This action might not be possible to undo. Are you sure you want to continue?
Simultaneous Calibration to Interest Rate
and FX Smile
Dr. Christian Fries
email@christianfries.de
Fabian Eckstaedt
email@fabianeckstaedt.de
August 2006
Revised December 2008
Abstract
In this paper we present a MarkovFunctional hybrid interest rate / foreign
exchange model which allows calibration to given market volatility surfaces in
both dimensions simultaneously. This is achieved by extending the approach
introduced in [8] by a functional for the foreign exchange rate (FX) which
allows a fast, yet accurate calibration to a given market FX volatility surface.
This calibration procedure comes as an additional step to the known calibration
of the LIBOR functional, resulting in an efﬁcient implementation.
CONTENTS
Contents
1 Introduction 3
1.1 Outline of the Paper . . . . . . . . . . . . . . . . . . . . . . . 4
2 Preliminaries and some Notation 4
3 The MarkovFunctional Model 4
3.1 The Interest Rate MarkovFunctional Model under the Termi
nal Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1.1 Calibration to Caplets . . . . . . . . . . . . . . . . . 7
3.2 Calibration of the Model Dynamic . . . . . . . . . . . . . . . 8
3.2.1 Instantaneous Volatility of the Markovian Driver . . . 8
3.2.2 Drift of the Markovian Driver . . . . . . . . . . . . . 8
3.3 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 9
4 The Hybrid MarkovFunctional Model 10
4.1 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.1.1 Role of the Drift of the Markovian Driver . . . . . . . 11
4.1.2 Speciﬁcation of the driving processes . . . . . . . . . 12
4.2 The TwoFactor CrossCurrency Model (stochastic FX rates) . 12
4.2.1 The General Calibration Procedure . . . . . . . . . . 13
4.3 Calibration of a Flat FX Volatility Surface . . . . . . . . . . . 14
4.3.1 The Functional Form . . . . . . . . . . . . . . . . . . 14
4.3.2 The Drift Equation for ρ = 0 . . . . . . . . . . . . . . 15
4.3.3 The Numerical Solution of the Drift Equation . . . . . 15
4.4 Calibration of a general FX Volatility Surface . . . . . . . . . 16
4.4.1 The Functional Form . . . . . . . . . . . . . . . . . . 16
4.4.2 The Drift Equation . . . . . . . . . . . . . . . . . . . 16
4.4.3 How this functional generates a smile effect and how
to choose the parameters . . . . . . . . . . . . . . . . 16
4.5 Free Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 17
5 Numerical Results 18
5.1 The Interest Rate Calibration . . . . . . . . . . . . . . . . . . 18
5.2 The FX Calibration . . . . . . . . . . . . . . . . . . . . . . . 20
5.3 Numerical Performance of the Model . . . . . . . . . . . . . 22
6 Conclusion and Outlook 23
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
2 Version 1.2 (January 17, 2009)
1 INTRODUCTION
1 Introduction
MarkovFunctional models were ﬁrst considered by Hunt, Kennedy and Pelsser
[12] (2000), see also [10], [16] and, independently, by Balland and Hughston
[2] (2000). The LIBOR MarkovFunctional model models a market rate, the
forward LIBOR as a functional of a low dimensional Markovian process. Thus,
as for low dimensional short rate models, a Markovfunctional model may be
implemented in a low dimensional lattice, resulting in a fast and robust pric
ing algorithm. Since we are free to choose the underlying Markov process, we
may restrict ourselves to a class of processes for which large timestep con
ditional expectations are given by a convolution with a transition probability
that is known analytically. This enables us to use large timesteps by numer
ical integration with an analytic integration kernel, thus further improving the
efﬁciency of the overall algorithm.
Since the LIBOR MarkovFunctional model models a market rate, the mar
ket prices of corresponding options (given as implied Black volatilities) are the
natural calibration products, similar to a LIBOR market model [3]. Due to the
setup of MarkovFunctional models it is possible to infer the shape of the func
tional directly from such market option prices, hence reproducing the market
implied terminal distribution of the underlying forward rate. Thus, the calibra
tion to implied volatility is fast, robust and exact. From the LIBOR functional
we may then calculate the functional of the numéraire, which ﬁnally deﬁnes
the (single currency) model.
Alternatively, the numéraire functional of the interest rate Markov func
tional model may be calibrated to options of other rates, e.g., swap rates, as
long as there is a formula expressing the numéraire in terms of the modeled
rate (i.e., as long as the model is closed in some sense), see [7].
Furthermore, the underlying Markov process leaves enough freedom to
capture important market features like forward rate autocorrelations. These
maybe used to calibrate to swaption prices or bermudan option prices.
The hybrid MarkovFunctional model extends the singlecurrency interest
rate Markovfunctional model to a consistent model of the joint evolution of
interest rates and another stochastic underlying. As the second underlying
we consider the FX rate but this could also be an equity process. The hy
brid MarkovFunctional model was ﬁrst introduced independently by Fries &
Rott [8] and also by Antonov & Lee[1] in 2004. These approaches focus on the
general theory and provide specimen examples of how to model a ﬂat volatil
ity surface for the second underlying. In [8], the functionals considered for
the second underlying were given by a oneparameter family. In this paper
we replace this functional by a more general family of functionals which al
lows robust calibration to a given FX volatility surface exhibiting smile/skew
effects.
The model can be extended straight forwardly to a full ﬂedged cross cur
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
3 Version 1.2 (January 17, 2009)
3 THE MARKOVFUNCTIONAL MODEL
rency model featuring stochastic foreign interest rate (or a full ﬂedged equity
hybrid model featuring stochastic dividend curves).
1.1 Outline of the Paper
In Section 3 we introduce the single currency MarkovFunctional Model as
published by Hunt, Kennedy and Pelsser in [12]. We suggest further enhance
ment to the model to improve the accuracy of the Swaption smile as replicated
in the model.
In Section 4 we reconsider the approach by Fries and Rott[8]. They have al
ready pointed out that a ﬂat volatility surface can be captured by an exponential
functional form modeled over a normal distributed driving process. We intro
duce a functional form which will ﬁt a general volatility smile in the second
underlying (FX). The calibration procedure is an additional step which retains
the calibration to the volatility smile in the ﬁrst underlying (LIBOR).
