METHODOLOGY FOR CALLABLE SWAPS AND BERMUDAN EXERCISE INTO SWAPTIONS

PATRICK S. HAGAN BLOOMBERG LP 499 PARK AVENUE NEW YORK, NY 10022 PHAGAN1@BLOOMBERG.NET 212-893-4231 Abstract. Here we present a methodology for obtaining quick decent prices for callable swaps and Bermudan exercise into swaps using the LGM model. Key words. Bermudans, callable swaps

1. Introduction. This is part of three related papers: Evaluating and hedging exotic swap instruments via LGM explains the theory and usage of the LGM model in detail. This paper, Methodology for Callable Swaps and Bermudan Exercise Into swaptions, details the methodology, including all steps of the pricing procedure. Finally, Procedure for pricing Bermudans and callable swaps, breaks down the method into a procedure and set of algorithms. This paper has three appendices. The rst appendix discusses handling Bermudan options on amortizing swaps (as opposed to bullet swaps). Amortizers require a slightly more sophisticated deal characterization step, which results in selecting a dierent set of vanilla instruments for calibration. Once the deals are selected, the calibration and evaluation steps are identical to those of the bullet Bermudans. The second appendix discusses American swaptions. With the appropriate pre-processing step, American swaptions can be priced by by using the Bermudan pricing engine. The third appendix is used to point out the modications that are needed if the two legs are in dierent currencies. 1.1. Notation. In our notation today is always t = 0, and (1.1a) D(T ) = todays discount factor for maturity T.

For any date t in the future, let Z(t; T ) be the value of $1 to be delivered at a later date T , (1.1b) Z(t; T ) = zero coupon bond, maturity T , as seen at t.

These discount factors and zero coupon bonds are the ones obtained from the currencys swap curve. Clearly D(T ) = Z(0; T ). We use distinct notation for discount factors and zero coupon bonds to remind ourselves that discount factors D(T ) are not random; we can always obtain the current discount factors from the stripper. Zero coupon bonds Z(t; T ) are random, at least until time catches up to date t. Also, we use N(z) and G(z) to be the standard (cumulative) normal distribution and Gaussian density, respectively: Z
z
2 1 G(z) = ez /2 2

(1.2)

N (z) =

G(z 0 )dz 0 ,

2. Deal denition and representation. Bermudans arise mainly from two sources. The rst is a direct Bermudan swaption, also called an exercise into Bermudan. The other (more common) source is a cancellable swap, which is invariably priced as a swap plus a Bermudan swaption to enter the opposite swap. Bermudans from both sources (and virtually any other Bermudan that arises) t into following deal
1

structure. After dening this deal structure, we will show how to t the most common types of Bermudans into the structure. Our Bermudan structure contains the following information: Payment information: (2.1a) (2.1b) (2.1c) Exercise information: (2.1d) (2.1e) (2.1f) (2.1g) (2.1h) (2.1i) P orR = payer or receiver ag tex [1, 2, ..., J] = exercise (notication) dates tset [, 1, 2, ..., J] = settlement date if exercised at ex j if irst [, 1, 2, ..., J] = rst coupon payment received if exercised at ex j f ee[, 1, 2, ..., J] = exercise fee (paid on tset ) j rf p[, 1, 2, ..., J] = reduction in rst coupon payment received if exer at ex j t[0, 1, 2, ..., n] = paydates C[, 1, 2, ..., n] = full payments for each interval N [, 1, 2, ..., n] = notionals for each interval

(In my notation, means this element of the array is not used. In my opinion, the indexing is simpler and less confusing if we waste the rst entry in all the vectors except t, but this is only a personal preference.) The Bermudan can be exercised on any of the notication dates tex for j = 1, 2, ..., J. Suppose rst the j P orR ag is set to receiver. Then, if the Bermudan is exercised at date tex , the owner receives all the j payments starting with payment i = if irst . However, the rst payment received is reduced by rf pj (which j may be zero): (2.2a) (2.2b) Ci rf pj Ci received at ti received at ti for i = if irst , j for i = if irst + 1, ..., n. j

In return, the owner pays the notional plus the exercise fee at the settlement date (2.2c) Nif irst + f eej
j

paid at tset . j

The full payments Ci include the xed legs interest, notional payments and prepayments, as well as adjustments for basis spreads and any margins. The oating leg is mainly accounted for by paying the notional Nj on settlement. Suppose now the P orR ag is set to payer. If the Bermudan is exercised at date tex , one receives the j payment (2.3a) and makes the payments (2.3b) (2.3c) Ci rf pj Ci paid at ti paid at ti for i = if irst , j for i = if irst + 1, ..., n. j Nif irst f eej
j

received at tset . j

In the next section we show how real deals, both the exercise into and callable swap Bermudans, can be put into the above deal structure. From then on we work exclusively with deal structure.
2

2.1. Swap. Let us rst dene the swap, and then dene the exercise features of the two types of Bermudan. We assume that the swap exchanges a xed leg against a standard oating leg plus a margin; we also assume that the legs are in the same currency. (This latter assumption is dropped in Appendix C). 2.1.1. Fixed leg. Let (2.4a) (2.4b) tth < tth < tth < tth < tth 0 1 2 n1 n t0 < t1 < t2 < tn1 < tn

be the xed legs theoretical and actual dates. In our notation, (2.5) is period i, and (2.6a) (2.6b) (2.6c) Ni f ix Ri = notional for period i, = xed rate for period i, ti1 < t ti

ai = cvg(ti1 , ti , f ix ) = day count fraction for period i.

The xed leg payments are (2.6d)

f Ni i Ri ix

paid at ti ,

for i = 1, 2, ..., n

2.1.2. Funding (oating) leg. Let the oating legs theoretical and actual dates be (2.7a) (2.7b) th < th < th < th < th 0 1 2 n1 m 0 < 1 < 2 < n1 < m tth = th , 0 0 t0 = 0 , tth = th , n m tn = m .

where the beginning and end dates of the two legs must agree: (2.7c) (2.7d)

Let the j th oating period be j1 < t < j , and let (2.8a) (2.8b) (2.8c) (2.8d)
f Nj lt = notional for the j th period,

mj = margin for the j th period bsj = oating rates basis spread for j th period j = cvg( j1 , j , f lt ) = day count fraction for period j

The oating leg pays the oating rate plus a margin, (2.9)
f Nj lt j [rj lt + morig ] f j

paid at j ,

j = 1, 2, ..., m.

Prior to xing, the j th oating leg payment is worth the same as the payments (2.10a) (2.10b) {1 +
f Nj lt = f j [bsj +mj ]} Nj lt

paid at j1 , paid at j ,

for j = 1, 2, ..., m. This is just the denition of the (forward) basis spread bs.
3

2.1.3. Bond model of a swap. Floating leg dates often occur with a dierent frequency (usually more frequent) than the xed leg dates. We are going to replace the oating leg payments with the equivalent payments based on the xed rate schedule. Unless the basis spreads and margins are identically zero, this will result in an invisibly small approximation. Suppose rst that the oating leg intervals are equal to or shorter than the xed leg intervals. Based on the theoretical dates, we can assign every oating leg interval j to a xed leg interval (2.11a) j Ii if and only if tth < th tth . i1 j i

It makes no sense for the oating leg notional to change when the xed rate notional does not change. We f restrict ourselves to deals whose oating rate notional Nj lt is constant and equal to the xed rate notional Ni within each xed rate interval: (2.11b)
f Nj lt = Ni

for all j Ii . paid at ti1 , paid at ti , paid at j for all j Ii

The net swap payments (xed minus oating) for interval i are: (2.12a) (2.12b) (2.12c)
f Ni i Ri ix

+ Ni

Ni

Ni j [bsj +morig ] j

We move the basis spread and margin to the xed leg, approximating the swap payments for interval i as (2.13a) (2.13b)
ef Ni i Ri f

+ Ni

Ni

paid at ti1 , paid at ti ,

for i = 1, 2, ..., n. Here the eective xed rate for interval i is: P jIi j [bsj +mj ]D( j ) ef f f ix (2.13c) Ri = Ri . i D(ti ) Suppose now that the oating leg intervals occur less frequently than the xed leg intervals. Based on the theoretical dates, we again assume that we can assign every xed leg interval i to a oating leg leg interval (2.14a) i Ij if and only if th < tth th . j1 i j

f We again assume that the xed rate notionals Ni are constant and equal to the oating rate notional Nj lt within each oating rate interval :

(2.14b)

f Ni = Nj lt

for all i Ij .

We move the basis spread and margin to the xed leg. This once again leads to approximating the swap payments for interval i as (2.15a) (2.15b)
ef Ni i Ri f

+ Ni

Ni

paid at ti1 , paid at ti ,

for i = 1, 2, ..., n. Here the eective xed rate for interval i is: (2.15c)
ef f Ri f = Ri ix

j [bsj +mj ]D( j ) P . iIj i D(ti )

4

2.1.4. Optionality: Exercise into Bermudan swaptions. Let us rst consider an exercise into Bermudan option.. It is not uncommon for a Bermudan to be exercisable more frequently than once a period, as in a semi-pay, monthly call deal. So we need to allow for intra-period exercises. The optionality can be dened by (i) a payer/receiver ag (2.16a) (ii) a set of notication dates, (2.16b) tex , tex , ..., tex 1 2 J P orR,

(iii) a set of theoretical and actual settlement (start)-upon-exercise dates, (2.16c) (2.16d) (iv) a set of exercise fees (2.16e) f ee1 , f ee2 , ..., f eeJ . tth,set , tth,set , ..., tth,set 1 2 J tset , tset , ..., tset 1 2 J

Suppose the payer/receiver ag is receive. Then if the deal is exercised on the notication date ex , j the owner receives the swap starting from the settlement date tset . Specically, dene if irst as the i with j j (2.17a) tth tth,set < tth . i1 i j

For the rst payment, the owner receives the interest that accrues from the settlement date tset to the rst j coupon date ti at if irst . This is less that the full coupon if the settlement date tset is after the interval starts j j at ti1 . So if the deal is exercised at tj , then the owner of the option gets ex (2.17b) (2.17c) (2.17d) where (2.17e) In return, the owner pays (2.17f) Ni + f eej at tset j with i = if irst . j f irst = cvg(tset , ti ) j j with i = if irst . j
ef Ni f irst Ri f + Ni Ni1 j ef Ni i Ri f + Ni Ni+1 ef Nn n Rn f + Nn

at ti at ti at tn

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

If the payer/receiver ag is payer. Then if the deal is exercised at tex , the owner recieves j (2.18a) and pays (2.18b) (2.18c) (2.18d)
ef Ni f irst Ri f + Ni Ni1 j ef Ni i Ri f + Ni Ni+1 ef Nn n Rn f + Nn

Ni f eej

at tset j

with i = if irst , j

at ti at ti at tn
5

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

We can t this deal into the above Bermudan structure by dening the full payments (2.19a) (2.19b)
ef Ci = Ni i Ri f + Ni Ni+1 ef Cn = Nn n Rn f + Nn ,

for i = 1, ..., n 1,

and dening the if irst as the index i for which j (2.19c) tth tth,set < tth i1 i j for j = 1, ..., J

and dening the reduction in the rst payment as (2.19d) ef ef rf pj = (i f irst )Ni Ri f = Ni Ri f cvg(ti1 , ti ) cvg(tset , ti ) j j for j = 1, ..., J

with i = if irst .The notionals Ni , pay/rec ag P orR, exercise fees f eej , and exercise and settlement dates, j ex and set , are copied into the structure unchanged. j j Aside. Best practices is for a deals conrm to specify i) the theoretical settlement-upon-exercise dates tth,set ; j ii) the business day rules and holiday calendars needed to obtain the actual settlement dates from the theoretical dates (these should be identical to the rules for the xed leg), and iii) that the notication date must occur at least N business days (or calendar days) before the actual settlement date. Then regardless of whether holidays are added or subtracted after the deal is struck, the settlement dates always relate to the payment dates in the same way without one day gaps opening up. Confusingly, the settlement (start)-upon-exercise dates are often called the exercise dates and the exercise (notication) dates are simply known as notication dates.. 2.1.5. Optionality: callable swaps. Let us now a callable swap. Again, Bermudans may be callable more frequently than once a period. If a swap is called mid-period, the xed and oating legs accrued interest must be paid, as well as any exercise fee, on the settlement date. Then no further payments are received. This is equivalent to a non-callable swap, plus a Bermudan swpation to enter the opposite swap. Consider a callable swap. Let it be a payer or receiver, according to (2.20a) The callability is dened by (i) a set of notication dates, (2.20b) tex , tex , ..., tex 1 2 J P orR.

