Sunteți pe pagina 1din 44

Quanto pricing with Copulas

Michael N. Bennett, Joanne E. Kennedy


University of Warwick

May 29, 2003

Abstract
We study the practical problem of pricing a particular multi-asset
option, a quanto FX option. The Black model, which corresponds to a
jointly lognormal distribution of asset prices at expiry, is inconsistent with
the implied volatility smile for each of the three relevant currency pairs.
We demonstrate a practical methodology for constructing a model for
the joint distribution that is calibrated to all relevant implied volatilities.
The margins of this distribution are determined separately in an initial
stage. To calibrate the joint distribution to the implied volatility smile on
the remaining FX rate, we perturb the dependence structure associated
with the Black model (the Normal copula) in order to influence the tail
dependence characteristics of the resulting joint distribution.
We calibrate our model to a number of real-life scenarios corresponding
to several maturities and currency set-ups. We find that a well-known ad-
hoc adjustment to the Black pricing formula often gives lower quanto call
prices than those calculated under our transformed copula model. The
relative difference in quanto prices with strikes furthest away from at-the-
money is occasionally large (10-15%).

1
1 Introduction
The price of any European multi-asset option may be written as an integral
involving the joint density of the asset prices at expiry. Information regarding
this joint density is given by the market prices of financial options written
on the relevant underlying assets. Under the Black model, asset prices are
jointly lognormal (under the appropriate equivalent martingale measure) and
analytical formulae are available. However, both the marginal distributions and
the dependence structure associated with this model may not be consistent with
information given by the prices of vanilla options, since implied volatilities are
strike-dependent. Constructing an alternative model with realistic properties
that may be calibrated to all relevant market data is a very general problem
in finance. Quanto FX options, spread options, basket options etc. are all
examples of such multi-asset pricing problems. It is extremely important to
practitioners to find an effective solution.
In our pricing methodology, we employ a copula function as a model for
the dependence structure. A copula provides the link between the multivariate
joint distribution and the univariate marginals. This allows us to separate the
modelling of the marginal distributions from the modelling of the dependence
structure, permitting a two-stage calibration. The concept of obtaining implied
marginal distributions from market prices of vanilla options is certainly not new
(see, for example, [Dupire, 1994]) and there is a growing literature in this area.
However, there currently exists no literature concerning the practical extraction
of the entire joint distribution of asset prices from implied volatilities.
The application of copulas in finance has recently attracted a great deal of
interest. [Embrechts et al, 2001] provides a survey of copulas and dependence
concepts in relation to finance, with particular focus on elliptical, Archimedean
and Marshall-Olkin copulas and applications in insurance risk and market risk
(VaR). In relation to option pricing, practical interest has recently focused on
credit derivatives, where simple parametric copulas are used to capture the de-
pendence between default and asset prices (see, for example,
[Schönbucher at al, 2001]). However, in this case the dependence structure is
given exogenously and is not implied from prices of vanilla options. [Rosenberg, 2003]
applies a similar methodology to the FX problem we consider, estimating the
dependence function via a non-parametric method based on historical return
data (see Section 3.2).
Considering the standard quanto FX option, market prices of relevant vanilla
options provide both information regarding the marginal distributions of the two

