1 StochasticCounterparty05

Section 1
Risk Management at
Counterparty level
01-Prisco.qxd 11/12/05 11:05 AM Page 1
3
The past two decades have witnessed a significant increase in over-
the-counter trading volumes as well as the dramatic fall of several
notable institutions such as Barings, Long Term Capital, Enron,
WorldCom and Parmalat. These factors have brought more pres-
sure on financial institutions to quantify credit risk accurately in
order to set limits and price credit lines appropriately. A recent
report by the Counterparty Risk Management Policy Group (CRMPG),
comprised of senior officials from major financial institutions, iden-
tified counterparty risk as . . . probably the single most important
variable in determining whether and with what speed financial distur-
bances become financial shocks, with potential systemic traits (CRPMG
2005).
The first level of portfolio credit risk analysis is to obtain a con-
solidated picture of the exposures that result from all the transac-
tions with a counterparty. The success of a credit risk assessment
depends on the ability to consolidate and measure accurately the
exposures in both the trading and banking books. Often, this is a
major systems exercise that involves the collection of all positions
with that counterparty (across all instruments, systems and geog-
raphies), the application of netting and credit mitigation agree-
ments and the management of collateral. Once consolidated,
exposures can be managed by monitoring limits against each coun-
terparty at different levels in the portfolio. Credit lines are set and
1
Modelling Stochastic
Counterparty Credit Exposures
for Derivatives Portfolios
Ben De Prisco; Dan Rosen
Algorithmics Inc; Fields Institute, University of Toronto
3
COUNTERPARTY CREDIT RISK MODELLING
4
controlled by the credit officer, limiting the maximum amount that
could be lost if any counterparty were to default. The sizes of
the credit lines are generally determined taking into account the
probability of default, collateralisation and mitigation, risk appetite
of the institution, etc. Managing credit limits on a counterparty-
by-counterparty basis has proven to be a simple, effective and
actionable risk management tool.
In the case of over-the-counter derivatives portfolios, the calcu-
lation of exposures is an involved matter. This can be thought of at
three levels.
First, exposures of derivatives are stochastic. Risk managers not
only measure the current exposure but also potential future
changes at many time horizons, often until the maturity of all
contracts. Potential future exposures can change substantially
over time according to movements in market variables. Thus,
the underlying market processes have a substantial impact on
these exposures. Furthermore, exposures may be path depend-
ent. For example, at a point in time, the exposure of a portfolio
that contains American or barrier options depends on whether
the options have been exercised or the barriers hit. Finally, expos-
ure profiles over time are largely discontinuous due to contract
expiries, options exercises, coupons, etc.
Second, credit mitigation techniques are commonly used to
manage counterparty risk and to reduce credit charges and capi-
tal. Exposure measures require detailed modelling of netting,
collateral and other mitigation techniques. Furthermore, collat-
eral values may be stochastic and path dependent. For example,
the collateral amount posted within a margin agreement can
depend on the level of exposure at a previous time, when the last
margin call was made.
Third, exposures and collateral may be dependent on credit
events in addition to market prices. Credit-state dependent
exposures arise, eg, when the portfolio contains credit derivatives
or credit-dependent collateral triggers. Furthermore, in some
cases, the exposure to a counterparty may be highly linked (or
correlated) to its credit-worthiness. Managing such wrong-way
exposures also requires consideration of joint market moves and
credit events.
MODELLING STOCHASTIC COUNTERPARTY CREDIT EXPOSURES FOR DERIVATIVES PORTFOLIOS
5
Various simplified methodologies have been introduced to calcu-
late counterparty exposures, such as add-on methods and analyt-
ical approximations. However, given the complex stochastic nature
of exposure profiles, as well as the detailed modelling required,
Monte Carlo (MC) simulation remains the most general and reli-
able approach.
Exposure is only one component of credit risk the other com-
ponents are the probabilities of default and migration, recovery
rates, credit spreads and correlations of credit events. Therefore,
limits management on its own does not provide a full picture
of the portfolio risk and can result in overly conservative poli-
cies. Counterparty exposure measurement and limits must
be consistent and readily integrated with credit pricing, capital
allocation and business decision support tools. The intricate sto-
chastic nature of derivatives exposures is now recognised by prac-
titioners and regulators alike. Thus, it is an important component
of credit pricing, economic capital and the Basel II regulatory
framework (Basel Committee of Banking Supervision, BCBS 2004,
2005).
This chapter presents a general overview of current counter-
party exposure measurement practices and recent modelling
advances. Our focus is on a cohesive presentation of the quantita-
tive modelling aspects of counterparty exposures and mitigation
techniques as well as on important implementation details. The
chapter is organised as follows.
The second section introduces the key definitions, the modelling
framework and exposure measures.
The third and fourth sections present the main concepts behind
MC simulation and analytical approximations and discuss their
advantages and limitations.
The fifth section identifies and discusses various advanced mod-
elling issues:
the impact on exposures measurement of the underlying scen-
ario models and the time steps in the simulation;
the modelling of collateral agreements with include lags and
credit-state dependent thresholds;
the treatment of credit derivatives and wrong-way exposures.
6
Several books and articles provide general presentations of coun-
terparty exposure measurement and modelling (eg, Arvanitis and
Gregory 2001; Schonbucher 2004) and credit mitigation (Wakeman
1996; Levy and Clarke 2000). Comprehensive treatments of MC
simulation in finance can be found, eg, in Glasserman (2004) and
Dupire (2004). We further provide references on various specific
issues treated in each section. Finally, the chapters in this book
provide further modelling details on counterparty exposures,
pricing credit risk and economic capital.
BASIC CONCEPTS POTENTIAL FUTURE EXPOSURES,
CREDIT MITIGATION AND RISK MANAGEMENT
This chapter focuses on default credit risk the risk of loss that
will be incurred in the event that a counterparty to a set of trans-
actions defaults ie, the counterparty fails to honour its contrac-
tual payments.
1
More specifically, our main objective is to model
pre-settlement counterparty credit risk. This is the risk that the
counterparty defaults before the final settlement of the transac-
tions cashflows. This loss depends on the replacement (eco-
nomic) value of the contracts at the time of default. In contrast,
settlement counterparty risk refers to the loss realised when con-
tractual payments are not received on the settlement date due to
default.
We define counterparty exposure as the economic loss incurred on
all outstanding transactions if the counterparty defaults.
2
It is
essentially the cost of replacing or hedging the set of contracts at
the time of default. Exposure measures account for netting and col-
lateral, but are generally not adjusted by possible recoveries.
Risk managers are interested in both the current exposure as
well as the potential future exposure (PFE) the potential future
changes in exposures during the contracts lives. This is particu-
larly important for derivatives whose values can change substan-
tially over time according to the state of the market. PFEs take into
account the aging of the portfolio as well as the movement of
underlying market factors that directly affect contract values at
future times. For example, consider a pay-fix interest rate swap
with negative mark-to-market value and current exposure of zero.
If future interest rates rise, the contract can have a positive value,
resulting in potential additional exposure to the holder.
7
Consider a portfolio of transactions with a single counterparty and
assume that the future is described by a discrete set of times and
scenarios.
3
The set of times is denoted by {t
0
, t
1
, . . ., t
N
T}, with t
0
denoting today and T the horizon date (typically the longest matu-
rity). Ascenario is a path containing all the market information up to
T. Denote by S
j
(t
k
) the state of scenario j at t
k
and by S
j
(t
k
) [S
j
(t
0
),
S
j
(t
1
), . . ., S
j
(t
k
)] the path (history) of the scenario up to t
k
.
The PFE profile
For a single position p with a counterparty, we define the PFE at
time t
k
and scenario S
j
, as the maximum of zero and the contract
value (if it were replaced in the market)
4
(1)
where V(p; S
j
(t
k
), t
k
) is the mark-to-market value of the contract at t
k
,
based on all market conditions realised by time t
k
in scenario S
j
.
Equation (1) models the fact that when the contract is out-of-the-
money (V(p; S
j
(t
k
), t
k
) 0), the contract is replaced at par and no
loss is incurred if the counterparty defaults. We refer to the matrix
of PFE values for every scenario S
j
and time t
k
as the counterparty
PFE profile.
More generally, PFE profiles can be expressed in terms of dis-
counted values to a given (previous) time:
(1)
where PV
t
() is the discount function to time t. This is required for a
credit valuation, or to compute expected losses and capital in a
multi-step setting.
Consider a portfolio P consisting of m positions, p
i
, i 1, . . ., m.
The market value of the portfolio at a given time and scenario can
be positive or negative and is given by
5
(2)
If there are no netting agreements in place, positive and negative
exposures cannot be netted out when the counterparty defaults
V P S t V p S t
j k i j k
i
m
( ; , ) ( ; , )
1
PFE p S t PV V p S t t
D t j k t j k k ( )
( ) ( , ( ( )]) ; , max[ ; ), 0
PFE p S t V p S t , t
j k j k k
( ; ) , ( ( ) )] ; max[ ; 0
8
and we stand to lose on all in-the-money transactions. The gross
credit exposure profile for P is
(3)
PFEs with netting and collateral
Credit mitigation techniques are commonly used to manage counter-
party risk for derivatives transactions and to reduce credit charges
and capital. The most common credit mitigation techniques are netting
and collateral. Under a master agreement (eg, International Swaps and
Derivatives Association, ISDA), the two parties are allowed to net a
set of positions (explicitly covered by the agreement) in the event of a
default by one of them. The enforceability of such agreements may
further depend on the laws in the jurisdictions of the defaulted party.
For modelling purposes, we introduce the concept of a netting
set a group of positions with a counterparty that are subject to a
legally enforceable netting arrangement. If P denotes a netting set,
containing a set of positions p
i
, i 1, . . ., m, the PFE (at each time
and scenario) is given by the maximum of zero and the mark-
to-market of the portfolio:
(4)
In some cases, netting agreements only cover certain types of transac-
tion (eg, fixed income), while other agreements allow cross-product
netting, where transactions of different product categories are
included within a netting set.
Counterparty credit exposures are further mitigated with the use
of collateral. When collateral with value C(P; S
j
, t
k
) is posted by the
counterparty against a netting set P, the counterparty exposure is
(5)
When netting is not allowed and collateral amounts C
i
(p
i
; S
j
, t
k
) are
placed against each position, the exposure to each position is
and the PFE profile is given by Equation (3).
PFE p S t V p S t C p S t
N
i j k i j k i i j k
( ; , ) , ( ; , ) ( ; , )] max[0
PFE P S t V P S t C P S t
N
j k j k j k
( ; , ) , ( ; , ) ( ; , )] max[0
PFE P S t V P S t V p S t
N
j k j k i j k
j
m
( ; , ) , ( ; , )] , ( ; , )
max[ max 0 0
1
]
]]
]
]
PFE P S t PFE p S t
G
j k i j k
i
m
( ; , ) ( ; , )
1

