EVALUATING AND HEDGING EXOTIC SWAP INSTRUMENTS VIA LGM

PATRICK S. HAGAN

Abstract. Here we use the one factor LGM model to price the standard IR exotic deals: callable swaps (including amortizing swaps), callable inverse oaters, callable super- oaters, callable range notes, autocaps and revolvers, and range notes. We lay out the complete pricing of these deals: how to represent the deals, how to select the calibration instruments, how to determine the appropriate calibration strategies and algorithms, the dierent deal evaluation algorithms for each deal type, and the usage of adjusters to obtain the best possible prices and hedges. We then extend this analysis to the world of bonds. By incorporating a second, credit spread factor into the LGM model, we extend our valuation/hedging analysis to bonds with embedded options

Key words. LGM, interest rate models, calibration, exotic options

(Part head:)Introduction In this paper we specify how to price and hedge common exotic interest rate deals: callable swaps (including amortizing swaps), callable inverse oaters, callable capped- oaters and super-oaters, callable range notes, autocaps, revolvers, and captions. Here we introduce our notation and the mathematics of swaps and vanilla options. In Part II, we work out best practices for using the one factor LGM mode for pricing and hedging: We introduce the LGM model, we determine how to select eective calibration instruments, we derive ecient calibration strategies and algorithms, we then discuss evaluation methodologies and determine ecient algorithms for evaluating the prices of exotic instruments under the calibrated model. In each of the subsequent sections, we treat a dierent exotic: callable swaps (with Bermudan and American), callable amortizing swaps, callable inverse oaters, callable capped-oaters and super-oaters, callable range notes and accrual, autocaps, and revolvers. Throughout these sections we use best practices in choosing the calibration instruments, calibrating the model, evaluating the deal, and using internal adjustors to obtain risks to the proper instrumentsand clean up the prices. In Part III, we extend the model to include a second factor for credit. The extended LGM model is used to price bond, focussing on bonds with embedded options.

1. Discount factors, zeros, and FRAs. Suppose at date t, one agrees to loan out $1 at date T , and get repaid the next day:

(1.1a) (1.1b)

1 1 + f (t, T )T

paid at T, received at T + T.

By denition, the fair interest rate to charge is

(1.1c)

f (t, T ) = instantaneous forward rate for date T as seen at date t.

1

f(t,T)

Tst

Tend

Instantaneous rate for date T as seen at date t Now suppose at date t one agrees to loan out $1 on Tst , with the money repaid on Tend . Economically this is equivalent to loaning out $1 on Tst , getting repaid $1 plus interest the next day, re-loaning out the $1 plus interest, getting repaid $1 plus interest plus interest on the interest, .... Clearly, if one agrees at date t to loan out (1.2a) the agreement should specify getting repaid (1.2b) e
Tend
Tst

paid at Tst ,

f (t,T 0 )dT 0

received at Tend

for the deal to be fair. Alternatively, we can re-phrase this as (1.3a) (1.3b) e
Tend
Tst

f (t,T 0 )dT 0

paid at Tst , received at Tend .

This type of single payment deal is equivalent to a FRA (forward rate agreement). Suppose we imagine that we are at date t, and we ask how much would I need to pay immediately to receive $1 at date T . Clearly the fair amount is (1.4a) (1.4b) By denition, (1.5) (t; T ) = e Z
T
t

Tend
t

f (t,T 0 )dT 0

paid at t, received at Tend .

f (t,T 0 )dT 0

= value at t of $1 paid at T,

is the value of a zero coupon bond for maturity T on date t. Today is always t = 0 in our notation. Discount factors are todays values of the zero coupon bonds: (1.6a) (0; T ) = e D(T ) = Z
2
T
0

f0 (T 0 )dT 0

where (1.6b) f0 (T ) = f (0, T ) = todays instantaneous forward rate curve.

The discount factors and todays forward curve are not random. We can always get their values by stripping (t; T ) will uctuate until now catches up to date the current swap curve. On the other hand, f (t, T ) and Z t. This is why we use dierent notation for discount factors and zero coupon bonds in general. 2. Swaps. 2.1. Fixed leg. Consider a swap with start date t0 , xed leg pay dates t1 , t2 , . . . , tn , and xed rate Rf ix . The xed leg makes the payments (2.1a) (2.1b) where (2.1c) i = cvg(ti1 , ti , ) i Rf ix 1 + n Rf ix paid at ti paid at tn , for i = 1, 2, . . . , n 1,

is the coverage (day count fraction) for interval i computed according to the appropriate day count basis . On any given day t, these payments have the value (2.2) f ix (t) = Rf ix V
n X i=1

(t; tn ). (t; ti ) + Z i Z

The mechanics of the swap market is covered in Appendix A. There we explain how to correctly construct date sequences and compute coverages.for the major currencies. 2.2. Floating leg. Let us now consider the swaps oating leg. Floating legs usually have a dierent frequency than the xed legs, so let this legs start and pay dates be (2.3) The oating leg pays (2.4a) (2.4b) where (2.4c) ) j = cvg( j 1 , j , j rj 1+ m rm paid at j for j = 1, 2, . . . , m 1, paid at m = tn , t0 = 0 , 1 , . . . , m = tn

. Here rj is generally is the coverage for interval j computed according to the appropriate day count basis the Libor or Euribor oating rate for interval j . This rate is set on the xing date ; for most oating legs, the xing date is two London business days before the interval starts on j 1 . j rj The oating rate represents a deal in which one invests 1 unit of the ccy at j 1 and receives 1 + ix units back at j . In principle, the value of the two payments must be the same at the xing date f , so j (2.5a) ( f ix ; j 1 ) = (1 + (t; j ). Z j rj ) Z j

In theory, then, the value of the interest payment j rj would be (2.5b)

ix f ix f ix Vjtheor ( f j ) = Z ( j ; j 1 ) Z ( j ; j )

ix on the xing date f j . If the value of the oating rate payment is the dierence of two freely tradeable securities (the zero coupon bonds) at the xing time, then the value must equal this dierence for all earlier times as well. So in principle, the value of the j th oating interest rate payment is

(2.6)

(t; j 1 ) Z (t; j ) Vjtheor (t) = Z

ix for t < f j

true for any date t, at least until the rate is xed. At any date t, the forward fair or true rate rj (t) is dened so that the value of the interest payment exactly equals the theoretical value:

(2.7a) So (2.7b)

true (t; j ) = theoretical value of interest rate payment = Z (t; j 1 ) Z (t; j ). j rj (t)Z

true rj (t) =

(t; j 1 ) Z (t; j ) Z . j Z (t; j )

In practice, oating rates are not set at the fair rate, they are set at the fair rate plus a small oset sj , the forward basis spread, due to credit considerations and supply and demand. The value of the (forward) basis spread depends on which index is used for the oating rate (3m USD Libor, 1m FedFunds, 6m Euribor, etc.), and on the starting date. The value of the oating rate payment paid at j is (2.8a) j (t) = Z (t; j 1 ) Z (t; j ) + (t; j ). V j sj Z

By denition, the forward rate for the oating rate is dened by, (2.8b) so (2.8c)
f wd true = rj (t) + sj = rj f wd (t; j ) = value of interest rate payment, (t)Z j rj

(t; j 1 ) Z (t; j ) Z + sj . (t; j ) jZ

Basis spread curves are obtained by stripping basis swaps. One can show that forward basis spreads are Martingales in the appropriate forward measures. Since they are very small, usually just 1 2 bps, and since they seldom vary, one always assumes they are constant. That is, one assumes that the gamma of the forward spread is inconsequential. Summing these payments together, the value of the oating leg is (2.9) f lt (t) = Z (t; t0 ) + V
m X j =1

(t; j ). j sj Z

This is true regardless of the model used. The value of the receiver swap (receive the xed leg, pay the oating leg) is (2.10a) rec (t) = Rf ix V
n X i=1

(t; tn ) Z (t; t0 ) (t; ti ) + Z i Z

m X j =1

(t; j ). j sj Z

The value of the payer swap (pay the xed leg, receive the oating leg) is (2.10b) rec (t) = Z (t; t0 ) + pay (t) = V V
m X j =1

(t; j ) Rf ix j sj Z
4

n X i=1

(t; tn ). (t; ti ) Z i Z

2.3. Handling the basis spread. Basis spreads are a nuisance. They are large enough that one cannot neglect them entirely (except for USD 3m Libor), and small enough that they are nearly irrelevent. One way of handling them is to treat them as another, very small, xed leg. There is really nothing wrong with that approach, although one usually has twice as many xed leg pay dates. We use a second, common approach in which each intervals xed rate is adjusted to account for the value of the basis spreads. I.e, if the basis spread is 0.625bps in an interval, then we subtract 0.625bps from the xed legs rate instead of adding it to the oating leg. More precisely, todays value of the swap is, (2.11) which is the same as (2.12a) rec (0) = V
n X i=1

rec (0) = V

n X i=1

i Rf ix D(ti ) + D(tn ) D(t0 )

m X j =1

j sj D( j ).

i Rf ix Si D(ti ) + D(tn ) D(t0 )./

Here Si is the basis spread expressed with the same frequency and day count basis as the xed leg. If the oating leg frequency is the same or higher than the xed leg frequency, then P j sj D( j ) (2.12b) Si =
j Ii

i D(ti )

where j Ii represents the oating leg intervals which are part of the ith xed leg interval. That is, th the oating leg intervals whose theoretical dates th xed leg theoretical interval j are contained in the i th th th ti1 < j ti . If the xed leg frequency is shorter than the oating leg frequency (this is rare), then the same Si is used for all xed leg intervals forming part of each oating leg interval. So, (2.12c) j sj D( j ) Si = P , i D(ti )
iIj

th th where, i Ij represents the xed leg intervals i with th j 1 < ti j . Either way, we approximate the swap values as