(iii) a set of theoretical and actual settlement-upon-exercise dates, (2.20c) (2.20d) (iii) a set of exercise fees (2.20e) f ee1 , f ee2 , ..., f eeJ . tth,set , tth,set , ..., tth,set 1 2 J tset , tset , ..., tset 1 2 J

The value of the cancellable swap is the value of the full (non-cancellable) swap plus the value of the cancelation feature. We assume that the non-cancellable swap is priced elsewhere. Here we only price the canacellation feature.
6

Suppose the payer/receiver ag is payer, and suppose the cancellation feature is exercised on the notication date ex . Dene if irst as the rst coupon after the settlement-upon-exercise date: j j (2.21a) tth tth,set < tth . i1 i j

Cancelling the payer swap is equivalent to recieving all the xed rate payments, and making all the oating leg payments, starting with payment i if irxt . So the owner receives the xed leg payments j (2.21b) (2.21c)
ef Ni i Ri f + Ni Ni+1 ef Nn n Rn f + Nn

at ti at tn

for i = if irst , ..., n 1 j for i = n,

and makes the oating leg payments, which are equivalent to (2.21d) Ni at ti1 .

with

At the settlement date, the owner also pays the accrued xed leg interest, receives the accrued oating rate interest, and pays any exercise fee. So the owner must also pay ef true (2.22a) f eej + Ni set Ri f ri at tset j j set = cvg(ti1 , tset ) j j with i = if irst . j

(2.22b) Here we use the true rate (2.22c)

true ri =

Z(t, ti1 ) Z(t, ti ) i Z(t, ti )

ef for interval i, instead of the forward oating rate, because the basis spread is already incorpoarted into Ri f . The oating rate payment at ti1 along with the settlement payments are now equivalent to n o ef f at tset (2.23a) Ni + f eej + Ni set Ri f rj lt j j

where

(2.23b)

f rj lt =

Z(t, ti1 ) Z(t, ti ) Z(t, ti1 ) Z(t, tset ) j i Z(t, ti ) set Z(t, tset ) j j

is the dierence between the rates for the full and partial intervals. Virtually every desk neglect this correction. We can do better by estimating this dierence from todays curve: (2.24)
f rj lt =

D(ti1 ) D(ti ) D(ti1 ) D(tset ) j . i D(ti ) set D(tset ) j j

In summary, if the owner of a payer swap cancels (by providing notication at tex ), the cancellation is j equivalent to receiving the payments (2.25a) (2.25b)
ef Ni i Ri f + Ni Ni+1 ef Nn n Rn f + Nn

at ti at tn
7

for i = if irst , ..., n 1 j for i = n,

and making the payment (2.25c) where (2.25d) n o ef f Ni + f eej + Ni set Ri f rj lt j

f rj lt =

at tset , j

D(ti1 ) D(ti ) D(ti1 ) D(tset ) j . i D(ti ) set D(, tset ) j j

and the payments (2.26b) (2.26c) As before, (2.26d)

Similarly, suppose the owner of a receiver swap cancels (by providing notication at tex ). Cancellation j is equivalent to making receving the payment n o ef f (2.26a) Ni f eej + Ni set Ri f rj lt at tset , j j
ef Ni i Ri f + Ni Ni+1 ef Nn n Rn f + Nn

at ti at tn

for i = if irst , ..., n 1 j for i = n,

f rj lt =

D(ti1 ) D(ti ) D(ti1 ) D(tset ) j . i D(ti ) set D(tset ) j j

As previously stated, we assume that the value of the non-cancellable swap is calculated elsewhere, and here only price the cancellation feature. We can t this cancellation feature into our Bermudan structure by dening the full payments to be (2.27a) (2.27b)
ef Ni i Ri f + Ni Ni+1 ef Nn n Rn f + Nn

at ti at tn

for i = 1, ..., n 1 for i = n.

by dening the payer/receiver ag to be receiver if a payer swap is cancellable, and to be payer if a receiver swap is cancellable, by dening if irst so that j (2.27c) by dening the exercise fee to be (2.27d) f eej = f eej Ni set j (
ef Ri f

tth tth,set < tth , i1 i j )

D(ti1 ) D(ti ) D(ti1 ) D(tset ) j + i D(ti ) set D(tset ) j j

for j = 1, ..., J.

Here the + sign is to be taken for callable payer swaps, and the sign for callable receivers. For callable swaps, the reduction in the rst payment to be zero (2.27e) rf pj = 0 for j = 1, ..., J

and the notionals Ni , the xed leg pay dates ti , the execise and settlement dates tex and tset are to be copied j j into the structure without change.with i = if irst .The notionals Ni , pay/rec ag P orR, exercise fees f eej , j and exercise and settlement dates, ex and set , are copied to the structure unchanged. j j
8

Aside. Best practices is for a deals conrm to specify i) the theoretical settlement-upon-exercise dates (call dates) tth,set ; j ii) the business day rules and holiday calendars needed to obtain the actual call dates from the theoretical dates, and iii) that the notication date must occur at least N business days (or calendar days) before the actual settlement date. The settlement-upon-call dates are often called the call dates and the exercise (notication) dates are simply known as notication dates. 3. The LGM (Linear Gauss Markov) model. 3.1. Basic LGM. We value these deals using calibrated LGM models. This model is chosen because it is very reliable as well as being very easy to work with. As explained fully in Evaluating and hedging exotic swap instruments via LGM, the one factor LGM model has a single state variable x, and uses the numeraire (3.1a) N (t, x) = 1 +H(t)x+ 1 H 2 (t)(t) 2 e D(t)

Let V f ull (t, x) be the actual value of any deal. Throughout we one use only the reduced value (3.1b) V (t, x) = V f ull (t, x) . N (t, x)

At t = 0, x = 0, the numeraire is 1, so todays full values and reduced values are identical. As we shall see, the full values at other dates are not relevent. The LGM model can be summarized in two relations: First, the (reduced) value V (t, x) of any deal can be determined from its value at any later date T via the expected value Z 2 1 e(Xx) /2 V (T, X)dX, (3.2a) V (t, x) = 2 where (3.2b) = (T ) (t).

Second, the (reduced) value of a zero coupon bond with maturity ti is (3.2c) Z(t, x; ti ) = D(ti )eH(ti )x 2 H
1 2

(ti )(t)

as can be determined by substituting V (T, X) = 1/N (T, X) in the expected value. Here the functions H(T ) and (t) are found by the calibration step. They are equivalent to the mean reversion parameters (t) and the local vol (t) in the Hull-White model. 3.2. Invariances. Recall from Evaluating and hedging exotic swap instruments via LGM that all deal prices remain the same if we replace (3.3a) (3.3b) H(T ) H(T ) + C, H(T ) KH(T ),
9

(t) (t) (t) (t)/K 2

for any constants C and K. These invariances need to be recognized (and exploited!) in the calibration step.

3.3. Swaption value. Calibration is a procedure for choosing the functions H(T ) and (t) so that the LGM prices match the actual market prices for a selected set of swaptions, caplets, and oorlets. Here we obtain a closed form expression for the LGM price of these instruments. Consider a (receiver) swap with start date t0 , xed leg pay dates t1 , t2 , . . . , tn , and xed rate Rf ix . The xed leg makes the payments (3.4a) (3.4b) 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 (3.5a) Vf ix (t, x) = Rf ix
n X i=1

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

As discussed in, Evaluating and hedging exotic swap instruments via LGM, the value of the oating leg is (3.5b) Vf lt (t, x) = Z(t, x; t0 ) +
n X i=1

i Si Z(t, x; ti ),

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 (3.5c) Vrec (t, x) =
n X i=1

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

where the strike Rf ix and eective spread Si are known constants. Consider a European option on this swap (a swaption), and let ex be the exercise date. Under the one factor LGM model, todays value for the swaption is Z 2 1 Vrec ( ex , X) if positive opt (3.6) Vrec (0, 0) = p eX /2 ex dX, 0 if negative 2 ex where ex = ( ex ) Integrating yields the exact pricing formulas ! n X y + [Hi H0 ] ex opt f ix p i R Si Di N (3.7a) Vrec (0, 0) = ex i=1 ! ! y + [Hn H0 ] ex y p D0 N p +Dn N ex ex
n X i=1

where y is obtained by solving (3.7b)

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 . Hi = Hi H0 = H(Ti ) H(T0 ).

10

Observe that the swaption value depends on (t) only through ex = ( ex ), and on H(T ) only through (3.8)

at the pay dates. Using Newtons method in the calibration procedure requires the derivatives of the prices with respect to the model parameters. We observe that ! n X f ix y + [Hi H0 ] ex opt rec (0, 0) = p V p [Hi H0 ] i R Si Di G (3.9a) ex ex i=1 ! y + [Hn H0 ] ex p + [Hn H0 ] Dn G ex ! n X p y + [Hi H0 ] ex opt f ix p i R Si Di G (3.9b) V (0, 0) = ex H0 rec ex i=1 ! p y + [Hn H0 ] ex p ex Dn G ex ! p f ix y + [Hi H0 ] ex opt p (3.9c) V (0, 0) = ex i R Si Di G Hi rec ex ! p f ix opt y + [Hn H0 ] ex p (3.9d) V (0, 0) = ex 1 + n R Sn Dn G Hn rec ex 3.4. Bermudan payo. Recall the Bermudan structure has the payment information t[0, 1, 2, ..., n] = paydates, C[, 1, 2, ..., n] = full payments for each interval, N [, 1, 2, ..., n] = notionals for each interval, (3.10a) (3.10b) (3.10c) and the exercise information: (3.10d) (3.10e) (3.10f) (3.10g) (3.10h) (3.10i) P orR = payer or receiver ag t [1, 2, ..., J] = exercise (notication) dates tset [, 1, 2, ..., J] = settlement date if exercised at ex j
ex

if irst [, 1, 2, ..., J] = rst coupon payment received if exercised at ex j f ee[, 1, 2, ..., J] = exercise fee (paid on tset ) j rf p[, 1, 2, ..., J] = reduction in rst coupon payment received if exer at ex j

If the P orR ag is set to receiver., then the payo on the j th exercise date is (3.11a) Pj (tex , x) = (Ci0 rf pj ) Z(tex , x; ti0 ) + j j
n X

i0 +1

Ci Z(tex , x; ti ) (Ni0 + f eej ) Z(tex , x; tset ) j j j

where i0 = if irst for simplicity. Similarly, if the P orR ag is set to payer. then the j th payo is j (3.11b) Pj (tex , x) = (Ci0 rf pj ) Z(tex , x; ti0 ) j j
n X

i0 +1

Ci Z(tex , x; ti ) + (Ni0 f eej ) Z(tex , x; tset ) j j j

Since we know the value of the (reduced) zero coupon bond, (3.11c) Z(t, x; T ) = D(T )eH(T )x 2 H
11
1 2

(T )(t)

we can write these payos explicitly. For receivers, (3.12a) n X 2 s s 2 1 2 1 1 s Pj (tex , x) = (Ci0 rf pj ) Di0 eHi0 x 2 Hi0 j + Ci Di eHi x 2 Hi j Ni0 + f eej Dj eHj x 2 (Hj ) j , j j
i0 +1 j

and for payers, (3.12b) Pj (tex , x) = (Ci0 rf pj ) Di0 eHi0 x 2 Hi0 j j Here (3.12c) (3.12d) 4. Evaluating the deal. Di = D(ti ), j = (tex ), j
1 2

i0 +1 j

n X

s s 2 2 1 1 s Ci Di eHi x 2 Hi j + Ni0 f eej Dj eHj x 2 (Hj ) j . j

s Dj = D(tset ) j

Hi = H(ti ),

s Hj = H(tset ) j

4.1. Rollback. Let us assume that the calibration procedure has given us (t) and H(T ). We now show how to evaluate the Bermudan. For each exercise j we break x into a grid of points, (4.1a) xk = hj (k mx )
(j)

for k = 0, 1, ..., 2mx .

(Below, we show how to choose the spacing hj and width hj mx of the grid). We dene (4.1b) Pj,k = P (tex , xk ) j
(j)

as the payo if the deal is exercised at tex . If P orR is receiver, eqs. 3.12a - 3.12d allow us to calculate the j payo as (4.2a) n X (j) (j) 2 s (j) s 2 1 2 1 1 s Pj,k = (Ci0 rf pj ) Di0 eHi0 xk 2 Hi0 j + Ci Di eHi xk 2 Hi j Ni0 + f eej Dj eHj xk 2 (Hj ) j j
i0 +1 j

at each k = 0, 1, ..., 2mx . Similarly, if P orR is payer, we calculate (4.2b) n X 2 s (j) s 2 1 2 1 1 s Pj,k = (Ci0 rf pj ) Di0 eHi0 x 2 Hi0 j Ci Di eHi x 2 Hi j + Ni0 f eej Dj eHj xk 2 (Hj ) j . j
i0 +1 j

at each k = 0, 1, ..., 2mx . Rollback is a backwards induction scheme. We rst use 4.2a - 4.2bto obtain the payo PJ,k at the last exercise date. Then (4.3) VJ,k = V ( ex , xk ) = max {PJ,k , 0} J at each k = 0, 1, ..., 2mx ,

is the value of the deal on the last exercise at tex , assuming that it has not been exercised at an earlier J exercise date.
12

Now suppose that we know the value of the deal at some exercise date tex , assuming that it was not j exercised on any of the exercise dates before tex . That is, we know j (4.4) Vj,k = V (tex , xk ), j
(j)

where xk = hj (k mx )

(j)

for k = 0, 1, ..., 2mx .

We now go to j 1. We rst break x into a grid of points (see below) (4.5) xk

(j1)

= hj1 (k mx )

for k = 0, 1, ..., 2mx .

(j1)

We use the Gaussian convolution formula to nd the value of the deal at each node xk (4.6) V + (tex , xk j1
(j1)

at tex : j1

1 )= 2

ey

/2

Vj (tex , xk j

(j1)

+y

q j j1 )dy.

