# On Cross-Currency Models with

**Stochastic Volatility and Correlated
**

Interest Rates

Lech A. Grzelak,

∗

Cornelis W. Oosterlee,

†

June 1, 2010

Abstract

We construct multi-currency models with stochastic volatility and correlated

stochastic interest rates with a full matrix of correlations. We ﬁrst deal with a

foreign exchange (FX) model of Heston-type, in which the domestic and foreign

interest rates are generated by the short-rate process of Hull-White [HW96]. We

then extend the framework by modeling the interest rate by a stochastic volatility

displaced-diﬀusion Libor Market Model [AA02], which can model an interest

rate smile. We provide semi-closed form approximations which lead to eﬃcient

calibration of the multi-currency models. Finally, we add a correlated stock to the

framework and discuss the construction, model calibration and pricing of equity-

FX-interest rate hybrid payoﬀs.

Key words: Foreign-exchange (FX); stochastic volatility; Heston model; stochastic interest rates;

interest rate smile; forward characteristic function; hybrids; aﬃne diﬀusion; eﬃcient calibration.

1 Introduction

Since the ﬁnancial crisis, investors tend to look for products with a long time horizon,

that are less sensitive to short-term market ﬂuctuations. When pricing these exotic

contracts it is desirable to incorporate in a mathematical model the patterns present in

the market that are relevant to the product.

Due to the existence of complex FX products, like the Power-Reverse Dual-

Currency [SO02], the Equity-CMS Chameleon or the Equity-Linked Range Accrual

TRAN swaps [Cap07], that all have a long lifetime and are sensitive to smiles or skews

in the market, improved models with stochastic interest rates need to be developed.

The literature on modeling foreign exchange (FX) rates is rich and many stochastic

models are available. An industrial standard is a model from [SO02], where log-normally

distributed FX dynamics are assumed and Gaussian, one-factor, interest rates are used.

This model gives analytic expressions for the prices of basic products for at-the-money

options. Extensions on the interest rate side were presented in [Sch02b; Mik01], where

the short-rate model was replaced by a Libor Market Model framework.

A Gaussian interest rate model was also used in [Pit06], in which a local volatility

model was applied for generating the skews present in the FX market. In another

paper, [KJ07], a displaced-diﬀusion model for FX was combined with the interest rate

Libor Market Model.

Stochastic volatility FX models have also been investigated. For example, in [HP09]

the Sch¨obel-Zhu model was applied for pricing FX in combination with short-rate

processes. This model leads to a semi-closed form for the characteristic function.

However, for a normally distributed volatility process it is diﬃcult to outperform the

Heston model with independent stochastic interest rates [HP09].

∗

Delft University of Technology, Delft Institute of Applied Mathematics, Delft, The Netherlands,

and Rabobank International, Utrecht. E-mail address: L.A.Grzelak@tudelft.nl.

†

CWI – Centrum Wiskunde & Informatica, Amsterdam, the Netherlands, and Delft University of

Technology, Delft Institute of Applied Mathematics. E-mail address: C.W.Oosterlee@cwi.nl.

1

Research on the Heston dynamics in combination with correlated interest rates has

led to some interesting models. In [And07] and [Gie04] an indirectly imposed correlation

structure between Gaussian short-rates and FX was presented. The model is intuitively

appealing, but it may give rise to very large model parameters [AAA08]. An alternative

model was presented in [AM06; AAA08], in which calibration formulas were developed

by means of Markov projection techniques.

In this article we present an FX Heston-type model in which the interest rates are

stochastic processes, correlated with the governing FX processes. We ﬁrst discuss the

Heston FX model with Gaussian interest rate (Hull-White model [HW96]) short-rate

processes. In this model a full matrix of correlations is used.

This model, denoted by FX-HHW here, is a generalization of our work in [GO09],

where we dealt with the problem of ﬁnding an aﬃne approximation of the Heston equity

model with a correlated stochastic interest rate. In this paper, we apply this technique

in the world of foreign exchange.

Secondly, we extend the framework by modeling the interest rates by a market model,

i.e., by the stochastic volatility displaced-diﬀusion Libor Market Model [AA02; Pit05].

In this hybrid model, called FX-HLMM here, we incorporate a non-zero correlation

between the FX and the interest rates and between the rates from diﬀerent currencies.

Because it is not possible to obtain closed-form formulas for the associated characteristic

function, we use a linearization approximation, developed earlier, in [GO10].

For both models we provide details on how to include a foreign stock in the multi-

currency pricing framework.

Fast model evaluation is highly desirable for FX options in practice, especially during

the calibration of the hybrid model. This is the main motivation for the generalization

of the linearization techniques in [GO09; GO10] to the world of foreign exchange. We

will see that the resulting approximations can be used very well in the FX context.

The present article is organized as follows. In Section 2 we discuss the extension of

the Heston model by stochastic interest rates, described by short-rate processes. We

provide details about some approximations in the model, and then derive the related

forward characteristic function. We also discuss the model’s accuracy and calibration

results. Section 3 gives the details for the cross-currency model with interest rates driven

by the market model and Section 4 concludes.

2 Multi-Currency Model with Short-Rate Interest

Rates

Here, we derive the model for the spot FX, ξ(t), expressed in units of domestic

currency, per unit of a foreign currency.

We start the analysis with the speciﬁcation of the underlying interest rate processes,

r

d

(t) and r

f

(t). At this stage we assume that the interest rate dynamics are deﬁned via

short-rate processes, which under their spot measures, i.e., Q−domestic and Z−foreign,

are driven by the Hull-White [HW96] one-factor model:

dr

d

(t) = λ

d

(θ

d

(t) −r

d

(t))dt +η

d

dW

Q

d

(t), (2.1)

dr

f

(t) = λ

f

(θ

f

(t) −r

f

(t))dt +η

f

dW

Z

f

(t), (2.2)

where W

Q

d

(t) and W

Z

f

(t) are Brownian motions under Q and Z, respectively. Parameters

λ

d

, λ

f

determine the speed of mean reversion to the time-dependent term structure

functions θ

d

(t), θ

f

(t), and parameters η

d

, η

f

are the volatility coeﬃcients.

These processes, under the appropriate measures, are linear in their state variables,

so that for a given maturity T (0 < t < T) the zero-coupon bonds (ZCB) are known to

be of the following form:

P

d

(t, T) = exp (A

d

(t, T) +B

d

(t, T)r

d

(t)) ,

P

f

(t, T) = exp (A

f

(t, T) +B

f

(t, T)r

f

(t)) ,

(2.3)

2

with A

d

(t, T), A

f

(t, T) and B

d

(t, T), B

f

(t, T) analytically known quantities (see for

example [BM07]). In the model the money market accounts are given by:

dM

d

(t) = r

d

(t)M

d

(t)dt, and dM

f

(t) = r

f

(t)M

f

(t)dt. (2.4)

By using the Heath-Jarrow-Morton arbitrage-free argument, [HJM92], the dynamics

for the ZCBs, under their own measures generated by the money savings accounts, are

known and given by the following result:

Result 2.1 (ZCB dynamics under the risk-free measure). The risk-free dynamics of the

zero-coupon bonds, P

d

(t, T) and P

f

(t, T), with maturity T are given by:

dP

d

(t, T)

P

d

(t, T)

= r

d

(t)dt −

T

t

Γ

d

(t, s)ds

dW

Q

d

(t), (2.5)

dP

f

(t, T)

P

f

(t, T)

= r

f

(t)dt −

T

t

Γ

f

(t, s)ds

dW

Z

f

(t), (2.6)

where Γ

d

(t, T), Γ

f

(t, T) are the volatility functions of the instantaneous forward rates

f

d

(t, T), f

f

(t, T), respectively, that are given by:

df

d

(t, T) = Γ

d

(t, T)

T

t

Γ

d

(t, s)ds + Γ

d

(t, T)dW

Q

d

(t), (2.7)

df

f

(t, T) = Γ

f

(t, T)

T

t

Γ

f

(t, s)ds + Γ

f

(t, T)dW

Z

f

(t). (2.8)

Proof. For the proof see [MR97].

The spot-rates at time t are deﬁned by r

d

(t) ≡ f

d

(t, t), r

f

(t) ≡ f

f

(t, t).

By means of the volatility structures, Γ

d

(t, T), Γ

f

(t, T), one can deﬁne a number

of short-rate processes. In our framework the volatility functions are chosen to be

Γ

d

(t, T) = η

d

exp (−λ

d

(T −t)) and Γ

f

(t, T) = η

f

exp (−λ

f

(T −t)). The Hull-White

short-rate processes, r

d

(t) and r

f

(t) as in (2.1), (2.2), are then obtained and the term

structures, θ

d

(t), θ

f

(t), expressed in terms of instantaneous forward rates, are also

known. The choice of speciﬁc volatility determines the dynamics of the ZCBs:

dP

d

(t, T)

P

d

(t, T)

= r

d

(t)dt +η

d

B

d

(t, T)dW

Q

d

(t),

dP

f

(t, T)

P

f

(t, T)

= r

f

(t)dt +η

f

B

f

(t, T)dW

Z

f

(t), (2.9)

with B

d

(t, T) and B

f

(t, T) as in (2.3), given by:

B

d

(t, T) =

1

λ

d

e

−λ

d

(T−t)

−1

, B

f

(t, T) =

1

λ

f

e

−λ

f

(T−t)

−1

. (2.10)

For a detailed discussion on short-rate processes, we refer to the analysis of Musiela and

Rutkowski in [MR97]. In the next subsection we deﬁne the FX hybrid model.

2.1 The Model with Correlated, Gaussian Interest Rates

The FX-HHW model, with all processes deﬁned under the domestic risk-neutral

measure, Q, is of the following form:

dξ(t)/ξ(t) = (r

d

(t) −r

f

(t)) dt+

σ(t)dW

Q

ξ

(t), ξ(0) > 0,

dσ(t) = κ(¯ σ −σ(t))dt+ γ

σ(t)dW

Q

σ

(t), σ(0) > 0,

dr

d

(t) = λ

d

(θ

d

(t) −r

d

(t))dt+ η

d

dW

Q

d

(t), r

d

(0) > 0,

dr

f

(t) =

λ

f

(θ

f

(t) −r

f

(t)) −η

f

ρ

ξ,f

σ(t)

dt+ η

f

dW

Q

f

(t), r

f

(0) > 0.

(2.11)

3

Here, the parameters κ, λ

d

, and λ

f

determine the speed of mean reversion of the latter

three processes, their long term mean is given by ¯ σ, θ

d

(t), θ

f

(t), respectively. The

volatility coeﬃcients for the processes r

d

(t) and r

f

(t) are given by η

d

and η

f

and the

volatility-of-volatility parameter for process σ(t) is γ.

In the model we assume a full matrix of correlations between the Brownian motions

W(t) =

W

Q

ξ

(t), W

Q

σ

(t), W

Q

d

(t), W

Q

f

(t)

T

:

dW(t)(dW(t))

T

=

¸

¸

¸

1 ρ

ξ,σ

ρ

ξ,d

ρ

ξ,f

ρ

ξ,σ

1 ρ

σ,d

ρ

σ,f

ρ

ξ,d

ρ

σ,d

1 ρ

d,f

ρ

ξ,f

ρ

σ,f

ρ

d,f

1

¸

dt. (2.12)

Under the domestic-spot measure the drift in the short-rate process, r

f

(t), gives rise to an

additional term, −η

f

ρ

ξ,f

**σ(t). This term ensures the existence of martingales, under
**

the domestic spot measure, for the following prices (for more discussion, see [Shr04]):

χ

1

(t) := ξ(t)

M

f

(t)

M

d

(t)

and χ

2

(t) := ξ(t)

P

f

(t, T)

M

d

(t)

,

where P

f

(t, T) is the price foreign zero-coupon bond (2.9), respectively, and the money

savings accounts M

d

(t) and M

f

(t) are from (2.4).

To see that the processes χ

1

(t) and χ

2

(t) are martingales, one can apply the Itˆo

product rule, which gives:

dχ

1

(t)

χ

1

(t)

=

σ(t)dW

Q

ξ

(t), (2.13)

dχ

2

(t)

χ

2

(t)

=

σ(t)dW

Q

ξ

(t) +η

f

B

f

(t, T)dW

Q

f

(t). (2.14)

The change of dynamics of the underlying processes, from the foreign-spot to the

domestic-spot measure, also inﬂuences the dynamics for the associated bonds, which,

under the domestic risk-neutral measure, Q, with the money savings account considered

as a num´eraire, have the following representations

dP

d

(t, T)

P

d

(t, T)

= r

d

(t)dt +η

d

B

d

(t, T)dW

Q

d

(t), (2.15)

dP

f

(t, T)

P

f

(t, T)

=

r

f

(t) −ρ

ξ,f

η

f

B

f

(t, T)

σ(t)

dt +η

f

B

f

(t, T)dW

Q

f

(t), (2.16)

with B

d

(t, T) and B

f

(t, T) as in (2.10).

2.2 Pricing of FX Options

In order to perform eﬃcient calibration of the model we need to be able to

price basic options on the FX rate, V (t, X(t)), for a given state vector, X(t) =

[ξ(t), σ(t), r

d

(t), r

f

(t)]

T

:

V (t, X(t)) = E

Q

M

d

(t)

M

d

(T)

max(ξ(T) −K, 0)

F

t

,

with

M

d

(t) = exp

t

0

r

d

(s)ds

.

Now, we consider a forward price, Π(t), such that:

E

Q

max(ξ(T) −K, 0)

M

d

(T)

F

t

=

V (t, X(t))

M

d

(t)

=: Π(t).

4

By Itˆo’s lemma we have:

dΠ(t) =

1

M

d

(t)

dV (t) −r

d

(t)

V (t)

M

d

(t)

dt, (2.17)

with V (t) := V (t, X(t)). We know that Π(t) must be a martingale, i.e.: E(dΠ(t)) = 0.

Including this in (2.17) gives the following Fokker-Planck forward equation for V :

r

d

V =

1

2

η

2

f

∂

2

V

∂r

2

f

+ρ

d,f

η

d

η

f

∂

2

V

∂r

d

∂r

f

+

1

2

η

2

d

∂

2

V

∂r

2

d

+ρ

σ,f

γη

f

√

σ

∂

2

V

∂σ∂r

f

+ρ

σ,d

γη

d

√

σ

∂

2

V

∂σ∂r

d

+

1

2

γ

2

σ

∂

2

V

∂σ

2

+ρ

ξ,f

η

f

ξ

√

σ

∂

2

V

∂ξ∂r

f

+ρ

x,d

η

d

ξ

√

σ

∂

2

V

∂ξ∂r

d

+ρ

ξ,σ

γξσ

∂

2

V

∂ξ∂σ

+

1

2

ξ

2

σ

∂

2

V

∂ξ

2

+

λ

f

(θ

f

(t) −r

f

) −ρ

ξ,f

η

f

√

σ

∂V

∂r

f

+λ

d

(θ

d

(t) −r

d

)

∂V

∂r

d

+κ(¯ σ −σ)

∂V

∂σ

+ (r

d

−r

f

)ξ

∂V

∂ξ

+

∂V

∂t

.

This 4D PDE contains non-aﬃne terms, like square-roots and products. It is therefore

diﬃcult to solve it analytically and a numerical PDE discretization, like ﬁnite diﬀerences,

needs to be employed. Finding a numerical solution for this PDE is therefore rather

expensive and not easily applicable for model calibration. In the next subsection we

propose an approximation of the model, which is useful for calibration.

2.2.1 The FX Model under the Forward Domestic Measure

To reduce the complexity of the pricing problem, we move from the spot measure,

generated by the money savings account in the domestic market, M

d

(t), to the forward

FX measure where the num´eraire is the domestic zero-coupon bond, P

d

(t, T). As

indicated in [MR97; Pit06], the forward is given by:

FX

T

(t) = ξ(t)

P

f

(t, T)

P

d

(t, T)

, (2.18)

where FX

T

(t) represents the forward exchange rate under the T-forward measure, and

ξ(t) stands for foreign exchange rate under the domestic spot measure. The superscript

should not be confused with the transpose notation used at other places in the text.

By switching from the domestic risk-neutral measure, Q, to the domestic T-forward

measure, Q

T

, the discounting will be decoupled from taking the expectation, i.e.:

Π(t) = P

d

(t, T)E

T

max

FX

T

(T) −K, 0

|F

t

. (2.19)

In order to determine the dynamics for FX

T

(t) in (2.18), we apply Itˆo’s formula:

dFX

T

(t) =

P

f

(t, T)

P

d

(t, T)

dξ(t) +

ξ(t)

P

d

(t, T)

dP

f

(t, T) −ξ(t)

P

f

(t, T)

P

2

d

(t, T)

dP

d

(t, T)

+ξ(t)

P

f

(t, T)

P

3

d

(t, T)

(dP

d

(t, T))

2

+

1

P

d

(t, T)

(dξ(t)dP

f

(t, T))

−

P

f

(t, T)

P

2

f

(t, T)

(dP

d

(t, T)dξ(t)) −

ξ(t)

P

2

d

(t, T)

dP

d

(t, T)dP

f

(t, T). (2.20)

After substitution of SDEs (2.11), (2.15) and (2.16) into (2.20), we arrive at the following

FX forward dynamics:

dFX

T

(t)

FX

T

(t)

= η

d

B

d

(t, T)

η

d

B

d

(t, T) −ρ

ξ,d

σ(t) −ρ

d,f

η

f

B

f

(t, T)

dt

+

σ(t)dW

Q

ξ

(t) −η

d

B

d

(t, T)dW

Q

d

(t) +η

f

B

f

(t, T)dW

Q

f

(t). (2.21)

5

Since FX

T

(t) is a martingale under the T-forward domestic measure, i.e.,

P

d

(t, T)E

T

(FX

T

(T)|F

t

) = P

d

(t, T)FX

T

(t) =: P

f

(t, T)ξ(t), the appropriate Brownian

motions under the T−forward domestic measure, dW

T

ξ

(t), dW

T

σ

(t), dW

T

d

(t) and

dW

T

f

(t), need to be determined.

A change of measure from domestic-spot to domestic T-forward measure requires a

change of num´eraire from money savings account, M

d

(t), to zero-coupon bond P

d

(t, T).

In the model we incorporate a full matrix of correlations, which implies that all processes

will change their dynamics by changing the measure from spot to forward. Lemma 2.2

provides the model dynamics under the domestic T-forward measure, Q

T

.

Lemma 2.2 (The FX-HHW model dynamics under the Q

T

measure). Under the T-

forward domestic measure, the model in (2.11) is governed by the following dynamics:

dFX

T

(t)

FX

T

(t)

=

σ(t)dW

T

ξ

(t) −η

d

B

d

(t, T)dW

T

d

(t) +η

f

B

f

(t, T)dW

T

f

(t), (2.22)

where

dσ(t) =

κ(¯ σ −σ(t)) +γρ

σ,d

η

d

B

d

(t, T)

σ(t)

dt +γ

σ(t)dW

T

σ

(t), (2.23)

dr

d

(t) =

λ

d

(θ

d

(t) −r

d

(t)) +η

2

d

B

d

(t, T)

dt +η

d

dW

T

d

(t), (2.24)

dr

f

(t) =

λ

f

(θ

f

(t) −r

f

(t)) −η

f

ρ

ξ,f

σ(t) +η

d

η

f

ρ

d,f

B

d

(t, T)

dt +η

f

dW

T

f

(t),

(2.25)

with a full matrix of correlations given in (2.12), and with B

d

(t, T), B

f

(t, T) given

by (2.10).

The proof can be found in Appendix A.

From the system in Lemma 2.2 we see that after moving from the domestic-spot

Q-measure to the domestic T-forward Q

T

measure, the forward exchange rate FX

T

(t)

does not depend explicitly on the short-rate processes r

d

(t) or r

f

(t). It does not contain

a drift term and only depends on dW

T

d

(t), dW

T

f

(t), see (2.22).

Remark. Since the sum of three correlated, normally distributed random variables,

Q = X+Y +Z, remains normal with the mean equal to the sum of the individual means

and the variance equal to

σ

2

Q

= σ

2

X

+σ

2

Y

+σ

2

Z

+ 2ρ

X,Y

σ

X

σ

Y

+ 2ρ

X,Z

σ

X

σ

Z

+ 2ρ

Y,Z

σ

Y

σ

Z

,

we can represent the forward (2.22) as:

dFX

T

/FX

T

=

σ +η

2

d

B

2

d

+η

2

f

B

2

f

−2ρ

ξ,d

η

d

B

d

√

σ

+2ρ

ξ,f

η

f

B

f

√

σ −2ρ

d,f

η

d

η

f

B

d

B

f

1

2

dW

T

F

. (2.26)

Although the representation in (2.26) reduces the number of Brownian motions in

the dynamics for the FX, one still needs to ﬁnd the appropriate cross-terms, like

dW

T

F

(t)dW

T

σ

(t), in order to obtain the covariance terms. For clarity we therefore prefer

to stay with the standard notation.

Remark. The dynamics of the forwards, FX

T

(t) in (2.22) or in (2.26), do not depend

explicitly on the interest rate processes, r

d

(t) and r

f

(t), and are completely described by

the appropriate diﬀusion coeﬃcients. This suggests that the short-rate variables will not

enter the pricing PDE. Note, that this is only the case for models in which the diﬀusion

coeﬃcient for the interest rate does not depend on the level of the interest rate.

In next section we derive the corresponding pricing PDE and provide model

approximations.

6

2.3 Approximations and the Forward Characteristic Function

As the dynamics of the forward foreign exchange, FX

T

(t), under the domestic forward

measure involve only the interest rate diﬀusions dW

T

d

(t) and dW

T

f

(t), a signiﬁcant

reduction of the pricing problem is achieved.

In order to ﬁnd the forward ChF we take, as usual, the log-transform of the forward

rate FX

T

(t), i.e.: x

T

(t) := log FX

T

(t), for which we obtain the following dynamics:

dx

T

(t) =

ζ(t,

σ(t)) −

1

2

σ(t)

dt +

σ(t)dW

T

ξ

(t) −η

d

B

d

dW

T

d

(t) +η

f

B

f

dW

T

f

(t),

(2.27)

with the variance process, σ(t), given by:

dσ(t) =

κ(¯ σ −σ(t)) +γρ

σ,d

η

d

B

d

σ(t)

dt +γ

σ(t)dW

T

σ

(t), (2.28)

where we used the notation B

d

:= B

d

(t, T) and B

f

:= B

f

(t, T), and

ζ(t,

σ(t)) = (ρ

x,d

η

d

B

d

−ρ

x,f

η

f

B

f

)

σ(t) +ρ

d,f

η

d

η

f

B

d

B

f

−

1

2

η

2

d

B

2

d

+η

2

f

B

2

f

.

