You are on page 1of 13

SINGLE CURRENCY INTEREST RATE MODELS

PATRICK S. HAGAN Abstract. Here we analyze one factor and multifactor interest rate models. For the HW, BK, and generalized BK models, we derive zero coupon bond formulas, lightning evaluation methods, calibration strategies and methods, and approximate swaption./caplet formuals. Key words. interest rate models, calibration methods, lightning

(Part head:)One factor models 1. Linear Gauss Markov (LGM) model. The numeraire is (1.1) N (t, x) =
2 2 1 +H (t)x+ 1 2 H (t) (t) . e D(t)

The complete LGM model can be written as (1.2a) (t, x) V (t, x) , V N (t, x) Z (T, X ) 2 V p e(X x) /2[ (T ) (t)] dX. 2 [ (T ) (t)]

(1.2b)

(t, x) = V

The value of a zero coupon bond is (1.3)

2 2 (t, x; T ) = Z (t, x; T ) = D(T )eH (T )x 1 2 H (T ) (t) . Z N (t, x)

By denition, the instantaneous forward rate for maturity T , which we denote as f (t, x; T ), satises (1.3) Z (t, x; T ) = e
T
t

f (t,x;T 0 )dT 0

D(t, T ) = e

T
t

f0 (T 0 )dT 0

where f0 (T ) is todays instantaneous forward rate for maturity T. So equation (1.3) shows that for the LGM model (1.4) f (t, x; T ) = f0 (T ) + H 0 (T )x + 2 H 0 (T )H (T ) (t).

We choose the sign of H 0 (T ) to be positive, setting x x, H 0 (T ) H 0 (T ) if necessary. Then as x increases, the forward rates f (t, x; T ) increase. 1.0.1. Scaling. Interest rates in G7 countries change by 0.80% or so over the course of a year. Thus, a 6.2% rate environment would typically remain within the range 5.4% to 7% during the next year. Equivalently, the standard deviation of H 0 (T )x should grow by 1% or less per year. Accordingly, we scale to be O(1%) and take H (T ), H 0 (T ), and x to be O(1). Note that the last term in (1.4) is O(104 ). 1.0.2. Model invariances. Inspection reveals that LGM has two invariances: All market prices remain unchanged if H (T ) H (T ) + C for any constant C ; All market prices remain unchanged if (T ) (T )/C 2 , H (T ) CH (T ) for any constant C . 1.1. Swaption & caplet prices.
1

1.1.1. Exact, closed form expressions for swaption prices. Consider a swaption with (1.5a) (1.5b) (1.5c) and let (1.5d) i = cvg (ti1 , ti ) tex = exercise date, ts t0 = start (settlement) date, t1 , t2 , . . . , tn = xed leg pay dates,

be the coverages (accrual fractions) of the xed leg. Then the xed leg payments are (1.6a) (1.6b) where (1.6c) si = oating basis spread for period i. (Rf si )i paid at ti , for i = 1, 2, . . . , n 1 1+(Rf sn )n paid at tn ,

The oating leg (less the basis spread) is worth (1.6d) At any t, x the value of the receiver swap is (1.7) Vswap (t, x) =
n X (Rf si )i Z (t, x; ti ) + Z (t, x; tn ) Z (t, x; ts ). i=1

1 paid at t0 = ts .

On the exercise date tex , the value of the receiver swaption is clearly [Vswap (t, x)]+ . Since N (0, 0) = 1, todays value of the receiver swaption Z 2 [Vswap (tex , X )]+ eX /2 (tex ) p dX. (1.8) Vrec (0, 0) = N (tex , X ) 2 (tex )

where

Substituting (1.3) and (1.7) into (1.8) now yields (1.9a) #+ Z X 2 /2 "X n e 2 2 1 2 2 1 2 2 e Hi X 1 H H X H H X H n s i e + D e n e D e s e 2 2 2 p Vrec (0, 0) = (Rf si )i Di e dX n s 2 e i=1 Di = D(ti ), Ds = D(ts ), Hi = H (ti ), Hs = H (ts ), and e = (tex ).

