Sei sulla pagina 1di 31

Factor Dependence of Bermudan Swaption Prices: Fact or Fiction?

Leif Andersen and Jesper Andreasen General Re Financial Products First Version: November 5, 1999 This Version: March, 2000

Abstract
This paper investigates the effect of interest rate correlation in the pricing of Bermudan swaptions. Investigating both Gaussian Markov models and Libor Market models, we find that Bermudan swaption prices change only moderately (and in fact typically decrease slightly) when the number of factors in the underlying interest rate model is increased from one to two. Our findings are markedly different from those of Longstaff, Schwarz, and Santa-Clara (1999) who conclude that single-factor interest rate models significantly undervalue Bermudan swaptions. We argue that the conclusions of Longstaff, Schwarz, and Santa-Clara are due to idiosyncrasies in the authors choices of model dynamics and calibration methodology. Our study highlights the importance of using a reasonable set of calibration instruments when applying and comparing interest rate models. JEL Classification: G12, G13, E43 Keywords: Bermudan swaptions, multi-factor Gaussian models, multi-factor Libor market models, model calibration

1. Introduction
Although Bermudan swaptions are among the most actively traded fixed income derivatives, there is currently no consensus on what type of model is appropriate to accurately value these contracts. To incorporate the early exercise feature in a numerically efficient way, most practitioners tend to use relatively simple Markov-style one-factor models, such as those proposed in Hull and White (1990) and Black et al (1990). In a recent paper, Longstaff, Schwarz, and Santa-Clara (LSS) (1999) investigate the effect on Bermudan swaption prices from raising the number of stochastic factors (and thus decreasing correlation between different parst of the yield curve). Using a string interest rate model, the authors detect high sensitivity of Bermudan swaptions prices to the number of factors used in their model, and suggest that the one-factor models commonly used on Wall Street practice significantly underestimate the true swaption price. Indeed, estimating the impact on the market as a whole, the authors conclude that the present value cost to swaption holders of following myopic singlefactor exercise strategies is in the order of several billion dollars. An informal survey of practitioners and academics revealed broad support for the qualitative (although not necessarily quantitative) results of LSS: multi-factor models are generally believed to give rise to at least somewhat higher Bermudan swaption prices than one-factor models. Given that a Bermudan swaption can be viewed as a type of complicated best-of chooser

Communicating author. Address: 630 Fifth Avenue, Suite 450, New York, NY 10111. Phone: (212) 3072339; Fax: (212) 307-2320; E-mail: landerse@genre.com.

option granting the holder the right to choose between a set of European swaptions, it certainly seems natural to assume that increasing the number of model factors (and hence decreasing correlation between the various underlyings of the contract) would lead to an increase in the price. While the decorrelation effect is, in fact, real, it turns out that it is typically offset by more subtle effects related to the forward dynamics of interest rate models. Indeed, it is clear that decreasing interest rate correlation cannot be done in a vacuum: any change in correlations will impact par yield volatilities and will require a recalibration of the model to ensure its consistency with quoted prices. As we will show, the process of calibrating the model to market will induce volatility effects that often more than offset the nave decorrelation effect described above. We argue that the conclusions in LSS are a by-product of both a calibration procedure with too much slack as well as the authorschoice of highly non-stationary interest rate dynamics. In our own experiments, we consider two types of models: a mean-reverting Gaussian model, and a log-normal Libor Market (LM) Model. For both model types, we consider the effects of pricing Bermudan swaptions using one and two driving factors. The Gaussian model can conveniently be implemented in a finite difference grid (an ADI scheme was used for the two-factor version), allowing for Bermudan swaption pricing by standard backwards induction. The LM model was set up in the Monte Carlo framework of Andersen and Andreasen (1998), with Bermudan swaptions priced as in Andersen (2000). For both model types, we find that standard Bermudan swaptions have little sensitivity to the number of factors driving the yield curve; and, in fact, the prices of standard Bermudan swaptions decrease slightly as the number of factors driving the yield curve is increased. The numerical results supporting this conclusion have been prepared with two main objectives in mind: first, the one- and two-factor models have been made comparable by calibrating them to hit the same consistent set of cap and swaption prices which span the Bermudan swaption in a reasonable way. Second, we wish for our models to be as time-stationary as possible, particularly when it comes to the evolution of interest rate volatility. Both objectives are, we feel, prerequisites for performing a realistic comparison of one- and two-factor models1. As is well-known, however, both objectives cannot be satisfied fully: it is generally not possible to make a model with a certain interest rate correlation structure perfectly match European prices generated by a time-stationary model with a different correlation structure without making some of the parameters of the fitted model time-dependent. This means that the model will not have a completely stationary volatility structure, and this again may induce effects that influence the pricing of Bermudan swaptions. To ensure that problems with time-stationarity do not lead to biased conclusions we perform two sets of comparisons: first, we calibrate a onefactor model to European interest rate options generated by a time-stationary two-factor model. Then we reverse the order and calibrate a two-factor model with time-stationary correlation structure (but non-stationary volatility) to a time-stationary one-factor model. In the same spirit, we use flat initial yield curves to minimize curve and interest rate skew effects. The paper is organized as follows: In Section 2, we briefly introduce some notation and define the payouts of European and Bermudan swaptions. In Section 3, we
1

Of course, if one has a firm belief that interest rate volatilities will change over time in a specific manner, this requires that the volatility structure be non-stationary (in a controlled fashion). In practice, timestationarity appears to be a reasonable and robust hypothesis in most markets.

discuss the model setup of LSS (1999), and point out the characteristics which we believe are responsible for their conclusions about Bermudan swaption pricing (and which we wish to avoid in our own setup). Section 4 contains our results for the Gaussian model, and includes a break-down of the effects associated with a change in the number of model factors. In Section 5, we repeat some of the experiments in Section 4 for the lognormal LM model, thereby making sure that the qualitative results of Section 4 are not unique to Gaussian models. Section 5 further considers the effect of using a one-factor exercise strategy in a two-factor model and also briefly discusses the pricing of so-called constant tenor Bermudan swaptions, for which the factor effect is shown to typically be opposite that of standard Bermudan swaptions. Finally, Section 6 contains our conclusions. To avoid cluttering the main text of the paper with too much notation and experimental data, some technical background information and many numerical results have been relegated to appendices.

2. Notation and payout definitions


Briefly speaking, a swaption is the right, but not the obligation, to enter into a fixed-for-floating interest rate swap. To be specific, introduce a uniformly spaced maturity structure 0 = T0 < T1 <... < TK + 1 , with Ti = i Typically, is either 0.25 or 0.5 . years. Consider now a European swaption S s ,e maturing at some date Ts , s {1,2,..., K} . The swaption gives the holder the right to enter into a fixed-floating interest rate swap where fixed cashflows c> 0 paid at Tk , k = s + 1, s + 2,..., e , are swapped against floating Libor (paid in arrears) on a $1 notional. The constant c is the coupon rate, and Ts and Te are the start- and end-dates of the underlying swap, respectively; clearly, we require TK + 1 Te > Ts . Let P (t , T ) denote the time t price of $1 received for sure at time T. Assuming that the option holder pays the fixed leg of the swap (a payer swaption), we can write the time Ts price of the swaption as
S s ,e

F ( T ) = G P ( T , T H
e 1 s s k =s

k+ 1

I )F (T ) c]J , [ K
+ k s

where x + max( x ,0) , and where we have introduced discretely compounded (Libor) forward rates Fk (t ) 1 P (t , Tk ) / P (t , Tk + 1 ) 1 . A little thought shows that the European swaption price can alternatively be written

S s ,e (Ts ) = Bs ,e (Ts ) s ,e (Ts ) c , Bs ,e (t ) = (t , Tk + 1 ) , P


+ k =s

e 1

where s ,e defines the forward par yield

s ,e (t ) =

P (t , Ts ) P (t , Te ) . Bs ,e (t )

(1)

A standard model (e.g. Jamshidian 1997, or Andersen and Andreasen 1998) for swaption pricing assumes that the par yield satisfies an SDE of the type d s ,e (t ) =... dt + s ,e (t ) (t , Ts , Te ) dW (t ) (2)

where W is a vector-valued Brownian motion, > 0 is constant, and is a deterministic vector-valued function of time. Notice that we have left the drift-term unspecified in (2), as it depends on the probability measure in which the dynamics are specified. The term volatility2 (also known as the implied volatility) function consistent with (2) is

v (t ; Ts , Te ) = (Ts t ) 1

z
t

Ts

|| (u; Ts , Te )||2 du

(3)

A model is defined to have time-stationary volatility if


(t , Ts , Te ) = (Ts t , Te Ts )

for all t Ts < Te < TK + 1 ; i.e. if the implied volatility of a swaption with option life Ts t and swap tenor Te Ts is independent of calendar time t. A reasonable degree of timestationarity of volatility is generally a desirable feature of interest rate models, particularly when the model is intended for the pricing of options that are sensitive to the future evolution of interest rate volatility (such as forward starting options or Bermudan options). Unlike European swaptions which have a single exercise date ( Ts ), a Bermudan swaption allows its holder to choose any one exercise date freely in the set {Ts , Ts + 1 ,..., Te 1} . Assuming that the option holder elects to exercise at time Tl , s l < e , he would notify the option seller at time Tl and would receive the European swaption payout Sl ,e (Tl ) at this time. The structure defined here is often known as a Te noncall Ts Bermudan swaption, and is by far the most common in practice. We point out, however, that other variations of Bermudan swaptions exist. One such variation is the constant tenor Bermudan swaption which at exercise pays a swap with a fixed number of payment periods , independent of when exercise takes place (unlike a regular Bermudan swaption where the length of the underlying swap shrinks as one moves from one exercise date to the next). So assuming that exercise takes place at some Tl {Ts , Ts + 1 ,..., Te 1}, the constant tenor Bermudan swaption pays out Sl ,l + (Tl ) . Our primary focus in this paper is on regular Bermudan swaptions, but we shall briefly consider a constant tenor Bermudan swaption in Section 5. For later use, let us introduce the concept of Bermudan swaption core rates, which we define as the set of forward par yields { s ,e , s + 1,e ,..., e 1,e }; that is, the core rates are simply the par yields of the set of swaps into which a regular Bermudan swaption can be exercised. In many settings it is convenient to approximate the forward par yields with continuously compounded forward zero-coupon rates. These rates are defined by
While the word volatility is normally associated with log-normal models where = 1, we use it here in a broader sense.
2

Y (t , Ti , Tj ) = (Tj Ti ) 1 ln

P (t , Ti ) , t Ti < Tj . P (t , Tj )

(4)

While the approximation Y (t , Ti , Tj ) i , j (t ) is not very precise, it is nevertheless useful for qualitative discussions of par yield volatilities, particularly in a Gaussian setting.