By applying the Feynman-Kac theorem we can obtain the characteristic function of the

forward FX rate dynamics. The forward characteristic function:

φ

T

:= φ

T

(u, X(t), t, T) = E

T

e

iux

T

(T)

F

t

,

with ﬁnal condition, φ

T

(u, X(T), T, T) = e

iux

T

(T)

, is the solution of the following

Kolmogorov backward partial diﬀerential equation:

−

∂φ

T

∂t

=

κ(¯ σ −σ) +ρ

σ,d

γη

d

√

σB

d

(t, T)

∂φ

T

∂σ

+

1

2

σ −ζ(t,

√

σ)

∂

2

φ

T

∂x

2

−

∂φ

T

∂x

+

ρ

x,σ

γσ −ρ

σ,d

γη

d

√

σB

d

(t, T) +ρ

σ,f

γη

f

√

σB

f

(t, T)

∂

2

φ

T

∂x∂σ

+

1

2

γ

2

σ

∂

2

φ

T

∂σ

2

.

This PDE contains however non-aﬃne

√

σ-terms so that it is nontrivial to ﬁnd the

solution. Recently, in [GO09], we have proposed two methods for linearization of these

non-aﬃne

1

square-roots of the square root process [CIR85]. The ﬁrst method is to

project the non-aﬃne square-root terms on their ﬁrst moments. This is also the approach

followed here.

The approximation of the non-aﬃne terms in the corresponding PDE is then done

as follows. We assume:

σ(t) ≈ E

σ(t)

=: ϕ(t), (2.29)

with the expectation of the square root of σ(t) given by:

E

σ(t)

=

2c(t)e

−ω(t)/2

∞

¸

k=0

1

k!

((t)/2)

k

Γ

1+

2

+k

Γ(

2

+k)

, (2.30)

and

c(t) =

1

4κ

γ

2

(1 −e

−κt

), =

4κ¯ σ

γ

2

, (t) =

4κσ(0)e

−κt

γ

2

(1 −e

−κt

)

. (2.31)

1

According to [DPS00] the n-dimensional system of SDEs:

dX(t) = µ(X(t))dt + σ(X(t))dW(t),

is of the aﬃne form if:

µ(X(t)) = a

0

+ a

1

X(t), for any (a

0

, a

1

) ∈ R

n

×R

n×n

,

σ(X(t))σ(X(t))

T

= (c

0

)

ij

+ (c

1

)

T

ij

X(t), for arbitrary (c

0

, c

1

) ∈ R

n×n

×R

n×n×n

,

r(X(t)) = r

0

+ r

T

1

X(t), for (r

0

, r

1

) ∈ R ×R

n

,

for i, j = 1, . . . , n, with r(X(t)) being an interest rate component.

7

Γ(k) is the gamma function deﬁned by:

Γ(k) =

∞

0

t

k−1

e

−t

dt.

Although the expectation in (2.30) is a closed form expression, its evaluation is rather

expensive. One may prefer to use a proxy, for example,

E(

σ(t)) ≈ β

1

+β

2

e

−β

3

t

, (2.32)

in which the constant coeﬃcients β

1

, β

2

and β

3

can be determined by asymptotic equality

with (2.30) (see [GO09] for details).

Projection of the non-aﬃne terms on their ﬁrst moments allows us to derive the

corresponding forward characteristic function, φ

T

, which is then of the following form:

φ

T

(u, X(t), t, T) = exp(A(u, τ) +B(u, τ)x

T

(t) +C(u, τ)σ(t)),

where τ = T −t, and the functions A(τ) := A(u, τ), B(τ) := B(u, τ) and C(τ) := C(u, τ)

are given by:

B

(τ) = 0,

C

(τ) = −κC(τ) + (B

2

(τ) −B(τ))/2 +ρ

x,σ

γB(τ)C(τ) +γ

2

C

2

(τ)/2,

A

(τ) = κ¯ σC(τ) +ρ

σ,d

γη

d

ϕ(τ)B

d

(τ)C(τ) −ζ(τ, ϕ(τ))

B

2

(τ) −B(τ)

+(−ρ

σ,d

η

d

γϕ(τ)B

d

(τ) +ρ

σ,f

γη

f

ϕ(τ)B

f

(τ)) B(τ)C(τ),

with ϕ(t) = E(

σ(t)), and B

i

(τ) = λ

−1

i

e

−λ

i

τ

−1

**for i = {d, f}. The initial
**

conditions are: B(0) = iu, C(0) = 0 and A(0) = 0.

With B(τ) = iu, the complex-valued function C(τ) is of the Heston-type, [Hes93],

and its solution reads:

C(τ) =

1 −e

−dτ

γ

2

(1 −ge

−dτ

)

(κ −ρ

x,σ

γiu −d) , (2.33)

with d =

(ρ

x,σ

γiu −κ)

2

−γ

2

iu(iu −1), g =

κ −γρ

x,σ

iu −d

κ −γρ

x,σ

iu +d

.

The parameters κ, γ, ρ

x,σ

are given in (2.11).

Function A(τ) is given by:

A(τ) =

τ

0

κ¯ σ +ρ

σ,d

γη

d

ϕ(s)B

d

(s) −ρ

σ,d

η

d

γϕ(s)B

d

(s)iu

+ρ

σ,f

γη

f

ϕ(s)B

f

(s)iu

C(s)ds + (u

2

+iu)

τ

0

ζ(s, ϕ(s))ds, (2.34)

with C(s) in (2.33). It is most convenient to solve A(τ) numerically with, for example,

Simpson’s quadrature rule. With correlations ρ

σ,d

, ρ

σ,f

equal to zero, a closed-form

expression for A(τ) would be available [GO09].

We denote the approximation, by means of linearization, of the full-scale FX-HHW

model by FX-HHW1. It is clear that eﬃcient pricing with Fourier-based methods can

be done with FX-HHW1, and not with FX-HHW.

By the projection of

σ(t) on its ﬁrst moment in (2.29) the corresponding PDE is

aﬃne in its coeﬃcients, and reads:

−

∂φ

T

∂t

= (κ(¯ σ −σ) + Ψ

1

)

∂φ

T

∂σ

+

1

2

σ −ζ(t, ϕ(t))

∂

2

φ

T

∂x

2

−

∂φ

T

∂x

+(ρ

x,σ

γσ −Ψ

2

)

∂

2

φ

T

∂x∂σ

+

1

2

γ

2

σ

∂

2

φ

T

∂σ

2

, with: (2.35)

φ

T

(u, X(T), T, T) = E

T

e

iux

T

(T)

F

T

= e

iux

T

(T)

,

and ζ(t, ϕ(t)) = Ψ

3

+ρ

d,f

η

d

η

f

B

d

(t, T)B

f

(t, T) −

1

2

η

2

d

B

2

d

(t, T) +η

2

f

B

2

f

(t, T)

.

8

The three terms, Ψ

1

, Ψ

2

, and Ψ

3

, in the PDE (2.35) contain the function ϕ(t):

Ψ

1

:= ρ

σ,d

γη

d

B

d

(t, T)ϕ(t),

Ψ

2

:= (ρ

σ,d

γη

d

B

d

(t, T) −ρ

σ,f

γη

f

B

f

(t, T)) ϕ(t),

Ψ

3

:= (ρ

x,d

η

d

B

d

(t, T) −ρ

x,f

η

f

B

f

(t, T)) ϕ(t).

When solving the pricing PDE for t →T, the terms B

d

(t, T) and B

f

(t, T) tend to zero,

and all terms that contain the approximation vanish. The case t → 0 is furthermore

trivial, since

σ(t)

t→0

−→E(

σ(0)).

Under the T-forward domestic FX measure, the projection of the non-aﬃne terms

on their ﬁrst moments is expected to provide high accuracy. In Section 2.5 we perform

a numerical experiment to validate this.

It is worth mentioning that also an alternative approximation for the non-aﬃne terms

**σ(t) is available, see [GO09]. This alternative approach guarantees that the ﬁrst two
**

moments are exact. In this article we stay, however, with the ﬁrst representation.

2.4 Pricing a Foreign Stock in the FX-HHW Model

Here, we focus our attention on pricing a foreign stock, S

f

(t), in a domestic market.

With this extension we can in principle price equity-FX-interest rate hybrid products.

With an equity smile/skew present in the market, we assume that S

f

(t) is given by

the Heston stochastic volatility model:

dS

f

(t)/S

f

(t) = r

f

(t)dt+

ω(t)dW

Z

S

f

(t),

dω(t) = κ

f

(¯ ω −ω(t))dt+ γ

f

ω(t)dW

Z

ω

(t),

dr

f

(t) = λ

f

(θ

f

(t) −r

f

(t)))dt+ η

f

dW

Z

f

(t),

(2.36)

where Z indicates the foreign-spot measure and the model parameters, κ

f

, γ

f

, λ

f

, θ

f

(t)

and η

f

, are as before.

Before deriving the stock dynamics in domestic currency, the model has to be

calibrated in the foreign market to plain vanilla options. This can be eﬃciently done

with the help of a fast pricing formula.

With the foreign short-rate process, r

f

(t), established in (2.11) we need to determine

the drifts for S

f

(t) and its variance process, ω(t), under the domestic spot measure. The

foreign stock, S

f

(t), can be expressed in domestic currency by multiplication with the

FX, ξ(t), and by discounting with the domestic money savings account, M

d

(t). Such

a stock is a tradable asset, so the price ξ(t)S

f

(t)/M

d

(t) (with ξ(t) in (2.11), S

f

(t)

from (2.36) and the domestic money-saving account M

d

(t) in (2.4)) needs to be a

martingale.

By applying Itˆo’s lemma to ξ(t)S

f

(t)/M

d

(t), we ﬁnd

d

ξ(t)

S

f

(t)

M

d

(t)

ξ(t)

S

f

(t)

M

d

(t)

= ρ

ξ,S

f

σ(t)

ω(t)dt +

σ(t)dW

Q

ξ

(t) +

ω(t)dW

Z

S

f

, (2.37)

where Q and Z indicate the domestic-spot and foreign-spot measures, respectively. To

make process ξ(t)S

f

(t)/M

d

(t) a martingale we set:

dW

Z

S

f

(t) = dW

Q

S

f

−ρ

ξ,S

f

σ(t)dt,

where σ(t) is the variance process of FX deﬁned in (2.11).

Under the change of measure, from foreign to domestic-spot, S

f

(t) has the following

dynamics:

dS

f

(t)/S

f

(t) = r

f

(t)dt +

ω(t)dW

Z

S

f

(t)

=

r

f

(t) −ρ

ξ,S

f

σ(t)

ω(t)

dt +

ω(t)dW

Q

S

f

(t). (2.38)

9

The variance process is correlated with the stock and by the Cholesky decomposition

we ﬁnd:

dω(t) = κ

f

(¯ ω −ω(t))dt +γ

f

ω(t)

ρ

S

f

,ω

d

W

Z

S

f

(t) +

1 −ρ

2

S

f

,ω

d

W

Z

ω

(t)

=

κ

f

(¯ ω −ω(t)) −ρ

S

f

,ω

ρ

S

f

,ξ

γ

f

ω(t)

σ(t)

dt +γ

f

ω(t)dW

Q

ω

(t). (2.39)

S

f

(t) in (2.38) and ω(t) in (2.39) are governed by several non-aﬃne terms. Those,

however, do not matter since the foreign stock is assumed to be already calibrated to the

foreign market data. The Monte Carlo simulation for pricing exotic options is deﬁned

in the domestic market. This implies that the presence of the non-aﬃne terms is not

complicating in this setting.

2.5 Numerical Experiment for the FX-HHW Model

In this section we check the errors resulting from the various approximations of the

FX-HHW1 model. We use the set-up from [Pit06], which means that the interest rate

curves are modeled by ZCBs deﬁned by P

d

(t = 0, T) = exp(−0.02T) and P

f

(t = 0, T) =

exp(−0.05T). Furthermore,

η

d

= 0.7%, η

f

= 1.2%, λ

d

= 1%, λ

f

= 5%.

We choose

2

:

κ = 0.5, γ = 0.3, ¯ σ = 0.1, σ(0) = 0.1. (2.40)

The correlation structure, deﬁned in (2.12), is given by:

¸

¸

¸

1 ρ

ξ,σ

ρ

ξ,d

ρ

ξ,f

ρ

ξ,σ

1 ρ

σ,d

ρ

σ,f

ρ

ξ,d

ρ

σ,d

1 ρ

d,f

ρ

ξ,f

ρ

σ,f

ρ

d,f

1

¸

=

¸

¸

¸

100% −40% −15% −15%

−40% 100% 30% 30%

−15% 30% 100% 25%

−15% 30% 25% 100%

¸

. (2.41)

The initial spot FX rate (Dollar, $, per Euro, e) is set to 1.35. For the FX-HHW model

we compute a number of FX option prices with many expiries and strikes, using two

diﬀerent pricing methods.

The ﬁrst method is the plain Monte Carlo method, with 50.000 paths and 20T

i

steps,

for the full-scale FX-HHW model, without any approximations.

For the second pricing method, we have used the ChF, based on the approximations

in the FX-HHW1 model in Section 2.3. Eﬃcient pricing of plain vanilla products is then

done by means of the COS method [FO08], based on a Fourier cosine series expansion of

the probability density function, which is recovered by the ChF with 500 Fourier cosine

terms.

We also deﬁne the experiments as in [Pit06], with expiries given by T

1

, . . . , T

10

, and

the strikes are computed by the formula:

K

n

(T

i

) = FX

T

i

(0) exp

0.1δ

n

T

i

, with (2.42)

δ

n

= {−1.5, −1.0, −0.5, 0, 0.5, 1.0, 1.5},

and FX

T

i

(0) as in (2.18) with ξ(0) = 1.35. This formula for the strikes is convenient,

since for n = 4, strikes K

4

(T

i

) with i = 1, . . . , 10 are equal to the forward FX rates for

time T

i

. The strikes and maturities are presented in Table B.1 in Appendix B.

The option prices resulting from both models are expressed in terms of the implied

Black volatilities. The diﬀerences between the volatilities are tabulated in Table 2.1. The

approximation FX-HHW1 appears to be highly accurate for the parameters considered.

We report a maximum error of about 0.1% volatility for at-the-money options with a

maturity of 30 years and less than 0.07% for the other options.

In the next subsection the calibration results to FX market data are presented.

2

The model parameters do not satisfy the Feller condition, γ

2

> 2κ¯ σ.

10

T

i

K

1

(T

i

) K

2

(T

i

) K

3

(T

i

) K

4

(T

i

) K

5

(T

i

) K

6

(T

i

) K

7

(T

i

)

6m -0.0003 -0.0002 0.0000 0.0002 0.0003 0.0004 0.0005

1y -0.0001 -0.0001 -0.0001 -0.0001 -0.0001 -0.0001 -0.0001

3y 0.0005 0.0004 0.0002 -0.0001 -0.0003 -0.0006 -0.0009

5y 0.0006 0.0004 0.0002 0.0000 -0.0003 -0.0007 -0.0010

7y 0.0008 0.0006 0.0004 0.0003 0.0001 -0.0001 -0.0003

10y -0.0002 -0.0003 -0.0003 -0.0005 -0.0007 -0.0009 -0.0012

15y -0.0012 -0.0010 -0.0009 -0.0009 -0.0009 -0.0009 -0.0010

20y 0.0009 0.0009 0.0009 0.0008 0.0008 0.0007 0.0006

25y -0.0015 -0.0011 -0.0008 -0.0006 -0.0005 -0.0004 -0.0004

30y 0.0010 0.0011 0.0012 0.0012 0.0012 0.0012 0.0012

Table 2.1: Diﬀerences, in implied volatilities, between the FX-HHW and FX-HHW1

models. The corresponding FX option prices and the standard deviations are tabulated

in Table B.4. Strike K

4

(T

i

) is the at-the-money strike.

2.5.1 Calibration to Market Data

We discuss the calibration of the FX-HHW model to FX market data. In the

simulation the reference market implied volatilities are taken from [Pit06] and are

presented in Table B.2 in Appendix B. In the calibration routine the approximate model

FX-HHW1 was applied. The correlation structure is as in (2.41). In Figure 2.1 some of

the calibration results are presented.

1.1 1.2 1.3 1.4 1.5

0.08

0.09

0.1

0.11

0.12

0.13

Strike

i

m

p

l

i

e

d

B

l

a

c

k

V

o

l

a

t

i

l

i

t

y

Calibration results

1Y Market

1Y FX−HHW

0.6 0.8 1 1.2 1.4 1.6

0.09

0.1

0.11

0.12

0.13

0.14

0.15

0.16

0.17

0.18

Strike

i

m

p

l

i

e

d

B

l

a

c

k

V

o

l

a

t

i

l

i

t

y

Calibration results

10Y Market

10Y FX−HHW

0.5 1 1.5

0.15

0.16

0.17

0.18

0.19

0.2

0.21

0.22

0.23

0.24

0.25

Strike

i

m

p

l

i

e

d

B

l

a

c

k

V

o

l

a

t

i

l

i

t

y

Calibration results

20Y Market

20Y FX−HHW

Figure 2.1: Comparison of implied volatilities from the market and the FX-HHW1

model for FX European call options for maturities of 1, 10 and 20 years. The strikes

are provided in Table B.1 in Appendix B. ξ(0) = 1.35.

Our experiments show that the model can be well calibrated to the market data. For

long maturities and for deep-in-the money options some discrepancy is present. This is

however typical when dealing with the Heston model (not related to our approximation),

since the skew/smile pattern in FX does not ﬂatten for long maturities. This was

sometimes improved by adding jumps to the model (Bates’ model). In Appendix B in

Table B.3 the detailed calibration results are tabulated.

Short-rate interest rate models can typically provide a satisfactory ﬁt to at-the-

money interest rate products. They are therefore not used for pricing derivatives that

are sensitive to the interest rate skew. This is a drawback of the short-rate interest rate

models. In the next section an extension of the framework, so that interest rate smiles

and skews can be modeled as well, is presented.

3 Multi-Currency Model with Interest Rate Smile

In this section we discuss a second extension of the multi-currency model, in which

an interest rate smile is incorporated. This hybrid model models two types of smiles, the

smile for the FX rate and the smiles in the domestic and foreign ﬁxed income markets.

11

We abbreviate the model by FX-HLMM. It is especially interesting for FX products

that are exposed to interest rate smiles. A description of such FX hybrid products can

be found in the handbook by Hunter [Hun05].

A ﬁrst attempt to model the FX by stochastic volatility and interest rates driven by

a market model was proposed in [TT08], assuming independence between log-normal-

Libor rates and FX. In our approach we deﬁne a model with non-zero correlation between

FX and interest rate processes.

As in the previous sections, the stochastic volatility FX is of the Heston type, which

under domestic risk-neutral measure, Q, follows the following dynamics:

dξ(t)/ξ(t) = (. . . )dt+

σ(t)dW

Q

ξ

(t), S(0) > 0,

dσ(t) = κ(¯ σ −σ(t))dt+ γ

σ(t)dW

Q

σ

(t), σ(0) > 0,

(3.1)

with the parameters as in (2.11). Since we consider the model under the forward measure

the drift in the ﬁrst SDE does not need to be speciﬁed (the dynamics of domestic-forward

FX ξ(t)P

f

(t, T)/P

d

(t, T) do not contain a drift term).

In the model we assume that the domestic and foreign currencies are independently

calibrated to interest rate products available in their own markets. For simplicity, we

also assume that the tenor structure for both currencies is the same, i.e., T

d

≡ T

f

=

{T

0

, T

1

, . . . , T

N

≡ T} and τ

k

= T

k

− T

k−1

for k = 1 . . . N. For t < T

k−1

we deﬁne the

forward Libor rates L

d,k

(t) := L

d

(t, T

k−1

, T

k

) and L

f,k

(t) := L

f

(t, T

k−1

, T

k

) as

L

d,k

(t) :=

1

τ

k

P

d

(t, T

k−1

)

P

d

(t, T

k

)

−1

, L

f,k

(t) :=

1

τ

k

P

f

(t, T

k−1

)

P

f

(t, T

k

)

−1

. (3.2)

For each currency we choose the DD-SV Libor Market Model from [AA02] for the interest

rates, under the T-forward measure generated by the num´eraires P

d

(t, T) and P

f

(t, T),

given by:

dL

d,k

(t) = σ

d,k

φ

d,k

(t)

v

d

(t)

µ

d

(t)

v

d

(t)dt + dW

d,T

k

(t)

,

dv

d

(t) = λ

d

(v

d

(0) −v

d

(t))dt +η

d

v

d

(t)dW

d,T

v

(t),

(3.3)

and

dL

f,k

(t) = σ

f,k

φ

f,k

(t)

v

f

(t)

µ

f

(t)

v

f

(t)dt + d

´

W

f,T

k

(t)

,

dv

f

(t) = λ

f

(v

f

(0) −v

f

(t))dt +η

f

v

f

(t)d

´

W

f,T

v

(t),

(3.4)

with

µ

d

(t) = −

N

¸

j=k+1

τ

j

φ

d,j

(t)σ

d,j

1 +τ

j

L

d,j

(t)

ρ

d

k,j

, µ

f

(t) = −

N

¸

j=k+1

τ

j

φ

f,j

(t)σ

f,j

1 +τ

j

L

f,j

(t)

ρ

f

k,j

, (3.5)

where

φ

d,k

= β

d,k

L

d,k

(t) + (1 −β

d,k

)L

d,k

(0),

φ

f,k

= β

f,k

L

f,k

(t) + (1 −β

f,k

)L

f,k

(0).

The Brownian motion, dW

d,T

k

, corresponds to the k-th domestic Libor rate, L

d,k

(t),

under the T-forward domestic measure, and the Brownian motion, d

´

W

f,T

k

relates to the

k-th foreign market Libor rate, L

f,k

(t) under the terminal foreign measure T.

In the model σ

d,k

(t) and σ

f,k

(t) determine the level of the interest rate volatility

smile, the parameters β

d,k

(t) and β

f,k

(t) control the slope of the volatility smile, and

λ

d

, λ

f

determine the speed of mean-reversion for the variance and inﬂuence the speed at

which the interest rate volatility smile ﬂattens as the swaption expiry increases [Pit05].

Parameters η

d

, η

f

determine the curvature of the interest rate smile.

12

The following correlation structure is imposed, between

FX and its variance process, σ(t): dW

T

ξ

(t)dW

T

σ

(t) = ρ

ξ,σ

dt,

FX and domestic Libors, L

d,j

(t): dW

T

ξ

(t)dW

d,T

j

(t) = ρ

d

ξ,j

dt,

FX and foreign Libors, L

f,j

(t): dW

T

ξ

(t)d

´

W

f,T

j

(t) = ρ

f

ξ,j

dt,

Libors in domestic market: dW

d,T

k

(t)dW

d,T

j

(t) = ρ

d

k,j

dt,

Libors in foreign market: d

´

W

f,T

k

(t)d

´

W

f,T

j

(t) = ρ

f

k,j

dt,

Libors in domestic and foreign markets: dW

d,T

k

(t)d

´

W

f,T

j

(t) = ρ

d,f

k,j

dt.