(1.9b)

Recall that H (t) is an increasing function. Thus Hi Hs = H (ti ) H (ts ) is positive. As X increases, all the positive terms in (1.9a) become smaller. So if we dene x as the unique break-even point (1.10a)
n X 2 2 2 2 1 2 1 2 (Rf si )i Di e(Hi Hs )x 2 (Hi Hs ) e + Dn e(Hn Hs )x 2 (Hn Hs ) e = Ds , i=1

then the integrand is positive for x < x . Consequently, we need to integrate (1.9a) from to x , which yields ! ! ! n X x + Hi e x + Hn e x + Hs e p p p (1.10b) Vrec = (Rf si )i Di N + Dn N Ds N . e e e i=1
2

Here N() is the standard (cumulative) normal distribution function. We can simplify this expression by dening y = x Hs e . Then the swaption value is ! n X y + (Hi Hs ) e p (1.11a) (Rf si )i Di N Vrec = e i=1 ! ! y + (Hn Hs ) e y p Ds N p , +Dn N e e where (1.11b)
n X 2 2 1 2 1 2 (Rf si )i Di e(Hi Hs )y 2 (Hi Hs ) e + Dn e(Hn Hs )y 2 (Hn Hs ) e = Ds . i=1 n X (Rf si )i Di Dn + Ds . i=1

(1.11c)

Vpay = Vrec

This expression depends only on the model parameters (t), H (t) on the dates: (1.12a) (1.12b) Hi Hn = H (ti ) H (tn ), e = (tex ). i = 1, 2, . . . , n

That is, the swaption value depends ony on H at the pay dates (and stard date) and on on the exercise dates. This is the key to fast, reliable calibration of the model. One arranges the series of instruments (swaptions, caplets, ...) so that one only has to calibrate one or two (new) parameters at a time. This is detailed below. A oorlet is a one period receiver swaption, so ! ! y + (H1 Hs ) e y p (1.13a) Vf loor = (1 + 1 [Rf s1 ])D1 N Ds N p , e e (1.13b) Vcap = Vf loor F + K, y=
2 2 log[(1 + 1 [Rf s1 ])D(t1 )/D(ts )] 1 2 (H1 Hs ) e . (H1 Hs )

(1.13c)

So the value of the oorlet is given by Blacks formula (1.14a) Vf loor = F N(d1 ) K N(d2 ), d1,2 = Vcap = Vf loor F + K

(1.14b) where (1.14c) (1.14d)

2 tex log F/K 1 2 B , B tex

F = (1 + 1 [Rf s1 ])D(0, t1 ),

K = D(0, ts ),

p B tex = [H (t1 ) H (ts )] (tex )


3

1.1.2. Derviatives of exact swaption prices. Dierentiating yields ! n X y + (Hi Hs ) e p p Vrec = (Hi Hs ) (Rf si ) i Di G (1.15a) e e i=1 ! y + (Hn Hs ) e p , +(Hn Hs )Dn G e ! n p X y + (Hi Hs ) e p Vrec = e (Rf si ) i Di G (1.15b) Hs e i=1 ! p y + (Hn Hs ) e p , e Dn G e ! p y + (Hi Hs ) e p Vrec = e (Rf si ) i Di G (1.15c) , Hi e ! p y + (Hn Hs ) e p Vrec = e (1 + [Rf si ] n )Dn G (1.15d) , Hn e where (1.15e) G(z ) =
2 1 ez /2 2

The derivative of payer swaptions is identical to the derivative of the receiver swaptions. 1.1.3. Approximate swaption values via equivalent vol techniques.. The rst approximate formula can be derived by switching to the level (1.16) L(t, x) =
n X i=1

i Z (t, x; ti )

as the numeraire, and the swap rate Rs (t, x) as the state variable. The swap value and swap rate Rs are (1.17) Vsw (t, x) = [Rf Rs ]L(t, x). Rs (t, x) = Z (t, x; ts ) + Pn i=1 i si Z (t, x; ti ) Z (t, x; tn ) P . n i=1 i Z (t, x; ti )
1 2 2 2 1 2 2 2 (Hn Hs ) (t)

