Patrick S.

Hagan
IN THE TRENCHES

Convexity Conundrums: Pricing

CMS Swaps, Caps, and Floors*
Bear, Stearns & Company 383 Madison Avenue New York, NY 10179
phagan@bear.com
1 Introduction
Im sure weve all been there: Were in hot competition with another
bank over a deal. As the deal evolves, our trading team starts getting
pushed around the market, and it dawns on us that the other banks
pricing is better than ours, at least for this class of deals. We could fix
this problem by inventing a universal method for achieving the best
possible prices for all deal types. That topic will be covered in a future
column, next to the column on Elvis sightings. Here we focus on a single class of deals, the constant maturity swaps, caps, and floors. We
develop a framework that leads to the standard methodology for pricing these deals, and then use this framework to systematically improve
the pricing.
Let us start by agreeing on basic notation. In our notation, today is
always t = 0. We use
Z(t; T) = value at date t of a zero coupon bond with maturity T , (1.1a)
D(T) Z(0, T) = today s discount factor for maturity T.

(1.1b)

We distinguish between zero coupon bonds and discount factors to

remind ourselves that discount factors are not random, we can always
obtain the current discount factors D(T) by stripping the yield curve,
while zero coupon bonds Z(t, T) remain random until the present catches
up to date t. We also use
cvg(tst , tend , dcb)

(1.2)

to denote the coverage (also called the year fraction or day count fraction) of
the period tst to tend , where dcb is the day count basis (Act360, 30360, . . .)

specified by the contract. So if interest accrues at rate R , then

cvg(tst , tend , dcb)R is the interest accruing in the interval tst to tend .
1.1 Deal definition
Consider a CMS swap leg paying, say, the N year swap rate plus a margin
m. Let t0 , t1 , . . . , tm be the dates of the CMS leg specified in the contract.
(These dates are usually quarterly). For each period j, the CMS leg pays
j (R j + m)

paid at t j

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

(1.3a)

where R j is the N year swap rate and

j = cvg(t j1 , t j , dcbpay )

(1.3b)

is the coverage of interval j. If the CMS leg is set-in-advance (this is standard), then R j is the rate for a standard swap that begins at t j1 and ends
N years later. This swap rate is fixed on the date j that is spot lag business
days before the interval begins at t j1 , and pertains throughout the interval, with the accrued interest j (R j + m) being paid on the intervals end

tj

dj

t j1

tj
interval j

Fig. 1.1. jth interval of a set-inadvance CMS leg.

* The views presented here are soley the views of the author, and do not necessarily reflect the views of Bear-Stearns or any of its affiliates or subsidiaries.

38

Wilmott magazine

date, t j . Although set-in-advance is the market standard, it is not

uncommon for contracts to specify CMS legs set-in-arrears. Then R j is the
N year swap rate for the swap that begins on the end date t j of the interval, not the start date, and the fixing date j for R j is spot lag business
days before the interval ends at t j . As before, j is the coverage for the jth
interval using the day count basis dcbpay specified in the contract. Standard practice is to use the 30360 basis for USD CMS legs.
CMS caps and floors are constructed in an almost identical fashion.
For CMS caps and floors on the N year swap rate, the payments are
j [R j K]+
j [K R j ]+

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

paid at t j

(cap),

Crudely speaking, the level L(t) represents the value at time t of receiving
$1 per year (paid annually or semiannually, according to the swaps frequency) for N years. With this definition, the value of the swap is
Vsw (t) = [Rs (t) Rfix ]L(t),

(1.9a)

Z(t; s0 ) Z(t; sn )
.
L(t)

(1.9b)

where
Rs (t) =

Clearly the swap is worth zero when Rfix equals Rs (t), so Rs (t) is the par
swap rate at date t. In particular, todays level

(1.4a)
L0 = L(0) =

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

paid at t j

(floor),

(1.4b)

where the N year swap rate is set-in-advance or set-in-arrears, as specified

in the contract.
1.2 Reference swap
The value of the CMS swap, cap, or floor is just the sum of the values of
each payment. Any margin payments m can also be valued easily. So all
we need do is value a single payment of the three types,
Rs

paid at tp ,

(1.5a)

[Rs K]+

paid at tp ,

(1.5b)

[K Rs ]+

paid at tp .