(2.13a)

rec (t) = V

n X i=1

(t; tn ) Z (t; t0 ). (t; ti ) + Z i Rf ix Si Z

n X i=1

(2.13b)

rec (t) = Z (t; t0 ) Z (t; tn ) pay (t) = V V

(t; ti ). i Rf ix Si Z

for all dates t, where the strike Rf ix and eective spread Si are known constants. We are neglecting any evolution of the basis spreads and any minor dierences due to the dierences between the legs day count bases and frequencies. We will use this approach throughout. Computationally, it would be just as easy to modify the code to add another xed leg, but this would make the formulas messier and debugging more dicult. 2.4. Swap rate and level. At any time t, the swap rate Rsw (t) is dened to be the break even rate, the value of Rf ix which would make the swap worth zero. Clearly, (t, x; tn ) + Pn i Si Z (t; ti ) (t, x; t0 ) Z Z sw i=1 (2.14a) R (t) = , L(t)
5

where the level (also known as the PV01, the DV01, or the annuity ) is (2.14b) L(t) =
n X i=1

(t; ti ). i Z

In particular, todays swap rate and level are (2.15a) (2.15b) and the swap values are (2.15c) (2.15d) R0 =

We can re-write the swap values in terms of the swap rate and level as rec (t) = Rf ix Rsw (t) L(t), pay (t) = Rsw (t) Rf ix L(t). (2.14c) V V Pn D(t0 ) D(tn ) + i=1 i Si D(ti ) , L0 n X L0 = i D(ti ),
i=1

is the value of the receiver swaption on the exercise date. Swaption prices are almost always quoted in terms of Blacks model.To introduce this model, suppose we choose the level L(t) as our numeraire. (It is just the sum of a bunch zero coupon bonds, and hence is a tradable instrument). There exists a probability measure in which the value of all tradeable instruments (including the swaption) divided by the numeraire is a Martingale. So ) ( opt rec V ( T ) opt rec (3.2a) V (t) = L(t) E for any T > t Ft L(T ) If we evaluate the expected value at T = tex , we see that n o + opt rec (t) = L(t) E Rf ix Rsw (tex ) Ft (3.2b) V Moreover, the swap rate (3.3a) R
sw

3. Swaptions. A swaption is a European option on a swap. Consider a receiver swaption with notication date tex . If one exercises on this date, one obtains the receiver swap. Clearly + opt ex rec (3.1) V (t ) = Rf ix Rsw (tex ) L(tex )

rec (t) = Rf ix R0 L0 , V pay (t) = R0 Rf ix L0 . V

(t; ti ) (t, x; tn ) + Pn i Si Z (t, x; t0 ) Z Z i=1 (t) = L(t)

is clearly a tradeable market instrument (a bunch of zero coupon bonds) divided by the numeraire. So the swap rate is also a Martingale. By the Martingale representation model, then, we conclude that (3.3b) dRsw = A(t, )dW,

where dW is Brownian motion, and A(t, ) is some measureable coecient. At this point we know that o n + opt rec (t) = L(t) E Rf ix Rsw (tex ) Ft (3.4a) V
6

where (3.4b) dRsw = A(t, )dW,

for some A(t, ). Fundamental theory can take us no further. We now have to model A(t, ). Black proposed that A(t, ) = Rsw , so that the swap rate is log normal: dRsw = Rsw dW. o n + Finding the expected value E Rf ix Rsw (tex ) Rsw (t) under this model yields Blacks formula, (3.5) opt rec (t) = Rf ix N(d1 ) Rsw (t)N(d2 ) L(t), V d1,2 = 2 (tex t) log Rf ix /Rsw (t) 1 ex 2 t t

(3.6a) with

(3.6b)

Todays market price of the swaption is (3.7a) where (3.7b) and (3.7c) P D0 Dn + n i=1 i Si Di R = , L0
0

0 mkt 0 0 rec V (0) = Rf ix N(d0 1 ) R N(d2 ) L , d0 1,2 =

2 log Rf ix /R0 1 2 tex tex

L0 =

n X i=1

i Di

Here Di = D(ti ) are todays discount factors. A payer swaption is a European option to pay the xed leg and receiver the oating leg. The value of the payer swaption is obtained by reversing Rf ix and R0 in the above formulas: 0 mkt f ix pay V (3.8a) (0) = R0 N(d0 N(d0 2) R 1) L mkt rec (3.8b) (0) Rf ix R0 L0 . =V

If one analyzes Blacks formula, one discovers that the receiver and payer swaption values are both increasing functions of the volatility . Instead of quoting swaption prices in terms of dollar values, one can just as well quote the price in terms of the value of that needs be inserted into Blacks formula to obtain the market price. This value of the volatility is known as the implied volatility. 3.1. Caplets and oorlets. Consider a oorlet for the interval 0 to 1 . The oating rate r for the interval is set on the xing date tex two (London) business days before the interval starts at 0 , and the oorlet pays the dierence between the strike (xed rate) and the oating rate at the end of its period, provided this dierence is positive: (3.9a) (Rf ix r)+ paid at 1 .

Here is the coverage (day count fraction) of the interval 0 to 1 . As above, the value of the oating rate payment is (3.9b) (tex ; 0 ) Z (tex ; 1 ) + (tex ; 1 ), Z s1 Z
7

on the xing date, where s1 is the (forward) basis spread for the interval. The oorlets payo is (3.9c) h i+ (tex ; 1 ) + Z (tex ; 1 ) Z (tex ; 0 ) . f loorlet (tex ) = (Rf ix s1 ) Z V

This is the same payo as a 1 period receiver swaption. Similarly, caplet payos are identical to the payos of 1 period payer swaptions. The analysis of caplets and oorlets parallels the analysis for swaptions exactly. We dene the forward or FRA rate as (3.10a) RF RA (t) = (t, x; 0 ) Z (t, x; 1 ) + (t; 1 ) Z s1 Z , Z (t; 1 )

where the forward FRA rate is a Martingale in this measure. Modeling this rate as log normal (3.10c) again yields Blacks formula, (3.11a) with (3.11b) (t; 1 ), f loorlet (t) = Rf ix N(d1 ) RF RA (t)N(d2 ) Z V d1,2 = 2 (tex t) log Rf ix /RF RA (t) 1 ex 2 t t dRF RA = RF RA dW.

(t; 1 ) as our numeraire. The value of the oorlet is and choose the zero coupon bond Z n o (t; 1 ) E Rf ix RF RA (tex ) + f loorlet (tex ) = Z (3.10b) V Ft ,

Todays market price of the oorlet is (3.12a) where (3.12b) f ix 0 0 fmkt V N(d0 D( 1 ), loorlet (0) = R 1 ) R N(d2 ) d0 1,2 =
2 log Rf ix /R0 1 2 tex , tex

and where todays forward FRA rate is (3.12c) R0 = D0 D1 + s1 D1 . D1

The value of the caplet is obtained by reversing Rf ix and R0 in the above formulas: (3.13a) (3.13b) mkt f ix caplet V (0) = R0 N(d0 N(d0 D ( 1 ) 2) R 1) mkt rec (0) Rf ix R0 D( 1 ). =V

Note that the caplet and oorlet values are special cases of the payer and receiver swaptions with n = 1. As before, the implied vol is the value of the volatility which makes the above formulas match the actual market values of the oorlet and caplet.
8

3.2. Digital caplets/oorlets. Digital caplets/oorlets are identical to a regular caplets/oorlets, except that the payo is 1 paid at the end date 1 if the oating rate ends up in the money: Z (tf ix ; 1 ) if r Rf ix dig (3.14a) , Vcaplet (0) = 0 if r < Rf ix 0 if r Rf ix Vfdig (3.14b) (0) = . loorlet Z (tf ix ; 1 ) if r < Rf ix Following the above line of reasoning shows that the value of these digitals is (3.15a) with (3.15b) Todays value of the digitals is (3.16a) with (3.16b) d=
2 ex log Rf ix /RF RA (t) + 1 2 t tex

(t; 1 ), dig (t) = N(d)Z V caplet

dig V f loorlet (t) = N(+d)Z (t; 1 )

d=

2 (tex t) log Rf ix /RF RA (t) 1 ex 2 t t dig V f loorlet (0) = N(+d)D ( 1 )

dig (0) = N(d)D( 1 ), V caplet

Again, the implied digital caplet vol is the value of which makes the above theoretical digital prices match their market prices. (Part head:)Pricing exotics via LGM 4. The LGM model. A modern interest rate model consists of three parts: a numeraire, a set of random evolution equations in the risk neutral world, and the Martingale pricing formula. The one factor It starts at 0 today and satises LGM model has a single state variable, X. (4.1) = (t)dW , dX (0) = 0. X

given by

This is the evolution under the risk neutral measure induced by the numeraire, which will be named shortly. (t) is Gaussian with the transition density Clearly X n o (T ) X + dX (t) = x (4.2a) p(t, x; T, X )dX = prob X < X X
2 1 1 e 2 (X x) / p(t, x; T, X ) = 2

Here (4.2b) ( ) = Z

2 (t0 )dt0 ,
0

= (T ) (t) =

2 (t0 )dt0 .