We will conclude by presenting some numerical results obtained from its
implementation in a two dimensional Lattice.
2 Preliminaries and some Notation
The ideas presented in this paper rely on knowledge of the general L
1
theory
of option pricing. A full and rigorous treatment of the background theory can
be found in e.g. [6],[10]. The notation used here corresponds to the one in [6].
Let (Ω, {F
t
}, F, P) denote a ﬁltered probability space where F
t
is the aug
mented natural ﬁltration generated by a Brownian motion W.
A pure discount bond or zero coupon bond paying unit amount at time T
is an {F
t
}adapted stochastic process deﬁned on the ﬁltered probability space
(Ω, {F
t
}, F, P) for 0 ≤ t ≤ T denoted by P(T). T is called the maturity of
the Bond. P(T; t) denotes the time t value of the Bond and deﬁnes a {F
t
}
measurable random variable, while P(T; t, ω) denotes its value for some path
ω ∈ Ω.
Respectively the forward LIBOR for the time period [S, T] will be denoted
by L(S, T), its value at time t by L(S, T; t). L(S, T; t, ω) denotes its value for
some path ω ∈ Ω. L(S, T) is determined by
1 +L(S, T; ·)(T −S) :=
P(S; ·)
P(T; ·)
3 The MarkovFunctional Model
The MarkovFunctional framework was ﬁrst introduced by Hunt, Kennedy and
Pelsser [12] and, independently, by Balland and Hughston [2] in 2000. In this
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
4 Version 1.2 (January 17, 2009)
3 THE MARKOVFUNCTIONAL MODEL
chapter we summarize the ideas given in [12], [10], [16] and [8]. Moreover we
describe some enhancements and give suggestions for further optimizations.
The deﬁning characteristics are:
• The numéraire, and hence, pure discount bond prices are a function of
some low dimensional process which is Markovian in the corresponding
martingale measure
⇒Implementation is efﬁcient.
• The freedom to choose the functional form connecting the driving pro
cess to the bond prices
⇒Possible to ﬁt the marginal distributions of market interest rates, smiles
& skews.
• The freedom to choose the law of the driving process
⇒Allows realistic modelling, to capture the joint distribution of the con
sidered interest rates.
Deﬁnition 1 (MarkovFunctional):
An interest rate model is said to be Markovfunctional if there exists some
numéraire N, a corresponding equivalent martingale measure Q and some
realvalued stochastic process X such that:
• The process X is a Markov process under the measure Q.
• The numéraire N is a price process, a stochastic process of the form
1
N(t, X
t
(ω)) 0 ≤ t ≤
ˆ
T,
where (t, ·) −→N(t, ·) is a deterministic function.
ˆ
T is the model horizon.
Remark 2 (Notation): We use the same letter N both for the stochastic
process N : [0, T] × Ω → R and the functional N : [0, T] × R → R. This
ambiguity is actually not relevant, since in the following we postulate that the
stochastic process is represented by the functional. In all cases it is clear from
the arguments: N(t) denotes the stochastic process at time t, while N(t, ξ)
denotes the functional at time t and state ξ.
Remark 3 (Numéraire): The numéraire might be relaxed to a (speciﬁc) path
dependent process, depending on {X(s)}
s<t
. This is relevant for an implemen
tation in the spot rate measure
2
, see [8].
1
N(t, X
t
(ω)) := N(t, ·) ◦ X
t
(ω)
2
We actually think that implementation in the spot rate measure has considerable numerical
advantages over the terminal measure. Further research is in preparation.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
5 Version 1.2 (January 17, 2009)
3 THE MARKOVFUNCTIONAL MODEL
For complete speciﬁcation of a MarkovFunctional Model it is sufﬁcient to
know
(i) The law of the process X under Q.
(ii) The functional form of the numéraire ξ →N(t, ξ) for 0 ≤ t ≤
ˆ
T.
From this we can recover the discount factors prior to the time horizon via
the martingale property for numéraire rebased assets under Q by the funda
mental valuation formula
3
:
P(S; t) = P(S; t, X(t)),
where
4
P(S; t, ξ) = N(t, ξ) E
Q
_
1
N(S, X(S))
¸
¸
{X(t) = ξ}
_
.
3.1 The Interest Rate MarkovFunctional Model under the
Terminal Measure
We consider a tenor structure, i.e., a time discretization of the interval [0,
ˆ
T]
into n subintervals given by 0 =: T
0
< T
1
< . . . < T
n
:=
ˆ
T. Deﬁne
L
i
= L(T
i
, T
i+1
) 0 ≤ i ≤ n −1
Speciﬁcation of the model:
(i) dX(t) = σ(t)dW(t) under Q, X(0) = x
0
:= 0,
(ii) N(t, X(t)) = P(T
n
; t, X(t)),
where W is a QBrownian motion, adapted to the ﬁltration {F
t
}.
Free Parameters: The free parameters of the model are
(i) the speciﬁcation of the driving process X, i.e. the deterministic function
σ(t) (the instantaneous volatility of the Markovian driver), and,
(ii) the speciﬁcation of the numéraire functional ξ → N(T
i
, ξ) for all 0 ≤
i ≤ n.
For the time being we leave t →σ(t) unspeciﬁed. It will be analyzed later.
Under the assumption of a given σ, we derive the numéraire functional from
LIBOR’s functionals. The LIBOR functional are inferred, i.e. calibrated from
market option prices (caplets). This gives us the LIBOR MarkovFunctional
Model.
3
See [10]
4
Again, we use the same symbol P, both for the stochastic process and for the functional.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
6 Version 1.2 (January 17, 2009)
3 THE MARKOVFUNCTIONAL MODEL
3.1.1 Calibration to Caplets
By deﬁnition the LIBOR L
i
, seen on its ﬁxing date T
i
is given by
1 +L
i
(T
i
)(T
i+1
−T
i
) =
P(T
i
; T
i
)
P(T
i+1
; T
i
)
=
1
P(T
i+1
; T
i
)
Applying the valuation formula conditional to {X(T
i
) = ξ} we get:
1 +L
i
(T
i
, ξ)(T
i+1
−T
i
) =
1
N(T
i
, ξ)E
Q
_
1
N(T
i+1
,X(T
i+1
))
¸
¸
{X(T
i
) = ξ}
_,
i.e.,
N(T
i
, ξ) =
1
E
Q
_
1
N(T
i+1
,X(T
i+1
))
¸
¸
{X(T
i
) = ξ}
_
(1 +L
i
(T
i
, ξ)(T
i+1
−T
i
))
Thus, given the functional ξ →L
i
(T
i
, ξ) representing the LIBOR rate, this
gives a backward induction step T
i+1
→ T
i
to calculate N(T
i
) from N(T
i+1
).