(1.18) So (1.19) Rs (t, x) =

Ds +

Pn

i=1

to R s (t) Rs (t, X (t)). Applying Itos formula to (1.18) yields Let us change variables from X (1.20a) where
0 (t, x) Rs 0 s = {drift}dt + (t)Rs ) dW , dR (t, X

i si Di e(Hi Hs )x 2 (Hi Hs ) (t) Dn e(Hn Hs )x 2 Pn 2 2 2 (Hi Hs )x 1 2 (Hi Hs ) (t) i=1 i Di e

(Hn Hs )Dn e(Hn Hs )x 2 (Hn Hs ) (t) = Pn 2 2 2 (Hi Hs )x 1 2 (Hi Hs ) (t) i=1 i Di e Pn 2 2 2 (Hi Hs )x 1 2 (Hi Hs ) (t) i=1 (Hi Hs ) (Rs (t, x) si ) i Di e + Pn 2 2 2 (Hi Hs )x 1 2 (Hi Hs ) (t) i=1 i Di e
4

1 2

We now use the level L(t, x) as the numeraire. Then (1.21a) where L0 is todays level and (1.21b) 0 s (tex )]+ Vrec = L0 E{ [Rf R }, Rs (0) = Rs
0 s = (t)Rs ) dW . dR (t, X

Here x = x(t, Rs ) is dened implicitly by (1.19). This is precisely the setup needed to apply the equivalent vol technique. Those results show that the swaption value is given by Blacks formula (1.22a)
0 Vrec = L0 {Rf N(d1 ) Rs N(d2 )}, 0 1 2 tex log Rf /Rs 2 B , B tex

(1.22b)

d1,2 =

0 where L0 is todays level and Rs is todays swap rate,

(1.22c)

L0 =

n X i=1

i Di ,

0 Rs =

Ds +

and where the equivalent volatility is (1.22d)

Pn i=1 i si Di Dn P , n i=1 i Di

We neglect all O(2 ) terms. Then (1.23a) where xf is dened by (1.23b) Rf =

0 p log(Rf /Rs ) B tex = (tex ) 1 + O(2 ) . x(0, Rf ) 0 p log(Rf /Rs ) B tex = (tex ), xf

Ds +

Pn

i=1

This can be re-written as (1.23c) Rf


0 Rs

Pn

0 i=1 (Rs

Note that as xf 0, this goes to (1.23d)

si )i Di 1 e(Hi Hs )xf + Dn 1 e(Hn Hs )xf Pn (Hi Hs )xf i=1 i Di e

i si Di e(Hi Hs )xf Dn e(Hn Hs )xf Pn . (Hi Hs )xf i=1 i Di e

1.1.4. Alternative approximations. A more accurate approximation can be obtained from the exact solution ! n X p y Vrec = (Rf si ) i Di N p + (Hi Hs ) e (1.24a) e i=1 ! ! p y y . +Dn N p + (Hn Hs ) e Ds N p e e
5

P 0 p (Hn Hs )Dn + n i=1 (Hi Hs )[Rs si ]i Di Pn B tex (tex ) . Ds + i=1 i si Di Dn

After neglecting the small 2 e terms, y is determined by (1.24b)


n X i=1

(Rf si ) i Di e(Hi Hs )y + Dn e(Hn Hs )y = Ds .

We expand (1.24a) in powers of . This yields (1.25)


n p X 0 i Di Vrec =N y/ e Rf Rs i=1

We can re-write this as ) ( n p X (Rf si ) i Di + Dn Ds (1.26) Vrec = N y/ e


i=1

) ( n p X p 2 2 2 (Rf si ) i Di (Hi Hs ) + Dn (Hn Hs ) + 1 2 y e G y/ e


i=1

) ( n p X p (Rf si ) i Di (Hi Hs ) + Dn (Hn Hs ) + e G y/ e


i=1

where y is dened by (1.27b)

Using (1.24b), our second approximation for swaption prices is " n # p i X p h p 0 i D Rf Rs N y/ e + ( e /y )G y/ e , (1.27a) Vrec =
i=1 n X i=1

( n ) p h i h i e p X (Hi Hs )y (Hn Hs )y (Rf si ) i Di 1 e + Dn 1 e + + G y/ e y i=1

(Rf si ) i Di e(Hi Hs )y + Dn e(Hn Hs )y = Ds .

