cutting

edge. cRedit Risk

PR

One of the

most active areas of risk management today is counterparty credit risk management (CCRM). Managing counterparty risk is particularly challenging because it requires the simultaneous evaluation of all the trades facing a given counterparty. For multi-asset portfolios, this typically comes with extraordinary computational challenges. Indeed, for portfolios other than those comprising simple vanilla instruments, computationally intensive Monte Carlo (MC) simulations are often the only practical tool available for this task. Standard approaches for the calculation of risk require repeating the calculation of the profit and loss of the portfolio under hundreds of market scenarios. As a result, in many cases these calculations cannot be completed in a practical amount of time, even employing a vast amount of computer power. Since the total cost of through-the-life risk management can determine whether it is profitable to execute a new trade, solving this technology problem is critical to allow a securities firm to remain competitive. Following the introduction of adjoint methods in finance (Giles & Glasserman, 2006), a computational technique dubbed adjoint algorithmic differentiation (AAD) (Capriotti, 2011, and Capriotti & Giles, 2010 and 2011) has recently emerged as tremendously effective for speeding up the calculation of sensitivities in MC in the context of the so-called pathwise derivatives method (Broadie & Glasserman, 1996). Algorithmic differentiation (AD) (Griewank, 2000) is a set of programming techniques for the efficient calculation of the derivatives of functions implemented as computer programs. The main idea underlying AD is that any such function no matter how complicated can be interpreted as a composition of basic arithmetic and intrinsic operations that are easy to differentiate. What makes AD particularly attractive, when compared with standard (finite-difference) methods for the calculation of derivatives, is its computational efficiency. In fact, AD exploits the information on the structure of the computer code in order to optimise the calculation. In particular, when one requires the derivatives of a small

FO R

RE

OD UC

NO

Adjoint algorithmic differentiation can be used to implement the calculation of counterparty credit risk efficiently. Luca Capriotti, Jacky Lee and Matthew Peacock demonstrate how this powerful technique can be used to reduce the computational cost by hundreds of times, thus opening the way to real-time risk management in Monte Carlo

number of outputs with respect to a large number of inputs, the calculation can be highly optimised by applying the chain rule through the instructions of the program in opposite order with respect to their original evaluation (Griewank, 2000). This gives rise to AAD. Surprisingly, even if AD has been an active branch of computer science for several decades, its impact in other research fields has been fairly limited until recently. Interestingly, in contrast to the usual situation in which well-established ideas in applied mathematics or physics have often been borrowed by quants, AAD has been introduced in MC applications in natural science (Sorella & Capriotti, 2010) only after its rediscovery in quantitative finance. In this article, we demonstrate how this powerful technique can be used for highly efficent computation of price sensitivities in the context of CCRM.

Counterparty credit risk management

As a typical task in the day-to-day operation of a CCRM desk, here we consider the calculation of the credit valuation adjustment (CVA) as the main measure of a dealers counterparty credit risk. For a given portfolio of trades facing the same investor or institution, the CVA aims to capture the expected loss associated with the counterparty defaulting in a situation in which the position, netted for any collateral agreement, has a positive mark-tomarket for the dealer. This can be evaluated at time T0 = 0 as:
VCVA = E I ( c T ) D ( 0, c ) LGD ( c ) NPV ( c ) C R c

TI
(
O

ON
(

Real-time counterparty credit risk management in Monte Carlo

where tc is the default time of the counterparty, NPV(t) is the net present value of the portfolio at time t from the dealers point of view, C(R(t)) is the collateral outstanding, typically dependent on the rating R of the counterparty, LGD (t) is the loss given default, D(0, t) is the discount factor for the interval [0, t], and I(tc T) is the indicator that the counterpartys default happens before the longest deal maturity in the portfolio, T. Here, for simplicity of notation we consider the unilateral CVA, and the generalisation to bilateral CVA (Brigo & Capponi, 2010) is straightforward. The quantity in (1) is typically calculated on a discrete time grid of horizon dates T0 < T1 < ... < TN as, for instance:
VCVA ; E I (Ti1 < c Ti ) D ( 0,Ti ) i=1

( ( )))

(1)

NO T

LGD (Ti ) NPV (Ti ) C R Ti

In general, the quantity above depends on several correlated random market factors, including the interest rate, the counterpar-

( ( )))

(2)

86

Risk June 2011

V = E Q P ( R, X )
NO

with payout given by:

(3)

P = P (Ti , R (Ti ) , X (Ti ))

i=1 NR

RE

Here the rating of the counterparty entity including default, R(t), is represented by an integer r = 0, ... , NR for simplicity; X(t) is the realised value of the M market factors at time t. Q = Q(R, X) represents a probability distribution according to which R = (R(T1), ~ ... , R(TN ))t and X = (X(T1), ... , X(TN ))t are distributed; Pi(; r) is a rating-dependent payout at time Ti.1 The expectation value in (3) can be estimated by means of MC by sampling a number NMC of random replicas of the underlying rating and market state vector, R[1], ... , R[NMC] and X[1], ... , X[NMC], according to the distribution Q(R, X), and evaluating the payout P(R, X) for each of them. This leads to the central limit theorem (Kallenberg, 1997) estimate of the option value V as:
0 0

% P (Ti , R (Ti ) , X (Ti )) = Pi ( X (Ti ) ;r ) r, R(Ti )

r=0

OD UC
(5) (6)
V ()

where:

(4)

TI
P ( R, X ) = E P k

V () (8) = E P ( R, X ) k k by defining appropriate estimators that can be sampled simultaneously in the same MC simulation. This can be achieved by observing that whenever the payout function is regular enough (for example, Lipschitz-continuous, see Glasserman, 2004), one can rewrite equation (8) by taking the derivative inside the expectation value, as:

N MC iMC =1 __ __ with standard error S/NMC, where S2 = EQ[P(R, X)2] EQ[P (R, X)]2 is the variance of the sampled payout. In the following, we will make minimal assumptions about the particular model used to describe the dynamics of the market factors. In particular, we will only assume that for a given MC sample the value at time Ti of the market factors can be obtained from their value at time Ti1 by means of a mapping of the form X(Ti) = Fi(X(Ti1), Z X) where Z X is a N X dimensional vector of correlated standard normal random variates, X(T0) is todays value of the market state vector, and Fi is a mapping regular enough for the pathwise derivatives method to be applicable (Glasserman, 2004), as is generally the case for practical applications. As an example of a counterparty rating model generally used in practice, here we consider the rating transition Markov chain model of Jarrow, Lando & Turnbull (1997) in which the rating at time Ti can be simulated as:

V;

N MC

P ( R [iMC ] , X [iMC ])

PR

ON

tys default time and rating, the recovery amount, and all the market factors the net present value of the portfolio depends on. As such, its calculation requires a MC simulation. To simplify the notation and generalise the discussion beyond the small details that might form part of a dealers definition of a specific credit charge, here we consider expectation values of the form:

In this setting, MC samples of the payout estimator in (6) can be generated according to the following standard algorithm. For i = 1, ... , NO: n Step 1. Generate a sample of N X + 1 jointly normal random variables (Z R, Z X) (Z R, Z X , ... , Z X )t distributed according to i i i i,1 i,N (Z R, Z X; i), a (N X+1)-dimensional standard normal probability i density function with correlation matrix i, for example, with the first row and column corresponding to the rating factor. n Step 2. Iterate the recursion X(Ti) = Fi(X(Ti1), Z X ). _ ~ n Step 3. Set ZR = Sij=1 Z R/i and calculate R(Ti) according to i j (7).2 n Step 4. Calculate the time Ti payout estimator P(Ti, R(Ti), X(Ti)) in (5), and add this contribution to the total estimator in (4). The calculation of risk can be obtained in a highly efficient way by implementing the pathwise derivatives method (Broadie & Glasserman, 1996) according to the principles of AAD (Capriotti, 2011, and Capriotti & Giles, 2010 and 2011). The pathwise derivatives method allows the calculation of the sensitivities of V (6) with respect to a set of N parameters = (1, ... , N ), say:
X

where P = P(Z R, Z X ) is the distribution of the correlated normal variates used in the MC simulation, which is independent of .3 The calculation of equation (9) can be performed by applying the chain rule, and calculating the average value of the pathwise derivatives estimator: P ( R, X ) NO M P ( R, X ) Xl (Ti ) P ( R, X ) k = + (10) k k k i=1 l=1 Xl (Ti )

(9)

FO

where ZR is a standard normal variate, and Q(Ti, r) is the quantilei threshold corresponding to the transition probability from todays rating to a rating r at time Ti. Note that the discussion below is not limited to this particular model, and it could be applied with minor modifications to other commonly used models describing the default time of the counterparty and its rating (Schnbucher, 2003). Here we consider the rating transition model (7) for its practical utility, as well as for the challenges it poses in the application of the pathwise derivatives method, because of the discreteness of its state space.

