cutting

edge. inteReSt RAte deRiVAtiVeS

Smooth calibration of Markov functional models for pricing exotic interest rate derivatives
The Libor market model is widely used but often criticised for its slowness. Nick Denson and Mark Joshi develop an accurate and stable calibration procedure that allows for the effective use of a control variate
as a vector covariance matrix. The LMM cannot be represented as a low-dimensional Markov process for two reasons. The first is that the drift is state-dependent. The second is that even if we set all drifts to zero, the timedependence of the volatility loadings will cause the set of attainable forward-rate curves to be high-dimensional. For these reasons, it is usual to implement the LMM using Monte Carlo simulation. However, if we modify the LMM appropriately, we can obtain a Markov-functional model.
Markov-functional model

The Libor

market model (LMM) is commonly used for pricing exotic interest rate derivatives. However, its popularity is limited by the fact that its highdimensionality necessitates (quasi)-Monte Carlo simulation for pricing. Such simulation is often criticised because of its low convergence rate and the difficulties of pricing early-exercisable derivatives. Lattice models do not suffer from these problems, but provide less flexibility in terms of correlation structures and calibration to market data. One solution (Piterbarg, 2004) is to use a Markov functional approximation to the LMM as a control variate. Thus the part that is able to be priced on a lattice is priced quickly, while the remaining part has lower variance and thus the Monte Carlo simulation converges quickly. The key features of the control variate are that it can be evaluated using both partial differential equations (PDEs) and simulations, and that the simulation can be made to be highly correlated with the LMM simulation. There has been little discussion of how to carry out the fitting. In this article, we solve the problem of how to fit a Markov functional approximation to a given calibration of the LMM in such a way that the control variate varies smoothly with the calibration. One virtue of this approach is that it allows the possibility of using the control variates for vegas as well as prices and deltas.
Displaced-diffusion Libor market model

We take as our starting point the Markov-functional model of Pietersz, Pelsser & Van Regenmortel (2003). However, we can encompass a more general volatility structure. We consider the case of an F-factor model for generality but it is the case where F = 2 that is important for us. n F-factor model. We require the volatility structure to be separable, that is: where ni is an F-dimensional row vector and C(t) is an F F matrix. The volatility structure gets its name because the forward rate and time dependence are separated out. Separability (see Pietersz, Pelsser & Van Regenmortel, 2003) is often used to refer to the case where C(t) is diagonal but this is an overly strong restriction. We illustrate using a two-factor model. Our volatility structure is:

i ( t ) = i C ( t )

(2)

i ( t ) = vi,1

We briefly recall the displaced-diffusion LMM. It is based on the idea of evolving discrete market observable forward/Libor rates. We have a set of tenor dates 0 < T0 < T1 < ... < Tn and n corresponding forward rates f0, f1, ... , f n1, where f i(t) represents the forward rate at time t for the interval [Ti, Ti+1). Let P(t, T) denote the price at time t of the zero-coupon bond paying one at its maturity T. Throughout this article, we will work in the terminal measure, which corresponds to taking P(t, Tn) as the numeraire. In the displaced-diffusion LMM, the forward rates are assumed to have the following evolution:

This volatility form differs from Piterbargs (2004) since he sets c21(t) = 0. The extra non-zero term allows better fitting of pseudosquare roots. We define an F-dimensional vector of Markov factors via:
dX ( t ) = C ( t ) dW ( t )

c11 ( t ) c12 ( t ) vi, 2 c (t ) c (t ) 22 21

(3)

where W(t) is the projection of the Brownian motion in (1) on to the first F co-ordinates. If we can consider an LMM driven only by the first F factors then its dynamics can be written as:

fi ( t ) + i

dfi ( t )

= i ( t ) dt + i dX ( t )

(4)

where the sis are deterministic D-(1 D n 1) dimensional row vectors, the ais are constant displacement coefficients, W is a standard D-dimensional Brownian motion and the drift term is given by no-arbitrage arguments. For more details on the LMM, we refer the reader to Brace (2007) or Joshi (2003). A calibration procedure results in a covariance matrix of the rates for each step of the simulation. Two procedures that result in the same covariance matrices will yield the same prices. In what follows, we will therefore regard a calibration of the LMM

fi ( t ) + i

dfi ( t )