To intepret this formula, consider a normal asset whose forward value F satises (1.28a) dF = N dW.

The value of European call option on a normal asset is F K F K + N tex G . (1.28b) Vrec = (F K )N N tex N tex Suppose we set (1.29a) and (1.29b)
n X 0 i Di , F K = Rf Rs i=1

Then the normal call option price is identical to our approximate swaption price If we are reasonably close to the money, then (1.30) y 1 . Pn F K i=1 (Rf si ) i Di (Hi Hs ) + Dn (Hn Hs )
6

F Kp N tex = e. y

1.2. Auto-calibration strategies. LGM has two model parameters, H (t) and (t), to t. One can t both functions H (t) and (t), which requires calibrating to two series of swaptions/caplets. Alternatively, one can choose one of the functions H (t) or (t) a priori, and then use a single series of swaptions/caplets to calibrate the other. Throughout we use piecewise constant interpolation for H 0 (t) and 0 (t) = 2 (t). So, after calibration, H (t) and (t) are piecewise linear functions. 1.2.1. Construction of the reference instruments. Each reference swaption can be specied by its exercise date tex , the nal pay date of its swap tend , and its strike Rf . Due to call-put parity, calibrating to a receiver is equivalent to calibrating to a payer, and all other parameters (the start date, the frequency of the xed leg, the day count basis, ...) are set by market practices. In particular, if the length of the swap is a non-integral number of periods, then the stub is in front, as is standard. Each reference caplet can be specied by its xing date tex and the strike Rf . The length of the caplet is set by the market. The set of reference instruments used in each strategy depend on the following sets of dates. For exotics which have discrete exercise dates (like Bermudans), let (1.31a) 1 < 2 < . . . < m

be the exotic deals exercise dates which are after today (the eval date). If the exotic has continuous exercises (like American deals) or frequent exercises (more than, say, once a month), then the exercise dates should be limited to either one per month, per quarter, or per six month period, depending on the situation and the traders instinct. (Remember, these dates are only going to be used to construct the calibration instruments, not to evaluate the deal). In this case one should ensure (by adding extra dates if needed) that the rst and last possible exercise dates are included in the reference set. Once the set of exercise dates has been obtained, dene the start date as the standard spot date of the rst exercise date: (1.31b) tst = spot( 1 ).

Also dene the pay dates for a standard xed leg in the market with theoretical end date tlast and start date tst , (1.31c)
p p p p tst = tp 0 < t1 < . . . < tn = tlast < tn+1 < tn+2 ,

and include two extra periods after tlast . Finally, dene the standard cap dates for a cap which has a theoretical end date at tlast and start date at tst : (1.31d)
c c tst = tc 0 < t1 < . . . < t n = tlast .

For legs with fractional periods, tc 1 is the last regular cap date which is strictly after tst , and tst is inserted into the date series as an extra caplet start date. For any given exotic, the standard calibration methods use one or two of the following series for reference instruments. Diagonal swaptions. This is a series of m swaptions, one swaption for each exercise date k . These swaptions share the common end date tlast = tp n whenever possible. If this end date would result in a swap shorter than, say, ten months, then one or two extra periods need to be added to make the swap at least ten months long. So swaption k is dened by the exercise date, start date, and end date (1.32a) tk ex = k ,
k tk st = spot(tex ),

(1.32b)

tk end

tp n+1

tp if tst tp n = tlast n 10 mos p if tn 10 mos < tst tp n+1 10 mos p tp if t 10 mos < tst n+2 n+1
7

for k = 1, 2, . . . , m. The xed legs of these swaptions all share a common set of pay dates. Namely, all dates in (1.31c) which are strictly after the start date of the swaption and on or before tlast ; the p k last few swaptions may also included the dates tp n+1 and/or tn+2 . The strikes Rf of these reference swaptions (and all other calibration instruments) are chosen to match the strike of the exotic deal using the standard strike matching method described below. Caplets (exercise dates). This is a series of m caplets, one for each of the above exercise dates. k k Caplet k is the caplet whose xing date is tk ex = k , which makes the start date tst = spot{tex }. The caplets each have the market-standard length (usually 3 months), and the strikes are chosen by the standard method below. Caplets (pay dates). This is a series of n caplets determined by the last pay date tlast , not by the exercise schedule. The xing dates of the caplets are (1.33a)
c tk ex = xing date{tk1 },

k = 1, 2, . . . , n 1,
c tk end = tk .