The induction start is trivially given by N(T
n
, ξ) = 1 ∀ξ.
The induction step from T
i+1
→ T
i
: We calibrate the model to digital
caplets which is the same as calibrating the model to caplets, see [8].
Assume that the numéraire functionals N(T
k
) are known for k ≥ i +1. Let
V
market
K,Ti
(T
0
) denote the market price of the digital caplet with ﬁxing date T
i
,
paying
V
market
K,T
i
(T
i+1
) =
_
1, if L(T
i
) ≥ K
0, if L(T
i
) < K
Furthermore assume that these prices are monotone in K
5
and known for arbi
trary strikes K.
We model the functional ξ → L(T
i
, ξ) as a monotone function in ξ. For a
ﬁxed x
∗
∈ R the payoff of a digital caplet with strike L(T
i
, x
∗
) and payment
date T
i+1
is then given by
ξ →V
model
L
i
(T
i
,x
∗
),T
i
(T
i+1
, ξ) =
_
1, if ξ ≥ x
∗
0, if ξ < x
∗
.
Hence the model price is
V
model
L
i
(T
i
,x
∗
),T
i
(T
0
) = N(0)E
Q
_
1
[x
∗
,∞)
N(T
i+1
, X(T
i+1
))
 {X(T
0
) = x
0
}
_
.
Note that the right hand side can be calculated for any given x
∗
from the in
formation available from the previous induction step, namely ξ →N(T
i+1
, ξ),
5
The prices are monotone in K if the market prices do not allow arbitrage.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
7 Version 1.2 (January 17, 2009)
3 THE MARKOVFUNCTIONAL MODEL
and does not depend on L
i
(T
i
). We will thus write V
model
x
∗
,Ti
:= V
model
L
i
(T
i
,x
∗
),T
i
.
The requirement to have the model calibrated to market prices for each strike
K = L
i
(T
i
, x
∗
) implies:
V
model
x
∗
,T
i
(T
0
) = V
market
K,T
i
(T
0
).
From this we ﬁnd L
i
(T
i
, x
∗
) = K by inverting the marketprice formula.
As a result this method gives functionals ξ → L
i
(T
i
, ξ), monotone in ξ,
calibrated to caplets of all strikes.
3.2 Calibration of the Model Dynamic
3.2.1 Instantaneous Volatility of the Markovian Driver
The as yet unspeciﬁed parameter σ controls the autocorrelation between the
different time steps and therefore the mean reversion of the model. If we
choose σ
t
= exp(at) the parameter a will be referred to as the meanreversion
parameter of the model. The parameter a induces a mean reversion on T
i
→
L
i
(T
i
, X
T
i
). For a detailed discussion see [16], [12].
The instantaneous volatility t → σ of the Markovian driver acts as scaling
to the instantaneous volatility of L
i
(t), while the functional itself only deter
mines the terminal distribution, i.e., the distribution of L
i
(T
i
). Thus, σ deter
mines when volatility is accumulated. In other words, σ determines the forward
volatility of the LIBOR rates and thus the forward volatility of all other rates
being a function of LIBOR rates, e.g., swap rates.
Hence, σ can be used to calibrate products which depend on the forward
volatility structure like swaptions with different maturities or bermudan options
with different underlyings.
6
3.2.2 Drift of the Markovian Driver
The deﬁnition of the Markovian driver can be generalized to include a drift
term:
dX(t) = µ(t, X
t
)dt +σ(t)dW(t) under Q, X(0) = x
0
= 0.
It indirectly determines where the mean reversion leads to, the meanreversion
level, thus leaving us free to calibrate for more than one additional option (e.g.,
swaption or bermudan option) at each time T
i
making the model more realis
tic.
7
6
Of course, σ has no effect on the rate calibrated through the speciﬁcation of the numéraire
functional (in our example in Section 3.1.1 L
i
(T
i
).
7
Note: Since the drift term makes the driver path dependent, its handling is much more
straightforward in the spot measure than in the terminal measure.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
8 Version 1.2 (January 17, 2009)
3 THE MARKOVFUNCTIONAL MODEL
For our discussion the ﬁnal extension of the single currency model is to
allow a(t) in σ(t) = exp(a(t)t) to be time dependent, enabling us to cali
brate swaptions to each option maturity T
i
in stead of a global least square ﬁt
approach.
3.3 Further Reading
The interested reader will ﬁnd several papers about a multidimensional exten
sion of the single currency model, for example [11], [9], [17], where a higher
dimensional driving process for the numéraire process and therefore for the
bonds is considered. This leads to even greater ﬂexibility in the joint distribu
tion. For an introduction to the Markovfunctional model see [6]. See [7] for
an indepth discussion of the calibration of the model dynamics.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
9 Version 1.2 (January 17, 2009)
4 THE HYBRID MARKOVFUNCTIONAL MODEL
4 The Hybrid MarkovFunctional Model
After this review of the single currency model we are now ready to introduce
a twodimensional Markovfunctional model in the sense that we consistently
model two underlyings at the same time. Here we will consider the second
underlying to be the stochastic FX rate but we will see that this is arbitrary and
we could just as well model an equity process.
The Hybrid Markovfunctional model was ﬁrst introduced independently
by C. Fries and M. Rott [8] and by A. Antonov and H. Lee [1] in 2004. Antonov
and Lee work with an exponential functional form for the second process. Fries
and Rott also consider this approach as one example but point out that this
will only model a ﬂat volatility surface and suggest further research on this
topic. In this chapter we extend the approach by Fries and Rott by introducing
a functional form to ﬁt smile effects and considering the degrees of freedom
given.