We choose the numeraire to be (t, x) = N 1 +H (t)x+ 1 H 2 (t) (t) 2 . e D(t)

9

 (0, 0) = 1. Note that the value of the numeraire is 1 today: N The last part of the model is the Martingale valuation formula. Suppose at time t the economy is in (t) = x. If V (t, x) is the value of any freely tradeable security, then V (t, x)/N (t, x) is a Martingale: state X ) ( (t) = x (t, x) = N (t, x) E V (T, X ) (4.3) V X (T, X ) N (t, x) Z V (T, X ) 1 2 N = for any T > t. e 2 (X x) / dX 2 N (T, X ) If the security throws o cash payments, then we would need to modify this formula appropriately. The LGM model can be written most simply in terms of the reduced prices (t, x) V (4.4a) V (t, x) . (t, x) N (0, 0) = V (0, 0). As Since the value of the numeraire is 1 today, values and reduced values are equal today: V we shall see, we only have to calculate the reduced prices V (t, x) and never have to calculate the full prices (t, x). This simplies our formulas substantially. In terms of the reduced prices V (t, x), the LGM model is V Z 2 1 1 V (T, X )e 2 (X x) / dX (4.4b) V (t, x) = for any T > t. 2 with (4.4c) as always. 4.1. Zero coupon bonds and the forward curves.. Let us go a bit further before summarizing. The (reduced) value of a zero coupon bond is Z (t, x; T ) 2 1 1 1 Z = e 2 (X x) / dX (4.5) Z (t, x; T ) = (t, x) (T, X ) 2 N N Substituting for the numeraire and carrying out the integration yields the (reduced) zero coupon price (4.6) Z (t, x; T ) = D(T )eH (T )x 2 H
1 2

= (T ) (t)

(T ) (t)

At t = 0, the state variable is x = 0, by denition. Since Z (0, 0; T ) = D(T ), the LGM model automatically matches todays discount curve D(T ). At t, x the instantaneous forward rate for maturity T , namely f (t, x; T ), is dened via (4.7a) (t, x; T ) = Z (t, x; T )N (t, x; T ) = e Z D(T ) = e
T
0

Similarly, the discount factor can be written in terms of todays instantaneous forward rate f0 (T ) as (4.7b)
f0 (T 0 )dT 0

T
t

f (t,x;T 0 )dT 0

.
2

So eqs. 4.7a, 4.7b) show that for the LGM model, (4.7c)
2

f (t, x; T ) = f0 (T ) + H 0 (T )x + [H 0 (T )] (t)

The last term [H 0 (T )] (t) is a small convexity correction; although it is needed for pricing, it does not aect the qualitative behavior of the model. The other terms show that at any date t, the forward curve is made up of todays forward curve f0 (T ) plus an amount x of the curve H 0 (T ). The amount x of the shift is a Gaussian random variable with mean zero and variance (t). The curve H 0 (T ) is a model parameter ; as we shall see, it replaces the mean reversion coecient (t) in the Hull-White model. The other model parameter is the variance (t). It takes the place of the volatility (t). As always, model parameters have to be set a priori during the calibration procedure by combining both theoretical reasoning (guessing) and calibration of vanilla instruments.
10

4.2. Aside: Connection to the Hull White model. Under the Hull White model, deals are valued according to the expectation n o (t0 )dt0 (t, r) = E e tT r (4.8a) V V (T, r (T )) r (t) = r for any T > t under the risk neutral measure. In this measure, the short rate r (t) is presumed to evolve according to (4.8b) dr = [(t) (t) r] dt + (t)dW.

The mean reversion (t) and local volatility (t) are the model parameters of this model. Given the mean reversion and local volatility, (t) is then chosen to match todays discount curve D(T ). In appendix A we derive the LGM model from the Hull White model. Thus, the LGM model is exactly the Hull-White model written in a more convenient form. There we also examine the connection between the Hull-White and LGM parameters. This shows that Hull-White parameters can be written in terms of the LGM parameters by (4.9a) (4.9b) H 00 (t) . H 0 (t) q (t) = H 0 (t) 0 (t). (t) =

As we shall see, the value of any vanilla option depends only on the value of the variance at the exercise date, (tex ), and on mean reversion function H (tj ) at the deals pay dates tj . Calibration determines the functions (t) and H (t) fairly directly. Obtaining the mean reversion parameter (t) requires dierentiating H (t) twice, which is an inherently noisy procedure. Similarly, obtaining (t) also requires dierentiating (t). This is why calibrating directly on the Hull-White model (instead of the LGM formulation of the model) is often an inherently unstable procedure. 4.3. Model invariances. The LGM parameters can be written in terms of the Hull-White parameters as (4.10a) (4.10b) H (t) = A Z
t t0
0

e Z
t

(t)dt

dt0 + B,
( )d

1 (t) = 2 A

0 +2 2 1 (t )e

t0
0

dt0 ,

where A and B are arbitrary positive constants. Since dierent A and B yield the same Hull-White model, and thus yield the identical prices, the LGM model has two invariances. First, all market prices remain unchanged if we change the model parameters by: (4.11a) H (T ) CH (T ), (t) (T )/C 2 .

for any positive constant C . To prove this, note that if we make the above transformation and then transform the internal variables x and X by (4.11b) x x/C, X X/C,

we obtain the same transition probabilities and zero coupon bond prices that we started with. Second, all market prices remain unchanged if (4.12a) H (T ) H (T ) + K (T ) (T )

for any constant K . To prove this, note that if we make the above transformation, and then transform the internal variables x and X by (4.12b) x x + K, X X + K,

11

 (t, x; T ) = N (t, x)Z (t, x; T ) as we obtain the same transition probabilities and zero coupon bond prices Z before. It is critical to pin down these invariances (by arbitrarily choosing some value of H (t) and of (t)) before calibration. Otherwise convergence would be innitely slow, with numerical roundo determining which of the equivalent sets of model parameters is chosen. 4.4. Scaling. On average, interest rates in G7 countries change by 80 bps or so over the course of a year. Equivalently, the standard deviation of H 0 (T )x should be about 1% or less each year. We choose to use the time scale of years (so T = 1 means an elapsed time of 1 year) and we scale H (T ) and H 0 (T ) to be O(1). Then x is of order O(1% t) at date t, and (t) is of the order of O(104 t). This makes the last term in (4.7c) O(104 t) also. More precisely, suppose we have chosen H (0) = 0, and have scaled H (T ) so it increases by 1 or so every year. Then: H (T ) O(T ), (t) O 0.64 104 t (4.13a) H 0 (T ) O(1), (4.13b) (4.13c) x, X 0.8 102 t

4.5. Summary of the LGM model. The complete LGM model can be summarized as Z 2 1 1 V (T, X )e 2 (X x) / dX1 dX2 for any T > t, (4.14a) V (t, x) = 2 with (T ) (t), and with the (reduced) zero coupon bond formula being (4.14b) and with (4.14c) x=0 at t = 0 Z (t, x; T ) = D(T )eH (T )x 2 H
1 2

H 0 (T )H (T ) (t) 0.64 104 tT

(T ) (t)

Consequently, (0) = 0. These equations are the only facts about the model we need to price any security. This model automatically reproduces the discount curve D(T ). The functions H (T ) and (t) are model parameters, which are set during the calibration step, where the model prices are matched to the market prices of selected vanilla instruments, usually caplets and swaptions. Once the model is calibrated, H (T ) and (t) are known functions, and the price of exotic deals can be determined from the Martingale formula 4.14a, using the zero coupon formula 4.14b to calculates the payos. Later we will present the calibration and pricing steps in exquisite detail. 5. Calibration. 5.1. Calibration and hedging. Model calibration is the most critical step in pricing. It determines not only the price obtained for an exotic deal, but also the hedges of the exotic. To see this, suppose we have some model M. It invariably contains unknown mathematical parameters which are set by calibration. To calibrate, one selects a set of vanilla instruments whose volatilities (prices) are known from market quotes. Let these volatilities be 1 , 2 , . . . , n . The calibration procedure picks the model parameters by matching the models yield curve to todays discount factors D(T ); and
12

 matching the models price of the selected vanilla instruments to their maket volatilities, either exactly of in a least squares sense. The calibrated model M0 is a function of todays discount factors and these n volatilities. The calibrated model is now used to price the exotic deal. The only step in this procedure which uses market information is the calibration step. This means that the price of the deal is a function of todays yield curve D(T ) and the n volatilities 1 , 2 , . . . , n . The price of the deal depends on no other market information. Consider what happens at the nightly mark-to-market. The model is calibrated and deal is priced as above. Next the vega risks are calculated by bumping the vols in the volatility matrix (cube) one-by-one. After each bump, the deal is priced using the identical software, and the dierence between the new price and the base price is the bucket vega risk for the the bumped volatility. Unless the bumped vol is one of the n vols used in calibration, it has no eect on the calibration of the model, so it does not aect the price of the instrument. An exotic deal only has vega risks to the n vanilla instruments used in calibration. After the vega risks are calculated for all the deals on the books, enough of each vanilla instrument is bought/sold to neutralize the corresponding bucket vega risk. This means that in the normal course of events, an exotic deal will be hedged by a linear combination of the vanilla instruments used during calibration. If the span of the vanilla instruments provide a good representation of the exotic, then the hedges should exhibit rock solid stability, with the day-to-day amounts of the hedges changing only as much as necessary to account for the actual changes in the market place. If the vanilla instruments do not provide a good representation of the exotic, then the hedges may exhibit instabilities, with day to day amounts of the hedges changing substantially even for relatively minor market changes. This latter is highly undesirable as the increased hedging costs gradually eliminate any initial prot from the exotic. (The nice term for this is leaking away your P&L). Indeed, in practice, even small improvements in the algorithm for matching the hedges to the exotics pay o disproportionately in the adroitness of the hedging. One can move most, and possibly all, the vega risk from the calibration instruments to a dierent (and presumably better) set of hedging instruments by using risk migration. This is also known as applying an external adjuster. As a side benet, this method also improves the pricing, often dramatically. This technique will also be discussed as it pertains to the dierent exotics. 5.2. Exact forumulas for swaption/caplet pricing. 5.2.1. Swaps. Consider a swap with start date t0 , xed leg pay dates t1 , t2 , . . . , tn , and xed rate Rf ix . The xed leg makes the payments (5.1a) (5.1b) i Rf ix 1 + n Rf ix paid at ti paid at tn , for i = 1, 2, . . . , n 1,

where i = cvg(ti1 , ti , ) is the coverage for period i according to the xed legs day count basis . On any given day t, the xed legs value is (5.2a) f ix (t, x) = Rf ix V
n X i=1

(t, x; tn ) (t, x; ti ) + Z i Z

As discussed earlier, the value of the oating leg is (5.2b) (t, x; t0 ) + f lt (t, x) = Z V
n X i=1

