Wilmott magazine 63

^
TECHNICAL ARTICLE 2
Pricing CMS Spread Options
and Digital CMS Spread
Options with Smile
We start by presenting the current approach used in different banks,
then we propose two different methods to take into account the smile.
The first method consists in changing the strike where the volatility of
each underlying is taken and represents only a partial modeling of the
smile. The second method takes into account the full smile of each
underlying and involves some numerical integration. These two meth-
ods are used to show the errors generated by the old approach. In the
last section, we extend the two methods to CMS underlyings, we give
some ideas how to generate an artificial smile and used the same
approach above.
II Notations
The following notations are used throughout this document.
Lets consider two assets F
1
and F
2
and an option of maturity T
depending on those two assets. We assume that under the T forward
probability, each F
i
(i = 1, 2) follows a lognormal process according to
the stochastic differential equation:
dF
i
(t)
F
i
(t)
= (F
i
(t), t)dt +
i
(t)dW
i
(t) (1)
I Introduction
This document deals with the smile of spread options in the Black
framework. The price of spread options is sensitive to the entire smile of
both underlyings. The classical approach uses the Black model without
smile. For each underlying, the corresponding at-the-money volatility is
taken. This approach ignores the effect of the smile and this is even more
of a problem when we deal with digital options, as in this case there is a
smile effect caused directly by the slope of the smile.
In general no closed formula exists for pricing a spread option when
the strike is different to zero. We dont focus in this paper on the numer-
al method. A very detailed survey on the valuation of spread option is
given in Carmona and Durrleman [2003].
Dempster and Hong [2001] propose to use the Fast Fourier Transform
(FFT) with stochastic volatility and interest rate environments
Alexander and Scourse [2003] propose to value spread options with a
bivariate normal mixture distribution.
An interesting study has been done see Cherubini and Luciano [2002]
where a non Gaussian copula has been proposed to associate the margin-
al distribution. This copula is calibrated using historical data.
The aim of this paper is to develop a simple approach, easy to imple-
ment with exogenous input smile with some application on CMS product.
Mourad Berrahoui
Commerzbank London
64 Wilmott magazine
Correlation between the two assets is represented by the fact that the two
standard Brownian processes in equation (1) satisfy:
E[dW
1
dW
2
] = dt (2)
A spread option also called crack spreads, due to their use in the oil
industry gives the holder the right to exchange F
2
for F
1
at expiry. The
payoff is:
payoff = Max(Q
1
F
1
Q
2
F
2
K; 0) (3)
where Q
1
is the quantity of asset F
1
, Q
2
the quantity of asset F
2
, and K the
strike.
III Current Approach without Smile
III.1 Spread option with zero strike
When K = 0, a closed-form formula exists (Margrabe, 1978). We assume
that the drift in equation (1) is deterministic. The price P of this option is:
P = Q
1
F
1
B(0, T)e
1
N(d
1
) Q
2
F
2
B(0, T)e
2
N(d
2
) (4)
where
d
1
=
ln(Q
1
F
1
/Q
2
F
2
) + (
1

2
+
2
/2)T

T
d
2
= d
1

i
=
_
T
0

i
(t)dt; i = 1, 2
=
_

2
1
+
2
2
2
1

i
=
_
1
T
_
T
0

i
(t)dt; i = 1, 2
and B(0, T) is the price of a zero coupon of maturity T.
III.2 Spread option with non-zero strike
III.2.1 Theoretical price of a spread option
To calculate the spread option price in the case where K = 0, it is neces-
sary to write equations (1) and (2) differently to use only independent
Brownian motions