4.1 The Model
In addition to the single currency Markovfunctional interest rate model we
introduce
8
(i) a process Y , which is a Markov process under the measure Q
(ii) a second underlying U as a price process, a stochastic process of the form
t →U(t, Y (t)), 0 ≤ t ≤
ˆ
T,
where (t, ·) −→U(t, ·) is a deterministic function.
9
NoArbitrage Constrain: To guarantee that the model is arbitragefree we
have to make sure that
U(t)
N(t)
is a martingale under Q, i.e.,
U(t)
N(t)
= E
Q
_
U(s)
N(s)
¸
¸
F
t
_
for all s > t (noarbitrage condition)
This can be satisﬁed by choosing the correct drift of the driving process Y .
Speciﬁcation: To completely specify the twodimensional MarkovFunctional
Model it is sufﬁcient to specify
(i) a single currency MarkovFunctional Model as introduced in the previ
ous chapter
8
This generalizes the ideas of the CrossCurrency model which can be found in [1] and [8].
9
To shorten notation we use the same symbol for the process t → U(t) and the functional
(t, η) →U(t, η) representing the process via U(t) = U(t, Y (t)).
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
10 Version 1.2 (January 17, 2009)
4 THE HYBRID MARKOVFUNCTIONAL MODEL
(ii) the law of the process Y under Q
(iii) the functional form of the second underlying η →U(t, η) for 0 ≤ t ≤
ˆ
T
where the choice of Y and U are restricted by the noarbitrage condition from
above.
From this we can price any {F
t
} measurable non pathdependent deriva
tive which depends on no more than the term structure of the interest rate and
the second underlying. For the valuation we use the fundamental valuation for
mula [10]. This is postulated by the martingale property for numéraire rebased
tradeable assets of the economy under Q .
Free Parameters: The desirable free parameters of the model are
(i) The free parameters of the onedimensional model.
(ii) The variance of the driving process Y .
(iii) The covariance between the processes X and Y .
(iv) The speciﬁcation of the functional U.
4.1.1 Role of the Drift of the Markovian Driver
In the single currency MarkovFunctional Model we could arbitrary choose a
driving process without imposing any restrictions on the functional form of the
LIBOR. The model would always remain arbitrage free and we ﬁrst naturally
opted for a zerodrift process.
In the twodimensional model the freedom to choose a Y drift is gone, as
for a given functional U we now need a corresponding drift for the model to
remain arbitrage free. If we would a priori insist on a speciﬁc drift, e.g. a zero
drift, specifying the functional form of U(
ˆ
T, ·) would automatically determine
all other functional forms U(t, ·), 0 ≤ t ≤
ˆ
T by the martingale property from
above. The only free parameter remaining for U would be the variance of the
process Y . The reason for the ’freedom’ to choose the drift for the process X
is due to the ability to choose the numéraire functional. Because of
P(T
i+1
; T
i
, X
T
i
) = N(T
i
, X
T
i
) E
Q
_
1
N(T
i+1
, X
T
i+1
)
¸
¸
F
t
_
T
i
→P(T
i+1
; T
i
, X
T
i
) is a Qmartingale for any functional formξ →N(T
i
, ξ).
Conversely, for any functional form ξ → P(T
i+1
; T
i
, ξ) we may choose the
numéraire ξ → N(T
i
, ξ) such that T
i
→ P(T
i+1
; T
i
, X
T
i
) is a Qmartingale.
This freedom is gone once the numéraire has been chosen.
One can either choose a zero drift and sacriﬁce the calibration of FX prod
ucts to solve the noarbitrage condition by the FX functional, or ,if one wants to
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
11 Version 1.2 (January 17, 2009)
4 THE HYBRID MARKOVFUNCTIONAL MODEL
calibrate a general FX volatility surface in addition to an existing IR Markov
Functional, one has to introduce a drift. We will consider the latter approach.
A three factor cross currency model with zerodrift Markovian drivers was
considered by Johnson and Dutra [13]. Their model features three driving pro
cesses modeling the three stochastic processes: domestic interest rates, foreign
interest rates and the FX process, all having zero drift. The focus is on calibra
tion of a general volatility surface in the domestic and foreign interest rate. The
calibration of the FX functional is sacriﬁced to satisfy the noarbitrage condi
tion. However, the model still allows calibration of ATM FX options. They
stress that this approach is especially applicable to the valuation of quantos.
Using Markovian driver without a drift has allows for a much faster numerical
implementation since: One does not have the additional step of solving a drift
condition and calibration decomposes into three independent onedimensional
problems.
4.1.2 Speciﬁcation of the driving processes
dX(t) = σ
x
(t) dW(t)
(1)
X
0
= x
0
dY (t) = µ(t, X(t), Y (t) dt +σ
y
(t) dW(t)
(2)
Y
0
= y
0
,
where W
(1)
and W
(2)
have instantaneous correlation ρ, i.e., dW(t)
(1)
dW(t)
(2)
=
ρ(t)dt.
4.2 The TwoFactor CrossCurrency Model (stochastic FX
rates)
In the twofactor cross currency model we model stochastic domestic inter
est rates and stochastic FX rates. The required foreign interest rates will be
assumed to be deterministic  for the time being.
In this case, we chose as the underlying U the foreign bond in domestic
currency (this is a traded asset for the domestic investor),i.e. U(T
i
, Y
T
i
) =
FX(T
i
, Y
T
i
)
˜
P(T
n
; T
i
) where
˜
P denotes the deterministic price process of the
foreign discount bond. That means we directly calibrate the functional form
FX(T
i
, ·).
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
12 Version 1.2 (January 17, 2009)
4 THE HYBRID MARKOVFUNCTIONAL MODEL
The drift condition can be rewritten in term of the FX functional as:
U(T
i
)
N(T
i
)
= E
Q
_
U(T
i+1
)
N(T
i+1
)
¸
¸
F
T
i
_
⇐⇒
FX(T
i
)
˜
P(T
n
; T
i
)
N(T
i
)
= E
Q
_
FX(T
i+1
)
˜
P(T
n
; T
i+1
)
N(T
i+1
)
¸
¸
F
T
i
_
⇐⇒
FX(T
i
)
˜
P(T
i+1
; T
i
)
N(T
i
)
= E
Q
_
FX(T
i+1
)
˜
P(T
i+1
; T
i+1
)
N(T
i+1
)
¸
¸
F
T
i
_
⇐⇒
FX(T
i
)
˜
P(T
i+1
; T
i
)
N(T
i
)
= E
Q
_
FX(T
i+1
)
N(T
i+1
)
 F
T
i
_
(drift condition)
4.2.1 The General Calibration Procedure
In the twodimensional model we have a dependency between the functional
form at time T
i
and the drift of the driving process up to time T
i
. This requires
a forward calibration and determination of the whole functional before calcu
lating any model prices. We can no longer calibrate the functional pointwise as
in the ﬁrst dimension.