This is the deals value at node xk on tex assuming it has not been exercised at tex or at any earlier j1 j1 exercise. We calculate this integral as the weighted sum,
2my

(j1)

(4.7a) with (4.7b)

(j1) (tex , xk ) j1

+ Vj1,k

X
i=0

wi Vj,k0 (i)

k 0 (i) =

hj1 (k mx ) + yi hj

j j1

+ mx ,

where the weights wi and yi will be specied shortly. Since k 0 will not be an integer, one should use piecewise linear interpolation (with at extrapolation) on Vj,k to get the Vj,k0 . Note that this sum over i has to be done for each node k, for k = 0, 1, ..., 2mx . (j1) + Now Vj1,k = V + (tex , xk ) is the value of the deal at tex assuming that the deal has not been j1 j1 ex exercised at tj1 or earlier. We now include the value of the exercise at tex . If the deal is exercised at tex , j1 j1 one gets the payo Pj1,k given by 4.2a - 4.2b with j j 1. Taking the maximum at each x, (4.7c) o n + Vj1,k = max Pj1,k , Vj1,k for k = 0, 1, ..., 2mx

now provides the the deals value at tex , including the exercise at tex . j1 j1 By looping over the rollback step, one obtains the value of the deal on the rst exercise date, V1,k = V1 (tex , xk ). A nal integration gives todays value of the deal: 1
2my

(4.8a) with (4.8b)

V (0, 0) =

X
i=0

wi V1,k0 (i)

k (i) =

yi

1 + mx h1

13

4.2. European options. Traditionally, Bermudan pricers also output the values of the European options that make up the Bermudan. This helps traders understand which exercise dates are the most valuable, and how much extra they are paying for the Bermudan over the most expensive European. Since we typically calibrate to these swaptions, the value of the European option should be the same as the market value. So for our case it is just a useful double check. The payo of the European option is (4.9a)
eur Vj,k = max {Pj ( j , xk ), 0} ,

and a single integration gives todays value of the j th European option of the range note
2my

(4.9b) with (4.9c)

Vjeur (0, 0) =

X
i=0

eur wi Vj,k0 (i)

k (i) =

yi

p hj

+ m.

4.3. Discretization and weights. One usually sets the x grid to be set number of points per standard deviation, with the width of the grid being a set multiple of the standard deviation. Recall that at tex the j variable x has mean 0 and variance j = (tex ). Setting the discretization as x points per standard j deviation, and extending the grid to Nx standard deviations, we have (4.10a) with (4.10b) hj = xk = hj (k mx ) q j /x ,
(j)

for k = 0, 1, ..., 2mx

mx = wx x .

Although some experimentation may be needed, typically Nx = 4 to 5.5 and x = 18 to 32 work well. To discretize the Gaussian and nd the weights wi , one again chooses the number of standard deviations and the number of points per standard deviation: (4.11a) (4.11b) yi = hy (i my ) for i = 0, 1, ..., 2my , hy = 1/y , my = Ny y ,

Typically Ny = 4 to 5.5 and y = 10 to 16 work well. One then generates a preliminary set of weights from Z yi +hy 2 |yi y| ey /2 wi = 1 (4.12a) dy hy 2 yi hy yi yi yi = 1+ N (yi + hy ) 2 N (yi ) + 1 N (yi hy ) hy hy hy 1 + {G(yi + hy ) 2G(yi ) + G(yi hy )} hy for i = 1, 2, ..., 2my 1. Here N (y) is the standard cumulative normal distribution, and G(y) is the Gaussian density, Z y 2 ey /2 (4.12b) G(y) = G(y)dy. , N (y) = 2
14

For i = 0 and i = 2my , we have special weights, (4.12c) Z y0 +hy Z y0 y2 /2 2 |y0 y| ey /2 e 1 dy + dy w2my = w0 = hy 2 2 y0 y0 y0 1 = 1+ N (y0 ) + {G(y0 + hy ) G(y0 )} . N (y0 + hy ) hy hy hy

4.3.1. Normalization of the weights. Once these weights are generated, one usually normalizes the weights, (4.13a) where the A and B are chosen so that
2my new 2 wi = (A + Byi )wi ,

(4.13b)

X
i=0

2my new wi

= 1,

X
i0

2 new yi wi = 1.

(By symmetry, all the odd moments are already zero.) If one calculates the moments with the original weights,
2my

(4.13c) we see that (4.13d)

M0 =

X
i=0

2my

wi ,

M2 =

X
i=0

2my 4 yi wi ,

M2 =

X
i=0

4 yi wi ,

A=

M4 M2 2, M0 M4 M2

B=

M2 M0 2. M0 M4 M2

4.3.2. Partial sums of the weights. We can speed up our integration routine if we have the weight generation routine also return a vector of partial sums, (4.14) Recall that the integration step
2my

Si =

k=0

i X

wk .

(4.15a) with (4.15b)

+ Vj1,k =

X
i=0

wi Vjk0 (i)

hj (k mx ) + yi k (i) = hj+1
0

j+1 j

+ mx

uses at-linear-at interpolation on the Vjk0 (i) . We can replace the sum over the i0 s with k 0 (i) < 0 and with k 0 (i) > 2mx : (4.16a)
+ Vj1,k = i 1 X

i=i +1

wi Vjk0 (i) + Vj,0 Si + Vj,2mx (1 Si )

15

where i is the largest i with (4.16b) and i is the smallest i with (4.16c) k0 (i ) > 2mx . k 0 (i ) < 0,

4.4. Accounting for the kinks. The Vj (x) in the integrand has a discontinuous rst derivative where the max switches from Vj+ (x) to Pj ( ex , x). This kink in the integrand is the dominant error, and by j eliminating this error, we can gain almost a full order of accuracy. Recall that in each step we take the maximum at each xk n o + (4.17a) Vj,k = max Pj,k , Vj,k for k = 0, 1, ..., 2mx ,
+ where Vj,k should be set identically zero for the last exercise j = J. We then evaluate the integral ! p Z hj1 (k mx ) + y j j1 1 + y 2 /2 e Vj + mx dy (4.17b) Vj1,k = hj 2

as the weighted sum,

2my

(4.18a) (4.18b)

+ Vj1,k

hj (k mx ) + yi k (i) = hj+1
0

X
i=0

wi Vjk0 (i) p j+1 j + mx

+ curve Pj,k and the curve Vj,k cross. Suppose that these curves cross in the interval K < k < K + 1. Then + + Pj,K Vj,K and Pj,K+1 Vj,K+1 have opposite signs. Dene n n o o + + (4.19a) MK = max Pj,K , Vj,K , mK = min Pj,K , Vj,K , n n o o + + mK+1 = min Pj,K+1 , Vj,K+1 , (4.19b) MK+1 = max Pj,K+1 , Vj,K+1

where piecewise linear interpolation is used to obtain Vjk0 (i) from theo grid points. n + Suppose that when we are taking the max, Vj,k = max Pj,k , Vj,k for all k, we record where the payo

Using a linear approximation (this is all we need to make the correction), these curves cross at the point (4.19c) k = K + MK mK MK+1 mK+1 =K +1 MK+1 mK+1 + MK mK MK+1 mK+1 + MK mK

in the interval. Our integration scheme is linear, so our base integration routine approximates the integrand as (4.20a) Vj,k = MK + (k K) (MK+1 MK ) for K < k < K + 1.

+ If we approximate both Pj ( j , xk ) and Vj,k as linear in k, then we would obtain MK + (k K)(mK+1 MK ) for K < k < k (4.20b) Vj,k = . mK + (k K)(MK+1 mK ) for k < k < K + 1

16

The error in the integrand is (4.20c) E(k) = (k K)(MK+1 mK+1 ) (K + 1 k) (MK mK ) for K < k < k . for K < k < k

+ We need to make the correction to Vj1,k of

(4.21a) where (4.21b)

1 + Cj1,k = 2

k(y)=K+1

ey
k(y)=K

/2

E(k(y))dy,

hj1 (k mx ) + y k(y) = hj

p j j1

+ mx .

The average value of the error over the interval is (4.22a) Eavg = 1 2 (MK mK ) (MK+1 MK ) . MK+1 mK+1 + MK mK

p p If hj / j j1 isnt too large, say, hj / j j1 1, we can correct the majority of the numerical error arising from the kink by evaluating the Gaussian at the midpoint and using the average. Thus, we should add the correction ! q hj (K + 1 mx ) hj1 (k mx ) hj Eavg + 2 p (4.22b) Cj1,k = p G if hj j j1 j j1 j j1 p + to Vj1,k for each k. On rare occasions, hj / j j1 may be too large to evaluate the Gaussian at the midpoint. For these cases, one should add the correction = Eavg N hj (K + 1 mx ) hj1 (k mx ) p j j1 ! Eavg N hj (K mx ) hj1 (k mx ) p j j1 q if hj > j j1 . !

C+ (4.22c) j1,k

Of course, the kinks should be corrected when evaluating the European options as well as the Bermudan option.nstruments, and then calibrating the model so that it matches the model prices against the market prices for these instruments. 4.5. Exotics evaluator. The evaluation step can be written in a way which is completely independent of the deal and the LGM model. Suppose we provide the evaluation routine with the following as inputs: (1) the number of exercises J and the values 1 , 2 , ..., J , (2) a pointer to a function which calculates the payo Pj,k = Pj (xk ), (3) the density of points 1/x and width Nx of the x grid to be used (4) the density of points 1/y and width Ny of the y discretization to be used The evaluation routine depends on no other input. This means that we can use the same evaluation function for dierent deal types just by writing new payo functions.
17

4.6. Writing the payo function. Recall that the payo functions are (4.23a) n X (j) (j) 2 s (j) s 2 1 2 1 1 s Pj,k = (Ci0 rf pj ) Di0 eHi0 xk 2 Hi0 j + Ci Di eHi xk 2 Hi j Ni0 + f eej Dj eHj xk 2 (Hj ) j j
i0 +1 j

if P orR is receiver, and (4.23b) Pj,k = (Ci0 rf pj ) Di0 e

2 Hi0 x 1 Hi0 j 2

i0 +1 j

n X

s (j) s 2 2 1 1 s Ci Di eHi x 2 Hi j + Ni0 f eej Dj eHj xk 2 (Hj ) j . j

if P orR is payer. Calculating these payos can be the most compute-intensive part of the calculation. (Calculating discount factors is especially worrisome since it is beyond our control. We can amelioriate this by ensuring that there are as few redundant calculations as possible. Before reaching the evaluator, one usually creates a second structure out of the Bermudan structure. The second structure contains the following vectors: (i) the rst payment upon each exercise, (4.24a) iF irst[, 1, .., J] : i0 = if irst j for j = 1, ..., J

(ii) the discounted full payments, (4.24b) DisP ay[, 1, ..., n] : Ci Di = Ci D(ti ) for i = 1, ..., n

(iii) the discounted amount exchanged for the xed leg at each exercise, s (4.24c) DisExP rice[, 1, 2, ..., J] : Ni0 f eej Dj = Ni0 f eej D(tset ) j j j (where the + sign is for receivers, and the sign is for payers), (iv) the mean reversion function on each pay date and on each settlement date, H[, 1, 2, ..., n] : Hi = H(ti ) s Hset[, 1, 2, ..., J} : Hj = H(tset ) j

for j = 1, ..., J

(4.24d) (4.24e)

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

(v) nally the value at the exercise dates: (4.24f) [, 1, 2, ..., J) : j = (tex ) j for j = 1, ..., J.

For completeness, the structure also contains (vi) the number of exercises and number of paydates: (4.24g) (4.24h) J = number of exercises n = number of payments

The payo functions can be calculated entirely from these pre-calculated vectors. Besides making the code more ecient, this enhances the soundness of the code because neither the discount curve D(t) nor the model parmeters (t) and H(T ), nor the original Bermudan structure needs to be passed down any further into the evaluator. The only thing inputs needed for the core evaluation routine are the new structure, a pointer to the function which calculates the payos Pj,k from the new structure, the vector [, 1, .., J] of variances, and the dicretization variables x , Nx , y , and Ny .
18

(The [, 1, .., J] vector should be passed outside of the second structure because the evaluator should just pass the structure to the payo function without relying on whats inside; the [, 1, .., J] vector should also be passed inside the structure so that the payo can be constructed solely from information stored within the structure.) One other comment about eciency. Normally one calls the payo function with the entire vector x[0, 1, ..., 2mx ] and it returns the vector Pj,k for k = 0, 1, 2, ..., 2mx . For many deal types, the payo vector can be caculated more eciently than the individual payos. 5. Calibration. The calibration procedure consists of three steps. First is to characterize the deal by extracting its essential features. Second is to select a set of vanilla calibration instruments based on the characterization and an over-all calibration strategy. The last part is applying the algorithms that choose (t) and H(T ) to match the LGM and market prices of the calibration instruments. Careful inspection will show that only the characterization step depends on the exotic being a Bermudan; the remaining two steps depend only on the features extracted by the the characterization step. This means that to handle the calibration step for other deal types (callable inverse oaters, callable capped oaters, callable range notes, ...) we just need to re-write the deal characterization part of the routine. 5.1. Deal characterization. We characterize deals by three quantities for each exercise. The rst is the exercise (notication) date itself, (5.1a) tex j for j = 1, 2, ..., J.