3. The LSS approach.


In this section we will briefly discuss the framework used in LSS (1999) and attempt to explain the effects leading the authors to conclude that Bermudan swaption prices increase in the number of stochastic factors. In translating the authors string model to a more standard framework, we will take some liberties in the conversion of notation. 3.1. LSS Setup. A key model assumption in LSS is the specification that the local volatility of a particular par yield is only a function of the absolute start- and end-dates of the swap not current time to the start- and end-dates. Using the notation introduced in the previous section, LSS thus makes the assumption that the time t local volatility of the par yield s ,e is independent of time t:

t , Ts , Te Ts , Te ,

g b

(5)

Recall that a time-stationary volatility specification sets t , Ts , Te = Ts t , Te Ts , whereby (5) necessarily implies a highly non-stationary volatility structure. The specification (5) can be argued to be at odds with conventional interest rate theory, as well as empirical observation: the implied volatility of European swaption prices typically decrease as either time to expiry ( Ts t ) or the tenor (Te Ts ) of the underlying swap increase. In the LSS approach, the e s core rates { s ,e , s + 1,e ,..., e 1,e } all equipped with an individual time-independent volatility, as in (5) constitute the main model primitives. Calibrating the model is merely a matter of choosing the constant core rate volatilities to match the time 0 prices of the corresponding European swaptions. Notice that once this calibration has been performed we are free to specify an arbitrary matrix of correlations between the various core rates without affecting the fit to the European option prices. While modifying correlations does not alter the present or future marginal distributions of the individual core rates, it certainly affects the overall dispersion of the core rates relative to each other. Drawing on the analogy between Bermudan swaptions and best-of chooser options, it is thus immediately obvious that decorrelating core rates will increase the prices of Bermudan swaptions in the LSS framework. 3.2. Some implications of the LSS setup. Besides implying non-stationary rate dynamics, the methodology employed by LSS has important consequences for prices of non-core instruments. As one can imagine,

g b

the fact that one can alter correlations in the LSS framework without re-calibrating volatilities implies that the calibration has significant slack. It is easy to see, that, say, if a model is fit to price a 10-year annual Bermudan swaption (and thereby only calibrated to fit to the 9 European swaptions (Ts , Te ) (110),(2,10),...,(9,10) ), there is no guarantee , that the model will give reasonable prices of, for instance, the swaptions (Ts , Te ) (1,2),(2,3),(3,4),...,(9,10) . As such, the calibrated model will generally not be consistent with the market for interest rate caps3. In general, the dynamics of all non-core rates in the LLS model are quite arbitrary and will depend strongly on the choice of correlation matrix. Specifically, the volatilities of non-core rates implied by the model will increase as the rates are de-correlated. To show this effect, it is enlightening to consider an example. For simplicity, we will consider the continuously compounded forward zero-coupon rates defined in (4), rather than forward par yields. Moreover, to avoid level dependence of volatility let us assume that the forward zero-coupon rates follow a Gaussian distribution. We now fix the horizon to Te years and consider the core rates Y Ti , Te where i = 1,2,K , e 1. Making an assumption similar to the one in LSS, , we assume that

(t , Ti , Te ) 2 var dY (t , Ti , Te ) = (Ti , Te ) 2 , t Ti .
To consider the effect of de-correlating the individual rates, we make the (very) simplistic assumption that
corr dY (t , Ti , Tm ), dY (t , Tj , Tn ) = , m n ,

where [ 1,1] is some constant, and t Ti < Tm Te , t Tj < Tn Te . The obvious relation for Ti < Tm ,

Y t , Ti , Tm Tm Ti = Y (t , Ti , Ti + 1 ) Y t , Ti + 1 , Tm Tm Ti + 1 , +
then implies that t , Ti , Tm

gb

gb
m

g bT T g = b t, T , T g + bt, T , T g b T T g + 2 b t , T , T g b t , T , T gb T T g.
2 2 2 m i i i+ 1 i 2 2 i+ 1 m i+ 1 i+ 1 i+ 1 m m i+ 1

(6)

This equation defines a recursive relation which can be used to back out the implied volatilities of non-core rates (i.e. rates where m < e ). Assuming t , Ti , Te = 0.01, e = 4 , and = 1, we get the following non-core rate volatility structure, as a function of :

An interest rate cap is, essentially, a strip of European swaptions, where each swaption (or caplet) only covers a single period (that is, Te = Ts + See Hull (1999) for further discussion of interest rate caps. ).

Table 1: Volatility of non-core zero-coupon rates in example; t < 1


(t ,1,3) (t ,2,3) (t ,1,2) = 0.5 0.01186 0.01303 0.01435 = 0.75 0.01088 0.01137 0.01189 = 0.99 0.01003 0.01005 0.01007

It is obvious from Table 1 that volatilities of non-core rates increase as the correlation is lowered. Had we, for instance, written a 2-year cap on the rates Y(11,2) and Y(2,2,3) , we , would have seen the price of the cap increase as the rates become more de-correlated. This effect, which is obviously undesirable in a model used for practical pricing, provides us another way of understanding the increase in Bermuda swaption value reported in LSS (1999). Specifically, it is well-known that Bermudan swaption prices can be bounded from above by an interest rate cap, and from below by the most expensive European swaption in the exercise set. In the LSS framework, reducing the correlation between interest rates does not affect the lower bound, but, as we have argued above, will increase the value of the upper bound. This again suggests that the price of the Bermudan swaption will increase as we decrease rate correlation. As was the case for the core rates, the volatilities of non-core rates in the LSS model are independent of calendar time, contradicting market evidence. 3.3. Extensions of the LSS calibration. Realizing the problems of having cap volatilities move around freely with changes in the correlation structure, we might be tempted to include caps or swaptions on non-core rates in the calibration set. Not surprisingly, this is, however, not possible in the LSS setup without altering the assumption (5) to allow for time-dependent par yield volatilities. Continuing the simple example from Section 3.2, consider for instance the task of correctly pricing a caplet on a rate Y (Ti , Ti , Ti + 1 ) . In other words, besides matching the term volatilities of the core rates, we insist on our process dynamics matching some given time 0 value of the term variance

Ti

(t , Ti , Ti + 1 ) 2 dt = 2 Te Ti 2

{b g z b t , T , T g dt b T T g z b t , T , T g dt} + b T T g z b t , T , T g dt 2 b T T gz b t , T , T g b t , T , T gdt ,
2 Ti 2 2 Ti + 1 2 0 i e e i+ 1 0 i+ 1 e 2 Ti + 1 2 e i+ 1 Ti i+ 1 e 1 Ti e i+ 1 0 i i+ 1 i+ 1 e

where the equality follows from a rearrangement of (6). Imagine now that we increase the correlation , for example by moving from a two-factor model to a one-factor model. In the expression on the right-hand side, the terms in curly brackets are term volatilities of core rates and must stay fixed for the model to remain in calibration. To ensure that the equality holds (and thereby that the caplet stays in calibration), we conclude that the future term variance

Ti + 1

Ti

t , Ti + 1 , Te dt
2

must necessarily increase when goes up. To ensure that the total term volatility of Y ( Ti + 1 , Te ) remains constant, the partial term variance ,

Ti

t , Ti + 1 , Te dt
2

must then decrease. Under the LSS dynamics, t , Ti + 1 , Te is not even a function of t, whereby such changes to the future and partial term variances are obviously not feasible, making calibration to the caplet impossible. The example above highlights the important fact that inclusion of non-core instruments such as caps in the calibration set will induce correlation-dependent changes in the evolution of core rate volatilities over time. As we shall see in the next section, such changes can have a significant impact on the pricing of Bermudan swaptions.