9
In general, a portfolio might have a combination of various net-
ting nodes (where netting is allowed within, but not across, the
nodes) and a series of positions where netting is not permitted. The
final PFE function for such portfolio can be obtained as a combina-
tion of Equations (3)(5). For example, the PFE of a portfolio made
up of two netting nodes P
1
and P
2
and a third sub-portfolio P
3
of
non-nettable positions is given by
A margin agreement is a contractual agreement, or provisions to an
agreement, under which the counterparty posts collateral when its
exposure exceeds a specified level. Bilateral agreements define
thresholds for both counterparties, where either one may post col-
lateral at a point in time, depending on whose transactions are in-
the-money.
Several components of margin agreements are given below.
The unsecured or margin threshold is the largest amount of an
exposure that remains outstanding until the party has the right
to call for collateral.
The margin call frequency is the period defined for making collat-
eral calls (eg, daily, weekly, etc).
The margin period or cure period is the time period from the last
exchange of collateral until a defaulting counterparty is closed
out and the resulting market risk is re-hedged. It generally
covers the time it takes to recognise the default event, the decision
to act and physically close the positions and liquidate collateral.
Some agreements explicitly stipulate waiting periods and some
jurisdictions require statutory periods.
To keep a model simple, it might seem reasonable to assume that
the exposure of a collateralised portfolio cannot exceed a margin
threshold. In this case, for a one-sided collateral agreement and full
netting, the PFE is given by Equation (5) with
(6)
where MT
CP
denotes the margin threshold amount for the counter-
party.
6
C P S t V P S t MT
j k j k CP
( ; , ) , ( ; , ) ] max[0
PFE P S t PFE P S t PFE P S t PFE P S t
j k
N
j k
N
j k
G
j k
( ; , ) ( ; , ) ( ; , ) ( ; , )
1 2 3
10
However, in practice, collateral can be delivered with a lag and it
can take a non-trivial period of two weeks or more to close-out the
portfolio when collateral is not delivered. Hence, Equation (6) is
only a lower bound, and we must consider the possibility that the
exposure can increase significantly over the close-out period. The
modelling of lag behaviour is further discussed in the Modelling
collateral section.
PFE calculations are further complicated with two-way collat-
eral agreements. In this case, the risk of over-collateralisation when
the receiving party has the right to re-use the collateral presents a
direct credit exposure to the delivering party. Consider an agree-
ment between bank B and a counterparty CP. Assume the portfolio
is out-of-the-money to B (V(P; S
j
, t
k
) 0) and B posts collateral to
CP. At a later time, the value increases and B is allowed to recall the
excess collateral. If CP enters into default, B may not be able to
recover the amount of over-collateralisation, thus effectively creat-
ing a loss. This occurs, for example when the CP is allowed to
re-use, or re-hypothecate, the collateral. Thus, the exposure of B to
CP is given by Equation (5) with
(7)
where MT
CP
and MT
B
are the margin threshold amounts for CP and
B, respectively.
There are several additional elements that need to be considered
when calculating collateral balances.
Independent amounts additional collateral paid/received usu-
ally on a deal by deal basis that serve to decrease/increase the
amount of collateral held beyond the static threshold amounts.
The additional collateral is returned as the deal matures.
Upfront amounts additional deposits paid/received upon initial
dealings with a new counterparty. These amounts are not included
in the calculation of collateral balance but are used to offset expos-
ure and, hence, result in a form of over-collateralisation.
Minimum transfer amounts since there are administration costs
in transferring collateral, these amounts represent a minimum
threshold amount under which no collateral transfers are to occur.
C P S t V P S t MT V P S t MT
j k j k CP j k B
( ; , ) , ( ; , ) ] , ( ; , ) ] max[ min[ 0 0
11
Rounding conventions collateral may be rounded to a mutually
agreeable level of significance.
These additional elements represent deterministic collateral adjust-
ments, and are handled in a straightforward manner. The Modelling
collateral section further focuses on two complications of modelling
collateral adjusted exposures that are not deterministic: lags and
credit-state dependent parameters.
PFE statistics and risk measures
Once the PFE profile for a contract or for a portfolio (with or with-
out mitigation) is calculated, various statistical measures can be
defined. Table 1 presents a set of measures commonly used in prac-
tice. We assume that the probabilities of the scenarios,
are known. The table gives an idea of the rich information pro-
duced from PFE profiles. In essence, quantile PFE measures are
used for risk management and credit limits. Expected exposures
are required for credit pricing (eg, CVA) and capital (eg, effective
EPE and effective maturity) and are part of the internal ratings
based models in the new Basel II Accord (BCBS 2005).
Figure 1 presents an example of a PFE profile and statistics
obtained from a MC simulation. The top left graph depicts the evo-
lution over time of portfolio mark-to-market values over a series of
MC scenarios. The upper right graph shows the corresponding
exposures (which are all non-negative). Finally, the plot in the bot-
tom depicts some of the measures in Table 1.
MONTE CARLO METHODS FOR PFEs
MC simulation is the most general and reliable approach used to
capture the complex stochastic nature of PFEs and for the model-
ling of credit mitigation. The PFE simulation methodology is
depicted in Figure 2. We will briefly describe each step.
7
Scenario generation
The calculation of PFEs requires a multi-step MC simulation,
where scenarios describe the joint evolution paths of all market

i i
i
i q , , . . ., ;

1 1
12
Table 1 PFE profile statistics and risk measures
Measure Definition Formula
Peak exposure
8
High percentile of the
distribution of exposures
at a particular date
Maximum Maximum peak exposure
peak that occurs at a given date
exposure or any prior date
Expected Average of the distribution
exposure
9
of exposures at a
particular future date
(before the longest-
maturity transaction
portfolio)
Effective Maximum expected
expected exposure at a given
exposure date or any prior date
(also defined recursively
as the greater of the
expected exposure at
that date and the
effective expected
exposure at the
previous date)
EPE Weighted average
over time of expected
exposures (weights are
the proportion that
an individual expected
exposure represents of
the entire time interval)
Effective Weighted average over
EPE time of effective expected
exposure (weights are the
proportion that an
individual exposure
represents of the
entire time interval)
10
CVA Credit risk premium of
counterparty portfolio
11
Effective Ratio of (discounted)
maturity
12
EPE over the life of the
portfolio divided by
the (discounted) EPE
over one year
M
EPE t T
EPE t year
D t
D t
( )
( )
( )
( )
0
0
1
CVA Spread EPE T T t
D t