(3.6)

We prescribe a zero correlation between the remaining processes, i.e., between

Libors and their variance process,

dW

d,T

k

(t)dW

d,T

v

(t) = 0, d

´

W

f,T

k

(t)d

´

W

f,T

v

(t) = 0,

Libors and the FX variance process,

dW

d,T

k

(t)dW

T

σ

(t) = 0, d

´

W

f,T

k

(t)dW

T

σ

(t) = 0,

all variance processes,

dW

T

σ

(t)dW

d,T

v

(t) = 0, dW

T

σ

(t)d

´

W

f,T

v

(t) = 0, dW

d,T

v

(t)d

´

W

f,T

v

(t) = 0,

FX and the Libor variance processes,

dW

T

ξ

(t)dW

d,T

v

(t) = 0, dW

T

ξ

(t)d

´

W

f,T

v

(t) = 0.

The correlation structure is graphically displayed in Figure 3.1.

Figure 3.1: The correlation structure for the FX-HLMM model. Arrows indicate non-

zero correlations. SV is Stochastic Volatility.

Throughout this article we assume that the DD-SV model in (3.3) and (3.4) is

already in the eﬀective parameter framework as developed in [Pit05]. This means

that approximate time-homogeneous parameters are used instead of the time-dependent

parameters, i.e., β

k

(t) ≡ β

k

and σ

k

(t) ≡ σ

k

.

With this correlation structure, we derive the dynamics for the forward FX, given

by:

FX

T

(t) = ξ(t)

P

f

(t, T)

P

d

(t, T)

, (3.7)

(see also (2.18)) with ξ(t) the spot exchange rate and P

d

(t, T) and P

f

(t, T) zero-coupon

bonds. Note that the bonds are not yet speciﬁed.

13

When deriving the dynamics for (3.7), we need expressions for the zero-coupon bonds,

P

d

(t, T) and P

f

(t, T). With Equation (3.2) the following expression for the ﬁnal bond

can be obtained:

1

P

i

(t, T)

=

1

P

i

(t, T

m(t)

)

N

¸

j=m(t)+1

(1 +τ

j

L

i,j

(t)) , for i = {d, f}, (3.8)

with T = T

N

and m(t) = min(k : t ≤ T

k

) (empty products in (3.8) are deﬁned to

be equal to 1). The bond P

i

(t, T

N

) in (3.8) is fully determined by the Libor rates

L

i,k

(t), k = 1, . . . , N and the bond P

i

(t, T

m(t)

). Whereas the Libors L

i,k

(t) are deﬁned

by System (3.3) and (3.4), the bond P

i

(t, T

m(t)

) is not yet well-deﬁned in the current

framework.

To deﬁne continuous time dynamics for a zero-coupon bond, interpolation techniques

are available (see, for example, [Sch02a; Pit04; DMP09; BJ09]). We consider here the

linear interpolation scheme, proposed in [Sch02a], which reads:

1

P

i

(t, T

m(t)

)

= 1 + (T

m(t)

−t)L

i,m(t)

(T

m(t)−1

), for T

m(t)−1

< t < T

m(t)

. (3.9)

In our previous work, [GO10], this basic interpolation technique was very satisfactory

for the calibration. By combining (3.9) with (3.8), we ﬁnd for the domestic and foreign

bonds:

1

P

d

(t, T)

=

1 + (T

m(t)

−t)L

d,m(t)

(T

m(t)−1

)

N

¸

j=m(t)+1

(1 +τ

j

L

d

(t, T

j−1

, T

j

)) ,

1

P

f

(t, T)

=

1 + (T

m(t)

−t)L

f,m(t)

(T

m(t)−1

)

N

¸

j=m(t)+1

(1 +τ

j

L

f

(t, T

j−1

, T

j

)) .

When deriving the dynamics for FX

T

(t) in (3.7) we will not encounter any dt-terms (as

FX

T

(t) has to be a martingale under the num´eraire P

d

(t, T)).

For each zero-coupon bond, P

d

(t, T) or P

f

(t, T), the dynamics are determined under

the appropriate T-forward measures (for P

d

(t, T) the domestic T-forward measure, and

for P

f

(t, T) the foreign T-forward measure). The dynamics for the zero-coupon bonds,

driven by the Libor dynamics in (3.3) and (3.4), are given by:

dP

d

(t, T)

P

d

(t, T)

= (. . . )dt −

v

d

(t)

N

¸

j=m(t)+1

τ

j

σ

d,j

φ

d,j

(t)

1 +τ

j

L

d,j

(t)

dW

d,T

j

(t), (3.10)

dP

f

(t, T)

P

f

(t, T)

= (. . . )dt −

v

f

(t)

N

¸

j=m(t)+1

τ

j

σ

f,j

φ

f,j

(t)

1 +τ

j

L

f,j

(t)

d

´

W

f,T

j

(t), (3.11)

and the coeﬃcients were deﬁned in (3.3) and (3.4).

By changing the num´eraire from P

f

(t, T) to P

d

(t, T) for the foreign bond, only the

drift terms will change. Since FX

T

(t) in (3.7) is a martingale under the P

d

(t, T) measure,

it is not necessary to determine the appropriate drift correction.

By taking Equation (2.20) for the general dynamics of (3.7) and neglecting all the

dt-terms we get

dFX

T

(t)

FX

T

(t)

=

σ(t)dW

T

ξ

(t) +

v

d

(t)

N

¸

j=m(t)+1

τ

j

σ

d,j

φ

d,j

(t)

1 +τ

j

L

d,j

(t)

dW

d,T

j

(t)

−

v

f

(t)

N

¸

j=m(t)+1

τ

j

σ

f,j

φ

f,j

(t)

1 +τ

j

L

f,j

(t)

dW

f,T

j

(t). (3.12)

Note that the hat in

´

W, disappeared from the Brownian motion dW

f,T

j

(t) in (3.12) which

is an indication for the change of measure from the foreign to the domestic measure for

the foreign Libors.

14

Since the stochastic volatility process, σ(t), for FX is independent of the domestic

and foreign Libors, L

d,k

(t) and L

f,k

(t), the dynamics under the P

d

(t, T)-measure do not

change

3

and are given by:

dσ(t) = κ(¯ σ −σ(t))dt +γ

σ(t)dW

T

σ

(t). (3.13)

The model given in (3.12) with the stochastic variance in (3.13) and the correlations

between the main underlying processes is not aﬃne. In the next section we discuss a

linearization.

3.1 Linearization and Forward Characteristic Function

The model in (3.12) is not of the aﬃne form, as it contains terms like φ

i,j

(t)/(1 +

τ

i,j

L

i,j

(t)) with φ

i,j

= β

i,j

L

i,j

(t) + (1 − β

i,j

)L

i,j

(0) for i = {d, f}. In order to derive

a characteristic function, we freeze the Libor rates, which is standard practice (see for

example [GZ99; HW00; JR00]), i.e.:

L

d,j

(t) ≈ L

d,j

(0) ⇒ φ

d,j

≡ L

d,j

(0),

L

f,j

(t) ≈ L

f,j

(0) ⇒ φ

f,j

≡ L

f,j

(0). (3.14)

This approximation gives the following FX

T

(t)-dynamics:

dFX

T

(t)

FX

T

(t)

≈

σ(t)dW

T

ξ

(t) +

v

d

(t)

¸

j∈A

ψ

d,j

dW

d,T

j

(t) −

v

f

(t)

¸

j∈A

ψ

f,j

dW

f,T

j

(t),

dσ(t) = κ(¯ σ −σ(t))dt + γ

σ(t)dW

T

σ

(t),

dv

i

(t) = λ

i

(v

i

(0) −v

i

(t))dt + η

i

v

i

(t)dW

i,T

v

(t),

with i = {d, f}, A = {m(t) + 1, . . . N}, the correlations are given in (3.6) and

ψ

d,j

:=

τ

j

σ

d,j

L

d,j

(0)

1 +τ

j

L

d,j

(0)

, ψ

f,j

:=

τ

j

σ

f,j

L

f,j

(0)

1 +τ

j

L

f,j

(0)

. (3.15)

We derive the dynamics for the logarithmic transformation of FX

T

(t), x

T

(t) =

log FX

T

(t), for which we need to calculate the square of the diﬀusion coeﬃcients

4

.

With the notation,

a :=

σ(t)dW

T

ξ

(t), b :=

v

d

(t)

¸

j∈A

ψ

d,j

dW

d,T

j

(t), c :=

v

f

(t)

¸

j∈A

ψ

f,j

dW

f,T

j

(t),

(3.16)

we ﬁnd, for the square diﬀusion coeﬃcient (a +b −c)

2

= a

2

+b

2

+c

2

+2ab −2ac −2bc.

So, the dynamics for the log-forward, x

T

(t) = log FX

T

(t), can be expressed as:

dx

T

(t) ≈ −

1

2

(a +b −c)

2

+

σ(t)dW

T

ξ

(t) +

v

d

(t)

¸

A

ψ

d,j

dW

d,T

j

(t)

−

v

f

(t)

¸

A

ψ

f,j

dW

f,T

j

(t), (3.17)

with the coeﬃcients a, b and c given in (3.16). Since

N

¸

j=1

x

j

2

=

N

¸

j=1

x

2

j

+

¸

i,j=1,...,N

i=j

x

i

x

j

, for N > 0,

3

In [GO10] the proof for this statement is given when a single yield curve is considered.

4

As in the standard Black-Scholes analysis for dS(t) = σS(t)dW(t), the log-transform gives

d log S(t) = −

1

2

σ

2

dt + σdW(t).

15

we ﬁnd:

a

2

= σ(t)dt,

b

2

= v

d

(t)

¸

j∈A

ψ

2

d,j

+

¸

i,j∈A

i=j

ψ

d,i

ψ

d,j

ρ

d

i,j

dt =: v

d

(t)A

d

(t)dt, (3.18)

c

2

= v

f

(t)

¸

j∈A

ψ

2

f,j

+

¸

i,j∈A

i=j

ψ

f,i

ψ

f,j

ρ

f

i,j

dt, =: v

f

(t)A

f

(t)dt, (3.19)

ab =

σ(t)

v

d

(t)

¸

j∈A

ψ

d,j

ρ

d

j,x

dt,

ac =

σ(t)

v

f

(t)

¸

j∈A

ψ

f,j

ρ

f

j,x

dt,

bc =

v

d

(t)

v

f

(t)

¸

j∈A

ψ

d,j

¸

k∈A

ψ

f,k

ρ

d,f

j,k

dt,

with ρ

d

j,x

, ρ

f

j,x

the correlation between the FX and j-th domestic and foreign Libor,

respectively. The correlation between the k-th domestic and j-th foreign Libor is ρ

d,f

k,j

.

By setting f

t,

σ(t),

v

d

(t),

v

f

(t)

**:= (2ab − 2ac − 2bc)/dt, we can express the
**

dynamics for dx

T

(t) in (3.17) by:

dx

T

(t) ≈ −

1

2

σ(t) +A

d

(t)v

d

(t) +A

f

(t)v

f

(t) +f

t,

σ(t),

v

d

(t),

v

f

(t)

dt

+

σ(t)dW

T

ξ

(t) +

v

d

(t)

¸

A

ψ

d,j

dW

d,T

j

(t) −

v

f

(t)

¸

A

ψ

f,j

dW

f,T

j

(t).

The coeﬃcients ψ

d,j

, ψ

f,j

, A

d

and A

f

in (3.15), (3.18), and (3.19) are deterministic and

piecewise constant.

In order to make the model aﬃne, we linearize the non-aﬃne terms in the drift in

f(t,

σ(t),

v

d

(t),

v

f

(t)) by a projection on the ﬁrst moments, i.e.,

f

t,

σ(t),

v

d

(t),

v

f

(t)

≈ f

t, E(

σ(t)), E(

v

d

(t)), E(

v

f

(t))

=: f(t). (3.20)

The variance processes σ(t), v

d

(t) and v

f

(t) are independent CIR-type processes [CIR85],

so the expectation of their products equals the product of the expectations. Function

f(t) can be determined with the help of the formula in (2.30).

The approximation in (3.20) linearizes all non-aﬃne terms in the corresponding PDE.

As before, the forward characteristic function, φ

T

:= φ

T

(u, X(t), t, T), is deﬁned as the

solution of the following backward PDE:

0 =

∂φ

T

∂t

+

1

2

(σ +A

d

(t)v

d

+A

f

(t)v

f

+f(t))

∂

2

φ

T

∂x

2

−

∂φ

T

∂x

+λ

d

(v

d

(0) −v

d

)

∂φ

T

∂v

d

+λ

f

(v

f

(0) −v

f

)

∂φ

T

∂v

f

+κ(¯ σ −σ)

∂φ

T

∂σ

+

1

2

η

2

d

v

d

∂

2

φ

T

∂v

2

d

+

1

2

η

2

f

v

f

∂

2

φ

T

∂v

2

f

+

1

2

γ

2

σ

∂

2

φ

T

∂σ

2

+ρ

x,σ

γσ

∂

2

φ

T

∂x∂σ

, (3.21)

with the ﬁnal condition φ

T

(u, X(T), T) = e

iux

T

(T)

. Since all coeﬃcients in this PDE are

linear, the solution is of the following form:

φ

T

(u, X(t), t, T) = exp

A(u, τ) +B(u, τ)x

T

(t) +C(u, τ)σ(t)

+D

d

(u, τ)v

d

(t) +D

f

(u, τ)v

f

(t)

, (3.22)

16

with τ := T −t. Substitution of (3.22) in (3.21) gives us the following system of ODEs

for the functions A(τ) := A(u, τ), B(τ) := B(u, τ), C(τ) := C(u, τ), D

d

(τ) := D

d

(u, τ)

and D

f

(τ) := D

f

(u, τ):

A

(τ) = f(t)(B

2

(τ) −B(τ))/2 +λ

d

v

d

(0)D

1

(τ) +λ

f

v

f

(0)D

2

(τ) +κ¯ σC(τ),

B

(τ) = 0,

C

(τ) = (B

2

(τ) −B(τ))/2 + (ρ

x,σ

γB(τ) −κ) C(τ) +γ

2

C

2

(τ)/2,

D

d

(τ) = A

d

(t)(B

2

(τ) −B(τ))/2 −λ

d

D

d

(τ) +η

2

d

D

2

d

(τ)/2,

D

f

(τ) = A

f

(t)(B

2

(τ) −B(τ))/2 −λ

f

D

f

(τ) +η

2

f

D

2

f

(τ)/2,

with initial conditions A(0) = 0, B(0) = iu, C(0) = 0, D

d

(0) = 0, D

f

(0) = 0 with A

d

(t)

and A

f

(t) from (3.18), (3.19), respectively, and f(t) as in (3.20).

With B(τ) = iu, the solution for C(τ) is analogous to the solution for the ODE for

the FX-HHW1 model in Equation (2.33). As the remaining ODEs involve the piecewise

constant functions A

d

(t), A

f

(t) the solution must be determined iteratively, like for the

pure Heston model with piecewise constant parameters in [AA00]. For a given grid

0 = τ

0

< τ

1

< · · · < τ

N

= τ, the functions D

d

(u, τ), D

f

(u, τ) and A(u, τ) can be

expressed as:

D

d

(u, τ

j

) = D

d

(u, τ

j−1

) +χ

d

(u, τ

j

),

D

f

(u, τ

j

) = D

f

(u, τ

j−1

) +χ

f

(u, τ

j

),

for j = 1, . . . , N, and

A(u, τ

j

) = A(u, τ

j−1

) +χ

A

(u, τ

j

) −

1

2

(u

2

+u)

τ

j

τ

j−1

f(s)ds,

with f(s) in (3.20) and analytically known functions χ

k

(u, τ

j

), for k = {d, f} and

χ

A

(u, τ

j

):

χ

k

(u, τ

j

) :=

λ

k

−δ

k,j

−η

2

k

D

k

(u, τ

j−1

)

(1 −e

−δ

k,j

s

j

)

(η

2

k

(1 −

k,j

e

−δ

k,j

s

j

)),

and

χ

A

(u, τ

j

) =

κ¯ σ

γ

2

(κ −ρ

x,σ

γiu −d

j

)s

j

−2 log

(1 −g

j

e

−d

j

s

j

)

(1 −g

j

)

+v

d

(0)

λ

d

η

2

d

(λ

d

−δ

d,j

)s

j

−2 log

(1 −

d,j

e

−δ

d,j

s

j

)

(1 −

d,j

)

+v

f

(0)

λ

f

η

2

f

(λ

f

−δ

f,j

)s

j

−2 log

(1 −

f,j

e

−δ

f,j

s

j

)

(1 −

f,j

)

,

where

d

j

=

(ρ

x,σ

γiu −κ)

2

+γ

2

(iu +u

2

), g

j

=

(κ −ρ

x,σ

γiu) −d

j

−γ

2

C(u, τ

j−1

)

(κ −ρ

x,σ

γiu) +d

j

−γ

2

C(u, τ

j−1

)

,

δ

k,j

=

λ

2

k

+η

2

k

A

k

(t)(u

2

+iu),

k,j

=

λ

k

−δ

k,j

−η

2

k

D

k

(u, τ

j−1

)

λ

k

+δ

k,j

−η

2

k

D

k

(u, τ

j−1

)

,

with s

j

= τ

j

−τ

j−1

, j = 1, . . . , N, A

d

(t) and A

f

(t) are from (3.18) and (3.19).

The resulting approximation of the full-scale FX-HLMM model is called FX-LMM1

here.

3.2 Foreign Stock in the FX-HLMM Framework

We also consider a foreign stock, S

f

(t), driven by the Heston stochastic volatility

model, with the interest rates driven by the market model. The stochastic processes

17

of the stock model are assumed to be of the same form as the FX (with one, foreign,

interest rate curve) with the dynamics, under the forward foreign measure, given by:

dS

T

f

(t)

S

T

f

(t)

=

ω(t)dW

f,T

S

f

(t) +

v

f

(t)

N

¸

j=m(t)+1

τ

j

σ

f,j

φ

f,j

(t)

1 +τ

j

L

f,j

(t)

dW

f,T

j

(t),

dω(t) = κ

f

(¯ ω −ω(t))dt +γ

f

ω(t)dW

f,T

ω

(t). (3.23)

Variance process, ω(t), is correlated with forward stock S

T

(t).

We move to the domestic-forward measure. The forward stock, S

T

f

, and forward

foreign exchange rate, FX

T

(t), are deﬁned by

S

T

f

(t) =

S

f

(t)

P

f

(t, T)

, FX

T

(t) = ξ(t)

P

f

(t, T)

P

d

(t, T)

. (3.24)

The quantity

S

T

f

(t)FX

T

(t) =

S

f

(t)

P

f

(t, T)

ξ(t)

P

f

(t, T)

P

d

(t, T)

=

S

f

(t)

P

d

(t, T)

ξ(t), (3.25)

is therefore a tradable asset. So, foreign stock exchanged by a foreign exchange rate and

denominated in the domestic zero-coupon bond is a tradable quantity, which implies

that S

T

f

(t)FX

T

(t) is a martingale. By Itˆo’s lemma, one ﬁnds:

d

S

T

f

(t)FX

T

(t)

S

T

f

(t)FX

T

(t)

=

dFX

T

(t)

FX

T

(t)

+

dS

T

f

(t)

S

T

f

(t)

+

dFX

T

(t)

FX

T

(t)

dS

T

f

(t)

S

T

f

(t)

. (3.26)

The two ﬁrst terms at the RHS of (3.26) do not contribute to the drift. The last term

involves all dt-terms, that, by a change of measure, will enter the drift of the variance

process dω(t) in (3.23).

3.3 Numerical Experiments with the FX-HLMM Model

We here focus on the FX-HLMM model covered in Section 3 and consider the errors

generated by the various approximations that led to the model FX-HLMM1. We have

performed basically two linearization steps to deﬁne FX-HLMM1: We have frozen

the Libors at their initial values and projected the non-aﬃne covariance terms on a

deterministic function. We check, by a numerical experiment, the size of the errors of

these approximations.

We have chosen the following interest rate curves P

d

(t = 0, T) =

exp(−0.02T), P

f

(t = 0, T) = exp(−0.05T), and, as before, for the FX stochastic

volatility model we set:

κ = 0.5, γ = 0.3, ¯ σ = 0.1, σ(0) = 0.1. (3.27)

In the simulation we have chosen the following parameters for the domestic and foreign

markets:

β

d,k

= 95%, σ

d,k

= 15%, λ

d

= 100%, η

d

= 10%, (3.28)

β

f,k

= 50%, σ

f,k

= 25%, λ

f

= 70%, η

f

= 20%. (3.29)

In the correlation matrix a number of correlations need to be speciﬁed. For the

correlations between the Libor rates in each market, we prescribe large positive values,

as frequently observed in ﬁxed income markets (see for example [BM07]), ρ

d

i,j

=

90%, ρ

f

i,j

= 70%, for i, j = 1, . . . , N (i = j). In order to generate skew for FX, we

prescribe a negative correlation between FX

T

(t) and its stochastic volatility process,

σ(t), i.e., ρ

ξ,σ

= −40%. The correlation between the FX and the domestic Libors is

set as ρ

d

ξ,k

= −15%, for k = 1, . . . , N, and the correlation between FX and the foreign

18

Libors is ρ

f

ξ,k

= −15%. The correlation between the domestic and foreign Libors is

ρ

d,f

i,j

= 25% for i, j = 1, . . . , N (i = j). The following block correlation matrix results:

C =

C

d

C

d,f

C

ξ,d

C

T

d,f

C

f

C

ξ,f

C

T

ξ,d

C

T

ξ,f

1

¸

¸

, (3.30)

with the domestic Libor correlations given by

C

d

=

1 ρ

d

1,2

. . . ρ

d

1,N

ρ

d

1,2

1 . . . ρ

d

2,N

.

.

.

.

.

.

.

.

.

.

.

.

ρ

d

1,N

ρ

d

2,N

. . . 1

¸

¸

¸

¸

¸

¸

=

1 90% . . . 90%

90% 1 . . . 90%

.

.

.

.

.

.

.

.

.

.

.

.

90% 90% . . . 1

¸

¸

¸

¸

¸

N×N

, (3.31)

the foreign Libors correlations given by:

C

f

=

1 ρ

f

1,2

. . . ρ

f

1,N

ρ

f

1,2

1 . . . ρ

f

2,N

.

.

.

.

.

.

.

.

.

.

.

.

ρ

f

1,N

ρ

f

2,N

. . . 1

¸

¸

¸

¸

¸

¸

¸

=

1 70% . . . 70%

70% 1 . . . 70%

.

.

.

.

.

.

.

.

.

.

.

.

70% 70% . . . 1

¸

¸

¸

¸

¸

N×N

, (3.32)

the correlation between Libors from the domestic and foreign markets given by:

C

df

=

1 ρ

d,f

1,2

. . . ρ

d,f

1,N

ρ

d,f

1,2

1 . . . ρ

d,f

2,N

.

.

.

.

.

.

.

.

.

.

.

.

ρ

d,f

1,N

ρ

d,f

2,N

. . . 1

¸

¸

¸

¸

¸

¸

¸

=

1 25% . . . 25%

25% 1 . . . 25%

.

.

.

.

.

.

.

.

.

.

.

.