(t, x; ti ), i Si Z

where Si is the oating rates basis spread, adjusted to the xed legs day count basis and frequency. The value of the receiver swap is (5.2c) rec (t, x) = V
n X i=1

(t, x; tn ) Z (t, x; t0 ). (t, x; ti ) + Z i Rf ix Si Z

13

where the strike Rf ix and eective spread Si are known constants. Under the LGM model, the reduced value of the swap is clearly (5.3a) where (5.3b) Z (t, x; t) = Di eHi x 2 Hi (t)
1 2

Vrec (t, x) =

n rec (t, x) X V i Rf ix Si Z (t, x; ti ) + Z (t, x; tn ) Z (t, x; t0 ) = N (t, x) i=1

is the reduced value of the zero coupon bonds. Here (5.3c) Di = D(ti ), Hi = H (ti )

are the discount factors and values of H (t) at the swaps pay dates ti . Payer swap values are the negative of the receiver swap values. 5.2.2. Swaptions. A swaption is a European option on a swap. Consider a receiver swaption with notication date te . If one exercises on this date, one receives the xed leg and pays the oating leg, so clearly (5.4)
opt (te , xe ) = Vrec

" n X
i=1

#+ f ix i R Si Z (te , xe ; ti ) + Z (te , xe ; tn ) Z (te , xe ; t0 )

for any T > 0. Evaluating the integral at T = te yields (5.6) 1 =p 2 e 1 =p 2 e 1 =p 2 e

is the (reduced) value of the receiver swaption on the exercise date. Under the LGM model, todays value of the swaption is given by Z 2 1 1 opt (5.5) Vrec (0, 0) = p e 2 X / (T ) Vrec (T, X )dX 2 (T )
opt rec V (0, 0) #+ " n Z X 2 1 i Rf ix Si Z (te , X ; ti ) + Z (te , X ; tn ) Z (te , X ; t0 ) dX e 2 X / e

Z Z

eX ey

/2 e

/2 e

" n X
i=1

" n X
i=1

i=1

2 2 1 1 i Rf ix Si Di e(Hi H0 )y 2 (Hi H0 ) e + Dn e(Hn H0 )y 2 (Hn H0 ) e D0

2 2 2 1 1 1 i Rf ix Si Di eHi X 2 Hi e + Dn eHn X 2 Hn e D0 eH0 X 2 H0 e

#+

dX #+

dy

Here e = (te ) and Hi = H (ti )as always. We now assume wolog that H (T ) is an increasing function, so that H 0 (T ) > 0. We can always ensure this is so by using the invariance H (T ) H (T ), (t) (t) to change the sign of H 0 (T ). The quantity inside the square brackets [ ]+ is clearly a decreasing function of y since Hi H0 > 0 for all i. So dene the break-even point y as the unique solution of (5.7a)
n X i=1

i (Rf ix Si ) Di e(Hi H0 )y

2 1 2 (Hi H0 ) e

+ Dn e(Hn H0 )y

2 1 2 (Hn H0 ) e

= D0 .

14

Working out the interval shows that the receiver swaptions price is ! n X y + [ H H ] i 0 e opt rec p (0, 0) = i (Rf ix Si ) Di N (5.7b) V e i=1 ! ! y + [Hn H0 ] e y p D0 N p +Dn N e e

under the 1 factor LGM model. A payer swaption is a European option to pay the xed leg and receiver the oating leg. Repeating the above derivation for a payer swaption shows that its value today is ! n ! X y y [Hi H0 ] e p p i (Rf ix Si ) Di N e e i=1 ! y [Hn H0 ] e p Dn N e
n X i=1

(5.7c)

opt pay V (0, 0)

= D0 N

Using N (x) = 1 N (x), this can be written as (5.7d)

opt opt pay rec V (0, 0) = V (0, 0) + D0

i (Rf ix Si ) Di Dn ,

which is call/put parity. Caplets and oorlets. Consider a oorlet for the interval 0 to 1 . The oating rate r for the interval is set on the xing date te two (London) business days before the interval starts at 0 , and oorlet pays the dierence between the strike (xed rate) and the oating rate at the end of its period, provided this dierence is positive. As discussed earlier, the oorlets payo is can be written as h i+ f loorlet (te ) = 1 + (te ; 0 ) , (te ; 1 ) Z (5.8) V Rf ix s1 Z independent of the model being used. This is the same payo as a 1 period receiver swaption, so the value of a oorlet is the special case of a receiver swaption, with n = 1. Under the LGM model, we can solve for y explicitly when n = 1, obtaining f loorlet (0, 0) = (1 + (5.9a) V Rf ix s1 )D1 N (d 1 ) D0 N (d2 ) 1+ Rf ix s1 2 log 1 2 (H1 H0 ) e 1+ [R0 s1 ] p = . (H1 H0 ) e R0 = D ( 0 ) D ( 1 ) + s1 . D( 1 )

with

(5.9b)

d 1,2

and R0 is the forward FRA rate, dened by (5.9c)

Note that the and s1 are the orginal coverage and basis spread for the oating leg; there is no adjustment to a xed leg frequency and day count basis. Similarly, caplets are single period payer swaptions. Under the LGM model, there value is caplet (0, 0) = D0 N (d1 ) (1 + (5.10a) V Rf ix s1 )D1 N (d2 )
15

with 1+ R0 s1 2 log 1 (H1 H0 ) e 1+ [Rf ix s1 ] 2 p = , (H1 H0 ) e

(5.10b)

d1,2

and again D ( 0 ) D ( 1 ) + s1 . D( 1 )

(5.10c)

R0 =

5.2.3. Summary of 1 factor LGM exact pricing formulas. Under the LGM model, the prices of vanilla swaptions, caplets, and oorlets depend on (t) only through (tex ), its value at the notication date. The swaption prices depend on H (T ) only through the dierences H (tj ) H (t0 ) for the pay dates tj of the xed leg. This will be the key to creating lightning fast, stable calibration schemes. Under the one factor LGM model, the exact pricing formulas for swaptions are
n X i=1

(5.11a)

opt rec V (0, 0) =

i Rf ix Si Di N

! y + [Hi H0 ] ex p ex ! ! y + [Hn H0 ] ex y p D0 N p +Dn N ex ex

(5.11b)

y opt pay V (0, 0) = D0 N p ex

(5.11c)

opt rec =V (0, 0) + D0

! f ix y + [Hi H0 ] ex p i R Si Di N ex i=1 ! y + [Hn H0 ] ex p Dn N ex

n X n X i=1

i Rf ix Si Di Dn .

Here y is obtained by solving

n X i=1

(5.11d)

2 2 1 1 i Rf ix Si Di e(Hi H0 )y 2 (Hi H0 ) ex + Dn e(Hn H0 )y 2 (Hn H0 ) ex = D0 .

Newtons method requires the derivatives of the prices with respect to the model parameters. We observe
16

that ! y + [Hi H0 ] ex p (5.12a) ex ! p f ix opt y + [Hn H0 ] ex opt p (5.12b) V (0, 0) = V (0, 0) = ex 1 + n R Sn Dn G Hn rec Hn pay ex ! n X y + [ H H ] i 0 ex opt opt rec pay p (0, 0) = p (0, 0) = [Hi H0 ] i Rf ix Si Di G (5.12c) p V V ex ex ex i=1 ! y + [Hn H0 ] ex p + [Hn H0 ] Dn G ex ! n X p f ix opt y + [Hi H0 ] ex opt p i R Si Di G (5.12d) V (0, 0) = V (0, 0) = ex H0 rec H0 pay ex i=1 ! p y + [Hn H0 ] ex p ex Dn G ex p opt opt V (0, 0) = V (0, 0) = ex i Rf ix Si Di G Hi rec Hi pay The caplet/oorlet prices are given by (5.13a) with 1+ R0 s1 log 1 (H1 H0 )2 ex 1+ [Rf ix s1 ] 2 p = (H1 H0 ) ex caplet (0, 0) = D0 N (d1 ) (1 + V [Rf ix s1 ])D1 N (d2 )

(5.13b) and (5.13c)

d1,2

f loorlet (0, 0) = (1 + [Rf ix s1 ])D1 N (d V 1 ) D0 N (d2 ) opt (0, 0) + (1 + [Rf ix s1 ])D1 D0 =V caplet

with d 1,2 = d2,1 : log (5.13d) d 1,2 = 1+ [Rf ix s1 ] 1 2 2 (H1 H0 ) ex 1+ [R0 s1 ] p . (H1 H0 ) ex R0 = D ( 0 ) D ( 1 ) + s1 . D( 1 )

Here, R0 is the forward FRA rate (5.13e)

The caplet/oorlet prices are clearly Blacks formulas for call/put prices for an asset with forward value D0 , adj strike (1 + 1 Rf ix )D1 , and implied volatility satisfying (5.14) p imp tex = (H1 H0 ) ex
17

where

5.2.4. Approximate vanilla pricing formulas for the 1 factor LGM model. It is useful to develop approximate formulas for one factor LGM model, even though we have exact closed form formulas. Recall that market prices for swaptions are usually quoted in terms of Blacks formula 0 mkt 0 0 rec V (5.15a) (0) = Rf ix N(d0 1 ) R N(d2 ) L , 0 mkt f ix pay V (5.15b) (0) = R0 N(d0 N(d0 2) R 1) L mkt (0) L0 Rf ix R0 = Vrec d0 1,2 =
2 log Rf ix /R0 1 2 tex tex