( )
( ) ( )
0
0
EPE t
t t
t t t
k
k
Eff
E
l l l
l
k
( ) ( ) ( )
1
1
1 o

EPE t
t t
t t t
k
k
E
l l l
l
k
( ) ( ) ( )
1
1
1 o

Eff
E
k
j k
E
j
Eff
E
k
E
k
t t
t t
( ) max ( )
max ( ) ( )
0
1
]
]
]
]
,

E
k i k i
i
t PFE S , t ( ) ( )
MPE t Q t
k
j k
E
j
( ) max ( )
0
]
]
Q t PFE t Q t
E
k k
E
k
( )|Pr{ ( , ) ( )} 1
13
(and perhaps credit) factors affecting exposures and collateral. The
scenario generation step consists of:
identifying the set of factors affecting each portfolio;
defining and calibrating the joint factor processes. Evolution
models must match todays conditions, be routed in historical
behaviour and be rich enough to account for future plausible
events;
identifying a discrete set of simulation times {t
0
, t
1
, . . ., t
N
T};
sampling, from the underlying joint process, a set of factor scen-
arios up to T{S
1
(T), S
2
(T), . . ., S
q
(T)} (typically q will be several
thousand).
Figure 1 PFE profiles and statistical measures (T = 2,200 days)
Value mark-to-future Exposure mark-to-future
01/02/1999
(0)
05/02/2002
(1,100)
09/02/2005
(2,200)
20
15
10
5
0
01/02/1999
(0)
05/02/2002
(1,100)
09/02/2005
(2,200)
M
i
l
l
i
o
n
s
20
10
0
10
20
5
M
i
l
l
i
o
n
s
01/02/1999 (0) 05/02/2002 (1,100) 09/02/2005 (2,200)
Effective expected exposure
Effective EPE (T)
EPE (T)
0
1
2
3
4
5
6
7
8
9
Expected exposure
Maximum peak exposure (T)
Peak exposure (95%)
Exposure measures
14
Consider, eg, a counterparty portfolio consisting of swaps, FX
forwards and options in several currencies. Each scenario is
the joint realisation, at various points in time, of interest rates in
each currency and FX rates. The underlying model from which
scenarios are drawn might be a joint diffusion model (perhaps
with mean reversion for some factors). The Risk factor pro-
cesses and scenarios section gives an example for interest rate
derivatives.
Instrument simulation (valuation)
As a second step, every instrument is valued over every scenario
and point in time. This essentially entails the computation of a
three-dimensional matrix, or cube, V(p
i
; S
j
, t
k
) (the value of every
instrument i, under scenario j and time k).
PFE (P, S, t) PFE (P, S, t)
PFE(P, S, t)
Counterparties
S
c
e
n
a
r
i
o
s
Instruments
T
i
m
e
1. Risk factor scenario
generation
2. Instrument
simulation
(valuation)
3. CP aggregation,
mitigation
(netting, collateral)
4. PFE statistics
V(p, S, t)
Figure 2 PFE MC simulation steps
15
This step requires the use of efficient valuation models and
algorithms. For a typical derivatives portfolio of a medium (large)
bank, a simulation might require the valuation of tens (or hun-
dreds) of thousands of instruments, over thousands of scenarios
and tens or perhaps hundreds of time steps. In general, front-office
pricing functions are too slow and do not cope well with time evo-
lution. The performance of the pricing functions must be optimised
for vanilla instruments and approximating models used for
complex derivatives.
In particular, structured products require either lattice-based or
MC methods for a single valuation. Thus, it is generally necessary to
resort to various approximations. One type of approximation recon-
structs the pricing function (parametrically or non-parametrically)
at each time. This may require limited simulation or the reuse of
intermediate information of the pricing model itself. Examples of
these tools include the use of multi-step versions of simulation
grids (Chishti 1999) and regression within MC pricing (Longstaff
and Schwartz 2001; Glasserman and Yu 2004).
PFE profiles
PFE profiles are then computed in a post-processing simulation
step by aggregating all the transactions with a counterparty and
accounting for netting and collateral. This amounts to applying
Equations (3)(7) or perhaps more sophisticated models of margin
agreements (see the Modelling collateral section) to the pre-
computed cube V(p
i
; S
j
, t
k
). Credit mitigation techniques have a
great impact on reducing exposures and capital and thus it is import-
ant to model them in detail. However, their intricate modelling
might also require reasonable simplifications.
The result of this step is the PFE profile for a counterparty port-
folio P, PFE(P; S
j
, t
k
).
PFE statistics
The final step consists of computing the PFE statistics to be used for
limits and risk management, as well as regulatory and economic
capital (see Table 1 and Figure 1 earlier). Statistics such as the mean,
variance or a quantile are computed using standard MC estimators
on the PFE simulation.
16
APPROXIMATION METHODS FOR PFEs
In practice, the calculation of PFEs requires balancing speed, accur-
acy and complexity. Therefore, various methodologies are used in
addition to MC techniques.
Add-on approaches are currently the most commonly used meth-
ods within limits systems. They are simple, fast and reasonably
accurate for deal-based exposures. The central idea is that, for a given
contract, the exposure is given by its current exposure (positive
mark-to-market) plus an add-on factor. The Basel I credit capital
regulation presents the most common example of add-ons (BCBS
1995) and many institutions have implemented more sophisticated
add-on factor tables. Their biggest shortcoming is that individual
exposures are calculated separately and the rules for portfolio aggre-
gation are not effective to capture netting, collateral and natural off-
sets (eg, long and short positions). Although more sophisticated
add-on methodologies have been developed (eg, Rowe 1995; Rowe
and Mulholland 1999), it is now widely recognised that simple add-
ons cannot accurately capture the stochastic nature of exposures and
the richness of netting and collateral behaviour.
Analytical approximations for PFEs address some deficiencies of
add-ons, at a much reduced cost compared to MC methods. To gain
analytical tractability, however, these methods make simplifying
assumptions regarding the number of risk factors, their distribu-
tions and the functional form (and distribution) of portfolio values.
This section introduces several analytical methods to estimate
mean and peak exposures. We formulate the problem at a given
time t
k
and assume that the value of the portfolio can be modelled
as a function of the level of the risk factors, V(P; S, t
k
) V(P; s(t
k
), t
k
)
(ie, it is state- but not path-dependent).
Mean exposures
Assume that the distribution of portfolio values at t
k
, is normal
with mean and variance
V
(t
k
),
V
(t
k
):
V(P; t
k
) N(
V
(P; t
k
),
V
(P; t
k
)) (8)
For a netting set, the mean exposure is given by
13
(9)

E
V V
V V V
P E V P x n x x
N n
V V
( ) { , ( )]} ( ) ( )
( ) (

max[ d
/
0

V
)
17
where n(x) denotes the standard normal density
Asimilar closed-form formula is obtained for the second moment:
(10)
For a portfolio consisting of m positions, where netting is not
applicable, the mean exposure is the sum of each contracts mean
exposures
(11)
Thus, we require two parameters the mean and variance of port-
folio values (
V
,
V
). It is common to approximate the mean port-
folio value by its future value at t
k
under a base scenario of mean
risk-factor values S
,
V
(P) V(P; S
, t
k
).
14
The estimation of the
standard deviation is more involved and is discussed below.
Peak exposures
It is useful to formulate the problem in terms of two general con-
cepts: the risk-factor probability density level sets and the portfolio value
(or exposure) manifolds at t
k
. Figure 3 depicts these concepts for a
two-dimensional risk-factor space. The dots in the diagram depict
joint movements in both risk factors, assuming they are jointly
Gaussian. The circles emanating out from the centre represent dif-
ferent density level curves that encapsulate a certain percentile of fac-
tor movements.
The portfolio value (or exposure) manifold is the surface created by
plotting portfolio value (or exposure) as a function of the risk factors
(right-hand side of Figure 3). The manifold is depicted by iso-exposure
lines, which capture the set of joint risk-factor movements that pro-
duce a given value (or exposure). In Figure 3, it runs from its lowest
values in the top right corner to its highest values in the bottom left.

E E
i
i
m
G
P p ( ) ( )
1
E PFE P N n
V V V
V
V
V
{ ( ) } ( ) ( ) ( )
2 2 2

|
(
'
`
J
J
n x
e
x
( )
2
2
2
/

18
The two most common analytical methods for approximating
peak exposures are polynomial approximations and reliability methods.
Both have been applied broadly to short-horizon market value-at-
risk (VAR) (eg, Pritsker 1997). For PFEs, they must be applied con-
secutively at each time, and require some modifications.
Polynomial approximations
Consider a netting node P with m positions (p
j
, j 1, . . ., m), where
the portfolio value depends on a vector of market factors X (x
1
,
x
2
, . . ., x
n
)
T
.
15
A polynomial approximation of the value manifold
can be written as
(12)
where V
0
V(P; X
0
) denotes the portfolio value under a base scen-
ario and x
j
is the change in factor j from the base.
Alinear (delta) approach approximates portfolio values by a linear
manifold (Figure 4) by considering only the first-order terms in
Equation (12). This reduces the problem to a multi-step RiskMetrics
style calculation.
16
In this case,
j
(P) is the sensitivity of the portfolio to
each factor and are obtained by summing the deltas for each position
(13)

j j i
i
m
P p ( ) ( )
1
V P X V P x P x x
j j
j
n
jl j l
l
n
j
n
( ; ) ( ) ( )
0
1 1 1
1
2

Figure 3 Risk-factor density level sets and portfolio value manifolds
4
4
3
2
1
0
R
i
s
k
-
f
a
c
t
o
r

2
1
2
3
4
3 2 1 0
Risk-factor 1
99%
95%
66%
1 2 3 4 4
4
3
2
1
0
1
2
3
4
3 2 1 0
Risk-factor 1
166
137
106
195
1 2 3 4
19
(second-order portfolio sensitivities
jl
(P) are similarly obtained
from position sensitivities).
If denotes the covariance of factor changes, then the portfolios
value variance is given by
V
(P)
2
(P)
T
(P)(P) (14)
and the peak exposure for the netting set is
(15)
where Z
is the -quantile of the standard normal distribution.

The linear approximation for peak exposures is ineffective for
option portfolios and, more importantly, for portfolios that do not
net where position exposures (instead of values) are needed to
arrive at the portfolios exposure. However, it is not a limitation for
mean exposures (see Equation (11)).
Quadratic approximations consider the first- and second-order terms
in Equation (12) (see Figure 5). For market VAR, localised sensitivities
Q P V Z P
E
V
( ) [ , ( )] max 0
0

Figure 4 Linear manifold approximation
106
137
166
195
4 3 2 1 0
Risk-factor 1
R
i
s
k
-
f
a
c
t
o
r

2
1 2 3 4
4
3
2
1
0
1
2
3
4
20
(deltas and gammas) may be sufficient to model the portfolio accu-
rately. However, they may not be robust for PFEs at long time horizons
where the risk-factor ranges are much larger. Studer (1999) proposes to
fit a quadratic function to each instruments profile by simulating the
value of the function at the current and extreme risk-factor levels.
Fitted quadratic functions capture each positions behaviour at
extreme factor movements and also address portfolios that are not
allowed to net. Since exposure is a non-linear function of value
(similar to a call option payoff), a quadratic form can be directly fit-
ted to each positions exposure (instead of value in Equation (12))
and the coefficients added-up to obtain a quadratic portfolio expos-
ure manifold.
As with the linear approximation, the quadratic peak exposure
calculation is a multi-step quadratic VAR problem. In this case, a
simple formula such as Equation (15) is not available, but several
numerical methods can be used to estimate quantiles, such as
Johnson transformations, CornishFisher expansions and Moment
Figure 5 Quadratic manifold approximation
4 3 2 1 0
Risk-factor 1
R
i
s
k
-
f
a
c
t
o
r