% R (Ti ) = I ZiR > Q (Ti ,r )

~

NR

r=1

(7)

where we have allowed for an explicit dependence of the payout on the model parameters.4 Due to the discreteness of the state space of the rating factor, the pathwise estimator for its related sensitivities is not well defined. However, as we will show below, one can express things in such a way that the rating sensitivities are incorporated in the explicit term P(R, X)/k. In the following, we will show how the calculation of the pathwise derivatives estimator (10) can be implemented efficiently by means of AAD. We begin by briefly reviewing this technique.
Adjoint algorithmic differentiation

NO T

Griewank (2000) contains a detailed discussion of the computational cost of AAD. Here, we will only recall the main results in order to clarify how this technique can be beneficial for the effi1 The discussion below applies also to the case in which the payout at time Ti depends on the history of the market factors X up to time Ti 2 Here we have used the fact that the payout (5) depends on the outturn value of the rating at time Ti and not on its history 3 For simplicity of notation, we exclude the case in which includes the elements of the correlation matrix in (Z R, Z X; ). The extension to this case is straightforward and can be performed along the lines of Capriotti & Giles (2010) _ 4 Here and in the following we will use the standard AD notation k to indicate the sensitivity of the payout with respect to the model parameter k

risk.net/risk-magazine

87

cutting edge. cRedit Risk

cient implementation of the pathwise derivatives method. The interested reader can find in Capriotti (2011) and Capriotti & Giles (2010 and 2011) several examples illustrating the intuition behind these results. To this end, consider a function: mapping a vector X in R to a vector Y in R through a sequence of steps:
n m

Cost [ FUNCTION _ b ] A Cost [ FUNCTION ]

Y = FUNCTION ( X )

(11)

RE

which corresponds to the adjoint mode equation for the intermediate function V = V(U): V Ui = Vk k (14) Ui k _ _ _ namely a function of the form U = V (U, V ). Starting from the _ adjoint of the outputs, Y , we can apply this to each step in the calculation, working from right to left:

PR

FO

with i = 1, ... , n. In the adjoint mode, the cost does not increase with the number of inputs, but it is linear in the number of (linear combinations of the) rows of the Jacobian that need to be evaluated independently. In particular, if the full Jacobian is required, one needs to repeat the adjoint calculation m times, setting the vec_ tor Y equal to each of the elements of the canonical basis in Rm . Furthermore, since the partial (branch) derivatives depend on the values of the intermediate variables, one generally first has to compute the original calculation storing the values of all the intermediate variables such as U and V, before performing the adjoint mode sensitivity calculation. One particularly important theoretical result (Griewank, 2000) is that given a computer program performing some highlevel function (11), the execution time of its adjoint counterpart:

(15) X LU V L Y _ until we obtain X , that is, the following linear combination of the rows of the Jacobian of the function X Y: m Y Xi = Y j j (16) Xi j=1

NO T

(with suffix _b for backward or bar) calculating the linear combination (16) is bounded by approximately four times the cost of execution of the original one, namely:

X = FUNCTION _ b ( X,Y )

88

Risk June 2011

OD UC
(17)

Here, each step can be a distinct high-level function or even an individual instruction. The adjoint mode of AD results from propagating the derivatives of the final result with respect to all the intermediate variables the so-called adjoints until the derivatives with respect to the independent variables are formed. Using the standard AD notation, the adjoint of any intermediate variable Vk is defined as: m Y Vk = Y j j (13) Vk j=1 _ where Y is the vector in Rm. In particular, for each of the intermediate variables Ui, using the chain rule we get: m m Y Y V Ui = Y j j = Y j j k Ui j=1 k Vk Ui j=1

X L U V L Y

(12)

with wA [3, 4]. Thus, one can obtain the sensitivity of a single output, or of a linear combination of outputs, to an unlimited number of inputs for a little more work than the original calculation. As also discussed at length in Capriotti (2011) and Capriotti & Giles (2010 and 2011), AAD can be straightforwardly implemented by starting from the output of an algorithm and proceeding backwards, applying systematically the adjoint composition rule (14) to each intermediate step, until the adjoints of the inputs (16) are calculated. As already noted, the execution of such backward sweep requires information that needs to be calculated and stored by executing beforehand the steps of the original algorithm the so-called forward sweep. A simple illustration of this procedure is discussed in the Appendix.
AAD and counterparty credit risk management

