
{P

n
=0}
1
P

n
_
P

n

n
+v

n
_
du
_
(18)
23
Note that on {L
t
= n} it holds P
n
(t, T) = e
t,T
n
> 0 a.s. Hence
t
P
0
(t, T) are P-local
martingales if and only if 1
{L
t
=0}
_
f
0
(t, t) r
t

t
_
= 0 P L(dt)-a.e. and P

0

0
+
v

0
= 0 PL(dt, dT)-a.e.
Now, setting b
0
:= 0, a
0
:= 0 and
0
:= 0, we (may) assume we have shown that u
j
= P

j
r
t
,
j = 0, . . . , n and a
j
= U

j
r
t
1
{L
t
=j1}

t
, j = 1, . . . , n P L(dt)-a.e. for all T T

is
equivalent to
1
{L
t
n}
_
f
L
t
(t, t) r
t

t
_
= 0 PL(dt)
1
{L
t
<n}
_
F
L
t
(t, t)
t
_
= 0 PL(dt)
P

j

j
+v

j
= 0 PL(dt, dT), j = 0 . . . , n
P

j

F
j
+v

F
j
= 0 PL(dt, dT), j = 0 . . . , n 1.
(19)
We continue by induction assuming (19) holds. Then we have
a
n+1
= 1
{Lt=n}
F
n
(t, t) +U

n+1
r
t
+
_
T
t
_
P

n

F
n
+v

F
n
_
du.
Thus, a
n+1
= U

n+1
r
t
1
{L
t
=n}

t
P L(dt)-a.e. holds for all T T

if and only if
1
{L
t
=n}
_
F
L
t
(t, t)
t
_
= 0 P L(dt)-a.e. and P

n

F
n
+ v

F
n
= 0 P L(dt, dT)-a.e.
Suppose the latter two conditions hold, then
1
{L
t
<n+1}
u
n+1
(t, T) = 1
{L
t
<n+1}
_
P

n+1
r
t
e
t,T
n
_
T
t
(e
t,u
n
)
1
_
P

n+1

n+1
+

n+1
v
n+1
_
du
_
.
Together with (18), this shows that (under the conditions in (19) above) u
n+1
= P

n+1
r
t
P(dt)-a.e. if and only if 1
{L
t
n+1}
_
f
L
t
(t, t) r
t

t
_
PL(dt)-a.e. and P

n+1

n+1
+

n+1
v
n+1
= 0 PL(dt, dT)-a.e. Repeating this step until n + 1 = N, this proves that P
is an EMM if and only if (19) holds for n = N. Then the claim follows, since f
N
= f
N
(F
N
= 0).
Theorem 21 suggest to dene r
t
:= f
L
t
(t, t) and
Q
t
:= F
L
t
(t, t) and

x
n
:=
x,Q
n
:=
v

x
n
P

n
for each n = 0, . . . , N and x = f, F in order to set up the model directly under an EMM Q.
However, in general, for
x
n
satisfying (ii) of assumption 2,
x,Q
n
does not automatically
satisfy (iii) of assumption 2. Nevertheless, we can exclude these cases by imposing a
growth condition on the volatilities
F
n
. Interestingly, we need not to further restrict
f
n
beyond assumption (ii) of 2.
Lemma 22. Assume there exists a predictable process
t
with
_
T

0

2
t
dt < a.s. and such
that
_
_

F
n
(t, T)
_
_

t
F
n
(t, T) for all t T T

and all n = 0, . . . , N 1. Then

_
T

0
__
u
0

Q,x
n
(s, u)

ds
_
2
du < a.s.
for each n = 0, . . . , N and x = f, F.
24
Proof of lemma 22. We start with a preliminary result.
Lemma 23. Assume there exists a predictable process
t
such that
_
_

F
n
(t, T)
_
_

t
F
n
(t, T)
for all t T T

and all n = 0, . . . , N 1. Then

v
n
(t, T) P
n
(t, T)
_
n

j=0
_
T
_
T
t

j
(t, v)
2
dv
_
1/2
+ (n 1)
t
_
for all t T T

and n = 0, . . . , N.
By lemma 23, we can rst estimate
Q,x
n
as follows.