(1.5c)

Here the reference rate Rs is the par rate for a standard swap that
starts at date s0 , and ends N years later at sn . To express this rate mathematically, let s1 , s2 , . . . , sn be the swaps (fixed leg) pay dates. Then a swap
with rate Rfix has the fixed leg payments
j Rfix

paid at s j

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

(1.6a)

where
j = cvg(tj1 , t j , dcb sw )

(1.6b)

is the coverage (fraction of a year) for each period j, and dcb sw is the standard swap basis. In return for making these payments, the payer receives
the floating leg payments. Neglecting any basis spread, the floating leg is
worth 1 paid at the start date s0 , minus 1 paid at the end date sn . At any
date t, then, the value of the swap to the payer is
Vsw (t) = Z(t; s0 ) Z(t; sn ) Rfix

n


j Z(t; s j ).

(1.7)

j=1

Wilmott magazine

j Z(t; s j ).

(1.8)

n


j D(s j ),

(1.10a)

j=1

and todays (forward) swap rate

R0s = Rs (0) =

D0 Dn
L0

(1.10b)

are both determined by todays discount factors.

2 Valuation
According to the theory of arbitrage free pricing, we can choose any
freely tradeable instrument as our numeraire. Examining 1.8 shows that
the level L(t) is just the value of a collection zero coupon bonds, since the
coverages j are just fixed numbers. These are clearly freely tradeable
instruments, so we can choose the level L(t) as our numeraire.1 The usual
theorems then guarantee that there exists a probability measure such
that the value V(t) of any freely tradeable deal divided by the numeraire
is a Martingale. So
 

V(T) 
Ft
for any T > t,
V(t) = L(t)E
(2.1)
L(T) 
provided there are no cash flows between t and T .
It is helpful to examine the valuation of a plain vanilla swaption. Consider a standard European option on the reference swap. The exercise
date of such an option is the swaps fixing date , which is spot-lag business days before the start date s0 . At this exercise date, the payoff is the
value of the swap, provided this value is positive, so
Vopt ( ) = [Rs ( ) Rfix ]+ L( )

(2.2)

on date . Since the Martingale formula 2.1 holds for any T > t, we can
evaluate it at T = , obtaining

Vopt (t) = L(t)E

 
 

Vopt ( ) 
Ft = L(t)E [Rs ( ) Rfix ]+  Ft .

L( )

In particular, todays value of the swaption is

 

Vopt (t) = L0 E [Rs ( ) Rfix ]+  F0 .

(2.3)

(2.4a)

L(t) =

jD j =

j=1

j=1

The level of the swap (also called the annuity, PV01, DV01, or numerical duration) is defined as
n


n


39

PATRICK S.HAGAN

Moreover, 1.9b shows that the par swap rate Rs (t) is the value of a freely
tradable instrument (two zero coupon bonds) divided by our numeraire.
So the swap rate must also a Martingale, and
E { Rs ( )| F0 } = Rs (0) R0s .

(2.4b)

To complete the pricing, one now has to invoke a mathematical

model (Blacks model, Hestons model, the SABR model, . . . ) for how Rs ( )
is distributed around its mean value R0s . In Blacks model, for example,
the swap rate is distributed according to
Rs ( ) = R0s e x

1 2

2
,

Z( ; tp )/L( ) = G(Rs ( )),

2.1 CMS caplets

The payoff of a CMS caplet is

D(tp )/L0 = G(R0s ),

(2.11b)

for some function G(Rs ). The convexity correction is then just the expected value

 


G(Rs ( ))
+
1  F0
cc = D(tp )E [Rs ( ) K]
(2.12)
G(R0 )
s

[Rs ( ) K]+

paid at tp .

