Calibration, Simulation and Hedging in A HESTON LIBOR Market Model With Stochastic Basis

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 26

Electronic copy available at: http://ssrn.

com/abstract=1704415
Cal i br at i on, Si mul at i on and Hed gi n g i n a
HESTON LI BOR Mar k et Model w i t h
St ochast i c Basi s
AHSAN AMIN
1



This Version: NOV 04, 2010

ABST RACT
We follow Mercurio's extension of the LIBOR market model with stochastic Basis
spreads and model the joint evolution of forward rates belonging to the discount curve
and corresponding spreads with FRA rates. We consider Heston stochastic-volatility
dynamics and show how to calculate the swaption pricing problems in general. We
present several different modeling approaches with different stochastic volatility
structures and possibilities of correlation. We also describe how to calibrate the models to
the swaption matrix with smile. We end the article by giving an overview of how to
calculate basis risk deltas and OIS rate deltas of derivative products in this model.

1 . I nt r oduct i on

During the credit crunch, the basis between different market rates of the same tenor
increased in an unprecedented way and their volatilities increased remarkably. Several
explanations of this scenario have been put forward explaining it in terms of credit and
liquidity constraints. One solution adopted by the banks is employing two different
forwarding and discounting curves. Forward rates are projected using one curve which

1
The author can be reached at anan2999@yahoo.com
Electronic copy available at: http://ssrn.com/abstract=1704415
includes default risk premium and effects of liquidity constraints while since most deals
are collateralized, the discounting is done using risk free rates. This paper addresses the
need for a comprehensive modeling of default free rates and basis curve and describes
how such a model can be calibrated and addresses some pricing and hedging issues.
2 . Over vi ew of Mar k et Model w i t h st ochast i c basi s

In a multi-curve setting we have to deal with existence of several yield curves. There is
one discount curve and as many forward curves as market tenors. We accomplish this,
following Mercurio, by modeling OIS curve as the discount curve and rest of the forward
curves are modeled by adding stochastic spread term structure for each forward curve
with different tenor.
The discount curve is the OIS zero-coupon curve, stripped from market OIS swap
rates and defined for every possible maturity. The rationale behind this assumption is that
in the interbank derivatives market, most deals are collateralized to mitigate the credit
risk of both parties and since OIS swaps have very negligible counterparty risk, they are
the best suitable proxy for the risk free rate. The OIS curve can be stripped from OIS
swap rates using single curve bootstrapping methods. For the EUR market, EONIA
swaps are quoted up to 30 years. In cases where OIS swap rates are not available for
long maturities, one has to model the spread between OIS forward rates and
corresponding forward LIBOR rates. It may be possible to construct the curve by adding
quotes of cross-currency swaps.
We follow Mercurio in how to define OIS discount rates and FRA rates.
The discount curve is the OIS zero coupon curve, stripped from market OIS swap rates
and defined for every possible maturity T
P

( 0 , I) = P
0IS
( 0, I)
where
P

( t, I) stands for the discount factor at time t for maturity T which is assumed to
coincide with the corresponding OIS-based zero-coupon for maturity T. The subscript D
stands for discount curve.
Consider times t, I
1
and I
2
, t I
1
< I
2
. The time-t FRA rate FRA( t; I
1
, I
2
) is defined
as the fixed rate to be exchanged at time I
2
for the LIBOR rate I( I
1
, I
2
) so that the swap
has zero value at time t.
Denoting by

1
the T-forward measure with numeraire the zero-coupon bond P

( t, I)
by no-arbitrage pricing, we immediately have

FRA( t; I
1
, I
2
) = E

1
2
[ I( I
1
, I
2
) |
t
]

The FRA rate is a martingale under the corresponding pricing measure. These properties
will prove to be very convenient when pricing swaps and options on LIBOR rates.
As Mercurio mentioned this definition of FRA rate slightly differs from that implied by
the actual market contract and we have to calculate a convexity correction to reconcile
the two definitions. We refer the reader to look at his paper [15] to know about the
difference in more detail.
we define
I
k
( t) = FRA( t; I
1
, I
2
) = E

1
2
[ I( I
1
, I
2
) |
t
]

the IRS value of a fixed rate payer is given as
IRS( t, K; I
u
, , I
b
) = o
k
P

( t, I
k
) I
k
( t) Ko
k
P

( t, I
k
)
b
u+1
b
k=u+1


The corresponding swap rate is given as
S
u,b
( t) =
o
k
P

( t, I
k
) I
k
( t)
b
k=u+1
o
k
P

( t, I
k
)
b
k=u+1


As Mercurio mentions in his paper that above equation can be used with spot starting
swaps to strip the values of FRA rates I
k
( 0) since values of OIS discount bonds are all
known from OIS curve.
We follow Mercurio's notation in the stochastic basis LMM set up. Let us fix a given
tenor x and consider a time structure J
x
= { 0 < I
0
x
, , I
M
x
} compatible with x
{ 1m, 3m, 6m, 1y} . Let us then define the OIS forward rate

F
k
x
( t) = F

( t; I
k-1
x
, I
k
x
) =
1
o
k
x
_
P

( t, I
k-1
x
)
P

( t, I
k
x
)
1_
where o
k
x
is the corresponding year fraction for the interval (I
k-1
x
, I
k
x
) and denote by
S
k
x
( t) the spread at time t between FRA rate I
k
x
( t) = FRA( t; I
k-1
x
, I
k
x
) and the OIS
forward rate F
k
x
( t) , that is
S
k
x
( t) = I
k
x
( t) F
k
x
( t)
By definition both F
k
x
( t) and I
k
x
( t) are martingales under the forward measure