(5.15c) and (5.15d) R0 =

Pn D0 Dn + i=1 i Si Di , L0

L0 =

n X i=1

i Di

This provides a good way to use market quotes of the implied volatility to obtain initial guesses for calibration. One can re-write these quotes more simply in terms of the implied normal volatility. Under the Gaussian (normal) swap rate model, the value of the swaption is f ix f ix R R0 R R0 mkt Vrec (0) = L0 Rf ix R0 N (5.16a) N ex G N ex N ex 0 0 f ix 0 R Rf ix R R mkt 0 f ix (5.16b) R R N ex G N Vrec (0) = L N ex N ex mkt = Vrec (0) L0 Rf ix R0 . The equivalent vol work shows that the implied normal or absolute vol N , is approximately Pn p i (R0 Si )Di (Hi H0 ) + Dn (Hn H0 ) Pn (5.16c) N tex ex i=1 . i=1 i Di

Here Di = D(ti ) are todays discount factors at the pay dates. By using equivalent vol techniques (or direct asymptotics), one discovers that under the LGM model, the implied (Black) volatility of the swaption is approximately: p Pn 0 ex i (Hi H0 ) + Dn (Hn H0 ) i=1 i (R Si )D P . (5.15e) B tex n f ix 0 R R i=1 i Di

5.2.5. Forward volatility. Forward volatility is a key concern of calibrated models. Suppose that we calibrate a model and then ask what the swaption volatilities will look like at a date t in the future. If the volatilities are increasing with t, we may be buying future volatility at too dear a price, and if volatilities are decreasing with t, we may be selling future volatility too cheaply. If we repeat the above equivalent vol analysis at a date t in the future, then we discover that the (normal) swaption volatility at that date is s Pn ( ex ) (t) i=1 i (R0 Si )Di [H (ti ) H (t0 )] + Dn [H (tn ) H (t0 )] Pn (5.17a) N , ex t i=1 i Di where (5.17b) P D0 Dn + n i=1 i Si Di R = . 0 L
0

If H (t) is decreasing exponentially, then ( ex ) should be increasing exponentially to compensate.

18

6. Calibration strategy. The most critical aspect of pricing is choosing the right set of vanilla instruments for calibrating the model. Even small improvements in matching the vanilla instruments to the exotic deal often lead to signicant improvements in the price and the stability of the hedge. For each type of exotic, the best calibration strategy often cannot be determined from purely theoretical considerations. Instead, one needs to determine which method leads to the best (the most market t) prices and risks. Here we briey discuss calibration strategies, illustrating the dierent strategies with a simple Bermuda swap (Bermudan swaptions and callable swaps are considered much more carefully in a later section). Consider a Bermudan receiver with start date t0 , end date tn , and strike Rf ix . Let the xed leg dates be t0 , t1 , . . . , tn , and let the exercise dates be 1 , 2 , . . . , n . If the Bermudan is exercised at j , then the holder receives the xed leg payments (6.1a) (6.1b) i Rf ix 1 + n Rf ix paid at ti , paid at tn , i = j, j + 1, . . . n 1

where i = cvg(ti1 , ti , ) is the coverage for interval i. In return, the holder makes the oating leg payments, which are worth the same as (6.1c) (6.1d) 1 i Si paid at tj 1 , paid at ti ,

i = j, j + 1, . . . n.

Here we have adjusted the basis spread to the xed legs frequency and day count basis as discussed above. Therefore, if the Bermudan is exercised at j , one receives/makes the payments 1 f ix i R Si 1 + n Rf ix Sn
n X i=j

(6.2a)

at tj 1 , at ti , at tn .

for i = j, j + 1, . . . n 1,

Clearly at any point t, x the j th payo is worth: (6.2b) pay (t, x) = V j (t, x; ti ) + Z (t, x; tn ) Z (t, x; tj 1 ). i (Rf ix Si )Z

6.1. Characterizing the exotic. The rst step in calibration is to characterise the exotic, extracting its essential features. If the Bermudan is exercised on exercise date j , one receives a swap worth (6.3)
n X i=j

( j , x; ti ) + Z ( j , x; tn ) Z ( j , x; tj 1 ) i (Rf ix Si )Z

at j

Suppose we evaluate this swap using todays yield curve with a parallel shift of size , (6.4) ( j , x; ti ) D(ti )eti = Di eti . Z

The shift j at which the j th swap is at-the-money is found by solving: (6.5)

n X i=j

i (Rf ix Si )Di e j (ti tj1 ) + Dn e j (tn tj1 ) = Dj 1

The Bermudan is characterized by a) the set of exercise dates 1 , 2 , . . . , n ; b) the set of parallel shifts j for j = 1, 2..., n; and c) the length tn t0 of the longest swap.

19

The second step is to select a calibration strategy and choose the calibration instruments. As we shall see, since we have two functions of time to calibrate, we can calibrate two seperate series of vanilla instruments. Under the LGM model, the prices of vanilla swaptions, caplets, and oorlets depend on (t) only through (tex ), its value at the notication date. The swaption prices depend on H (T ) only through the dierences H (tj ) H (t0 ) for the pay dates tj of the xed leg. This will be the key to creating lightning fast, stable calibration schemes. The trick is to calibrate on vanilla instruments whose pay dates line up exactly. We now go through the various calibration strategies for this Bermudan. 6.2. Calibration to the diagonal with constant mean reversion.. Consider the swaptions with exercise (notication) dates j , start dates tj 1 , end dates tn , and strike Rf ix for j = 1, 2, . . . , n. These are called the diagonal set of swaptions for the Bermudan. More briey, the j into n j swaptions struck at Rf ix . Surely we should calibrate our LGM model to these swaptions, for if the model failed to price these swaptions accurately, we could have no condence in the Bermudan price. In the normal course of business, our hedges will be a linear combination of our calibration instruments, in this case the diagonal swaptions. As market conditions change, the particular combination will also change. Intuitively, the Bermudan should be well represented by a (dynamic) linear combination of these swaptions, so one expects the amount of the hedges to be stable. For this calibration technique, we take the mean reversion coecient to be constant. Empirically taking = 0 gives decent prices for exotic deals, although these prices can usually be improved by taking the mean reversion to be positive (generally < 6%) or slightly negative (generally > 1%). Recall that H (T )/H (T ) = , so that H (T ) = AeT + B for some constants A and B . At this point we use the model invariants H (T ) CH (T ) and H (T ) H (T ) + K to set (6.6) H (T ) = 1 eT ,

is given implicitly by Here yj n X i=j

without loss of generality, where T is measured in years. With H (T ) known, we need to determine (t) by calibrating to the diagonal swaptions. Under the LGM model, the value of the j th swaption is ! n X + [ H H ] y i j 1 j j p i (Rf ix Si )Di N (6.7a) Vjmod = j i=j ! ! + [Hn Hj 1 ] j yj yj p +Dn N Dj N p . j j i (Rf ix Si )Di e(Hi Hj1 )yj 2 (Hi Hj1 )
1 2

(6.7b)

+ Dn e(Hn Hj1 )yj 2 (Hn Hj1 )

= Dj 1 ,

and j = ( j ) is the value of on the exercise date of the j th swaption. For future reference, we note that ! n X + [ H H ] y i j 1 j j p V mod = p [Hi Hj 1 ] i (Rf ix Si )Di G (6.7c) j j j i=j ! + [Hn Hj 1 ] j yj p + [Hn Hj 1 ] Dn G j The market value of the j th swaption is (6.8a) n o j 0 R R , Vjmkt = L0 N d N dj f ix j j 1 2
20

where the level and current swap rate for the j th swap are (6.8b) and (6.8c) dj 1,2 = L0 j =
n X i=j

i Di ,

0 Rj

Dj 1 Dn +

L0 j

Pn

i=j

i Si Di

0 2 1 log Rf ix /Rj 2 j j

j j

1/2

Here j is the implied volatility of p the j th swaption obtained from, e.g., the volatility cube. For each swaption j , the value j needs to be chosen to match the LGM value Vjmod of the swaption to p its market value Vjmkt . This value is unique since Vjmod is an increasing function of j . Since the derivative p p Vjmod / j is known explicitly, j can be found easily by using a global Newtons method. (Note: It is p usually more ecient to solve for j instead of j ). Since (0) = 0, once the swaptions are calibrated, we have (t) at 0, 1 , 2 , . . . , n . Piecewise linear interpolation should be used to get values of (t) for dates t between these points. As we shall see, if 1 , 2 , . . . , n are the Bermudans exercise dates, evaluating the Bermudan does not require knowing (t) at intermediate dates. p 6.2.1. Aside: Initial guess. An accurate initial guess for j can be found from the equivelent vol formula 5.15e. This yields Pn q q i=j i Di 0 j j j Rf ix Rj Pn (6.9) f ix Si )Di (Hi Hj 1 ) + Dn (Hn Hj 1 ) i=j i (R 6.2.2. Aside: Global Newtons method for one parameter ts. Suppose one is trying to solve f (z ) = target (6.10) for z . Normally one starts from an intial guess z0 , and expands f (zn+1 ) = f (zn + z ) f (zn ) + f 0 (zn )z to obtain a Newtons method: target f (zn ) (6.11) z = zn+1 zn = . f 0 (zn ) Provided this algorithm converges, it converges very rapidly. Unfortunately, this algorithm sometimes diverges. The global Newton method diers in only one respect: after calculating the Newton step z , one checks to see if taking this step decreases the error. If it does, one accepts the step. If it does not, then one cuts the step in half, and then again checks to see if the error decreases. Eventually the error will decrease, and the step is accepted. The next Newton step is then calculated. 6.2.3. Aside: Infeasible market prices. Since Z t 2 (t0 )dt0 , (6.12) (t) =
0

clearly (t) must be an increasing function of t : (6.13)

0 = (0) 1 2 n .