(2.6)

On the swaps fixing date , the par swap rate Rs is set and the payoff is
known to be [Rs ( ) K]+ Z( ; tp ) , since the payment is made on tp . Evaluating 2.1 at T = yields
 

[Rs ( ) K]+ Z( ; tp ) 
CMS
Vcap
(t) = L(t)E
(2.7a)
 Ft .
L( )


 
[Rs ( ) K]+ Z( ; tp ) 
 F0 .
L( )

(2.7b)
=

The ratio Z( ; tp )/L( ) is (yet another!) Martingale, so its average value

is todays value:
 

E Z( ; tp )/L( ) F0 = D(tp )/L0 .
(2.8)
By dividing Z( ; tp )/L( ) by its mean, we obtain
 

Z( ; tp )/L( ) 
CMS
F0 ,
Vcap
(0) = D(tp )E [Rs ( ) K]+
D(tp )/L0 

which can be written more evocatively as

 

CMS
Vcap
(0) = D(tp )E [Rs ( ) K]+  F0


 

Z( ; tp )/L( )
+
+ D(tp )E [Rs ( ) K]
1  F0 .
D(tp )/L0

over the swap rate Rs ( ). The second step is to evaluate this expected
value.
In the appendix we start with the street-standard model for expressing L( ) and Z( ; tp ) in terms of the swap rate Rs . This model uses bond
math to obtain
Rs
1
G(Rs ) =
.
(2.13a)
1
(1 + Rs /q)
1
(1 + Rs /q)n
Here q is the number of periods per year (1 if the reference swap is annual, 2 if it is semi-annual, . . .), and

In particular, todays value is

(2.9)

(2.10)

The first term is exactly the price of a European swaption with

notional D(tp )/L0 , regardless of how the swap rate Rs ( ) is modeled. The

40

(2.11a)

(2.5)

where x is a normal variable with mean zero and unit variance. One
completes the pricing by integrating to calculate the expected value.

CMS
Vcap
(0) = L0 E

last
term is the
convexity correction. Since Rs ( ) is a Martingale and

Z( ; tp )/L( ) / Z(t; tp )/L(t) 1 is zero on average, this term goes to zero

linearly with the variance of the swap rate Rs ( ), and is much, much
smaller than the first term.
There are two steps in evaluating the convexity correction. The first
step is to model the yield curve movements in a way that allows us to rewrite the level L( ) and the zero coupon bond Z( ; tp ) in terms of the
swap rate Rs . (One obvious model is to allow only parallel shifts of the
yield curve.) Then we can write

tp s0
s1 s0

(2.13b)

is the fraction of a period between the swaps start date s0 and the pay
date tp . For deals set-in-arrears  = 0. For deals set-in-advance, if the
CMS leg dates t0 , t1 , . . . are quarterly, then tp is 3 months after the start
date s0 , so  = 12 if the swap is semiannual and  = 14 if it is annual.
In the apprendix we also consider increasingly sophisticated models
for expressing L( ) and Z( ; tp ) in terms of the swap rate Rs , and obtain
increasingly sophisticated functions G(Rs ).
We can carry out the second step by replicating the payoff in 2.12 in
terms of payer swaptions. For any smooth function f (Rs ) with f (K) = 0,
we can write
f  (K)[Rs K]+ +


[Rs x]+ f  (x)dx =

f (Rs )
0

for Rs > K
.
for Rs < K

(2.14)

Choosing

f (x) [x K]

G(x)

1
,
G(R0s )

(2.15)

Wilmott magazine

and substituting this into 2.12, we find that


 

cc = D(tp ) f  (K) E [Rs ( ) K]+  F0


 

+
f  (x) E [Rs ( ) x]+  F0 dx .

where
fatm (x) [x

(2.16)

Together with the first term, this yields

CMS
Vcap
(0) =





D(tp )

1 + f  (K) C(K) +
C(x)f  (x)dx ,
L0
K

(2.17a)

 

C(x) = L0 E [Rs ( ) x]+  F0

(2.17b)

is the value of an ordinary payer swaption with strike x.