Q,x
n
(s, u)

=
|v
n
(s, u)

x
n
(s, u)|
P
n
(s, u)

v
n
(s, u)
x
n
(s, u)
P
n
(s, u)

_
n

j=0
_
u
_
u
s

j
(t, v)
2
dv
_
1/2
+ (n 1)
s
_

x
n
(s, u)
Using this estimation and H olders inequality, we observe
_
T

0
__
u
0

x,Q
n
(s, u)

ds
_
2
du 4
_
T

0
_
_
u
0
n

j=0
_
u
_
u
s

j
(s, v)
2
dv
_
1/2

x
n
(s, u) ds
_
2
du
+4
_
T

0
__
u
0
(n 1)
s

x
n
(s, u) ds
_
2
du
4n
2
n

j=0
_
T

0
__
u
0
u
_
u
s

j
(s, v)
2
dv ds
___
u
0

x
n
(s, u)
2
ds
_
du
+4n
2
_
T

0
__
u
0

2
s
ds
___
u
0

x
n
(s, u)
2
ds
_
du
4n
2
T

j=0
__
T

0
_
T

j
(s, v)
2
dv ds
__
T

0
_
u
0

x
n
(s, u)
2
ds du
+4n
2
__
T

2
s
ds
__
T

0
__
u
0

x
n
(s, u)
2
ds
_
du < .
Proof of lemma 23. The lemma holds for n = 0 because
v
0
(t, T) = P
0
(t, T)
_
_
_
t,T
0
_
_
_ P
0
(t, T)
_
T
_
T
t

0
(t, v)
2
dv
_
1/2
for all t T T

. We continue by induction assuming the lemma holds for n. Then, by

25
the denition of v
n+1
and lemma 19,
v
n+1
(t, T) 1
{L
t
=n+1}
P
n+1
(t, T)
_
_
_
t,T
n+1
_
_
_
+1
{L
t
n}
_
T
t
e
t,u,T
n+1
_
P

n
_
F
n
_
_
_
t,u,T
n+1
_
_
_ +
_
_

F
n
_
_
_
+F
n
v
n

_
du
1
{L
t
=n+1}
P
n+1
(t, T)
_
T
_
T
t

n+1
(t, v)
2
dv
_
1/2
+1
{L
t
n}
_
T
t
e
t,u,T
n+1
P

n
F
n
_
_
T
_
T
u

n+1
(t, v)
2
dv
_
1/2
+
t
_
du
+1
{L
t
n}
_
T
t
e
t,u,T
n+1
F
n
P

n
_
n

j=0
_
u
_
u
t

j
(t, v)
2
dv
_
1/2
+ (n 1)
t
_
du
1
{L
t
=n+1}
P
n+1
(t, T)
_
T
_
T
t

n+1
(t, v)
2
dv
_
1/2
+1
{L
t
n}
_
n+1

j=0
_
T
_
T
t

j
(t, v)
2
dv
_
1/2
+n
t
_
_
T
t
e
t,u,T
n+1
F
n
P

n
du
P
n+1
(t, T)
_
n+1

j=0
_
T
_
T
t

j
(t, v)
2
dv
_
1/2
+n
t
_
.
Finally, we come back to the question when f
n
and F
n
as considered in assumption 2
satisfy the conditions of theorem 8.
Lemma 24. Let f
n
(0, T) and F
n
(0, T) be in C
0,r
(T) for all n,
x
n
and
x
n
be in C
0,r
(T)
for all t [0, T

] a.s. for all n and x = f, F, and assume there exists processes

t
and

t
with E
_
_
T

0

t
+
2
t
dt
_
< such that |
x
n
|
t
and
x
n

t
, for all n, T t and
x = f, F. Then f
n
(t, T) and F
n
(t, T) are integrable and
lim
TT
0
E[ |x
n
(t, T) x
n
(t, T
0
)| ] = 0
for all t T
0
T