Instead we choose a family of reasonable functionals with enough free pa
rameters to allow a good ﬁt to market prices. We proceed by forward induction
over the maturities. Obviously we set FX(0) to be the current FX spot rate
observed in the market. For the calibration of FX(T
i+1
, ·) a multidimensional
optimizer
10
is used, where in each iteration step we calculate FX(T
i+1
, ·) from
the parameters proposed by the optimizer, then the drift µ(T
i
, ·, ·) is adjusted
to fulﬁll the drift condition, then model prices of FXoptions with different
strikes can be calculated and compared to the corresponding market prices,
which deﬁned the objective function of the optimizer. As an initial guess for
the functional form we may use the functional FX(T
i
) of the previous calibra
tion step.
In other words, the calibration proceeds as follows:
For each i = 0, . . . , n −1:
• Optimize the FX(T
i+1
) functional with a multidimensional optimizer,
where the objective function (calibration error) is calculated as follows:
– set the FX functional’s parameters from the optimizer’s guess.
– solve and set corresponding drift.
– calculate model prices for selected options on FX(T
i+1
) (the cali
bration products)
– calculate the calibration error as the deviation from the correspond
ing market prices (e.g., root mean square).
10
One can for example use the LevenbergMarquardt algorithm.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
13 Version 1.2 (January 17, 2009)
4 THE HYBRID MARKOVFUNCTIONAL MODEL
• set FX(T
i+1
) using the optimal parameters
• solve and set corresponding drift.
• proceed with the next maturity
4.3 Calibration of a Flat FX Volatility Surface
4.3.1 The Functional Form
As a functional form
11
in the case of a ﬂat volatility surface we choose:
η →FX(T
i
, η) := a(T
i
) exp(b(T
i
) η)
and the driving process is given, as stated above, by
dY (t) = µ(T
i
, X
T
i
, Y
T
i
)dt +σ
y,i−1
_
∆T
i−1
dW
(2)
(t) for T
i
≤ t < T
i+1
,
Y (0) = y
0
Implied Volatility Smile: This setup is consistent with a lognormal dy
namic of the FX which implies a ﬂat FX volatility. Consider
dFX(t) = FX(t) · µ(t)dt +FX(t) · σ(t)dW(t), (4.1)
then we get by Ito’s formula
d(log(FX(t))) = (µ(t) −1/2σ
2
(t))dt +σ(t)dW(t),
i.e.,
log FX(t) ∼ N(log(FX(0)) +
_
t
0
µ(s) −1/2σ
2
(s)ds,
_
t
0
σ
2
(s)ds).
This can be rewritten in the proposed Markov functional form
FX(T
i
) = a(T
i
) exp(b(T
i
)
ˆ
Y (T
i
)
with
d
ˆ
Y (t) = (µ(t) −1/2σ
2
(t))dt +σ(t)dW(t)
and
a(T
i
) = FX(0), b(T
i
) =
_
T
i
0
σ(s)ds
_
T
i
0
σ
y
(s)ds
.
The drift of the dynamical model (4.1) is such that the drift condition is ful
ﬁlled, hence the dynamical model (4.1) produces the same distribution for
FX(T
i
) as the proposed Markov functional model.
11
[1] and [8] also use an exponential functional form but with fewer degrees of freedom.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
14 Version 1.2 (January 17, 2009)
4 THE HYBRID MARKOVFUNCTIONAL MODEL
4.3.2 The Drift Equation for ρ = 0
In the case of zero instantaneous correlation, the drift equation can be simpli
ﬁed to
FX(T
i
)
˜
P(T
i+1
; T
i
)
N(T
i
)
= E
Q
_
FX(T
i+1
)
N(T
i+1
)
¸
¸
F
T
i
_
= E
Q
_
FX(T
i+1
)
¸
¸
F
T
i
_
E
Q
_
1
N(T
i+1
)
¸
¸
F
T
i
_
⇐⇒ FX(T
i
, Y
T
i
)
˜
P(T
i+1
; T
i
)
P(T
i+1
; T
i
, X
T
i
)
= E
Q
_
FX(T
i+1
, Y
T
i+1
)
¸
¸
F
T
i
_
.
The last form is advantageous, since the expectation is only a one dimensional
integral (whereas the general case involves a two dimensional integral).
Analytic Solution of the Drift Equation: For ρ = 0 and the above func
tional form FX(t, ·) we can ﬁnd an analytic solution for µ. It is
E
Q
(FX(T
i+1
, Y (T
i+1
))  {X(T
i
) = ξ, Y (T
i
) = η)
= a
T
i+1
exp(b
T
i+1
· (η +µ
T
i
(ξ, η) · ∆T
i
) +
b
2
T
i+1
2
· σ
2
T
i
∆T
i
)
⇐⇒ FX((T
i
, η)
˜
P(T
i+1
; T
i
)
P(T
i+1
; T
i
, ξ)
·
1
a
T
i+1
= exp(b
T
i+1
(η +µ
Ti
(ξ, η) · ∆T
i
) +
b
2
T
i+1
2
· σ
2
T
i
∆T
i
)
⇐⇒ log(FX((T
i
, η)
˜
P(T
i+1
; T
i
)
P(T
i+1
; T
i
, ξ)
1
a
T
i+1
) ·
1
b
T
i+1
= η +µ
Ti
(ξ, η) · ∆T
i
+
b
T
i+1
2
· σ
2
T
i
∆T
i
⇐⇒ µ
T
i
(ξ, η) · ∆T
i
= log(FX((T
i
, η) ·
˜
P(T
i+1
; T
i
)
P(T
i+1
; T
i
, ξ)
·
1
a
T
i+1
) ·
1
b
T
i+1
−η −
b
T
i+1
2
· σ
2
T
i
∆T
i
4.3.3 The Numerical Solution of the Drift Equation
Within a numerical implementation of the model the analytical solution of the
drift equation will not provide an arbitragefree model due to the discretization
error. The analytic drift is a good guess in the center of the grid. Here the
approximation error of the numerical implementation is small. At the sides, the
extrapolation error implies that the analytic drift formula no longer fulﬁlls the
(numerical) drift equation. The ﬁrst evidence for a drift which is not consistent
with the drift equation with respect to the integration and extrapolation method
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
15 Version 1.2 (January 17, 2009)
4 THE HYBRID MARKOVFUNCTIONAL MODEL
used is an error in the FXforward. Furthermore the above calculation holds
only for ρ = 0.