1
x
k
and
hence their difference S
k
x
( t) is a

1
x
k
-martingale.
We model the joint evolution of rates F
k
x
( t) and the spreads S
k
x
( t) .
We adopt a displaced diffusion Heston model as originally proposed by Andersen and
Andreasen [2]. We present different modeling approaches for volatility and correlation
term structure of the OIS rates and the Basis spreads.
In a complete model construction, we have to specify different sets of correlations, for
example, among different OIS discount rates, between OIS discount rates and stochastic
basis of different curves and between stochastic basis of different curves. We also have to
specify a common or separate stochastic volatility structure for each set. We present three
modeling options for joint evolution of OIS discount rates and basis spreads.

a. Both OIS rates and Basis spreads are modeled as displaced diffusion processes
with a common Heston Stochastic Volatility. Discount rates and basis spreads can
be correlated with each other but are uncorrelated with the common driving
stochastic volatility.

b. OIS rates and Basis spreads are based on displaced diffusion processes with their
own distinct stochastic volatility processes i.e. there is one stochastic volatility
driving diffusion for OIS rates and there is another stochastic volatility diffusion
for basis spreads but OIS rates and Basis spreads are uncorrelated with each other.
Both OIS rates and basis spreads are also uncorrelated with their stochastic
volatility factors.

c. In the third scenario both OIS rates and Basis spreads are based on displaced
diffusion processes with their own distinct stochastic volatility process and OIS
rates and Basis spreads are correlated with each other. Both OIS rates and basis
spreads are uncorrelated with their respective stochastic volatility factors.
3 . St ochast i c Basi s M ar k et M odel w i t h common st ochast i c
vol at i l i t y for OI S r at es and basi s spr eads

The dynamics of F
k
x
and S
k
x
on the T-forward measure

1
k
x
associated with zero coupon
bond P

( t, I
k
) are given as

JF
k
x
( t) = I( t) ([F
k
x
( t) + ( 1 [) F
k
x
( 0) )z
k
P
( t) ( I( t) p
k
( t) Jt + JZ
k
P
)

JS
k
x
( t) = I( t) ([S
k
x
( t) + ( 1 [) S
k
x
( 0) )z
k
S
( t) ( I( t) p
k
S
( t) Jt + JZ
k
S
)

Usually for simulation of the whole curve it is preferred to use the spot LIBOR measure

J
associated with times J = { I
0
x
, , I
M
x
} whose numeraire is the discretely balanced
bank account B

J

B

J
=
P

( t, I
[( t) -1
x
)
P

( I
]-1
x
, I
]
x
)
[( t) -1
]=0

where [( t) = m if I
m-2
x
< t I
m-1
x
, m 1

For a lognormal model with [ = 1, the equations on the spot measure associated with
become
JF
k
x
( t) = z
k
P
( t) I( t) F
k
x
( t)
p
k,]
P,P
z
]
P
( t) F
]
x
( t)
1 + oF
]
x
( t)
Jt + z
k
P
( t) I( t) F
k
x
( t) JZ
k
P,J
k
]=[( t)

JS
k
x
( t) = z
k
S
( t) I( t) S
k
x
( t)
p
k,]
P,S
z
]
P
( t) F
]
x
( t)
1 + oF
]
x
( t)
Jt + z
k
S
( t) I( t) S
k
x
( t) JZ
k
S,J
k
]=[( t)


In the case of general [, it would be expedient to reformulate the equation in terms of
general displaced diffusion form as

F

k
x
( t) = F
k
x
( t) +
( 1 [)
[
F
k
x
( 0) onJ z
k
P
( t) = [z
k
P
( t)

S

k
x
( t) = S
k
x
( t) +
( 1 [)
[
S
k
x
( 0) onJ z
k
S
( t) = [z
k
S
( t)
The evolution equations now become
JF

k
x
( t) = z
k
P
( t) I( t) F

k
x
( t)
p
k,]
P,P
z
]
P
( t) F

]
x
( t)
1 + oF
]
x
( t)
Jt + z
k
P
( t) I( t) F

k
x
( t) JZ
k
P,J
( t)
k
]=[( t)

JS

k
x
( t) = z
k
S
( t) I( t) S

k
x
( t)
p
k,]
P,S
z
]
P
( t) F

]
x
( t)
1 + oF
]
x
( t)
Jt + z
k
S
( t) I( t) S

k
x
( t) JZ
k
S,J
( t)
k
]=[( t)