= i ( f ,t ) dt + i ( t ) dW ( t )

(1)

and so if the drifts were dropped the forward rates would be Markovian in X(t). Our Markov-functional model takes this equation and defines the forward rates at time t as a function of X(t) to be the values that would be obtained by evolving from zero to t using a single-step discretisation. We use the iterative variant of predictor-corrector for stepping (see Hunter, Jckel & Joshi, 2001, and Joshi & Stacey, 2008). The deflated price of a derivative, V, satisfies a PDE:

with i = /xi, where xi represents the Markov factors. For the coefficients, let cij(t) be the element in the ith row of the jth column of C(t) in (2), and then:

F V = qij ( t ) i jV t i, j =1

(5)

risk-magazine.net

81

cutting edge. inteReSt RAte deRiVAtiVeS

qij ( t ) =

1 2

cif ( t ) c jf ( t )
f =1

squares norm of the difference:

VC ( i ) B i

n Analytic calibration. The innovation of our article lies in how we fit the Markov-functional models calibration to that of the LMM. We make both the evolution of the LMM and that of the fitted control variate vary smoothly with the calibration. For the LMM, a calibration is a covariance matrix for each step and the evolution is determined by a choice of pseudo-square root for each step. We use the spectral decomposition: that is, if the covariance matrix for step i has eigenvalues lj in decreasing order, with corresponding eigenvectors ej, then we take:
bj = j ej

including only the first F columns of Bi. Since the kth column of C(i) only contributes to the kth column of VC(i), we have a separate least-squares optimisation problem for each k and i. If we let b denote the target column and let c denote the relevant column of C(i), we have to minimise: where V is a n F matrix, c is an F-vector and b is a n-vector. This can be solved using standard linear algebra techniques for solving an over-determined system. Since it is low-dimensional linear algebra, the calibration is performed effectively instantaneously within a computer program. We could also perform the calibration starting with any evolution interval. Consider the evolution over [Ti1, Ti). We exactly fit the first F columns of Bi by setting: with C(i) equal to the identity matrix (note that some elements of nj will temporarily be set to zero as elements of Bi are zero due to reset forward rates). We perform a least-squares fit to the n i 1 pseudo-square roots after the evolution time Ti. To fit the pseudo-square roots before Ti1, we work backwards considering one evolution interval at a time. We calculate the matrix C(i 1) by solving the system of equations, as above, including nj for j = i, ... , n 1. Once C(i 1) has been calculated, we set ni1 so that the first F columns of the i 1th row of Bi1 are exactly matched. We repeat, working backwards until all pseudo-square roots have been fitted. This approach gives the flexibility of exactly fitting the first F columns of a particular pseudo-square root. We see that the choice of pseudo-square root exactly fitted does affect the accuracy of the overall calibration. However, the calibration method is very fast so it is possible to iterate over all possible choices and select the best within a fraction of a second.
Pricing j = b ij,1 K b ij, F Vc b

to be column j of our pseudo-square root, Bi. As long as the eigenvalues are distinct, both they and the eigenvectors are analytic (and therefore smooth) functions of the matrix (Andrew, Chu & Lancaster, 1993). This approach also maximises the impact of the first F factors. For fitting the Markov functional model, we take C(t) to be constant over each interval [Ti1, Ti). Let C(i) represent the matrix of constant values over that interval, with the convention that T1 = 0. For an F-dimensional Markov-functional model and a product that depends on n forward rates, there are Fn + F2 n parameters to be fitted: the Fn forward rate specific scalars and F2 n piecewise-constant time-dependent values. We will denote the pseudo-square roots arising from our separable fit as Ai. One possible methodology, as in Piterbarg (2004), is to use a global least-squares numerical optimisation. However, there is no guarantee in that case that the matrix Ai varies smoothly with Bi. Here, we present a calibration method that is both analytic and stable. It has the features of using spectral methods to fit one matrix and inductively fitting the other matrices one by one. For the rest of this section, we shall address the problem of fitting an F-dimensional Markov-functional model to a D-dimensional LMM, where F D. Naturally, we will only attempt to fit the first F columns of the matrices Bi. However, the fact that we are using the pseudo-square root defined by spectral decomposition means that these will be the dominant factors. Consider the evolution of the forward rates over [0, T0), with the corresponding pseudo-square roots B 0 and A0. By looking at equations (1) and (4), we see that if we have: by an appropriate choice of C values, then we can exactly calibrate the first F columns of A0 to B 0, simply by setting: for j = 0, ... , n 1 with:
j = b 0 K b 0 F j, j,1 dX ( t ) = dW ( t )