and.the start and end dates or the caplet are (1.33b)


c tk st = tk1 ,

The strikes of these deals are determined by the standard method below. Row swaptions. These swaptions all have a common exercise date tex , and pay dates which are part of swaption pay date series tp k in (1.31c). Let tex = 1 be the exotics rst exercise date after today (the eval date). If this exercise date tex is earlier than some minimum date tmin (usually three or six months after today) then take tex = tmin . Also set the start date tst to be spot-of-the exercise date. Then the row swaptions are dened by (1.34) tk ex = tex , tk st = spot{tex },
p tk end = tk

for k = kmin , kmin +1 , ..., n.

Here tp kmin is the rst pay date at least, say, 10 months after the common start date tst . If none of p these swaptions are over ten months long, then we need to change the end date to tp n+1 or tn+2 . The swaption series will then consist of a single swaption with exercise date tex , start date tst , and an p end date of either tp n+1 or tn+2 as needed so that the swap is at least 10 months long. The strikes Rf are chosen by the standard method given below. Short swaptions (exercise dates). Like the diagonal swaptions, there is one short reference swaption for each exercise date k in (1.31a). Instead of ending at the last pay date tlast , these swaps only use enough of the date sequence tp j so that they are at least 10 months long. So swaption k is dened by the exercise date, start date (1.35a) and has the xed leg pay dates (1.35b) tp j for j = ik , ik + 1, . . . , nk tk ex = k , tk st = spot{ k },

k p where tp ik is the rst date pay date strictly after tst and tnk is the rst pay date in the sequence that k is at least 10 months after tst . By construction, these swaptions also share the common set of pay dates (1.31c). The strikes Rf are chosen by the standard method below. Short swaptions (pay dates). This is a series of n options into (roughly) one year swaps which is determined by the last pay date tlast , and not by the exercise schedule. These swaptions are given by

(1.36)

p tk ex = spot{tk },

p tk st = tk ,

p tk end = tk+

for k = 0, 1, . . . , n 1,

where is one or 2, depending on whether the dates tp k are a semi-annual or annual series.
8

Swaption column (exercise dates). The one year tenor of the short swaptions is an arbitrary choice. After choosing a tenor (like 5 years), these swaptions are constructed exactly like the short swaptions p (exercise) except that each nk is chosen so that the swap from tk st to tnk is longer than the prescribed tenor less a fudge factor of, say 2 months. This generally requires extending the date series tp k by including other dates after tp . n+2 Swaption column (pay dates). After choosing a tenor (like 5 years), these swaptions are constructed p exactly like the short swaptions (payment) except that is chosen so that the swap from tk st = tk p to tk+ is longer than the prescribed tenor less a fudge factor of, say 2 months.This generally requires p extending the date series tp k by including other dates after tn+2 . Choosing strikes of reference instruments is not straightforward for non-Bermudan deals or for Bermudans with amortizing or step-up features. The standard method of choosing these strikes is to rst characterize the exotic by setting the vols to zero and, for each exercise date k of the exotic, parallel shifting the yield curve until the payo obtained by exercising the exotic is at-themoney. The characterization is the set of exercise dates and the corresponding set of parallel shifts. The strike of the swaption/caplet is matched to this characterization by rst interpolating to nd the parallel shift for the exercise date (xing date) ex of the swaption (caplet), and then nding selecting the strike Rf so that the swaption/caplet is at the money when this shift is applied to the yield curve. The interpolation should use a linear interpolation method with at extensions beyong the end points. Generally autocalibration of the Hull-White model employs one of the following strategies. We emphasize that the purpose is to obtain Hi = H (ti ) and k = ( k ) at a series of points ti and k ; piecewise linear interpolation (with linear extrapolation beyond the end points) is used to obtain H (t) and (t) at other points t. 1.2.2. Fixed tau/diagonal calibration method. For this method the user abritrarily chooses a = 1/ . Since A and B are arbitrary see eq. one arbitrarily chooses (1.37) Hi = H (tp i) = eti 1 e
p