25% 25% . . . 1

¸

¸

¸

¸

¸

N×N

, (3.33)

and the vectors C

ξ,d

and C

ξ,f

, given by:

C

ξ,d

=

ρ

d

ξ,1

ρ

d

ξ,2

.

.

.

ρ

d

ξ,N

¸

¸

¸

¸

¸

¸

=

−15%

−15%

.

.

.

−15%

¸

¸

¸

¸

¸

N×1

, C

ξ,f

=

ρ

f

ξ,1

ρ

f

ξ,2

.

.

.

ρ

f

ξ,N

¸

¸

¸

¸

¸

¸

¸

=

−15%

−15%

.

.

.

−15%

¸

¸

¸

¸

¸

N×1

. (3.34)

Since in both markets the Libor rates are assumed to be independent of their variance

processes, we can neglect these correlations here.

Now we ﬁnd the prices of plain vanilla options on FX in (3.7). The simulation is

performed in the same spirit as in Section 2.5 where the FX-HHW model was considered.

In Table 3.1 we present the diﬀerences, in terms of the implied volatilities between the

models FX-HLMM and FX-HLMM1. While the prices for the FX-HLMM were obtained

by Monte-Carlo simulation (20.000 paths and 20 intermediate points between the dates

T

i−1

and T

i

for i = 1, . . . , N), the prices for FX-HLMM1 were obtained by the Fourier-

based COS method [FO08] with 500 Fourier series terms.

T

i

K

1

(T

i

) K

2

(T

i

) K

3

(T

i

) K

4

(T

i

) K

5

(T

i

) K

6

(T

i

) K

7

(T

i

)

2y 0.0019 0.0014 0.0009 0.0005 0.0000 -0.0005 -0.0010

3y 0.0029 0.0025 0.0021 0.0016 0.0011 0.0006 0.0002

5y 0.0032 0.0028 0.0023 0.0017 0.0010 0.0005 0.0000

7y 0.0030 0.0028 0.0025 0.0021 0.0018 0.0014 0.0010

10y 0.0039 0.0032 0.0025 0.0018 0.0012 0.0005 -0.0003

15y 0.0038 0.0029 0.0021 0.0013 0.0005 -0.0004 -0.0014

20y 0.0002 -0.0009 -0.0018 -0.0027 -0.0034 -0.0040 -0.0044

25y 0.0008 0.0004 -0.0014 -0.0025 -0.0034 -0.0040 -0.0046

30y 0.0011 0.0007 0.0000 -0.0009 -0.0018 -0.0021 -0.0024

Table 3.1: Diﬀerences, in implied Black volatilities, between the FX-HLMM and FX-

LMM1 models. The corresponding strikes K

1

(T

i

), . . . , K

7

(T

i

) are tabulated in Table B.1.

The prices and associated standard deviations are presented in Table B.5.

The FX-HLMM1 model performs very well, as the maximum diﬀerence in terms of

implied volatilities is about 0.2%−0.5%.

19

3.3.1 Sensitivity to the Interest Rate Skew

Approximation FX-HLMM1 was based on freezing the Libor rates. By freezing the

Libors, i.e.: L

d,k

(t) ≡ L

d,k

(0) and L

f,k

(t) ≡ L

f,k

(0) we have

φ

d,k

(t) = β

d,k

L

d,k

(t) + (1 −β

d,k

)L

d,k

(0) = L

d,k

(0), (3.35)

φ

f,k

(t) = β

f,k

L

f,k

(t) + (1 −β

f,k

)L

f,k

(0) = L

f,k

(0). (3.36)

In the DD-SV models for the Libor rates L

d,k

(t) and L

f,k

(t) for any k, the parameters

β

d,k

, β

f,k

control the slope of the interest rate volatility smiles. Freezing the Libors to

L

d,k

(0) and L

f,k

(0) is equivalent to setting β

d,k

= 0 and β

f,k

= 0 in (3.35) and (3.36)

in the approximation FX-HLMM1.

By a Monte Carlo simulation, we obtain the FX implied volatilities from the full

scale FX-HLMM model for diﬀerent values of β and by comparing them to those from

FX-HLMM1 iwith β = 0 we check the inﬂuence of the parameters β

d,k

and β

f,k

on

the FX. In Table 3.2 the implied volatilities for the FX European call options for FX-

HLMM and FX-HLMM1 are presented. The experiments are performed for diﬀerent

combinations of the interest rate skew parameters, β

d

and β

f

.

FX-HLMM (Monte Carlo simulation) FX-HLMM1 (Fourier)

strike β

f

= 0.5 β

d

= 0.5 β

d

= 0

(2.42) β

d

= 0 β

d

= 0.5 β

d

= 1 β

f

= 0 β

f

= 1 β

f

= 0

0.6224 0.3198 0.3191 0.3198 0.3199 0.3196 0.3156

(0.0020) (0.0017) (0.0017) (0.0015) (0.0018)

0.7290 0.3149 0.3143 0.3148 0.3151 0.3146 0.3112

(0.0021) (0.0016) (0.0019) (0.0015) (0.0018)

0.8538 0.3102 0.3096 0.3101 0.3104 0.3097 0.3069

(0.0021) (0.0017) (0.0020) (0.0015) (0.0018)

1.0001 0.3058 0.3053 0.3056 0.3061 0.3052 0.3030

(0.0021) (0.0017) (0.0022) (0.0015) (0.0017)

1.1714 0.3016 0.3011 0.3015 0.3020 0.3008 0.2993

(0.0020) (0.0017) (0.0024) (0.0015) (0.0016)

1.3721 0.2977 0.2973 0.2977 0.2982 0.2968 0.2960

(0.0022) (0.0016) (0.0026) (0.0016) (0.0017)

1.6071 0.2941 0.2938 0.2943 0.2948 0.2931 0.2930

(0.0024) (0.0017) (0.0028) (0.0017) (0.0018)

Table 3.2: Implied volatilities of the FX options from the FX-HLMM and FX-HLMM1

models, T = 10 and parameters were as in Section 3.3. The numbers in parentheses

correspond to the standard deviations (the experiment was performed 20 times with

20T time steps).

The experiment indicates that there is only a small impact of the diﬀerent β

d,k

−

and β

f,k

−values on the FX implied volatilities, implying that the approximate model,

FX-HLMM1 with β

d,k

= β

f,k

= 0, is useful for the interest rate modelling, for the

parameters studied. With β

d,k

= 0 and β

f,k

= 0 the implied volatilities obtained by the

FX-HLMM model appear to be somewhat higher than those obtained by FX-HLMM1,

a diﬀerence of approximately 0.1%−0.15%, which is considered highly satisfactory.

4 Conclusion

In this article we have presented two FX models with stochastic volatility and

correlated stochastic interest rates. Both FX models were based on the Heston FX

model and diﬀer with respect to the interest rate processes.

In the ﬁrst model we considered a model in which the domestic and foreign interest

rates were driven by single factor Hull-White short-rate processes. This model enables

pricing of FX-interest rate hybrid products that are not exposed to the smile in the ﬁxed

income markets.

For hybrid products sensitive to the interest rate skew a second model was presented

in which the interest rates were driven by the stochastic volatility Libor Market Model.

20

For both hybrid models we have developed approximate models for the pricing of

European options on the FX. These pricing formulas form the basis for highly eﬃcient

model calibration strategies.

The approximate models are based on the linearization of the non-aﬃne terms in the

corresponding pricing PDE, in a very similar way as in our previous article [GO09] on

equity-interest rate options. The approximate models perform very well in the world of

foreign exchange.

These models can also be used to obtain an initial guess when the full-scale models

are used.

Acknowledgments

The authors would like to thank to Sacha van Weeren and Natalia Borovykh from

Rabobank International for fruitful discussions and helpful comments.

References

[AA00] L.B.G. Andersen, J. Andreasen, Volatility Skews and Extensions of the Libor

Market Model. Appl. Math. Finance, 1(7): 1–32, 2000.

[AA02] L.B.G. Andersen, J. Andreasen, Volatile Volatilities. Risk, 15(12): 163–168, 2002.

[And07] J. Andreasen, Closed Form Pricing of FX Options Under Stochastic Rates and

Volatility, Presentation at Global Derivatives Conference 2006, Paris, 9-11 May 2006.

[AM06] A. Antonov, T. Misirpashaev, Eﬃcient Calibration to FX Options by Markovian

Projection in Cross-Currency LIBOR Market Models. SSRN working paper, 2006.

[AAA08] A. Antonov, M. Arneguy, N. Audet, Markovian Projection to a Displaced

Volatility Heston Model. SSRN working paper, 2008.

[BJ09] C. Beveridge, M. Joshi, Interpolation Schemes in the Displaced-Diﬀusion Libor

Market Model and the Eﬃcient Pricing and Greeks for Callable Range Accruals. SSRN

working paper, 2009.

[BM07] D. Brigo, F. Mercurio, Interest Rate Models- Theory and Practice: With Smile,

Inﬂation and Credit. Springer Finance, second edition, 2007.

[Cap07] O. Caps, On the Valuation of Power-Reverse Duals and Equity-Rates Hybrids.

Presentation at Frankfurt MathFinance Conference, Derivatives and Risk Management in

Theory and Practice, Frankfurt, 26-27 May 2007.

[CIR85] J.C. Cox, J.E. Ingersoll, S.A. Ross, A Theory of the Term Structure of Interest

Rates. Econometrica, 53(2): 385–407, 1985.

[DMP09] M.H.A. Davis, V. Mataix-Pastor, Arbitrage-Free Interpolation of the Swap

Curve. IJTAF, 12(7): 969–1005, 2009.

[DPS00] D. Duffie, J. Pan, K. Singleton, Transform Analysis and Asset Pricing for Aﬃne

Jump-Diﬀusions. Econometrica, 68: 1343–1376, 2000.

[FO08] F. Fang, C.W. Oosterlee, A Novel Pricing Method for European Options Based on

Fourier-Cosine Series Expansions. SIAM J. Sci. Comput., 31: 826-848, 2008.

[GKR96] H. Geman, N. El Karoui, J.C. Rochet, Changes of Num´eraire, Changes of

Probability Measures and Pricing of Options. J. App. Prob., 32: 443–548, 1996.

[Gie04] A. Giese, On the Pricing of Auto-Callable Equity Securities in the Presence of

Stochastic Volatility and Stochastic Interest Rates. Presentation at MathFinance Workshop:

Derivatives and Risk Management in Theory and Practice, Frankfurt, 2004.

[GZ99] P. Glasserman, X. Zhao, Fast Greeks by Simulation in Forward LIBOR Models. J.

Comp. Fin., 3(1): 5–39, 1999.

[GO09] L.A. Grzelak, C.W. Oosterlee, On the Heston Model with Stochastic Interest

Rates. Techn. Report 09-05, Delft Univ. Techn., the Netherlands, SSRN working paper,

2009.

[GO10] L.A. Grzelak, C.W. Oosterlee, An Equity-Interest Rate Hybrid Model With

Stochastic Volatility and Interest Rate Smile. Techn. Report 10-01, Delft Univ. Techn.,

the Netherlands, SSRN working paper, 2010.

[HP09] A.v. Haastrecht, A. Pelsser, Generic Pricing of FX, Inﬂation and Stock Options

Under Stochastic Interest Rates and Stochastic Volatility. SSRN working paper, 2009.

[HJM92] D. Heath, R.A. Jarrow, A. Morton, Bond Pricing and the Term Structure of

Interest Rates: A New Methodology for Contingent Claims Valuation. Econometrica, 1(60):

77–105, 1992.

21

[Hes93] S. Heston, A Closed-form Solution for Options with Stochastic Volatility with

Applications to Bond and Currency Options. Rev. Fin. Studies, 6(2): 327–343, 1993.

[HW96] J. Hull, A. White, Using Hull-White Interest Rate Trees. J. Derivatives, 3(3): 26–

36, 1996.

[HW00] J. Hull, A. White, Forward Rate Volatilities, Swap Rate Volatilities and the

Implementation of the Libor Market Model. J. Fixed Income, 10(2): 46–62, 2000.

[Hun05] C. Hunter, Hybrid Derivatives. The Euromoney Derivatives and Risk Management

Handbook, 2005.

[JR00] P. J¨ ackel, R. Rebonato, Linking Caplet and Swaption Volatilities in a BGM

Framework: Approximate Solutions. J. Comp. Fin., 6(4): 41–60, 2000.

[KJ07] A. Kawai, P. J¨ ackel, An Asymptotic FX Option Formula in the Cross Currency

Libor Market Model. Wilmott, 74-84, March 2007.

[Mik01] P. Mikkelsen, Cross-Currency LIBOR Market Models. Center for Analytica Finance

Aarhus School of Business, working paper no.85. 2001.

[MR97] M. Musiela, M. Rutkowski, Martingale Methods in Financial Modelling. Springer

Finance, ﬁrst edition, 1997.

[Pit05] V.V. Piterbarg, Time to Smile, Risk, 18(5): 71-75, 2005.

[Pit04] V.V. Piterbarg, Computing Deltas of Callable Libor Exotics in Forward Libor

Models. J. Comp. Fin., 7(2): 107–144, 2004.

[Pit06] V.V. Piterbarg, Smiling Hybrids, Risk, 19: 66-71, 2006.

[Sch02a] E. Schl¨ ogl, Arbitrage-Free Interpolation in Models of a Market Observable Interest

Rates. In K. Sandmann and P. Sch¨onbucher, editors, Advances in Finance and Stochastic:

Essays in Honour of Dieter Sondermann. 197–218. Springer Verlag, Heidelberg, 2002.

[Sch02b] E. Schl¨ ogl, A Multicurrency Extension of the Lognormal Interest Rate Market

Models. Finance and Stochastics, Springer, 6(2): 173-196, 2002.

[SO02] J. Sippel, S. Ohkoshi, All Power to PRDC notes. Risk, 15(11): 531-533, 2002.

[Shr04] S. Shreve, Stochastic Calculus for Finance II: Continuos-Time Models. New York:

Springer, 2004.

[TT08] A. Takahashi, K. Takehara, Fourier Transform Method with an Asymptotic

Expansion Approach: an Application to Currency Options. IJTAF, 11(4): 381–401, 2008.

22

A Proof of Lemma 2.2

Since the domestic short rate process, r

d

(t), is driven by one source of uncertainty

(only one Brownian motion dW

Q

d

(t)), it is convenient to change the order of the

state variables, from dX(t) = [dFX

T

(t)/FX

T

(t), dσ(t), dr

d

(t), dr

f

(t)]

T

to dX

∗

(t) =

[dr

d

(t), dr

f

(t), dσ(t), dFX

T

(t)/FX

T

(t)]

T

and express the model in terms of the inde-

pendent Brownian motions d

W

Q

(t) = [d

W

d

(t), d

W

f

(t), d

W

σ

(t), d

W

ξ

(t)]

T

, i.e.:

dr

d

dr

f

dσ

dFX

T

/FX

T

¸

¸

¸

¸

= µ(X

∗

)dt +

η

d

0 0 0

0 η

f

0 0

0 0 γ

√

σ 0

−η

d

B

d

η

f

B

f

0

√

σ

¸

¸

¸

¸

H

d

W

Q

d

d

W

Q

f

d

W

Q

σ

d

W

Q

ξ

¸

¸

¸

¸

¸

, (A.1)

which, equivalently, can be written as:

dX

∗

(t) = µ(X

∗

)dt +AHd

W

Q

(t), (A.2)

where µ(X

∗

) represents the drift for system dX

∗

(t) and H is the Cholesky lower-

triangular matrix of the following form:

H =

1 0 0 0

H

2,1

H

2,2

0 0

H

3,1

H

3,2

H

3,3

0

H

4,1

H

4,2

H

4,3

H

4,4

¸

¸

¸

¸

∆

=

1 0 0 0

ρ

f,d

H

2,2

0 0

ρ

σ,d

H

3,2

H

3,3

0

ρ

ξ,d

H

4,2

H

4,3

H

4,4

¸

¸

¸

¸

. (A.3)

The representation presented above seems to be favorable, since the short-rate process

r

d

(t) can be considered independently of the other processes.

The matrix model representation in terms of orthogonal Brownian motions results

in the following dynamics for the domestic short rate r

d

(t) under measure Q:

dr

d

(t) = λ

d

(θ

d

(t) −r

d

(t))dt +ζ

1

(t)d

W

Q

(t),

and for the domestic ZCB:

dP

d

(t, T)

P

d

(t, T)

= r

d

(t)dt +B

d

(t, T)ζ

1

(t)d

W

Q

(t),

with ζ

k

(t) being the k’th row vector resulting from multiplying the matrices A and H.

Note, that for the 1D Hull-White short rate processes ζ

1

(t) =

η

d

, 0, 0, 0

.

Now, we derive the Radon-Nikod´ ym derivative [GKR96], Λ

T

Q

(t),:

Λ

T

Q

(t) =

dQ

T

dQ

=

P

d

(t, T)

P

d

(0, T)M

d

(t)

. (A.4)

By calculating the Itˆo derivative of Equation (A.4) we get:

dΛ

T

Q

Λ

T

Q

= B

d

(t, T)ζ

1

(t)d

W

Q

(t), (A.5)

which implies that the Girsanov kernel for the transition from Q to Q

T

is given by

B

d

(t, T)ζ

1

(t) which is the T-bond volatility given by η

d

B

d

(t, T), i.e.:

Λ

T

Q

= exp

−

1

2

T

0

B

2

r

(s, T)ζ

2

1

(s)ds +

T

0

B

r

(s, T)ζ

1

(s)d

W

Q

(s)

. (A.6)

So,

d

W

T

(t) = −B

d

(t, T)ζ

T

1

(t)dt + d

W

Q

(t).

Since the vector ζ

T

1

(t) is of scalar form, the Brownian motion under the T-forward

measure is given by:

d

W

Q

(t) =

d

W

T

d

(t) +η

d

B

d

(t, T)dt, d

W

T

f

(t), d

W

T

σ

(t), d

W

T

ξ

(t)

T

.

23

Now, from the vector representation (A.2) we get that:

Hd

W

Q

=

η

d

B

d

+ d

W

T

d

dt

ρ

d,f

η

d

B

d

dt+ ρ

d,f

d

W

T

d

+H

2,2

d

W

T

f

ρ

σ,d

η

d

B

d

dt+ ρ

σ,d

d

W

T

d

+H

3,2

d

W

T

f

+H

3,3

d

W

T

ξ

ρ

ξ,d

η

d

B

d

dt+ ρ

ξ,d

d

W

T

d

+H

4,2

d

W

T

f

+H

4,3

d

W

T

ξ

+H

4,4

d

W

T

σ

¸

¸

¸

¸

¸

¸

¸

. (A.7)

Returning to the dependent Brownian motions under the T-forward measure, gives us:

dFX

T

(t)

FX

T

(t)

=

σ(t)dW

T

ξ

(t) −η

d

B

d

(t, T)dW

T

d

(t) +η

f

B

f

(t, T)dW

T

f

(t),

dσ(t) =

κ(¯ σ −σ(t)) +γρ

σ,d

η

d

B

d

(t, T)

σ(t)

dt +γ

σ(t)dW

T

σ

(t),

dr

d

(t) =

λ

d

(θ

d

(t) −r

d

(t)) +η

2

d

B

d

(t, T)

dt +η

d

dW

T

d

(t),

dr

f

(t) =

λ

f

(θ

f

(t) −r

f

(t)) −η

f

ρ

ξ,f

σ(t) +η

d

η

f

ρ

d,f

B

d

(t, T)

dt +η

f

dW

T

f

(t),

with full matrix of correlations given in (2.12).

B Tables

In this appendix we present tables with details for the numerical experiments.

T

i

K

1

(T

i

) K

2

(T

i

) K

3

(T

i

) K

4

(T

i

) K

5

(T

i

) K

6

(T

i

) K

7

(T

i

)

6m 1.1961 1.2391 1.2837 1.3299 1.3778 1.4273 1.4787

1y 1.1276 1.1854 1.2462 1.3101 1.3773 1.4479 1.5221

3y 0.9515 1.0376 1.1315 1.2338 1.3454 1.4671 1.5999

5y 0.8309 0.9291 1.0390 1.1620 1.2994 1.4531 1.6250

7y 0.7358 0.8399 0.9587 1.0943 1.2491 1.4257 1.6274

10y 0.6224 0.7290 0.8538 1.0001 1.1714 1.3721 1.6071

15y 0.4815 0.5844 0.7093 0.8608 1.0447 1.2680 1.5389

20y 0.3788 0.4737 0.5924 0.7409 0.9265 1.1587 1.4491

25y 0.3012 0.3868 0.4966 0.6377 0.8188 1.0514 1.3500

30y 0.2414 0.3174 0.4174 0.5489 0.7218 0.9492 1.2482

Table B.1: Expiries and strikes of FX options used in the FX-HHW model. Strikes

K

n

(T

i

) were calculated as given in (2.42) with ξ(0) = 1.35.

T

i

K

1

(T

i

) K

2

(T

i

) K

3

(T

i

) K

4

(T

i

) K

5

(T

i

) K

6

(T

i

) K

7

(T

i

)

6m 0.1141 0.1049 0.0966 0.0902 0.0872 0.0866 0.0868

1y 0.1223 0.1098 0.0982 0.0895 0.0859 0.0859 0.0865

3y 0.1294 0.1135 0.0989 0.0878 0.0834 0.0836 0.0846

5y 0.1344 0.1184 0.1038 0.0927 0.0876 0.0871 0.0883

7y 0.1429 0.1268 0.1123 0.1012 0.0952 0.0937 0.0943

10y 0.1643 0.1479 0.1334 0.1218 0.1143 0.1107 0.1099

15y 0.2093 0.1913 0.1756 0.1627 0.1529 0.1465 0.1429

20y 0.2296 0.2119 0.1968 0.1844 0.1750 0.1684 0.1646

25y 0.2397 0.2231 0.2092 0.1980 0.1895 0.1837 0.1802

30y 0.2509 0.2348 0.2217 0.2113 0.2035 0.1981 0.1948

Table B.2: Market implied Black volatilities for FX options as given in [Pit06]. The

strikes K

n

(T

i

) were tabulated in Table B.1.

24

T

i

K

1

(T

i

) K

2

(T

i

) K

3

(T

i

) K

4

(T

i

) K

5

(T

i

) K

6

(T

i

) K

7

(T

i

)

6m 0.0012 -0.0012 -0.0025 -0.0023 -0.0001 0.0020 0.0022

1y 0.0013 -0.0008 -0.0018 -0.0009 0.0014 0.0016 -0.0014

3y 0.0016 -0.0007 -0.0017 -0.0008 0.0018 0.0022 -0.0014

5y 0.0011 -0.0006 -0.0012 -0.0007 0.0010 0.0013 -0.0014

7y 0.0007 -0.0003 -0.0006 -0.0003 0.0006 0.0010 -0.0008

10y 0.0004 -0.0001 -0.0001 -0.0002 0.0002 0.0005 -0.0002

15y 0.0011 -0.0005 -0.0009 -0.0004 0.0003 0.0009 -0.0005

20y 0.0094 0.0039 0.0002 -0.0019 -0.0024 -0.0016 0.0002

25y 0.0143 0.0059 -0.0002 -0.0043 -0.0063 -0.0064 -0.0051

30y 0.0165 0.0070 0.0000 -0.0048 -0.0074 -0.0082 -0.0074

Table B.3: The calibration results for the FX-HHW model, in terms of the diﬀerences

between the market (given in Table B.2) and FX-HHW model implied volatilities.