To apply the Markov-functional model as a control variate, we price a product V using:

Vcontrol = VLMM (VMC VPDE )

Let V denote the matrix with columns nj (which is equal to A0). Now that we have set all the forward rate specific scalars by exactly fitting to the first F columns of the first pseudo-square root, we need to fit the remaining n 1 pseudo-square roots. Our remaining freedom lies in our ability to pick the matrices C(i). For each i, we want to choose C(i) to minimise the sum-of-

1 0 C (0) = M 0

0 K 0 1 K 0 M O M 0 K 1

(6)

where VMC and VPDE represent the deflated price in the Markovfunctional model implemented via a Monte Carlo simulation and PDE method respectively. Using different numerical implementation methods, we should still arrive at the same price. However, the variance of the price will be reduced because of the correlation between VLMM and VMC (see section 4.1 of Glasserman, 2003). The Markov-functional Monte Carlo implementation needs to be as close as possible to that of the LMM. This means the same random numbers must be used for the first two factors when evolving rates. We also use the LMM exercise times for the Monte Carlo Markov-functional model as this leads to the largest variance reduction. For the lattice method, we apply the exercise strategy from the LMM. The slight difference between Monte Carlo and lattice exercise strategies will result in a bias. However, our results show this bias to be very small. We consider the prices of at-the-money payer Bermudan swaptions starting in six months, which can be exercised on any of the reset dates. Forward rates are semiannual and initially flat, calculated using a continuously compounded rate of 5%, that is, P(t,

82

Risk August 2010

A. Prices and standard errors in basis points for different tenor Bermudan swaptions
two-factor 5Y LMM control 10Y LMM control 20Y LMM control 226 225 479 481 852 853 (2.66) (0.01) (5.63) (0.10) (9.58) (0.48) 20x 53x 183x 220 221 480 482 850 853 three-factor (2.62) (0.07) (5.73) (0.25) (9.65) (0.86) 11x 23x 39x 219 220 476 477 839 838 Five-factor (2.60) (0.08) (5.44) (0.29) (9.80) (0.97) 10x 19x 31x

B. Model vegas and standard errors in basis points for the Bermudan swaptions priced in table A
two-factor 5Y LMM control 10Y LMM control 20Y LMM control 10.8 10.8 24.4 24.5 46.1 46.2 (1.5E-01) (1.9E-03) (3.7E-01) (1.7E-02) (7.7E-01) (9.6E-02) 8x 21x 80x 10.6 10.5 24.9 25.0 46.2 46.3 three-factor (1.5E-01) (9.3E-03) (3.8E-01) (2.9E-02) (7.8E-01) (1.1E-01) 7x 13x 16x 10.5 10.7 24.2 24.4 45.5 45.6 Five-factor (1.4E-01) (5.6E-03) (3.6E-01) (3.0E-02) (8.2E-01) (1.7E-01) 5x 12x 26x