, n = 0, . . . , N and x = f, F.
Under the conditions of lemma 24, the EMM drift restrictions obtained in theorem
21 follow straighforward from applying It os lemma to the products
t
P
n
f
n
(t, T) and

t
P
n
F
n
(t, T), which must be martingales under every EMM by theorem 8.
Proof of lemma 24. From the It o isometry it is clear that f
n
(t, T) and F
n
(t, T) are inte-
26
grable. Further
E[ |x
n
(t, T) x
n
(t, T
0
)| ] E
_ _
t
0
|
x
n
(s, T)
x
n
(s, T
0
)| ds
_
+E
_

_
t
0

x
n
(s, T)
x
n
(s, T
0
)dW
s

_
E
_ _
t
0
|
x
n
(s, T)
x
n
(s, T
0
)| ds
_
+E
_ _
t
0

x
n
(s, T)
x
n
(s, T
0
)
2
ds
_
1/2
.
Then the claim follows with a dominated convergence argument letting T T
0
.
As a concatenation of lemma 24 and theorem 8 we obtain
Corollary 25. Under the conditions of lemma 24 the following assertions are equivalent.
(i) P is an EMM.
(ii) P

n

f
n
+v
n

f
n
= 0 and P

n

F
n
+v
n

F
n
= 0 PL(dt, dT)-a.e. for each n, and f
L
t
(t, t) =
r
t
and F
L
t
(t, t) =
t
are predictable versions of the short rate and the loss intensity.
(iii)
t
P
n
f
n
(t, T) and
t
P
n
F
n
(t, T) are in M(P) for all n and T, and f
L
t
(t, t) = r
t
and
F
L
t
(t, t) =
t
are predictable versions of the short rate and the loss intensity.
References
N. Bennani. The forward loss model: A dynamic term structure approach for the pricing
of portfolio credit derivatives. Working paper, The Royal Bank of Scotland, 2005.
Damiano Brigo, Andrea Pallavicini, and Roberto Torresetti. Calibration of CDO tranches
with the dynamical generalized-poisson loss model. Working paper, Banca IMI, Milano,
May 2006.
Pierre Collin-Dufresne, Robert S. Goldstein, and J. Spencer Martin. The determinants of
credit spread changes. Journal of Finance, 2001.
Mark Davis and Violet Lo. Infectious defaults. Quantitative Finance, 1:382387, 2001.
Gregory R. Duee. Estimating the price of default risk. The Review of Financial Studies,
12(1):197226, 1995.
Gregory R. Duee. The relation between treasury yields and corporate bond yield spreads.
The Journal of Finance, 53(6):22252242, 12 1998.
Darrel Due and Kenneth Singleton. An econometric model of the term structure of
interest rate swap yields. Journal of Finance, 52(4):12871321, 1997.
27
Darrel Due, Andreas Eckner, Horel Guillaume, and Leandro Saita. Frailty correlated
default. Working paper, Stanford University, 2006.
R. Frey and J. Backhaus. Portfolio credit risk models with inter-
acting default intensities : a Markovian approach. Working pa-
per, Department of Mathematics, Universitt Leipzig, 2004. URL
http://www.mathematik.uni-leipzig.de/MI/frey/interacting-intensities-final.pdf.
Kay Giesecke and Lisa Goldberg. A top down approach to multi-name
credit. Working paper, Cornell University, September 2005. URL
http://www.stanford.edu/dept/MSandE/people/faculty/giesecke/topdown.pdf.
David Heath, Robert Jarrow, and Andrew Morton. Bond pricing and the term structure
of interest rates: A new methodology for contingent claims valuation. Econometrica,
60:77105, 1992.
L.P. Hughston and A. Rafailidis. A chaotic approach to interest rate modelling. Finance
and Stochastics, 9(1):4365, 2005.
Farshid Jamshidian. Pricing of contingent claims in the one-factor term structure model.
Working paper, Sakura Global Capital, July 1987.
Robert A. Jarrow, David Lando, and Stuart M. Turnbull. A Markov model for the term
structure of credit risk spreads. The Review of Financial Studies, 10(2):481523, 1997.
David. X. Li. On default correlation: A Copula function approach. working paper 99-07,
Risk Metrics Group, April 2000.
Francis Longsta and Arvind Rajan. An empirical analysis of the pricing of collateralized
debt obligations. Working paper, Anderson School, UCLA, April 2006.
Dilip Madan and Haluk Unal. Pricing the risks of default. Review of Derivatives Research,
2(2/3):121160, 1998.
Philip Protter. Stochastic Integration and Dierential Equations. Springer, second edition,
2004.
P. Sch onbucher. Portfolio losses and the term structue of loss transition rates: A new
methodology for the pricing of portfolio credit derivatives. Working paper, Department
of Mathematics, ETH Z urich, September 2005.
Jakob Sidenius, Vladimir Piterbag, and Leif Andersen. A new framework for dynamic
credit portfolio loss modelling. Working paper, Royal Bank of Scotland and Barclays
Capital and Bank of America, October 2005.
28
A Proof of Theorem 2
Proof of theorem 2. It is clear that P
0
= . . . = P
Lt1
= U
1
= . . . = U
Lt
= 0 and
P
Lt
(t, T) = e
t,T
Lt
(0, 1]. Hence if L
t
= n, then (B), (P) and (U) hold.
On the other hand, if L
t
< n, we may assume that for some n
0
L
t
, we have P
n
(t, T), U
n
(t, T)
[0, 1], U
n
(t, T) is nondecreasing in T, and P
n
and U
n
satisfy U
n
(t, T) =
_
T
t
P
n1
(t, u)F
n1
(t, u)du
and P
n
(t, T) = 1
{Lt=n}