2
relevant FX rates at expiry and also partial information on the dependence struc-
ture (see Section 3.1). Calibration of marginal distributions is accomplished in
an initial stage. Given these marginal distributions, we find that the depen-
dence structure associated with the Black model (the bivariate Normal copula)
is insufficiently flexible to permit calibration to the final set of available im-
plied volatilities. We accomplish this final calibration step by modifying the tail
dependence structure of the Normal copula using a suitable parametric trans-
formation. This transformation is based on a generic parametric transformation
of a bivariate copula first suggested in [Genest, 2000]. The resulting model is
calibrated to the prices of vanilla options for various maturities and currency
set-ups.
We have only looked at one multi-asset option in this paper. As mentioned
above, this methodology is clearly applicable to other multi-asset options such
as spread options. Also, this methodology may have more general application in
many real-world statistical problems where the Normal copula does not provide
a sufficiently accurate model for the dependence structure under consideration,
for example in scenario analysis or stress-testing.
The rest of this paper is organised as follows: In Section 2, we describe the
problem of pricing a standard quanto FX call option and outline current market
practice. In Section 3, we give an overview of our proposed pricing methodology,
introduce the notion of dependence modelling with copulas and describe how an
appropriate transformation of the Normal copula provides a suitable model for
the dependence structure, which may be calibrated to market prices of vanilla
options. We present the results of calibration to actual market data for several
currency set-ups and maturities in Section 4. In Section 5, the prices of a quanto
FX option under our proposed model are compared against prices calculated
using the Black model and those given by an alternative ad-hoc approximation.
Our conclusions are presented in Section 6.

2 Motivation
2.1 Quanto Call Option Pricing Problem
i
Consider the standard quanto FX call option pricing problem. Let Dt,T denote
the time-t value in currency i of the zero coupon discount bond with maturity
T. Let Qi be the equivalent martingale measure associated with this numeraire.
Also let Xti,j , (i 6= j) denote the value in currency i of one unit of currency
j at time t. For an arbitrage-free economy we must have Xti,j = (Xtj,i )−1 and
Xti,j = Xtk,j /Xtk,i .

3
For the quanto pricing problem we only need to consider three currencies.
A standard quanto FX call option is a contract which pays the holder a total
of κ(XTi,j − K)+ in the “quanto” currency k, where κ is a known constant,
the “conversion factor.” Note that κ is simply a scaling factor. Currency i is
commonly referred to as the “money” currency and currency j as the “dealt”
currency. Since we may renumber currencies at our convenience, without loss of
generality we choose currency one to be the quanto currency, currency two to be
the money currency and currency three to be the dealt currency. By standard
arbitrage pricing theory, the value of this quanto call option at time zero is given
by
C0quanto = κD0,T
1
EQ1 [(XT2,3 − K)+ ], (1)

where Q1 is the equivalent martingale measure associated with the numeraire


1
Dt,T . Therefore the value of the quanto option is completely determined if the
distribution of XT2,3 under Q1 is known.
This quanto FX option must be priced consistently with the prices of vanilla
options written on the three relevant underlying FX rates, XT1,2 , XT1,3 and XT2,3 .
The prices of vanilla options on each of these FX rates are determined by the
market for a finite set of strikes.
The prices of vanilla options on XT1,j (j = 2, 3) allow us to recover the
marginal distributions of XT1,j under Q1 , under suitable assumptions regarding
the shape of the distribution. This subproblem has received a great deal of
attention recently in the literature and we proceed by choosing a simple param-
eterisation of the distribution, described in Section 3.1.1.
Prices of vanilla options on XT2,3 give information about the distribution
of this FX rate under Q2 . A change of measure shows these options provide
partial information regarding the joint distribution of (XT1,2 , XT1,3 ) under Q1 ,
since XT2,3 = XT1,3 /XT1,2 .
Under a suitable model for the joint density, we may calculate the distri-
bution of XT2,3 under Q1 and hence the price of the quanto FX option. The
problem is therefore to formulate a model such that we capture the dependence
structure implied by the third set of option prices (on XT2,3 ) while incorporating
the information provided about the marginal distributions given by the first two
sets of prices (on XT1,2 and XT1,3 respectively).
In the following, we choose to separate the modeling of the marginal distri-
butions and the dependence structure by specifying the dependence structure
in terms of a copula function, described in Section 3.2. Note that this allows us
to interchange the method we have used in Section 3.1.1 to obtain the marginal
distributions referred to above with other more sophisticated methods. Once we
have calibrated the marginal distributions to the first two sets of option prices,

4
the problem remains to choose a suitable parameterisation of the dependence
structure such that the associated joint density is well calibrated to the third
set of vanilla option prices.