The second quantity is the length of the swap obtained upon exercise, (5.1b)
j

= tn tset j

for j = 1, 2, ..., J.

The last last piece of information determines how far the underlying is from being at-the-money for each exercise. There are several dierent measures of this distance. The one I prefer is to determine the parallel shift j needed, (5.2) D(ti ) D(ti )e j ti

so that todays value of the j th payo is at-the-money. Suppose that if the deal is exercised at tex . The receiver gets j (5.3a) (5.3b) and in return pays (5.3c) (5.3d) Ni0 + f eej j Ni0 f eej j paid at tset j paid at tset j if P orR is reciever, if P orR is payer. Ci0 rf pj Ci paid at ti0 , paid at ti for i = i0 + 1, ..., n,

Here we are using the abbreviation i0 = if irst for the rst paydate after settlement. Clearly this payo is at j the money when (5.4)
set (Ci0 rf pj ) D(ti0 )e j (ti0 tj ) +

i=i0 +1 j

n X

set Ci D(ti )e j (ti tj ) = Ni0 f eej D(tset ). j j

The idea behind characterization is that the most natural set of vanilla instruments for representing the Berumdan are the swaptions (one for each exercise date) which
19

(a) have the same exercise date, (b) have the same length of the underlying swap, (c) are at-the-money for the same parallel shift of the yield curve. Using these swaptions in calibration implies that our vega risks will be to these swaptions, and in the normal course of events, our Bermudan would then be hedged by a linear combination of these swaptions. This is eminently reasonable for Bermudan options on bullet swaps (and like-shaped underlyings). It is less reasonable for Bermudan options on amortizing swaps, and perhaps for zero coupon swaps. Would a 10 year option on a 20 year amorting swap be better represented by an 10 into 20 bullet swaption, or a 10 into 10 bullet? In appendix A we develop a more robust method of characterizing the option based on the duration and convexity of the payo. This method should be used for options on amortizers or zero-coupon swaps. Here we calibrate based on the above characterization. In Appendix A we point out the dierences needed for amortizers. 5.2. Calibration instruments. 5.2.1. Diagonal swaptions. Most decent calibration methods use the Bermudans diagonal swaptions, which we construct here. For each of the exercise dates tex , let Tjset be the currencys standard spot date: j (5.5) Tjset = SpotDate(tex , ccy) j for j = 1, 2, ..., J.

Let tth and tact be the theoretical and actual end dates of the Bermudan. The diagonal swaptions are Berm end the swaptions with exercise date tex , start date Tjset , and the end date Tn for j = 1, 2, ..., J. It remains to j diag of these swaptions and to construct the payments. choose the strike Rj Let us create a standard xed leg and oating leg schedules based on the theoretical end date tth : end (5.6a) (5.6b) T0 , T1 , T2 , ..., Tn = tact . end
f f f f T0 lt , T1 lt , T2 lt , ..., Tmlt = tact . end

set The longest diagonal swap is j = 1, which starts at T1 . The schedule should be carried back far enough so f lt that T0 and T0 are on or before this start date:

(5.7)

set T0 T1 < T1 ,

f f set T0 lt T1 < T1 lt

For each swaption j, let i1 be the index of the rst pay date after Tjset : j (5.8) Ti1 1 Tjset < Ti1 . j j j R j diag
diag i R j diag 1 + n Rj

Then the xed leg payments for swaption j are: (5.9a) (5.9b) (5.9c) Here, (5.9d) i = cvg(Ti1 , Ti , x) for i = 1, 2, .., n at Ti at Ti at Tn for i = i1 j for i = i1 + 1, i1 + 2, ..., n 1 j j for i = n

is the xed leg day count fraction for the full periods, and (5.9e) j = cvg(Tjset , Ti1 , x) j
20

is the day count fraction for the rst period, which is may be a stub. (The argument x means to used the xed legs day count basis). We now constuct the oating leg, converting the basis spreads from the oating legs frequency and basis to the xed legs frequency and basis. Consider a oating leg that starts at, say, k0 . This oating leg is equivalent to (5.10a) (5.10b) 1 k bsk
f at Tk0lt f at Tk lt

for k = k0 + 1, ..., m.

f lt f Here, bsk is the basis spread for the period beginning at Tk1 and ending at Tk lt , and

(5.10c)

f lt f k = cvg(Tk1 , Tk lt , t)

for k = 1, 2, .., m

is the day count fraction for the full oating point period. We convert the basis spreads to the xed legs frequency and day count basis in the usual way. If the oating leg frequency is the shorter than, or equal to, the xed leg frequency, dene (5.11a) Si = P
kIi f k bsk D(Tk lt ) i D(Ti )

where k Ii are the oating leg intervals that are part of the ith xed leg interval: (5.11b) k Ii
f th if and only if Ti1 < Tk lt,th Tith .

If the oating leg frequency is longer than the xed leg frequency (this is rare), dene (5.12a)
f k bsk D(Tk lt ) Si = P iIk i D(Ti )

where i Ik are the xed leg intervals that are part of the k th oating leg interval: (5.12b) i Ik
f lt,th f if and only if Tk1 < Tith Tk lt,th .

We the j th swaption, we approximate the oating leg payments as being equivalent to (5.13a) (5.13b) (5.13c) 1 j Si i Si at Tjset at Ti at Ti for i = i1 j for i = i1 + 1, i1 + 2, ..., n j j

The net payments for the swaption are (5.14a) (5.14b) (5.14c) (5.14d) 1 diag j Rj Si diag i Rj Si diag 1 + n Rj Sn at Tjset at Ti at Ti at Tn
21

for i = i1 j for i = i1 + 1, i1 + 2, ..., n j j for i = n

We now choose the strikes of the diagonal swaptions. The strike swaption j is set so that the swaption is in the money at the same shift as the Bermudan:
diag Rj set Dj Dn e j (Tn Tj
set

(5.15a)

+ j Si1 Di1 e j j
j

set j (Ti1 Tj ) j

j Di1 e j where (5.15b)

set j (Ti1 Tj )

Pn

i=i1 +1 j

i Di e j (Ti Tj

Pn

i=i1 +1 j

i Si Di e j (Ti Tj
set )

set

set Dj = D(Tjset ),

Di = D(Ti ),

etc.

After constructing the diagonal swaptions, we obtain their market price via Blacks formula, n o Xn diag diag sw (5.16a) Mkt j Rj N (d1 ) Rj N (d2 ) , = j Di1 + i Di 1 j
i=ij +1 sw where Rj is the (break even) swap rate for the j th diagonal swap, P set Dj Dn + j Si1 Di1 + n 1 +1 i Si Di j j i=ij sw Pn , (5.16b) Rj = 1 + j Dij i=i1 +1 i Di
j

and where (5.16c)

d1,2 =

Here is the log normal volatility obtained from, for example, the GetVol function. 5.2.2. Row swaptions. Some calibration methods use the Bermudans row swaptions. Let tex be 1 set the earliest exercise date of the Bermudan, and let T1 be the corresponding spot date. The j th row swaption set is the swaption with start date T1 and end date Tj . Its equivalent payments are: (5.17a) (5.17b) (5.17c) (5.17d) 1 row 1 Rj Si row i Rj Si row 1 + j Rj Sj
set at T1 at Ti

diag sw log Rj /Rj 1 2 tex j p ex 2 . tj

for i = i1 1 for i = i1 + 1, i1 + 2, ..., j 1 1 1 for i = j

at Ti at Tj

Here the dates Ti , day count fractions j , i and equivalent basis spreads Si are the precisely the same quantities calculated for the diagonal swaptions. If the exercise date tex is too near today, say less than 3 months, then one should choose replace it with 1 the rst exercise date tex which is, say, at least 3 months from today. j set The diagonal swaptions are dened for j = imin , imin + 1, ..., n where Timin T1 is the shortest inteval which makes a decent swap (say 10 months). We choose the strike : P set set (T 1 T set ) set D1 Dj e 1 (Tj T1 ) + 1 Si1 Di1 e 1 i1 1 + j 1 +1 i Si Di e 1 (Ti T1 ) 1 1 i=i1 row , (5.18) Rj = P (T 1 T set ) set 1 Di1 e 1 i1 1 + j 1 +1 i Di e 1 (Ti T1 ) i=i 1
1

These strikes are all at the money at the same parallel shift 1 at the Bermudans rst payo for t1 . ex
22

sw where Rj is the (break even) swap rate for the j th row swap, P set D1 Dj + 1 Si1 Di1 + j 1 +1 i Si Di 1 1 i=i1 sw (5.19b) Rj = , Pj 1 Di1 + i=i1 +1 i Di 1
1

The market value of these swaptions are Xj row (5.19a) Mktj = 1 Di1 + 1

i=i1 +1 1

i Di

row sw Rj N (d1 ) Rj N (d2 ) ,

and where (5.19c)

d1,2 =

Again the implied vol needs to be obtained from, e.g., GetVol.

row sw log Rj /Rj 1 2 tex j p ex 2 . tj

5.2.3. Column swaptions. For the calibration strategies which use a column of swaptions, we choose the swaptions which have exercise date tex , start date Tjset , and end date Tiend where iend is the rst index j j j such that Tiend Tjset makes a decent swap (is at least, say, 10 months long). For each j = 1, 2, ..., J, the j equivalent payments for swaptions j is: (5.20a) (5.20b) (5.20c) (5.20d) 1 col j Rj Si col i Rj Si col 1 + i Rj Si at Tjset at Ti at Ti at Ti for i = i1 j for i = i1 + 1, ..., iend 1 1 j for i = iend j

Here the dates Ti , day count fractions j , i and equivalent basis spreads Si are the precisely the same col quantities calculated for the diagonal swaptions. We choose the strike Rj so that each swaption is at the money for the same parallel shift as the Bermudan,
set Dj Diend e j
set j (Tiend Tj ) j set Piend j (Ti1 Tj ) set j j + j Si1 Di1 e j j + i=i1 +1 i Si Di e j (Ti Tj ) j , set Piend j (Ti1 Tj ) set j j j Di1 e + i=i1 +1 i Di e j (Ti Tj ) j j

(5.20e)

col Rj =

sw where Rj is the (break even) swap rate for the j th column swap,

After constructing the column swaptions, we obtain their market price via Blacks formula, Xiend col j sw Rj N (d1 ) Rj N (d2 ) , i Di (5.21a) Mktcol = j Di1 + j 1 j
i=ij +1

(5.21b) and where (5.21c)

sw Rj

Piend j set j j Dj Diend + j Si1 Di1 + i=i1 +1 i Si Di j j = , Piend j j Di1 + i=i1 +1 i Di j

j

d1,2 =

Here is the log normal volatility obtained from, for example, the GetVol function.
23

col sw log Rj /Rj 1 2 tex 1 p ex 2 . t1

5.2.4. Caplets. For the calibration strategies which use caplets (oorlets), we choose the swaptions which have exercise date tex , start, start date Tjset , and end date Tjend , where the end date is either 3 months j or 6 months from the start date, depending on the currency. For each j = 1, 2, ..., J, the equivalent payments for caplet j is: (5.22a) (5.22b) where (5.22c) 1 cap 1 + j Rj bsj j = cvg(Tjset , Tjend , t) at Tjset at Tjend

for j = 1, 2, .., J

is the appropriate day count fraction. Here bsj is the basis spread for the oating rate set for start date cap Tjset . We choose the strike Rj so that each swaption is at the money for the same parallel shift as the Bermudan, end end set set Dj + 1 j bsj Dj e j (Tj Tj )
end j Dj e j (Tj
end T set ) j

(5.22d)

cap Rj

F where Rj RA is the (break even) rate for the j th diagonal swap is end set Dj 1 j bsj Dj F RA = , (5.23b) Rj end j Dj

After constructing the column swaptions, we obtain their market price via Blacks formula, cap end F (5.23a) Mktcap = j Dj Rj N (d1 ) Rj RA N (d2 ) , j

and where (5.23c) d1,2 =

cap F log Rj /Rj RA 1 2 tex j p ex 2 . tj

Here is the log normal caplet volatility obtained from, for example, the GetVol function. 5.3. Calibration to the diagonal swaptions. Having constructed the universe of possible calibration instruments, we now go through the calibration strategies and algorithms one by one. Pricing Bermudans accurately requires calibrating the model to the diagonal swaptions. For if our model doesnt correctly price the European swaptions that make up the Bermudan, how could we believe the price obtained for the Bermudan? In this section we present the strategies for calibrating on diagonal swaptions. These are: calibration to the diagonal with a constant mean revesion ; calibration to the diagonal with a known function H(T ); calibration to the diagonal with a linear (t), and calibration to the diagonal with a known (T ). For instruments other than Bermudans, it may be appropriate to calibrate to other series of vanilla instruments. So following are sections devoted to calibating on a series of caplets, to calibrating on a column of swaptions, and to calibrating to a row of swaptions Since the LGM model has two model parameters, (t) and H(T ), we can calibrate jointly to two distinct series of vanilla instrumetns. In the nal section we present calibration strategies which calibrate jointly to the diagonal swaptions plus another series of instruments. These are: calibration to the diagonal swaptions and a row swaptions, calibration to the diagonal swaptions and a column swaptions, and calibrating to the diagonal swaptions and to caplets. For completeness, we also calibrate on a row and column of swaptions, on a row of swaptions and to caplets.
24

5.3.1. Calibration to the diagonal swaptions with constant . For this calibration strategy, the mean reversion coecient is a user-supplied constant (Where to obtain good wake-up values for is discussed below. Empirically is usually between 1% and +5%.. 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 (5.24) H(T ) = 1 eT ,

without loss of generality, where T is measured in years. With H(T ) known, we compute (5.25a) (5.25b)
s Hj

1 eTj = = 1 eTi Hi = H(Ti ) = H(Tjset )

set

for j = 1, 2, ..., J for i = 1, 2, ..., n.