This formula replicates the value of the CMS caplet in terms of European swaptions at different strikes x. At this point some pricing systems
break the integral up into 10bp or so buckets, and re-write the convexity
correction as the sum of European swaptions centered in each bucket.
These swaptions are then consolidated with the other European swaptions in the vanilla book, and priced in the vanilla pricing system. This
replication method is the most accurate method of evaluating CMS legs.
It also has the advantage of automatically making the CMS pricing and
hedging consistent with the desks handling of the rest of its vanilla book.
In particular, it incorporates the desks smile/skew corrections into the
CMS pricing. However, this method is opaque and compute intensive.
After briefly considering CMS floorlets and CMS swaplets, we develop simpler approximate formulas for the convexity correction, as an alternative
to the replication method.
2.2 CMS floorlets and swaplets
Repeating the above arguments shows that the value of a CMS floorlet is
given by


 K

D(tp )

CMS
1 + f  (K) P(K)
Vfloor
(0) =
P(x)f  (x)dx ,
(2.18a)
L0

+

F0

(2.18b)

(2.19b)

paid at tp ,

(2.20)

the payoff of a CMS caplet minus a CMS floorlet is equal to the payoff of a
CMS swaplet minus K . Therefore, the value of this combination must be
equal at all earlier times as well:
CMS
CMS
CMS
Vcap
(t) Vfloor
(t) = Vswap
(t) KZ(t; tp )

(2.21a)

CMS
CMS
CMS
Vcap
(0) Vfloor
(0) = Vswap
(0) KD(tp ).

(2.21b)

In particular,

Accordingly, we can price an in-the-money caplet or floorlet as a swaplet

plus an out-of-the-money floorlet or caplet.

3 Analytical formulas
The function G(x) is smooth and slowly varying, regardless of the model
used to obtain it. Since the probable swap rates Rs ( ) are heavily concentrated around R0s , it makes sense to expand G(x) as
G(x) G(R0s ) + G(R0s )(x R0s ) + .

(3.1a)

For the moment, let us limit the expansion to the linear term. This makes
f (x) a quadratic function,
G(R0s )
(x R0s )(x K),
G(R0s )

(3.1b)

and f (x) a constant. Substituting this into our formula for a CMS caplet
(2.17a), we obtain



D(tp )
CMS
Vcap
(0) =
C(K) + G(R0s ) (K R0s )C(K) + 2
C(x)dx , (3.2)
L0
K
where we have used G(R0s ) = D(tp )/L0 . Now, for any K the value of the
payer swaption is
 

C(K) = L0 E [Rs ( ) K]+  F0 ,

(3.3a)

is the value of the ordinary receiver swaption with strike x. Thus, the
CMS floolets can also be priced through replication with vanilla
receivers. Similarly, the value of a single CMS swap payment is


 R0s

D(tp )
CMS
0


Vswap (0) = D(tp )Rs +
C(x)fatm (x)dx +
P(x)fatm (x)dx ,
L0

R 0s
(2.19a)

Wilmott magazine

is the same as f (x) with the strike K replaced by the par swap rate R0s .
Here, the first term in 2.19a is the value if the payment were exactly
equal to the forward swap rate R0s as seen today. The other terms represent the convexity correction, written in terms of vanilla payer and
receiver swaptions. These too can be evaluated by replication.
It should be noted that CMS caplets and floorlets satisfy call-put parity. Since

f (x)

where f (x) is the same function as before (see 2.15), and where
P(x) = L0 E [x Rs ( )]

G(x)
1
G(R0s )

[Rs ( ) K]+ [K Rs ( )]+ = Rs ( ) K

as the value of the CMS caplet, where


R0s ]

41

PATRICK S.HAGAN

so the integral can be re-written as



C(x)dx = L0 E

 


[Rs ( ) x]+ dx F0
K

2  
1
= L0 E [Rs ( ) K]+  F0 .
2

(3.3b)

caplets and floorlets, the volatility K for strike K should be used, since
the swap rates Rs ( ) near K provide the largest contribution to the
expected value. For in-the-money options, the largest contributions come
from swap rates Rs ( ) near the mean value R0s . Accordingly, call-put parity should be used to evaluate in-the-money caplets and floorlets as a CMS
swap payment plus an out-of-the-money floorlet or caplet.