Since each j is calibrated seperately, it may happen that j < j 1 . (In practice this happens very, very rarely, but it does happen). One should test to see that the condition j j 1 is true after each j is found, and when this conditiona is violated, one should replace j by j 1 , its minimum feasible value: (6.14) j j 1 if j < j 1 . This means that the j th swaption will be priced at the closest possible price to the market price attainable within the calibrated LGM model, but it will not match the price exactly.
21

6.2.4. Aside: Where do the 0 s come from?. Suppose we set , calibrate the model to the diagonal, and then price the Bermudan. The resulting Bermudan price is a slightly increasing function of . Selecting the right ensures that we match the market price for the Bermudan. Desks often use a matrix to keep track of the needed to price a y NC x Bermudan correctly. That is, they ll in the s for the liquid Bermudans, and use continuity obtain the other entries in the matrix. Empirically, the change very, very slowly. market makers keep track of the mean reversion 6.3. Calibration to the diagonal with H (T ) specied. Suppose that H (T ) is specied a priori. (A possible source of such curves H (T ) is indicated below). Typically H (T ) is given at discrete points H (T1 ), H (T2 ), . . . , H (TN ). In that case, piecewise linear interpolation is used between nodes. This is equivalent to assuming that all shifts of the forward rate curve are piecewise constant curves. See 4.7c. With H (T ) set, we can use the preceding procedure and formulas to calibrate on the diagonal swaptions. This determines the value of (t) at 1 , 2 , . . . , n . As above, one adds the point (0) = 0, one ensures that the j = ( j ) are increasing, and uses piecewise linear interpolation to obtain (t) at other values of t. 6.3.1. Origin of the H (T ). Suppose one had the set of Bermudan swaptions 30 NC 20, 30 NC 15, 30 NC 10, 30 NC 5 and 30 NC 1. Wouldnt it be nice if the same curve H (T ) were used for each of these Bermudans? The 30 NC 10 Bermudan includes the 30 NC 15 and the 30 NC 20 Bermudans. It would be satisfying if our valuation procedure for the 30 NC 15 and 30 NC 20 assigned the same price to these Bermudans regardless of whether they were individual deals or part of a larger Bermudan. One could arrange this by rst using a constant , lets call it 4 , to calibrate and price the 30 NC 20 Bermudan. Without loss of generality, we could select (6.15a) H 0 (T ) = e4 (T30 T ) 4 (T30 T ) 1 H (T ) = e 4 for T20 T T30 .

We would calibrate on the diagonal to nd (t) at expiry dates m , m+1 , . . . beyond 20 years, and then price the 30 NC 20 Bermudan. Selecting the right value of 4 would match the Bermudan price to its market value. Neither the swaption prices nor the Bermudan prices depend on H (T ) or (t) for dates before the 20 year point. To price the 30 NC 15, one could use the H (T ) obtained from 4 for years 20 to 30, and choose a dierent kappa, say 3 , for years 15 to 20: (6.15b) H 0 (T ) = e3 (T20 T ) e4 (T30 T20 ) 1 4 (T30 T20 ) e4 (T30 T20 ) 1 e H (T ) = e 3 4
3 (T20 T )

for T15 T T20 .

Calibrating would produce the same (t) values for years 20 to 30 as before. In addition, for each 3 it would determine (t) for years 15 to 20. By selecting the right 3 , one could match the 30 NC 15 Bermudans market price. Continuing in this way, one produces the values of (t) and H (T ) for years 10 to 15, for years 5 to 10, and nally for years 1 to 5. This (t) and H (T ) would then yield a model which matches all the diagonal swaptions and happens to correctly price all the liquid, 30y co-terminal Bermudans. These (t)s turn out to be extremely stable, only varying very rarely, and then by small amounts. Typically a desk would remember the (t)s as a function of the co-terminal points, relying on the same (t)s for years. In general, if Tn is the co-terminal point and T0 , T1 , . . . , Tn1 are the no call points, then H (T ) is: (6.16) H (T ) =
n n n X ek (Tk Tk1 ) 1 Y i (Ti Ti1 ) ej (Tj T ) 1 Y i (Ti Ti1 ) e e j k i=j +1 k=j +1 i=k+1

for Tj 1 T Tj

After H (T ) and (t) have been found, one can use the invariants to re-scale them if desired.
22

6.4. Calibration to the diagonal with linear (t). This is an idea pioneered by Solomon brothers. Let us use a constant local volatility . Then Z t 2 dt0 = 2 t (6.17) (t) =
0

is linear. By using the invariance (t) (t)/C , H (T ) CH (T ) we can choose to be any arbitrary constant without aecting any prices. So we choose (6.18) (t) = 2 0 t,

where t is measured in years, and the dimensionless constant 0 is typically 102 . For this calibration, we use the other invariant to set Hn = H (tn ) = 0. We now determine the values of Hi for other values of i by calibrating on the diagonal swaptions, starting with the last swaption. Recall that the price of the j th diagonal swaption is ! n X + [Hi Hj 1 ] j yj mod f ix p = Rf ix i (R Si )Di N (6.19a) Vj j i=j ! ! + [Hn Hj 1 ] j yj yj p +Dn N Dj N p j j and its derivative with respect to Hj 1 is (6.19b) ! n q X + [Hi Hj 1 ] j yj mod f ix p V = j i (R Si )Di G Hj 1 j j i=j ! q y + [Hn Hj 1 ] j p j Dn G j
1 2

is given implicitly by Here yj n X i=j

(6.19c)

i (Rf ix Si )Di e(Hi Hj1 )yj 2 (Hi Hj1 )

+ Dn e(Hn Hj1 )yj 2 (Hn Hj1 )

= Dj 1 .

Consider the last swaption, j = n. It depends on n = ( n ), on Hn , and Hn1 . Of these, n is known, mod Hn has been set to zero, so only Hn1 is unknown. Since Vn is a decreasing function of Hn1 , there is a unique value of Hn1 which matches the model price to the market price. This can be found easily using a global Newtons method. We can then move onto the j = n 1 swaption. This swaption depends on Hn2 , which is unknown, and j 1 , Hn1 , and Hn , which are known. Working backwards like this, we can calibrate all of the swaptions, and for each calibration there will only be a single unknown parameter, Hj 1 . This calibration procedure will yield H0 , H1 , ..., Hn on the dates t0 , t1 , ..., tn . One uses linear interpolation/extrapolation to get H (t) at other values of t. 6.4.1. Infeasible values. In deriving the swaption formulas, we assumed that H (T ) was an increasing function of T . (This assumption was stronger than we needed: inspection of the above argument shows that one only needs to assume that there is a unique break-even point y , with in-the-moneyness on the left.) Since we are calibrating the Hj s seperately, it may happen that Hj 1 may exceed Hj . (In practice, this has never happened to my knowledge. Still one must be prepared.) After each Hj 1 is found, one should check to see that (6.20) Hj 1 Hj .

If this condition is violated, one should reset Hj 1 = Hj . This means the j th swaption would not match its market price exactly. Instead it would be the closest feasible price.
23

6.5. Calibration to diagonals with prescribed (t). Suppose (t) is a known function which is increasing and has (0) = 0. We could carry out the preceding calibration procedure to determine H (T ) from the diagonal swaptions; the procedure doe not depend on (t) being linear. 6.6. Calibration to caplets with constant mean reversion. One can also calibrate the LGM model to a series of caplets/oorlets instead of swaptions. This is inadvisable for pricing the Bermudan, since it means that we would invariably be hedging swaption risks with caplets. However, we present the caplet (oorlet) calibration methodology here, since this methodology is sensible for other deal types, such as autocaps and revolvers. Recall that the Bermudan has exercise date 1 , 2, ..., n . To hedge each of these vega risks, we choose caplets/oorlets whose xing dates exactly correspond to these exercise dates. So consider the n oorlets with (6.21) j = xing date, t0 j = start date, t1 j = end date,
f ix Rj = strike

for j = 1, 2, . . . , n. Here standard market practice for the currency determines the oorlets start dates and end dates in terms of the xing date j . f ix appropriately. Surely we shouldnt For each of these reference caplets, we need to select the strike Rj f ix f ix just blindly pick the oorlet strikes Rj equal to the Bermudans strike Rj . Unless the yield curve happens to be at, the short dated oorlets are likely to be way in the money, and the long dated ones are likely to be way out of the money. Rather, to match the Bermudan as closely as possible, we need to choose oorlets which are at-the-money precisely when the swaps underlying the Bermudan are at the money. Consider the Bermudans payo from exercising at j . Recall that this payo would be at-the-money under a parallel shift j of the yield curve. To match the Bermudan as closely as possible, we need to pick f ix the strike Rj so that the oorlet is at-the-money for the same parallel shift j . The oorlets payo is at (6.22a) i+ o hn f ix 0 ( j ; t1 sj Z 1+ j Rj j ) Z ( j ; tj )

1 where j = cvg(t0 j , tj ) is the day count fraction, and

(6.22b)

sj = oating rates basis spread fo start date t0 j.

See equation 3.9a et seq. Under the shifted yield curve, (6.23) ( j , x; t) D(t)e j t , Z

the value of the oorlets payo is n o 0 f ix j (t1 j tj ) D (t0 ). sj (6.24a) 1+ j Rj D(t1 j )e j For this to be at-the-money, we need to choose (6.24b) for the caplets strike. 6.6.1. Calibration to caplets. For this calibration method, we choose a constant mean reversion . This means that H (T ) = AeT + B for some constants A and B , and using the invariants we can set (6.25) H (T ) = 1 eT ,
24
f ix Rj + j (tj tj ) D(t1 D(t0 j )e j) = + sj 1 j D(tj )
1 0

without loss of generality. Here T is measured in years. We now obtain (t) by calibrating to the oorlets (6.26a) with strikes (6.26b)
f ix = Rj + j (tj tj ) D(t1 D(t0 j )e j) + sj 1 j D(tj )
1 0

j = xing date,

t0 j = start date,

t1 j = end date,

f ix Rj = strike

for j = 1, 2, ..., n. The market value of the j th oorlet is (6.27a) n o f ix j j 0 Vjmkt = , j D(t1 j ) Rj N d1 Rj N d2
0 Rj = 1 D(t0 j ) D (tj ) + sj , 1 j D(tj )

where the current break-even rate for the j th oorlet is (6.27b) and (6.27c) dj 1,2 =

f ix 0 2 /Rj 1 log Rj 2 j j

j j

1/2

with

Here j is the implied volatility of the j th oolet obtained from, e.g., the volatility cube. Under the LGM model, the value of the j th oorlet is h i f ix 0 j j j Rj sj )D(t1 (6.28a) Vjmod (0, 0) = (1 + j )N d1 D (tj )N d2 h i f ix sj 1+ j Rj 1 0 2 0 1 j log 2 Hj Hj 1+ j Rj sj 1 p = . 0 j Hj Hj

(6.28b) and its derivative is (6.28c) Here (6.28d)

j d 1,2

1 f ix 0 j p Vjmod = Hj Hj )D(t1 (1 + j Rj j )G d1 . j j = ( j ),
0 Hj = H (t0 j ), 1 Hj = H (t1 j ).