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

The normalizing factor 1 e is arbitrary, but this choice works well enough. Note that H (t) is not purely exponential since linear interpolation is used between the nodes ti . We now calibrate on the diagonal series of m swaptions dened in (1.32a), (1.32b). Recall that the value k of a swaption depends only on (t) on the exercise date tk ex , and H (t) on the start date tst , and pay dates ti . See eqs. (1.11a), (1.11b). Since H (t) is already known from (1.37), we need only use global Newton to th determine the unique value of k = (tk diagonal swaption to its market price, for each ex ) which ts the k k. Fixed tau methods give good prices for exotics, especially swaption-like exotics like Bermudan swaptions, callable inverse oaters, and other callable deals. Unfortunately, the exotic prices depend on the arbitrary input , with the best price being obtained when = 1/ is very near zero or even negative. However, it is often dicult to calibrate for long dated deals when = 1/ is too small, and the nal exotics price depends on the arbitrary input . 1.2.3. Fixed tau/caplet, xed tau/short swaption, and xed tau/swaption column methods. The xed tau/caplet method againn requires the user to choose = 1/ , and (1.37) is used to obtain H (t) p p on the set of dates tp 0 , t1 , . . . , tn . The LGM model is then calibrated on the series of caplets dened above in caplet(exercise dates). Since H (t) is already known, for each k we need to use global Newton to determine th the unique value of k = (tk caplet to its market price. See eqs. (1.14a) - (1.14b). This ex ) which ts the k method has the same strengths and weaknesses as the xed tau/diagonal methods, except that one gets good pricing for deals which have strong cap-like characteristics, like captions, chooser caps, and autocaps. The xed tau/short swaption method is identical to the xed tau/diagonal method except that it calibrates against the short swaptions (exercise dates) swaptions dened in (1.35a), (1.35b). This calibration scheme should be reserved for exotics which can be represented well in terms of a series of 1 year swaptions.
9

The xed tau/swaption column method is identical to the xed tau/diagonal method except it calibrates to the swaption column (exercise dates) instruments dened above. Again, it should be used for deals which are represented fairly well in terms of a column of swaptions. 1.2.4. Fixed sigma/diagonal. This is really a xed calibration; it is called xed due to a historical quirk. This calibration method xes (t) to be a constant 0 , typically chosen to be 0.0001. Then (1.38a) (t) = 2 0t

Recall that the LGM model has two invariances. We eliminate one by choosing 0 ; we eliminate the other by setting H (tp n ) to zero: (1.38b) H (tp n ) = 0.

Due to these invariances, the market price of all instruments, both exotic and vanilla, are independent of the value of 0 chosen. The model is calibrated to the diagonal swaptions in (1.32a), (1.32b) to determine H (t) at their start dates tk st : (1.39) H1 = H (t1 st ), H2 = H (t2 st ), ...
p Hm = H (tm st ), and Hm+1 = H (tn ) = 0.

Piecewise linear interpolation (linear extrapolation beyond the end points) is used to nd H (t) for other values of t. We calibrate backwards, starting with the mth swaption, continuing to the (m 1)th swaption, ..., and ending with the rst swaption. Recall that the value of the mth swaption depends on the volatility (tm ex ) at its exercise date, and the value of H (t) at the start date and all pay dates. Given Hm at the start date tm st , the values at the rest of the dates are known by linear interpolation: (1.40a) H (t) = Hm + (t tm st ) Hm+1 Hm m tp n tst for t tm st .