Strikes K

n

(T

i

) are given in Table B.1.

T

i

method K

1

(T

i

) K

2

(T

i

) K

3

(T

i

) K

4

(T

i

) K

5

(T

i

) K

6

(T

i

) K

7

(T

i

)

6m MC 0.1907 0.1636 0.1382 0.1148 0.0935 0.0748 0.0585

std dev 0.0004 0.0004 0.0005 0.0004 0.0004 0.0004 0.0004

COS 0.1908 0.1637 0.1382 0.1147 0.0934 0.0746 0.0583

1y MC 0.2566 0.2209 0.1870 0.1553 0.1264 0.1008 0.0785

std dev 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007 0.0007

COS 0.2567 0.2210 0.1870 0.1554 0.1265 0.1008 0.0786

3y MC 0.3768 0.3281 0.2805 0.2349 0.1923 0.1538 0.1200

std dev 0.0014 0.0015 0.0015 0.0015 0.0015 0.0015 0.0014

COS 0.3765 0.3279 0.2804 0.2349 0.1926 0.1543 0.1207

5y MC 0.4216 0.3709 0.3205 0.2713 0.2246 0.1816 0.1432

std dev 0.0021 0.0021 0.0021 0.0020 0.0020 0.0019 0.0018

COS 0.4212 0.3706 0.3203 0.2713 0.2249 0.1822 0.1441

7y MC 0.4368 0.3878 0.3383 0.2895 0.2426 0.1986 0.1587

std dev 0.0018 0.0018 0.0018 0.0018 0.0018 0.0017 0.0016

COS 0.4362 0.3873 0.3380 0.2893 0.2425 0.1987 0.1590

10y MC 0.4310 0.3871 0.3420 0.2967 0.2521 0.2096 0.1702

std dev 0.0033 0.0033 0.0033 0.0033 0.0033 0.0031 0.0030

COS 0.4311 0.3873 0.3423 0.2971 0.2528 0.2106 0.1714

15y MC 0.3894 0.3553 0.3195 0.2826 0.2455 0.2092 0.1744

std dev 0.0038 0.0037 0.0037 0.0036 0.0036 0.0036 0.0035

COS 0.3900 0.3560 0.3202 0.2834 0.2463 0.2100 0.1754

20y MC 0.3362 0.3109 0.2838 0.2553 0.2260 0.1966 0.1677

std dev 0.0037 0.0037 0.0037 0.0037 0.0037 0.0036 0.0036

COS 0.3358 0.3104 0.2833 0.2548 0.2254 0.1960 0.1672

25y MC 0.2809 0.2626 0.2425 0.2211 0.1987 0.1757 0.1526

std dev 0.0048 0.0048 0.0048 0.0048 0.0047 0.0046 0.0045

COS 0.2814 0.2630 0.2429 0.2215 0.1990 0.1759 0.1529

30y MC 0.2322 0.2191 0.2046 0.1888 0.1720 0.1545 0.1367

std dev 0.0050 0.0050 0.0050 0.0050 0.0049 0.0048 0.0048

COS 0.2319 0.2188 0.2042 0.1883 0.1714 0.1539 0.1359

Table B.4: Average FX call option prices obtained by the FX-HHW model with 20

Monte-Carlo simulations, 50.000 paths and 20 × T

i

steps; MC stands for Monte Carlo

and COS for Fourier Cosine expansion technique ([FO08]) for the FX-HHW1 model with

500 expansion terms. The strikes K

n

(T

i

) are tabulated in Table B.1.

25

T

i

method K

1

(T

i

) K

2

(T

i

) K

3

(T

i

) K

4

(T

i

) K

5

(T

i

) K

6

(T

i

) K

7

(T

i

)

2y MC 0.3336 0.2889 0.2456 0.2046 0.1667 0.1327 0.1030

std dev 0.0008 0.0009 0.0010 0.0010 0.0011 0.0011 0.0012

COS 0.3326 0.2880 0.2450 0.2043 0.1667 0.1330 0.1037

3y MC 0.3786 0.3299 0.2823 0.2366 0.1939 0.1553 0.1213

std dev 0.0006 0.0007 0.0008 0.0009 0.0011 0.0012 0.0013

COS 0.3768 0.3282 0.2808 0.2354 0.1931 0.1548 0.1212

5y MC 0.4243 0.3738 0.3234 0.2743 0.2274 0.1843 0.1457

std dev 0.0012 0.0013 0.0014 0.0015 0.0016 0.0016 0.0016

COS 0.4222 0.3717 0.3215 0.2727 0.2265 0.1838 0.1457

7y MC 0.4399 0.3914 0.3424 0.2938 0.2470 0.2031 0.1631

std dev 0.0013 0.0014 0.0015 0.0016 0.0018 0.0019 0.0021

COS 0.4379 0.3893 0.3402 0.2918 0.2453 0.2017 0.1621

10y MC 0.4363 0.3928 0.3482 0.3031 0.2587 0.2162 0.1764

std dev 0.0012 0.0016 0.0019 0.0023 0.0026 0.0027 0.0028

COS 0.4338 0.3905 0.3461 0.3014 0.2576 0.2157 0.1767

15y MC 0.3964 0.3632 0.3280 0.2917 0.2550 0.2186 0.1834

std dev 0.0008 0.0010 0.0012 0.0014 0.0016 0.0019 0.0023

COS 0.3944 0.3613 0.3265 0.2907 0.2545 0.2190 0.1848

20y MC 0.3417 0.3171 0.2907 0.2629 0.2342 0.2052 0.1768

std dev 0.0010 0.0013 0.0015 0.0018 0.0021 0.0025 0.0030

COS 0.3416 0.3176 0.2918 0.2647 0.2367 0.2085 0.1806

25y MC 0.2886 0.2715 0.2525 0.2321 0.2107 0.1887 0.1664

std dev 0.0011 0.0014 0.0016 0.0019 0.0023 0.0027 0.0033

COS 0.2883 0.2715 0.2532 0.2335 0.2127 0.1913 0.1697

30y MC 0.2396 0.2281 0.2152 0.2011 0.1858 0.1699 0.1534

std dev 0.0012 0.0015 0.0018 0.0021 0.0024 0.0029 0.0035

COS 0.2393 0.2279 0.2152 0.2014 0.1866 0.1710 0.1548

Table B.5: Average FX call option prices obtained by the FX-HLMM model with 20

Monte-Carlo simulations, 50.000 paths and 20 × T

i

steps; MC stands for Monte Carlo

and COS for the Fourier Cosine expansion technique ([FO08]) for the FX-HLMM1 model

with 500 expansion terms. Values of the strikes K

n

(T

i

) are tabulated in Table B.1.

26

Research on the Heston dynamics in combination with correlated interest rates has led to some interesting models. In [And07] and [Gie04] an indirectly imposed correlation structure between Gaussian short-rates and FX was presented. The model is intuitively appealing, but it may give rise to very large model parameters [AAA08]. An alternative model was presented in [AM06; AAA08], in which calibration formulas were developed by means of Markov projection techniques. In this article we present an FX Heston-type model in which the interest rates are stochastic processes, correlated with the governing FX processes. We ﬁrst discuss the Heston FX model with Gaussian interest rate (Hull-White model [HW96]) short-rate processes. In this model a full matrix of correlations is used. This model, denoted by FX-HHW here, is a generalization of our work in [GO09], where we dealt with the problem of ﬁnding an aﬃne approximation of the Heston equity model with a correlated stochastic interest rate. In this paper, we apply this technique in the world of foreign exchange. Secondly, we extend the framework by modeling the interest rates by a market model, i.e., by the stochastic volatility displaced-diﬀusion Libor Market Model [AA02; Pit05]. In this hybrid model, called FX-HLMM here, we incorporate a non-zero correlation between the FX and the interest rates and between the rates from diﬀerent currencies. Because it is not possible to obtain closed-form formulas for the associated characteristic function, we use a linearization approximation, developed earlier, in [GO10]. For both models we provide details on how to include a foreign stock in the multicurrency pricing framework. Fast model evaluation is highly desirable for FX options in practice, especially during the calibration of the hybrid model. This is the main motivation for the generalization of the linearization techniques in [GO09; GO10] to the world of foreign exchange. We will see that the resulting approximations can be used very well in the FX context. The present article is organized as follows. In Section 2 we discuss the extension of the Heston model by stochastic interest rates, described by short-rate processes. We provide details about some approximations in the model, and then derive the related forward characteristic function. We also discuss the model’s accuracy and calibration results. Section 3 gives the details for the cross-currency model with interest rates driven by the market model and Section 4 concludes.

2

Multi-Currency Model with Short-Rate Interest Rates

Here, we derive the model for the spot FX, ξ(t), expressed in units of domestic currency, per unit of a foreign currency. We start the analysis with the speciﬁcation of the underlying interest rate processes, rd (t) and rf (t). At this stage we assume that the interest rate dynamics are deﬁned via short-rate processes, which under their spot measures, i.e., Q−domestic and Z−foreign, are driven by the Hull-White [HW96] one-factor model:

Q drd (t) = λd (θd (t) − rd (t))dt + ηd dWd (t), Z drf (t) = λf (θf (t) − rf (t))dt + ηf dWf (t),

(2.1) (2.2)

Q Z where Wd (t) and Wf (t) are Brownian motions under Q and Z, respectively. Parameters λd , λf determine the speed of mean reversion to the time-dependent term structure functions θd (t), θf (t), and parameters ηd , ηf are the volatility coeﬃcients. These processes, under the appropriate measures, are linear in their state variables, so that for a given maturity T (0 < t < T ) the zero-coupon bonds (ZCB) are known to be of the following form:

Pd (t, T ) = exp (Ad (t, T ) + Bd (t, T )rd (t)) , Pf (t, T ) = exp (Af (t, T ) + Bf (t, T )rf (t)) ,

(2.3)

2

T ) dPf (t. one can deﬁne a number of short-rate processes. expressed in terms of instantaneous forward rates.4)
By using the Heath-Jarrow-Morton arbitrage-free argument. In the next subsection we deﬁne the FX hybrid model. T ) = ηd exp (−λd (T − t)) and Γf (t. T )
T t T Z Γf (t.1 (ZCB dynamics under the risk-free measure).
ξ(0) > 0. t).1
The Model with Correlated. T ) = ηf exp (−λf (T − t)). T ) = Γd (t. T ). t Q Γd (t. Gaussian Interest Rates
The FX-HHW model. T ). s)ds + Γd (t. The risk-free dynamics of the zero-coupon bonds. given by: Bd (t. Z = rf (t)dt + ηf Bf (t. rd (0) > 0. are then obtained and the term structures. rf (0) > 0.6)
where Γd (t. respectively. By means of the volatility structures.8)
dff (t. T ) and Pf (t. Q. ff (t.1). T ). (2. T ) = rf (t)dt − Pf (t. T ) = Γf (t. T ) are the volatility functions of the instantaneous forward rates fd (t. with maturity T are given by: dPd (t.3). Q ηf dWf (t). T ) as in (2. T ). T ) = 1 e−λd (T −t) − 1 . T )dWf (t).
(2. T ) and Bd (t. [HJM92].
(2. In the model the money market accounts are given by: dMd (t) = rd (t)Md (t)dt. T ) = 1 λf e−λf (T −t) − 1 . In our framework the volatility functions are chosen to be Γd (t. T )dWd (t). (2. θd (t). λd Bf (t. the dynamics for the ZCBs. Γf (t. T )
t
Proof. T ). Q ηd dWd (t). rf (t) ≡ ff (t. The Hull-White short-rate processes. and dMf (t) = rf (t)Mf (t)dt. The choice of speciﬁc volatility determines the dynamics of the ZCBs: dPd (t.7) (2. Z Γf (t.2). are known and given by the following result: Result 2. Γf (t. (2. T )dWd (t).
(2.9)
with Bd (t. σ(0) > 0. Q σ(t)dWσ (t). Bf (t. T ) = rd (t)dt − Pd (t. Pd (t.11)
3
. Af (t. s)ds + Γf (t. T ) and Bf (t.f σ(t) dt+
Q σ(t)dWξ (t). Γd (t. that are given by:
T
dfd (t. T ).
2. T )
t T
Q Γd (t. T ) analytically known quantities (see for example [BM07]). T ) Pf (t. T ). s)ds dWf (t). (2.with Ad (t.10)
For a detailed discussion on short-rate processes. are also known. T ) Pd (t. rd (t) and rf (t) as in (2. is of the following form: dξ(t)/ξ(t) = dσ(t) = drd (t) = drf (t) = (rd (t) − rf (t)) dt+ κ(¯ − σ(t))dt+ γ σ λd (θd (t) − rd (t))dt+ λf (θf (t) − rf (t)) − ηf ρξ. s)ds dWd (t). T ) dPf (t. T )dWf (t). T ). For the proof see [MR97]. θf (t). with all processes deﬁned under the domestic risk-neutral measure. we refer to the analysis of Musiela and Rutkowski in [MR97]. The spot-rates at time t are deﬁned by rd (t) ≡ fd (t. T )
Q = rd (t)dt + ηd Bd (t. t). under their own measures generated by the money savings accounts.5)
(2.

d ρσ. ρd.13) (2. (2.d ρξ.4). we consider a forward price.10). The ¯ volatility coeﬃcients for the processes rd (t) and rf (t) are given by ηd and ηf and the volatility-of-volatility parameter for process σ(t) is γ. To see that the processes χ1 (t) and χ2 (t) are martingales. V (t.f σ(t).15)
=
Q rf (t) − ρξ. T ) Pf (t. θd (t). under the domestic risk-neutral measure. and λf determine the speed of mean reversion of the latter three processes. rf (t).14)
The change of dynamics of the underlying processes. This term ensures the existence of martingales. respectively. also inﬂuences the dynamics for the associated bonds. Md (t)
4
.9). Wf (t) : 1 ρ dW(t)(dW(t))T = ξ. respectively. T )
Q = rd (t)dt + ηd Bd (t. λd . Md (t) Md (t)
where Pf (t. one can apply the Itˆ o product rule. T )dWd (t).f ρσ. θf (t).16)
with Bd (t. under the domestic spot measure. Q Q σ(t)dWξ (t) + ηf Bf (t. gives rise to an additional term. T ) as in (2. which. In the model we assume a full matrix of correlations between the Brownian motions T Q Q Q Q W(t) = Wξ (t). 0) Ft Md (T ) = V (t. which gives: dχ1 (t) χ1 (t) dχ2 (t) χ2 (t) = =
Q σ(t)dWξ (t). 0) Ft .Here. T )dWf (t). have the following representations e dPd (t. X(t) = [ξ(t). rf (t)]T : V (t. and the money savings accounts Md (t) and Mf (t) are from (2. T ) and Bf (t. from the foreign-spot to the domestic-spot measure. T ) is the price foreign zero-coupon bond (2. X(t)) =: Π(t). with the money savings account considered as a num´raire.σ 1 ρσ. their long term mean is given by σ . Md (T )
t
rd (s)ds . Wd (t).
Now.12)
Under the domestic-spot measure the drift in the short-rate process. X(t)) = EQ with Md (t) = exp
0
Md (t) max(ξ(T ) − K. T ) Pd (t.f 1
(2. T )dWf (t). T ) dPf (t.f ρξ.
(2. for a given state vector. the parameters κ.f dt. rd (t).f ρξ.2
Pricing of FX Options
In order to perform eﬃcient calibration of the model we need to be able to price basic options on the FX rate. Π(t). −ηf ρξ. σ(t).d ρσ.f ρξ.d 1 ρd. Q. see [Shr04]): χ1 (t) := ξ(t) Mf (t) Pf (t. for the following prices (for more discussion. T ) σ(t) dt + ηf Bf (t.
(2.f ηf Bf (t.σ ρξ. X(t)). T ) and χ2 (t) := ξ(t) .
2. such that: EQ max(ξ(T ) − K. Wσ (t).

to the domestic T-forward measure. T ) ξ(t) − 2 (dPd (t. T )dWf (t).11). o In order to determine the dynamics for FXT (t) in (2.18). T ) (2. like ﬁnite diﬀerences. we move from the spot measure. to the forward FX measure where the num´raire is the domestic zero-coupon bond. We know that Π(t) must be a martingale.20) Pf (t. It is therefore diﬃcult to solve it analytically and a numerical PDE discretization. T ) dξ(t) + dPf (t. T )) Pd (t. Md (t). generated by the money savings account in the domestic market.20). T ) Pd (t.d γηd σ + γ 2 σ 2 + ρξ. the discounting will be decoupled from taking the expectation. T )dWd (t) + ηf Bf (t.e. Md (t) Md (t) (2.2. Including this in (2. T ) 1 2 +ξ(t) 3 (dPd (t. Pd (t.19)
After substitution of SDEs (2.: Π(t) = Pd (t. T )) + (dξ(t)dPf (t.: E(dΠ(t)) = 0.1 The FX Model under the Forward Domestic Measure
To reduce the complexity of the pricing problem. X(t)).f γηf σ 2 ∂rf ∂rd ∂rf 2 ∂rd ∂σ∂rf √ ∂2V √ ∂2V √ ∂2V 1 ∂2V +ρσ. T )dξ(t)) − 2 dPd (t. 0 |Ft . which is useful for calibration. Pit06].σ γξσ ∂ξ∂σ 2 ∂ξ ∂rf ∂V ∂V ∂V ∂V +λd (θd (t) − rd ) + κ(¯ − σ) σ + (rd − rf )ξ + .21)
Q Q Q + σ(t)dWξ (t) − ηd Bd (t.16) into (2. T ) − ρξ. and ξ(t) stands for foreign exchange rate under the domestic spot measure.f ηf ξ σ + ρx. By switching from the domestic risk-neutral measure. T )ET max FXT (T ) − K. T ) Pf (t. ∂rd ∂σ ∂ξ ∂t This 4D PDE contains non-aﬃne terms. T ) Pd (t. T ) Pd (t. we apply Itˆ’s formula: dFXT (t) = Pf (t. QT .f ηf σ +ρξ. T )dPf (t. T ) .d ηd ξ σ ∂σ∂rd 2 ∂σ ∂ξ∂rf ∂ξ∂rd 2 2 √ ∂V 1 ∂ V ∂ V + ξ 2 σ 2 + λf (θf (t) − rf ) − ρξ.d FXT (t) σ(t) − ρd. T ) (2. Finding a numerical solution for this PDE is therefore rather expensive and not easily applicable for model calibration. In the next subsection we propose an approximation of the model. i. 2. The superscript should not be confused with the transpose notation used at other places in the text.f ηf Bf (t. we arrive at the following FX forward dynamics: dFXT (t) = ηd Bd (t. T ) Pd (t.17) gives the following Fokker-Planck forward equation for V : rd V = √ ∂2V ∂2V 1 2 ∂2V 1 2 ∂2V ηf 2 + ρd. Pd (t.17)
with V (t) := V (t. i.By Itˆ’s lemma we have: o dΠ(t) = 1 V (t) dV (t) − rd (t) dt. T ). needs to be employed.15) and (2. (2. As e indicated in [MR97. T ) − ξ(t) 2 dPd (t. T ) Pd (t.f ηd ηf + ηd 2 + ρσ.e. T ) Pf (t.18)
where FXT (t) represents the forward exchange rate under the T -forward measure. T ) ξ(t) Pf (t.
5
. Q. like square-roots and products. T ) dt (2. T ) ηd Bd (t. T ). (2. the forward is given by: FXT (t) = ξ(t) Pf (t.