For each caplet j , we determine j by using a global Newton procedure to match the oorlet price under the LGM model price to its market price. Since the model price is an increasing function of j , this solution is unique. Alternatively, we note that Vjmod is just Blacks formula: (6.29a) with (6.29b) (6.29c) (6.29d)
Vjmod = D(t1 j ) {F N (d1 ) K N (d2 )}

1 2 j 1,2 = log F/K , d 2 j

25

0 1 j Rj sj K = D(t0 j )/D (tj ) = 1 +

j F =1+ j Rf ix sj

and (6.29e) q j = j 1 H0 . Hj j

So one can use an existing implied vol routine to obtain , and then j . After tting all the caplets, one needs to ensure that (6.30) 0 = (0) 1 n .

It is conceivable that j < j 1 for some j ; if this ever occurs, then one raises j up to equal j 1 . As always, we use linear interpolation to ll in the values at other times t. 6.7. Calibration to caplets with H (T ) specied. There are occasionally times when one wishes to calibrate the model with the curve H (T ) is specied a priori. Typically H (T ) is given at discrete points H (T1 ), H (T2 ), . . . , H (TN ), and piecewise linear interpolation is used between nodes. (This is equivalent to assuming that all shifts of the forward rate curve are piecewise constant curves. See 4.7c. With H (T ) set, we choose the same oorlets as above and calibrate to determine the value of (t) at 1 , 2 , . . . , n . One then ensures that the (t)0 s are increasing, and uses piecewise linear interpolation to obtain (t) at other values of t. 6.8. Calibration to caplets with linear (t). For this calibration strategy, we assume a constant local volatility so that (t) is linear in t. By using the invariance (t) (t)/C 2 ,we can choose (t) to be (6.31) (t) = 2 0 t,

with

for any constant 0 , without loss of generality. Here t is measured in years, and we choose the dimensionless constant 0 to be 102 . We now choose the same oorlets as above, and determine the H (T ) by calibrating to the caplets. The value of the j th oorlet is Under the LGM model, the value of the j th oorlet is h i j D(t0 )N d j j Rf ix sj )D(t1 )N d (6.32a) V mod (0, 0) = (1 +
j j j 1 j 2

(6.32b) and (6.32c)

j d 1,2

h i f ix sj 1+ j Rj 1 0 2 0 1 j log 2 Hj Hj 1+ j Rj sj 1 p = . 0 j Hj Hj
0 Hj = H (t0 j ), 1 Hj = H (t1 j ).

j = ( j ),

0 With j known, the only unknown is Hj = H (t1 j ) H (tj ) for each oorlet. Since the price is an increasing function of Hj , we can use a global Newton routine (or an implied vol routine) to nd the unique value of Hj which matches the model price to the market price. We can then take H (T ) to be piecewise linear, with 0 0 0 H (0) = 0, with breaks in the slope at t0 2 , t3 , ..., tn , and with the slopes chosen to match the Hj s implied by the market prices. This yields

(6.33a)

H (T ) = (H1 )T
k 1 X j =2

for T t0 2
0 for t0 k T tk+1

(6.33b)

H (T ) = (Hk )(T t0 k) +

0 0 (Hj ) t0 j +1 tj + (H1 )t2

26

for k = 1, 2, ..., n 1 and (6.33c) H (T ) = (Hn )(T t0 n) +

n 1 X j =2

0 0 (Hj ) t0 j +1 tj + (H1 )t2

for t0 n T

After calibration, one should ensure that H (T ) is increasing, modifying the values of H (T ) as needed. 6.9. Calibration to caplets with prescribed (t). Suppose (t) is a known function which is increasing and has (0) = 0. We could carry out the preceding calibration procedure to determine H (T ) from the caplets; the procedure doe not depend on (t) being linear. 6.10. Calibration to diagonal swaptions and a row of swaptions. The one factor LGM model has two model parameters, (t) and H (T ), so it can be calibrated to two distinct sequences of vanilla instruments. Here we simultaneously calibrate the LGM model on the diagonal swaptions (6.34a) exercise date j , start date tj 1 , end date tn for j = 1, 2, ...n,

and on the row of swaptions with common exercise date 1 , common start date t0 , and varying end dates: (6.34b) exercise date 1 , start date t0 , end date tk for k = 1, 2, ...n.

We choose the strike of the diagonal swaptions to be the xed rate Rf ix of the Bermudan, for clearly the diagonal swaptions would be at-the-money whenever the respective Bermudan payo is at-the-money. To f ix select the strike Rk of the k th row swaption, note that this swaption has the payo (6.35a)
row Vk (t, x) = k X i=1 f ix i (Rk Si )Z (t, x; ti ) + Z (t, x; tk ) Z (t, x; t0 ),

f ix at any t, x. We choose Rk so that this swaption is at-the-money under same the parallel shift 1 ,

(6.35b)

(t, x; ti ) D(t)e 1 ti , Z Pk D0 Dk e 1 (tk t0 ) + i=1 i Si Di e 1 (ti t0 ) = . Pk 1 (ti t0 ) i=1 i Di e 1 = ( 1 ) = 2 0 1

as the Bermudan exercise with the exercise date 1 . This requires (6.36)
f ix Rk

We rst use the multiplicative invariant to set (6.37a)

without loss of generality, where we arbitrarily choose 0 = 0.01. We use the additive invariant to set (6.37b) H0 = H (t0 ) = 0.

Under the LGM model, todays value of the k th row swaption is ! k X y + [Hi H0 ] 1 f ix row p Vk (0, 0) = i Rk Si Di N (6.38a) 1 i=1 ! ! y + [Hk H0 ] 1 y p D0 N p , +Dk N 1 1
27

and its derivative with respect to Hk is (6.38b) i p h f ix row Vk (0, 0) = 1 1 + k Rk Sk Dk G Hk

k X i=1

Here y is the solution of (6.38c)

f ix i (Rk Si )Di e(Hi H0 )y
2 1 2 (Hi H0 ) 1

y + [Hk H0 ] 1 p 1

+ Dk e(Hk H0 )y

2 1 2 (Hk H0 ) 1

= D0 .

The k = 1 swaption depends on H0 , H1 , and 1 , of which only H1 is unknown. Since the swaptions price is an increasing function of H1 , we can use a global Newton scheme to obtain the unique value of H1 at which the swaptions price will match its market value. The k = 2 swaption then introduces one new parameter, H2 , which we select by using the global Newton scheme to match this swaption to its market value. In very rare situations, it may occur that H2 < H1 , in which case we set H2 = H1 . (In these cases we cannot match the swaption price exactly, so we set H2 to its closest feasible value). We continue taking the swaptions in turn, with each swaption j introducing one additional value of Hj to be set by calibrating the swaption to its market value. We then ensure that Hj Hj 1 , adjusting the value of Hj if needed, and then continue to the next swaption. In this way we nd (6.39) H0 = H (t0 ), H1 = H (t1 ), . . . , Hn = H (tn ).

We use linear interpolation/extrapolation to get H (T ) as other values of T . Having found H (T ), we can now use the calibration to the diagonal with H (T ) specied method (and code!) to get (t). 6.10.1. Aside: Using 2 . For some deals, 1 is too short for the 1 into k swaptions to be used for calibration. Instead of blindly taking 1 as the exercise date, many rms take the Bermudans rst exercise date which is at least, say, 6 months from today. 6.11. Calibration to diagonal swaptions and caplets. A Bermudan swaption can be viewed as the most expensive of its component European swaptions, plus an option to switch to a dierent swaption should market conditions change. The component swaptions are just the diagonal swaptions, so calibrating to the diagonals accounts for this part of the pricing. On any exercise date, switch option is the option to exercise immediately, or to delay the exercise decision until the next exercise date. Since these delays are short, typically six months, one may believe that the switch option can best be represented by short underlyings. Accordingly, one could argue that one should calibrate to either a column of caplets or a column of 1 year underlyings, as well as the diagonal swaptions. Here we calibrate on the caplets and swaptions simultaneously; in the next section we calibrate to the diagonal swaptions and the swaptions with one year underlyings. 6.12. Calibration to diagonal swaptions and a column of swaptions. One may could argue that caplet and swaption markets have distinct indentities, and that mixing the two markets introduces small, but needless, noise. Instead one could calibrate on the diagonal swaptions and a column of swaptions with 1 year tenors. (In most currencies, these are the swaptions with the shortest underlying avaliable). 6.13. Other calibration strategies. There are many other simple calibration strategies; although they are not overly appropriate for pricing a Bermudan, they may well be appropriate for other deal types. Calibrate on swaptions with constant or specied H (T ). Suppose we have chosen a constant mean reversion parameter , or have otherwise specied H (T ). Then the calibration procedure just needs to nd (t). Suppose we have selected an arbitrary set of n swaptions to be our calibration instruments. In LGM valuation of each swaption the only unknown parameter is (t) at the swaptions exercise date. Using a global Newtons method to calibrate each swaption to its market value thus determines (t) and the exercise dates 1 , 2 , . . . , n of the n swaptions. After obtaining the j = ( j ), we need to ensure that ( j ) are
28

non-decreasing, alterring the oending values if necessary. We then include the value 0 = (0) = 0, and use piecewise linear interpolation to obtain (t) at other dates. Note that this method fails if two swaptions share the same exercise date ; calibration would either yield the same , in which case one of the swaptions is redundant, or diering , in which case our data is contradictory. If the exercise dates of any two swaptions are too close, say within 1-2 months, the results may be problematic. For this reason one usually ensures that the swaption exercise dates are, say, at least 21 2 months apart, excluding instruments from the calibration set to achieve this spacing, if necessary. Calibrate on swaptions with specied (t). Suppose we have chosen a linear (t), or otherwise specied parameter (t). The calibration procedure just needs to nd H (T ). Suppose we have selected an arbitrary set of n swaptions to be our calibration instruments. We can then arrange the swaptions in increasing order of their nal pay dates. Let these nal pay dates be T1 , T2 , . . . , Tn . Suppose we use our invariance to set H0 = H (0) = 0, and we use piecewise linear interpolation: (6.40a) (6.40b) H (T ) = 1 T H (T ) =
k 1 X

for T < T1 , for Tk1 < T < Tk , for Tn < T,

(6.40c)