Stochastic volatility process is governed by Heston dynamics
JI( t) = (0 I( t) )Jt + eI( t) JZ
v
( t)
where
I( 0) = 0 = 1
The correlations are defined as
_
JZ
k
P,J
( t) JZ
]
P,J
( t) = p
k,]
P,P
( t) Jt
JZ
k
S,J
( t) JZ
]
S,J
( t) = p
k,]
S,S
( t) Jt
JZ
k
P,J
( t) JZ
]
S,J
( t) = p
k,]
P,S
( t) Jt

and
_
JZ
k
P,J
( t) JZ
v
( t) = 0
JZ
k
S,J
( t) JZ
v
( t) = 0


4 . Anal yt i c for mul as for cal i br at i on t o sw apt i ons

Let us consider a payer swaption which gives the user right to enter at time I
u
x
an IRS
ending at I
b
x
. For the sake of simplicity we assume that both floating and fixed legs have
the same frequency. The swaption payoff at I
u
x
is given as

|S
u,b
( I
u
x
) K]
+
o
]
x
P

( I
u
x
, I
]
x
)
b
]=u+1

where swap rate is given as
S
u,b
( t) =
o
k
P

( t, I
k
) I
k
( t)
b
k=u+1
o
k
P

( t, I
k
)
b
k=u+1

The swap numeraire is defined as
C

u,b
= o
k
P

( t, I
k
x
)
b
k=u+1

The swpation price is calculated by the expectation given below where E

u,b
is the
expectation with respect to swap numeraire
PS( t, K; I
u
x
, , I
b
x
) = o
k
P

( t, I
k
x
)
b
k=u+1
E

u,b
]|S
u,b
( I
u
x
) K]
+
|
t

The swap rate can be decomposed as
S
u,b
( t) =
k
( t) I
k
x
( t) =
b
k=u+1

k
( t) F
k
x
( t) +
k
( t) S
k
x
( t)
b
k=u+1
b
k=u+1

where

k
( t) =
o
k
x
P

( t, I
k
x
)
o
]
S
P

( t, I
k
S
)
b
]=u+1

Taking derivatives of swap rate with respect to OIS forward rate and basis spread, the
total differential can be broken down as

JS
u,b
( t) = _( F
k
x
( t) + S
k
x
( t) ) _
o
k
( t)
oF
]
x
( t)
b
]=u+1
JF
]
x
( t) _ +
k
( t) JF
k
x
( t)
b
k=u+1
+
k
( t) JS
k
x
( t) _
which can be written in shorthand as

JS
u,b
( t) =
oS
u,b
( t)
oF
k
x
( t)
b
k=u+1
JF
k
x
( t) +
oS
u,b
( t)
oS
k
x
( t)
b
k=u+1
JS
k
x
( t)
We can approximate the dynamics of the swap rate on
JS
u,b
( t) = [S
u,b
( t) _
oS
u,b
oF
k
x
( t)
b
k=u+1
( F
k
x
( t) )
[S
u,b
( t)
I( t) z
k
P
( t) JZ
k
u,b
( t)
+
oS
u,b
( t)
oS
k
x
( t)
b
k=u+1
( S
k
x
( t) )
[S
u,b
( t)
I( t) z
k
S
( t) JZ
k
u,b
( t) _

JS
u,b
( t) = [S
u,b
( t) I( t) z
u,b
S
( t) JZ
k
u,b
( t)

Here Z
k
u,b
is the brownian motion associated with the swap numeraire C

u,b
. The effective
swaption volatility used in the diffusion above is given as

z
u,b
S
( t) = _z
u,b
P
( t)
2
+ z
u,b
S
( t)
2
+ z
u,b
P,S
( t)
2

where
z
u,b
P
( t) = _ w
]
P
( t) w
k
P
( t) p
k,]
P,P
( t)
b
]=u+1
b
k=u+1

z
u,b
S
( t) = _ w
]
S
( t) w
k
S
( t) p
k,]
S,S
( t)
b
]=u+1
b
k=u+1


z
u,b
S,P
( t) = _2 w
]
P
( t) w
k
S
( t) p
k,]
P,S
( t)
b
]=u+1
b
k=u+1


w
k
P
( t) =
oS
u,b
( 0)
oF
k
x
( 0)
( F
k
x
( 0) )
[S
u,b
( 0)
z
k
P
( t)
w
k
S
( t) =
oS
u,b
( 0)
oS
k
x
( 0)
( S
k
x
( 0) )
[S
u,b
( 0)
z
k
S
( t)
Here we mention that ( F
k
x
( t) ) = [F
k
x
( t) + ( 1 [) F
k
x
( 0) and so on for basis
spreads and swap rates.
Once we have calculated the effective volatility and we have the knowledge of
parameters of stochastic volatility process, the price of swaption can easily be calculated
by transform techniques relevant to Heston model. Andersen and Andreasen were first to
propose these formulas in the context of fixed income derivatives and their paper is a
good reference for the formulas.
5 . Cal i br at i on t echni ques

A modeler can use his judgment about scenarios of credit and liquidity constraints to
project volatilities of basis spreads in the future and this can be added to historical
evolution of basis spread volatilities.
For each swaption smile in the swaption matrix, we calculate the effective constant
volatility referred as z
u,b
S
i n equat i on under the relevant forward swap measure while
simultaneously optimizing globally for skew parameter, mean reversion of stochastic
volatility and volatility of stochastic volatility.
JS
u,b
( t) = ( [S
u,b
( t) + ( 1 [) S
u,b
( 0) ) I( t) z
u,b
S
JZ
k
u,b
( t)