W
1
and

W
2
, as follows:
dF
1
F
1
=
1
dt +
1
_
d

W
1
t
+
_
1
2
d

W
2
t
_
(5)
dF
2
F
2
=
2
dt +
2
dW
1
t
(6)
The price P is the discounted expectation of payoff (3) under the T for-
ward probability Q
T
where T is the maturity of the option we want to
price:
P = B(0, T) E
QT
[Max(Q
1
F
1
(T) Q
2
F
2
(T) K; 0)] (7)
There is no closed-form formula but two different numerical methods
are available to calculate P: Monte Carlo and semi-analytical.
III.2.2 Montecarlo approach
We simulate the two processes F
1
and F
2
. The price P will corresponds to
the mean of (7) over the set of Monte Carlo paths.
III.2.3 Semi-analytical approach
Different approaches exist:
Apply a conditioning technique to turn the two-dimensional inte-
gral into a single one ( Ravindran [1993] , Shimko [1994] )
P = B(0, T) E
QT
{E
QT
[Max(Q
1
F
1
(T) Q
2
F
2
(T) K; 0)|F
2
(T)]} (8)
Fast Fourier transform (Carr and Madan [1999] , Dempster and Hong
[2001])
IV New Approach with Smile
IV.1 A simple way to take into account a partial smile
The problem with the formulas presented in the last section in the pres-
ence of smile is what volatility to use for each index. In general, we use
the volatility at the money for each underlying.
In some special cases it is possible to determine a strike at which to
take the volatility of each underlying rather than the money. Lets
assume that the asset F
2
is less volatile. So the spread option become sim-
ply a mono-underlying option and the volatility to use for F
1
correspond
to the strike F
2
(0) + K. On the other hand, if we suppose that the asset F1
is less volatile, then the volatility to use for F
2
corresponds to the strike
F
1
(0) K.
On the basis of this reasoning, we propose to use in general:
Vol(F
1
) = Vol(Strike = ATM(F
2
) + K)
Vol(F
2
) = Vol(Strike = ATM(F
1
) K)
We will show later how accurate this approximation is in comparison to
the habit to use at the money volatility in case of a deeply in/out money
option.
Just to give an example, imagine that we try to price a spread USD CMS
20Y and USD Libor 3M at 06/17/2003 (Libor 3M = 1.02%, Swap20Y =
4.299%) with strike equal to 3.279% (4.299%1.02%). When the option is at
money (as it is the case at the beginning of the trade), there is no differ-
ence between the two methods. However, when the spread moves, the
option becomes deeply out or in the money and the more convex the
smile, the greater the difference between the two methods. Even if the
option was dealt at zero strike, because the smile for the indexes LIBOR3M
and CMS20Y is quite different, the two methods still lead to different
prices.
^
Wilmott magazine 65
TECHNICAL ARTICLE 2
IV.2 How to take into account the entire smile
The formula given for the price of a spread option in the previous sec-
tions cannot be extended to calculate a price with smile. For this, we
need a more general expression for the price which does not assume that
F
1
and F
2
follow lognormal distributions. The following formula is true
independently of the distribution of the underlying:
C = B(0, T)
_
+
0
Prob(Q
1
F
1
(T) > x + K, Q
2
F
2
(T) x) dx (9)
where Prob(. . .) is the bivariate cumulative distribution with correlation
equal to .
In order to prove (9), we need the following proposition.
Proposition
The spread option payoff is a sum of product of digital options:
Max(Q
1
F
1
(T) Q
2
F
2
(T) K; 0) =
_
+
0
1
{Q1 F1 (T)>x+K}
1
{Q2 F2 (T)x}
dx (10)
Proof
We have just to change the boundary of the integral in (10) by
x < Q
1
F
1
(T) K
x Q
2
F
2
(T)
.
(9) is then obtained by taking the expectation of (10).
The integral in (9) can be calculated numerically using simple meth-
ods: Trapezoidal rule, Simpsons rule. . ., or high-order methods: Gauss,
Gauss-Kronrod.
All those methods involve approximating (9) in the discrete form:
P = B(0, T)