We now determine j = (tex ) for each j by calibrating to diagonal j. j Recall that if the j th diagonal swaption is exercised at its notication date tex , the payments are j (5.26a) (5.26b) (5.26c) (5.26d) 1 diag j Rj Si diag i Rj Si diag 1 + n Rj Sn at Tjset at Ti at Ti at Tn for i = i1 j for i = i1 + 1, i1 + 2, ..., n j j for i = n,

Under the LGM model, the value of this swaption is thus diag = j Rj Si1 Di1 N j j ! y + Hi1 j j p j ! n X y + Hi j diag p + i Rj Si Di N j 1 +1 i=ij ! ! y + Hn j y set p +Dn N Dj N p j j

(5.27a)

Vjdiag (0, 0)

where y is obtained by solving

n 2 X Hi1 y 1 Hi1 j 2 1 2 diag diag j j (5.27b) j Rj Si1 Di1 e + i Rj Si Di eHi y 2 Hi j j j i=i1 +1 j

+Dn eHn y and where we have used (5.27c)

set Hi = Hi Hj = H(Ti ) H(Tjset )

2 1 Hn j 2

set = Dj ,

25

! y + Hi1 j j p (5.28) j ! n X y + Hi j diag p + Hi i Rj Si Di G j 1 +1 i=ij ! y + Hn j p +Hn Dn G . j p We can use a global Newtons scheme to compute the value of j which sets the theoretical price to the market price: diag p Vjdiag (0, 0) = Hi1 j Rj Si1 Di1 G j j j j (5.29) Vjdiag (0, 0) = Mktdiag j Repeating for all j gives us (0) = 0 and (tex ) for j = 1, 2, ..., J. We use piecewise linear interpolation to j get values of (t) at other values of t. It should be noted that evaluating the Bermudan does not require (t) at any other dates. Re-scaling H(T ) and (t). At this point we have both H(T ) and (t). Many rms nd it convenient to use a standard scaling for H(T ) and (t), to aid intuition if for no other reason. One can use the invariances (5.30a) (5.30b) H(T ) H(T ) + C, H(T ) KH(T ), (t) (t) (t) (t)/K 2

We also have a formula for the derivative

to re-scale these quantities, if desired. For example, many people choose to set H(0) = 0 and H(tend ) = tend , where tend is the nal pay date of the deal in years. p Aside: Initial guess. An accurate initial guess for j can be found from the equivelent vol formula. This yields s j (5.31) ex Rsw Rdiag tj j j P j Di1 + n 1 +1 i Di i=ij j Pn diag diag j Rj Si1 Di1 Hi1 + i=i1 +1 i Rj Si Di Hi + Dn Hn j j j
j

where is the swaptions implied vol from the marketplace. Aside: Global Newtons method for one parameter ts. Suppose one is trying to solve f (z) = target

(5.32a)

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 ) (5.32b) 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 the 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.
26

Aside: Infeasible market prices. Since (5.33) (t) = Z

t

2 (t0 )dt0 ,

clearly (t) must be an increasing function of t : (5.34) 0 = (0) 1 2 J .

Since each j is calibrated seperately, it may happen that j < j1 . (In practice this happens very, very rarely, but it does happen). One should test to see that the condition j j1 is true after each j is found, and when this condition is violated, one should replace j by j1 , its minimum feasible value: (5.35) j j1 if j < j1 .

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. 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 . We should plan to have a matrix of wake-up values, perhaps by currency, for this strategy. I can obtain the current matrix. 5.3.2. Calibration to the diagonal swaptions 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 ), and piecewise linear interpolation is used between nodes. Piecewise linear interpolation is equivalent to assuming that all shifts of the forward rate curve are by piecewise constant curves. 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 tex , tex , . . . , tex . As above, one adds the point (0) = 0, one ensures that 1 2 J the j = (tex ) are increasing, and one re-scales (t), H(T ) to taste. If one needs (t) for other values of t, j one uses piecewise linear interpolation. Origin of the H(T ). Suppose one had the set of 30 NC 20, 30 NC 15, 30 NC 10, 30 NC 5 and 30 NC 1 Bermudan swaptions. 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 (5.36a) 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.
27

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: (5.36b) H 0 (T ) = e3 (T20 T ) e4 (T30 T20 ) e 1 4 (T30 T20 ) e4 (T30 T20 ) 1 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. I will obtain the current t)s to use for the wakeup value for this strategy. In general, if Tn is the co-terminal point and T0 , T1 , . . . , Tn1 are the no call points, then H(T ) is: (5.37) H(T ) =
n n n X ek (Tk Tk1 ) 1 Y ej (Tj T ) 1 Y i (Ti Ti1 ) e ei (Ti Ti1 ) j k i=j+1 k=j+1 i=k+1

for Tj1 T Tj

5.3.3. Calibration to the diagonal swaptions with linear (t). This is an idea pioneered by Solomon brothers. Let us use a constant local volatility. Then (5.38a) (t) = Z
t

2 dt0 = 2 t 0

is linear. By using the invariance (t) (t)/K 2 , H(T ) KH(T ) we can choose 0 to be any arbitrary constant without aecting any prices. So we choose (5.38b) (t) = 2 t, 0

where t is measured in years, and the dimensionless constant 0 is, say, (5.38c) 0 = 102 .

We use the second invariance to set Hn = H(Tn ) = 0. We shall calibrate the diagonal swaptions to determine set the values of H(T ) on the settlement dates, Hj = H(Tjset ). For other values of T, we assume that H(T ) is piecewise linear: (5.39a) (5.39b) (5.39c)
set H(T ) = H1 + set T T1 set set for T Tjset , set set H2 H1 T2 T1 set T Tj1 set set set set Hj1 for Tj1 T Tjset H(T ) = Hj1 + set set Hj Tj Tj1 set T TJ set set set H(T ) = HJ + for TJ T set Hn HJ Tn TJ

28

where

The value of the j th diagonal swaption can be written as ! ! n X q + hij1 j q + hi j diag diag diag p p i Rj Si Di N (5.40a) Vj (0, 0) = j Rj Si1 Di1 N + j j j j i=i1 +1 j ! ! q + hset j q j set p +Dn N p Dj N j j hi = H(Ti ) H(Tn ) hset = H(Tjset ) H(Tn ) j for i = 1, 2, ..., n for j = 1, 2, ..., J

(5.40b) (5.40c)

and where q is determined implicitly by

n X hi1 q 1 h21 j 1 2 2 i diag diag j (5.40d) j Rj Si1 Di1 e j + i Rj Si Di ehi q 2 hi j j j i=i1 +1 j set +Dn = Dj ehj
set

q 1 h2 j 2 i

set The last swaption J only depends on hset and on hi for the paydates after TJ . Since Hn = 0, these J set values are given in terms of hJ

(5.41a)

hi = hset J

Tn T set Tn TJ

for i i1 . J

analytically. Now suppose that we have calibrated all the swaptions after j to obtain hset , hset , ..., hset . Since we are j+1 j+2 J set using piecewise interpolation, this determines H(T ) for all T Tj+1 . We now calibrate on swaption j to set obtain hset . The value of this swaption depends on hset and any paydates between Tjset and Tj+1 : j j (5.42a) hi =
set Tj+1 T set T Tjset set hj + set h set Tj+1 Tjset Tj+1 Tjset j+1

There is a unique value of hset which matches the LGM price to the market price for the last swaption. This J can be easily found by a global Newtons scheme, since we have the derivative ! diag Tn Ti1 q + hi1 J 1 VJ diag J p p J 1 = J RJ Si1 (5.41b) set DiJ G J Tn TJ J hset J J ! n1 T T X q + hi J n i diag p + i RJ Si set Di G Tn TJ J i=i1 +1 J ! q + hset J set pj DJ G J

for i1 i i1 j j+1

Since hi for i i1 are known from previous steps in the calibration, the only unknown parameter is hset . A j+1 j global Newtons scheme can be used to eciently determin thunique value of this parameter which matches
29

the j th swaptions LGM price to its market price. Note that the derivative of the value with respect to hj is ! set diag Tj+1 Ti1 q + hi1 j 1 Vj j diag J p p = j Rj Si1 D 1G (5.42b) set J Tj+1 Tjset iJ j hset j j ! i1 1 j+1 T set Ti X q + hi j j+1 diag p i Rj Si Di G + set Tj+1 Tjset j i=i1 +1 j ! q + hset j set pj Dj G j Continuing, we can calibrate on the swaptions one at a time (backwards) to obtain
set set set H1 , H2 , ..., HJ , Hn

(5.43)

on the dates tset , ..., tset , tn . One uses linear interpolation/extrapolation to get H(t) at other values of t. Of 1 J course, after nding the (t), H(T ), one can use the invariances to scale them to taste. Infeasible values. In deriving the swaption formulas, we assumed that H(T ) was an increasing function set set set of T . Since we are calibrating the Hj s seperately, it may happen that Hj may exceed Hj+1 . (In practice, set this has never happened to my knowledge. Still one must be prepared.) After each Hj is found, one should check to see that (5.44)
set set Hj Hj+1 .

If this condition is violated, one should reset Hj1 = Hj . This means the j th swaption would not match its market price exactly. Instead it would be the closest feasible price. Initial guess. The equivalent vol techniques yields v u ex sw diag u t R R t j j j (5.45) j Pn diag diag set j Rj Si1 Di1 hi1 + i=i1 +1 i Rj Si Di hi Dj hset j j j j j Pn 1 + j Dij i=i1 +1 i Di
j

where is the swap rate and is the swaptions implied vol from the marketplace. Since this is linear in the hs, one can solve to get a decent initial guess for hset . j 5.3.4. Calibration to diagonal swaptions with prescribed (t). The preceding calibration procedure did not depend on (t) being linear; it just depended on (t) being known. So suppose that (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. 5.4. Calibration to caplets. There are many exotic structures which are more naturally priced and hedged in terms of caplets. Autocaps and revolvers, for example. Even though these calibration methods shouldnt be used for pricing Bermudans, we present them here for completeness. We will also make use of these calibration methods later for joint calibrations to the diagonal swaptions and caplets. For each j = 1, 2, ..., J , the equivalent payments for caplet j are: (5.46a) (5.46b) 1 cap 1 + j Rj bsj
30

sw Rj

at Tjset at Tjend .

Here (5.46c) j = cvg(Tjset , Tjend , t) for j = 1, 2, .., J

is the appropriate day count fraction and bsj is the basis spread for the oating rate set for start date Tjset . Caplet and oorlets are one period swaptions. If we specialize the swaption formulas 3.7a, 3.7b to one period, we nd that the LGM price for the caplet is cap F end end (5.47a) Vjcap (0, 0) = Dj 1 + j Rj bsj N (dlgm ) Dj 1 + j Rj RA bsj N dlgm 1 2
F where Rj RA is the break-even caplet rate F Rj RA

(5.47b)

end set Dj 1 j bsj Dj = , end j Dj

and where dlgm and dlgm are given by 1 2 cap end 1 + j Rj bsj set 2 F RA 1 Hj Hj j log 2 1 + j Rj bsj end p = . set j Hj Hj
end Hj = H(Tjend ),

(5.47c) Here, (5.47d)

dlgm 1,2

set Hj = H(Tjset ),

j = (tex ). j

We observe this is Blacks formula for a European option on an asset with forward price, cap (5.48a) F = 1 + j Rj bsj , with strike (5.48b) F set end K = 1 + j Rj RA bsj = Dj /Dj ,

5.4.1. Calibration to caplets with constant mean reversion . For this calibration strategy, the mean reversion coecient is a user-supplied constant. 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 to set (5.49a) H(T ) = 1 eT ,

and with settlement date Tjend . Suppose we use an implied volatility routine to nd the implied (price) vol cap,price which matches this caplet to its market value. Then j end q set j = cap,price tex . (5.48c) Hj Hj j

without loss of generality, where T is measured in years. With H(T ) known, matching the caplets to their market price requires q cap,price tex j (5.49b) j = for j = 1, 2, .., J end set Hj Hj
31

This determines (t) at the exercise dates tex , tex , ..., tex . Again, it may happen that j < j1 for some j, 1 2 J in which case we need to make the replacement (5.50) j j1 if j < j1 .