2.2 Current market practice


In the market, the standard approach to pricing quanto FX options is based
on a Black-type model where we assume correlated lognormal dynamics for the
i,j
forward FX rates. Let Mt,T denote the forward FX rate at time t with maturity
T quoted in units of currency i per unit of currency j. Then by no arbitrage,
i,j j
Mt,T = Dt,T Xti,j /Dt,T
i

and
i,j i,j −1 i,j k,j k,i
Mt,T = (Mt,T ) , Mt,T = Mt,T /Mt,T .
1,j
Under the Black model, we assume the forward rates Mt,T follow the process
1,2 1,2 1,3 1,3
dMt,T = σ (1) Mt,T dWt1 , dMt,T = σ (2) Mt,T dWt2

under the equivalent martingale measure Q1 corresponding to the numeraire


1
Dt,T , where σ (1) and σ (2) are (strike-independent) volatilities of vanilla options
on XT1,2 and XT1,3 respectively and Wt1 and Wt2 are standard Brownian motions
under Q1 such that
dWt1 dWt2 = ρ dt.
It is straightforward to show that under Q2 ,
2,3 2,3
dMt,T = σ (3) Mt,T dWt3

where
(σ (3) )2 = (σ (1) )2 + (σ (2) )2 − 2ρσ (1) σ (2) (2)
and Wt3 is a standard Brownian motion under Q2 .
Consider a contract which pays VT = κ(XT2,3 − K) at time T in currency
one. Taking the expectation under Q1 , the usual pricing formula yields
(1) 2
2,3 ((σ ) −ρσ (1) σ (2) )(T )
V0 = κ(M0,T e − K),

which is zero if the strike K is given by the quanto forward rate


2,3 ∗ 2,3 ((σ (1) 2
) −ρσ (1) σ (2) )T
M0,T = M0,T e .

In the Black model, XT2,3 is lognormal under Q1 , thus evaluating the expec-
tation (1), we find the price of the quanto option is given by

C0quanto = κD0,T
1 2,3
[M0,T Φ(d∗1 ) − KΦ(d∗2 )] (3)

5
where ∗
2,3
log(M0,T /K) 1 (3) √ √
d∗1 = √ + σ T, d∗2 = d∗1 − σ (3) T
σ (3) T 2
and Φ(·) is the standard cumulative Normal distribution function. Note that
under the Black model, σ (3) is the volatility of vanilla options on XT2,3 , given
by Equation (2). Thus given the (strike-independent) volatilities σ (1) , σ (2) and
σ (3) , we may infer the “implied” correlation ρ under the Black model.
The above model is a good benchmark against which to compare more so-
phisticated models. To understand the deficiencies of this model, it is necessary
to look at the implied volatilities corresponding to the vanilla options written
on the relevant underlying assets. Recall that it is standard market convention
to quote a vanilla FX option price in units of its implied volatility, that is, the
corresponding value of the volatility parameter which gives the correct market
price when entered into the Black formula. Thus the Black formula is simply
used as a 1-1 mapping of market prices to Black volatilities.1
In practice the Black volatility corresponding to a quoted vanilla option
price is dependent on the strike of the option, indicating that the assumptions
underlying the Black model do not hold (this well-known feature is termed
the volatility “smile”). Therefore the assumption that XT1,2 , XT1,3 are jointly
lognormal under Q1 may be inappropriate. In general, it is not possible to
obtain an analytical form for the distribution of XT2,3 . To incorporate the effect
of the volatility smile, practitioners often adopt an ad-hoc approach and modify
the Black formula (3) as follows. The value of ρ may be determined as usual from
the at-the-money volatilities of options on XT1,2 , XT1,3 and XT2,3 using Equation
(2), then an ad-hoc approximation for the price of the quanto option may be
found by substituting the actual strike-dependent volatility σ̃ (3) (K) for σ (3) in
Equation (3), where σ̃ (3) (K) is the Black volatility corresponding to the price of
the vanilla option XT2,3 with the same strike K as the quanto option in question.
In Section 5.2 we compare these “Black-adjusted” prices with those obtained
by the copula approach, described below. Note that this ad-hoc modification
of the Black formula does not provide a model for the joint distribution of XT1,2
and XT1,3 . Therefore the only comparison with the copula approach available to
us is a simple comparison of calculated quanto prices for a range of strikes.
1 Recall that the “at-the-money” FX option is that which has strike given by the current