i
w
i
Prob(Q
1
F
1
(T) > x
i
+ K, Q
2
F
2
(T) x
i
) (11)
where w
i
is a series of quadrature weights.
We are now faced with the problem of calculating the probability in
(11) in presence of smile. The probability that one asset is above a fixed
strike can be retrieved easily from prices of call options. Here we need to
calculate a bivariate probability. By no arbitrage, we can found see
Cherubini and Luciano [2002] a lower and upper limit
P1 MIN(P1, P2) Prob(F
1
(T) > x + K, F
2
(T) x)
P1 MAX(P1 + P2 1, 0)
with
P1 = Prob(F
1
(T) > x + K)
P2 = Prob(F
1
(T) > x)
Theses limits represent the financial application of the minimal and
maximal copulas of the Frechet-Hoeffding inequality.
Copulas helps us to calculate the bivariate probability knowing the
marginal distribution for each underlying (call spread price), and for that
the following assumption is needed:
Gaussian copula assumption
Prob(F
1
(T) > x
1
, F
2
(T) x
2
|Full smile)
= Prob(F
1
(T) > x
1
, F
2
(T) x
2
|
1
=
1
(T, x
1
));
2
=
2
(T, x
2
)))
(12)
with x
1
and x
2
such that
Prob(F
1
(T) > x
1
|
1
=
1
(T, x
1
)) = Prob(F
1
(T) > x
1
|Full smile) (13)
Prob(F
2
(T) > x
2
|
2
=
2
(T, x
2
)) = Prob(F
2
(T) > x
2
|Full smile) (14)

1
(T, x
1
)denotes the implied volatility of F
1
(T) at strike x
1
and
2
(T, x
2
) the
implied volatility of F
2
(T) at strike x
2
.
This assumption means that we are using a Gaussian Copula to repre-
sent the joint distribution of the random variables F
1
(T) and F
2
(T).
The following algorithm, which relies on the Gaussian copula
assumption, can then be used to calculate the price of a spread option
with smile as in (11).
Algorithm
Calculate Prob(F
1
(T) > x
1
|Full smile), i = 1, 2, from the price of a call
spread.
Solve equations (13) and (14) for x
1
and x
2
.
Estimate from historical data for F
1
(t) and F
2
(t).
Calculate the joint distribution (normal bivariate) of F
1
(T) and F
2
(T)
using (12).
IV.3 Extension to CMS spread options
IV.3.1 Introduction
If we want to use the model we have proposed above, we need the smile
surface for each underlying. This smile is more or less known in the mar-
ket when the underlying is the short rate (Libor 1M,. . .,12M). But, when
the underlying is CMS, the smile is unknown. One idea is to use the swap-
tion smile with the swap maturity equal to the tenor of this CMS.
Unfortunately this strategy is not arbitrage free -in theory- particularly
when the CMS cap/floor and swap are liquid. In the last part of this sec-
tion, we propose an idea to build this smile using the prices of CMS
caps/floors and swaps. To introduce this idea, first we present the issues
involved in pricing CMS products, with a specific section about the tim-
ing adjustment necessary for CMS products with fixings in advance.
Then we expose a simple approach, widely used in banks, to price CMS
swaps and caps/floors using the whole smile of swaptions. This approach
is based on a simple idea of replication, which can be used for any com-
plex European payoff.
IV.3.2 Issues in pricing CMS products
Let us denote SR
t
the swap rate at time t. Its value at time t is:
SR
t
=
B(t, T
0
) B(t, T
N
)
N

i=1
B(t, T
i
)
i
66 Wilmott magazine
The swap is starts a time T
0
and its payments occur at times T
i
(i = 1,
. . . , N) with T = T
0
< T
1
< . . . < T
N
.
B(t, T
i
) is the price at time t of the bond which pay 1 unit at time T
i
.

i
=
Ti Ti1
365
if SR
t
is expressed in basis Act/365.
SR
t
is then a martingale (i.e. a driftless process) under the numeraire
SM
T
defined as:
SM
T
=
N

i=1
B(T, T
i
)
i
Prices of FRAs and caplets are given by:
FRA(t) = B(t, T) E
QT
[SR
T
|F
t
]
Caplet(t) = B(t, T) E
QT
[Max(SR
T
K; 0)|F
t
]
Q
T
denotes the T forward measure. Under this measure, SR
t
is not driftless
and it is difficult to calculate its drift.
The price of a physical swaption is given by:
Swaption(t) = E
SMT
[Max(SR
T
K, 0)|F
t
]
N

i=1
B(t, T
i
)
i
where E
SMT
denotes the expectation with respect to numeraire SM
T
.
We can apply the Black formula in this case, because SR
t
is driftless.
From this short analysis, we can see that if we can express the payoff
of FRAs/caps in terms of the payoff of the swaption, then pricing becomes
simple. It is the idea of the replication, which we develop now.
Note that in order to price a cash swaption, which is a more common
product than physical swaptions, one has to use instead of the physical
swap measure, the cash swap measure where the numeraire is:
SCashM
t
=
N