_
T
t
P
n
(t, u) (f
n
(t, u) +F
n
(t, u)) du +U
n
(t, T) for all n n
0
, This
implies in particular U
n
(t, t) = 0 and P
n
(t, t) = 1
{Lt=n}
for all n n
0
.
Then U
n
0
+1
(t, T) :=
_
T
t
P
n
0
(t, u)F
n
0
(t, u)du and P
n
0
+1
(t, T) :=
_
T
t
e
t,u,T
n
0
+1
P
n
0
(t, u)F
n
0
(t, u)du
are well-dened and nonnegative, U
n
0
+1
(t, T) is nondecreasing in T, U
n
0
+1
(t, t) = 0 and
P
n
0
+1
(t, t) = 1
{Lt=n
0
+1}
by construction. Further
P
n
0
+1
(t, T) U
n
0
+1
(t, T)
_
T
t
P
n
0
(t, u) (f
n
0
(t, u) +F
n
0
(t, u)) du
= P
n
0
(t, t) P
n
0
(t, T) +U
n
0
(t, T) 1
and, using the abbreviation f
n
(t, T) := f
n
(t, T) +F
n
(t, T),
P
n
0
+1
(t, T) =
_
T
t
_
e
t,T
n
0
+1
e
t,v
n
0
+1
_
(e
t,v
n
0
+1
)
1
P
n
0
(t, v)F
n
0
(t, v)dv +
_
T
t
P
n
0
(t, v)F
n
0
(t, v)dv
=
_
T
t
__
T
v
e
t,u
n
0
+1
f
n
0
+1
(t, u)du
_
(e
t,v
n
0
+1
)
1
P
n
0
(t, v)F
n
0
(t, v)dv +U
n
0
+1
(t, T)
=
_
T
t
e
t,u
n
0
+1
_
u
t
(e
t,v
n
0
+1
)
1
P
n
0
(t, v)F
n
0
(t, v)dv f
n
0
+1
(t, u)du +U
n
0
+1
(t, T)
= 1
{Lt=n
0
+1}

_
T
t
P
n
0
+1
(t, u)f
n
0
+1
(t, u)du +U
n
0
+1
(t, T),
This way (P), (U), B(t, T) > 0 (since P
Lt
(t, T) > 0) and B(t, t) = 1 follow by induction
in n
0
. By denition we have B(t, T) = 1

N
n=0

n
(t, T). Hence it remains to show ().
But by the above, we have for all n N

n
(t, T) =
_
T
t
P
n
(t, u)f
n
(t, u)du,
which is nondecreasing in T.
B Proofs of Lemmata used in Proof of Theorem 13
Proof of lemma 15. We show the lemma in three steps. By localisation, we can always
assume without restriction
_
T

s
F
2
n
(s, u)du C < for all s T

(cf. proposition 3).