4. Bermudan Swaption Pricing in a Gaussian Model.


In this section, we will attempt to quantify the effect of rate correlation on Bermudan swaption prices in a Gaussian Markov model. In calibration of the model, we pay particular attention to the problems of the LSS approach: non-stationary interest rate dynamics and excessive slack in the calibration. 4.1. A Gaussian two-factor model. The model used in this section is a Gaussian Markov model, under which the dynamics of the short rate4 r is given as
r (t ) = a (t ) + x1 (t ) + x2 (t )

(7)

where a is a time-dependent function chosen so that the model matches the initial yield curve, and x1 , x2 are two correlated Ornstein-Uhlenbeck processes under the risk-neutral probability measure. That is,
dxi (t ) = i (t ) xi (t )dt + i (t )dWi (t ), i = 1,2 , dW1 (t )dW2 (t ) = (t )dt ,

(8)

where W1 and W2 are Brownian motions under the risk-neutral measure, and 1 , 2 , , 1 , and 2 are time-dependent deterministic functions. It can be shown that any arbitrage-free two-factor Markovian interest rate model with Gaussian interest rates can be described this way. Under (7)-(8), the instantaneous correlation between instantaneous forward rates f,
f (t , T ) = ln P (t , T ) / , T
4

This rate is defined as r ( t ) = ln P (t , T ) / |T = t . T

depend only on the quantities ( ,2 1 , 2 / 1 ) ; see Appendix A. This again implies that if the quantities ( ,2 1 , 2 / 1 ) are held constant for all times then the correlation structure of the forward rates will be time-stationary. We notice that if the correlation structure is fixed this way, then the two-factor model does not have more free parameters to vary than a one-factor model. In the model (7)-(8), European caps can be priced exactly in closed-form, and European swaptions can be priced efficiently by 1-dimensional numerical integration5. Pricing of Bermudan swaptions involve the numerical solution of a 2-dimensional partial differential equation (PDE). For maximum precision, the solution of the PDE has been implemented as an ADI (Alternating Directions Implicit) finite difference scheme, see e.g. Mitchell and Griffiths (1981) or Morton and Mayers (1994). 4.2. Test Description. To test the effect of rate correlation on Bermudan swaption prices, we first define a reasonably parameterized time-stationary two-factor model, where all parameters 1 ,2 , , 1 , 2 have been set to constants in such a way that the model is reasonably consistent with current conditions in the US. We then calibrate a one-factor version of the model with time-dependent short rate volatility (but constant mean reversion speed ) as closely as possible to the two-factor model. The Bermudan swaption prices of both models are then compared. The calibration of the one-factor model to the two-factor model can be done in numerous ways, but in no circumstances will we be able to perfectly match every single swaption price that can be generated by the two-factor model. As we are primarily interested in Bermudan swaptions with final maturity ( Te ) of 10 years, we here elect to match the model as closely as possible (in the sense of sum of least-squares) to i) 9 European at-the-money swaptions with final maturity Te = 10 and option maturities ranging from 1 year to 9 years, in annual steps: Ts = 1,2,...,9 ; and ii) forward starting at-the-money caps with terminal maturity of 10 years and time to first rate-set ranging from 1 year to 9 years, in annual steps. While i) is similar to the calibration set in LSS, notice that we specifically include caps in our calibration to prevent non-core rate volatilities from being overly affected by correlation changes. The model calibration used here is fairly standard among practitioners when using a onefactor model to price Bermudan swaptions. To make our testing symmetric and to eliminate effects of time-dependent parameters we also perform the test the other way around. That is, we first use a timestationary one-factor model to price the Bermudan swaptions. We then calibrate a twofactor model with stationary correlation structure to the European swaption and cap prices of the one-factor model and price the Bermudan swaptions in this model. For ease of reference, let us denote this test Test II, and the test described above Test I. We use the same calibration instruments in both tests. In all tests, we exclusively consider semi-annual structures (that is, = 0.5 ) and assume that the initial yield curve is flat at 6%, continuously compounded. One-factor
The latter result can be obtained by realizing that the price of a swap at any time can be written as a function of the jointly Gaussian variables ( x1 , x 2 ) . By conditioning on one of the variables and using the results in Jamshidian (1989) we find that the European swaption price can be found by numerical integration in one variable.
5

prices are computed in a Crank-Nicolson finite difference grid with 50 time steps and 50 spatial steps; two-factor prices are computed in an ADI finite difference grid with 50 time steps and 50 steps in each of the two spatial directions. For the instruments we are considering, numerical experiments suggest that the error associated with the chosen discretization is consistently less than 0.5 basis points (one basis point = 1/10,000). 4.3. Test results. We first turn to Test I and set the two-factor parameters to the constant values in Table B.1 in Appendix B. The values chosen are, as mentioned earlier, roughly consistent with current (December 1999) conditions in the US market. The correlation structure implied by the selected process parameters is displayed in Table B.2 in Appendix B. Table B.1 further contains the values of the best-fit one-factor model; notice that the onefactor short-rate volatility is now an increasing function of time, making the model (slightly) non-stationary, as discussed above. As documented in Table B.3, the quality of the fit to the 18 cap and swaptions in the calibration set is excellent, with virtually all price errors less than one or two basis point. In Figure 1 below, we have used the formulas in Appendix A to graph the volatilities of instantaneous forward rates in the one- and two-factor models, at two different points in time. Figure 1: Volatility of f(t,T) as a function of T- t (Test I)
0.015 0.013 2-factor, t = 0 & t = 5 0.011 1-factor, t = 0 1-factor, t = 5 0.009 0.007 0.005 0.003 0 2 4

Volatility

T-t

10

As the figure shows, the steepness of the forward volatility term structure is higher in the one-factor model6 is higher than in the two-factor model, a relationship that can be understood by realizing that par yields can be viewed as weighted sums of forward rates (recall that the variance of a sum is highest, when all components are perfectly
6

Alternatively, we can say that the mean reversion in the one-factor model is higher than in the two-factor model.

10

correlated). Notice also that while the two-factor volatilities are perfectly time-stationary, the one-factor volatilities increase slightly as calendar time progresses. This is a typical result of calibrating the model to both caps and swaptions. To translate Figure 1 into more useful form, consider now switching from instantaneous forward rates to the continuously compounded zero-coupon rates Y (t , Ti , Te ) defined in (4). These rates are good approximations of the Bermudan swaption core rates and allow for closed computation of their instantaneous volatilities (t , Ti , Te ) by the formulas in Appendix A. More importantly, we can compute the integrals of 2 in closed form to generate the zero-coupon yield term volatilities (similar to equation (3)) which are more directly linked to swaption prices. Figure 2 below graphs the difference in zero-coupon yield term volatilities for the one- and two-factor models in Test I, at various points in time t. As our Bermudan swaption maturity is 10 years, we have fixed Te to 10 years in the figure. Figure 2: Difference in Term Volatilities of Y (t , T ,10) for 1- and 2-factor models in Test I
0.001 t=5 0.0008

1-fac. Vol - 2-fac. Vol

t=4 t=3 t=2 t=1 t=0

0.0006 0.0004 0.0002 0 1 -0.0002

While the term volatilities are virtually identical at time 0 (indicating that the model is correctly calibrated to the European swaptions) we see that as time progresses, the term volatilities of the core rates in the one-factor model increase relative to those of the twofactor model. In other words, the two models imply different future dynamics of the core rates. This is consistent with the qualitative discussion given in Section 3.3. To see how the effect in Figure 2 might affect Bermudan swaption prices, consider a Bermudan payer swaption with swap maturity Te and only two exercise dates: T1 , T2 . While the term volatilities v (0; T1 , Te ) and v (0; T2 , Te ) (see (3)) are independent of rate correlation (for a properly calibrated model), we know from the figure above that the future term volatility v (T1 ; T2 , Te ) will increase with increasing rate correlation. As argued earlier, this requires that the partial term variance

11

T1

1 T1 0

|| (u; T2 , Te )||2 du

(9)

will decrease. Now, standing at the first exercise date T1 , the holder of the Bermudan swaption has the right to choose between taking a swap starting at that date or taking a European swaption maturing at T2 . So the value of the Bermudan swaption at the first exercise date must be the maximum of the swap and swaption values. Raising correlation causes an increase in the future volatility v (T1; T2 , Te ) , and hence the value of the European swaption, which again raises the value of the Bermudan swaption at T1 . The decrease of the partial variance in (9), on the other hand, will lower the dispersion of 2,e (T1 ) somewhat, which will tend to offset the effect of the increase in v (T1; T2 , Te ) . The sensitivity of Bermudan prices to partial variance (9) is, however, often limited as effects from partial variance typically are absorbed by the effect of the term volatility v (0; T1 , Te ) . For most Bermudan swaptions the effect of an increase in the future volatility v (T1; T2 , Te ) can be expected to be stronger than the more secondary effect of lowering the partial volatility (9). Consequently, we would typically see that the net effect of the two volatility changes to be an increase in the Bermudan swaption value. Drawing on earlier discussion (in Section 3.1), we have identified two opposing effects on Bermudan swaption prices from increasing interest rate correlations: i) ii) a lowered dispersion of the underlying European swaptions relative to each other, resulting in a decrease of the price an increase in term volatilities of core rates as calendar time passes, leading to an increase in the price

Other subtle effects might also come into play, but whether the net effect on Bermudan swaption prices is positive or negative is an empirical question, to which we now turn. In Tables 2 and 3 below, we report the Test I prices of Bermudan receiver and payer swaptions, respectively. The tables consider various coupon rates c, with the first exercise date Ts varying between 1 and 5 years. For all numbers in the tables, the terminal swaption maturity Te is 10 years. Table 2: 10-year Bermudan Receiver Swaption Prices in Test I
Ts = 1 Ts = 2 Ts = 3 Ts = 4 Ts = 5

c 4% 5% 6% 7% 8%

1-fac. 93.7 206.1 419.6 782.0 1289.2

2-fac. 90.8 200.8 412.6 772.7 1284.5

1-fac. 93.6 204.2 404.3 722.7 1146.2

2-fac. 90.6 198.6 397.4 712.8 1139.1

1-fac. 91.9 195.5 374.9 645.0 995.4

2-fac. 89.2 190.8 368.0 635.8 988.1

1-fac. 87.6 180.6 335.3 558.9 844.6

2-fac. 85.6 177.0 329.4 551.5 838.6

1-fac. 80.3 160.4 288.3 467.9 695.1

2-fac. 79.0 157.8 283.8 462.8 691.1

All prices in basis points.