Putting this together yields

CMS
Vcap
(0) =

 



D(tp )
C(K) + G (R0s )L0 E Rs ( ) R0s [Rs ( ) K]+  F0 (3.4a)
L0

for the value of a CMS caplet, where the convexity correction is now the
expected value of a quadratic payoff. An identical arguments yields the
formula
CMS
Vfloor
(0) =

 



D(tp )
P(K) G (R0s )L0 E R0s Rs ( ) [K Rs ( )]+  F0
(3.4b)
L0

for the value of a CMS floorlet. Similarly, the value of a CMS swap payment works out to be
CMS
Vswap
(0) = D(tp )R0s + G (R0s )L0 E



2  
Rs ( ) R0s  F0 .

(3.4c)

To finish the calculation, one needs an explicit model for the swap
rate Rs ( ). The simplest model is Blacks model, which assumes that the
swap rate Rs ( ) is log normal with a volatility . With this model, one
obtains

 2  2
CMS
e 1
Vswap
(0) = D(tp )R0s + G (R0s )L0 R0s
(3.5a)
for the CMS swaplets,
CMS
Vcap
(0) =


D(tp )
2
C(K) + G (R0s )L0 (R0s )2 e N (d3/2 )
L0

R0s (R0s + K)N (d1/2 ) + R0s KN (d1/2 )

4 Conclusions
The standard pricing for CMS legs is given by 3.5a3.5d with G(Rs ) given
by 2.13a. These formulas are adequate for many purposes. When finer
pricing is required, one can systematically improve these formulas by
using the more sophisticated models for G(Rs ) developed in the Appendix, and by adding the quadratic and higher order terms in the expansion 3.1a. In addition, 3.4a3.4b show that the convexity corrections are
essentially swaptions with quadratic payoffs. These payoffs emphasize
away-from-the-money rates more than standard swaptions, so the convexity corrections can be quite sensitive to the markets skew and smile.
CMS pricing can be improved by replacing Blacks model with a model
that matches the market smile, such as Hestons model or the SABR
model. Alternatively, when the very highest accuracy is needed, replication can be used to obtain near perfect results.

Appendix A. Models of the yield curve

A.1 Model 1: Standard model
The standard method for computing convexity corrections uses bond
math approximations: payments are discounted at a flat rate, and the
coverage (day count fraction) for each period is assumed to be 1/q, where
q is the number of periods per year (1 for annual, 2 for semi-annual, etc).
At any date t, the level is approximated as

(3.5b)
L(t) = Z(t, s0 )

j=1

for CMS caplets, and

CMS
Vfloor
(0) =


D(tp )
2
P(K) G (R0s )L0 (R0s )2 e N (d3/2 )
L0
R0s (R0s + K)N (d1/2 ) + R0s KN (d1/2 )

d =

ln R0s /K + 2
.


Z(t, s j )
1/q
Z(t, s0 )
,
Z(t, s0 )
[1 + Rs (t)/q] j
n

(A.1)

j=1

which works out to



(3.5c)

for CMS floorlets. Here

(3.5d)

The key concern with Blacks model is that it does not address the
smiles and/or skews seen in the marketplace. This can be partially mitigated by using the correct volatilities. For CMS swaps, the volatility ATM
for at-the-money swaptions should be used, since the expected value 3.4c
includes high and low strike swaptions equally. For out-of-the-money

42

n


L(t) =



1
Z(t, s0 )
.
1
Rs (t)
(1 + Rs (t)/q)n

(A.2a)

Here the par swap rate Rs (t) is used as the discount rate, since it represents the average rate over the life of the reference swap. In a similar
spirit, the zero coupon bond for the pay date tp is approximated as
Z(t; tp )

Z(t, s0 )
(1 + Rs (t)/q)

where
=

tp s0
s1 s0

(A.2b)

(A.2c)

Wilmott magazine

is the fraction of a period between the swaps start date s0 and the pay
date tp . Thus the standard bond math model leads to
G(Rs ) =

Z(t; tp )
Rs

L(t)
(1 + Rs /q)

1
1

1
(1 + Rs /q)n

(A.3)

This method a) approximates the schedule and coverages for the reference swaption; b) assumes that the initial and final yield curves are
flat, at least over the tenor of the reference swaption; and c) assumes a
correlation of 100% between rates of differing maturities.

where x is the amount of the parallel shift. The level and swap rate Rs are
given by
n

D(s j ) (s j s0 )x
L(t)
=
e
j
(A.10a)
Z(t; s0 )
D(s0 )
j=1
Rs (t) =

Z(t; s j ) Z(t; s0 )

j

k=1

1
,
1 + k Rs (t)

j=1

n


Z(t; s0 )
L(t) =
Rs (t)