Step 1:
_
T

0
_
T

s
x
n
(s, u)
2
du ds < a.s. for n = 0, . . . , N 1.
Clearly, x
n
(s, u)
2
4(
_
_

F
n
(s, u)
_
_
2
+ F
2
n
(s, u) v
n
(s, u)
2
). It thus suces to show
29
_
T

0
_
T

s
F
2
n
(s, u) v
n
(s, u)
2
du ds < . With H olders inequality, we obtain
s,u
n

2

u
_
u
s

n
(s, v)
2
dv, and setting x
1
(s, u) := 0, we notice that
v
n
(s, u)
2
u
_
u
s

n
(s, v)
2
dv + 4u
_
u
s

n
(s, v)
2
+x
n1
(s, v)
2
dv
5u
_
u
s

n
(s, v)
2
+x
n1
(s, v)
2
dv (20)
for n = 0, . . . , N. Using
_
T

s
F
2
n
(s, u)du C < for all s T

and (20), we nd
_
T

0
_
T

s
F
2
n
(s, u) v
n
(s, u)
2
du ds 5
_
T

0
_
T

s
F
2
n
(s, u) u
_
u
s

n
(s, v)
2
+x
n1
(s, v)
2
dv du ds
5T
_
T

0
_
T

s
F
2
n
(s, u)du
_
T
s

n
(s, v)
2
+x
n1
(s, v)
2
dv ds
5CT
_
T

0
_
T

n
(s, v)
2
+x
n1
(s, v)
2
dv ds.
For n = 0 the righthandside is a.s. nite, which implies
_
T

0
_
T

s
x
0
(s, u)
2
du ds < a.s.,
then Step 1 follows by induction in n. Step 1
Step 2:
_
T

0
_
T

s
v
n
(s, u)
2
du ds < a.s. for n = 0, . . . , N.
We set x
1
:= 0. With (20) and the last estimate in Step 1, we immediately see that
_
T

0
_
T

s
v
n
(s, u)
2
du ds 5
_
T

0
_
T

s
u
_
u
s

n
(s, v)
2
+x
n1
(s, v)
2
dv du ds
5T

_
T

0
_
T

s
_
T

n
(s, v)
2
+x
n1
(s, v)
2
du dv ds
5T
2
_
T

0
_
T

n
(s, v)
2
+x
n1
(s, v)
2
dv ds < a.s.
for n = 0 . . . , N. Then Step 2 follows from Step 1. Step 2
Step 3:
_
T

0
_
T

s
|y
n
(s, u)| du ds < a.s. for n = 0, . . . , N 1.
First of all |y
n
(s, u)| |
F
n
(s, u)| + F
n
(s, u) |u
n
(s, u)| + |v

F
n
(s, u)|. By the Step 2
and H olders inequality
_
T

0
_
T

s
|v

F
n
(s, u)|du ds < a.s. It is thus sucient to show
_
T

0
_
T

s
F
n
(s, u) |u
n
(s, u)| du ds < a.s. for n = 0, . . . , N 1. Setting y
1
:= 0, it holds
|u
n
(s, u)| f
n
(s, s) +
s
+
_
u
s
_
2 |
n
| +
u
2

2
+ |y
n1
| + |

n
v
n
|
_
(s, v)dv (21)
for n = 0, . . . , N. Using this and
_
T

s
F
n
(s, u)du C < , we obtain
_
T

0
_
T

s
F
n
|u
n
| (s, u)du ds
_
T

0
_
T

s
F
n
(s, u)
_
u
s
_
2 |
n
| +
u
2

2
+ |y
n1
| + |

n
v
n
|
_
dv du ds
+
_
T

0
_
f
n
(s, s) +
s
_
_
T

s
F
n
(s, u)du ds
C
_
T

0
_
T

s
_
2 |
n
| +
T

2

n

2
+ |y
n1
| + |

n
v
n
|
_
(s, v)dv ds
+C
_
T

0
f
n
(s, s) +
s
ds
30
Clearly,
_
T
0
_
T
s
|

n
v
n
| (s, v)dv ds < a.s. for n = 0, . . . , N is again due to H olders inequal-
ity and Step 2. Thus for n = 0 the righthandside above is nite a.s., and Step 3 follows
by induction in n. Step 3
Now, the proof of lemma 15 is direct from Step 1 and Step 3.
Proof of lemma 16. For n = 0 the claim is trivially satised. First, we check the lower
bound assuming the lower bound rule holds for n, i.e.
n
P

n
. Then

n+1
1
{L
t
=n+1}
P

n+1

_
T
t
e
t,u,T
n+1
F
n
P

n
(t, u)du = P

n+1
.
For the upper bound we need an explicit formula for
n
, which is given in the following
lemma.
Lemma 26.
0
= P