In other words we find global best fit values of , e onJ [ such that all swaptions smiles
are as closely calibrated as possible while calculating one effective value of volatility
z
u,b
S
for each ATM swaption. This way we fix stochastic volatility parameters of the
process and all we are left to do is optimize for time dependent volatilities of LIBOR
rates. We mention that even though we are optimizing stochastic volatility parameters
with constant volatility coefficient for each swaption, this results in parameters that are
very close to the case if we had performed a full scale calibration with time dependent
volatility coefficients z
u,b
S
( t) .
Once we have calculated constant volatility z
u,b
S
for each ATM swaption in the swaption
matrix, we use parameter averaging proposed by Piterbarg [19] while using relationships
between OIS volatilities, spread volatilities and swaption volatilities. Briefly in
optimization we slowly perturb the OIS volatilities and spread volatilities and calculate
corresponding time dependent swaption volatilities z
u,b
S
( t) and equate them to constant
swaption volatilities z
u,b
S
using parameter averaging. The optimization problem is solved
when all z
u,b
S
( t) are equated to z
u,b
S
by parameter averaging formula for each a and b in
the swpation matrix.
We have to emphasize that strong smoothing constraints have to be applied on OIS
LIBOR volatilities and Basis spread volatilities and these constraints are a part of the
optimization procedure.
The advantage of the parameter averaging technique is that each time we perturb the
volatility coefficients of OIS rates and Basis spreads, we just have to solve a simple
riccati equation. In full scale Heston optimization, we have to solve a complete set of
riccati equation coupled with integrations and so parameter averaging is several orders of
magnitude faster.
We mention that in the above calibration procedure, we used a single universal value of
skew parameter [ but it is not any more difficult to use parameter averaging in skew
parameter dimension and reader can easily implement a time dependent skew parameter
and we refer to Piterbarg [19] for more details about how to implement this.

6 . St ochast i c Basi s M ar k et M odel w i t h di st i nct st ochast i c
vol at i l i t y pr ocesses for di scount r at es and basi s spr eads

In this setup OIS rates and Basis spreads are based on displaced diffusion processes with
their own distinct stochastic volatility processes i.e. there is one stochastic volatility
driving diffusion for OIS rates and there is another stochastic volatility diffusion for basis
spreads but OIS rates and Basis spreads are uncorrelated with each other. Both OIS rates
and basis spreads are also uncorrelated with their stochastic volatility factors.

JF

k
x
( t) = z
k
P
( t)
1
I( t) F
k
x
( t)
p
k,]
P,P
z
]
P
( t) F

]
x
( t)
1 + oF
]
x
( t)
Jt + z
k
P
( t)
1
I( t) F
k
x
( t) JZ
k
P,J
k
]=[( t)

JS

k
x
( t) = z
k
S
( t)
1
I
S
( t) S
k
x
( t) JZ
k
S,J

We notice that since Basis spreads are uncorrelated with OIS rates and with discount
money market numeraire they do not have any drift in the simulation stochastic
differential equations. This makes simulation algorithm cheaper to compute.
JI( t) = (0 I( t) )Jt + eI( t) JZ
v
( t)
JI
S
( t) =
S
(0
S
I
S
( t) )Jt + e
S
I
S
( t) JZ
v
S
( t)
I( 0) = 0 = 1
I
S
( 0) = 0
S
= 1
The correlations are defined as
_
JZ
k
P,J
( t) JZ
]
P,J
( t) = p
k,]
P,P
( t) Jt
JZ
k
S,J
( t) JZ
]
S,J
( t) = p
k,]
S,S
( t) Jt
JZ
k
P,J
( t) JZ
]
S,J
( t) = 0

and
_
JZ
k
P,J
( t) JZ
v
( t) = 0
JZ
k
S,J
( t) JZ
v
S
( t) = 0



7 . Cal i br at i on for mul as for sw apt i ons

The approximate stochastic differential equation for the swap rate with respect to
measure associated with swap numeraire is given as
JS
u,b
( t) = [S
u,b
( t) _
oS
u,b
( t)
oF
k
x
( t)
b
k=u+1
( F
k
x
( t) )
[S
u,b
( t)
I( t) z
k
P
( t) JZ
k,P
u,b
( t)
+
oS
u,b
( t)
oS
k
x
( t)
b
k=u+1
( S
k
x
( t) )
[S
u,b
( t)
I
S
( t) z
k
S
( t) JZ
k,S
u,b
( t) _
In shorthand notation this is written as
JS
u,b
( t) = [S
u,b
( t) I( t) z
u,b
P
( t) JZ
k,P
u,b
( t) + [S
u,b
( t) I
S
( t) z
u,b
S
( t) JZ
k,S
u,b
( t)
where
z
u,b
P
( t) = _ w
]
P
( t) w
k
P
( t) p
k,]
P,P
( t)
b
]=u+1
b
k=u+1

z
u,b
S
( t) = _ w
]
S
( t) w
k
S
( t) p
k,]
S,S
( t)
b
]=u+1
b
k=u+1

We identify the above stochastic differential equation as a Double Heston process and
these formulas can easily be solved by double Heston techniques. The characteristic
function can be calculated as

E( t, s) = c
A( t,s) +B
F
( t,s) v( t) +B
S
( t,s) v
S
( t)

JA
Jt
= 0B
S
0
S
B
S

JB
P
Jt
= .5[
2
z
u,b
P
( t)
2
( s s
2
) + B
P
.5e
2
( B
P
)
2

JB
S
Jt
= .5[
2
z
u,b
S
( t)
2
( s s
2
) + B
S
.5e
S
2
( B
S
)
2

We used the Andersen and Andreasen notation above and after we have solved for the
characteristic function, we can integrate over the characteristic function to get the price of
the swaption. The details are in Andersen and Andreasen's paper [2].