1

n

k=1

1
[1 + k Rs (t)]

D(s0 ) D(sn )e(sn s0 )x

L(Rs )
=
Z(t; s0 )
D(s0 )Rs

in terms of the swap rate. Thus this model yields

G(Rs ) =

n

k=1

(A.7)

1
(1 + j Rs )

A.3 Model 3: Parallel shifts

This model takes into account the initial yield curve shape, which can be
significant in steep yield curve environments. We still only allow parallel
yield curve shifts, so we approximate

j D(s j )e(s j s0 )x + D(sn )e(sn s0 )x = D(s0 ).

(A.12b)

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

Z(t; s j )
D(s j ) [h(s j )h(s0 )]x

e
,
Z(t; s0 )
D(s0 )

(A.9)

(A.13)

where x is the amount of the shift, and h(s) is the effect of the shift on
maturity s. As above, the shift x is determined implicitly in terms of the
swap rate Rs via
Rs

n


j D(s j )e[h(s j )h(s0 )]x + D(sn )e[h(sn )h(s0 )]x = D(s0 ).

(A.14a)

j=1

Then
D(s0 ) D(sn )e[h(sn )h(s0 )]x
L(Rs )
=
Z(t; s0 )
D(s0 )Rs

(A.14b)

Wilmott magazine

n


This models limitations are that it allows only parallel shifts of the yield
curve and it presumes perfect correlation between long and short term
rates.

(A.8)

This approximates the yield curve as flat and only allows parallel shifts,
but has the schedule right.

Z(t; s j )
D(s j ) (s j s0 )x

e
Z(t; s0 )
D(s0 )

(A.12a)

A.4 Model 4: Non-parallel shifts

We can allow non-parallel shifts by approximating
1

Z(t; tp )
Rs e(tp s0 )x

,
D(sn ) (sn s0 )x
L(t)
1 D(s
e
0)

j=1

1
,
(1 + 1 Rs (t))

Z(t; tp )
Rs

L(t)
(1 + 1 Rs )

Rs

(A.6)

where  = (tp s0 )/(s1 s0 ) as before. Then

G(Rs ) =

(A.11b)

where x is determined implicitly in terms of Rs by

In the same spirit, we can approximate

Z(t; tp ) = Z(t; s0 )

(A.11a)

determines the parallel shift x implicitly in terms of the swap rate Rs .

With x determined by Rs , the level is given by

k=1

We can establish the following identity by induction:

j D(s j )e(s j s0 )x + D(sn )e(sn s0 )x = D(s0 )

j=1

(A.4)

where k is the coverage of the kth period of the reference swaption. At

any date t, the level is then

j

n
n



1
.
L(t) =
j Z(t; s j ) = Z(t; s0 )
j
(A.5)
1 + k Rs (t)
j=1
j=1

(A.10b)

Turning this around,

Rs

A.2 Model 2: Exact yield model

We can account for the reference swaptions schedule and day count
exactly by approximating

D(s0 ) D(sn )e(sn s0 )x

.
n

(s
j s 0 )x
j D(s j )e

43

PATRICK S.HAGAN

determines the level in terms of the swap rate. This model then yields
G(Rs ) =

Z(t; tp )
Rs e[h(tp )h(s0 )]x

,
D(sn ) [h(sn )h(s0 )]x
L(t)
1 D(s
e
0)

(A.15a)

where x is determined implicitly in terms of Rs by

Rs

n


j D(s j )e[h(s j )h(s0 )]x + D(sn )e[h(sn )h(s0 )]x = D(s0 ).

(A.15b)

j=1

To continue further requires selecting the function h(s j ) which determines the shape of the non-parallel shift. This is often done by postulating a constant mean reversion,

1

1 e (ss0 ) .
h(s) h(s0 ) =
(A.16)

Alternatively, one can choose h(s j ) by calibrating the vanilla swaptions

which have the same start date s0 and varying end dates to their market
prices.

FOOTNOTE
1. We follow the standard (if bad) practice of referring to both the physical instrument
and its value as the numeraire.

W
44

Wilmott magazine