0
,
1
= P

1
+1
{L
t
=0}
e
t,T
1
and for n 2 it holds

n
= P

n
+1
{Lt=n1}
e
t,T
n
+
n1

j=1
1
{Lt=j1}

T
t
e
t,uj
j
F
j
(t, u
j
)

T
uj
e
t,uj,uj+1
j+1
F
j+1
(t, u
j+1
)

T
un2
e
t,un2,un1
n1
F
n1
(t, u
n1
) e
t,un1,T
n
du
n1
du
j+1

du
j
. (22)
With lemma 26, the upper bound is easily checked for n = 0, 1. Further, for each k =
0, . . . , N 1, we have
_
T
t
e
t,v
k
F
k
(t, v)dv
_
T
t
e
t,v
k
f
k
(t, v)dv = 1 e
t,T
k
1 for all t T,
and similarly
_
T
u
e
t,u,v
k
F
k
(t, v)dv 1 e
t,u,T
k
1 for all t u T. Using these estimates
iteratively, we observe that every multi-integral in (22) is bounded from above by one.
Thus

n
P

n
+1
{L
t
=n1}
+
n1

j=1
1
{L
t
=j1}
= 1
{L
t
<n}
P

n
.
Proof of lemma 17. From H olders inequality
_
_

s,u,T
n
_
_
2
T
_
T
u

n
(s, v)
2
dv, and by lo-
calisation, we may assume
_
T
s
F
2
n1
(s, u)du C < for all s T. Then
_
T
0
_
T
s
F
2
n1
(s, u)
_
_

s,u,T
n
_
_
2
du ds
_
T
0
_
T
s
F
2
n1
(s, u) T
_
T
u

n
(s, v)
2
dv du ds
T
_
T
0
__
T
s
F
2
n1
(s, u)du
__
T
s

n
(s, v)
2
dv ds
C
_
T
0
_
T
s

n
(s, v)
2
dv ds < a.s.
Proof of lemma 18. By localisation, we may always assume
_
T
s
F
n1
(s, u)du C < .
We will treat every summand of the integrand separately. First
_
T
0
_
T
s
F
n1
(s, u)

s,u,T
n

du ds
_
T
0
_
T
s
F
n1
(s, u)
__
T
u
|
n
(s, v)| dv
_
du ds

_
T
0
__
T
s
F
n1
(s, u)du
__
T
s
|
n
(s, v)| dv ds
C
_
T
0
_
T
s
|
n
(s, v)| dv ds < a.s.
31
Second, since
_
_