Thus we can write the price of the mth swaption in terms of Hm = H (tm st ) and known quantities, and use a global Newton routine to determine Hm by matching the swaption to its market value. We then move onto the m 1 swaption, writing (1.40b)
1 ) H (t) = Hm1 + (t tm st

Hm Hm1 m1 tm st tst

m1 for tst t tm st

m1 for times earlier than tm ) by tting st , and using (1.40a) for later times. This allows us to determine H (tst m2 th the (m 1) swaption to its market value. Repeating this procedure, we obtain H (tst ), . . . , H (t1 st ). The xed sigma technique yields excellent prices, although they may be slightly too agressive, and is an excellent choice for swaption-like exotics such as Bermudan swaptions, callable inverser oaters, callable range notes, and other callables. Another substantial advantage is that the exotic prices do not depend on any adjustable parameters (since 0 scales out). This is important for a nightly MTM application.

1.2.5. Fixed sigma/caplet, xed sigma/short swaption, and xed sigma/swaption column. For all three methods, we again x (t) to be a constant 0 , typically 0.0001, and set (1.41a) (t) = 2 0 t.

The market prices of all instruments are independent of the value of a0 chosen. We eliminate the second LGM invariance by setting H (tp n ) to zero: (1.41b) H (tp n ) = 0.
10

There are actually two distinct xed sigma/caplet methods. The rst calibrates to the caplet (exercise dates) set of caplets. This method determines H (t) on the start dates of the caplets (1.42)
2 m p H (t1 st ), H (tst ), . . . , H (tst ), and H (tn ) = 0.

Piecewise linear interpolation (linear extrapolation beyond the end points) is used to nd H (t) for other values of t. As above, this is accomplished by a backwards bootstrap scheme, rst calibrating the mth caplet m1 to determine H (tm ), ..., and nally calibrating the rst st ), then calibrating caplet m 1 to determine H (tst 1 caplet to determine H (tst ). At each stage we are determining a single parameter by tting a single market price. The second xed sigma/caplet method is identical except that it uses the caplet (pay dates) set of caplets, and determines H (t) on the dates
c c p H (tc 0 ), H (t1 ), . . . , H (tn ), and H (tn ) = 0

by using a backwards bootstrapping scheme. Both the xed sigma/caplet techniques yield excellent prices for exotic deals with strong caplike characteristics, and the exotic prices do not depend on any adjustable parameters. Similarly, the xed sigma/short swaptions method uses either the short swaptions (exercise dates) set of swaptions or the short swaptions (pay dates) set of swaptions, and the xed sigma/swaption column method uses either the swaption column (exercise dates) or swaption column (pay dates) sets of swaptions. In all four cases, a backward bootstrap calibration method is used to consecutively determine H (t) at the start dates tk st of the swaptions, starting with the last swaption and working backwards. Exotic prices obtained from the calibrated model will not depend on any arbitrary parameters, and excellent pricing can be expected if the calibration instruments are representative of the exotic deal. 1.2.6. Fixed sigma/swaption row method. d 1.2.7. Row swaption/diagonal swaptions method. d 1.2.8. Short swaptions/diagonal swaptions method. d 1.2.9. Column swaptions/diagonal swaptions method. d 1.2.10. Caplets/diagonal swaptions method. d 1.3. LGM formulas. Here we summarize the formulas of the LGM model. 2. Black Karasinski. 2.1. General theory. Using the money market as a numeraire, the Black Karasinski model is usually written as o n T ( ))d ( ,X (T )) (t) = x , V (T, X (2.1a) V (t, x) = E e t r X where the state variable evolves according to (2.1b) = [(t) (t)X ] dt + (t) dW , dX (0) = 0, X

and the short rate is given by (2.1c) = log r. X

. We prefer the equivalent formulation in terms of the state variable X n T o ( ))d ( ,Y (T )) (t) = y , (2.2a) V (t, y ) = E e t r V (T, Y Y
11

where the state variable evolves according to (2.2b) and the short rate is given by (2.2c)
) = A(t)eb(t)Y r(t, Y