8 . Cal i br at i on t echni ques

In this calibration, we have two sets of stochastic volatilities coupled with time dependent
volatilities for each of the basis spreads and OIS rates. A full scale optimization only to
swaptions may be delicate, instable and too computationally intensive. We suggest that
stochastic volatility parameters and time dependent volatility coefficients of the basis
spreads to be computed from historical data. A modeler can use his judgment about
scenarios of credit and liquidity constraints to project volatilities of basis spreads in the
future and this can be used to refine the parameters calculated from historical evolution of
basis spread volatilities.
We also emphasize that this is just a way to solve the problem and other equally better
approaches may be possible. We follow the following steps to solve the problem.

1. Given the historical data on movement of basis spreads, find out the stochastic
volatility parameters and time dependent volatility coefficients for the diffusion
equation of basis spreads. We propose a filtering technique coupled with
maximum likelihood optimization in the appendix to work out these parameters
from the historical data.
2. Fine tune the above calculated data based on the trader's judgment about scenarios
of credit and liquidity constraints in the future.
3. We decompose each swaption volatility into two components one of which is
Basis swaption volatility and the other is OIS swaption volatility. The time
dependent basis swaption volatility can be calculated from the above calculated
parameters of the Basis swap volatilities.
4. Given the swaption prices in the swaption matrix with different strikes, optimize
for the stochastic volatility parameters of OIS rates but keeping their time time
dependent volatilities constant. This would give us a set of time homogenous
volatilities along with the volatility of volatility as well as mean reversion
parameter used in the evolution of stochastic volatility for the OIS rates. This step
gives us global fit to stochastic volatility parameters and a local constant volatility
for each swaption. We want the reader to note that simultaneously optimizing for
stochastic volatility parameters and time dependent volatilities is computationally
intensive without much improvement to calibration. It is very easy to optimize for
the stochastic volatility parameters while keeping time dependent OIS swaption
volatilities constant and it gives results very close to full scale optimization so we
recommend this procedure.
5. Once we have found the constant OIS swaption volatility, we optimize for the
volatilities of OIS LIBOR rates using an augmented version of Piterbarg's
parameter averaging technique tailored to Double Heston model. We equate the
Laplace transform of constant swaption volatility with the Laplace transform of
time dependent swaption volatility obtained by optimizing for the volatility
parameters of the OIS LIBOR rates. This parameter averaging saves us from doing
costly integrations each time we perturb the volatility coefficients of OIS LIBOR
rates and makes optimization process extremely fast.
Here we mention how to use Piterbarg's parameter averaging technique for volatility in
the context of Double Heston model. We assume that the reader is familiar with
parameter averaging in simple Heston model. We refer the reader to Piterbarg [19] for the
reference and just give cursory details.
In parameter averaging, given the parameters of both Heston diffusion equations, we first
solve for constant asset price volatility z that solves for the ATM call price given the
parameters of stochastic volatility. This is done by standard Heston algorithms. We
determine the Laplace transform given by E
0
( t, w
i
) using this homogenous volatility.
Then we calculate the Laplace transform given by E( t, w ) from knowledge of time
dependent volatility coefficients of the asset price. The appropriate time dependent asset
price is the one that solves for
E
0
( t, w
i
) = E( t, w )
We have to use a root finding procedure in case we are dealing with a single option but in
general we are dealing with the whole swaption matrix and a very large number of asset
price volatilities so we have to use some optimization technique. The advantage of the
technique is that each time we perturb the volatility coefficients, we just have to solve a
simple riccati equation. In full scale Heston optimization, we have to solve a complete set
of riccati equation coupled with integrations and so parameter averaging is several orders
of magnitude faster.
We set
= I( 0) _ z
u,b
P
( t)
2
Jt + I
S
( 0) _ z
u,b
s
( t)
2
Jt
1
0
1
0

w =
1
2
+

2
8

w
i
= wz
2

The Laplace transform in case of time dependent volatilities is given as
E( t, w ) = c
A( t,w) +B
F
( t,w) v( t) +B
S
( t,w) v
S
( t)

where
JA
Jt
= 0B +
S
0
S
B
S


JB
P
Jt
= z
u,b
P
( t)
2
w + B
P
+ .5e
2
( B
P
)
2


JB
S
Jt
= z
u,b
S
( t)
2
w + B
S
+ .5e
S
2
( B
S
)
2


the homogenous case is solved with time dependent volatilities set to one and by using
parameter w
i
instead of w. In this case the Laplace transform is given as
E
0
( t, w
i
) = c
A
0
(t,w
|
)+B
0
F
(t,w
|
)v( t) +B
0
S
(t,w
|
)v
S
( t)


JA
0
Jt
= 0B
0
+
S
0
S
B
0
S

JB
0
P
Jt
= w
i
+ B
0
P
+ .5e
2
( B
0
P
)
2

JB
0
S
Jt
= w
i
+ B
0
S
+ .5e
S
2
( B
0
S
)
2

9 . St ochast i c Basi s Mar k et Model w i t h separ at e
st ochast i c vol at i l i t y and non-zer o cor r el at i ons bet w een
OI S r at es and basi s spr eads

The diffusions of Basis spreads and OIS forwards have non-zero correlations and are
driven by separate stochastic volatility processes. There is zero correlation between
diffusions of Basis spreads and OIS forwards with their respective stochastic volatility
processes.
JF
k
x

( t) = z
k
P
( t) I( t) F
k
x

( t)
p
k,]
P,P
z
]
P
( t) F
]
x

( t)
1 + oF
]
x
( t)
Jt + z
k
P
( t) I( t) F
k
x

( t) JZ
k
P,J
k
]=[( t)