12

Table 3: 10-year Bermudan Payer Swaption Prices in Test I


Ts = 1 Ts = 2 Ts = 3 Ts = 4 Ts = 5

c 4% 5% 6% 7% 8%

1-fac. 1393.2 872.3 492.2 253.8 121.2

2-fac. 1391.5 865.4 481.8 245.7 116.9

1-fac. 1233.2 800.2 469.9 249.1 120.5

2-fac. 1229.4 792.2 459.8 241.5 116.4

1-fac. 1067.8 709.1 429.9 236.2 117.4

2-fac. 1063.3 702.2 421.7 229.4 113.5

1-fac. 904.1 610.7 379.3 215.0 110.5

2-fac. 899.9 605.6 373.7 209.9 107.3

1-fac. 741.6 508.3 322.0 187.2 99.3

2-fac. 738.3 504.7 318.6 184.3 97.2

All prices in basis points.

It is obvious from Tables 2 and 3 that the error involved in using a best-fit one-factor model to price Bermudan swaptions is quite small, typically less than 10 basis points. We notice, however, that the fitted one-factor model consistently produces Bermudan swaption prices that are slightly higher than the two-factor model. In other words, it appears that the volatility effect ii) is slightly larger than the direct correlation effect i). In any case, the price differences reported above are many orders of magnitude below those reported in LSS (1999), and justify the common practice of using simple one-factor models in practical pricing of standard Bermudan swaptions. To check whether the conclusion above is consistent and not a byproduct of the slight non-stationarity in the parameters of the fitted one-factor model, we now perform Test II. Specifically, we choose constant parameters of the one-factor model that are broadly consistent with the US market of December 1999 (see Table B.4 in Appendix B), and again assume a flat initial curve of 6%, continuously compounded. We then calibrate a two-factor model to the cap and European swaption prices of the one-factor model. In this calibration, we fix the quantities ( ,2 1 , 2 / 1 ) so that the two-factor model hits the same correlation structure as in Test I, but we allow the two-factor model to have time-dependent factor volatilities. The resulting fitted parameters are given in Table B.4 and the quality of the fit in terms of prices of caps and European swaptions is reported in Table B.5. As in Test I, we see a tendency of the one-factor model to have higher future term volatility than the two-factor model. The difference between the two models is shown below in Figure 3, which is almost indistinguishable from Figure 2.

13

Figure 3: Difference in Term Volatilities of Y (t , T ,10) for 1- and 2-factor models in Test II
0.001

t=5
0.0008

t=4 t=3 t=2 t=1 t=0

1-fac. Vol - 2-fac. Vol

0.0006

0.0004

0.0002

0 1 -0.0002 2 3 4 5 6 7 8 9

The Bermudan swaption prices of Test II are given in Tables 4 and 5 below. Table 4: 10-year Bermudan Receiver Swaption Prices in Test II
Ts = 1 1-fac. 2-fac. 96.5 93.3 213.2 207.1 432.4 424.1 793.0 786.5 1296.2 1293.2 Ts = 2 1-fac. 2-fac. 96.2 93.1 210.3 204.2 415.1 406.3 730.8 722.6 1151.9 1146.0 Ts = 3 1-fac. 2-fac. 94.2 91.3 200.4 195.1 382.0 374.4 650.4 642.8 1000.3 994.0 Ts = 4 1-fac. 2-fac. 89.4 87.0 183.9 179.7 338.6 332.9 561.1 555.1 846.4 841.2 Ts = 5 1-fac. 2-fac. 81.5 79.8 161.9 159.2 289.3 285.8 468.8 464.8 696.1 692.7

c 4% 5% 6% 7% 8%

All prices in basis points.

Table 5: 10-year Bermudan Payer Swaption Prices in Test II


Ts = 1 Ts = 2 Ts = 3 Ts = 4 Ts = 5

c 4% 5% 6% 7% 8%

1-fac. 1402.4 886.2 503.8 261.2 125.1

2-fac. 1399.9 879.4 494.9 253.7 120.6

1-fac. 1241.6 810.5 479.1 255.7 124.3

2-fac. 1236.0 802.1 469.7 248.3 119.9

1-fac. 1075.0 717.2 436.5 240.8 120.4

2-fac. 1069.0 709.3 428.5 234.3 116.3

1-fac. 907.0 615.0 383.4 217.7 112.2

2-fac. 902.02 609.0 377.2 212.8 109.0

1-fac. 742.9 510.3 324.2 188.9 100.2

2-fac. 739.7 506.5 320.4 185.7 98.2

All prices in basis points.

14

The differences in Bermudan prices of the one- and two-factor models of Test II are similar to those of Test I, i.e. small and with the one-factor model producing slightly higher (between 2 and 10 basis points) Bermudan swaption prices than the two-factor model. It thus appears safe to conclude that the observed effects are robust and not tied to the small non-stationarities induced by the fitting procedure. Finally, we point out that while the conclusions reached above are specific to the examples chosen, the magnitude and direction of the factor effect on standard Bermudan swaptions are, in our experience, typical for a large range of realistic market conditions.

5. Bermudan Swaption Pricing in a Log-Normal Libor Market Model.


The Gaussian model in Section 4 is convenient in the sense that Bermudan swaptions can be priced bias-free in a finite difference grid. The model, however, suffers from certain short-comings when it comes to real-life application. First, the assumed Gaussian distribution of interest rates does not correspond well to empirical data, and can result in negative interest rates. Second, the limited number of free parameters makes accurate calibration to a large set of target prices difficult7. In this section we will turn to the so-called Libor Market (LM) model which addresses both these issues. The added realism of this model class, however, comes at the price of making finite difference methods impractical; instead we here rely on the approximate method in Andersen (2000) to compute the early exercise premium of Bermudan swaptions. 5.1. The Log-Normal LM Model. Let us first define a right-continuous mapping function n(t ) on the maturity structure in Section 2 as Tn ( t ) 1 t < Tn ( t ) The basic model primitives of the LM model is the set of discrete forward rates Fk (t ) , k = 1,2,..., K (see Section 2). Assuming log-normal dynamics8, the no-arbitrage evolution of forward rates can be represented by the following set of stochastic differential equations, k = n(t ),..., K :
dFk (t ) = Fk (t )k (t ) k (t )dt + dW (t ) .

(9)

In (9), k (t ) is a bounded m-dimensional deterministic function, and W (t ) is a mdimensional Brownian motion. The term k (t ) depends on the probability measure used and is typically a function of all forward rates F j (t ) , j = n(t ),..., k . For instance, if we
7

Using a large set of calibration instruments is, for example, necessary if one wishes to apply a single, calibrated model to many kinds of exotic interest rate derivatives. Models with few free parameters are less convenient and normally require that the user make the calibration depend on the derivative to be priced. Also, to accurately hit the prices of more non-core instruments than just caps and floors, more free parameters are required. 8 Extensions to these dynamics to allow for volatility skews and smiles are discussed in Andersen and Andreasen (1998).

15

use the probability measure (the spot measure) induced by the strategy of "rolling over" an initial investment of $1 at each time in the maturity structure, it can be shown that (Jamshidian 1997, Brace et al 1997, Miltersen et al 1997)

k (t ) =

j =n(t )

F j (t )j (t ) j 1 + Fj (t ) j

(9) defines a system of up to K Markov state variables, where it is not unusual to have K > 40 . Under the dynamics (9), caps can be priced in closed form, and excellent closedform approximations for European swaptions exist (see Andersen and Andreasen 1998 for details). Valuation of other derivatives generally require Monte Carlo simulation due to the very high number of state variables; algorithms for discretizing (11) for Monte Carlo simulation can be found in Andersen and Andreasen (1998). 5.2. Test Description. As in Section 4, we first pick a reasonably parameterized, time-stationary twofactor LM model and proceed to fit a one-factor model to the cap and swaption prices generated by the chosen two-factor model. Appendix C contains a brief discussion of how to define the dynamics of a two-factor model by specifying the total volatility || k || as well as the correlation between the spot rate and various points on the forward curve. Calibration of a one-factor model is then a matter of determining a scalar volatility function s: + + such that k (t ) = s(t , Tk t ) will approximate the two-factor dynamics as well as possible. Notice that s is a function of calendar time, i.e. we allow non-stationarities in the one-factor model, just as we did in Test I in Section 4. While parameterization of the volatility function s in a few parameters is an option, we here wish to be a bit more ambitious, and instead attempt to fit s on nodes in a pre-specified grid. Writing s = s(t , ) , we define a grid {ti , j }i =1,..., Nt ; j =1,..., N and a corresponding matrix S, given by

Sij = s(ti , j ) .
Assuming that values of s outside of the grid-points can be constructed by interpolation, knowledge of the matrix S completely specifies the dynamics of the one-factor model. We can then calibrate the one-factor model by repeatedly perturbing the elements of S until the prices of caps and swaptions in the one-factor model correspond to those generated by the two-factor model. This matrix-based calibration procedure has, potentially, many more free parameters (the elements of the matrix) than was the case for the calibration used in Section 4. As a consequence, we are able to perform a larger calibration than in the previous section; specifically, we can include swaptions on noncore rates in the calibration9. In the calibration used here, S is set to have N t = 8 rows and N = 8 columns and covers the region [0,10] [0,10]; see Table D.3 in Appendix D. As calibration
9

The large number of parameters also implies a risk of overfitting the model. To avoid this problem, one normally regularizes the optimization problem by adding terms penalizing lack of smoothness of S.