(18)

When applied to the pathwise derivatives method, AAD allows the simultaneous calculation of the pathwise derivatives estimators for an arbitrarily large number of sensitivities at a small fixed cost. Here, we describe in detail the AAD implementation of the pathwise derivatives estimator (10) for the CCRM problem (1). As noted above, the sensitivities with respect to parameters affecting the rating dynamics need special care due to the discrete nature of the state space. However, setting these sensitivities aside for the moment, the AAD implementation of the pathwise derivatives estimator consists of Steps 14 described above plus the following steps of the backward sweep. For i = NO, ... , 1: _ n Step 4 . Evaluate the adjoint of the payout:

( X (Ti ) , ) = P (Ti , R (Ti ) , X (Ti ) ,, P ) _ with P _ 1. = n Step 3. Nothing to do: the parameters do not affect this nondifferentiable step. _ n Step 2. Evaluate the adjoint of the propagation rule in Step 2:
( X (Ti1 ) , ) + = Fi ( X (Ti1 ) ,, Z X , X (Ti ) , )

where + = is the standard addition assignment operator. _ n Step 1 . Nothing to do: the parameters_ do not affect this step. A few comments are in order. In Step 4 , the adjoint of the payout function is defined while keeping the discrete rating variable _ _ ) constant. This provides the derivatives X l(Ti_ = P/Xl(Ti), and k = P/k. In defining the adjoint in Step 2, we have taken into account that the propagation rule in Step 2 is explicitly dependa ent on both X(Ti) and the model parameters . As _ result, its _ adjoint counterpart produces contributions to both and X(Ti). Both the adjoint of the payout and of the propagation mapping can be implemented following the principles of AAD as discussed in Capriotti (2011) and Capriotti & Giles (2011). In many situations, AD tools can also be used as an aid or to automate the implementation, especially for simpler, self-contained functions. _ _ In the backward sweep above, Steps 1 and 3 have been skipped because we have assumed for simplicity of exposition that the parameters do not affect the correlation matrices i and the rat_ ing dynamics. If correlation risk is instead required, Step _ also 2 produces the adjoint of the random variables Z X, and Step 1 contains the adjoint of the Cholesky decomposition, possibly with the support of the binning technique, as described in Capriotti and Giles (2010).

TI

ON

Rating transition risk

The risk associated with the rating dynamics can be treated by noting that (5) can be expressed more conveniently as:
% % P Ti , ZiR , X (Ti ) = Pi ( X (Ti ) ;0 )
NR

ON

% % % + Pi ( X (Ti ) ;r ) Pi ( X (Ti ) ;r 1) I Z iR > Q (Ti ,r;)

r=1

)(

(19)

1 speed-up in the calculation of risk for the cVA of a portfolio of five commodity swaps over a fiveyear horizon, as a function of the number of risks calculated (empty dots)
160 140 120 Speed-up/RCPU 100 80 60 40 20

so that the singular contribution to the pathwise derivatives estimator reads:

% % % k P Ti , Zi , X (Ti ) = Pi ( X (Ti ) ;r ) Pi ( X (Ti ) ;r 1)

% ZiR = Q (Ti ,r;) k Q (Ti ,r;)

r=1

This estimator cannot be sampled in this form with MC. Nevertheless, it can be integrated out using the properties of Diracs delta along the lines of Joshi & Kainth (2004), giving after straightforward computations:
k =
NR

OD UC
0 0

(20)

TI

100 200

NR

Z i

, ZiX ,i ZiX ,iX

r=1

RE

% % Pi ( X (Ti ) ;r ) Pi ( X (Ti ) ;r 1) _ 1 where Z* is such that (Z* + Sji =1Z R)/i = Q(Ti, r; ), and (Z X, X ) j i i is a N X-dimensional standard normal probability density function X with correlation matrix i obtained by removing the first row and column of i; here Q(Ti, r; ) is not stochastic and can be evaluated (for example, using AAD) once per simulation. The final result is rather intuitive as it is given by the probability-weighted sum of the discontinuities in the payout.