T ) σ(t) dt + γ σ
T σ(t)dWσ (t).26). e In the model we incorporate a full matrix of correlations. T )dWf (t). the model in (2. T ) dt + ηd dWd (t).. like T T dWF (t)dWσ (t).d ηd Bd σ 1 √ T +2ρξ. need to be determined. dWξ (t). The dynamics of the forwards. Since the sum of three correlated. i. Note. Lemma 2.10). A change of measure from domestic-spot to domestic T-forward measure requires a change of num´raire from money savings account.d ηd Bd (t. Under the Tforward domestic measure. FXT (t) in (2.
(2. rd (t) and rf (t).11) is governed by the following dynamics: dFXT (t) = FXT (t) where dσ(t) = drd (t) = drf (t) = κ(¯ − σ(t)) + γρσ. see (2.Since FXT (t) is a martingale under the T -forward domestic measure. that this is only the case for models in which the diﬀusion coeﬃcient for the interest rate does not depend on the level of the interest rate.2 (The FX-HHW model dynamics under the QT measure).f ηd ηf Bd Bf 2 dWF . and with Bd (t.12). dWd (t) and T dWf (t). From the system in Lemma 2. T ) dt + ηf dWf (t).
(2.22) as: dFXT /FXT = √ 2 2 2 2 σ + ηd Bd + ηf Bf − 2ρξ. and are completely described by the appropriate diﬀusion coeﬃcients. Pd (t.Z σX σZ + 2ρY. Q = X + Y + Z.24)
2 T λd (θd (t) − rd (t)) + ηd Bd (t. in order to obtain the covariance terms.22)
(2. T ).26)
Although the representation in (2.2 we see that after moving from the domestic-spot Q-measure to the domestic T-forward QT measure. which implies that all processes will change their dynamics by changing the measure from spot to forward. It does not contain T T a drift term and only depends on dWd (t). In next section we derive the corresponding pricing PDE and provide model approximations. T )dWd (t) + ηf Bf (t. dWf (t).f Bd (t. The proof can be found in Appendix A. For clarity we therefore prefer to stay with the standard notation.f ηf Bf σ − 2ρd. T ). the forward exchange rate FXT (t) does not depend explicitly on the short-rate processes rd (t) or rf (t).22) or in (2.Z σY σZ .
λf (θf (t) − rf (t)) − ηf ρξ.2 provides the model dynamics under the domestic T-forward measure. dWσ (t). T )FXT (t) =: Pf (t. T ) given by (2.
we can represent the forward (2. the appropriate Brownian T T T motions under the T −forward domestic measure. Bf (t.
(2. one still needs to ﬁnd the appropriate cross-terms. Remark.25) with a full matrix of correlations given in (2. to zero-coupon bond Pd (t. This suggests that the short-rate variables will not enter the pricing PDE. Md (t).e.f
T σ(t) + ηd ηf ρd. T )ET (FXT (T )|Ft ) = Pd (t.23) (2. T T T σ(t)dWξ (t) − ηd Bd (t. QT . normally distributed random variables.26) reduces the number of Brownian motions in the dynamics for the FX.22).Y σX σY + 2ρX. T )ξ(t). do not depend explicitly on the interest rate processes. Remark. remains normal with the mean equal to the sum of the individual means and the variance equal to
2 2 2 2 σQ = σX + σY + σZ + 2ρX.
6
. Lemma 2.

d γηd σBd (t. X(T ).2. σ(t). Recently. This is also the approach followed here. 4κ
=
4κ¯ σ . 2 d d
By applying the Feynman-Kac theorem we can obtain the characteristic function of the forward FX rate dynamics. . In order to ﬁnd the forward ChF we take.d ηd Bd σ σ(t) dt + γ
T σ(t)dWσ (t). We assume: − = σ(t) ≈ E σ(t) =: ϕ(t). for (r0 .
is of the aﬃne form if: µ(X(t)) σ(X(t))σ(X(t))
T
= = =
a0 + a1 X(t). j = 1.3
Approximations and the Forward Characteristic Function
As the dynamics of the forward foreign exchange. ij
T r0 + r1 X(t).31)
to [DPS00] the n-dimensional system of SDEs: dX(t) = µ(X(t))dt + σ(X(t))dW(t).f ηf Bf ) σ(t) + ρd. t. σ(t)) = (ρx. The ﬁrst method is to project the non-aﬃne square-root terms on their ﬁrst moments.
with ﬁnal condition. . σ) − ∂σ 2 ∂x2 ∂x 2 T √ √ ∂ φ 1 ∂ 2 φT + ρx. c1 ) ∈ Rn×n × Rn×n×n . FXT (t). given by: dσ(t) = κ(¯ − σ(t)) + γρσ.
r(X(t))
for i. a signiﬁcant reduction of the pricing problem is achieved.28)
where we used the notation Bd := Bd (t. T ) + γ2σ . T ) = eiux (T ) .
(2. T ) + ρσ.f γηf σBf (t.30)
1 2 γ (1 − e−κt ). and ζ(t. T ).27)
with the variance process.f ηd ηf Bd Bf − 1 2 2 2 2 η B + ηf Bf . n. γ2
(t) =
4κσ(0)e−κt . The approximation of the non-aﬃne terms in the corresponding PDE is then done as follows. X(t). for arbitrary (c0 . the log-transform of the forward rate FXT (t). as usual.
(2.e. for any (a0 . k! Γ( 2 + k) k=0
∞
(2. a1 ) ∈ Rn × Rn×n . with r(X(t)) being an interest rate component. (2. T ) and Bf := Bf (t. The forward characteristic function: φT := φT (u.d ηd Bd − ρx.: xT (t) := log FXT (t). i. ∂x∂σ 2 ∂σ 2 √ This PDE contains however non-aﬃne σ-terms so that it is nontrivial to ﬁnd the solution. γ 2 (1 − e−κt )
(2. for which we obtain the following dynamics: dxT (t) = ζ(t. T. r1 ) ∈ R × Rn . 1 σ(t)) − σ(t) dt + 2
T T T σ(t)dWξ (t) − ηd Bd dWd (t) + ηf Bf dWf (t).
7
. we have proposed two methods for linearization of these non-aﬃne1 square-roots of the square root process [CIR85]. T ) σ + σ − ζ(t. under the domestic forward T T measure involve only the interest rate diﬀusions dWd (t) and dWf (t).σ γσ − ρσ.d γηd σBd (t.29) ∂φT ∂t
with the expectation of the square root of σ(t) given by: E and c(t) =
1 According
σ(t) =
2c(t)e−ω(t)/2
1+ 1 k Γ 2 +k ( (t)/2) . T ) = ET eiux
T T
(T )
Ft . (c0 )ij + (c1 )T X(t). . φT (u. is the solution of the following Kolmogorov backward partial diﬀerential equation: √ √ ∂φT 1 ∂ 2 φT ∂φT κ(¯ − σ) + ρσ. in [GO09]. .

One may prefer to use a proxy. With B(τ ) = iu. β2 and β3 can be determined by asymptotic equality with (2. φT . ϕ(τ )) B 2 (τ ) − B(τ ) σ + (−ρσ.σ γσ − Ψ2 ) φT (u.f equal to zero. τ ) are given by: B (τ ) = 0.Γ(k) is the gamma function deﬁned by:
∞
Γ(k) =
0
tk−1 e−t dt. Simpson’s quadrature rule. ϕ(t)) 2
∂φT ∂ 2 φT − 2 ∂x ∂x with: (2.σ γiu − d) . (2. a closed-form expression for A(τ ) would be available [GO09].
and ζ(t. With correlations ρσ. τ ) and C(τ ) := C(u. the complex-valued function C(τ ) is of the Heston-type. T ) = exp(A(u.σ iu − d . of the full-scale FX-HHW model by FX-HHW1.d ηd γϕ(s)Bd (s)iu σ
τ
+ρσ. γ 2 (1 − ge−dτ ) (2. t.d .29) the corresponding PDE is aﬃne in its coeﬃcients.32)
in which the constant coeﬃcients β1 .30) is a closed form expression. X(T ).σ Function A(τ ) is given by:
τ
κ − γρx. C(0) = 0 and A(0) = 0. and its solution reads: C(τ ) = with d = 1 − e−dτ (κ − ρx.f ηd ηf Bd (t.σ iu + d are given in (2. f }. for example.
. τ ) + B(u. τ )σ(t)).30) (see [GO09] for details). and Bi (τ ) = λ−1 e−λi τ − 1 for i = {d. for example.35)
∂ 2 φT 1 ∂ 2 φT + γ2σ .33)
(ρx.11).σ γB(τ )C(τ ) + γ 2 C 2 (τ )/2. and the functions A(τ ) := A(u. ρσ.33). B(τ ) := B(u.
A(τ ) =
0
κ¯ + ρσ. By the projection of σ(t) on its ﬁrst moment in (2. with ϕ(t) = E( σ(t)). its evaluation is rather expensive. and not with FX-HHW. It is clear that eﬃcient pricing with Fourier-based methods can be done with FX-HHW1. [Hes93]. E( σ(t)) ≈ β1 + β2 e−β3 t .d ηd γϕ(τ )Bd (τ ) + ρσ. T ) . τ ).34)
with C(s) in (2. X(t). ρx. ϕ(t)) = Ψ3 + ρd.f γηf ϕ(s)Bf (s)iu C(s)ds + (u2 + iu)
0
ζ(s.
Although the expectation in (2.σ γiu − κ)2 − γ 2 iu(iu − 1). We denote the approximation. where τ = T −t. T ) + ηf Bf (t. C (τ ) = −κC(τ ) + (B 2 (τ ) − B(τ ))/2 + ρx. τ )xT (t) + C(u. Projection of the non-aﬃne terms on their ﬁrst moments allows us to derive the corresponding forward characteristic function. which is then of the following form: φT (u. and reads: − ∂φT ∂t = (κ(¯ − σ) + Ψ1 ) σ + (ρx. T. T ) − 8
1 2
2 2 2 2 ηd Bd (t.d γηd ϕ(τ )Bd (τ )C(τ ) − ζ(τ. T )Bf (t. by means of linearization.d γηd ϕ(s)Bd (s) − ρσ. T ) = ET eiux
T
∂φT + ∂σ
1 σ − ζ(t. The initial i conditions are: B(0) = iu.f γηf ϕ(τ )Bf (τ )) B(τ )C(τ ). ϕ(s))ds. g =
The parameters κ. A (τ ) = κ¯ C(τ ) + ρσ. ∂x∂σ 2 ∂σ 2 = eiux
T
(T )
FT
(T )
.
(2. It is most convenient to solve A(τ ) numerically with. κ − γρx. γ.

d ηd Bd (t. in a domestic market.37)
where Q and Z indicate the domestic-spot and foreign-spot measures. the terms Bd (t.
(2.
When solving the pricing PDE for t → T . we assume that Sf (t) is given by the Heston stochastic volatility model: dSf (t)/Sf (t) = dω(t) = rf (t)dt+ κf (¯ − ω(t))dt+ γf ω
Z ω(t)dWSf (t).4
Pricing a Foreign Stock in the FX-HHW Model
Here.d γηd Bd (t. := (ρx.11) we need to determine the drifts for Sf (t) and its variance process. we ﬁnd o d ξ(t) Mfd (t)
S (t) ξ(t) Mfd (t) S (t) Q σ(t)dWξ (t) +
= ρξ. T ) − ρσ.f ηf Bf (t.36) and the domestic money-saving account Md (t) in (2. This can be eﬃciently done with the help of a fast pricing formula.The three terms. we focus our attention on pricing a foreign stock. Sf (t).38)
. With the foreign short-rate process. Md (t). with the ﬁrst representation. however. Such a stock is a tradable asset.Sf 9
σ(t) ω(t) dt +
Q ω(t)dWSf (t). Sf (t) has the following dynamics: dSf (t)/Sf (t) = rf (t)dt + =
Z ω(t)dWSf (t)
rf (t) − ρξ.36)
drf (t) = λf (θf (t) − rf (t)))dt+
where Z indicates the foreign-spot measure and the model parameters. Ψ2 . and by discounting with the domestic money savings account.
where σ(t) is the variance process of FX deﬁned in (2. Under the T-forward domestic FX measure.f γηf Bf (t.11). The case t → 0 is furthermore t→0 trivial. the projection of the non-aﬃne terms on their ﬁrst moments is expected to provide high accuracy. T )ϕ(t). T )) ϕ(t). in the PDE (2.11). ω(t). Sf (t). T )) ϕ(t). To make process ξ(t)Sf (t)/Md (t) a martingale we set:
Q Z dWSf (t) = dWSf − ρξ. It is worth mentioning that also an alternative approximation for the non-aﬃne terms σ(t) is available. can be expressed in domestic currency by multiplication with the FX. are as before.
2. λf . θf (t) and ηf . The foreign stock. This alternative approach guarantees that the ﬁrst two moments are exact. the model has to be calibrated in the foreign market to plain vanilla options. Before deriving the stock dynamics in domestic currency. respectively. Sf (t) from (2. In this article we stay. T ) tend to zero.Sf
σ(t)dt. := (ρσ. T ) − ρx.4)) needs to be a martingale. With this extension we can in principle price equity-FX-interest rate hybrid products. since σ(t) −→ E( σ(0)).Sf
σ(t) ω(t)dt +
Z ω(t)dWSf . so the price ξ(t)Sf (t)/Md (t) (with ξ(t) in (2. In Section 2. from foreign to domestic-spot. T ) and Bf (t. By applying Itˆ’s lemma to ξ(t)Sf (t)/Md (t). Ψ1 .5 we perform a numerical experiment to validate this. see [GO09].35) contain the function ϕ(t): Ψ1 Ψ2 Ψ3 := ρσ. and Ψ3 . under the domestic spot measure.
(2. Z ηf dWf (t). With an equity smile/skew present in the market.
(2. and all terms that contain the approximation vanish. κf . established in (2. γf . Z ω(t)dWω (t). rf (t). ξ(t).d γηd Bd (t. Under the change of measure.

18) with ξ(0) = 1. e) is set to 1. per Euro.5. For the second pricing method. −1.f −15% −15% 1 −40% 100% 30% 30% −15% 30% 100% 25% −15% 30% . deﬁned in (2. which means that the interest rate curves are modeled by ZCBs deﬁned by Pd (t = 0. The option prices resulting from both models are expressed in terms of the implied Black volatilities. We also deﬁne the experiments as in [Pit06]. Eﬃcient pricing of plain vanilla products is then done by means of the COS method [FO08]. T ) = exp(−0. .f 100% ρσ. (2. .f ρξ. −0.f −40% = ρd.1. We choose2 : κ = 0.12).1δn δn Ti . Furthermore. 0.5
Numerical Experiment for the FX-HHW Model
In this section we check the errors resulting from the various approximations of the FX-HHW1 model.05T ).42)
= {−1. ¯ The correlation structure. The diﬀerences between the volatilities are tabulated in Table 2.ω dWSf (t) + Z 1 − ρ2 f .d 1 ρd.5}.000 paths and 20Ti steps.
2.ω ρSf . We use the set-up from [Pit06]. however. λf = 5%.7%. We report a maximum error of about 0.1 in Appendix B. . which is recovered by the ChF with 500 Fourier cosine terms. The approximation FX-HHW1 appears to be highly accurate for the parameters considered.The variance process is correlated with the stock and by the Cholesky decomposition we ﬁnd: dω(t) = κf (¯ − ω(t))dt + γf ω
Z ω(t) ρSf . based on the approximations in the FX-HHW1 model in Section 2.f ρξ. and the strikes are computed by the formula: Kn (Ti ) = FXTi (0) exp 0. . For the FX-HHW model we compute a number of FX option prices with many expiries and strikes. σ = 0.5. .1. based on a Fourier cosine series expansion of the probability density function. . In the next subsection the calibration results to FX market data are presented.σ ρξ. λd = 1%. The Monte Carlo simulation for pricing exotic options is deﬁned in the domestic market. σ
10
. 25% 100% (2. with (2. This implies that the presence of the non-aﬃne terms is not complicating in this setting.35. γ = 0.
2 The
model parameters do not satisfy the Feller condition. do not matter since the foreign stock is assumed to be already calibrated to the foreign market data.3. γ 2 > 2κ¯ . This formula for the strikes is convenient. is given by:
1 ρξ. Those. without any approximations. strikes K4 (Ti ) with i = 1. 1.39) are governed by several non-aﬃne terms. using two diﬀerent pricing methods. 10 are equal to the forward FX rates for time Ti . σ(0) = 0. $. with expiries given by T1 . for the full-scale FX-HHW model. 1. ηf = 1.ω dWω (t) S Q ω(t)dWω (t).0.f ρξ. . T10 .39)
= κf (¯ − ω(t)) − ρSf .07% for the other options.5.3.35.σ 1 ρσ.02T ) and Pf (t = 0. 0.d ρξ. since for n = 4.ξ γf ω
ω(t) σ(t) dt + γf
Sf (t) in (2.5.40)
The initial spot FX rate (Dollar.d ρσ.1% volatility for at-the-money options with a maturity of 30 years and less than 0.2%. The strikes and maturities are presented in Table B.d ρσ.38) and ω(t) in (2. we have used the ChF.1. T ) = exp(−0.41)
(2. The ﬁrst method is the plain Monte Carlo method. .
and FXTi (0) as in (2. ηd = 0.0. with 50.

18 0.0008 0.3 the detailed calibration results are tabulated.0007 -0.4 1. between the FX-HHW and FX-HHW1 models.0009 -0.11
0.0005 -0.0005 -0.0012
K7 (Ti ) 0.0008 -0.13 0.2 1.
Calibration results 1Y Market 1Y FX−HHW Calibration results Calibration results
0.0003 0.0004 -0.0002 -0.13
10Y Market 10Y FX−HHW
0.0009 -0. Short-rate interest rate models can typically provide a satisfactory ﬁt to at-themoney interest rate products. In the simulation the reference market implied volatilities are taken from [Pit06] and are presented in Table B.21 0.0007 -0.2 0.24 0. is presented.0010 0.1: Comparison of implied volatilities from the market and the FX-HHW1 model for FX European call options for maturities of 1.0002 -0. The correlation structure is as in (2.0012
K5 (Ti ) 0.0012
K4 (Ti ) 0. the smile for the FX rate and the smiles in the domestic and foreign ﬁxed income markets.17
0.0003 -0.3 1.0003 -0.0008 -0.0003 -0.5
Figure 2.41).1
0. in which an interest rate smile is incorporated.17 0. This was sometimes improved by adding jumps to the model (Bates’ model).
3
Multi-Currency Model with Interest Rate Smile
In this section we discuss a second extension of the multi-currency model.0009 0.0015 0.09
0.0012
K6 (Ti ) 0.0012 -0.35.0009 0. The corresponding FX option prices and the standard deviations are tabulated in Table B.0004 0.0004 0. In the next section an extension of the framework.16
0.0001 -0.0003 -0.0010 -0.23
implied Black Volatility
20Y Market 20Y FX−HHW
0.1 0.0004 0.0009 -0.0002 0.0012
Table 2.0001 -0. For long maturities and for deep-in-the money options some discrepancy is present.19 0.11
0.0006 0.2 Strike 1.0011
K3 (Ti ) 0.6 0.15 1.0005 0.18 0.1 in Appendix B.15 0.1: Diﬀerences.0000 -0.0006 -0.0009 0.0001 -0.12 0.1 1.0009 0.0001 0.0004 -0. ξ(0) = 1.0001 0. This is a drawback of the short-rate interest rate models.16 0.14 0.0001 0. 10 and 20 years.0007 -0.0001 -0.8 1 1.6 0.2 in Appendix B. 2. so that interest rate smiles and skews can be modeled as well. In Appendix B in Table B.5 1 Strike 1. In the calibration routine the approximate model FX-HHW1 was applied. This hybrid model models two types of smiles. Our experiments show that the model can be well calibrated to the market data.12
implied Black Volatility implied Black Volatility
0.0002 0.0006 -0.5. Strike K4 (Ti ) is the at-the-money strike.0001 0.0003 -0.0009 -0.0003 -0. They are therefore not used for pricing derivatives that are sensitive to the interest rate skew.25 0.0009 -0.0010
K2 (Ti ) -0.08
0. in implied volatilities.22 0. This is however typical when dealing with the Heston model (not related to our approximation). In Figure 2.09 0.0003 -0.0012 0.0011 0.1 Calibration to Market Data
We discuss the calibration of the FX-HHW model to FX market data.0006 0. since the skew/smile pattern in FX does not ﬂatten for long maturities.0000 0. 11
.0005 0.4.0006 -0.4 Strike 1.0004 0.0001 -0.1 some of the calibration results are presented. The strikes are provided in Table B.5 0.0001 -0.0002 -0.0010 0.Ti 6m 1y 3y 5y 7y 10y 15y 20y 25y 30y
K1 (Ti ) -0.0008 -0.

T1 .j (t) k. .k (t) = σf.
12
. the stochastic volatility FX is of the Heston type.k (t) := Lf (t.e.j
N
µf (t) = −
j=k+1
τj φf. For t < Tk−1 we deﬁne the forward Libor rates Ld.j (t)σf.k (t) + (1 − βf.k (t) + (1 − βd.k Lf. Ld.
(3.T dLd. TN ≡ T } and τk = Tk − Tk−1 for k = 1 .T The Brownian motion. . T ) do not contain a drift term).3)
dvd (t) = and
λd (vd (0) − vd (t))dt + ηd
d. we also assume that the tenor structure for both currencies is the same.2)
For each currency we choose the DD-SV Libor Market Model from [AA02] for the interest rates.5)
where φd. Tk ) and Lf.
µd (t) = −
j=k+1
τj φd. e given by:
d.k )Lf. T ) and Pf (t.
S(0) > 0.k (t) := 1 τk Pd (t.k φf. Tk−1 . f.k (t) and σf. Tk ) (3..k (t) control the slope of the volatility smile.T vd (t)dWv (t). Pf (t. Pd (t. assuming independence between log-normalLibor rates and FX. Tk−1 ) −1 . Tk ) as Ld.k (t) = σd.k (t) and βf. = βf.We abbreviate the model by FX-HLMM. the parameters βd.k (t) vf (t) µf (t) vf (t)dt + dWk (t) . .
(3. 1 + τj Ld.k (t) vd (t) µd (t) vd (t)dt + dWk (t) .k (t) under the terminal foreign measure T .j (t) k.
f.k (0).k (0).1)
dσ(t) = κ(¯ − σ(t))dt+ γ σ
with the parameters as in (2. ηf determine the curvature of the interest rate smile.
d. and the Brownian motion. follows the following dynamics: dξ(t)/ξ(t) = (. N.j f ρ . .k )Ld. . Td ≡ Tf = {T0 . λf determine the speed of mean-reversion for the variance and inﬂuence the speed at which the interest rate volatility smile ﬂattens as the swaption expiry increases [Pit05]. Tk−1 ) −1 . dWk relates to the k-th foreign market Libor rate. In our approach we deﬁne a model with non-zero correlation between FX and interest rate processes.T vf (t)dWv (t).k (t). which under domestic risk-neutral measure.k (t) := 1 τk Pf (t.k (t) determine the level of the interest rate volatility smile. Tk−1 . . Parameters ηd . T )/Pd (t. . . 1 + τj Lf.j (t)σd.11). )dt+
Q σ(t)dWξ (t).k Ld. Q σ(t)dWσ (t).T dLf. A description of such FX hybrid products can be found in the handbook by Hunter [Hun05]. A ﬁrst attempt to model the FX by stochastic volatility and interest rates driven by a market model was proposed in [TT08]. Tk ) Lf. corresponds to the k-th domestic Libor rate.T under the T-forward domestic measure.j
(3. T ).k φf. under the T -forward measure generated by the num´raires Pd (t.j d ρ . For simplicity. It is especially interesting for FX products that are exposed to interest rate smiles. i.
(3. Lf. As in the previous sections. Q. dWk .k = βd. In the model σd. and λd . Since we consider the model under the forward measure the drift in the ﬁrst SDE does not need to be speciﬁed (the dynamics of domestic-forward FX ξ(t)Pf (t. σ(0) > 0. In the model we assume that the domestic and foreign currencies are independently calibrated to interest rate products available in their own markets.4)
dvf (t) = with
N
λf (vf (0) − vf (t))dt + ηf
f.k (t) := Ld (t.k φd.

T T f. i.T dWk (t)dWjf.T (t) = ρd. dWk (t)dWσ (t) = 0.T f. k. T ) (3. ξ.j ρf dt. Lf. k. dWσ (t)dWv (t) = 0.T f.T d.The following correlation structure is imposed. i.j
We prescribe a zero correlation between the remaining processes.T T f. T ) .T (t) = T dWξ (t)dWjf.T (t) = f.
T d. given by: FXT (t) = ξ(t) Pf (t. Note that the bonds are not yet speciﬁed. Throughout this article we assume that the DD-SV model in (3. dWk (t)dWv (t) = 0.f dt.6)
d.T dWk (t)dWv (t) = 0.T dWk (t)dWjf.j (t): Libors in domestic market: Libors in foreign market: Libors in domestic and foreign markets:
T T dWξ (t)dWσ (t) = T dWξ (t)dWjd.j (3.. ξ.σ dt.T (t) = d.T (t) =
ρξ.
The correlation structure is graphically displayed in Figure 3..e.T dWk (t)dWjd. 13
.j ρd dt. With this correlation structure. βk (t) ≡ βk and σk (t) ≡ σk .
T d. between FX and its variance process.T d.
all variance processes. dWξ (t)dWv (t) = 0.3) and (3.j (t): FX and foreign Libors. k.1: The correlation structure for the FX-HLMM model. σ(t): FX and domestic Libors. we derive the dynamics for the forward FX. Arrows indicate nonzero correlations. SV is Stochastic Volatility.
FX and the Libor variance processes.
Figure 3.7)
(see also (2.T dWξ (t)dWv (t) = 0. This means that approximate time-homogeneous parameters are used instead of the time-dependent parameters.
Libors and the FX variance process.
d. ρd dt.
d. between Libors and their variance process.T f.18)) with ξ(t) the spot exchange rate and Pd (t.j ρf dt.1. Pd (t.T dWσ (t)dWv (t) = 0.T f. T ) and Pf (t. dWv (t)dWv (t) = 0.T T T dWk (t)dWσ (t) = 0.e. Ld.4) is already in the eﬀective parameter framework as developed in [Pit05]. T ) zero-coupon bonds.