16

instruments, we pick the same caps as in Section 4, but now extend the set of swaptions to include all 49 positive pairs (Ts , Te Ts ) contained in the grid of S (see Table D.5 in Appendix D). Pricing of all instruments10 were performed using a log-Euler Monte Carlo scheme (see Andersen and Andreasen 1998 for details) with time-discretization equal to the tenor structure spacing (0.5 in our test). All numbers in Appendix D were generated from 500,000 simulation trials with antithetic sampling (1,000,000 total paths). Upon calibration of the one-factor model, we can proceed to price Bermudan swaptions. As in Section 4, we start out by comparing one- and two-factor prices of Bermudan swaptions with a final swap maturity of Te = 10 years. Pricing of the Bermudan swaptions were done using the algorithm in Andersen (2000). Specifically, we used an exercise strategy based on the spread between intrinsic value and the most expensive outstanding European swaption (Strategy III in Andersen, 2000). The optimal exercise barrier was parameterized as a piece-wise linear function with 6 kink-points; the location of the exercise barrier was performed in a pre-simulation with 15,000 paths and antithetic sampling. The subsequent Monte Carlo pricing of the Bermudan swaption used 250,000 trials with antithetic sampling. The test described above can, as in Section 4, be performed in reverse with calibration of a two-factor model to a time-stationary one-factor model. As one would expect from the results in Section 4, the results of this reverse test are consistent with the test discussed above, and have consequently been omitted. Instead, we attempt to investigate further the consequences of following a one-factor exercise strategy in a twofactor world. Also, we spend some time comparing one- and two-factor prices of constant tenor Bermudan swaption (see Section 2 for definition of this instrument) with a constant swap tenor of 10 years. We notice that the grid-based calibration outlined above spans a 20-year time-horizon and as such is large enough to cover this instrument. 5.3. Test Results. As in Section 4, we set interest rates to be flat at 6%, continuously compounded. With =0.5, this implies that Fi (0) = 0.06091, for all i. The parameters used to specify the two-factor LM model is contained in Table D.1, based on the functional forms suggested in Appendix C. The correlation structure implied by the chosen volatilities is listed in Table D.2. The results of fitting a one-factor volatility matrix S to the cap and swaption prices implied by the two-factor model are summarized in Table D.3; the resulting cap and swaption prices are displayed in Tables D.4 and D.5. We notice that the fit is extremely accurate, with all cap prices being within 1/10 basis points of each other, and all but one swaption price being within 1 basis point. The graph below lists the instantaneous volatility of various points on the forward curve, at two different times t (taken directly from Table D.3):

10

In the actual calibration of the one-factor model, computational considerations required us to rely heavily on the closed-form cap and swaption formulas in Andersen and Andreasen (1998). The prices reported in Appendix C were, however, all generated by Monte Carlo simulation.

17

Figure 4: Volatility of Fi (t ) as a function of Ti t


0.2100 0.2000 0.1900
1-factor, t = 0 1-factor, t = 6 2-factor, t = 0 & t = 6

Volatility

0.1800 0.1700 0.1600 0.1500 0.1400 0 2 4

Ti-t

10

Qualitatively, Figure 4 is similar to Figure 1: the one-factor forward volatilities increase slightly with calendar time t, and exhibit a steeper term structure than the two-factor volatilities. In Tables 6 and 7, we list the results of pricing regular Bermudan swaption with 10 year final maturity.

Table 6: 10-year Bermudan Receiver Swaption Prices in LM Model


Ts = 1 1-fac. 2-fac.
66.9 (0.2) 194.3 (0.3) 435.1 (0.3) 812.6 (0.4) 1317.0 (0.4) 65.8 (0.2) 192.6 (0.3) 431.6 (0.3) 808.0 (0.4) 1312.7 (0.4)

c 4% 5% 6% 7% 8%

Ts = 2 1-fac. 2-fac.
66.8 (0.2) 193.5 (0.3) 422.7 (0.3) 760.6 (0.4) 1187.5 (0.4) 65.8 (0.2) 191.3 (0.3) 419.6 (0.3) 756.1 (0.4) 1183.7 (0.4)

Ts = 3 1-fac. 2-fac.
66.2 (0.2) 186.7 (0.3) 393.8 (0.3) 685.6 (0.4) 1043.5 (0.3) 65.3 (0.2) 185.7 (0.3) 392.6 (0.4) 684.0 (0.4) 1042.6 (0.3)

Ts = 4 1-fac. 2-fac.
64.2 (0.2) 174.4 (0.3) 355.1 (0.3) 600.2 (0.4) 895.1 (0.3) 63.3 (0.2) 173.0 (0.3) 353.0 (0.3) 597.9 (0.4) 892.9 (0.3)

Ts = 5 1-fac. 2-fac.
60.0 (0.2) 156.1 (0.3) 307.1 (0.3) 506.5 (0.3) 743.8 (0.3) 59.2 (0.2) 155.0 (0.3) 305.8 (0.3) 505.1 (0.3) 742.3 (0.3)

All prices in basis points. Numbers in parentheses denote standard deviations (250,000 antithetic paths).

18

Table 7: 10-year Bermudan Payer Swaption Prices in LM Model


Ts = 1 Ts = 2 Ts = 3 Ts = 4 Ts = 5

c 4% 5% 6% 7% 8%

1-fac.
1365.0 (0.2) 844.9 (0.5) 505.2 (0.5) 307.7 (0.4) 192.2 (0.4)

2-fac.
1363.2 (0.2) 840.5 (0.5) 501.2 (0.5) 304.0 (0.5) 190.1 (0.4)

1-fac.
1198.1 (0.3) 777.8 (0.5) 486.0 (0.5) 303.2 (0.5) 191.4 (0.4)

2-fac.
1196.2 (0.3) 773.9 (0.5) 482.5 (0.5) 300.0 (0.5) 189.3 (0.4)

1-fac.
1032.8 (0.3) 691.1 (0.5) 447.8 (0.5) 287.7 (0.5) 185.2 (0.4)

2-fac.
1032.7 (0.3) 689.5 (0.5) 446.8 (0.5) 286.2 (0.5) 184.4 (0.4)

1-fac.
872.2 (0.3) 598.1 (0.5) 399.3 (0.5) 263.8 (0.5) 174.1 (0.4)

2-fac.
870.8 (0.3) 595.6 (0.5) 397.0 (0.5) 261.7 (0.5) 172.6 (0.4)

1-fac.
715.1 (0.3) 499.8 (0.4) 341.9 (0.4) 231.3 (0.4) 156.6 (0.4)

2-fac.
714.2 (0.3) 498.4 (0.4) 340.3 (0.5) 230.1 (0.4) 155.1 (0.4)

All prices in basis points. Numbers in parentheses denote standard deviations (250,000 antithetic paths).

As was the case for the Gaussian model in Section 4, the one- and two-factor Bermudan prices are almost identical, but with the one-factor prices being slightly higher than the two-factor prices. Interestingly, the differences between the one- and two-factor prices are somewhat smaller for the LM model (1-5 basis points) than for the Gaussian model in Section 4. Whether this effect is a consequence of the differences in model dynamics or the tighter calibration used in the LM tests is difficult to say11. In any case, our previous conclusions about the limited, and negative, sensitivity of Bermudan swaption to the number of factors appears to hold for the LM model as well. To summarize our experiments with both Gaussian and lognormal models, we conclude that a one-factor model that is calibrated closely to the European swaption and cap prices of a two-factor model will produce Bermudan swaption prices that are very close to those of the true two-factor model. This conclusion appears stable across models and strikes. Now consider the case where an investor lives in a two-factor world but exercises her Bermudan swaptions according to a one-factor model that remains in constant calibration to the market prices of caps and European swaptions. The Bermudan swaptions will in this case be exercised whenever the one-factor model shows a higher value of the swap starting today, than the Bermudan swaption that starts at the next exercise date. Since, as discussed above, the PV of the Bermudan swaption in the calibrated one-factor model can be expected to be virtually the same as the true twofactor model price, the investor will tend to exercise her options very close to the true optimal timing. Hence, we would expect the present value of the loss of using a calibrated one-factor model in a two-factor enviroment to be very small. To investigate the loss empirically, one would need to track a continuously re-calibrated one-factor model up to the point where the Bermudan swaption either expires or is exercised. The
11

The technique used to price swaptions in the LM model is known to provide only a lower bound on the Bermudan price, although this bound seems to be very tight (see Andersen 1999). As the magnitude of the low bias should be higher in a two-factor model than in a one-factor model (and hence should help widen the spread between one- and two-factor Bermudan prices), this effect is probably small in the examples used here. The least-squares method used in LSS (1999) to price Bermudan swaptions also provides a lower bound only.