To guarantee that the numerical model is arbitragefree we need to solve
the drift equation numerically for every conditional state (x, y) at every time
step T
i
→T
i+1
with a one dimensional root ﬁnder.
4.4 Calibration of a general FX Volatility Surface
4.4.1 The Functional Form
To ﬁt smile effects we suggest the same form as the one for a ﬂat calibration
with two local correction terms
12
:
η →FX(T
i
, η)FX(T
i
, η) = a(T
i
) · exp(b(T
i
) · η)
+d
1
(T
i
) · exp(−c
1
(T
i
) · (η −m
1
(T
i
))
2
)
+d
2
(T
i
) · exp(−c
2
(T
i
) · (η −m
2
(T
i
))
2
),
where
m = state of the correction
d = impact of the correction
c = radius of the correction (c > 0).
The most signiﬁcant impact on the functional is given between the inﬂection
points of the correction terms: m±
_
1
2c
, see Figure 4.1 for an example.
Keep in mind that these additional terms also have an impact on the drift.
Here the radius is much larger. Hence, a change of one of these correction
terms will change all FX option prices even if c is large.
4.4.2 The Drift Equation
With the correction terms we have not been able to ﬁnd an analytic solution for
the drift equation, and, as we have seen in the ﬂat case, the analytic solution
is not the best choice anyway. So we are using the numerical drift calculation
from above.
4.4.3 How this functional generates a smile effect and how to choose the
parameters
The local correction terms have the effect that if d is positive the probability
density function of the r.v. FX shifts mass from the left of m to the right and
vice versa if d is negative.
12
Pelsser discusses a functional with a similar correction term in the context of the one di
mensional model in [16].
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
16 Version 1.2 (January 17, 2009)
4 THE HYBRID MARKOVFUNCTIONAL MODEL
1
0
1
2
3
4
5
6
7
8
2 1,5 1 0,5 0 0,5 1 1,5 2
Figure 4.1: Plot of exp(x) (black), the two correction terms (green, blue) and
the resulting FX functional (red).
With the two correction terms placed on each side of ATM we can create
the fat tails which generate the smile effect of the volatility surface.
To ﬁt the volatility smile we use an ATM, one in and one out of the money
FXoption. It seems to be best to chose m
1
and m
2
in the middle of these
three in terms of the driving process Y . a
T
i
is chosen such that the ATM strike
corresponds to Y
T
i
= 0. We chose the corresponding c to be a function of
the distance between the point m
1/2
and ATM, c
1/2
=
const
m
2
1/2
. To stabilize the
calibration procedure the constant left to determine c has to be large enough to
ensure that the correction terms does not effect the functional form at the sides
of the grid, but small enough to have a smooth effect around m
1/2
.
If c exceeds its lower limit the modeling range changes according to the
change in the correction term, which has the consequence that the solver will
become unstable.
4.5 Free Parameters
Remark 4 (σ, ρ): We are still left with the free parameter σ which can be used
to calibrate autocorrelation dependent products and the parameter ρ to choose
the correlation between the FX and the domestic interest rate processes. If
there is no need to adjust σ one should use σ(t) = exp(at) with a such that the
parameter b does not move too far away from its value at T
1
over time. This
guarantees ﬁrstly that the ratio of σ of the FX rate and the driving process stays
the same over time and secondly that the model has greater numerical quality.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
17 Version 1.2 (January 17, 2009)
5 NUMERICAL RESULTS
5 Numerical Results
5.1 The Interest Rate Calibration
For the interest rate calibration we observe the typical calibration accuracy of
the standard MarkovFunctional approach: The neglectable calibration devia
tion s due to the difference of the interpolation/extrapolation of between given
market strikes for the market date and given grid points for the model.
In Figure 5.2 the graph on the left shows some caplet market data in terms of
a implied volatility surface plotted against the option maturity and it’s absolute
moneyness. On the left we got the corresponding caplet prices. The data is
generated from market cap prices by a bootstrap algorithm. A preprocessing
of the data makes sure that it is arbitrage free. There are different methods,
but this is beyond the focus of this paper. The graphs in Figure 5.3 show the
difference between the preprocessed market data and the model output. The
left shows the difference in terms of the implied volatility and the right in terms
of caplet prices. The difference in the implied volatility is less than 0.1% and
just a fraction of a base point in caplet prices. Figure 5.4 shows the calibration
accuracy of a global calibration of the mean reversion (parameter a) to co
terminal swaptions. Here the difference between the market implied volatility
and the model implied volatility plotted on the yaxes is rather large reaching
1% for an option maturity, xaxes, of 26 years. Depending on the use case
one would like to improve this by a time dependent calibration of the mean
reversion.
3%
1%
2%
0,5
3,5
6,5
9,5
12,5
15,5
18,5
21,5
24,5
27,5
#NV
0,0%
10,0%
20,0%
30,0%
40,0%
50,0%
60,0%
70,0%
80,0%
90,0%
100,0%
Volatility
Moneyne
ss
Maturity
Preprossed Market Caplet Volatilities
(a) Implied Volatility
3%
1%
2%
0,5
3,5
6,5
9,5
12,5
15,5
18,5
21,5
24,5
27,5
#NV
0,0%
0,2%
0,4%
0,6%
0,8%
1,0%
1,2%
1,4%
1,6%
Price
Moneyne
ss
Maturity
Preprocessed Market Caplet Prices
(b) Prices
Figure 5.2: Caplet input market data
A further discussion of the variety of calibration methods for the one di
mensional MarkovFunctional Model can be found in [7].