s,u,T
n
_
_
2
T
_
T
u

n
(s, v)
2
dv,
_
T
0
_
T
s
F
n1
(s, u)
_
_

s,u,T
n
_
_
2
du ds T
_
T
0
_
T
s
F
n1
(s, u)
_
T
u

n
(s, v)
2
dv du ds
CT
_
T
0
_
T
s

n
(s, v)
2
dv ds <
Third, by the denition of x
n1
, we have x
n1
F
n1
v
n1
+
_
_

F
n1
_
_
. With (20) we
nd
Z
T
0
Z
T
s
F
n1

s,u,T
n

v
n1
duds
Z
T
0
Z
T
s
F
n1

T
Z
T
u
n(s, v)
2
dv 5u
Z
u
s
n(s, v)
2
+x
n1
(s, v)
2
dv

1
2
duds

5T
Z
T
0
Z
T
s
F
n1
Z
T
s
n(s, v)
2
dv
Z
T
s
n(s, v)
2
+x
n1
(s, v)
2
dv

1
2
duds

5TC
Z
T
0
Z
T
s
n(s, v)
2
+x
n1
(s, v)
2
dv
Z
T
s
n(s, v)
2
dv

1
2
ds

5TC
Z
T
0
Z
T
s
n(s, v)
2
+x
n1
(s, v)
2
dv ds
Z
T
0
Z
T
s
n(s, v)
2
dv ds

1
2
< a.s.
and on the other hand (remember
_
_

s,u,T
n
_
_
2
T
_
T
s

n
(s, v)
2
dv)
_
T
0
_
T
s
F
n1
_
_

F
n1
_
_
_
_

s,u,T
n
_
_
du ds
_
T
0
_
T
s
F
n1
_
_

F
n1
_
_
_
T
_
T
s

n
(s, v)
2
dv
_
1
2
du ds

_
T
_
T
0
__
T
s
F
n1
_
_

F
n1
_
_
du
_
2
ds
_
T
0
_
T
s

n
(s, v)
2
dv ds
_1
2

_
T
_
T
0
_
T
s
F
2
n1
du
_
T
s
_
_

F
n1
_
_
du ds
_
T
0
_
T
s

2
dv ds
_
1
2

_
TC
_
T
0
_
T
s
_
_

F
n1
_
_
du ds
_
T
0
_
T
s

2
dv ds
_
1
2
< a.s.
Proof of lemma 19. First,
_
T
t
F
n1
(t, u)
t,u,T
n
du
_
T
t
F
n1
(t, u)du
_
T
t

n
(t, v) dv <
a.s. holds for all t T T

. Recall e
t,u,T
n
= e
t,u,v
n
e
t,v,T
n
for t u v T. Then
_
T
t
e
t,u,T
n
P

n1
F
n1
(t, u)
t,u,T
n
du =
_
T
t
e
t,u,T
n
P

n1
F
n1
(t, u)
_
T
u

n
(t, v)dv du
=
_
T
t
__
v
t
e
t,u,v
n
P

n1
F
n1
(t, u)du
_
e
t,v,T
n

n
(t, v)dv
= 1
{L
t
<n}
_
T
t
e
t,v,T
n
P

n

n
(t, v)dv.
For the second result, we will carry out three Fubini-type transformations separately.
32
Admissibility is guaranteed by lemma 18.
_
T
t
e
t,u,T
n
P

n1
F
n1
(t, u)
t,u,T
n
du =
_
T
t
_
T
u
e
t,u,T
n
P

n1
F
n1
(t, u)
n
(t, v)dv du
=
_
T
t
_
v
t
e
t,u,v
n
P

n1
F
n1
(t, u)du e
t,v,T
n

n
(t, v)dv
= 1
{L
t
<n}
_
T
t
e
t,v,T
n
P

n

n
(t, v)dv
It is easy to verify that
1
2
_
_

t,u,T
n
_
_
2
=
_
T
u
_ _
T
v

n
(t, y)dy
_

n
(t, v)dv, and thus
Z
T
t
e
t,u,T
n
P

n1
F
n1
1
2

t,u,T
n

2
du =
Z
T
t
Z
T
u
e
t,u,T
n
P

n1
F
n1
(t, u)
Z
T
v
n(t, y)dy

n(t, v)dv du
=
Z
T
t
Z
v
t
e
t,u,v
n
P

n1
F
n1
(t, u)due
t,v,T
n
Z
T
v
n(t, y)dy

n(t, v)dv
= 1
{L
t
<n}
Z
T
t
Z
T
v
e
t,v,T
n
P

n
(t, v)n(t, y)

n(t, v)dy dv
= 1
{L
t
<n}
Z
T
t
e
t,y,T
n
n(t, y)

Z
y
t
e
t,v,y
n
P

n
n(t, v)dv

dy
not.
= 1
{L
t
<n}
Z
T
t
e
t,v,T
n
n(t, v)

Z
v
t
e
t,u,v
n
P

n
n(t, u)du

dv
_
T
t
e
t,u,T
n

t,u,T
n

x
n1
(t, u)du ds =
_
T
t
_
T
u
e
t,u,T
n

n
(t, v)

x
n1
(t, u)dv du
=
_
T
t
_
v
t
e
t,u,T
n

n
(t, v)

x
n1
(t, u)du dv
=
_
T
t
e
t,v,T
n

n
(t, v)