19

average payout from this strategy (as simulated under the "true" two-factor dynamics) could then be compared against the true two-factor Bermudan price and would constitute a measure of the sub-optimality of the one-factor model. The need to maintain the calibration of the one-factor model over time makes such an exercise quite onerous. As a less demanding alternative, consider a simpler experiment where the exercise rule generated in a calibrated one-factor model at a particular point in time is assumed to hold unchanged over the remaining life of the option. This effectively assumes that the onefactor model dynamics is never recalibrated beyond the date where the exercise rule is frozen. Using the simple, frozen one-factor strategy in a two-factor world would give us a conservative estimate of the sub-optimality of the one-factor model. In Tables 8 and 9 below, we have performed this experiment on the one- and two-factor LM models used in Table 6 and 7. That is, we have taken the time 0 exercise rule of the one-factor LM model and used it unchanged to price Bermudan swaptions in the two-factor LM model. Table 8: 10-year Bermudan Receiver Swaption Prices in LM Model Model A: 2-factor model with 2-factor exercise boundary Model B: 2-factor model with 1-factor exercise boundary
Ts = 1 Ts = 2 Ts = 3 Ts = 4 Ts = 5

c 4% 5% 6% 7% 8%

A
65.8 (0.2) 192.6 (0.3) 431.6 (0.3) 808.0 (0.4) 1312.7 (0.4)

B
66.1 (0.2) 192.1 (0.3) 430.9 (0.3) 808.0 (0.4) 1311.2 (0.5)

A
65.8 (0.2) 191.3 (0.3) 419.6 (0.3) 756.1 (0.4) 1183.7 (0.4)

B
65.9 (0.2) 191.4 (0.3) 419.0 (0.4) 755.4 (0.4) 1183.4 (0.4)

A
65.3 (0.2) 185.7 (0.3) 392.6 (0.4) 684.0 (0.4) 1042.6 (0.3)

B
65.5 (0.2) 185.7 (0.3) 392.7 (0.4) 683.3 (0.4) 1041.4 (0.3)

A
63.3 (0.2) 173.0 (0.3) 353.0 (0.3) 597.9 (0.4) 892.9 (0.3)

B
63.3 (0.2) 172.9 (0.3) 353.1 (0.3) 598.0 (0.4) 893.3 (0.3)

A
59.2 (0.2) 155.0 (0.3) 305.8 (0.3) 505.1 (0.3) 742.3 (0.3)

B
59.2 (0.2) 155.0 (0.3) 306.0 (0.3) 505.2 (0.3) 742.1 (0.3)

All prices in basis points. Numbers in parentheses denote standard deviations (250,000 antithetic paths).

Table 9: 10-year Bermudan Payer Swaption Prices in LM Model Model A: 2-factor model with 2-factor exercise boundary Model B: 2-factor model with 1-factor exercise boundary
Ts = 1 Ts = 2 Ts = 3 Ts = 4 Ts = 5

1363.2 1363.3 1196.2 1195.6 1032.7 1032.0 870.8 871.2 714.2 (0.2) (0.2) (0.3) (0.3) (0.3) (0.4) (0.3) (0.3) (0.3) 840.5 839.9 773.9 773.5 689.5 689.9 595.6 596.1 498.4 5% (0.5) (0.5) (0.5) (0.5) (0.5) (0.5) (0.5) (0.5) (0.4) 501.2 500.6 482.5 482.5 446.8 447.2 397.0 398.0 340.3 6% (0.5) (0.5) (0.5) (0.5) (0.5) (0.5) (0.5) (0.5) (0.5) 304.0 303.5 300.0 299.9 286.2 286.8 261.7 262.1 230.1 7% (0.5) (0.5) (0.5) (0.5) (0.5) (0.5) (0.5) (0.5) (0.4) 190.1 190.3 189.3 189.6 184.4 184.4 172.6 172.9 155.1 8% (0.4) (0.4) (0.4) (0.4) (0.4) (0.4) (0.4) (0.4) (0.4) All prices in basis points. Numbers in parentheses denote standard deviations (250,000 antithetic paths).

c 4%

B
714.0 (0.3) 498.7 (0.4) 339.9 (0.5) 229.9 (0.4) 155.2 (0.4)

20

It is obvious that the loss in value from following the "naive" one-factor exercise rule is generally miniscule12 further strengthening our arguments that Bermudan swaptions are affected only mildly by the number of factors. We also note that the contents of Tables 8 and 9 are consistent with the observations in Andersen (1998) who concludes that most Bermudan swaptions are not very sensitive to the exact location of the exercise rule We should point out that the conclusions reached so far all apply to regular Bermudan swaptions only. It is not inconceivable that other types of Bermudan swaptions will react differently to decorrelation of rates. To demonstrate this, we now turn to the pricing of a constant tenor Bermudan swaption, where exercise of the option pays a 10year swap starting on the day of exercise (so, in the notation of Section 2, = 10 / = 20). We will assume that the option can be exercised between Ts and Te 1, where we fix Te 1 = 9.5, and let Ts vary between 1 and 5 years. Our pricing results are shown below: Table 10: 10-year Const. Tenor Bermudan Receiver Swaption Prices in LM Model
Ts = 1 1-fac. 2-fac. Ts = 2 1-fac. 2-fac. Ts = 3 1-fac. 2-fac. Ts = 4 1-fac. 2-fac. Ts = 5 1-fac. 2-fac.

120.8 121.2 120.8 120.9 120.8 120.9 120.2 120.8 120.0 120.0 (0.3) (0.3) (0.3) (0.3) (0.3) (0.3) (0.3) (0.3) (0.3) (0.3) 301.9 304.1 301.8 304.9 300.8 304.2 299.4 301.2 295.8 297.0 5% (0.4) (0.5) (0.5) (0.5) (0.4) (0.5) (0.5) (0.5) (0.5) (0.5) 590.4 593.6 589.1 590.6 583.6 587.0 574.3 577.3 561.5 562.7 6% (0.5) (0.5) (0.5) (0.6) (0.5) (0.6) (0.5) (0.6) (0.6) (0.6) 988.3 989.2 979.9 981.0 960.4 963.5 935.8 938.4 906.7 908.3 7% (0.5) (0.7) (0.6) (0.6) (0.6) (0.6) (0.6) (0.6) (0.6) (0.6) 8% 1495.3 1497.8 1462.7 1465.0 1416.8 1421.0 1367.1 1370.7 1313.7 1316.2 (0.7) (0.8) (0.7) (0.7) (0.6) (0.7) (0.6) (0.6) (0.6) (0.6) All prices in basis points. Numbers in parentheses denote standard deviations (250,000 antithetic paths).

c 4%

Table 11: 10-year Const. Tenor Bermudan Payer Swaption Prices in LM Model
Ts = 1 Ts = 2 Ts = 3 Ts = 4 Ts = 5

c 4% 5% 6% 7% 8%

1-fac.
1545.4 (0.7) 1064.7 (0.6) 727.9 (0.6) 499.3 (0.6) 344.4 (0.7)

2-fac.
1568.2 (0.8) 1083.6 (0.8) 739.4 (0.7) 506.1 (0.7) 346.5 (0.7)

1-fac.
1492.4 (0.7) 1047.1 (0.6) 724.7 (0.6) 498.2 (0.6) 344.4 (0.6)

2-fac.
1515.6 (0.8) 1065.7 (0.8) 734.8 (0.7) 504.7 (0.7) 346.8 (0.7)

1-fac.
1428.7 (0.7) 1016.1 (0.6) 711.1 (0.6) 492.5 (0.6) 342.0 (0.7)

2-fac.
1452.7 (0.8) 1035.1 (0.7) 723.3 (0.7) 500.7 (0.7) 344.4 (0.7)

1-fac.
1362.9 (0.6) 979.6 (0.6) 690.9 (0.6) 482.8 (0.7) 337.7 (0.7)

2-fac.
1383.9 (0.7) 994.7 (0.7) 700.2 (0.7) 489.1 (0.7) 338.8 (0.7)

1-fac.
1294.1 (0.6) 935.8 (0.6) 665.0 (0.6) 467.4 (0.7) 329.4 (0.7)

2-fac.
1311.2 (0.6) 946.9 (0.7) 671.7 (0.7) 472.1 (0.7) 330.1 (0.7)

All prices in basis points. Numbers in parentheses denote standard deviations (250,000 antithetic paths).
12

In fact manyof the price errors arising from using a one-factor exercise rule are statistically insignificant. The fact that a few prices actually increase when using a one-factor exercise rule is due to sampling errors.