i=0
1
(1 + SR
T
)
i
IV.3.3 Replication of simple products on CMS
In this section we develop the idea of replicating the payoff of a CMS
swap or cap as a linear combination of swaptions with different strikes.
In addition to the mathematical argument of easy derivation given in the
last section, another motivation for doing this is that the only simple and
liquid way to hedge a product on CMS is using swaptions.
We want to write a linear payoff (swap/cap/floor) of the form:
Max(SR
T
K; 0)
in terms of a non-linear payoff (swaption with cash settlement) of the form:
Max(SR
T
K; 0)
N

i=1
1
(1 + SR
T
)
i
So the idea is the find a set of weights w
j
and strikes Kj such that:
Max(SR
T
K; 0) =

j
w
j
Max(SR
T
K
j
; 0)
N

i=1
1
(1 + SR
T
)
i
(15)
We choose the strikes Kj to be equally spaced, using a discretization step
. So we have:
K
j
= K + j; j = 1, . . . , M
In our experience, = 5 to 10 basis points is a good choice and M can be
chosen so that K is about 15%, but it really depends to what limit of strike
the trader wants to hedge its CMS products.
The calculus of the weight wi is straightforward.
IV.3.4 Timing adjustment for CMS products with fixings in advance
We have seen that the replication technique is based on swaptions with
cash settlement, so it can only be used to price CMS products in which
the swap rate is observed and paid at the same time. When we deal with
CMS product with fixings in advance, e.g. CMS vanilla caps/floors/swaps,
the price has to be adjusted.
If the swap rate is observed at time T and paid at T + , the forward
swap rate SR
0
has to be corrected by a timing adjustment (see Hull):

SR
0
R
0

R
T
1 + R
0

with R
0
is the value at time zero of the forward rate between T and
T + ,
R
is the volatility of this forward rate, is the at-the-money volatil-
ity of the forward swap rate and is the correlation between the forward
swap rate and the forward rate.
Example
Lets take the example given by Hull (see ref.).
SR
0
= 5%
R
0
= 5%
= 15%

R
= 20%
= 0.5
= 0.7
The forward rate has to be adjusted by 0.0000256 T.
IV.3.5 Building CMS smile by arbitrage
The process SR
T
can be written under the T-forward measure as follows:
SR
T
= E
QT
[SR
T
] exp
_