2
1 2 3 4
4
3
2
1
0
1
2
3
4
106
137
166
195
21
Generating functions with Fast Fourier transforms (eg, Mina and
Ulmer 1999; Brummelhuis et al 2002; Duffie and Pan 1997).
Reliability theory approximations
Reliability theory methods seek to speed up the computation by
targeting only particular areas of interest in the risk-factor space
(see De and Tamarachenko 2002; Cardenas et al 2002; De Prisco et al
2004). The main idea is that if the portfolio exposure manifold is
monotonic in each risk factor, only the density curve corresponding
to the desired quantile is relevant for simulation (see Figure 6). For
each time step, the method works as follows:
derive the probability density surface in n-dimensional risk-factor
space for the quantile of interest (the circle encapsulating joint
risk-factor movements with probability in Figure 6);
estimate the portfolios manifold along the density surface (the iso-
exposure lines that cross the density surface at various points in
Figure 6);
Figure 6 Reliability theory approximation
4 3 2 1 0
Risk-factor 1
R
i
s
k
-
f
a
c
t
o
r

2
1 2 3 4
4
3
2
1
0
1
2
3
4
1
()
124 137 145 161
22
run an optimisation along the density surface to capture the
largest exposure, using the portfolios manifold (the star in
Figure 6).
17
The maximum returned from the optimisation is the peak exposure
at that time step.
Reliability methods can be used in combination with analytical
methods for simulation of the portfolios manifold along the dens-
ity surface. One of the main shortcomings of the approach is that it
assumes that knowing the risk-factor quantile is enough to know
the portfolios exposure quantile.
18
Thus, the method works well for
portfolios with clear directional bets, but is ineffective for hedged
portfolios.
Remarks on analytical approximations
The principal benefit of analytical approaches is their speed
they run many orders of magnitude faster than MC techniques.
However, it is important to understand several shortcomings.
They are generally applied at each time step independently
of adjoining steps. Therefore, they may fail to capture path
dependencies (early exercises, physical settlements, collateral
behaviour).
They often require simplifying assumptions regarding the num-
ber of risk factors, their distributions and the portfolio mani-
folds, which may not be appropriate.
Approximations are sometimes specifically fitted to particu-
lar PFE measures and may not be well suited for obtaining full
distributions. For example, quadratic functions fitted to extreme
risk-factor points may ensure reasonable quantile results, but
the mean and standard deviation might be unreliable.
Reliability approaches often require full re-computation for
each quantile.
Some numerical methods may not always be applicable (eg,
CornishFisher expansions are known to generate non-monotonic
behaviour with extreme quantiles when the distribution is not
normal) (Jaschke 2002).
Analytical methods generally lose efficiency when extensive
roll-ups of exposures to higher portfolio levels are required (eg,
from a netting agreement to a counterparty to a country level).
23
Example: analytical approximations
We consider a stylised non-linear portfolio, with the objective to
provide some intuition on the behaviour of portfolio manifolds
and their approximations. In practice, counterparty portfolios may
have hundreds or thousands of positions (linear and non-linear,
long and short) across tens of risk factors. Some portfolios may pre-
sent simpler manifolds and others more complex.
Consider a non-linear portfolio with two equity barrier options
(an up-and-out call and a down-and-out put) on two stocks, expir-
ing in one year.
19
The portfolio is a function of the two underlying
stock values, modelled as Gaussian risk factors.
Figure 7 illustrates the portfolios manifold and iso-exposure
lines at two weeks and six months, and the associated 90%, 99%
Figure 7 Value manifold, iso-exposures and density level curves (90%, 99% and
99.9%) at two weeks (top) and 6 months (bottom)
3 2 1 0
0.75
0
.
5
0.5 0.5
0.5
0
.
5
0.5
0
.
5
0
.
5
0
.5
0
.
7
5
0
.
7
5
0
.
7
5
0
.
7
5
1
.
0
2
7
6
1
.0
2
7
6
1
.
0
2
7
6
1
.
5
0
.
7
5
0
.
7
5
0
.7
5
0
.5
0
.5
0
.
7
5
0
.8
4
8
4
9
Gaussian risk-factor 1
Level curve for the portfolio value 14 days horizon
G
a
u
s
s
i
a
n

r
i
s
k
-
f
a
c
t
o
r

2
1 2 3
3
2
1
0
1
2
3
3
4
4
2
2
0
Surface of the portfolio value in the
Gaussian space 182 days horizon
0
G
a
u
s
s
ia
n
r
is
k
-
fa
c
to
r
2
G
a
u
ssia
n
risk
-fa
c
to
r 1
2
2 4
4
0
0.2
0.4
0.6
0.8
0
V
a
l
u
e

o
f

p
o
r
t
f
o
l
i
o
1.2
1.4
1.6
2 1 0
Gaussian risk-factor 1
Level curve for the portfolio value 182 days horizon
G
a
u
s
s
i
a
n

r
i
s
k
-
f
a
c
t
o
r

2
1 2 3
3
2
1
0
1
2
3
3
3
2
2
0
Surface of the portfolio value in the
Gaussian space 14 days horizon
0
G
a
u
s
s
ia
n
r
is
k
-
fa
c
to
r
2
G
a
u
ssia
n
risk
-fa
c
to
r 1
1
2
3
0
0.3
0.4
0.5
0.6
0.7
V
a
l
u
e

o
f

p
o
r
t
f
o
l
i
o
0.8
0.9
1.0
1
1
2
3
1
24
and 99.9% risk-factor density-level curves. At two weeks, the mani-
fold is concave and monotonic within the confidence bounds. Both
quadratic functions and reliability methods work well. In contrast,
at six months it is not well approximated by a quadratic and peaks
towards the centre of the risk-factor space.
Figure 8 shows the 99% peak exposures from various methods.
The differences increase over time as the manifold becomes more
complex. At six months, analytical methods underestimate the
peak exposure by 50% and at one year by almost 270%.
ADVANCED MODELLING
This section presents advanced PFE modelling examples in three
main areas:
the impact on PFEs of the underlying scenario models and simu-
lation time steps;
the modelling of collateral agreements with lags and credit-state
dependent thresholds;
the integrated treatment of market and credit risk to model
portfolios that include credit derivatives and wrong-way
exposures.
Figure 8 Options portfolio: analytical and MC methods
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
2 7 14 30 91 182 273 300 330
Time horizon (days)

9
9
%

E
x
p
o
s
u
r
e
Reliability method
Quadratic method
MC
25
Risk factor processes and scenarios
Scenario generation for PFEs is inherently different from no-
arbitrage pricing. Pricing is generally done under the so-called risk-
neutral measure and risk-factor distributions are calculated from
market implied values (eg, implied volatilities). In contrast, PFE
requires the actual, or real, distributions, which are, in general, esti-
mated from historical price changes.
The market risk factors that affect PFEs of derivatives portfolios
include interest rate (IR) curves, equity indices, exchange rates,
commodity prices and implied volatilities. The underlying processes
describing their evolution through time constitute a significant
driving factor of PFEs. Ascenario generation model must generally
satisfy three principles.
The model should match todays prices, yield curves and for-
ward rates precisely.
It should be routed in historical behaviour, but rich enough to
account for future events, which may be plausible under current
circumstances. For example, an IR model should allow for paral-
lel shifts, steepening and curve inversions.
It must be robust and stable over long simulation horizons.
As an example, consider the evolution term structure curves,
which drive IR derivatives exposures. The simplest model to cap-
ture the mean reversion characteristics typically observed is the
one-factor extended Vasicek (Hull and White 1990) model
dr [(t) ar]dt dz
where the short rate, r, reverts to a long run level (t)/a with speed
a. This model can be fitted to todays term structure and is easy to
calibrate. At any point in time, the full term structure is constructed
from the value of the short rate.
In the single factor model, the entire term structure increases and
decreases simultaneously with the short rate.
20
This directly impacts
the quality of the scenarios. As an example, Figure 9 shows the term
structure of the EUR Interbank curve as of the 10th August 2001, as
well as a set of MC scenarios one year in the future.
The term structure moves almost in parallel with the MC scen-
arios.
21
Such behaviour is in direct contrast to the term structure
26
movements actually seen in the market. Figure 10 plots the day-to-
day EUR Interbank curve changes over three-years. The curve is
subject to considerably twisting as the nodes on the short-end
move with much greater volatility than the nodes on the long end
and their correlation is not one.
On the other extreme, curve evolution models typically used in
market risk take each node in the curve as a random process. Such
models are generally not appropriate for PFEs since they generate
very unlikely future scenarios. Amore robust model for PFE simula-
tion is a multi-factor statistical term structure model, based on two to
five (typically three) underlying principal factors describing histor-
ical curve movements (cf, Rebonato 1998; Reimers and Zerbs 1999).
Figure 9 EUR Interbank term structure and scenarios one year in the future with
one-factor HullWhite model
0.060
0.055
0.050
0.045
0.040
0 1,333 2,667 4,000
Term in days
5,333 6,667 8,000
0.080
0.050
0.060
0.070
0.030
0.040
0.020
0.010
0 1,333 2,667 4,000
Term in days
5,333 6,667 8,000
27
The scenario model can have a significant impact on PFEs. As an
example, we compare the PFEs of the two instruments in Table 2,
using a one-factor and a five-factor model, calibrated against the
same historical information.
22
Figures 11 and 12 show the mean
and 99% peak exposures generated with 10,000 MC scenarios.
The cap/floor peak exposures present over 100% difference. While
the discrepancy looks smaller for mean exposures, percentage-wise it
is still substantial (reaches over 40%). In contrast, swaption peak
exposure differences reach only 25% and the mean PFEs are almost
Figure 10 Historical daily changes in EUR Interbank term structure
0
60
40
20
0
20
B
a
s
i
s