) )

k Q

(Ti ,r;)

300 Nrisks

400

500

600

(21)

PR

Note: the full dots are the ratio of the CPU time required for the calculation of the CVA, and its sensitivities, and the CPU time spent for the calculation of the CVA alone. Lines are guides for the eye
Ne

Results

FO

where Wt is a standard Brownian motion; F T (t) is the price at time t of a futures contract expiring at T; sT and b define a simple instantaneous volatility function that increases approaching the contract expiry, as empirically observed for many commodities. The value of the futures price F T (t) can be simulated exactly for any time t so that the propagation rule in Step 2 reads for Ti T: 1 FT (Ti ) = FT (Ti1 ) exp i Ti Z i2 Ti (23) 2

As a numerical test, we present here results for the calculation of risk on the CVA of a portfolio of commodity derivatives. For the purpose of this illustration, we consider a simple one-factor lognormal model for the futures curve of the form: dFT (t ) = T exp ( (T t )) dWt (22) FT (t )

Note that although we consider here for simplicity of exposition a linear portfolio, the method proposed applies to an arbitrarily complex portfolio of derivatives, for which in general NPV(t) will be a non-linear function of the market factors Ft (t) and model parameters . _ For this example, the adjoint propagation rule in Step 2 simply reads:
j

NPV (t ) = D (t,t j ) Ft j (t ) K
j=1

(24)

1 FT (Ti 1) + = FT (Ti ) exp i Ti Z i2 Ti 2 i = FT (Ti ) F (Ti ) Ti Z i Ti _ with si related to_ this steps contribution to the adjoint of the futures volatility sT by:
e2T e2Ti e2Ti1 _ _ At the end of the backward path, F T (0) and s T contain the pathwise derivatives estimator (10) corresponding, respectively, to the sensitivity with respect to todays price and volatility of the futures contract with expiry T. The remarkable computational efficiency of the AAD implementation is illustrated in figure 1. Here, we plot the speed-up produced by AAD with respect to the standard finite-difference method. On a fairly typical trade horizon of five years, for a portfolio of five swaps referencing distinct commodities futures with monthly expiries, the CVA bears non-trivial risk to more than 600 parameters: 300 futures prices (FT (0)), and at-the-money volatilities (sT), say 10 points on the zero rate curve, and 10 points on the credit default swap curve of the counterparty used to calibrate the transition probabilities of the rating transition model (7). As illustrated in figure 1, the computation time required for T + = i 2Ti

NO T

where DTi = Ti Ti1 and:

i2 =

2 T 2T 2Ti e e e2Ti1 2Ti

is the outturn variance. In this example, we will consider deterministic interest rates. As an underlying portfolio for the CVA calculation, we consider a set of commodity swaps, paying on a strip of futures (for example, monthly) expiries tj, j = 1, ... , Ne the amount Ft (tj) K. The time t net present value for this portfolio reads:
j

risk.net/risk-magazine

89

cutting edge. cRedit Risk

Appendix: a simple example

As a simple example of adjoint algorithmic differentiation (AAD) implementation, we consider an algorithm mapping a set of inputs (1, ... , n) into a single output P, according to the following steps: n Step 1. Set Xi = exp(2/2 + iZ), for i = 1, ... , n, where Z is a constant. i n Step 2. Set P = (Sn Xi K)+, where K is a constant. i=1 The corresponding adjoint algorithm consists of Steps 1 and 2 (forward sweep), plus a backward sweep consisting of the adjoints of Steps 2 and 1, respectively: _ _ _ n Step 2. Set Xi = P I(Sn Xi K > 0), for i = 1, ... , n. i=1 _ _ n Step 1. Set i = Xi(i + Z), for i = 1, ... , n. We can immediately verify that the output of the adjoint algorithm above _ _ gives for P = 1 the full set of sensitivities with respect to the inputs, i = P/i. Note that, as described in the main text, the backward sweep requires information that is calculated during the execution of the forward sweep, Steps 1 and 2, for example, to calculate the indicator I(Sn Xi K) and the value of Xi. Finally, i=1 simple inspection shows that both the forward and the backward sweep have a computation complexity O(n), that is, all the components of the gradient of P can be obtained at a cost that is of the same order of the cost of computing P, in agreement with the general result (18). It is easy to recognise in this example a stylised representation of the calculation of the pathwise estimator for vega (volatility sensitivity) of a call option on a sum of lognormal assets.