21

As Tables 10 and 11 demonstrate, constant tenor Bermudan swaptions are typically priced higher in the two-factor LM model than in the one-factor model, particularly for payer swaptions. For constant tenor swaptions, the net effect of the correlation and volatility effects discussed earlier thus appears to favor the two-factor model. We notice that while the differences in the constant tenor Bermudan swaption prices generated by the one- and two-factor models are somewhat higher than was the case for regular Bermudan swaptions, overall the differences are still fairly modest

6. Conclusion
In this paper, we have analyzed the impact of correlation on the pricing of Bermudan swaptions. For such an analysis to be meaningful, it is important that alteration of the correlation structure does not materially change the underlying models ability to hit the prices of liquid, quoted instruments. In other words, we need to ensure that a change in correlation structure is followed by a re-calibration to market data. An important question is then: which market instruments should be included in the calibration? One view, exemplified by the methodology used in Longstaff, Schwarz, and Santa-Clara (1999), is that calibration to a limited set of core European swaptions (the European swaptions into which the Bermudan swaption can be exercised) is sufficient. While this approach might, at first glance, seem reasonable, this paper has shown that ignoring non-core swaptions such as those embedded in interest rate caps can have profound impact on model dynamics and will affect the prices of Bermudan swaptions. In general, one should approach with care the often-heard advice about trying to identify whether a particular exotic instrument is cap-like or swaption-like for the purpose of determining whether a model should be calibrated to either caps or swaptions. As this paper shows, one is normally best served by calibrating to both caps and swaptions. Our paper also highlights the importance of using models which imply a realistic evolution of the term structure of rate volatilities. In particular, we attempt to make volatilities in our models as time-stationary as possible, yet fit the market prices of liquid instruments to sufficient accuracy. Working in such a framework, and using a calibration that encompasses both caps and swaptions, we find that the prices of Bermudan swaptions are quite insensitive to the correlation structure. Indeed, by keeping the evolution of future interest rate volatilities under control (a consequence of calibrating to non-core rates) we actually find that the prices of regular Bermudan swaptions are typically (slightly) higher in a one-factor model than in a two-factor model. This conclusion appears to be quite robust with respect to both the moneys of the Bermudan swaption, as well as the interest rate model used. We also demonstrate that in a twofactor world, usage of an exercise strategy generated in a properly calibrated one-factor model typically leads to only a small loss in value, futher bolstering our claims about the limited sensitivity of Bermudan values to interest rate correlation. We point out that the conclusions reached for regular Bermudan swaptions are specific to this product and do not generally carry over to other types of contracts. For instance, we have demonstrated that the prices of constant tenor Bermudan swaptions are typically higher in a two-factor model than in a one-factor model.

22

References Andersen, L. (2000): A Simple Approach to the Pricing of Bermudan Swaptions in the Multi-Factor Libor Market Model, Journal of Computational Finance, 2, 3, 5-32. Andersen, L. and J. Andreasen (1998): Volatility Skews and Extensions of the Libor Market Model, working paper, General Re Financial Products, forthcoming in Applied Mathematical Finance. Black, F., E. Derman, and W. Toy (1990), A One-Factor Model of Interest Rates and its Application to Treasury Bond Options, Financial Analysts Journal, January-February, 33-39. Brace, A., M. Grate, and M. Musical (1997). "The Market Model of Interest Rate Dynamics," Mathematical Finance, 7, 127-155. Hull, J. and A. White (1990), Pricing Interest Rate Derivatives, The Review of Financial Studies, 3, 573-592. Hull, J. (1999), Options, Futures, and Other Derivatives, Fourth Edition, Prentice and Hall. Jamshidian, F. (1989): An Exact Bond Option Pricing Formula, Journal of Finance 44, 205-209. Jamshidian, F. (1997). "Libor and Swap Market Models and Measures," Finance and Stochastic, 1, 293-330. Longstaff, F., E. Schwarz, and E. Santa-Clara (1999): Throwing Away a Billion Dollars: The Cost of Suboptimal Exercise in the Swaptions Market, working paper, The Anderson School, UCLA. Miltersen, K., K. Sandmann, and D. Sondermann (1997). "Closed-Form Solutions for Term Structure Derivatives with Lognormal Interest Rates," Journal of Finance, 409430. Mitchell, A. R. and D. F. Griffiths (1980), The Finite Difference Method in Partial Differential Equations, John Wiley and Sons. Morton, K. W. and D. F. Mayers (1994), Numerical Solution of Partial Differential Equations, Cambridge University Press.

23

Appendix A. Technical Details of Gaussian Model Given the specification of the short rate process as in (7) and (8) the covariance structure of the instantaneous forward rates is given by the equations:
var df (t , T ) = 1 (t ) 2 e

2 1 ( u ) du
t

corr df (t , Ti ), df (t , Tj ) =

i
T

v (t , T , T )dt , v (t , Ti , Tj )

v (t , Ti , Ti )v (t , Tj , Tj )

where (t ) t i b2 ( u) 1 ( u ) gdu v t , Ti , Tj = 1 + 2 e + e 1 (t )

2 (t ) 2 t i b2 ( u) 1 ( u) gdu t j b2 ( u ) 1 ( u ) gdu e . 1 (t ) 2
T T

F z G H

zb
Tj t

2 ( u ) 1 ( u ) du

I J K

Assuming constant mean-reversion coefficients, the instantaneous variance of the forward zero-coupon yield, Y t , Ti , Te , Te > Ti , is given by

g b t , T , T g = [ (t ) b b t , T , T g + (t ) b b t , T , T g + 2 (t ) (t )b b t , T , T g (t )b b t , T , T g]/ b T T g
2 2 2 2 2 i e 1 1 i e 2 2 i e 1 1 i e 2 2 i e e i

where

bi t , Ti , Te =

e i b Ti t g e i b Te t g , i = 1,2 . i

The term volatility discussed in Figures 2 and 3 can be computed in closed form from the formula

v (t ; T , Te ) = (T t ) 1

z
t

(u, T , Te ) 2 du .

24

Appendix B Fitting Results of Gaussian Model Tests I and II Table B.1 shows the parameters for the Gaussian one- and two-factor models of Test I. The one-factor model has time-dependent volatility in order to be able to (approximately) match the cap and European swaption prices of the time-homogeneous two-factor model. Table B.1: Parameters for Test I 1-Factor Model NA 1 0.1574 2 NA 1 (t ) 2 (t ) 0.01596 NA 0.01676 NA 0.01669 NA 0.01709 NA 0.01695 NA 0.01708 NA 0.01715 NA 0.01726 NA 0.01786 NA 2-Factor Model -0.8330 1 0.1467 2 0.3037 1 (t ) 2 (t ) 0.02742 0.03457 0.02742 0.03457 0.02742 0.03457 0.02742 0.03457 0.02742 0.03457 0.02742 0.03457 0.02742 0.03457 0.02742 0.03457 0.02742 0.03457

t 1 2 3 4 5 6 7 8 9

Table B.2 shows the correlation structure of the instantaneous forward rates f(t,T) implied by the parameter choice of the two-factor model in Tests I and II. Table B.2: Correlation Structure of Gaussian Two-Factor Model
0 1.000 0.971 0.877 0.738 0.590 0.459 0.352 0.268 0.203 0.152 0.112 1 0.971 1.000 0.967 0.878 0.766 0.658 0.566 0.491 0.432 0.384 0.346 2 0.877 0.967 1.000 0.971 0.905 0.829 0.758 0.698 0.648 0.607 0.575 3 0.738 0.878 0.971 1.000 0.980 0.938 0.891 0.848 0.811 0.779 0.753 4 0.590 0.766 0.905 0.980 1.000 0.988 0.963 0.936 0.910 0.888 0.868 5 0.459 0.658 0.829 0.938 0.988 1.000 0.993 0.979 0.963 0.948 0.934 6 0.352 0.566 0.758 0.891 0.963 0.993 1.000 0.996 0.988 0.979 0.969 7 0.268 0.491 0.698 0.848 0.936 0.979 0.996 1.000 0.998 0.993 0.987 8 0.203 0.432 0.648 0.811 0.910 0.963 0.988 0.998 1.000 0.999 0.996 9 0.152 0.384 0.607 0.779 0.888 0.948 0.979 0.993 0.999 1.000 0.999 10 0.112 0.346 0.575 0.753 0.868 0.934 0.969 0.987 0.996 0.999 1.000

0 1 2 3 4 5 6 7 8 9 10

25

Table B.3. shows the European payer swaption and cap prices (in basis points) of the two-factor model and the calibrated one-factor model of Test I. All instruments are atthe-money and have terminal maturity of 10 years. For caps, the start date is here interpreted as the time the first rate-set takes place. Table B.3: European Style Interest Rate Option Prices of Test I Start Date (years) Product 1 Cap 2 Cap 3 Cap 4 Cap 5 Cap 6 Cap 7 Cap 8 Cap 9 Cap 1 Payer 2 Payer 3 Payer 4 Payer 5 Payer 6 Payer 7 Payer 8 Payer 9 Payer 1-factor price 630.3 572.7 503.5 429.5 353.6 278.4 205.1 134.1 65.7 223.7 269.1 278.6 267.5 242.9 208.5 166.7 118.1 62.9 2-factor price 634.8 572.9 501.9 427.7 352.6 278.2 205.2 134.3 65.7 223.6 269.0 278.4 267.5 243.1 208.8 166.8 118.2 62.8

Table B.4 shows the parameters for the Gaussian one- and two-factor models of Test II. The two-factor model has time-dependent volatility in order to be able to (approximately) match the cap and European swaption prices of the time-homogeneous one-factor model. The two-factor model is calibrated to the one-factor model in such a way that the correlation structure of the forward rates is the same as in Test I.