= (t) dW , dY

(0) = 0, Y

, log r = log A(t) + b(t)Y

. As written, the Black Karasinski model is not overly ecient: in terms of the short rate Y By using a money market numeraire, we must evaluate deals by stepping backward through the (t) explicitly (it is a simple Gaussian), this tree. Even though we know the transition density for Y knowledge is wasted by using the money market as a numeraire; There is no obvious closed form formula for zero coupon bonds, so we need to roll back through the tree for each payment. The objective of our analysis is to overcome both of these ineciencies by using singular perturbation methods to develop a closed form (approximate) formula for zero coupon bonds, and to switch from the money market numeraire to a zero coupon bond numeraire. The key to interest rate modeling is correct scaling, which enables us to understand and exploit the sizes of the dierent terms in the model. Recall that interest rates typically change by 0.80% or so over the course of a year. Thus, a 6.2% rate environment would typically remain within the range 5.4% to 7% during + 1 2 A(t)b2 (t)Y 2 + . We choose to the next year. Expanding (2.2c) shows that r A(t) + A(t)b(t)Y 2 scale (t) to be O(1). Then A(t) has to be of the orderThus Equivalently, the standard deviation of H 0 (T )x should grow by 1% or less per year. Accordingly, we scale to be O(1%) and take H (T ), H 0 (T ), and x to be O(1). Note that the last term in (1.4) is O(104 ). 2.1.1. Zero coupon bonds. We can eliminate the second problem by using singular perturbation methods to derive a closed form (approximate) formula for zero coupon bond prices. Imagine that we are at (t) = y. Dene some time t in the future, and that Y (2.3a) so (2.3b) The value of a zero coupon bond is Z (t, y ; T ) = e
T
t

(t0 ) Y (t) = Y (t0 ) y. q (t0 ) = Y (t0 ), dq (t0 ) = (t0 )dW

q (t) = 0.

A( )eb( )y d

Observe that (2.4a)

o n T 2 2 3 3 A( )eb( )y [b( ) q ( )+ 1 q 2 ( )+ 1 q 3 ( )+ ]d 2 b ( ) 6 e b ( ) (t) = 0 q E e t I1 = Z


T

b( )A( )eb( )y q ( )d

is Gaussian with mean zero and variance (2.4b) Thus (2.5a) o n T T 2 by 0 2 ( )y 1 2 T b( ) q ( )d (t) = 0 = e 2 t ( )( bAe dt ) d . q E e t A( )e
12

2 I1 = 2

( )

T 0 0

b(t )A(t )e

b(t0 )y

dt

!2

d .

Also, (2.5b) where (2.6) Thus the zero coupon bond prices are (2.7) Z (t, y ; T ) = e e
T 2 T b( )y [ ( ) (t)]( b( 1 )A( 1 )eb( 1 )y d 1 )d e+ t b( )A( )e 2 T 1 b( )A( )eb( )y [ ( ) (t)]d +O(4 ) 2 t T
t

1 2 2

) Z T b( )y 2 2 1 2 A( )e b ( ) q ( )d q (t) = 0 = 2 A( )eb( )y b2 ( )[ ( ) (t)]d t (t) = Z


t

2 ( )d .

A( )eb( )y d

Matching todays yield curve requires setting Z (0, 0; T ) equal to todays discount curve: (2.8a) D(0, T ) = e
T
0

T
0

A( ){12 b( ) ( )

T

2 2 4 b( 1 )A( 1 )d 1 + 1 2 b ( ) ( )+O ( )}d

Since D(0, T ) = e (2.8b) so (2.8c) Therefore (2.9)

f0 ( )d

f0 (T ) = A(T ) 1 2 b(T ) (

, where f0 (T ) is todays instantaneous forward rate curve, we have ( ) Z


T 0 2 2 4 A( )b( ) ( )d + 1 2 b (T ) (T ) + O ( ) ,

A(T ) = f0 (T ) 1 + b(T )

f0 ( )b( ) ( )d

1 2 2 2 b (T ) (T )

+ O ( ) .

Z (t, y ; T ) = D(t, T )e

T 2 T b( )y [ ( ) (t)]( b( 1 )A( 1 )eb( 1 )y d 1 )d e+ t b( )A( )e b( )y 1 2 T [ ( ) (t)]d +O(4 ) e 2 t b( )A( )e .

T
t

2 2 A( )[eb( )y 1]d + 1 2 [H2 (T )H2 (t)[H1 (T )H1 (t)] ] (t)

13

You might also like