JS
k
x

( t) = z
k
S
( t) I
S
( t) S
k
x

( t)
p
k,]
P,S
I( t) z

]
P
( t) F
]
x

( t)
1 + oF
]
x
( t)
Jt + z
k
S
( t) I
S
( t) S
k
x

( t) Z
k
S,J
k
]=[( t)


JI( t) = ( I( t) )Jt + eI( t) JZ
V
( t)

JI
S
( t) =
S
(
S
I
S
( t) )Jt + e
S
I
S
( t) JZ
v
S
( t)
The correlations are defined as
_
JZ
k
P,J
( t) JZ
]
P,J
( t) = p
k,]
P,P
( t) Jt
JZ
k
S,J
( t) JZ
]
S,J
( t) = p
k,]
S,S
( t) Jt
JZ
k
P,J
( t) JZ
]
S,J
( t) = p
k,]
P,S
( t) Jt

and
_
JZ
k
P,J
( t) JZ
v
( t) = 0
JZ
k
S,J
( t) JZ
v
S
( t) = 0


1 0 . Cal i br at i on for mul as for sw apt i ons

We again use the transform technique for this method as described by Grzelak and
Oosterlee[8]. The Characteristic function is described as

E( t, s) = c
A( t,s) +B
F
( t,s) v( t) +B
S
( t,s) v
S
( t)


JA
Jt
= 0B
S
0
S
B
S
+ .5[
2
z
u,b
P,S
( t) ( s s
2
) 0( t)

JB
P
Jt
= .5
2
z
u,b
P
( t)
2
( s s
2
) + B
P
.5e
2
( B
P
)
2


JB
S
Jt
= .5[
2
z
u,b
S
( t)
2
( s s
2
) + B
S
.5e
S
2
( B
S
)
2

where 0( t) is defined as
I( t) I
S
( t) = E [I( t) I
S
( t) = 0( t)
and calculated by the formula
E [I( t) = 2c( t) c
-w( t) / 2

1
k!
(
w( t)
2
)
2
(
1 + J
2
+ k)
(
J
2
+ k)

k=0

where indicates gamma function and rest of the variables are defined below
c( t) =
1
4
e
2
( 1 e
-xt
)
J =
4k0
s
2

( t) =
4I( 0) c
-kt
e
2
( 1 c
-kt
)

The price can be easily obtained by integrating over the characteristic function as shown
in [2].
Though it seems straightforward to see that a parameter averaging technique can be
found for this particular case, we do not describe it since we did not numerically verify it.

1 1 . Cal cul at i ng t he Basi s r i sk Gr eek s of exot i c
der i vat i ves.

In this part we show how to calculate deltas of different fixed income securities with
respect to OIS forward rates and Basis spreads. We use pathwise method of Glasserman
and Zhao for this purpose.
Consider a security which has a probabilistic payoff C( I) at expiry T and considering
money market process with time t value indicated as B
m
( t) , we have
C( t) = B
m
( t) E _
C( I)
B
m
( I)
|
t
_
The derivatives of the time zero value of the security with respect to basis spreads are
given as

C( 0)
S
n
x
( 0)
= B
m
( 0)

S
n
x
( 0)
E _
C( I; F
1
x
, . . , F
N
x
, S
1
x
, . . , S
N
x
)
B
m
( I)
|
t
_
C( 0)
S
n
x
( 0)
= B
m
( 0) E _
1
B
m
( I)
C( I)
S
n
x
( I)
S
n
x
( I)
S
n
x
( 0)
|
t
_

The above expression is valid since the money market process has no dependence on
Basis spreads and depends only on OIS forward rates. We also mention that unlike the
case of OIS forwards, the deltas of Basis spreads are independent of each other since
simulation equation of Basis spreads depend upon OIS forwards in drift but do not
depend upon other spreads.
The derivatives of present value of the security with respect to OIS forwards are given as

oC( 0)
oF
k
x
( 0)
= B
m
( 0) E _
1
B
m
( I)

oC( I)
oS
n
x
( I)
M
n=1
oS
n
x
( I)
oF
k
x
( 0)
+
1
B
m
( I)

oC( I)
oF
n
x
( I)
oF
n
x
( I)
oF
k
x
( 0)
M
n=1
+ C( I)
o( B
m
-1
( I) )
oF
n
x
( I)
1
n=1
oF
n
x
( I)
oF
k
x
( 0)
|
t
_
In the above equation we have to include derivative of terminal payoff with respect to
Basis spreads since evolution of basis spreads depends upon OIS forwards but usually
this correction is mild but we show it here for sake of completeness.
We mention that derivatives of payoff with respect to Basis spreads and OIS forwards
like
C( 1)
S
n
x
( 1)
and
C( 1)
P
n
x
( 1)
are calculated by differentiating the payoff at expiry while
S
n
x
( 1)
P
n
x
( 0)

and
P
n
x
( 1)
P
n
x
( 0)
are simulated simultaneously with OIS forward rates and Basis spreads.
We use the following shorthand notation for simulation of pathwise deltas.

nk
( t ) =
0P
n
x
( t)
0P
k
x
( 0)
,

nk
( t ) =
S
n
x
( t)
F
k
x
( 0)

( t ) =
S
n
x
( t)
S
n
x
( 0)