1
2

2
T + W
T
_
(16)
The application of the replication technique for FRAs gives the expecta-
tion value E
QT
[SR
T
] of SR
T
under the T-forward measure as:
E
QT
[SR
T
] =
FRA
B(0, T)
The price of the caplet/floorlet with strike K using the expression (16) of
the process SR
T
is simply given by Blacks formula:
Caplet = Black
_
E
Q
T [SR
T
], (K), T, K
_
^
Wilmott magazine 67
TECHNICAL ARTICLE 2
The unknown variable in this formula is the volatility (K). At the same
time this price can be obtained using the replication technique
described above. Hence we can imply the volatility (K) by:
(K) = Black
1
(Caplet)
We can apply this technique for every strike K and thus we build the CMS
smile.
We admit that it can be time consuming. At the first approximation we can take
swaption smile.
V Tests
V.1 Introduction
We first show the difference in price for short rate spread options
(Libor6MLibor3M), for given market data: yield curve and smile, with
the three methods:
The approach with taking at the money volatility for each index.
The same approach but with taking as strike for one index, the
money for the second index plus/minus the strike of the spread
option.
Pricing with full smile as described in this document.
Then we do similar tests on CMS products.
In all our tests, we use the following features:
Payment frequency : 6M
Day count : ACT/360
Yield curve:
V.2 Short rate spread option
We consider a spread option Libor6MLibor3M. First, we consider strike
zero and volatility flat. We compare the Margrabe closed-form formula
(without smile), the Monte Carlo approach (partial smile), and the full
smile method.
Libor 6M = 0.99 %
Libor 3M = 0.95 %
From the table below, we check that our model give the same results
as Margrabes formula in the case where the volatility is flat, for different
maturities.
ATM swap rate
1Y 1.14%
2Y 1.55%
5Y 2.63%
7Y 3.12%
10Y 3.61%
15Y 4.12%
20Y 4.36%
Volatility surface:
3% 4% 5% 6% 7% 8% 9%
1Y 26.50 21.90 23.30 25.50 26.70 27.80 29.10
2Y 25.80 20.30 19.10 21.30 22.70 23.80 25.20
5Y 23.90 19.50 15.50 15.00 15.60 16.40 17.60
7Y 22.70 18.70 14.80 13.30 13.40 14.00 14.70
10Y 21.50 17.90 14.10 12.10 12.10 12.60 13.10
15Y 20.20 17.00 13.50 11.20 10.90 11.30 11.60
20Y 19.30 16.40 13.10 10.60 10.30 10.70 11.00
Margrabe MC method Full smile
formula (10 000 path)
1Y 9 9 9
2Y 32 32 32
5Y 153 153 153
7Y 275 275 275
10Y 484 483 483
15Y 854 852 853
20Y 1181 1177 1176
TABLE 1: STRIKE = 0 , VOLATILITY IS
FLAT AT 20%, CORRELATION = 0.7
The small difference can be due to the numeric integration method
used in our implementation.
Now, we consider the same option but with smiled volatility. We
notice that the difference become significant when the maturity
increases.
Margrabe Full smile
1Y 12 12
2Y 42 41
5Y 176 170
7Y 278 275
10Y 436 436
15Y 665 690
20Y 860 901
TABLE 2: STRIKE = 0,
VOLATILITY WITH SMILE,
CORRELATION = 0.7
68 Wilmott magazine
V.3 Building CMS smile surface
We compare the swaption volatility smile (for 10Y fixed swap maturity) with
the CMS 10Y, after building the CMS smile as described in this document.
V.5 Impact of Smile in Digital CMS
We approximate a Digital option as a call spread with a strike shift equal
to 10 basis points. We compare the same three models again.
Margrabe Our approach
(in basis point)
1Y 4 4
2Y 24 23
5Y 131 126
7Y 217 216
10Y 347 356
15Y 552 583
20Y 725 773
TABLE 3: STRIKE = 0.20%
VOLATILITY WITH SMILE,
CORRELATION = 0.7
3% 4% 5% 6% 7% 8% 9%
1Y -0.2 0.1 0.3 0.3 0.4 0.5 0.6
2Y -0.4 -0.1 0.3 0.5 0.6 0.7 0.8
5Y -0.8 -0.5 -0.2 0.1 0.4 0.5 0.6
7Y -0.9 -0.7 -0.4 -0.1 0.1 0.2 0.3
10Y -0.9 -0.8 -0.5 -0.2 0.1 0.2 0.2
20Y -0.9 -0.9 -0.7 -0.4 -0.1 -0.1 0.1
TABLE 4: SMILE CMS 10 Y (VOLCMS10Y
VOLSWAPTION NX10Y)
In general the CMS smile is less than the swaption smile with the
same swap maturity.
V.4 Impact of smile in CMS spread option
We consider the spread option on CMS 20Y and CMS 2Y
with strike equal 2.5% (in but not far from the money).
The first column gives the price of the spread option
priced with the volatility at the money for each index and
the second column with partial smile. The third column
shows the price with full smile. In the first two columns, the
price is calculated with Monte Carlo.
It is clear from table 5 that the model with partial smile
is closer to the full smile model than the classical approach
without smile.
Those differences depends on
the convexity of the smile
how far is the strike of the spread option from the
money.
Vol at the money Partial smile Full smile
1Y 47 46 47
2Y 79 80 84
5Y 139 155 166
7Y 170 194 210
10Y 211 250 269
15Y 273 334 351
20Y 339 415 428
TABLE 5: SPREAD OPTION ON CMS 20Y
AND CMS 2Y WITH STRIKE = 2.50% AND
CORRELATION = 0.7
Vol at the money Partial smile Full smile
1Y 98 98 98
2Y 179 180 176
5Y 320 309 308
7Y 378 361 363
10Y 441 417 421
15Y 508 478 484
20Y 559 528 527
TABLE 6: CALL DIGITAL OPTION ON CMS20Y
AND CMS2Y WITH STRIKE = 1.50% AND
CORRELATION = 0.7
The graphs below show the differences between the prices with the
different models as presented in the tables.
Difference partial/without smile and full smile methods
(strike = 1.5%)
12
7
2
3
8
13
18
23
28
33
1Y 2Y 5Y 7Y 10Y 15Y 20Y
Maturity
Price (in b p)
Without smile method
Partil smile method
Digital spread option
Wilmott magazine 69
TECHNICAL ARTICLE 2
W
These graphs show that taking the volatility at the right strike (par-
tial smile) gives closer prices to the full smile method especially when
the option is deeply in or at the money.
For digital option at the money, the differences between the two
models however are significant.
VI Conclusions
In this document we have exposed two new methods to take into account
the smile for spread options and in particular digital spread options.
The most advanced of those two methods is a numerical integration
method based on a copula assumption, which uses the entire smile of
each underlying.
If the smile is not smooth enough, this method can lead to instabili-
ties. This is why, when this situation occurs, a parameterization of the
smile and then using a closed-form formula for Pr ob(F
i,t
> x
i
; smile(F
i,t
))
could be a worthwhile alternative. For a digital option, in this
case one needs to consider:
dC
dK