26

Table B.4: Parameters for Test II 1-Factor Model : NA 1 : 0.1574 2 : NA 1 (t ) 2 (t ) 0.01698 NA 0.01698 NA 0.01698 NA 0.01698 NA 0.01698 NA 0.01698 NA 0.01698 NA 0.01698 NA 0.01698 NA 2-Factor Model : -0.8330 1 : 0.1467 2 : 0.3037 1 (t ) 2 (t ) 0.02932 0.03696 0.02778 0.03502 0.02793 0.03522 0.02720 0.03429 0.02753 0.03471 0.02728 0.03439 0.02720 0.03429 0.02695 0.03397 0.02608 0.03288

t 1 2 3 4 5 6 7 8 9

Table B.5. shows the European payer swaption and cap prices (in basis points) of the one-factor model and the calibrated two-factor model of Test II. All instruments are atthe-money and have terminal maturity of 10 years. For caps, the start date is here interpreted as the time the first rate-set takes place. Table B.5: European Style Interest Rate Option Prices of Test II Start Date (years) Product 1 Cap 2 Cap 3 Cap 4 Cap 5 Cap 6 Cap 7 Cap 8 Cap 9 Cap 1 Payer 2 Payer 3 Payer 4 Payer 5 Payer 6 Payer 7 Payer 8 Payer 9 Payer 1-factor price 640.9 579.8 507.7 431.5 354.1 278.0 204.1 132.9 64.8 237.4 282.0 287.7 273.3 246.3 210.3 167.3 117.9 62.2 2-factor price 645.7 579.9 506.0 429.6 353.0 277.7 204.2 133.1 64.8 237.3 281.7 287.4 273.3 246.4 210.7 167.5 118.0 62.2

27

Appendix C Volatility Specification for two-factor LM Model Set i Ti t , and consider a two-factor time-stationary LM model where || i (t )|| = l ( i ) and i (t ) = (l1 ( i ) , l2 ( i )) T , for functions l1, l2 , and l. Let ( i , j ) denote the instantaneous correlation between Fi (t ) and F j (t ) . We then have

( i , j ) =

l1 ( i )l1 ( j ) + l2 ( i )l2 ( j ) l ( i )l ( j )

(C.1)

Assume now that the total volatility function l is known, but not the individual factor volatilities l1 and l2 . To separate out these functions, we need to specify correlation information. For this purpose, define g ( i ) ( i ,0) , from (C.1),
g ( i ) = l1 ( i )l1 (0) + l2 ( i )l2 (0) . l ( i )l (0)

(C.2)

Equation (C.2), along with the identity l ( i ) 2 = l1 ( i ) 2 + l2 ( i ) 2 , gives us the following equations for l1 and l2 :
l ( i ) l1 (0) g ( i ) l2 (0) 1 g ( i ) 2 , l (0) l ( i ) l2 ( i ) = l2 (0) g ( i ) + l1 (0) 1 g ( i ) 2 . l (0) l1 ( i ) =

e e

j j

(C.3a) (C.3b)

To eliminate l1 (0) and l2 (0) we can for instance, without loss of generality, set l2 (0) = l1 (0) whereby l1 (0) = l (0)(1 + 2 ) 1/ 2 and l2 (0) = l (0)(1 + 2 ) 1/ 2 . To summarize, a full specification of the factor volatility functions l1 and l2 can be accomplished by specifying a) the total volatility l( ; b) the correlation g( between ) ) the spot rate and all forward rates; and c) the ratio of factor volatilities of the spot rate. One useful specification of the functions l and g is

l ( ) = + (0 )e dl ; g ( ) = + (1 )e

d g

(C.4)

28

Appendix D Fitting Results of LM Model Calibration

Table D.1 shows the choice of parameters in the two-factor specification (C.4), in Appendix C. Table D.1: Two-Factor LM Model Parameters
0 dl dg

0.15 0.2 0.2 0 0.15 -0.5

Table D.2 shows the correlation function (Ti t , Tj t ) (see Appendix C) implied by the parameter choice above. Table D.2: Correlation Structure of Two-Factor LM Model
0 1.000 0.861 0.741 0.638 0.549 0.472 0.407 0.350 0.301 0.259 0.223 1 0.861 1.000 0.980 0.941 0.898 0.855 0.815 0.778 0.745 0.715 0.688 2 0.741 0.980 1.000 0.990 0.968 0.942 0.915 0.888 0.864 0.841 0.820 3 0.638 0.941 0.990 1.000 0.994 0.980 0.963 0.945 0.927 0.909 0.893 4 0.549 0.898 0.968 0.994 1.000 0.996 0.987 0.975 0.962 0.950 0.937 5 0.472 0.855 0.942 0.980 0.996 1.000 0.997 0.991 0.983 0.974 0.965 6 0.407 0.815 0.915 0.963 0.987 0.997 1.000 0.998 0.994 0.988 0.981 7 0.350 0.778 0.888 0.945 0.975 0.991 0.998 1.000 0.999 0.995 0.991 8 0.301 0.745 0.864 0.927 0.962 0.983 0.994 0.999 1.000 0.999 0.997 9 0.259 0.715 0.841 0.909 0.950 0.974 0.988 0.995 0.999 1.000 0.999 10 0.223 0.688 0.820 0.893 0.937 0.965 0.981 0.991 0.997 0.999 1.000

0 1 2 3 4 5 6 7 8 9 10

Table D.3 shows the one-factor grid volatility S, after calibration to the two-factor model defined above.

29

Table D.3: Calibrated Volatility Table for One-Factor LM Model


tB / 0 1 2 3 4 6 8 10 0 0.1999 0.2003 0.2010 0.2013 0.2013 0.2006 0.2001 0.2000 1 0.1903 0.1914 0.1934 0.1943 0.1946 0.1947 0.1944 0.1945 2 0.1816 0.1830 0.1854 0.1869 0.1876 0.1885 0.1888 0.1889 3 0.1739 0.1754 0.1781 0.1798 0.1806 0.1824 0.1831 0.1833 4 0.1674 0.1689 0.1715 0.1732 0.1742 0.1766 0.1774 0.1777 6 0.1574 0.1587 0.1610 0.1627 0.1638 0.1662 0.1668 0.1670 8 0.1519 0.1527 0.1542 0.1553 0.1560 0.1574 0.1578 0.1579 10 0.1503 0.1506 0.1512 0.1513 0.1514 0.1513 0.1511 0.1509

Table D.4 shows the prices of caps (in basis points) in the two-factor LM model, as well as the fitted one-factor LM model. All caps are at-the-money and have terminal maturity of 10 years. For caps, the start date is here interpreted as the time the first rate-set takes place. Numbers in parentheses denote sample standard deviation (from 500,000 antithetic Monte Carlo trials). Table D.4: Cap Prices Start Date (years) 1 2 3 4 5 6 7 8 9 1-factor price 590.1 (0.8) 543.5 (0.7) 485.7 (0.7) 421.4 (0.6) 353.3 (0.5) 283.2 (0.4) 211.9 (0.3) 140.6 (0.2) 69.9 (0.1) 2-factor price 590.1 (0.8) 543.5 (0.7) 485.6 (0.7) 421.4 (0.6) 353.3 (0.5) 283.0 (0.4) 211.8 (0.3) 140.6 (0.2) 69.9 (0.1)

Table D.5 shows the prices of European swaptions (in basis points) in the two-factor LM model, as well as the fitted one-factor LM model. All swaptions are at-the-money. Numbers in parentheses denote sample standard deviation (from 500,000 antithetic Monte Carlo trials). Row headers correspond to values of swaption maturity Ts ; column header correspond to swap tenors Ts Te .

30

Table D.5: European ATM Payer Swaption Prices 1-Factor Model:


1 1 2 3 4 6 8 10
41.6 (0.1) 54.4 (0.1) 61.5 (0.1) 65.8 (0.1) 69.5 (0.1) 69.4 (0.1) 67.4 (0.1)

2
79.1 (0.1) 103.5 (0.1) 117.2 (0.2) 125.6 (0.2) 132.9 (0.2) 133.0 (0.2) 129.4 (0.2)

3
112.8 (0.1) 147.8 (0.2) 167.7 (0.2) 179.9 (0.2) 190.9 (0.2) 191.4 (0.2) 186.6 (0.2)

4
143.4 (0.2) 188.1 (0.2) 213.6 (0.3) 229.5 (0.3) 244.0 (0.3) 245.2 (0.3) 239.4 (0.3)

6
196.8 (0.2) 258.7 (0.3) 294.4 (0.4) 316.9 (0.4) 338.3 (0.4) 340.9 (0.4) 333.5 (0.3)

8
242.2 (0.3) 318.9 (0.4) 363.5 (0.4) 391.8 (0.5) 419.4 (0.5) 423.5 (0.4) 415.0 (0.4)

10
281.6 (0.3) 371.3 (0.4) 423.8 (0.5) 457.3 (0.5) 490.2 (0.5) 495.7 (0.5) 486.2 (0.4)

2-Factor Model:
1 1 2 3 4 6 8 10 2 3 4
143.1 (0.2) 188.0 (0.2) 214.0 (0.3) 229.8 (0.3) 244.2 (0.3) 245.2 (0.3) 239.4 (0.3)

6
196.3 (0.2) 258.6 (0.3) 295.2 (0.4) 317.6 (0.4) 338.6 (0.4) 341.1 (0.4) 333.7 (0.3)

8
241.3 (0.3) 318.5 (0.4) 364.3 (0.4) 392.7 (0.5) 419.8 (0.5) 423.8 (0.4) 415.2 (0.4)

10
279.9 (0.3) 370.0 (0.4) 423.9 (0.5) 457.6 (0.5) 490.1 (0.5) 495.6 (0.5) 486.2 (0.4)

41.7 (0.1) 79.0 (0.1) 112.6 (0.1) 54.4 (0.1) 103.4 (0.1) 147.7 (0.2) 61.6 (0.1) 117.3 (0.2) 167.9 (0.2) 65.9 (0.1) 125.6 (0.2) 180.0 (0.2) 69.5 (0.1) 132.9 (0.2) 190.9 (0.2) 69.4 (0.1) 133.0 (0.2) 191.4 (0.2) 67.5 (0.1) 129.5 (0.2) 186.6 (0.2) Rows: swaption maturity Te . Columns: swap tenor Te Ts

31

Potrebbero piacerti anche