As usual, we append (0) = 0 and use piecewise linear interpolation to obtain (t) at other dates. Having found (t) and H(T ), one can use the invariances to normalize them according to test. 5.4.2. Calibration to caplets with H(T ) specied. The above calibration procedure does not depend on being constant. It depends only on H(T ) being known. If H(T ) is an externally supplied function, then we can carry out the same calibration to obtain (t). 5.4.3. Calibration to caplets with linear (t). For this calibration procedure, we assume the local volatility is constant Z t (5.51a) (t) = 2 dt0 = 2 t. 0
0

is linear. By using the multiplicative invariance (t) (t)/K 2 , H(T ) KH(T ) we can choose (5.51b) where t is measured in years and 0 is, say, (5.51c) 0 = 102 . (t) = 2 t, 0

Matching the caplets to their market prices requires (5.52)

end Hj

set Hj

cap,price j = 0

for j = 1, 2, .., J.

We now use the additive invariance to set H(Tn ) = 0 and take H(T ) to be piecewise linear (5.53a) (5.53b) (5.53c)
set T T1 set set for T Tjset , set T set H2 H1 T2 1 set T Tj1 set set set set Hj1 for Tj1 T Tjset H(T ) = Hj1 + set set Hj Tj Tj1 set T TJ set set set H(T ) = HJ + for TJ T set Hn HJ Tn TJ set H(T ) = H1 +

Starting at the last caplet j = J, we see that we must choose (5.54a)

set HJ =

set cap,price Tn TJ J end T set 0 TJ J

set set set since H(Tn ) = 0. Suppose our calibration procedure has produced Hj+1 , Hj+2 , ..., HJ , and Hn . We now set end set nd Hj . First, if Tj Tj+1 , then

(5.54b)

set set Hj = Hj+1

set cap,price Tj+1 Tjset j 0 Tjend Tjset

set if Tjend Tj+1 .

32

set On the other hand, if Tjend > Tj+1 , then H(Tjend ) is in the already-calibrated region of the curve, and can set set be found by piecewise linear interpolation on Hj+1 , Hj+2 , ..., Hn . In this case, we use

(5.54c)

set end Hj = Hj

cap,price j 0

set if Tjend > Tj+1 .

set set to set Hj . Continuting backwards in this way, we obtain Hj at all the settlement dates j. Having (t) and H(T ), we can now normalize them to taste.

5.4.4. Calibration to caplets with prescribed (t). The above procedure for determining H(T ) did not depend on (t) be linear in t; it only relied on (t) being a known function. Suppose that (t) is an externally supplied function. Then we can use the above procedure to nd H(T ) provided we make the replacement p ex tj cap,price j (5.55) cap,price q j 0 (tex ) j

5.5. Calibration to a column of swaptions. Recall that tthe equivalent payments for the j th column swaptions j are (5.56a) (5.56b) 1 col j Rj Si col i Rj Si col 1 + i Rj Si at Tjset at Ti at Ti at Ti for i = i1 j for i = i1 + 1, ..., iend 1 1 j for i = iend j

(5.56c) (5.56d)

Under the LGM model, the value of this swaption is thus col = j Rj Si1 Di1 N j j ! y + Hi1 j j p j ! iend j X col y + Hi j p + i Rj Si Di N j i=i1 +1 j ! ! y + Hiend j y set p j +Diend N Dj N p j j j
end

(5.57a)

Vjcol (0, 0)

where y is obtained by solving

ij 2 X col Hi1 y 1 Hi1 j 2 1 2 coil j j (5.57b) j Rj Si1 Di1 e + i Rj Si Di eHi y 2 Hi j j j i=i1 +1 j

+Diend e j and where we have used (5.57c)

2 Hiend y 1 Hiend j 2 j j

set = Dj ,

set Hi = Hi Hj = H(Ti ) H(Tjset )

33

diag These formulas are identical to the formulas for the diagonal swaptions, provided one replaces Rj with row Rj and replaces n with iend for each swaption j. With a little tinkering, one can use the same software j to calibrate each column swaption as used for the corresponding diagonal swaption. This gives us the methods

Calibration to a column of swaptions with constant mean reversion Calibration to a column of swaptions with H(T ) specied Calibration to a column of swaptions with linear (t) Calibration to a column of swaptions with prescribed (t). Written properly, these routines should work with an arbitrary iend , so one does not have to limit oneself j to a column of swaptions. Instead one can use any sequence of swaptions which has an increasing set of exercise dates tex and settlement dates Tjset . j 5.6. Calibration to a row of swaptions. Recall that the j th row swaption is the swaption with start set date T1 and end date Tj . Its equivalent payments are: (5.58a) (5.58b) (5.58c) (5.58d) 1 row 1 Rj Si row i Rj Si row 1 + j Rj Sj at T set at Ti at Ti at Tj

for i = i1 for i = i1 + 1, i1 + 2, ..., j 1 for i = j.

Here the dates Ti , day count fractions j , i and equivalent basis spreads Si are the precisely the same quantities calculated for the diagonal swaptions. We also abbreviate i1 = i1 for the index of the rst paydate 1 set after T1 . Under the LGM model, the value of the j th row swaption is row = j Rj Si1 Di1 N ! y + Hi1 ex p ex ! j X row y + Hi ex p + i Rj Si Di N ex i=i1 +1 ! ! y + Hj ex y set p D N p +Dj N ex ex

(5.59a)

Vjrow (0, 0)

where y is obtained by solving

j X row row 2 1 2 1 (5.59b) j Rj Si1 Di1 eHi1 y 2 Hi1 ex + i Rj Si Di eHi y 2 Hi ex i=i1 +1

+Dj eHj y

2 1 Hj ex 2

= Dset ,

and where we have used (5.59c)

set Dset = D(T1 ),

ex = (tex ), 1

set Hi = Hi H set = H(Ti ) H(T1 ),

Since all these swaptions have the same exercise date, they depend only on a single value of (t), namely ex . It makes no sense to calibrate (t) from these swaptions. This leaves two natural methods for calibrating a row of swaptions: Calibration to a row of swaptions with linear (t)
34

 Calibration to a row of swaptions with prescribed (t). In the rst case, we can use the multiplicative invariance to set (5.60a) ex = 2 tex 0 1

without loss of generality, where 0 = 102 . This puts us in the second case where ex is prescribed as an input. Since ex is known, we only need to nd H(T ) via calibration. We use the second invariance to set H set = 0, and prescribe H(T ) to be piecewise linear with nodes at the end dates Ti1 , Ti1 +1 , ..., Tn :
set T T1 for T Ti1 , set Hi1 Ti1 T1 T Ti1 H(T ) = Hi1 + (Hi Hi1 ) Ti Ti1 T Tn1 (Hn Hn1 ) H(T ) = Hn1 + Tn Tn1

(5.60b) (5.60c) (5.60d)

H(T ) =

for Ti1 T Ti ,
set for Tn1 T

i = i1 + 1, .., n 1

We note that the rst swaption j = i1 depends only on H(T ) that are determined by Hi1 . A global Newton scheme suces to nd Hi1 by matching this swaption against its market value. The next swaption depends on the same H(T ) values as before, along with one new value, Hj = H(Tj ) with j = i1 + 1. Again Hj can 1 be found by calibrating the j th row swaption. Iterating, we can determine H at all the nodes by calibrating to successive swaptions. Again, its concievable that one of these Hj is not increasing. In that case we have to replace it to ensure that H(T ) is non-decreasing: (5.61) Hj Hj1 if Hj < Hj1 .

5.7. Calibraton to two series of vanilla instruments. Since the LGM model has two model parameters, (t) and H(T ), we can calibrate to two series of vanilla instruments. Following are the most popular strategies 5.7.1. Calibration to diagonal swaptions and a row of swaptions. Recall that a row of swaptions set is a set of swaptions, all with the same exercise date tex and same start date T1 , but with varying end 1 dates. We use the multiplicative invariance to set (5.62) ex = 2 tex 0 1

without loss of generality, where 0 = 102 . Since this is the only value of (t) used by the row swaptions, we can now use the calibration to a row of swaptions with prescribed (t) routine described above to nd H(T ). Knowing H(T ), we can use the calibration to the diagonal swaptions with H(T ) specied described above to nd (t). At this point we can normalize (t), H(T ) to taste. After calibrating to a row of swaptions to determine H(T ), one does not have to use the diagonal swaptions to nd (t). Instead one could calibrate on the caplets or a column of swaptions. This gives us the methods 5.7.2. Calibration to caplets and a row of swaptions. After calibrating to the row of swaptions to determine H(T ), one can use the calibration to caplets with H(T ) specied routine described above to nd (t). 5.7.3. Calibration to a column and row of swaptions. After calibrating to the row of swaptions to determine H(T ), one can use the calibration to a column of swaptions with H(T ) specied routine described above to nd (t).
35

5.7.4. Calibration to diagonal swaptions and a column of swaptions. This calibration method simultaneously calibrating the j th diagonal and the j th column swaption to determine both j = (tex ) and j H(Tjset ). One starts at the last pair, j = J, and works backward. Recall that the j th diagonal and j th column swaption share identical exercise dates tex and settlement j dates Tjset . They dier only in the end date: the diagonal swaption goes all the way to Tn , while the column swaption stops at Tiend . j For the last pair of swaption, j = J, we usually have iend = n, and the two swaptions are identical. Even J if they are not identical, we should exclude the last column swap as being too similar to the diagonal swap. The value of the j th diagonal swaption can be written as (5.63a) Vjdiag (0, 0) = j
diag Rj

Si1 Di1 N j j

where

! ! n X q + hij1 j q + hi j diag p p i Rj Si Di N + j j i=i1 +1 j ! ! q + hset j q j set p +Dn N p Dj N j j

n X hi1 q 1 h21 j 1 2 2 i diag diag j (5.63b) j Rj Si1 Di1 e j + i Rj Si Di ehi q 2 hi j j j i=i1 +1 j set +Dn = Dj ehj
set

q 1 h2 j 2 i

Similarly, the value of the j th column swaption is (5.63c) Vjcol (0, 0) col = j Rj Si1 Di1 N j j ! u + hi1 j j p j end ! ij X col u + hi j p + i Rj Si Di N j 1 +1 i=ij ! ! u + hiend j u + hset j j j set p p +Diend N Dj N j j j
end

where

ij X col hi1 u 1 h21 j 1 2 2 i col j j (5.63d) j Rj Si1 Di1 e + i Rj Si Di ehi u 2 hi j j j i=i1 +1 j

+Diend e j Here we are using (5.63e) (5.63f) hi set hj = H(Ti ) H(Tn ) = H(Tjset ) H(Tn )
36

hiend u 1 h2end j 2
j i j

set = Dj ehj

set

u 1 h2 j 2 i

for i = 1, 2, ..., n for j = 1, 2, ..., J

We use piecewise linear interpolation for H(T ), with nodes at the start dates Tjset and the nal end date Tn : (5.64a) (5.64b) (5.64c)
set T T set T2 T set set H1 + set 1 set H2 set set T2 T1 T2 T1 set T Tj1 Tjset T set set Hj1 + set H(T ) = set set set H Tj Tj1 Tj Tj1 j

H(T ) =

for T Tjset ,
set for Tj1 T Tjset

H(T ) =

set T TJ Tn T set set HJ + T T set Hn Tn TJ n J

set for TJ T

Without loss of generality, we choose H(T ) to be 0 at the nal pay date. This means that H(T ) and h(T ) are identical in the above formulas. We use the second invariance to set the slope of H(T ) to be 1 in the nal interval: (5.65)
set set set HJ = H(TJ ) = TJ Tn

Hn = H(Tn ) = 0

set This determines all the values of H(T ) for T TJ , so the last swaption depends only on one unknown parameter, J = (tex ). We use our standard global Newton scheme to determine J . J set Suppose that we have already found H(T ) for T Tj+1 for some j. We now nd Hj = H(Tj ) and j th th by matching the j diagonal and j column swaption. These swaptions depend on j = (tex ) (which is j set unknown), H(T ) for T Tj+1 (which is known), and on

(5.66)

H(T ) =

set T Tjset Tj+1 T set Hj + set H set set Tj+1 Tjset Tj+1 Tjset j+1

for Tjset T Tjset ,

set set (which is determined by Hj , which is unknown). So there are two parameters to t, Hj and j and two swaption values to set to their market prices. We will use a global multi-factor Newtons method to nd these parameters. This requires dierentiating the swaption values:

(5.67a)

Vjdiag diag p = hij1 j Rj Si1 Di1 G j j j

i=ij +1

! q + hij1 j p j ! ! n X q + hset j q + hi j j diag set set p p + hj Dj G hi i Rj Si Di G j j 1 ! u + hi1 j j p j ! iend j X col u + hi j p + hi i Rj Si Di G j i=i1 +1 j ! ! u + hiend j u + hset j j set pj p +hiend Diend G hset Dj G j j j j j
37

(5.67b)

Vjcol col p = hi1 j Rj Si1 Di1 G j j j j

and (5.67c)
set diag Tj+1 Ti1 1 Vj j diag p = set Si1 Di1 G set set j Rj j j Tj+1 Tj j Hj

(5.67d)

From 5.63b, 5.63d, we deduce that ! ! n X q + hij1 j q + hi j diag diag p p j R i Rj Si Di G (5.68a) j Si1 Di1 G + j j j j i=i1 +1 j ! ! q + hset j q + hn j j set p p +Dn G = Dj G , j j col (5.68b)Rj Si1 Di1 G j j j u + hij1 j p j !
iend j