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
18 Version 1.2 (January 17, 2009)
5 NUMERICAL RESULTS
3%
1%
2%
0,5
3,5
6,5
9,5
12,5
15,5
18,5
21,5
24,5
27,5
#NV
0,50%
0,40%
0,30%
0,20%
0,10%
0,00%
0,10%
0,20%
0,30%
0,40%
0,50%
Volatility Diff.
Moneynes
s
Maturity
Deviation (MarketLattice) Caplet Volatilities
(a) Implied Volatility
3%
1%
2%
0,5
3,5
6,5
9,5
12,5
15,5
18,5
21,5
24,5
27,5
#NV
1,5
1
0,5
0
0,5
1
1,5
2
Price Diff. In BP
Moneynes
s
Maturity
Deviation (MarketLattice) Caplet Prices
(b) Prices
Figure 5.3: Caplet calibration error (deviation of model to input market data).
Deviation Swaption volatilities at ATM
1,20%
1,00%
0,80%
0,60%
0,40%
0,20%
0,00%
0,20%
0,40%
0,5 2,5 4,5 6,5 8,5 10,5 12,5 14,5 16,5 18,5 20,5 22,5 24,5 26,5 28,5
maturity
Figure 5.4: Swaption calibration error (in terms of implied volatility) calibra
tion error after a calibration of the mean reversion parameter a.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
19 Version 1.2 (January 17, 2009)
5 NUMERICAL RESULTS
5.2 The FX Calibration
Deﬁnition 5 (Standardized Moneyness):
The Standardized Moneyness
13
ˆ
M for an underlying S is deﬁned as
ˆ
M :=
ln(
K
FS(T
i
)
)
ˆ σ
i
√
T
i
where ˆ σ
i
is the implied Black ATM volatility for a call option with maturity T
i
on S, and FS(T
i
) denotes the forward as seen today of the underlying. K is
the value of the corresponding strike at time T
i
.
This moneyness term is close to a denomination of how many standard
deviations the strike K is apart from ATM. (If we have smile effects this is not
exactly the case but still a sufﬁcient measure.)
The following charts are plotted against the standardized moneyness.
In Figure 5.5 we show some FX call option’s market data plotted against
the option maturity and the standardized moneyness. The left is the implied
volatility and the right call option prices. The extrapolation of the market data
on moneyness is set to constant volatility. In Figure 5.6 we observe that the
suggested parametrization of the functionals overall leads to a very good ﬁt
of the market data with an maximum difference in implied FX volatility of
0.2% (left) and fractions of a base point in FX call option prices (right) in the
range of −160
ˆ
M to +160
ˆ
M. This is more than we need in most applications.
Figure 5.7 gives another detailed view of the market and model implied FX
option volatility plotted against the standardized moneyness for options with
different maturities.
0,5
3
5,5
8
10,5
13
15,5
18

1
6
0
%

8
0
%
0
%
8
0
%
1
6
0
%
6,00%
7,00%
8,00%
9,00%
10,00%
11,00%
12,00%
13,00%
14,00%
Volatility
Maturity std.
Moneyness
Market FX Volatilities
(a) Implied Volatility
0
,5
2
,5
4
,5
6
,5
8
,5
1
0
,5
1
2
,5
1
4
,5
1
6
,5
1
8
,5

3
4
0
%

2
2
0
%

1
0
0
%
2
0
%
1
4
0
%
2
6
0
%
0,00%
5,00%
10,00%
15,00%
20,00%
25,00%
30,00%
35,00%
40,00%
Price
Maturity
std.
Moneyness
Market FX Prices
(b) Prices
Figure 5.5: FX options input market data.
13
A further discussion of the concept can be found in [15].
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
20 Version 1.2 (January 17, 2009)
5 NUMERICAL RESULTS
0,5
3
5,5
8
10,5
13
15,5
18
1
6
0
%
8
0
%
0
%
8
0
%
1
6
0
%
1,00%
0,80%
0,60%
0,40%
0,20%
0,00%
0,20%
0,40%
0,60%
0,80%
1,00%
Volatility Diff.
Maturity
std.
Moneyness
Deviation (Market  Lattice) FX Volatilities
(a) Implied Volatility
0,5
3
5,5
8
10,5
13
15,5
18 
1
6
0
,0
%

6
0
,0
%
4
0
,0
%
1
4
0
,0
%
2
1,5
1
0,5
0
0,5
1
1,5
2
Price Diff. in BP
Maturity
std.
Moneyness
Deviation (MarketLattice) FX Prices
(b) Prices
Figure 5.6: FX option calibration error (deviation of model to input market
data).
FX Volatility at Maturity 0,5
9,6000%
9,8000%
10,0000%
10,2000%
10,4000%
10,6000%
10,8000%
11,0000%
11,2000%
400% 300% 200% 100% 0% 100% 200% 300% 400%
Input Model
(a) Implied Volatility for a 6M FX Option
FX Volatility at Maturity 0,5
9,6000%
9,8000%
10,0000%
10,2000%
10,4000%
10,6000%
10,8000%
11,0000%
11,2000%
400% 300% 200% 100% 0% 100% 200% 300% 400%
Input Model
(b) Implied Volatility for a 5Y FX Option
Maturity 10
10,40%
10,60%
10,80%
11,00%
11,20%
11,40%
11,60%
11,80%
12,00%
400% 300% 200% 100% 0% 100% 200% 300% 400%
Input Model
(c) Implied Volatility for a 10Y FX Option
Maturity 20
10,50%
11,00%
11,50%
12,00%
12,50%
13,00%
13,50%
400% 300% 200% 100% 0% 100% 200% 300% 400%
Input Model
(d) Implied Volatility for a 20YFXOption
Figure 5.7: FX option calibration error (deviation of model to input market
data).
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
21 Version 1.2 (January 17, 2009)
5 NUMERICAL RESULTS
However there is a slight twist in the volatility curve over time, which leads
to an unstable calibration (for the sample data after about 22 years). This ef
fect is due to the implementation with zero correlation between the domestic
interest rate and the FX rate.
A discussion of the implementation and suggestions for further optimiza
tion can be found in [7].
5.3 Numerical Performance of the Model
The numerical performance of the model depends on the efﬁciency of the nu
merical integration used for the conditional expectation operator.