FO R

the calculation of the CVA and its sensitivities is less than four times that spent for the computation of the CVA alone, as predicted by equation (18). As a result, even for this very simple application, AAD produces risk measures more than 150 times faster than finite differences, that is, for a CVA evaluation taking 10 seconds, AAD produces the full set of sensitivities in less than 40 seconds, while finite differences require approximately 1 hour and 40 minutes. Moreover, as a result of the analytic integration of the singularities introduced by the rating process, the risk measures produced by AAD are typically less noisy than those produced by finite differences. This is illustrated in table A, which shows the variance reduction on the sensitivities with respect to the thresholds Q(Ti, r) for a simple test case. Here, we have considered the calculation of a call option of the form (FT(Ti) C(R(Ti)))+ with a strike C(R(Ti)) linearly dependent on the rating, and Ti = 1. The variance reduction displayed in the table can be thought of as a further speed-up factor because it corresponds to the reduction in the computation time for a given statistical uncertainty on the sensitivities. This diverges as the perturbation in the finite-difference estimators d tends to zero, and may be very significant even for a fairly large value of d.

RE

PR
VR[Q(1, 3)] 12 125 1,350 16 165 1,640

OD UC
References
Brigo D and A Capponi, 2010 Bilateral counterparty risk with application to CDSs Risk March, pages 8590

AAD therefore makes possible real-time risk management in MC, allowing investment firms to hedge their positions more effectively, actively manage their capital allocation, reduce their infrastructure costs and ultimately attract more business. n
Luca capriotti is a director, Jacky Lee is a managing director and Matthew Peacock is a vice-president in the Quantitative strategies department of the investment banking division of credit suisse group in new York. they would like to thank Mike giles, Adam Peacock, nick seed and Mark stedman for numerous useful discussions, and Fredrik Akesson for a careful reading of the manuscript. the opinions and views expressed in this article are those of the authors, and do not necessarily represent those of credit suisse group. email: luca.capriotti@credit-suisse.com, shinghoi.lee@credit-suisse.com, matthew. peacock@credit-suisse.com

Conclusion

In conclusion, we have shown how AAD allows an extremely efficient calculation of counterparty credit risk valuations in MC. The scope of this technique is clearly not limited to this important application but extends to any valuation performed with MC. For any number of underlying assets or names in a portfolio, the proposed method allows the calculation of the complete risk at a computational cost that is at most four times the cost of calculating the profit and loss of the portfolio. This results in remarkable computational savings with respect to standard finite-difference approaches. In fact, AAD allows one to perform in minutes risk runs that would take otherwise several hours or could not even be performed overnight without large parallel computers.

Broadie M and P Glasserman, 1996 Estimating security price derivatives using simulation Management Science 42, pages 269285 Capriotti L, 2011 Fast Greeks by algorithmic differentiation Journal of Computational Finance 3(3), pages 335 Capriotti L and M Giles, 2010 Fast correlation Greeks by adjoint algorithmic differentiation Risk April, pages 7983 Capriotti L and M Giles, 2011 Algorithmic differentiation: adjoint Greeks made easy Available at http://ssrn.com/ abstract=1801522 Giles M and P Glasserman, 2006 Smoking adjoints: fast Monte Carlo Greeks Risk January, pages 9296 Glasserman P, 2004 Monte Carlo methods in financial engineering Springer, New York

NO T

A. Variance reduction (VR) on the sensitivities with respect to the thresholds Q(1, r) (NR = 3) for a call option with a rating-dependent strike
d 0.1 0.01 0.001 VR[(Q(1, 1)] 24 245 2,490 VR[Q(1, 2)]

Note: d indicates the perturbation used in the finite-difference estimators of the sensitivities. The specification of the parameters used for this example is available upon request

90

Risk June 2011

TI

ON

Griewank A, 2000 Evaluating derivatives: principles and techniques of algorithmic differentiation Frontiers in Applied Mathematics, Philadelphia Jarrow R, D Lando and S Turnbull, 1997 A Markov model for the term structure of credit spreads Review of Financial Studies 10, pages 481523 Joshi M and D Kainth, 2004 Rapid computation of prices and deltas of n-th to default swaps in the Li model Quantitative Finance 4, pages 266275 Kallenberg O, 1997 Foundations of modern probability Springer, New York Schnbucher P, 2003 Credit derivatives pricing models: models, pricing, implementation Wiley Finance, London Sorella S and L Capriotti, 2010 Algorithmic differentiation and the calculation of forces in quantum Monte Carlo Journal of Chemical Physics 133, pages 234111 110