p
o
i
n
t

c
h
a
n
g
e
s
40
60
80
500 1,000
Term in days
Daily changes of EURO curve from 1998/01/22 to 2001/08/10
1,500 2,000 2,500
Table 2 Instruments for PFE calculation
Cap/floor Swaption
5-year cap 4-year swaption
Monthly caplets 5.25% strike (ATM)
6.5% strike (30% OTM) 5-year underlying swap
Pay fixed; receive floating
28
the same (5% differences). This result can be explained as follows:
the cap/floor experiences a great deal of sensitivity to term struc-
ture twists due to monthly caplets, while the swaption shows less
sensitivity to twists, since the forward swap yield calculation
averages out the shocks to the term structure nodes.
Figure 11 Cap; PFEs from one-factor and five-factor models
0
0
2
4
6
8
10
12
14
T
h
o
u
s
a
n
d
s
16
18
20
183
One-factor model
Five-factor model
PFE for cap (mean & 95%)
366 549 732 915 1098 1281 1464 1647 1830 2013 2196
Figure 12 Swaption; PFEs from one-factor and five-factor models
M
i
l
l
i
o
n
s
PFE for swaption (99%)
Five-factor model
One-factor model
0
1
2
3
4
5
6
7
0 183 366 549 732 915 1098 1281 1464 1647 1830 2013 2196
29
Roll-off risk
Although, in principle, PFEs can be calculated using daily time
steps until the maturity of a portfolio, in practice, a smaller set of
time points are used, which are deemed to represent the profile
well. Managing exposures over the smaller set of time steps
assumes, either explicitly or implicitly, that PFEs can be interpol-
ated between these points. However, exposures are largely discon-
tinuous across time due to contract expiries, options exercises,
coupons, payments and amortisation.
The risk of not detecting these jumps and, more generally, of fail-
ing to capture future exposures that are large in magnitude but
short in duration is referred to as roll-off risk. Figure 13 shows, as an
example, the 95% peak exposure profile for a portfolio with daily
time steps (black) and monthly time steps (grey). The sparse simu-
lation clearly misses many peaks in the profile.
Solutions to capture roll-off risk include a combination of the fol-
lowing techniques.
Fixed non-homogeneous time grids define a fixed set of steps that
are smaller in the short term and larger towards the end of the
profile for example, daily time steps for a week, weekly steps
until three months, monthly up to one year, quarterly to year
Figure 13 Roll-off risk in 95% peak exposure profile
0
100
200
300
400
500
600
700
800
900
1
4
0
7
9
1
1
8
1
5
7
1
9
6
2
3
5
2
7
4
3
1
3
3
5
2
3
9
1
4
3
0
4
6
9
5
0
8
5
4
7
5
8
6
6
2
5
6
6
4
7
0
3
M
i
l
l
i
o
n
s
Credit exposure roll-off
30
three and then yearly until maturity (eg, a 15 year simulation
results in 46 steps). While they do not present a full picture of
exposures, these grids capture roll-off risks in smaller windows
as they get closer and, hence, provide enough time for manage-
ment to act as jumps approach. On the downside, PFEs may
change dramatically from day to day simply because the grid
captures jumps that did not appear before.
Endogenous time steps. Exposure peaks often occur at cashflow
and settlement dates. Therefore, they can be reasonably captured
by simulating on all cashflow and settlement dates and the dates
prior to them (or adding these to a grid).
Scenario banding and limited simulation. When portfolios have a
large number of cashflow dates, one must further discriminate
a priori (or dynamically at simulation) which of these steps may
have significant jumps. This can be achieved through some sim-
ple heuristics or, more generally, through the use of analytics
(see the Approximation Methods for PFEs section) or limited
simulation. An effective simulation methodology is scenario
banding. It identifies, in a multi-factor setting, a reduced number
of banded or extreme scenarios, which can be used to identify
jumps (Cartolano and Verma 2000).
Modelling collateral
The second section presents the basic framework for modelling
margin agreements and discusses some details of these contracts.
We now focus on two key areas that are not handled deterministi-
cally: margin lags and credit-state dependent parameters.
Collateral margins lag behaviour
In the case of instantaneous payment/receipt of collateral (when
the margin period is zero), PFEs are modelled through Equations
(5) and (6). Substituting Equation (6) into Equation (5) results in:
(16)
Therefore, if V(P; S
j
, t
k
) MT, then PFE
N
(P; S
j
, t
k
) is capped at MT
CP
.
PFE P S t V P S t
V P S t MT
N
j k j k
j k CP
( ; , ) , ( ; , )
, ( ; , ) ]}

max{
max[
0
0
31
Figure 14 shows a daily simulation 50 days into the future,
under a given scenario, of a portfolio with a CSAagreement speci-
fying a US$35 million threshold. The left-hand side presents the
results when collateral is posted instantaneously. The top line
shows the portfolio value evolution, the middle line the evolution
of the collateral balance and the bottom line the daily margin calls
paid/received. Daily changes in valuation bring about corre-
sponding changes in margin calls that affect changes in the collat-
eral held. The end result is a flat PFE at the threshold amount.
When the margin period is taken into account, Equation (5) can
be restated as
(17)
Credit exposure is a function of the portfolio value at t
k
and a
lagged collateral balance at an earlier time t
kMP
. The length of the
margin period is (t
k
t
kMP
). The lag period allows the PFE to rise
above its threshold. This is shown in the right-hand side graph in
Figure 14, where a lag of three days and a user lag of zero days are
assumed (receiving collateral requires three days while returning
collateral is instantaneous). The PFE exceeds the threshold as the
portfolios value drifts upwards while waiting for collateral to be
PFE P S t V P S t C P S t
N
j k j k j kMP
( ; , ) , ( ; , ) ( ; , )] max[0
Value
Credit exposure
Exposure, collateral balance & margin call Credit exposure with lag
Collateral balance
Margin calls
M
i
l
l
i
o
n
s
0
10
20
30
40
50
60
10
1999/02/01 (0) 1999/02/26 (25) 1999/03/23 (50)
Credit exposure
assuming
instantaneous receipt
of collateral
Credit exposure
assuming three-day lag
in receipt of
collateral
Figure 14 PFE of portfolio with margin agreement zero margin period (left) and
three-day margin period (right)
M
i
l
l
i
o
n
s
22
20
1999/02/01 (0) 1999/02/26 (25) 1999/03/23 (50)
24
26
28
30
32
34
36
38
40
42
44
32
received but does not fall below the threshold since excess collat-
eral is returned immediately.
23
There are three common approaches to model this lag behav-
iour: full simulation, analytical add-ons and post-simulation sampling.
Full simulation
This approach involves augmenting the original steps t
k
in the sim-
ulation to also include the t
kMP
dates. As the most straightforward
modelling approach, it captures full valuation, roll-offs and path
dependency. However, it is also most onerous from a calculation
perspective.
Analytical lag add-ons
Equation (17) can be restated as
(18)
with V(P; S
j
, t
k
) V(P; S
j
, t
k
) V(P; S
j
, t
kMP
). Alag-adjusted expos-
ure is similar to an exposure at t
kMP
and adjusted for a change in
valuation from t
kMP
to t
k
.
Assume that V(P; S
j
, t
k
) and [V(P; S
j
, t
kMP
) C(P; S
j
, t
kMP
)] are
normally distributed with respective means and standard devia-
tions of , and
V
,
V
, and correlation . Then,
(19)
with and .
The mean and peak exposures can be expressed as
24
(20)
(21)
where Z
is the -quantile of the standard normal distribution.

While
V
and
V
can be calculated directly from a simulation at
t
kMP
, , and are not directly observable without simulating at
t
k
as well. However, in most cases the margin period V(P; S
i
, t
k
) is
*
V
*
V
Q PFE P t Z
p
N
k V V
{ ( ; )}

E PFE P t N
N
k
V V
V
V
V
V
{ ( ; )}

2 2
2
2
exp
|
(
'
`
J
J

|
(
'
`
J
J

V V V V V

* *
2
V V V
*
[ ( ; , ) ( ; , ) ( ; , )] ~ ( , ) V P S t V P S t C P S t N
i k i kMP i kMP V V

*
V
*
V
PFE P S t V P S t V P S t C P S t
N
j k j k j kMP j kMP
( ; , ) , ( ; , ) ( ; , ) ( ; , )] max[0
33
short (one to two weeks), allowing us to make the following rea-
sonable assumptions:
0 (no drift in V(P; S
i
, t
k
));
is estimated from the distribution of V(P; S
i
, t
kMP
) and scaled
appropriately by
1 (a conservative assumption providing an upper bound).
This adjustment might not be accurate for longer margin periods
or where the period overlaps with a roll-off event. Furthermore,
since the lag-adjustment calculation uses general statistics (
V
,
, the scenario consistency is lost, resulting in some
possible inconsistencies on roll-ups to higher levels.
Post-simulation sampling
The logic above can be used with a slight twist to get the benefit of
scenario consistency. The idea is to perform a full simulation on the
t
kMP
dates and store the results of the simulation under each time
step and scenario. In a post-processing stage, we cycle over all scen-
arios to calculate as above for each time step. Then, we cycle
back to the original scenario results stored and add a randomly
sampled term from a distribution under each scenario in
order to represent V(P; S
i
, t
k
). Finally, exposure measures (mean,
quantiles) are calculated in the normal way.
Credit state dependent collateral
Some CSA agreements define thresholds, upfront amounts and
minimum transfer amounts as a function of both the counterpartys
and the Banks credit states. Table 3 gives an example of thresholds
in a margin agreement that depend on the credit ratings of both
parties. Should the counterparty continue with its AA status, the
bank extends US$40 million of uncollateralised credit (ie, collateral
is only called for beyond an exposure of US$40 million). If the
credit rating drops to B, it is now only comfortable enduring
US$5 million of exposure and will ask for US$35 million of extra
N
V
( ,
*
) 0
*
V
*
, ,
*
, )
V V V
t t
t t
k kMP
kMP
0
*
V
*
V
34
collateral. The same logic works for the banks own thresholds.
When the counterparty is not rated, thresholds may be defined as a
function of some financial ratios (eg, debt to cash-flow), which are
essentially used as proxies for its credit state (or financial health).
When thresholds are credit-state dependent, the collateral balance
must now be expressed as a function of the bank and counterparty
credit states, CS
CP
and CS
B
. Thus Equation (16) can be restated as
(22)
The computation of collateralised PFEs with Equation (22) requires
an integrated market and credit-risk model to capture the expos-
ures that are market dependent, as well as the collateral balances,
which depend on both the market levels and the credit states of
both parties. Thus, in addition to the market information, we
require the joint credit-state probabilities of both parties at each
time that collateral is called, t
kMP
.
Consider a one-sided collateral threshold with a rated counter-
party and assume independence of the underlying market and
credit processes. We can model the counterpartys credit-state evo-
lution assuming a Markov process with a credit migration matrix
such as those given by the credit rating agencies.
25
The relevant
probabilities for PFE calculations are
(23)
Pr{ for CS t y CS t Def CS t Def y Def
CP kMP CP kMP B kMP
( ) ( ) , ( ) },
CE P S t CS CS V P S t V P S t
C P S
N
i k B CP i k i kMP
( ; , , , ) , ( ; , ) ( ; , )
( ;