The evolution of pathwise deltas of OIS forwards is calculated as
J
nk
( t) =
nk
( t) jI( t) p
n
( t) Jt + I( t) z
n
P
( t)
1
JZ[ + F

n
x
( t) I( t)
op
n
( t)
oF
]
x
( t)
N
]=1

]k
( t)
with

nk
( 0) = 1{ n = k }
Pathwise deltas of Basis spreads with respect to OIS forwards are given as
J

nk
( t) = S

n
x
( t) I( t)
op
n
( t)
oF
]
x
( t)
N
]=1

]k
( t)

nk
( 0) = 0
The above deltas show up in equations since evolution of basis spreads depends upon the
OIS forward rates whenever there is a correlation between them.
The deltas of Basis spreads are very simple to calculate as they are not interdependent
upon each other
J
n

( t) =
n

( t) jI( t) p
n
( t) Jt + I( t) z
n
P
( t)
1
JZ[

( 0) = 1
1 2 . CONCLUSI ON
In this paper we have followed Mercurio's extension of the LIBOR market model with
stochastic Basis and discussed it in the case of Heston Stochastic Volatility Model with
displaced diffusion. We have modeled the joint evolution of forward rates belonging to
the discount curve and corresponding spreads with FRA rates. We have shown how to
calculate the swaption pricing problems in general. We have presented several different
modeling approaches to the problem with different stochastic volatility structures and
possibilities of correlation. We have described how to calibrate the models to the
swaption matrix with smile. We also give an overview of how to calculate basis spread
deltas and OIS rate deltas of derivative products in this model.
REFERENCES

[1] Ametrano, F. and M. Bianchetti (2009). Bootstrapping the Illiquidity, in Modelling
Interest Rates: Advances for Derivatives Pricing, edited by F. Mercurio, Risk Books.
[2] Andersen, L. and J. Andreasen (2002). Volatile volatilities. Risk, December, 163-168.
[3] Bianchetti, M. (2009) Two Curves, One Price: Pricing & Hedging Interest Rate
Derivatives Decoupling Forwarding and Discounting Yield Curves. Working Paper.
online at: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1334356
[4] Boenkost, W. and W. Schmidt (2005), Cross currency swap valuation. Working
Paper, HfBBusiness School of Finance & Management. Available online at:
www.frankfurt-school.de/dms/publications-cqf/FS_CPQF_Brosch_E
[5] Brace, A., D. Gatarek, and M. Musiela (1997). The market model of interest rate
dynamics, Mathematical Finance, 7, 127154.
[6] Fujii, M., Shimada, Y. and A. Takahashi (2009a). A note on construction of multiple
swap curves with and without collateral. CARF Working Paper Series F-154, available
online at http://ssrn.com/abstract=1440633.
[7] Fujii, M., Shimada, Y. and A. Takahashi (2009b). A Market Model of Interest Rates
with Dynamic Basis Spreads in the presence of Collateral and Multiple Currencies.
Working Paper, University of Tokyo and Shinsei Bank. Available online at:
http://www.e.u-tokyo.ac.jp/cirje/research/dp/2009/2009cf698.pdf
[8] Grzelak, L., and C.W. Oosterlee (2010). An Equity-Interest Rate Hybrid Model with
Stochastic Volatility and the Interest Rate Smile.
[9] Henrard, M. (2007). The Irony in the Derivatives Discounting. Wilmott Magazine,
July 2007, 92-98.
[10] Henrard, M. (2009). The Irony in the Derivatives Discounting Part II: The Crisis.
Preprint, Dexia Bank, Brussels.
[11] Heston, S.L. (1993) A Closed-Form Solution for Options with Stochastic Volatility
with Applications to Bond and Currency Options. The Review of Financial Studies
6, 327-343.
[12] Jamshidian, F. (1997) LIBOR and Swap Market Models and Measures. Finance and
Stochastics 1, 293-330.
[13] Mercurio, F. (2009) Interest Rates and The Credit Crunch: New Formulas and
Market Models. Available online at:
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1332205
[14] Mercurio, F., and M. Morini (2007). A Note on Correlation in Stochastic Volatility
Term Structure Models, Working Paper available at SSRN.com.
[15] Mercurio, F. (2010) LIBOR Market Models with Stochastic Basis.
[16] Miltersen, K.R., K. Sandmann and D. Sondermann (1997). Closed Form Solutions
forTerm Structure Derivatives with Log-Normal Interest Rates. The Journal of Finance
52, 409-430.
[17] Morini, M (2008). The puzzle in the interest rate curve: counterparty risk, Preprint.
Banca IMI, Milan.
[18] Morini, M (2009). Solving the Puzzle in the Interest Rate Market. Available online
at: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1506046
[19] Piterbarg, V. (2005). Stochastic volatility model with time-dependent skew, Applied
Mathematical Finance, 12(2), 147-185.
[20] Wu, L. and F. Zhang (2006). LIBOR market model with stochastic volatility,
Journal of Industrial and Management Optimization, Vol. 2, No. 2, 199227.