Note: two-, three- and five-factor LMMs were used with the control variate

Note: all caplet volatilities were increased by 1%

T) = er(Tt). We use the abcd volatility structure:

with a = 0.05, b = 0.09, c = 0.44, d = 0.12 and correlation between rates: with L = 0.5, b = 0.1 as this represents what we consider a fairly difficult market scenario. All displacements are equal to zero, that is, ai = 0 i. To calculate the exercise strategy, we use the least-squares algorithm with the Andersen method superimposed, as discussed in Bender, Kolodko & Schoenmakers (2006). A Sobol quasi-random number generator with 214 paths was used to determine the exercise strategy and then 10,000 Mersenne Twister paths were used to estimate the price and standard error. We optimised the fitting procedure to get the best overall fit for the Markov-functional model. Table A shows prices, standard errors and the multiple of standard error reduction using the control variate. The control variate works very well. For a five-year Bermudan swaption, a standard error reduction of 31 to 183 times is achieved. The method is also effective for longer-dated products. We achieved a standard error reduction of 10 times for a five-factor 20-year Bermudan swaption, which is an increase in speed of around 100 times. In line with expectations, the control variate works best for shorter-dated products and when coupled with a low-factor LMM. Somewhat surprising is the control variates effectiveness for longerdated products, which is due to the accurate calibration method.
Vega ij = L + (1 L ) e
Ti T j

c T t i ( t ) = a + b ( Ti t ) e ( i ) + d

best global fit. Focusing on the second row of the table, for example, we see that for a 10-year Bermudan swaption the standard error reduction for the vega calculation ranges from 12 to 21 times. The contribution of the third and above factors appears to be larger for the vega than the price and so we obtain smaller variance reductions. However, a factor of five translates into an increase in speed of around 25 times, so these are still significant.
Conclusion

We have presented a Libor Markov-functional model as a control variate for the displaced-diffusion LMM and introduced a new analytic calibration method. The calibration method is effective for vegas as it allows for small perturbations in the LMM volatility structure to be translated into small perturbations in the separable volatility structure. The control variate produces large standard error reductions for pricing, ranging from 10 to 20 times for the 20-year Bermudan swaptions. It also produces useful reductions for vega calculations, ranging from five to eight times for the 20-year product. These standard error reductions are significant as they dramatically reduce the number of required simulations for a desired level of accuracy. n
nick denson is a Phd student and Mark Joshi is an associate professor at the centre for Actuarial Studies at the university of Melbourne. e-mail: ndenson@gmail.com, mark@markjoshi.com

References
Andrew A, K-W Chu and P Lancaster, 1993 Derivatives of eigenvalues and eigenvectors of matrix functions SIAM Journal on Matrix Analysis and Applications 14(4), pages 903926 Bender C, A Kolodko and J Schoenmakers, 2006 Iterating cancellable snowballs and related exotics Risk September, pages 126130 Brace A, 2007 Engineering BGM Chapman & Hall/CRC Glasserman P, 2003 Monte Carlo methods in financial engineering Springer Hunter C, P Jckel and M Joshi, 2001 Getting the drift Risk July, pages 8184 Joshi M, 2003 The concepts and practice of mathematical finance Cambridge University Press Joshi M and A Stacey, 2008 New and robust drift approximations for the Libor market model Quantitative Finance 8(4), pages 427434 Pietersz R, A Pelsser and M Van Regenmortel, 2003 Fast drift approximated pricing in the BGM model Journal of Computational Finance 8(1), pages 93124 Piterbarg V, 2004 A practitioners guide to pricing and hedging callable Libor exotics in forward Libor models Journal of Computational Finance 8, pages 65119

To calculate the vega, we use a finite-difference approach, where we compute the price using the control variate method, then perturb the LMM volatility, recalibrate our Markov-functional model (which is performed very quickly) and then compute the new price. The crucial fact here is that the eigenvalues and eigenvectors of a matrix are analytic functions of its entries (see Andrew, Chu & Lancaster, 1993), provided the eigenvalues are distinct. The constraint of distinctness will generally be fulfilled since the first eigenvalue is usually dominant when we consider correlation matrices for forward rates. n Vega results. We present results for the computed model vega for the corresponding Bermudan swaption prices in table A. We computed a one-sided difference, where the change in price is shown in table B. Table B shows the vega estimate when d in the abcd volatility structure was perturbed by 1%. Once again we have used 214 Sobol paths to determine the exercise strategy and 10,000 Mersenne Twister paths to determine both the unperturbed and perturbed prices. We exactly fit to the pseudo-square root that produces the

risk-magazine.net

83