max[0
ii kMP B CP
t CS CS , , , )]
Table 3 Margin agreement with credit-dependent thresholds
Credit state Bank threshold Counterparty threshold
(in US$ million) (in US$ million)
AAA 50 50
AA 40 40
A 35 35
BBB 25 25
BB 10 10
B 5 5
C 1 1
35
It is important to highlight several characteristics about these
probabilities.
First, they are real-world probabilities (taken from historical
transition matrices) not risk-neutral probabilities.
Second, they are cumulative probabilities; Equation (23) requires
the probability that the counterparty is in a given credit state at a
point in time.
Third, while it is common to have one-year credit transition
matrices (eg, from rating agencies), the time steps in the PFE pro-
file can be much shorter and at arbitrary intervals. In practice,
one-year transition matrices must be scaled for different time
steps (eg, Kreinin and Sidelnikova 2001).
Finally, probabilities at t
kMP
are conditional on no previous
defaults (from either the counterparty or the bank). This follows
from the original definition of exposure at time t
k
as the amount
the bank would lose conditional on the counterpartys default at
that point in time (and the last time collateral was called was
t
kMP
). The cumulative (unconditional) credit migration probabili-
ties must be scaled to make them conditional on survival of both
firms to time t
kMP
.
26
Once these credit-state probabilities are obtained, the PFE profile at
a given time step is computed as follows.
for each scenario:
compute the uncollateralised exposures, using Equations (2)(4);
calculate (n
CP
1) collateral balances C(P; S
i
, t
kMP
, CS
B
, CS
CP
),
(one for each possible non-default counterparty rating);
calculate (n
CP
1) collateralised exposures from the un-
collateralised exposure and each collateral balance;
weight collateralised exposures by the conditional credit-state
probabilities.
PFE statistics are computed from the (n
CP
1)n
s
augmented col-
lateralised exposure scenarios.
When the CSA covers bilateral margin thresholds, we also
require the credit migration probabilities of the bank, as well as the
co-dependence between the bank and the counterparty credit
migrations. Afew points are worth highlighting.
36
The joint credit evolution of both parties can be modelled
through a multi-step version of a Merton-type or CreditMetrics
model
27
or using an intensity based (or reduced-form) model.
28
At a given scenario and time, only the credit state of one party is
required. If the exposure is positive (to the bank), only the credit
state of the counterparty is relevant at the last margin call.
Similarly, if the exposure is negative and the bank has posted
collateral, only the credit state of the bank is required to deter-
mine how much collateral has been posted.
Since probabilities are conditional on previous survival by both
parties, a conservative assumption that simplifies the model is
that both parties are uncorrelated.
The solutions outlined above assume independence between the
market and credit events. Exposures that are correlated to credit
events are generally referred to as wrong-way exposures and are
briefly discussed in the Wrong-way exposures section.
PFEs for portfolios with credit derivatives
The presence of credit derivatives complicates PFE modelling in
two ways. First, credit derivatives potentially impact the exposure
profiles of two or more entities simultaneously. For example, a
single-name credit default swap (CDS) affects the profiles of both the
counterparty to the CDS and the underlying issuer of the reference
asset. Second, counterparty PFEs now depend not only on the level
of the market rates, but also on the credit states of the underlying
issuers.
Exposure to the counterparty
In the previous PFE formulae, valuation was independent of the
credit quality of the trading counterparty. In contrast, the valuation
of credit derivatives directly involves the credit state of the under-
lying issuers. A firm is expected to pay more for protection when
the issuers credit rating is low than when it is high. This means
that the valuation for a credit derivative depends also on the
issuers credit state, V(P; S
i
, t
kMP
, CS
Iss
), where we have assumed
that the scenario S contains all the market information (but not the
actual credit events). Similarly, the PFE of a portfolio containing a
set of issuers depends on their joint credit states.
37
What remains to be calculated, then, are the values of the credit
derivatives under various states of issuer credit ratings (condi-
tional upon the counterparty defaulting at that point in time). Hille
et al (2005) describe an approach that is somewhat similar to the
method for capturing lag behaviour in the Modelling collateral
section. In summary, it involves the following steps (for each mar-
ket scenario and time step):
calculate spread curves and risk-neutral default probabilities for
each possible credit state of the underlying issuer;
calculate a set of CDS contract values using the risk-neutral
probabilities;
calculate the probabilities of the issuer in each credit state (con-
ditional upon the counterparty defaulting at that point in time);
augment the original market scenarios with the new distribution
of CDS values and probabilities;
calculate all exposure measures using the augmented distribution.
This approach works well when the CDS contracts with a given
counterparty have the same underlying issuer and, thus, the aug-
mented distribution covers only one set of issuer credit state prob-
abilities. When the CDS contracts with a given counterparty span
many underlying issues, the solution might require some form of
MC simulation on the joint issuers credit states.
Exposure to issuer
A CDS can also impact the exposure to its underlying issuer. For
example, consider a portfolio of loans, bonds and derivatives with
Ford Motor Co. In some way, buying credit protection for Ford in
the form of a CDS reduces our exposure to Ford. One simple way to
capture this impact within a limits system is to create an offsetting
instrument for the CDS. The CDS is allocated to the netting struc-
ture of the counterparty (as explained above), while the offsetting
instrument can be allocated to the netting structure of the issuer.
It is important to highlight that using CDS contracts to offset
exposures is not fully clean, since the CDS might also incorporate
recovery effects. For example, when protection is purchased, if the
reference asset is actually held, the CDS can fully offset its expo-
sure. Essentially, if the issuer defaults, the counterparty can deliver
38
the asset and not incur a loss. In contrast, when the bank does not
hold the exact reference asset but one with the same issuer and dif-
ferent seniority, then the net exposure to the issuer is not com-
pletely eliminated. We can calculate this as
(24)
where rr
CDS
is the recovery rate for the reference asset and rr
B
, the
recovery rate of the asset held. Thus, the full exposure amount is
offset if the CDS is contracted on a reference asset that is of equiva-
lent or more junior quality to the one held. Otherwise, we have not
purchased complete protection and only reduce the exposure to the
amount in Equation (24).
Finally, note that the use of credit derivatives to mitigate expos-
ures of issuers is similar to having an offsetting position within a
netting agreement. It effectively cancels a positive exposure to the
counterparty.
Wrong-way exposures
A wrong-way exposure (WWE) occurs when counterparty expos-
ures increase simultaneously with a weakening of the counter-
partys credit quality. For instance, the bank may negotiate a cross
currency swap with a Russian counterparty that requires the bank
to pay Russian rubles and receive US dollars. In the event that the
ruble depreciates, not only does this increase the value of the trans-
action but it also impacts the counterpartys ability to meet its
financial obligations (a lower ruble means less available financial
resources). The market scenarios leading to an increased exposure
to the counterparty are precisely those impairing its overall credit-
worthiness.
Exposures are often assumed to be independent of counterparty
defaults and stress tests or ad hoc analyses are used to handle
WWEs (eg, Rowe 1999). While such methods produce useful early
warning signals, we also require quantitative measures that can be
used for pricing, capital and limits.
Recall that counterparty exposure is the economic loss incurred
on all outstanding transactions if the counterparty defaults. This
NetExposure
Notional rr rr
rr
CDS B
B
max 0
1
,
( )
( )

]
]
]
39
implies that various PFE statistics require the underlying distribu-
tion of counterparty exposures conditional upon the counterparty
defaulting.
Consider a single time step. If we denote by Pr(PFECS
CP
Def )
the distribution of counterparty exposures conditional upon a coun-
terparty default, the mean and peak exposures are given by
29
(25)
When exposures are independent of the counterparty credit behav-
iour, conditioning has no effect. The market factor distributions can
be used directly to generate the set of (unconditional) scenarios and
probabilities S
j
,
j
, j 1, . . ., q;
j

j
1, which in turn are used
to compute the PFE distribution. In the case of WWEs, the co-
dependence between PFEs (or market scenarios) and the counter-
party health can have a significant impact on both mean and peak
exposures. To model this, an integrated market and credit frame-
work is required that defines the joint distributions of market factors
and credit events. The joint evolution of market and credit events
is modelled naturally through a structural (Merton-type) model.
30
Alternatively, reduced form models can also be used.
31
We can think of computing conditional PFE distribution in
two ways. Adirect approach generates conditional market scenarios
(with their probabilities) from the market distributions conditional
on the counterparty default, Pr(S x CS
CP
Def ). One simple
way of looking at this is to generate a large number of joint scenar-
ios on the market factors and credit events from the model and
keep those scenarios where the counterparty actually defaults (nor-
malising their probabilities). The resulting (conditional) distribution
of market scenarios is used to compute PFE statistics.
An indirect approach uses Bayes Theorem:
32
(26)
Pr
Pr
Pr
Pr
( )
( , )
( )
(
S x CS Def
S x CS Def
CS Def
CS Def
CP
CP
CP
CP