K=K0
=
C
K

K=K0
+
C

K=K0

d
dK

K=K0
with C(K, (K)) is the call option price and (K) is a paramet-
ric volatility function
(Example: SABR model).
Another method which we propose is to price spread
options taking the volatility at a different strike than the
money of each underlying the same time, as follows:
Vol(F
1
) = Vol(Strike = ATM(F
2
) + K)
Vol(F
2
) = Vol(Strike = ATM(F
1
) K)
This method is only a partial smile model but we show that it
is close to the first, full smile, method.
A separate section of this document is dedicated to deal-
ing with CMS underlyings and building the CMS smile.
Difference partial/without smile and full smile methods
(strike = 2.9%)
90
80
70
60
50
40
30
20
10
0
10
1Y 2Y 5Y 7Y 10Y 15Y 20Y
Maturity
Price (in b p)
Without smile method
Partial smile method
Digital spread option
Difference partial/without smile and full smile methods
(strike = 3.5%)
10
8
6
4
2
0
2
4
6
8
10
1Y 2Y 5Y 7Y 10Y 15Y 20Y
Maturity
Price (in b p)
Without smile method
Partial smile method
Digital spread option
IC. Alexander and A. Scourse [2003] Bivariate Normal Mixture
Spread Option Valuation ISMA Centre Discussion Papers in Finance
200315.
IR. Carmona and V Durrleman [2003] Pricing And Hedging Spread
Options in A Log-Normal Model Working paper.
IR. Carmona and V. Durrleman [2003] Pricing And Hedging Spread
Options SIAM Review, 45:4 627687.
IP. Carr and.D.B Madan [1999] Option Valuation Using the Fast
Fourier Transform, Journal of Computational Finance, 2, 4, 6173.
IU. Cherubini and E. Luciano [2002] Multivariate Option Pricing With
Copulas Working paper.
IS. Coutant, V. Durrleman, G. Rapuch and T. Roncalli [2001]
Copulas, multivariate risk-neutral distributions and implied dependence
functions Working paper.
IM. Dempster and G. Hong Pricing Spread Options with the Fast
Fourier Transform Risk 2001 Europe.
IJohn C. Hull Options, Futures & Other Derivatives.
IG. Rapuch and T. Roncalli [2001] Some remarks on two-asset
options pricing and stochastic dependence of asset prices Working
paper.
IK. Ravindran [1993] Low Fat spread Risk Magazine 6:10, 667.
IJ. Rosenberg [2003] Non parametric Pricing of Multivariate
Contingent Claims Research and Market Analysis Group, Federal
Reserve Bank of New York.
ID. Shimko [1994] Option on Futures Spreads: Hedging, Speculation
and Valuation The journal of Futures Markets 14:2,182213.
IPaul Wilmott, Derivatives The Theory and Practice of Financial
Engineering.
VII REFERENCES