T ) or Pf (t. The dynamics for the zero-coupon bonds. the dynamics are determined under the appropriate T-forward measures (for Pd (t. DMP09. 14
.T (t). f }. T ) Pd (t. T ) 1 Pf (t.12)
− vf (t)
j=m(t)+1
τj σf.8). driven by the Libor dynamics in (3. [Sch02a. it is not necessary to determine the appropriate drift correction.j (t) dWjd. 1 + τj Lf.12) which is an indication for the change of measure from the foreign to the domestic measure for the foreign Libors. Pit04. Tj−1 .3) and (3. )dt −
vf (t)
j=m(t)+1
(3. T ) the domestic T-forward measure. . T ) and Pf (t.m(t) (Tm(t)−1 )
j=m(t)+1 N
(1 + τj Ld (t. .T (t) in (3. we need expressions for the zero-coupon bonds. Pd (t.11)
and the coeﬃcients were deﬁned in (3. Whereas the Libors Li. We consider here the linear interpolation scheme. T ) to Pd (t.8) are deﬁned to be equal to 1).7) we will not encounter any dt-terms (as FXT (t) has to be a martingale under the num´raire Pd (t. )dt −
vd (t)
j=m(t)+1 N
τj σd. T )
N
= (.10)
= (.2) the following expression for the ﬁnal bond can be obtained: 1 1 = Pi (t. . e For each zero-coupon bond.4). To deﬁne continuous time dynamics for a zero-coupon bond. T ) the foreign T-forward measure). .j φf. N and the bond Pi (t.7) and neglecting all the dt-terms we get dFXT (t) FXT (t)
N
=
T σ(t)dWξ (t) + N
vd (t)
j=m(t)+1
τj σd. Since FXT (t) in (3.9) with (3. which reads: 1 = 1 + (Tm(t) − t)Li.m(t) (Tm(t)−1 ). T ). 1 + τj Lf. T ) for the foreign bond.j (t)) . interpolation techniques are available (see. 1 + τj Ld. proposed in [Sch02a]. [GO10]. . .T (t). By combining (3. T ).j φd. T )). Tm(t) ) (3.j (t)
Note that the hat in W . and for Pf (t. BJ09]).9)
In our previous work.j φd. . T ) measure. Pi (t. The bond Pi (t. Tj )) .7).4). the bond Pi (t. Pd (t.j (t) τj σf. T ) Pf (t.T (t). for example. for Tm(t)−1 < t < Tm(t) .m(t) (Tm(t)−1 )
j=m(t)+1 T
(1 + τj Lf (t.
(3. Tm(t) ) is not yet well-deﬁned in the current framework.8)
with T = TN and m(t) = min(k : t ≤ Tk ) (empty products in (3.7) is a martingale under the Pd (t.
=
1 + (Tm(t) − t)Lf. T ) dPf (t.8) is fully determined by the Libor rates Li.j (t) dWjf.j φf.k (t) are deﬁned by System (3. we ﬁnd for the domestic and foreign bonds: 1 Pd (t.j (t) (3.20) for the general dynamics of (3. T ) Pi (t. only the e drift terms will change.When deriving the dynamics for (3. Tm(t) ).
When deriving the dynamics for FX (t) in (3. k = 1.4).
j=m(t)+1
for i = {d. TN ) in (3. By taking Equation (2. With Equation (3.T (t) 1 + τj Ld. are given by: dPd (t. Tj )) . disappeared from the Brownian motion dWjf.j (t) dWjf. this basic interpolation technique was very satisfactory for the calibration.j (t) dWjd. Tm(t) )
N
(1 + τj Li.3) and (3.j (t)
(3.3) and (3. .k (t). T )
N
=
1 + (Tm(t) − t)Ld. Tj−1 . By changing the num´raire from Pf (t.

j (t)) with φi.j := τj σd.j (0) for i = {d.13) and the correlations between the main underlying processes is not aﬃne.j (0) .1
Linearization and Forward Characteristic Function
The model in (3.k (t) and Lf.T (t) − vf (t) ψf. the log-transform gives 1 d log S(t) = − 2 σ 2 dt + σdW (t). So.j dWjd. a :=
T σ(t)dWξ (t).16). in the standard Black-Scholes analysis for dS(t) = σS(t)dW (t).. for the square diﬀusion coeﬃcient (a + b − c)2 = a2 + b2 + c2 + 2ab − 2ac − 2bc.j = βi. Since
N N
xj
j=1
3 In
2
=
j=1
x2 + j
i.12) with the stochastic variance in (3. xT (t) = log FXT (t). Lf.13)
The model given in (3.j dWjf. can be expressed as: 1 2 dxT (t) ≈ − (a + b − c) + 2 − vf (t)
A T σ(t)dWξ (t) +
vd (t)
A
ψd.j )Li.e.j (t) ≈ Ld.j (0).
dvi (t) = λi (vi (0) − vi (t))dt + ηi
with i = {d. which is standard practice (see for example [GZ99. T )-measure do not change3 and are given by: dσ(t) = κ(¯ − σ(t))dt + γ σ
T σ(t)dWσ (t).j Li.j (0) (3. i.
[GO10] the proof for this statement is given when a single yield curve is considered.j (t) ≈ Lf. .j (0) .j (0) ⇒ φf.j Lf.N i=j
xi xj .j := τj σf. . JR00]).j (0) ⇒ φd.14)
This approximation gives the following FXT (t)-dynamics: dFXT (t) T ψd.
with the coeﬃcients a. 1 + τj Ld.T vi (t)dWv (t). A = {m(t) + 1. for N > 0. 1 + τj Lf.
(3.T (t).16)
we ﬁnd. .: Ld.j (t)/(1 + τi.j dWjf. σ(t). (3..
(3.T (t). xT (t) = log FXT (t).j (0) ψf.j ≡ Ld. FXT (t) ≈ σ(t)dWξ (t) + vd (t) j∈A j∈A dσ(t) = κ(¯ − σ(t))dt + σ γ
T σ(t)dWσ (t).j ≡ Lf.j=1. With the notation.j dWjd.j (0). the dynamics under the Pd (t. In order to derive a characteristic function. i. for which we need to calculate the square of the diﬀusion coeﬃcients4 .k (t). as it contains terms like φi.j Ld.Since the stochastic volatility process. Ld.T (t). the dynamics for the log-forward. for FX is independent of the domestic and foreign Libors. N }.j dWjd.j (t) + (1 − βi. we freeze the Libor rates. b :=
vd (t)
j∈A
ψd.. c :=
vf (t)
j∈A
ψf.17)
ψf.
4 As
15
. b and c given in (3.
3. the correlations are given in (3.6) and ψd. HW00. f }.T (t) (3.j Li.15)
We derive the dynamics for the logarithmic transformation of FXT (t). f }.j dWjf.T (t). In the next section we discuss a linearization.12) is not of the aﬃne form..

j . E( vd (t)). vd (t) and vf (t) are independent CIR-type processes [CIR85]. (3.T (t). and (3. Function f (t) can be determined with the help of the formula in (2. ψf. j. 2 2 2 ∂vd 2 ∂vf 2 ∂σ 2 ∂x∂σ with the ﬁnal condition φT (u.j ψd.
vf (t)
dt
T + σ(t)dWξ (t) +
ψd. j.j∈A i=j 2 ψf. (3. T ). = vd (t)
j∈A 2 ψd. τ )σ(t) +Dd (u. vf (t) := (2ab − 2ac − 2bc)/dt.15). j.17) by: dxT (t) ≈ − 1 2 σ(t) + Ad (t)vd (t) + Af (t)vf (t) + f t.22)
16
.j dWjf. (3.30).j
(3. σ(t).j dWjd. we linearize the non-aﬃne terms in the drift in f (t. f t.e.20) linearizes all non-aﬃne terms in the corresponding PDE.x respectively.k ρd.x ψd.j + i.σ γσ . The correlation between the k-th domestic and j-th foreign Libor is ρd. i. E( σ(t)).k
with ρd . k. the forward characteristic function. vd (t). Ad and Af in (3. the solution is of the following form:
T
(T )
. As before. τ ) + B(u.18)
c2
= vf (t)
ψf. X(t). t.
The coeﬃcients ψd. so the expectation of their products equals the product of the expectations.21)
1 2 ∂ 2 φT 1 2 ∂ 2 φT 1 ∂ 2 φT ∂ 2 φT + ηd vd + ηf vf + γ2σ + ρx.20)
The variance processes σ(t).j∈A i=j
ψd. σ(t). we can express the dynamics for dxT (t) in (3. φT := φT (u. E( vf (t)) =: f (t). τ )xT (t) + C(u. The approximation in (3. vd (t).we ﬁnd: a2 b2 = σ(t)dt.j
j∈A k∈A
vd (t) vf (t)
ψf. X(T ).19) are deterministic and piecewise constant. τ )vf (t) .x
(3.T (t) −
vf (t)
A
ψf. vd (t)
A
σ(t). Since all coeﬃcients in this PDE are
φT (u.f . i. is deﬁned as the solution of the following backward PDE: 0 = ∂φT 1 + (σ + Ad (t)vd + Af (t)vf + f (t)) ∂t 2 +λd (vd (0) − vd ) ∂ 2 φT ∂φT − 2 ∂x ∂x
∂φT ∂φT ∂φT + λf (vf (0) − vf ) + κ(¯ − σ) σ ∂vd ∂vf ∂σ (3.x j.j ρd dt. ρf the correlation between the FX and j-th domestic and foreign Libor.j ρd dt =: vd (t)Ad (t)dt. vf (t) ≈ f t. σ(t)..j + j∈A i.j By setting f t.j ρf dt. T ) = eiux linear. vf (t)) by a projection on the ﬁrst moments.
vd (t).j . τ )vd (t) + Df (u. T ) = exp A(u.j ρf dt.18). j.i ψd.i ψf. In order to make the model aﬃne. i. X(t).f dt. =: vf (t)Af (t)dt. vd (t).19)
ab = ac = bc =
σ(t) vd (t)
j∈A
σ(t) vf (t)
j∈A
ψf. t.

20) and analytically known functions χk (u. Substitution of (3. . Df (u. the solution for C(τ ) is analogous to the solution for the ODE for the FX-HHW1 model in Equation (2. Af (t) the solution must be determined iteratively. τj ). . .
3.σ γiu) + dj − γ 2 C(u.
and χA (u. k
gj =
(κ − ρx. τj ) := λk − δk. C(τ ) := C(u.j ) ηd λf +vf (0) 2 (λf − δf. N . Ad (t) and Af (t) are from (3. . τ ) can be expressed as: Dd (u. τ ). τj ):
2 2 χk (u. Df (0) = 0 with Ad (t) and Af (t) from (3. and A(u.j sj ) (ηk (1 − k. τ ).j sj
)).j = (ρx. σ 0.j e−δf. 2 λk + δk. for j = 1. τj ) − 1 2 (u + u) 2
τj
f (s)ds. The resulting approximation of the full-scale FX-HLMM model is called FX-LMM1 here. with the interest rates driven by the market model.
where dj = δk. τj−1 ) + χf (u. 2 2 Af (t)(B 2 (τ ) − B(τ ))/2 − λf Df (τ ) + ηf Df (τ )/2. C(0) = 0.
τj−1
with f (s) in (3. τj−1 )
with sj = τj − τj−1 . B(0) = iu.18). and f (t) as in (3.σ γiu) − dj − γ 2 C(u. τj ) = Dd (u. τj−1 ) + χd (u.19). Dd (τ ) := Dd (u.
2 λ2 + ηk Ak (t)(u2 + iu). τj ). τj−1 ) + χA (u.19). . 2 2 Ad (t)(B 2 (τ ) − B(τ ))/2 − λd Dd (τ ) + ηd Dd (τ )/2. .σ γiu − κ)2 + γ 2 (iu + u2 ).j e−δd.j − ηk Dk (u. τ ): A (τ ) B (τ ) C (τ ) Dd (τ ) Df (τ ) = = = = = f (t)(B 2 (τ ) − B(τ ))/2 + λd vd (0)D1 (τ ) + λf vf (0)D2 (τ ) + κ¯ C(τ ).j sj ) (1 − d. τj−1 ) . τj−1 ) .22) in (3.
with initial conditions A(0) = 0.2
Foreign Stock in the FX-HLMM Framework
We also consider a foreign stock.20). f } and χA (u.33). τj ) = Df (u. . τj ) = A(u.18) and (3. τ ) and Df (τ ) := Df (u. Dd (0) = 0.with τ := T − t. τj−1 ) (1 − e−δk. . N . (B 2 (τ ) − B(τ ))/2 + (ρx.21) gives us the following system of ODEs for the functions A(τ ) := A(u. driven by the Heston stochastic volatility model.j )sj − 2 log (1 − d.j − ηk Dk (u. respectively. τ ). With B(τ ) = iu. τ ) and A(u. τj ) = κ¯ σ (κ − ρx. (κ − ρx. τ ). As the remaining ODEs involve the piecewise constant functions Ad (t). τj ).j
=
2 λk − δk. The stochastic processes
17
. B(τ ) := B(u.j ) ηf
.σ γB(τ ) − κ) C(τ ) + γ 2 C 2 (τ )/2. Sf (t). j = 1. for k = {d. Df (u. like for the pure Heston model with piecewise constant parameters in [AA00]. For a given grid 0 = τ0 < τ1 < · · · < τN = τ . the functions Dd (u.j e −δk. τj−1 )
k. (3.j )sj − 2 log (1 − f.σ γiu − dj )sj − 2 log (1 − gj e−dj sj ) (1 − gj ) γ2 λd +vd (0) 2 (λd − δd.j sj ) (1 − f.j − ηk Dk (u.

T )
FXT (t) = ξ(t)
Pf (t. In order to generate skew for FX. Pf (t.σ = −40%. T We move to the domestic-forward measure. for i. i. under the forward foreign measure.e. The last term involves all dt-terms. λd = 100%. σf. ρξ. ¯ (3. as frequently observed in ﬁxed income markets (see for example [BM07]). ω(t). γ = 0. . T ) Pd (t. for k = 1. which implies T that Sf (t)FXT (t) is a martingale.T (t). j = 1.25)
is therefore a tradable asset.k = 15%. foreign stock exchanged by a foreign exchange rate and denominated in the domestic zero-coupon bond is a tradable quantity. Pf (t = 0.5. we i.k = 50%.of the stock model are assumed to be of the same form as the FX (with one.T ω(t)dWSf (t) +
vf (t)
j=m(t)+1
τj σf. .j φf.05T ). foreign. and forward T foreign exchange rate. . 1 + τj Lf. We check. that. σ(0) = 0. one ﬁnds: o
T d Sf (t)FXT (t) T Sf (t)FXT (t) T dFXT (t) dSf (t) + T + Sf (t) FXT (t)
=
dFXT (t) FXT (t)
T dSf (t) T Sf (t)
. (3. . are deﬁned by
T Sf (t) =
Sf (t) .3
Numerical Experiments with the FX-HLMM Model
We here focus on the FX-HLMM model covered in Section 3 and consider the errors generated by the various approximations that led to the model FX-HLMM1.29)
In the correlation matrix a number of correlations need to be speciﬁed. T ) .k = 25%. given by:
T dSf (t) T Sf (t) N
=
f. By Itˆ’s lemma. for the FX stochastic volatility model we set: κ = 0. will enter the drift of the variance process dω(t) in (3. ρd = i. βf. interest rate curve) with the dynamics.. by a numerical experiment. σ = 0. . So.23).j (t) dWjf. by a change of measure.28) (3. T ) = exp(−0. we prescribe large positive values. N . is correlated with forward stock S T (t). FX (t). ηf = 20%.j 90%.
(3. . σd.3. as before. N (i = j). T ) Sf (t) ξ(t) = ξ(t).k = 95%.j (t) (3. .02T ). . T )
(3. We have performed basically two linearization steps to deﬁne FX-HLMM1: We have frozen the Libors at their initial values and projected the non-aﬃne covariance terms on a deterministic function. Pf (t. ρf = 70%.
3. T ) Pd (t.1. We have chosen the following interest rate curves Pd (t = 0. T ) = exp(−0.T ω(t)dWω (t).27)
In the simulation we have chosen the following parameters for the domestic and foreign markets: βd. Pd (t.k 18
.
Variance process.23)
dω(t) = κf (¯ − ω(t))dt + γf ω
f. and. The forward stock.1. Sf . The correlation between the FX and the domestic Libors is set as ρd = −15%. λf = 70%. T )
(3. σ(t). ηd = 10%.26) do not contribute to the drift. and the correlation between FX and the foreign ξ.26)
The two ﬁrst terms at the RHS of (3. the size of the errors of these approximations.j prescribe a negative correlation between FXT (t) and its stochastic volatility process.24)
The quantity
T Sf (t)FXT (t) =
Sf (t) Pf (t. For the correlations between the Libor rates in each market.

0000 K4 (Ti ) 0. . K7 (Ti ) are tabulated in Table B. 1
N ×N
Cdf
. .0005 -0. N (i = j).. . . .0018 0.0012 0.N ρd 1. ..
(3. .0025 0. .. . ..2 1 . .d Cξ.0014 0..30)
with the domestic Libor correlations given by
1 ρd 1. ..N ρf 2. . .N ρd 2. .0010 0.1 we present the diﬀerences. d ρξ.5%.... . 25% 25% 1 . .. .f . In Table 3. ρd 1.5 where the FX-HHW model was considered. .N . 1
N ×N
.0027 -0.0028 0.f ρ1..34)
Since in both markets the Libor rates are assumed to be independent of their variance processes. . . . .k ρd. The following block correlation matrix results: i.f .0010 0. between the FX-HLMM and FXLMM1 models. ρf 2. ..f ρ1.0014 0... 25% 25% .000 paths and 20 intermediate points between the dates Ti−1 and Ti for i = 1.f CT ξ. . .0000 0. 90% .2 = . .0044 -0. . 25% . ... ..0008 0. . N ). .N .d = ρd ξ..0010 -0.0021 -0. −15%
N ×1
. . .0014 -0. .0004 -0. j = 1.0039 0. . in implied Black volatilities.0046 -0. . as the maximum diﬀerence in terms of implied volatilities is about 0. . ....0005 0.. The prices and associated standard deviations are presented in Table B.0009 0. . 70% .0021 0.. ..0002 0. ...33)
and the vectors Cξ.1: Diﬀerences.2 = .0038 0. The correlation between the domestic and foreign Libors is ξ.0040 -0. −15% N ×1
(3.0040 -0.N ρd.f 1..f = 25% for i.N 1 70% = .f 1. .0019 0. = .0032 0.2 1 ..0021 K7 (Ti ) -0.0005 0.0032 0.f
ρf ξ.0034 -0.2 Cf = . .d and Cξ. given by:
Cξ.0025 -0. . .. ρf 1. 1 (3.0028 0. .f Cξ.0029 0.0018 K6 (Ti ) -0. 1
N ×N
..
(3. . f ρξ. . 90% 90% 1 .. .. . . . ρd. d ρ2.0011 0.0007 K3 (Ti ) 0.2 1 .0003 -0.. .Libors is ρf = −15%.N . . .N
−15% −15% . the prices for FX-HLMM1 were obtained by the Fourierbased COS method [FO08] with 500 Fourier series terms.0013 -0.0005 0.0018 0.0009 0.j Cd C = CT d.0016 0. . .d Cd..1. .0023 0. 1 1 90% = . 70% 70% .0000 0. Now we ﬁnd the prices of plain vanilla options on FX in (3. . .1 f ρξ. .0011 K2 (Ti ) 0.N ρd.
Ti 2y 3y 5y 7y 10y 15y 20y 25y 30y K1 (Ti ) 0.7). Cξ. 70% 70% 1 .0025 0. .f Cf CT ξ. in terms of the implied volatilities between the models FX-HLMM and FX-HLMM1.. . . ρd.. . While the prices for the FX-HLMM were obtained by Monte-Carlo simulation (20. .0018 -0.1 ρd ξ.N .0021 0.0034 -0. .. . we can neglect these correlations here.
(3.N = −15% −15% .0029 -0. .5..N . .2% − 0.0025 0. . 1 1 25% = .f 2..0014 0.f 2. 90% 90% . . d.0005 -0. 1 f ρ1..0004 0.0002 0.2 . .0017 0.N . .0006 0. The FX-HLMM1 model performs very well. .. The corresponding strikes K1 (Ti ).. .0030 0. f ρ1. . The simulation is performed in the same spirit as in Section 2.0009 K5 (Ti ) 0. d ρ1.32)
the correlation between Libors from the domestic and foreign markets given by:
1 d. 19
.31)
the foreign Libors correlations given by:
1 ρf 1. .0024
Table 3.2 Cd = . .

0018) 0.k control the slope of the interest rate volatility smiles.0024) 0. the parameters βd. i.0017) 0.1% − 0.: Ld.5 βd = 0.3102 (0.3.3058 (0.0001 1.0024) 0.3008 (0. For hybrid products sensitive to the interest rate skew a second model was presented in which the interest rates were driven by the stochastic volatility Libor Market Model. Both FX models were based on the Heston FX model and diﬀer with respect to the interest rate processes.3198 (0. for the parameters studied.k )Ld.3151 (0.0015) 0.0016) 0.3030 0.3721 1.k (0) is equivalent to setting βd.36) in the approximation FX-HLMM1.k (0).0016) 0.3146 (0.0017) 0.2931 (0.3052 (0.3148 (0. In Table 3.3097 (0.42)
0.2943 (0.k Lf.k − and βf.3149 (0.6071
Table 3.2982 (0.3020 (0.3199 (0.2 the implied volatilities for the FX European call options for FXHLMM and FX-HLMM1 are presented.0018) 0.k −values on the FX implied volatilities.k (t) and Lf.0017) 0. φf.2977 (0. implying that the approximate model.3104 (0.3.3. The experiment indicates that there is only a small impact of the diﬀerent βd.3011 (0.5 βd = 0 βd = 0.k (t) ≡ Ld.k = 0 in (3. The experiments are performed for diﬀerent combinations of the interest rate skew parameters.6224 0. This model enables pricing of FX-interest rate hybrid products that are not exposed to the smile in the ﬁxed income markets. Freezing the Libors to Ld.0020) 0. With βd.3112 0. By a Monte Carlo simulation.3096 (0.3143 (0. By freezing the Libors.k (0) we have φd.0021) 0.k (0) and Lf.0017) 0.0015) 0.0016) 0.2977 (0.2941 (0.0026) 0.k (t) + (1 − βd.2930
strike (2.k on the FX.0017) 0.3053 (0.2973 (0.2: Implied volatilities of the FX options from the FX-HLMM and FX-HLMM1 models.k (0) and Lf.36)
In the DD-SV models for the Libor rates Ld.1
Sensitivity to the Interest Rate Skew
Approximation FX-HLMM1 was based on freezing the Libor rates.0019) 0.3069 0.k = 0.
FX-HLMM (Monte Carlo simulation) βf = 0.2960 0.35) and (3.k (0) = Ld.0017) 0.0018) FX-HLMM1 (Fourier) βd = 0 βf = 0 0. βd and βf .7290 0.0018) 0.8538 1.3198 (0.3015 (0.k = βf.k = 0 and βf.k (t) + (1 − βf.k .0017) 0. which is considered highly satisfactory.3191 (0.35) (3.3156 0.k (t) = βd.0016) 0. T = 10 and parameters were as in Section 3. 20
.15%. βf.k = 0 the implied volatilities obtained by the FX-HLMM model appear to be somewhat higher than those obtained by FX-HLMM1.k (0).k (t) ≡ Lf.0022) 0.k Ld. (3.0020) 0.3016 (0.0015) 0.k and βf.0017) 0.k (0) = Lf.2968 (0.0028) 0.
4
Conclusion
In this article we have presented two FX models with stochastic volatility and correlated stochastic interest rates. a diﬀerence of approximately 0.k = 0 and βf.2948 (0.0015) 0. In the ﬁrst model we considered a model in which the domestic and foreign interest rates were driven by single factor Hull-White short-rate processes.3061 (0. we obtain the FX implied volatilities from the full scale FX-HLMM model for diﬀerent values of β and by comparing them to those from FX-HLMM1 iwith β = 0 we check the inﬂuence of the parameters βd.k )Lf.0017) 0. FX-HLMM1 with βd.1714 1.2993 0.k (t) = βf.0022) 0. is useful for the interest rate modelling.0020) 0.5 βd = 1 βf = 0 βf = 1 0.3101 (0. The numbers in parentheses correspond to the standard deviations (the experiment was performed 20 times with 20T time steps).3196 (0.3056 (0.2938 (0.0021) 0.k (t) for any k.0021) 0.e.0015) 0.