SS x S x
P CS Def
CP

) ( )
( )
Pr
Q t PFE S Q CS Def
E
k
E
CP
( ) ( ) }
*
such that Pr{ 1
E
CP
E PFE S CS Def
*
[ ( ) ]
40
Equation (26) indicates that we can generate market scenarios
directly from the unconditional distribution and scale their (uncon-
ditional) probabilities, Pr(S x) (ie, the scenario weights
j
) by the
ratio of the conditional default probabilities Pr(CS
CP
Def S x) and
the unconditional default probabilities, Pr(CS
CP
Def ). Thus, market
scenarios producing large conditional default probabilities, receive
large weights.
Both, reduced form and structural type credit models can be
used with this formulation since they readily produce conditional
default probabilities. For example, within a one-step Merton-type
model, the conditional default probabilities are given by
33
(27)
where N and N
1
denote the normal distribution and its inverse.
We have assumed that the scenario S covers K credit drivers x
k
and
the
k
are the factor loadings for the counterpartys creditworthi-
ness. In this model, credit events are correlated to the market in
three steps:
a set of scenarios are constructed with joint market factors and
credit drivers;
the counterpartys creditworthiness index (akin to asset returns
in a Merton model) is constructed as a linear factor model of a
set of credit drivers (and an idiosyncratic component);
conditional default probabilities are computed in each scenario
(Equation 27).
The computation of wrong-way exposures, via a direct or indirect
method, present numerical challenges as it requires conditioning
on an unlikely event (default). In this sense, Levy and Levin (1999)
discuss a particular heuristic methodology to obtain conditional
probabilities and scenarios for foreign-exchange transactions.
Alternatively, variance reduction techniques can be used together
with MC methods to obtain robust WWE statistics. For example,
Kalkbrener et al (2004) and Glasserman and Li (2003) discuss the
application of importance sampling for credit portfolio problems,
which are directly applicable to WWEs.
Pr
Pr
( )
( ( ))
CS Def S x
N N CS Def x
CP
CP k k
k
K
k
k
K

1
1
1
1
( )
( )
22
41
CONCLUDING REMARKS
This chapter presents a general modelling framework for counter-
party exposures and mitigation techniques in derivatives portfolios
and introduces the key concepts behind MC and analytical methods.
The modelling of stochastic PFEs is complex. Exposures may be
path dependent and largely discontinuous and the underlying
market processes have a substantial impact on them. Furthermore,
we require detailed modelling of netting and collateral, whose
values may also be stochastic and path dependent. Finally, both
exposures and collateral may be dependent on credit events, in
addition to market prices, as is the case with credit derivatives,
margin thresholds and WWE. An integrated model for market and
credit risk is further required for credit-state dependent exposures
and capital.
As exposure is only one component of credit risk, counterparty
exposure measurement and limits must be consistent with credit
pricing, capital allocation (economic and regulatory) and business
decision tools. While the link between PFEs to pricing and capital is
beyond the scope of this presentation, some final remarks are
appropriate.
Pricing the credit risk of derivatives requires, in essence, the cal-
culation of expected exposures in a risk-neutral measure (calibrated
to CDS market prices, spreads, etc). It is important to highlight that
the value of an instrument on a stand-alone basis might differ from
its marginal value when it is part of a netting set. The latter
depends on its marginal expected exposure (stand-alone exposures
are not additive with netting and collateral).
The role of PFEs in regulatory capital has evolved from the add-
on approach in the Basel I Accord, to internal PFE models (using
effective expected potential exposure and effective maturity) in Basel II
(BCBS 2005).
Finally, the integration of PFEs and credit portfolio models to
calculate economic capital presents three modelling challenges.
First, credit portfolio models generally assume that exposures are
deterministic (eg, Gupton et al 1997; Credite Suisse Financial
Products 1997). The accurate computation of economic capital for
derivatives portfolios requires an integrated market and credit-
risk model that allows for exposures to be (jointly) stochastic and
42
perhaps correlated with credit events (see Iscoe et al 1999; Rosen
and Sidelnikova 2001; Canabarro et al 2003 and chapters in this
volume). As a simpler alternative, practitioners generally use
the standard credit portfolio models loan equivalent exposures
ideally, the equivalent deterministic exposure that yields similar
marginal capital for that counterparty. Loan-equivalents are
largely based on expected positive exposure measures (with some
adjustments).
Second, credit capital is generally computed with one-step port-
folio models. In contrast, counterparty exposures generally vary
though time and have jumps; thus, the timing of default can
have a substantial effect on credit losses. Several credit portfolio
models have thus been extended to a multi-step setting (eg,
Iscoe et al 1999; Finger 2000; Thompson et al 2005). Alternatively,
practitioners commonly apply a one step model with a reason-
able approximation for the exposure over the period (eg, the
EPE, as the average exposure over time, serves this purpose).
Finally, PFEs are generally defined as amounts that would be
lost if a counterparty defaults and can only be used directly in
capital calculation for default losses. To measure mark-to-market
losses (eg, in a CreditMetrics model), we further require valuing
the portfolio under various counterparty credit-state assump-
tions (see Picoult 2005).
ACKNOWLEDGEMENTS
We would like to acknowledge the support of The Fields Institute
and Algorithmics Inc. Special thanks to Alex Kreinin and Helmut
Mausser for their fruitful discussions and comments on the manu-
script. The examples and graphs in this chapter were generated
using Algorithmics software.
1 Mark-to-market credit risk also covers losses incurred due to credit downgrades (or upgrades)
and changes in credit spreads.
2 More generally, exposures can also be defined conditional not only on counterparty default,
but also as potential losses on credit migration events such as credit downgrades. This is
necessary for mark-to-market credit portfolio models; for example CreditMetrics (Gupton
et al 1997).
3 This is done for ease of exposition and to relate to practical implementations. It is straight-
forward to extend these definitions to a continuous time and state space.
43
4 Valuation is thus assumed at mid-market (ie the mid between buy and sell rates, unen-
cumbered by credit risk adjustments). Acredit valuation adjustment (CVA) is an adjustment to
the mid-market valuation, which further reflects the market value of the credit risk of both
the bank and the counterparty. Aone-sided credit valuation adjustment reflects only the market
value of the credit risk of the counterparty. These are required for mark-to-market exposures
in credit portfolio models.
5 From here on, we simplify the notation by removing the time dependence of the scenario
and understand that the information used is only until time t
k
unless otherwise stated.
6 For ease of exposition, we have assumed a zero minimum transfer amount and upfront col-
lateral (see Levy and Clarke 2000) for formulae including these effects.
7 See also Dembo et al (2000), Aguais and Rosen (2001).
8 Peak exposures can also be defined in terms of expected shortfall (ES). The 100% ES is the
average of the 100(1 )% largest exposures:
where the indicator function 1
x
equals one if the condition x is satisfied and zero otherwise.
This formula includes a correction term, in the case of discrete loss distributions, to ensure
that the total probability of the averaged losses is exactly (1 ) (see Acerbi and Tasche 2002).
9 There is no universal naming for some of these measures. We have chosen to stay close to the
Basel II nomenclature (BCBS 2005). For example, various sources refer to the expected expos-
ure, above, also as the expected positive exposure, and to the EPE as defined above, as the
average EPE or the time-weighted EPE.
10 Alternative formulae for time weighted averages are obtained, in the discrete case, by using
the trapezoidal rule; eg,
11 Formula is for a one-sided CVA (Picoult 2005). EPE
D(t
0
)
denotes a discounted EPE computed
from discounted PFEs (Equation (1)). Discounting is at risk free rates. Spread is the average
spread for the counterparty, given its rating. See also footnote 6.
12 This is defined in analogy to a bond and its duration.
13 To simplify notation, we have removed the time and portfolio dependence from the parameters.
14 We might be tempted to use todays portfolio value. This would be reasonable if the
processes had zero drifts and if there are no intermediate cashflows or settlements. The first
assumption might be appropriate for valuation (under a risk-neutral measure), but not for
risk management purposes and the second assumption is not appropriate in general.
15 The superscript T denotes the transpose of a vector.
16 Popularised for market VAR by JP Morgan (1995) and extended to credit exposures in
Zangari (1997).
17 When the manifold is monotonic, the iso-exposure line is tangent to the density curve at the
optimal point.
18 Monotonicity of the manifold is sufficient but not necessary for the method to work.
19 This example appears also in De Prisco et al (2004).
20 The model does not necessarily result in parallel term structure movements. The presence of
mean reversion speed results in the short end of the term structure moving to a greater
degree than the long end.
21 The lower the a parameter for the short rate, the more parallel the movements.
22 The use of a five-factor model is illustrative.
EPE t
t t
t t t t
EPE
k
k o
E
l
E
l l l
l
k
Eff
( )
(
1
2
1 1
1
( )
( ) ( )
]
]
( )

tt
t t
t t t t
k
k o
Eff
E
l Eff
E
l l l
l
k
)
1
2
1 1
1
( )
( ) ( )
]
]
( )