H (T ) =

i=1 n 1 X i=1

i (Ti Ti1 ) + k (T Tk1 ) i (Ti Ti1 ) + n (T Tn1 )

where T0 = 0. For the rst swaption, the slope 1 determines the value of H (T ) at all the swaptions pay dates. Since (t) is known, the LGM value of the swaption depends only on a single unknown quantity, . It is easily seen that the value is an increasing function of 1 , so one can use a global Newton scheme to nd the unique 1 which matches the swaptions price to its market value. The value of H (T ) at the second swaptions pay dates is determined by both 1 and 2 , of which only 2 is unknown at this stage. Again a global Newton scheme can be used to nd the 2 needed to calibrate the swaption to its market value. (In rare cases it may occur that 2 < 0; in this case we need to set 2 = 0, its minimum feasible value). We then continue in this way, calibrating the swaptions and obtaining the j s in succession. This method will fail only if two deals have the same nal pay date, and will work poorly if the nal pay dates are too near together. For this reason one usually ensures that the nal pay dates are, say, at least 2 1 2 months apart, excluding instruments from the calibration set to achieve this spacing, if necessary. 7. Evaluation of exotics. dldldl (Part head:)Valuation of exotic swaps and caps 8. Bermudan swaptions and callable swaps (bullet). 8.1. Deal denition and encapsulation. 8.1.1. Mid-period exercises. 8.1.2. Encapsulation of deal data. 8.2. Characterization. 8.3. Evaluation. 8.4. Comparision of results. 8.4.1. Constant mean reversion and market implied 0 s. 8.4.2. Given H (t)0 s matching the Bermudan market. 8.5. Extension to american swaptions.
29

9. Bermudan swaptions and callable swaps (amortizing). 9.1. Deal denition and encapsulation. 9.2. Characterization. 9.3. Evaluation. 9.4. Comparision of results. 10. Callable inverse oaters. 11. Callable capped oaters and superoaters. 12. Callable accrual swaps. 13. Autocaps. 14. Revolvers. 15. Captions (European options on caps). (Part head:)Valuation of bonds 16. Extending the LGM model to include credit. 17. Valuation of noncallable bullet bonds. 18. Valuation of callable bonds. Appendix A. Mechanics and standard practices of swap markets. A.1. Date arithmetic and business day conventions. A.2. Year fractions, day count bases, and adjusted interest rate payments. A.3. Fixing dates, spot dates, and oating legs. A.4. Notication dates and the swaption market. A.5. Cap markets. Appendix B. Connection between LGM and Hull White. The one factor LGM model is completely equivalent t the Hull-White model; it is just written in a vastly more convenient form. Here we show this by deriving the LGM model from the Hull-White model. This then determine the relationship between the Hull White model parameters and the LGM model parameters. B.1. Hull-White model. The Hull-White model is written using the money market numeraire. With (t, y ) of any tradeable security at t, y is this numeraire, the value V n o (t0 ))dt0 (t, y ) = E e tT r(t0 ,Y (B.1a) V V (T, Y ) Y (t) = y , where the short rate is given by (B.1b) ) = (t) + Y , r(t, Y

start at 0 and evolve according to and where the state variables Y (B.1c) = (t)Y dt + (t)dW , dY (0) = 0, Y

= r(t, Y ) = (t) + Y as the fundamental process. For convenience we (Often authors use the short rate R ). The model parameters are have seperated the short rate R(t) into its mean (t) and random component Y the mean reversion (t) and the local volatility (t). The other model parameter, (t), is reserved to make the model match todays discount curve. To derive the LGM model from Hull-White, dene the new variable (B.2) (t) = Y (t)e+ U
30
t
0

( )d

Then the short rate is given by (B.3a) evolves according to where U (B.3b) Here h(t) is dened by (B.4a) h(t) = e
t
0

, r(t) = (t) + h(t)U

= (t)dW , dU

(0) = 0 U

(t0 )dt0

and (t) is related to the Hull-White parameters by (B.4b) (t) = (t)e+

t
0

( )d

(t, u; T ) of a zero coupon bond can be calculated by solving the Feynman-Kac equation The value Z (B.5a) t + 1 2 Z = [(t) + h(t)u] Z uu = r(t, u)Z Z 2

for T < t, with the boundary condition (B.5b) Solving we obtain (B.6a) where (B.6b) and (B.6c) A(t, T ) = Z
T t

(T, u; T ) = 1 Z

at t = T.

(t, u; T ) = e[H (T )H (t)]uA(t,T ) Z Z

H ( ) =

h(t0 )dt0 + K,

(t0 )dt0

1 2

[H (T ) H (t0 )] 0 (t0 )dt0 .

Here the function (t) is given by (B.6d) ( ) = Z

2 (t0 )dt0

and K is an arbitrary constant. At t, u = 0, 0, the zero coupon bond must match todays discount factors D(T ). Recalling that the discount factors can be written as (B.7) we see that this requires (B.8) (T ) = f0 (T ) + H 0 (T ) Z
T

D(T ) = e

T
0

f0 (t0 )dt0

0 (t0 ) [H (T ) H (t0 )] dt0 .

Substituting for allows us to write the zero coupon bond formula as (B.9a)
2 2 (t, x; T ) = D(T ) e[H (T )H (t)]x 1 2 [H (T )H (t)] (t) , Z D(t)

31

where (B.9b) x(t) = u

0 (t0 )H (t0 )dt0 .

We have nearly derived the LGM model from the Hull-White model. Consider the variables Z t (t) = U X (B.10) 0 (t0 )H (t0 )dt0 0 Z t (t)e+ 0t ( )d =Y 0 (t0 )H (t0 )dt0
0

The Hull-White model can now be written as o n (t0 ))dt0 (t, x) = E e tT r(t0 ,X V (T, X ) X (t) = x , (B.12a) V where (B.12b) (t0 )) = (t) + H 0 (t)X (t) + H 0 (t) r(t0 , X Z
t

The new variable evolves according to , 1 = 2 (t)H (t) dt + (t)dW (B.11a) dX

(0) = 0, X

H (t0 ) 0 (t0 )dt0

(t) + H 0 (t)H (t) (t). = f0 (t) + H 0 (t)X Accordingly, the value of any freely tradeable security (between cash ows) satises the Feynman-Kac equation x + 1 0 (t)V xx = f0 (t) + H 0 (t)x + H 0 (t)H (t) 0 (t) V t H (t) 0 (t)V (B.13) V 2 (t, x; T ) satisfy this PDE. Another solution One can verify by direct substitution that the zero coupon bonds Z is 2 (t, x) = 1 eH (t)x+ 1 2 H (t) (t) , (B.14) N D(t) as can again be veried by direct substitution. Let us dene the reduced price (B.15a) V (t, x) = (t, x) V . (t, x) N

(t, x) = N (t, x)V (t, x) into equation ??, and using the fact that N (t, x) also satises ??, one Substituting V nds that V (t, x) satises (B.15b)
0 Vt + 1 2 (t)Vxx = 0

But this is the backwards Kolmogorov equation for n o (t) = x , (B.16a) V (t, x) = E V (T, X )| X (t) as evolving according to the process provided we take X (B.16b) Here (B.16c) ( ) = (t) = (t)dW . dX Z

2 (t0 )dt0 ,

as before. Together, B.14, B.15a and B.16a - B.16c are LGM model.
32

B.2. Connection between LGM and HW parameters. The LGM model parameters are the (accumulated) variance (t), and the response or mean reversion function H (t). The Hull-White model has the local vol function (t), the mean reversion rate (t), and (t). The mean reversion function H (t) can be obtained from the Hull-White mean reversion parameter by integrating: Z
t t0
0

(B.17a)

H (t) =

e
0

(t)dt

dt0 + K.

In principle, the mean reversion parameter can be obtained by dierentiating the observed mean reversion function: (B.17b) (t) = H 00 (t) . H 0 (t)

As we shall see, the value of a vanilla instrument depends only on the value of the mean reversion function Hj (t) at the pay dates t of the instrument, so calibration determines the function H (t) fairly directly. Obtaining the mean reversion parameter (t) from this function requires dierentiation, which is an inherently noisy procedure. This is why calibrating directly on the Hull-White formulation of the model (instead of LGM formulation) is often inherently unstable. The variance (t) can also be determined from the HW local volatility (t) : (B.18a) (t) = Z
t 0 +2 2 1 (t )e 0 t0
0

( )d

dt0 ,

We shall nd that prices of European options depend on the value of (t) at the exercise date of the option, and not on the value at other times. This allows us to cleanly determine (t) by calibrating to vanilla deals with dierent exercise dates. Determining the local volatility (t) from (t) requires dierentiating. This is another reason why calibrating the Hull-White formulation of the model is often unstable, while calibrating the LGM formulation is stable.

Alternatively, the HW parameters can be determined by dierentiating the M (t) matrix: q (B.18b) (t) = H 0 (t) 0 (t).

33