! set u + hi1 j Tj+1 Tij1 1 Vjcol j col p p 1 1 set = T set T set j Rj Sij Dij G j Hj j j+1 j end ( ) ! ij set X col Tj+1 Ti u + hi j p + max , 0 i Rj Si Di G set Tj+1 Tjset j i=i1 +1 j ( set ) ! ! Tj+1 Tiend u + hiend j u + hset j j j j set p p + max , 0 Diend G Dj G set j Tj+1 Tjset j j

! q + hij1 j p j set ! Ti Tj+1 set X Tj+1 Ti q + hi j diag p + i Rj Si Di G set Tj+1 Tjset j i=i1 +1 j ! q + hset j set pj Dj G j

i=i1 +1 j

col i Rj Si Di G +Diend G j

This shows that the four derivatives can be wriiten as: (5.69a)

u + hiend j pj j

u + hi j p j

! ! =
set Dj G

u + hset j pj j

Vjdiag diag set p = Hij1 Hj j Rj Si1 Di1 G j j j

! q + hij1 j p j ! n X diag q + hi j set p + Hi Hj i Rj Si Di G j i=i1 +1 j ! q + hn j set p + Hn Hj Dn G j

38

(5.69b)

Vjcol set col p = Hij1 Hj j Rj Si1 Di1 G j j j

and (5.69c)

diag 1 Vj p set = min j Hj

(5.69d)

1 Vjcol p set = min j Hj

! q + hij1 j p set Tj+1 Tj j ( ) ! n set X Ti Tj q + hi j diag p min , 1 i Rj Si Di G set Tj+1 Tjset j i=i1 +1 j ! q + hn j p Dn G j ( Ti1 Tjset j diag , 1 j Rj Si1 Di1 G set j j ) ! u + hi1 j j p , 1 j Si1 Di1 G set j j Tj+1 Tjset j ( ) ! iend j X col Ti Tjset u + hi j p min , 1 i Rj Si Di G set Tj+1 Tjset j i=i1 +1 j ( ) ! Tiend Tjset u + hiend j j pj min , 1 Diend G set j Tj+1 Tjset j Tij1 Tjset
col Rj

! u + hi1 j j p j end ! ij X col u + hi j set p + Hi Hj i Rj Si Di G j 1 +1 i=ij ! u + hiend j j set p + Hiend Hj Diend G j j j

p j , with the set Hiend Hj value of the diagonal swaption going up faster than the column swaption by roughly the ratio j set . Hn Hj set The value of the diagonal and column swaption both increase at roughtly the same rate as Hj decreases. set With the derivatives in hand, we can use a global Newton scheme to nd j and Hj . We require that Since H(T ) is increasing, the value of the diagonal and column swaption both go up with (5.70) j < j+1 ,
set set Hj < Hj+1 .

set set Suppose we start our search at the corner j = j+1 , Hj = Hj+1 . If Vjdiag > Mktdiag , then we decrease j , j diag diag diag set set = Mktj . If Vj < Mktdiag , then we decrease Hj , keeping keeping Hj at Hj+1 , until we match Vj j j = j+1 until Vjdiag = Mktdiag . Let us now imagine decreasing both j and Hj on a trajectory such that j Vjdiag remains equal to Mktdiag . On this trajectory Vjcol increases. So if we start with Vjcol Mktcol , then a j j unique solution exists. We use a global Newton scheme to nd it. Alternatively, if Vjcol > Mktcol , we can do j no better than keeping the current j and Hj . (These are the parameters which t the diagonal swaptions exactly, and come as close as possible to tting the column swaptions). set set Once weve found both j , Hj we step back to j 1, etc. until weve found j , Hj for j = 1, 2, ..., J. In the usual way, we use piecewise linear interpolation between the known values of (t) and H(T ), and use the invariances to rescale , H to taste.

39

5.7.5. Calibration to two columns of swaptions. The above calibration techniques does not depend on one set of swaptions being the diagonal swaptions. It just relies on there being J swaption pairs, with both members of each pair sharing the same exercise date tex and start date Tjset , and having distinctly j dierent end dates. For some exotics, like MBS traunches, it makes more sense to calibrate on two columns of swaptions, say the 1y and 10y tenors. With only trivial modications, the algorithm described above will calibrate these more general sets of swaption pairs.. 5.7.6. Calibration to diagonal swaptions and caplets. This calibration method simultaneously calibrates the j th diagonal swaption and the j th caplet to determine both j = (tex ) and H(Tjset ). As in j the preceding case, one starts at the last pair, j = J, and works backward. For parameter stability, we do not calibrate to the nal caplet, since in our view it may not be suciently dierent from the last diagonal swaption. Recall that the j th diagonal swaption and j th caplet share identical exercise (xing) dates tex and j settlement dates Tjset . They dier only in the end date: the diagonal swaption goes all the way to Tn , while the caplet stops at Tjend . As above, the value of the j th diagonal swaption can be written as Vjdiag (0, 0) ! q + hij1 j p = j Si1 Di1 N j j j ! n X q + hi j diag p + i Rj Si Di N j i=i1 +1 j ! ! q + hset j q j set p +Dn N p Dj N j j
diag Rj

(5.71a)

where

n X hi1 q 1 h21 j 1 2 2 i diag diag j (5.71b) j Rj Si1 Di1 e j + i Rj Si Di ehi q 2 hi j j j i=i1 +1 j set +Dn = Dj ehj
set

q 1 h2 j 2 i

Recall that the value of the j th caplet matches its market value when q cap,price tex j j = hend hset j j

(5.71c)

where cap,price is the implied price vol dened earlier. Here we are using j (5.72a) (5.72b) (5.72c) hi = H(Ti ) set hj = H(Tjset ) hend = H(Tjend ) j H(Tn ) H(Tn ) H(Tn ) for i = 1, 2, ..., n for j = 1, 2, ..., J for j = 1, 2, ..., J

We use piecewise linear interpolation for H(T ), with nodes at the start dates Tjset and the nal end date
40

Tn : (5.73a) (5.73b) (5.73c)

set T2 T T T set set set H1 + set 1 set H2 set set T2 T1 T2 T1 set T Tj1 Tjset T set set Hj1 + set H(T ) = set set set H Tj Tj1 Tj Tj1 j

H(T ) =

for T Tjset ,
set for Tj1 T Tjset

H(T ) =

set T TJ Tn T set set HJ + T T set Hn Tn TJ n J

set for TJ T

Without loss of generality, we choose H(T ) to be 0 at the nal pay date. Then H(T ) and h(T ) are equal in the above formulas. We use the second invariance to set the slope of H(T ) to be 1 in the nal interval: (5.74)
set set set HJ = H(TJ ) = TJ Tn

Hn = H(Tn ) = 0

set This determines all the values of H(T ) for T TJ , so the last swaption depends only on one unknown ex parameter, J = (tJ ). We use our standard global Newton scheme to determine J . set Suppose that we have already found H(T ) for T Tj+1 for some j. Consider the j th diagonal swaption. ex set It depends on j = (tj ) and H(T ) for T Tj+1 and on

(5.75)

H(T ) =

set T Tjset Tj+1 T set + set H set set T set Hj Tj+1 Tj+1 Tjset j+1 j

for Tjset T Tjset .

As before, the dierentiating the diagonal swaption value eventually yields ! q + hij1 j Vjdiag diag set p p = Hij1 Hj (5.76a) j Rj Si1 Di1 G j j j j ! n X diag q + hi j set p + Hi Hj i Rj Si Di G j 1 +1 i=ij ! q + hn j set p + Hn Hj Dn G j (5.76b)
diag 1 Vj p set = min j Hj

The value of the swaption increases as Fitting the caplet requires (5.77a)

! q + hij1 j p set Tj+1 Tj j ( ) ! n X Ti Tjset q + hi j diag p min , 1 i Rj Si Di G set Tj+1 Tjset j i=i1 +1 j ! q + hn j p Dn G j Ti1 Tjset j diag Si1 Di1 G set , 1 j Rj j j p j increases and increases as Hj decreases.
set Hj

end Hj

p cap,price tex j j p = j
41

set end set If Tjend Tj+1 , then Hj is known in terms of Hj ,

(5.77b)

set Tjend Tjset end set set Hj = Hj + Hj+1 Hj set Tj+1 Tjset p set q cap,price tex Tj+1 Tjset j j set p j ) = Hj+1 Tjend Tjset j
set Hj (

set if Tjend Tj+1 .

set end If Tjend > Tj+1 , then Hj is known from preceding calibraton step. Fitting to the caplet thus requires

(5.78a)

set Hj (

set if Tjend Tj+1 ,

(5.78b)

p p set These formulas describe the trajectory j , Hj ( j ) on which the caplet matches its market value. With p set set Hj = Hj ( j ), we use a 1 parameter global Newton method (starting at j = j+1 ) to choose j to match Vjdiag = Mktdiag . Note that along this trajectory, j (5.79a) (5.79b) 1 Vjdiag dVjdiag Vjdiag set set p p = p + Hj+1 Hj set d j j j Hj 1 Vjdiag dVjdiag Vjdiag end set p p = p + Hj Hj set d j j j Hj
set if Tjend Tj+1 set if Tjend > Tj+1

p q cap,price tex j j end p j ) = Hj j

set if Tjend > Tj+1 .

set set Once weve found both j , Hj we step back to j 1, etc. until weve found j , Hj for j = 1, 2, ..., J. In the usual way, we use piecewise linear interpolation between the known values of (t) and H(T ), and use the invariances to rescale , H to taste.

Appendix A. Bermudans on amortizing swaps. The notional of an amortizing swap steadily declines over the life of the swap. Given a calibrated LGM model, one can evaluate a Bermudan amortizer in exactly the same way as a Bermudan bullet swap. Since the nal price of a deal is determined largely by which instruments are used in calibration, the question is which instruments should be used to get the most adroit pricing and hedging? If we have an 5 year option on a 20 year amortizing swap, surely it will behave more like the 5 into 10 or a 5 into 12 vanilla swaption instead of a 5 into 20 swaption. There are two main approaches to calibrating the model for amortizing Bermudans: (A) For each exercise date, select the vanilla swaption whose behavior matches the behaviour of the Bermudan payo as closely as possible. These swaptions then replace the diagonal swaptions. (B) For each exercise date, calibrate the model to European options on the amortizing swap. To obtain the European options price, we construct a construct a basket of swaps that exactly reproduces the Bermudans payo; we then use LGM itself to value European option on the basket. A.1. Calibrating to the equivalent vanilla swaption. We need to select the vanilla swaption whose behaviour most nearly matches the Bermudans payo for each exercise. Consider the exercise at tex . j The Bermudans xed leg recieves (A.1a) (A.1b) Ci rf pj Ci at ti at ti for i = if irst , j for i = if irst + 1, ..., n, j
42

The Bermudans oating leg receives payments equivalent to (A.1c) Nif irst f eej
j

at tset , j

where the + sign is used if the xed leg (receiver) has the exercise privilege, and the sign is used if the payer has the option. Let us abbreviate (A.2) At any date t, the value of these legs is (A.3a) (A.3b) Vjf ix (t) = Cif irst rf pj Z(t, tif irst ) +
j j

Mj = Nif irst f eej .

j

i=if irst +1 j

n X

Ci Z(t, ti ),

Vjf lt (t) = Mj Z(t, tset ). j

A.1.1. Ratio matching. There are two main ideas for picking the most similar swaption, ratio matching and payo matching. Consider the ratio (A.4) Ratio = Vjf ix (tex ) j Vjf lt (tex ) j = Cif irst rf pj Z(tex , tif irst ) j
j j

Mj

Z(tex , tset ) j j

i=if irst +1 j

n X

Ci Z(tex , ti ) j . Mj Z(tex , tset ) j j

This ratio represents the dollars received per dollar spent upon exercising the option. Suppose we model the yield curve as being todays yield curve plus parallel shifts and tilts. Then, (A.5) D(T ) (T tjset )(T tset )2 + Z(t, T ) j set ) D(tset ) e Z(t, tj j D(T ) 1 2 set ) (T tset )2 + , = j set ) 1 (T tj 2 D(tj

where and are the amount of the parallel shift and tilt, respectively. Under these movements, the ratio becomes (A.6a) where (A.6b) D(tif irst ) j + Moneyness = Cif irst rf pj j D(tset ) j
n X

Ratio = Moneyness Sensitivity ( 1 2 ) Convexity + , 2 D(ti ) D(tset ) j

n X

Ci

i=if irst +1 j

(A.6c)

D(tif irst ) j + Sensitivity = tif irst tst Cif irst rf pj j j j D(tset ) j

i=if irst +1 j n X

(A.6d)

2 D(tif irst ) j Cif irst rf pj + Convexity = tif irst tst j j j D(tset ) j

D(ti ) ti tst Ci j D(tset ) j 2 D(ti ) ti tst Ci j D(tset ) j

i=if irst +1 j

Consider a standard bullet swap with exercise date tex and with a start date Tjref st which is spot-of-tex . j j ref Let the Tjref end be the theoretical end date, and let the strike be Rj . Assume that the swap has K periods,
43

the rst of which is generally a stub, and let the xed rate dates be s0 , s1 , . . . , sK . Clearly s0 = Tjref Then the ratio of the swaps xed leg to the oating leg is (A.7) Ratioref = j
K X

st

k=1

Z(tex , sk ) Z(tex , sK ) j j ref cvg(sk1 , sk ) Rj Sk + , Z(tex , s0 ) Z(tex , s0 ) j j

where Sk is the oating legs basis spread adjusted to the xed legs frequency and day count basis. Under the same set of yield curve movements as before, this ratio becomes (A.8a) where (A.8b) Moneynessref = j Sensitivityref = j Convexityref = j
K X