ES t PFE S t Q t
E
k i k
i
i
PFE S t Q t
E
k
i k
E
k

( ) ( , ) ( )
( , ) ( )
1
1
1
Pr ( , ) ( ) PFE t Q t
k
E
k
]
]
( ) ( )
44
23 This is a conservative assumption but one that is often preferred in practice. Without further
complication, we can assume that the return and receipt of collateral operate under the same
margin period rules.
24 See De Prisco and Kreinin (2003).
25 Multi-state credit models have been used to price derivatives and loans (cf, Jarrow et al 1997,
Aguais et al 2000) and also for credit portfolio management (Gupton et al 1997).
26 If we assume that the Bank has negligible default probabilities compared to the counterparty
or that both counterparties are uncorrelated, then conditioning can be done only on the
counterpartys survival.
27 For example, Iscoe et al (1999), Aguais et al (2000), Grundke (2004).
28 For example, Jarrow et al (1997), Duffie and Singleton (2003).
29 For simplicity we omit the time dimension in the notation.
30 This was first discussed in Iscoe et al (1999).
31 For example, Duffie and Singleton (2003), Jarrow and van Deventer (2005).
32 See Finger (1999).
33 See Finger (2000), Iscoe et al (1999).
REFERENCES
Acerbi, C. and D. Tasche, 2002, On the Coherence of Expected Shortfall, Journal of
Banking and Finance, 26(7), pp 1487503.
Arvanitis, A. and J. Gregory, 2001, Credit: The Complete Guide to Pricing, Hedging and Risk
Management (London: Risk Books).
Aguais, S. and D. Rosen, 2001, Credit Risk: Enterprise Credit Risk using Mark-to-Future
(Toronto: Algorithmics Inc.).
Aguais S., L. Forest, and D. Rosen, 2000, Building a Credit Risk Valuation Framework
for Loan Instruments, Algo Research Quarterly, 3(3), pp 2146. Also in 2001, Commercial
Lending Review, 16(4), pp 1230.
Basel Committee on Banking Supervision, 1995, Basel Capital Accord: Treatment of
Potential Exposure for Off-Balance-Sheet Items. Available at http://www.bis.org
Basel Committee on Banking Supervision, 2004, International Convergence of
Capital Measurement and Capital Standards: ARevised Framework. Available at http://
www.bis.org
Basel Committee on Banking Supervision, 2005, The Application of Basel II to Trading
Activities and the Treatment of Double Default Effects, July. Available at http://
www.bis.org
Brummelhuis, R., A. Cordoba, M. Quintanilla, and L. Seco, 2002, Principal Component
Value-at-Risk, Mathematical Finance, 12, pp 2343.
Canabarro, E., E. Picoult, and T. Wilde, 2003, Analysing Counterparty Risk, Risk,
pp 11722, September.
Cardenas, J., E. Fruchard, and J.F. Picorn, 2002, Linear yet Attractive Contour, Credit,
pp 3840, January.
45
Cartolano, P. and S. Verma, 2000, Using Scenario Banding to Stress Test Counterparty
Credit Exposures, Algo Research Quarterly, 3(1), pp 2736.
Chishti, A., 1999, Simulation of Fixed-income Portfolios using Grids, Algo Research
Quarterly, 2(2), pp 4150.
Credite Suisse Financial Products, 1997, CreditRisk: ACredit Risk Management Framework
(Zurich: Credit Suisse Group).
Counterparty Risk Management Policy Group II, 2005, Toward Greater Financial
Stability: APrivate Sector Perspective. Available at http://www.crmpolicygroup.org/
De, R. and T. Tamarachenko, 2002, VaR You Can Rely On, Risk, pp 7781, August.
De Prisco, B. and A. Kreinin, 2003, On Computing of Lag-Adjusted Exposure, Research
Paper, Algorithmics Inc.
De Prisco, B., I. Iscoe, A. Kreinin, and A. Nagi, 2004, ASemi-Analytical Method for VaR
and Credit Exposure Analysis, Research Paper, Algorithmics Inc. (to be published in
Annals of Operations Research).
Dembo, R.S., A. Aziz, D. Rosen, and M. Zerbs, 2000, Mark-to-Future: A Framework for
Measuring Risk and Reward (Toronto: Algorithmics Inc.).
Duffie, D. and J. Pan, 1997, An Overview of Value-at-Risk, Journal of Derivatives, 4(3), pp749.
Duffie, D. and K.J. Singleton, 2003, Credit Risk Pricing Measurement and Management
(Princeton series in Finance) (Princeton, NJ: Princeton University Press).
Dupire, B. (ed), 2004, Monte Carlo Methodologies and Applications for Pricing and Risk
Management (London: Risk Books).
Finger, C., 1999, Towards a Better Estimation of Wrong-way Credit Exposure, RiskMetrics
Journal, Summer, pp 2639.
Finger, C., 2000, AComparison of Stochastic Default Rate Models, Risk Metrics Journal, 1,
pp 4973.
Glasserman, P., 2004, Monte Carlo Methods in Financial Engineering (New York: Springer).
Glasserman, P. and J. Li, 2003, Importance Sampling for Portfolio Credit Risk, Research
Paper, Columbia University.
Glasserman, P. and B. Yu, 2002, Pricing American Options by Simulation: Regression
Now or Regression Later?, in H. Niederreiter (ed), Monte Carlo and Quasi-Monte Carlo
Methods (Berlin: Springer).
Grundke, P., 2004, Risk Measurement with Integrated Market and Credit Portfolio
Models, Working Paper, Department of Banking, University of Cologne.
Gupton, G.M., C. Finger, and M. Bhatia, 1997, CreditMetrics: The Benchmark for
Understanding Credit Risk (New York: JP Morgan, Inc.).
Hille, C., J. Ring, and H. Shimamoto, 2005, Modelling Counterparty Credit Exposure for
Credit Default Swaps, Risk, May, pp 659 (Also featured in Counterparty Credit Risk
(M. Pykhtin, Editor), 2005, Risk Books, London.).
Hull, J. and A. White, 1990, Pricing Interest Rate Derivative Securities, Review of
Financial Studies, 3(4), pp 57392.
Iscoe, I., A. Kreinin, and D. Rosen, 1999, An Integrated Market and Credit Risk Portfolio
Model, Algo Research Quarterly, 2(3), pp 2138.
46
Jaschke, S.R., 2002, The Cornish-Fisher-Expansion in the Context of Delta-Gamma-
Normal Approximations, Journal of Risk, 4(4), pp 3352.
Jarrow, R., D. Lando, and S. Turnbull, 1997, AMarkov Model for the Term Structure of
Credit Risk Spreads, The Review of Financial Studies, 10(2), pp 481523.
Jarrow, R. and D. van Deventer, 2005, Estimating Default Correlations using a Reduced-
form Model, Risk, pp 837, January.
Morgan, J.P., 1995, Introduction to RiskMetrics, Technical Document, 4th Edition,
New York.
Kalkbrener, M., H. Lotter, and L. Overbeck, 2004, Sensible and Efficient Capital
Allocation for Credit Portfolios, Risk, pp S19S24, January.
Kreinin, A. and M. Sidelnikova, 2001, Regularization Algorithms for Transition
Matrices, Algo Research Quarterly, 2(1/2), pp 2340.
Levin, R. and A. Levy, 1999, Wrong-way Exposure, Research Paper, JP Morgan Securities
Inc., Derivatives Research. Also featured in Risk, 12(7), pp 525.
Levy, A. and M. Clarke, 2000, Modelling Derivative Credit Risk in the Presence of
Collateral and Netting, Research Paper, JP Morgan Securities Inc. Derivatives Research.
Also featured in Risk, 13(1), pp 1003.
Longstaff, F.A. and S.E. Schwartz, 2001, Valuing American Options by Simulation:
ASimple Least-Squares Approach, The Review of Financial Studies, 14(1), pp 11347.
Mina, J. and A. Ulmer, 1999, Delta-gamma Four Ways, Working Paper, RiskMetrics.
Picoult, E., 2005, Calculating and Hedging Exposure, CVA and Economic Capital for
Counterparty Credit Risk, Counterparty Credit Risk (M. Pykhtin, Editor), Risk Books,
London.
Pritsker, M., 1997, Evaluating Value at Risk Methodologies: Accuracy versus
Computational Time, Journal of Financial Services Research, 12(2/3), pp 20142.
Rebonato, R., 1998, Interest Rate Options Models, 2nd edn (Wessex: John Wiley & Sons).
Reimers, M. and M. Zerbs, 1999, AMulti-factor Statistical Model for Interest Rates, Algo
Research Quarterly, 2(3), pp 5364.
Rosen, D. and M. Sidelnikova, 2001, Understanding Stochastic Exposures and LGDs in
Portfolio Credit Risk, in S. Carrillo Mendez, A. Sanchez Calle and L. Seco (eds),
Proceedings of the First Annual RiskLab International Conference, Madrid, Spain, October
2001. Also featured in 2002, Algo Research Quarterly, 5(1), pp 4356.
Rowe, D., 1995, Aggregating Credit Exposures: The Primary Risk Source Approach,
in Derivative Credit Risk: Advances in Measurement and Management (London: Risk
Publications), pp 1321.
Rowe, D. 1999, How to Right Wrong-way Exposures, Risk, 12(10), p 49, Oct.
Rowe, D. and M. Mulholland, 1999, Aggregating Market-driven Credit Exposures:
A Multiple Risk Source Approach, in Derivative Credit Risk: Further Advances in
Measurement and Management, 2nd edn (London: Risk Books), pp 1417.
Schonbucher, P., 2004, Chapter III.B.3: Credit Exposure, in C.A. Alexander and
E. Sheedy (eds), The Professional Risk Managers Handbook (Wilmington: PRMIA
Publications), pp 22541.
Studer, G., 1999, Market Risk Computation for Nonlinear Portfolios, Journal of Risk, 1(4),
pp 3353.
47
Thompson, K., A. McLeod, P. Teklos, and S. Gupta, 2005, Time for Multi-Period Capital
Models, Risk, 18(10), pp 748.
Wakeman, L., 1996, Credit Enhancement, in C.A. Alexander (ed), The Handbook of Risk
Management and Analysis (West Sussex: John Wiley & Sons), pp 30727.
Zangari, P., 1997, On Measuring Credit Exposure, RiskMetrics Monitor, First Quarter,
pp 322.

1 StochasticCounterparty05

Încărcat de

Informații document

Descriere originală:

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

1 StochasticCounterparty05

Încărcat de

Drepturi de autor:

Formate disponibile

Section 1

01-Prisco.qxd 11/12/05 11:05 AM Page 8

01-Prisco.qxd 11/12/05 11:05 AM Page 17

is the -quantile of the standard normal distribution.

is the -quantile of the standard normal distribution.

S-ar putea să vă placă și