APPENENDI X A
OPT I MI ZI NG FOR T HE VOLAT I LI T Y COEFFI CI ENT S AND
PARAMET ERS OF HEST ON ST OCHAST I C VOLAT I LI T Y
FROM HI ST ORI CAL DAT A USI NG NON-LI NEAR FI LT ERI NG
T HEORY

We describe the procedure below for univariate time series but it can be applied to term
structure of Basis spreads by using Principal Component Analysis and following the
procedure for each eigenvector as univariate time series. We assumed a constant
volatility coefficient z in the equations below which is justified if we assume time
homogenous volatilities for basis spreads.
We postulate that the dynamics of log of asset prices, y
t
, follow a stochastic differential
equation governed by a stochastic volatility variable :( t) . The stochastic differential
equation is given below
Jy
t
= .5z
2
:( t) Jt + z:( t) Jz( t) (1)
The dynamics of stochastic volatility variable are given by Heston dynamics and these
are governed by the equation below
J:( t) = ( 0 :( t) ) Jt + e:( t) Jz
2
( t) (2)

Though the technique we describe in what follows is for general Heston model but in the
interest rate model we described, we used the following parameters.
:( 0) = 0 = 1
We also assumed a zero correlation between asset price and its stochastic volatility.
From the Stochastic Differential Equation of Heston stochastic volatility, we find that
density g( :, t) of stochastic volatility evolves in time as given by the Forward
Kolmogorov partial differential equation below

og( :, t)
ot
=
o( ( : 0) g( :, t) )
o:
+
o
2
( .5e
2
:g( :, t) )
o:
2

(3)
At any time all we observe is the change in the asset price and we do not observe the
change in the stochastic volatility. It is a hidden markov processes. Our goal is to find the
parameters of evolution of the stochastic volatility given the observation of change in the
stock prices. We accomplish this goal by employing non-linear filtering theory and
optimization techniques.
We setup the notation and explanation for the most important conditional distribution in
the model.
y
t
~ r( . | :
t
, o)
This is the conditional distribution of y on stochastic volatility given a parameter set o. A
change of variables in the equation for asset price shows that
r( y
t
| :
t
, 0
t
, o) = _
Jy
t
+ .5:
t
Jt
:
t
Jt
_ =
1
2n:
t
Jt
cxp _.5(
Jy
t
+ .5:
t
Jt
:
t
Jt
)
2
_
The form of the probability density of the stochastic volatility at time t conditional on the
information at time t is given by
g( :
t
| y
t
, o) =
r( y
t
| :
t
, y
t-1
, o) g( :
t
| y
t-1
, o)
( y
t
| , y
t-1
, o)

Here ( y
t
| , y
t-1
, o) is likelihood of observing current value of stock price given the old
value of stock price and is given as
( y
t
| , y
t-1
, o) = _r( y
t
| :
t
, y
t-1
, o) g( :
t
| y
t-1
, o) J:
t

(4)
Inserting this in the previous equation gives the result
g( :
t
| y
t
, o) =
r( y
t
| :
t
, y
t-1
, o) g( :
t
| y
t-1
, o)
r( y
t
| :
t
, y
t-1
, o) g( :
t
| y
t-1
, o) J:
t


In what follows, we use a simpler and lighter notation to make the update step simpler to
understand in computational terms
g( :
t
| y
t
, o) =
_
Jy
t
+ .5:
t
Jt
:
t
Jt
_g( :
t
| y
t-1
, o) J:
t
_
Jy
t
+ .5:
t
Jt
:
t
Jt
_g( :
t
| y
t-1
, o) J:
t

0

(5)
In the above equation densities g( :
t
| y
t-1
, o) is the density obtained from prediction step
prior to filtering step. Again, we mention that this density was obtained from previous
(time level t-1) update(filtering) step and were fed to the Forward Kolmogorov equation
to project them to next(current) time level. g( :
t
| y
t
, o) is the density of the stochastic
volatility obtained in the update step.

Expected value of stochastic volatility can now easily be found as
E[ :( t) ] = _ :g( :
t
) J:
t

0

After observing current asset price y
t
i
and filtering the stochastic volatility variable, we
can find the distribution of future asset prices that can be given as
p( y
t+1
i
)
= _
1
z2n:
t+1
Jty
t+1
i
cxp_.5(
lny
t+1
i
lny
t
i
+ .5:
t+1
Jt
z:
t+1
Jt
)
2
_ g( :
t+1
) J:
t+1
Jy
t+1
i

0


We have converted from normal distribution to lognormal distribution and also changed
the notation as y
t
i
= cxp( y
t
) so y
t
i
is the actual asset price and not the log of asset price.
Please notice that densities of trend variable and the stochastic volatility variable have
been advanced to the next time step as in the prediction step by feeding them to their
respective Forward Kolmogorov equations.
The Optimization for the parameters of stochastic volatility process follows the following
steps
1. we set some initial starting guess for :( 0) , , 0, e onJ z.
2. For each observation in the historical data, we feed filtered density obtained from
the previous step and feed it to Forward Kolmogorov equation. This is prediction
step. Since the difference between two observations is at maximum a few days,
this one dimensional equation is extremely cheap to solve. This projected density
is fed to the filtering equation 5, and we get a filtered estimate of the distribution
of stochastic volatility given the observation of stock price. In the meanwhile we
calculate the likelihood of observing current value of stock price from equation
(4).
3. We continue the above steps for each observation in the historical data. We
calculate the sum of log of likelihood of stock price observation at each data for
the whole series. This is the maximum likelihood estimator to be optimized.
4. we perturb the parameters :( 0) , , 0, e onJ z which are the parameters to be
optimized as dictated by the optimization algorithm and repeat steps 1-3. The
optimization algorithm looks for the parameters such that value of log-likelihood
is maximized and these are the best fit stochastic volatility parameters.

You might also like