__
v
t
e
t,u,v
n
x
n1
(t, u)du
_
dv
Combining these three results and the denition of v
n
yields the claim.
Proof of lemma 20. We set x
1
:= 0 and y
1
:= 0. Using (20) we see that
_
T
0
v
n
(s, T)
2
ds 5T
_
T
0
_
T
s

n
(s, v)
2
+x
n1
(s, v)
2
dv ds < a.s.
for every n and T T

. In the proof of lemma 15 we argue

_
T
0
_
T
s
|

n
v
n
| (s, v)dv ds <
a.s. Then with (21) and lemma 15, we nd directly that for n = 0 . . . , N and all T T

_
T
0
|u
n
(s, T)| ds
_
T
0
_
T
s
_
2 |
n
| +
T
2

n

2
+ |y
n1
| + |

n
v
n
|
_
(s, v)dv ds
+
_
T
0
f
n
(s, s) +
s
ds < a.s.
Proof of lemma 26.
0
and
1
follow directly from (13). Using
1
= P

1
+1
{L
t
=0}
e
t,T
1
,
we nd

2
= 1
{L
t
=2}
P

2
+1
{L
t
=1}
e
t,T
2
+
_
T
t
e
t,u,T
2
F
1
(t, u)
_
P
1
(t, u) +1
{L
t
=0}
e
t,u
1
_
du
= P

2
+1
{L
t
=1}
e
t,T
2
+1
{L
t
=0}
_
T
t
e
t,u
1
F
1
(t, u)e
t,u,T
2
du,
33
i.e the lemma holds for n = 2. Then we may continue by induction and indeed

n+1
= 1
{Lt=n+1}
P

n+1
+1
{Lt=n}
e
t,T
n+1
+

T
t
e
t,un,T
n+1
F
n

n
(t, u
n
)du
n
= 1
{Lt=n+1}
P

n+1
+1
{Lt=n}
e
t,T
n+1

T
t
e
t,un,T
n+1
P

n
F
n
(t, u
n
)du
n
+1
{Lt=n1}

T
t
e
t,un
n
F
n
(t, u
n
)e
t,un,T
n+1
du
n
+
n1

j=1
1
{Lt=j1}

T
t
e
t,un,T
n+1
F
n
(t, u
n
)

un
t
e
t,uj
j
F
j
(t, u
j
)

un
uj
e
t,uj,uj+1
j+1
F
j+1
(t, u
j+1
)

un
un2
e
t,un2,un1
n1
F
n1
(t, u
n1
) e
t,un1,un
n
du
n1
du
j+1

du
j

du
n
= P

n+1
+1
{Lt=n}
e
t,T
n+1
+1
{Lt=n1}

T
t
e
t,un
n
F
n
(t, u
n
)e
t,un,T
n+1
du
n
+
n1

j=1
1
{Lt=j1}

T
t
e
t,uj
j
F
j
(t, u
j
)

T
uj
e
t,uj,uj+1
j+1
F
j+1
(t, u
j+1
)

un
un2
e
t,un2,un1
n1
F
n1
(t, u
n1
)

T
un1
e
t,un1,un
n
F
n
(t, u
n
) e
t,un,T
n+1
du
n

du
n1
du
j+1
du
j

= P

n+1
+1
{Lt=n}
e
t,T
n+1
+
n

j=1
1
{Lt=j1}

T
t
e
t,uj
j
F
j
(t, u
j
)

T
uj
e
t,uj,uj+1
j+1
F
j+1
(t, u
j+1
)

T
un1
e
t,un1,un
n
F
n
(t, u
n
) e
t,un,T
n+1
du
n
du
j+1

du
j
.
34
c/o University of Geneva
40 bd du Pont d'Arve
1211 Geneva 4
Switzerland
T +41 22 379 84 71
F +41 22 379 82 77
RPS@sfi.ch
www.SwissFinanceInstitute.ch
ECF-SFI 06 25.1.2006 16:05 Page 24