Ratioref = Moneynessref Sensitivityref ( 1 2 ) Convexityref + , j j j j 2 D(s ) D(s ) k K ref cvg(sk1 , sk ) Rj Sk + D(s0 ) D(s0 )

k=1

(A.8c)

k=1

K D(s ) X D(sK ) k ref (sk s0 ) cvg(sk1 , sk ) Rj Sk + (sK s0 ) D(s0 ) D(s0 )

(A.8d)

k=1

K D(s ) X D(sK ) k ref (sk s0 )2 cvg(sk1 , sk ) Rj Sk + (sK s0 )2 D(s0 ) D(s0 )

If the ratio of the reference swap matched the ratio of the amortizing swap reference under all possible movements of the yield curve, then clearly the value of the European option on the amortizer would equal ref the value of the vanilla swaption, With two free variables Rj and Tjref end , however, we can only choose the swaption which matches the moneyness and the sensitivity of the Bermudans payo. The two xed legs match only for parallel shifts, so although we still argue that the values should be nearly the same, we incur some risk in doing so. From a traders perspective, if we went long (short) the amortizing swaption and short (long) the bullet, we would be neutral for parallel shifts, but exposed to tilts. If the prices on the two were signicantly dierent, it would tempt enough traders to take the tilt risk, eliminating the mis-balance. A.1.2. Payo matching. In ratio matching, we essentially matched the oating leg exactly and matched the xed leg as well as possible. To get the options value right, however, all we have to do is mimic the forward value of the net payo (xed minus oating correctly). This allows us one more variable: the notional. A $2 swap with a 3y tenor may have the same sensitivity as a $1 swap with a 7y tenor.By varying the notional, we can match both the sensitivity and the convexity. Let (A.9) s0 = Tjref
st

be the standard spot date for tex . Consider the forward value of the payo for date s0 as seen at date tex : j j (A.10) Z(tex , tif irst ) j j Payo(tex ) = Cif irst rf pj + j j Z(tex , s0 ) j
n X

Ci

i=if irst +1 j

Z(tex , ti ) Z(tex , tset ) j j j Mj . Z(tex , s0 ) Z(tex , s0 ) j j

Once again we suppose that the yield curve undergoes parallel shifts and tilts: (A.11) Z(tj , T ) ex Z(tj , t0 ) ex = D(T ) (T s0 )(T s0 )2 e D(s0 ) D(T ) = 1 (T s0 ) ( 1 2 )(T s0 )2 + . 2 D(s0 )
44

Then the forward value of the payo becomes (A.12a) where (A.12b) D(tif irst ) j + FwdVal = Cif irst rf pj j D(s0 )
n X

Payo(tex ) = FwdVal (T t0 ) Sensitivity ( 1 2 ) Convexity + j 2 D(tset ) D(ti ) j Mj , D(s0 ) D(s0 )

Ci

i=if irst +1 j

(A.12c)

D(tif irst ) j Sensitivity = tif irst s0 Cif irst rf pj j j D(s0 ) n X D(tset ) D(ti ) set j (ti s0 ) Ci tj s0 Mj , + D(s0 ) D(s0 ) f irst
i=ij +1

(A.12d)

2 D(tif irst ) j Cif irst rf pj Convexity = tif irst s0 j j D(s0 ) n X 2 D(tset ) D(ti ) set j (ti s0 )2 Ci tj s0 Mj . + D(s0 ) D(s0 ) f irst
i=ij +1

ref Now consider a standard bullet swap with notional Mj , start date s0 = Tjref st , theoretical end date ref Tjref end , and strike Rj . Assume that the swap has K periods, the rst of which is generally a stub, and let the xed rate dates be s0 , s1 , . . . , sK . Clearly s0 = t0 . With three free variables (M, Rref , and tth ), we end can match the forward value, the sensitivity, and the convexity of the amortizing swap: (K ) X ref D(sk ) D(sK ) ref (A.13a) FwdValref = Mj cvg(sk1 , sk ) R Sk + 1 , j D(s0 ) D(s0 ) k=1

(A.13b)

Sensitivityref j

ref Mj

k=1

K X

D(sk ) D(sK ) (sk s0 ) cvg(sk1 , sk ) Rref Sk + (sK s0 ) D(s0 ) D(s0 )

2

(A.13c)

Convexityref j

ref Mj

(K X

k=1

D(sk ) D(sK ) (sk s0 ) cvg(sk1 , sk ) Rref Sk + (sK s0 )2 D(s0 ) D(s0 )

The forward value of the amortizing swap and the reference swap are now the same under reasonably large parallel shifts of the yield curve, and under not-too-large tilts of the yield curve. If we went long (short) the amortizing swaption and short (long) the bullet swaption, we would be delta and gamma neutral for parallel shifts. We also be delta neutral for tilts. We assume that this combination should be priced at zero, since even a small dierence would tempt traders to take on the residual yield curve risk. If we were agressive, we would claim that the value of the this reference swaption is the same as the value of the European option on the amortizing swap. Here we are less aggressive, and simply insist that we calibrate the LGM model to these reference swaptions in lieu of the diagonal swaptions.
45

A.2. Matching by baskets. Recall that if the Bermudan is exercised at tex , the payos xed leg is j (A.14a) (A.14b) Ci rf pj Ci at ti at ti for i = if irst , j for i = if irst + 1, ..., n, j We approximate the oating leg

and the payos oating leg is equivalent to the payment Mj at tset . j payment as a payment of (A.14c) Mj D(tset ) j D(Tjset ) at Tjset ,

where Tjset is the standard spot-of-tex for the currency. j Suppose we knew the market price of the European option on this amortizing swap. Then we would use our favorite calibration strategy (constant + diagonals, caplets+diagonals, . . . ), and calibrate the model to the European option on the amortizing swap in lieu of the diagonal swaptions. Unfortunately, there are no liquid quotes for the prices of European options on amortizers. Instead we are going to reconstruct the amortizing swap as a linear combination of standard bullet swaps; i.e., express the amortizing swap as a basket of ordinary swaps. We then use the LGM model itself to nd the value of a European option on the basket in terms of the market values of each swaptions. This European value is then to be fed into our calibration scheme. A.2.1. Constructing the basket. Dene swap k to be the bullet swap which starts on date Tjset , and ends on the kth paydate tk of the Bermudans payo. We assume (or approximate) the bullet swaps pay dates as being the same as the Bermudan pay dates. So we assume that swap k has xed leg pay dates (A.15) ti for i = if irst , if irst + 1, . . . , k j j

ref ref and start date Tjset . Let Mk and Rk be the notional and strike of the k th swap. Then its xed leg payments are ref ref (A.16a) for i = if irst , 2, . . . , k 1 at ti Mk i Rk Si j n o ref ref Mk (A.16b) 1 + k Rk Si at tk

and it oating leg payments are equivalent to (A.16c)

ref Mk

at Tjset .

Here Si is the basis spread for the ith interval, adjusted to the xed legs frequency and day count basis is the usual way, and (A.16d) (A.16e) i = cvg(Tjset , ti ) i = cvg(ti1 , ti ) for i = if irst , j for i = if irst + 1, 2, . . . , k 1. j

ref ref We wish to choose the notionals Mk and strikes Rk so that the sum of all the payments of these reference swaps equals the payments in the Bermudans payo. Equating the ith payment of all the swaps in the basket to the ith payment of the amortzing swap yields n X ref ref (A.17a) Mk i Rk Si + Miref = Ci rf pj for i = if irst j k=i

(A.17b)

n X k=i

ref ref Mk i Rk Si + Miref = Ci

46

for i = if irst + 1, 2, . . . , n j

Equating the oating legs yields (A.17c)

n X ref Mk = Mj

k=if irst j

D(tset ) j . D(Tjset )

ref Now, if all the strikes Rk were specied, then we could work backwards. We would rst determine the ref notional Mk as k = n needed to match the last payment, then the notional for k = n 1 needed to match the next to last payment, etc. Proceeding in this way, we would match all the xed leg payments, but the ref sum of these notionals Mk would not (unless we were very lucky) match Mj D(tset )/D(Tjset ). We use our j freedom to choose the strikes to get one more degree of freedom in choosing the notionals. Strike choice A. There are two obvious methods for choosing the reference strikes. The rst is setting each swaps strike equally far from the money, so that the same parallel shift is needed to bring each to the money:

(A.18a)

ref sw Rk = Rk +

for k = if irst , if irst + 1, . . . , n j j

sw where Rk is the forward (break-even) xed rate for swap k. Solving

(A.18b) (A.18c)

i i

n X k=i n X k=i

ref sw Mk (Rk Si ) + i ref sw Mk (Rk Si ) + i

n X k=i n X k=i

ref Mk + Miref = Ci rf pj ref Mk + Miref = Ci

for i = if irst j

for i = if irst + 1, 2, . . . , n j

ref determines the notionals Mk () in terms of . We then need to nd which enables the oating leg to be matched: n X D(tset ) j ref (A.18d) Mk () = Mj . D(Tjset ) f irst k=ij

This can be done by a quick global Newtons scheme, starting from = 0. Strike choice B. The second method is a variant of this scheme. It sets the strikes of the reference swaps to be the same number of standard deviations from the money: (A.19a)
ref sw Rk = Rk + atm k

for k = if irst , if irst + 1, . . . , n. j j

sw Here Rk is again the swap rate, and now atm is the at-the-money swaption volatility for the swaption with k exercise date tex , start date Tjset , and end date tk . Solving j

(A.19b) (A.19c)

i i

n X k=i n X k=i

ref Mk

sw (Rk

Si ) + i

n X k=i n X k=i

ref Mk atm + Miref = Ci rf pj k ref Mk atm + Miref = Ci k

for i = if irst j

ref sw Mk (Rk Si ) + i ref Mk ()

for i = if irst + 1, 2, . . . , n j

in terms of . We can then use a quick global Newtons scheme to nd determines the notionals which enables the oating leg to be matched: (A.20)
n X ref Mk () = Mj

k=if irst j

D(tset ) j . D(Tjset )

47

Which strike method to use. We dont have enough market experience to know whether the second method of choosing strikes oeres signicantly better pricing/hedging. If it does not, then we should use the simpler method, method A. Note that these can be programmed together, since it is just a matter of inserting the weights atm into the problem. k A.2.2. Pricing the European option on the basket. Now that we have replicated the amortizing swap as a basket of bullet swaps, we price the European option on the basket. We rst calibrate LGM model to reproduce the market price of each swaption in the basket, and then use the calibrated LGM model to price the European option on the basket. Once we obtain the price of the European option on the basket, we throw this calibration away. This calibration has no role in pricing our original Bermudan except to obtain the value of the European option on the amortizing swap. Recall that the k th swap has the xed leg payments, (A.20a) (A.20b) ref ref for i = if irst , 2, . . . , k 1 Mk i Rk Si at ti j n o ref ref Mk 1 + k Rk Si at tk
ref Mk

and it oating leg payments are equivalent to (A.20c) The LGM value of receiver swaption k is
LGM (A.21a) Vbasket = k X ref i Rk Si

at Tjset .

i=if irst j

yk + Hi ex yk + Hi ex yk j j q q +DK N D0 N q Di N ex ex j j ex j

whose yk is determined implicitly by solving, (A.21b)

k X

i=if irst j

2 ex 2 ex 1 1 ref set i Rk Si Di eHi yk 2 Hi j + Dk eHk yk 2 Hk j = Dj ,

and where we have used (A.21c)

set Hi = Hi Hj = H(ti ) H(Tjset ).

We calibrate these swaptions by using the calibration of a row of swaptions technique described above. (More to the point, we can use the same routines). We rst set (A.22a) (A.22b) et = (tex ) = 104 tex , j j j

H(Tjset ) = 0

without loss of generality. We then assume that H(T ) is piecewise linear with nodes at ti for i = if irst , if irst + j j 1, . . . , n. We rst calibrate on the swaption k = if irst , which determines H(ti ) for i = if irst . We then j j calibrate on swaption k = if irst + 1, which determines the next H(ti ). Continuing gives us all the values of j H(ti ). Once we have calibrate H(t), then the value of the European option on the basket is
48

(A.23a) V LGM

where y is the unique solution of: (A.23b) 2 Cif irst rf pj Dif irst exp{Hif irst yk 1 Hif irst ex } + j 2
j j j j

n X y + Hi ex yk + Hi ex j j q = Cif irst rf pj Dif irst N q + Ci Di N j j ex ex j j i=if irst +1 j yk Mj D(tset )N q j ex j

n X

Ci Di eHi yk 2 Hi j = Mj D(tset ). j

2 ex

i=if irst +1 j

LGM Once we have the value Vbasket of the European option on the amortizing swap, we can use this as the market price of the amortizing swap in our calibration.

Appendix B. American swaptions. Appendix C. Cross-currency swaptions.

49