[AAA08] A.B. [GKR96] H. An Equity-Interest Rate Hybrid Model With Stochastic Volatility and Interest Rate Smile.W. 2009. Presentation at Global Derivatives Conference 2006. 1999. [GZ99] P.C. Techn. Grzelak. Sci. second edition. Closed Form Pricing of FX Options Under Stochastic Rates and Volatility. Techn. The approximate models are based on the linearization of the non-aﬃne terms in the corresponding pricing PDE. M.Theory and Practice: With Smile.. 1(60): 77–105. [BM07] D. [CIR85] J.E. App.H. SIAM J. V. Transform Analysis and Asset Pricing for Aﬃne Jump-Diﬀusions. Presentation at Frankfurt MathFinance Conference. IJTAF. These pricing formulas form the basis for highly eﬃcient model calibration strategies. Jarrow.
References
[AA00] L. Grzelak. Presentation at MathFinance Workshop: Derivatives and Risk Management in Theory and Practice. Geman. [GO09] L. C. J. Heath. 2000. A Novel Pricing Method for European Options Based on Fourier-Cosine Series Expansions.A. 2006. Appl. 2009. [FO08] F. Interest Rate Models.For both hybrid models we have developed approximate models for the pricing of European options on the FX. Ingersoll. A. 68: 1343–1376.
21
. Econometrica. Markovian Projection to a Displaced Volatility Heston Model. Fang. Antonov. [AM06] A. Delft Univ. Andersen. Econometrica. Singleton.G. 2007. Beveridge. Oosterlee. Misirpashaev. SSRN working paper. Inﬂation and Credit. SSRN working paper.A. SSRN working paper. 1996. Brigo. Audet. 2010.A. Haastrecht. 2009. Techn. 1(7): 1–32. [DMP09] M. On the Valuation of Power-Reverse Duals and Equity-Rates Hybrids. Giese. J. SSRN working paper. Pelsser. 12(7): 969–1005. N. S. J. [HP09] A. [GO10] L. Ross. K. 2002.v. 1985. J. Comp. Risk.W. El Karoui. Inﬂation and Stock Options Under Stochastic Interest Rates and Stochastic Volatility. Andreasen. Mataix-Pastor. Econometrica. On the Heston Model with Stochastic Interest Rates.C. Pan.B. [And07] J. X. Glasserman. Changes of Num´raire.W. Generic Pricing of FX. Oosterlee. Techn. Arneguy. These models can also be used to obtain an initial guess when the full-scale models are used. Rochet. Andreasen. 2004. Andreasen.. Fast Greeks by Simulation in Forward LIBOR Models. Cox. 9-11 May 2006. T. Paris. SSRN working paper. Joshi. Prob. [BJ09] C. [DPS00] D. Acknowledgments The authors would like to thank to Sacha van Weeren and Natalia Borovykh from Rabobank International for fruitful discussions and helpful comments. 2000. C. 2008. The approximate models perform very well in the world of foreign exchange. Comput. Finance. Andersen. Springer Finance. Morton. Delft Univ. Fin. C. On the Pricing of Auto-Callable Equity Securities in the Presence of Stochastic Volatility and Stochastic Interest Rates. Duffie. 1992. in a very similar way as in our previous article [GO09] on equity-interest rate options. the Netherlands. F. Changes of e Probability Measures and Pricing of Options. R. [Gie04] A. J.G. Eﬃcient Calibration to FX Options by Markovian Projection in Cross-Currency LIBOR Market Models. Davis. A Theory of the Term Structure of Interest Rates..A. Bond Pricing and the Term Structure of Interest Rates: A New Methodology for Contingent Claims Valuation. SSRN working paper. 2009. Arbitrage-Free Interpolation of the Swap Curve. Report 09-05.. 2008. Caps. J. 31: 826-848. Interpolation Schemes in the Displaced-Diﬀusion Libor Market Model and the Eﬃcient Pricing and Greeks for Callable Range Accruals.. 15(12): 163–168. Frankfurt. Zhao. 3(1): 5–39. Volatility Skews and Extensions of the Libor Market Model. Math. the Netherlands. J. Volatile Volatilities. N. Report 10-01. 32: 443–548. Antonov.A. [Cap07] O. M. Mercurio. 26-27 May 2007. [HJM92] D. A. [AA02] L. Derivatives and Risk Management in Theory and Practice. Oosterlee. Frankfurt. 53(2): 385–407.

6(2): 327–343. J. Takehara. Schlogl. Wilmott. Comp. An Asymptotic FX Option Formula in the Cross Currency Libor Market Model. Arbitrage-Free Interpolation in Models of a Market Observable Interest Rates. Derivatives. Shreve. 6(4): 41–60. Risk. Springer Finance. Hull. editors. 2004. Studies. 7(2): 107–144. [Pit06] V. 2002. 2004. Jackel.85. 2002. Forward Rate Volatilities. 19: 66-71. Sandmann and P. Advances in Finance and Stochastic: o Essays in Honour of Dieter Sondermann. A. 2002. [HW00] J. Fixed Income. New York: Springer. Rebonato. ¨ [JR00] P. Springer. 6(2): 173-196. 10(2): 46–62. Swap Rate Volatilities and the Implementation of the Libor Market Model. Kawai. working paper no. Piterbarg. 15(11): 531-533. Time to Smile. Center for Analytica Finance Aarhus School of Business. S. Sch¨nbucher. Risk. 74-84. The Euromoney Derivatives and Risk Management Handbook. Rev. R. Heidelberg. White.V. Piterbarg. Springer Verlag. 3(3): 26– 36.V. Hull. IJTAF.
22
. K. 1993. Risk. Piterbarg. 2001. Rutkowski. [HW96] J. 18(5): 71-75. Takahashi. A Closed-form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options. 2000. A Multicurrency Extension of the Lognormal Interest Rate Market Models. P. Comp. Heston. Ohkoshi. 197–218. Smiling Hybrids. 2005. Stochastic Calculus for Finance II: Continuos-Time Models. Hybrid Derivatives. In K. ¨ [Sch02a] E. [SO02] J. [MR97] M. [Pit05] V. 2005. J. M. J.V. Fin. Musiela. 1997. Hunter. Computing Deltas of Callable Libor Exotics in Forward Libor Models. Sippel. [Hun05] C. Using Hull-White Interest Rate Trees. Linking Caplet and Swaption Volatilities in a BGM Framework: Approximate Solutions. Martingale Methods in Financial Modelling. [Shr04] S. [Pit04] V.[Hes93] S. 2006. Schlogl. White. Jackel.. [Mik01] P. ¨ [Sch02b] E. A. ¨ [KJ07] A. Finance and Stochastics. [TT08] A. Mikkelsen. March 2007. 1996.. J. ﬁrst edition. 2008. 2000. All Power to PRDC notes. Fourier Transform Method with an Asymptotic Expansion Approach: an Application to Currency Options. 11(4): 381–401. Cross-Currency LIBOR Market Models. Fin. Fin.

e.d 0 H2. T )ζ1 (t)dt + dWQ (t). 0 H4. from dX(t) = [dFXT (t)/FXT (t). dWξ (t) T
1 2
T 2 2 Br (s. T ) with ζk (t) being the k’th row vector resulting from multiplying the matrices A and H. the Brownian motion under the T -forward measure is given by: T T T T dWQ (t) = dWd (t) + ηd Bd (t.3)
The representation presented above seems to be favorable. T )ζ1 (t)dWQ (t).2 H3. dWf (t).d H4.2 0 0 H3.1 H= H3.: y Q ΛT (t) = Q dQT Pd (t. i. can be written as: dX∗ (t) = µ(X∗ )dt + AHdWQ (t). dWσ (t). T Since the vector ζ1 (t) is of scalar form. that for the 1D Hull-White short rate processes ζ1 (t) = ηd .2 H4. i.4) we get: o dΛT Q ΛT Q = Bd (t. dWf (t). drd (t). rd (t).1 H4. (A.1) Q dσ 0 0 γ σ √ dWσ 0 Q −ηd Bd ηf Bf 0 σ dFXT /FXT dWξ which. and for the domestic ZCB: dPd (t. T )ζ1 (s)ds + 0 0
T
Br (s. Note. The matrix model representation in terms of orthogonal Brownian motions results in the following dynamics for the domestic short rate rd (t) under measure Q: drd (t) = λd (θd (t) − rd (t))dt + ζ1 (t)dWQ (t).
T dWT (t) = −Bd (t. T )Md (t) (A. dFXT (t)/FXT (t)]T and express the model in terms of the independent Brownian motions dWQ (t) = [dWd (t). dσ(t). Pd (t. T )dt. equivalently.6)
.4)
By calculating the Itˆ derivative of Equation (A. dσ(t).2)
where µ(X∗ ) represents the drift for system dX∗ (t) and H is the Cholesky lowertriangular matrix of the following form:
1 H2. T ) = rd (t)dt + Bd (t. T ) = . drf (t)]T to dX∗ (t) = [drd (t). Now. ΛT (t). is driven by one source of uncertainty Q (only one Brownian motion dWd (t)).
23
.5)
which implies that the Girsanov kernel for the transition from Q to QT is given by Bd (t. dQ Pd (0. 0 .4 (A.A
Proof of Lemma 2.3 0 0 . T )ζ1 (t) which is the T -bond volatility given by ηd Bd (t. dWξ (t)]T .2
Since the domestic short rate process.4 ρξ.: ΛT = exp − Q So. 0. T )ζ1 (t)dWQ (t). 0. T ). T )ζ1 (s)dWQ (s) .d = 0 ρσ.3 H4.
(A. (A.2 0 0 H3.3 0 1 0 ∆ ρf. (A. drf (t).: Q dWd drd ηd 0 0 0 Q drf 0 ηf 0 0 dWf H = µ(X∗ )dt + √ .1 0 H2. it is convenient to change the order of the state variables. since the short-rate process rd (t) can be considered independently of the other processes. we derive the Radon-Nikod´m derivative [GKR96].2 H3.2 H4. dWσ (t).3 H4.e.

1268 0.4531 1.1344 0.1643 0.3868 0.1: Expiries and strikes of FX options used in the FX-HHW model.1429 0.1756 0.0001 0.0390 0.6224 0.f
T σ(t) + ηd ηf ρd.0982 0.f Bd (t.
24
.1948
Table B.
Ti 6m 1y 3y 5y 7y 10y 15y 20y 25y 30y K1 (Ti ) 1.7)
Returning to the dependent Brownian motions under the T-forward measure.1465 0.3 dWξ T T T T ρξ.7290 0.7409 0.1750 0.35.2113
K5 (Ti ) 0.2837 1. Strikes Kn (Ti ) were calculated as given in (2.1620 1.1223 0.2217
K4 (Ti ) 0.8399 0.1294 0.0966 0.3773 1.6274 1.0834 0.8309 0.2 dWf + H3.1980 0.0868 0.9265 0.42) with ξ(0) = 1. The strikes Kn (Ti ) were tabulated in Table B.2035
K6 (Ti ) 0.2414 K2 (Ti ) 1.3788 0.3454 1.1981
K7 (Ti ) 0.1135 0.2) we get that: T ηd Bd + dWd dt T T ρd.2680 1.f dWd + H2.4815 0. (A.3500 1.Now.1123 0.0376 0.1141 0.0865 0.3778 1.0883 0.5924 0.2462 1.9515 0.9492 K7 (Ti ) 1.f ηd Bd dt+ ρd.3174 K3 (Ti ) 1.7093 0.d ηd Bd dt+ ρσ.
λf (θf (t) − rf (t)) − ηf ρξ.2119 0.4491 1.1854 1.0902 0.1646 0.2391 1.1099 0.1627 0.2348
K3 (Ti ) 0.9587 0.1184 0.2491 1.1098 0.d ηd Bd (t.1529 0.3721 1.1714 1.5999 1.0859 0. T ) dt + ηf dWf (t).
B
Tables
In this appendix we present tables with details for the numerical experiments.1038 0.1968 0.6071 1.2397 0. T ) σ(t) dt + γ σ
T σ(t)dWσ (t).3299 1.2482
Table B.
Ti 6m 1y 3y 5y 7y 10y 15y 20y 25y 30y
K1 (Ti ) 0. from the vector representation (A.5389 1.1684 0.d ηd Bd dt+ ρξ.4671 1.2093 0.4273 1. T ) dt + ηd dWd (t).0895 0.0878 0.2338 1.0943 1.8538 0.7358 0.1837 0.2: Market implied Black volatilities for FX options as given in [Pit06].12).7218 K6 (Ti ) 1.1334 0.2 dWf + H4.1.4966 0. T )dWf (t).2231 0. gives us: dFXT (t) FXT (t) =
T T T σ(t)dWξ (t) − ηd Bd (t.
dσ(t) = drd (t) = drf (t) =
κ(¯ − σ(t)) + γρσ.1844 0.1802 0.0927 0.0514 0.3 dWξ + H4.1143 0.2296 0.3101 1.1913 0.5221 1.2092 0.4257 1.0447 0.0872 0.0876 0.d dWd + H4.4787 1.2994 1.
2 T λd (θd (t) − rd (t)) + ηd Bd (t.d dWd + H3.0859 0.0952 0.0943 0.5844 0.3012 0.1315 1. T )dWd (t) + ηf Bf (t.8188 0.6250 1.1587 1.0866 0.1012 0.4 dWσ
.0937 0.1218 0.0846 0.1479 0.6377 0.1107 0.4737 0.5489 K5 (Ti ) 1.4479 1.0871 0.1276 0.1049 0.1961 1.0836 0.2509
K2 (Ti ) 0.2 dWf Q HdW = T T T ρσ.
with full matrix of correlations given in (2.0989 0.9291 0.4174 K4 (Ti ) 1.1895 0.8608 0.1429 0.

0013 0.3109 0.0165
K2 (Ti ) -0.0048 0.0018 0.3560 0.0022 -0.0786 0.3553 0.1264 0.1637 0.0037 0.0025 -0.0010 0.4310 0.0036 0.1987 0.0063 -0.2349 0.0037 0.0012 0.0746 0.1816 0.0031 0.0005 0.0012 -0.1529 0.0009 0.1382 0.0047 0.0018 -0.1926 0.3202 0. MC stands for Monte Carlo and COS for Fourier Cosine expansion technique ([FO08]) for the FX-HHW1 model with 500 expansion terms.0015 0.0015 0.0005 0.0007 0.0050 0.0007 0.0007 0.0585 0.2188
K3 (Ti ) 0.0014 0.0016 0.0094 0.0007 0.1986 0.0046 0.0015 0.0006 0.1545 0.1754 0.0045 0.0005 0.0002 0.0004 -0.0785 0.0002 -0.1883
K5 (Ti ) 0.0007 0.2096 0.1987 0.2349 0.0008 -0.1907 0.4212 0.1966 0.3423 0.3871 0.0037 0.0007 0.1538 0.1672 0.3205 0.1147 0.0012 -0.0014 -0.1759 0.2567 0.0008 -0.4311 0.1720 0.0036 0.0016 -0.1757 0.2804 0.0003 -0.0004 0.2426 0.2967 0.3: The calibration results for the FX-HHW model.0039 0.2106 0.0008 -0.3362 0.1543 0.0004 0.3281 0.0002 -0.2) and FX-HHW model implied volatilities.2713 0.0018 0.
25
. in terms of the diﬀerences between the market (given in Table B.2626 0.0007 -0.0048 0.1702 0.0036 0.0002 -0.0009 -0.1008 0.3873 0.2210 0.3706 0.2971 0.2425 0.1539
K7 (Ti ) 0.1870 0.0005 0.2211 0.2713 0.3383 0.2809 0.1359
1y
3y
5y
7y
10y
15y
20y
25y
30y
Table B.0033 0.0016 0.0048 0.0033 0.0014 -0.1.2429 0.0013 0.0007 0.0001 -0.4: Average FX call option prices obtained by the FX-HHW model with 20 Monte-Carlo simulations.
Ti 6m
method MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS
K1 (Ti ) 0.1.1432 0.0051 -0.3768 0.0011 0.0033 0.0014 0. 50.0004 0.1148 0.2246 0.1587 0.3203 0.0001 0.0014 0.0035 0.2630 0.3104 0.1990 0.2209 0.0001 -0.3900 0.0007 -0.0074
K6 (Ti ) 0.0022 0.2834 0.0018 0.0009 -0.0037 0.0017 0.0043 -0.0050 0.2322 0.0020 0.2805 0.0020 0.3358 0.0007 0.Ti 6m 1y 3y 5y 7y 10y 15y 20y 25y 30y
K1 (Ti ) 0.1822 0.0059 0.0038 0.0018 0.1265 0.2319
K2 (Ti ) 0.2046 0.0017 -0.1526 0.2521 0.0015 0.1870 0.0002 0.0000
K4 (Ti ) -0.2566 0.0036 0.0064 -0.0014 -0.3873 0.2215 0.2548 0.3420 0.0934 0.1553 0.0004 0.0030 0.1714
K6 (Ti ) 0.0074
Table B.2425 0.0049 0.1714 0.0021 0.1636 0.3765 0.2893 0.3279 0.0002 -0.0036 0.2838 0.0018 0.2895 0.0024 -0.2191 0.0004 0.1008 0.0037 0.0583 0.2249 0.0004 0.2254 0.2455 0.1441 0.1923 0.0037 0.0015 0.1677 0.2092 0.0020 0.2042
K4 (Ti ) 0.0048 0.0018 0.0010 0.0004 0.0050 0.0018 0.1744 0.3709 0.4362 0.0050 0.2100 0.0016 0.0003 -0.0037 0.2528 0.0082
K7 (Ti ) 0.1888 0.4216 0.1382 0.0006 -0.2833 0. Strikes Kn (Ti ) are given in Table B.2814 0.1590 0.2260 0.0021 0.0070
K3 (Ti ) -0.0143 0.000 paths and 20 × Ti steps.0048 0.0003 -0.0033 0.1200 0.1960 0.0021 0.0006 -0.1367 0.1554 0.3195 0.2826 0.3878 0.3894 0.1207 0.2553 0.0023 -0.0748 0.0011 0.2463 0.0935 0.0048
K5 (Ti ) -0.0019 0. The strikes Kn (Ti ) are tabulated in Table B.0019 -0.0033 0.1908 0.0048 0.3380 0.4368 0.

2715 0.1838 0.3424 0.3944 0.2011 0.0016 0.1548 0.1548
3y
5y
7y
10y
15y
20y
25y
30y
Table B.0013 0.0015 0.2453 0.1212 0.1710
K7 (Ti ) 0.2186 0.0016 0.0026 0.0015 0.1213 0.2938 0.0021 0.2450 0.0023 0.0011 0.4399 0.4222 0.0019 0.2880 0.3215 0.1697 0.0029 0.3014 0.3928 0.1667 0.2727 0.0018 0.1664 0.0007 0.0016 0.0018 0.1848 0.0011 0.4363 0.3768 0.0030 0.2190 0.3280 0.1037 0.3613 0.1534 0.0009 0.2889 0. Values of the strikes Kn (Ti ) are tabulated in Table B.1866
K6 (Ti ) 0.2715 0.2052 0.2085 0.1.0013 0.1887 0.0016 0.2107 0.0016 0.1843 0.0016 0.3905 0.2354 0.1939 0.2335 0.0012 0.0010 0.3416 0.3031 0.2366 0.0014 0.0015 0.2883 0.2629 0.0013 0.0014 0.1330 0.3336 0.2321 0.2279
K3 (Ti ) 0.1913 0.0027 0.2808 0.1699 0.3632 0.Ti 2y
method MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS MC std dev COS
K1 (Ti ) 0.2031 0.2647 0.3282 0.2525 0.1931 0.1327 0.3893 0.3176 0.1457 0.0028 0.2587 0.0008 0.3171 0.3265 0.2550 0.2265 0.2907 0.2918 0.2127 0.0023 0.1457 0.0015 0.1030 0.0012 0.2907 0.2274 0.2918 0.0027 0.4243 0.3299 0.1858 0.0019 0.0012 0.0033 0.3402 0.3417 0.0024 0.2162 0.3461 0.0013 0.0010 0.3738 0.0016 0.0010 0.2152 0.2743 0.2017 0.0019 0.0012 0.3964 0.0018 0.2367 0.4338 0.2157 0.2545 0.2342 0.0008 0.0011 0.0025 0.1667 0.1553 0.0010 0.1621 0.2152
K4 (Ti ) 0.0012 0.2886 0.2576 0.2456 0.3234 0.0014 0.2043 0.5: Average FX call option prices obtained by the FX-HLMM model with 20 Monte-Carlo simulations. MC stands for Monte Carlo and COS for the Fourier Cosine expansion technique ([FO08]) for the FX-HLMM1 model with 500 expansion terms.4379 0.2014
K5 (Ti ) 0.0009 0.3326 0.2532 0.0006 0.2396 0.2917 0.0035 0.2281 0.2046 0.2393
K2 (Ti ) 0.1764 0.0012 0.000 paths and 20 × Ti steps. 50.0023 0.0019 0.3717 0.
26
.3914 0.1631 0.1767 0.0008 0.1806 0.2823 0.3786 0.1768 0.3482 0.2470 0.0014 0.0011 0.0021 0.0021 0.1834 0.