The two factor model used to generate the results above used a simple,
nonoptimized implementation of the convolution of a piecewise polynomial
functional against the gaussian kernel. Calibration was performed in a few
minutes, pricing in few seconds on a standard 2006 desktop computer.
Since the performance of the model depends so much on the numerical
methods used, detailed timings are hardly informative. Instead we make some
more qualitative remarks on the numerical performance of the two and three
factor Markov functional models.
For the single currency (one dimensional) model, the number of grid points
per maturity has to be at least as large as the number of calibration instruments
per maturity. Working with n grid points per maturity (we found n = 31
sufﬁcient) will require at most n calculations of the implied volatility and a
corresponding conditional expectation. Within the one dimensional model cal
ibration of the functionals and product pricing is almost instant. For the two
dimensional model the main factors are:
• The valuation of the calibration products involve a two dimensional in
tegration. If we precalculate the expectation operator as suggested in [8]
the time to calibrate the FX process is still linear in the number of time
steps.
• In each iteration to solve for the FX functional at T
i
, we recalculate the
drift of the second driving process and therefor have to rebuild the ex
pectation operator from T
i
to T
i−1
.
• The expectation operator is the most expensive in terms of computational
time. As the drift depends on both dimensions the rebuilt takes the num
ber of xstates longer than in one dimension and has to be redone in every
solver iteration.
The additional overhead of a two or three factor model can be parallelized.
Thus, a full ﬂedged threefactor Markov functional model can be implemented
without any increase of the real time calculation using the computational power
of modern multiprocessor (multicore) computers.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
22 Version 1.2 (January 17, 2009)
6 CONCLUSION AND OUTLOOK
6 Conclusion and Outlook
In this paper we have reviewed and extended the theory of a hybrid interest
rate / fx Markovfunctional model as it had been introduce by [8] in 2004. We
extended the setup such that it provides a consistent interest rate / fx hybrid
model with an almost perfect smile ﬁt in both underlyings at the same time.
As we pointed out earlier, there is still a lot of research to be done. Just one
example is that while we model the term structure under the terminal measure,
an implementation under the rolling spot measure as described in [8] might in
fact lead to even better outcomes. It makes better numerical features possible
as discussed in [7]. In other papers dealing with the MarkovFunctional Model
there are several extensions and improvements to the single currency model
which can now be combined with or even extended to the hybrid model.
Meanwhile more such extensions were published, among them we would
like to mention the recent work of [14], where an ndimensional Markov func
tional model is presented in spot measure. To cope with the then high di
mensional integrals a MonteCarlo implementation is proposed. This approach
could also be considered to implement a hybrid Markov functional model,
where high dimensional integrals can become a challenge, e.g., when consid
ering stochastic foreign interest rate.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
23 Version 1.2 (January 17, 2009)
REFERENCES
References
[1] ANTONOV, ALEXANDRE; LEE, HAN: Interest Rate Mod
elling Framework in Discrete Rolling Spot Measure. 2004.
http://www.ssrn.com.
[2] BALLAND, PHILIPPE; HUGHSTON, LANE P.: Markov Market Model
Consistent with Cap Smile. International Journal of Theoretical and
Applied Finance, Vol. 3, No. 2 (2000) 161181.
[3] BRIGO, DAMIANO; MERCURIO, FABIO: Interest Rate Models  The
ory and Practice. SpringerVerlag, Berlin, 2001. ISBN 3540417729.
[4] CHEYETTE,OREN: Markov Representation of the HeathJarrow
Morton Model. http://www.ssrn.com.
[5] ECKSTAEDT, FABIAN:: The valuation of Hybrid Options with a two
dimensional MarkovFunctional Model. Diploma Thesis. Bielefeld
University, 2006.
[6] FRIES, CHRISTIAN P.: Mathematical Finance: Theory, Modeling, Im
plementation. John Wiley & Sons, 2007. ISBN 0470047224.
[7] FRIES, CHRISTIAN; ECKSTAEDT, FABIAN: Markov Functional Mod
eling. In preparation, 2009.
[8] FRIES, CHRISTIAN P.; ROTT, MARIUS G.: CrossCurency
and Hybrid MarkovFunctional Models. 2004. http://www.christian
fries.de/ﬁnmath.
[9] GENHEIMER, FRANK: A Two Factor MarkovFunctional Model for
Pricing Interest Rate Derivatives. Diploma Thesis in Mathematics.
University of Mainz, 2003.
[10] HUNT, PHIL J.; KENNEDY, JOANNE E.: Financial Derivatives in The
ory and Practice,Revised Edition. John Wiley & Sons, 2004. ISBN 0
470863595.
[11] HUNT, PHIL J.; KENNEDY, JOANNE E.: LongstaffSchwarz, Effec
tive Model Dimensionality and Reducible MarkovFunctional Mod
els.2005. http://www.ssrn.com.
[12] HUNT, PHIL J.; KENNEDY, JOANNE E.; PELSSER, ANTOON:
MarkovFunctional Interest Rate Models. Finance and Stochastics,
Volume 4(4), 391408, 2000. Reprinted in Hughston, Lane (ed.): The
new Interest Rate Models, RISK Publications.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
24 Version 1.2 (January 17, 2009)
REFERENCES
[13] JOHNSON, SIMON; DUTRA, SERGIO: Crosscurrency Markov Func
tional model. Working paper, Commerzbank, 2006.
[14] KAISAJUNTTI, LINUS; KENNEDY, JOANNE: An nDimensional
Markovfunctional Interest Rate Model. 2008, http://www.ssrn.com.
[15] MEISTER, MARKUS: Smile Modeling in the LIBOR Market Model.
Diploma Thesis. University of Karlsruhe, 2004.
[16] PELSSER, ANTOON: Efﬁcient Methods for Valuing Interest Rate
Derivatives. Springer Finance, 2000. ISBN 1852333049.
[17] SEIFERT, THOMAS: Multidimensional Markovfunctional models in
option pricing. Dr. Thesis. Universitaet der Bundeswehr Muenchen
2004.
c 2006 Christian Fries & Fabian Eckstaedt
http://www.christianfries.de/ﬁnmath/markovfunctionalhybridwithsmile
http://www.fabianeckstaedt.de
25 Version 1.2 (January 17, 2009)
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.