forward FX rate. Options with higher (lower) instrinsic value are known as “in-the-money”
(“out-of-the-money”) options. Market convention also dictates that the strike of a vanilla FX
option is quoted in terms of the corresponding delta, i.e. the value of the delta parameter
which gives the correct value of the strike of the option under the Black model, given the
implied volatility of the option. By convention, the at-the-money volatility quote is used
in this calculation. For a more detailed description of market practice, see, for example,
[Malz, 1997].

6
3 Methodology
3.1 Overview of pricing methodology
The price of a standard quanto option may be evaluated directly given the dis-
tribution of XT2,3 under Q1 , which in turn may be computed from the joint
distribution function of XT1,2 and XT1,3 under Q1 . We model this joint distribu-
tion function such that it is well calibrated to all relevant vanilla FX options as
follows.
In our modeling framework we choose to specify the dependence structure
between XT1,2 and XT1,3 under Q1 using a function known as a copula (described
in Section 3.2). This allows us to separate the modeling of the implied marginal
distributions (incorporating information given by market quotes for the prices
of vanilla options on XT1,2 and XT1,3 ), from the modeling of the dependence
structure (for which we have only partial information given by the prices of
vanilla options on XT2,3 ).
In the first stage of this methodology, we choose to model the implied
marginal distributions by parameterising each distribution as a mixture of log-
normals (outlined in Section 3.1.1). Since we subsequently model the depen-
dence structure via a copula function, this method may be interchanged with
more sophisticated methods without affecting the remainder of the methodology.
Once the marginal distributions have been determined, the joint distribution
of XT1,2 and XT1,3 under Q1 may be computed under the assumption that the
dependence structure is modeled appropriately using a given copula. Under
this model for the joint distribution, prices of vanilla options on XT2,3 may be
calculated by numerical integration. The problem remains to choose a suitable
copula such that that the corresponding joint distribution can be calibrated to
market quotes for the prices of vanilla options on XT2,3 .
There are infinitely many models for the joint distribution which are well-
calibrated to all relevant market prices since the available prices of vanilla op-
tions on XT2,3 only provide partial information regarding the dependence struc-
ture. We proceed to model the dependence structure by perturbing the Normal
copula, which is the model for the dependence structure associated with the
Black pricing formula.
Under the Black pricing model, implied volatilities are independent of the
strike. The correlation parameter ρ controls the overall level of association
between XT1,2 and XT1,3 under Q1 and therefore the level of (flat) volatilities
of vanilla options on XT2,3 . In the general case implied volatilities are strike-
dependent, so the Black model can only be calibrated to at-the-money volatili-
ties (by choosing the value of ρ to satisfy (2)). We shall see that if we replace

7
the lognormal marginals by those actually implied by vanilla options on XT1,2
and XT1,3 , we find that the parameter ρ of the Normal copula still controls the
overall level of association and consequently the general level of volatilities of
options on XT2,3 (see Figure 3, Section 4.2).
Given the implied marginal distributions, we shall find that typically it is
necessary to alter the tail dependence characteristics of the Normal copula in
order to match the volatility smile of vanilla options on XT2,3 (see Section 3.3.3).
We choose to accomplish this by modeling the dependence structure using a
specific parametric transformation of the Normal copula (based on the generic
copula transformation suggested in [Genest, 2000]). We shall see that the pa-
rameter ρ of the original Normal copula still controls the overall level of volatil-
ities of options on XT2,3 under this new model for the joint distribution provided
the transformation function exhibits certain qualitative features.
Calibration to implied volatilities of options on XT2,3 is performed by a non-
linear weighted least squares optimisation procedure. The resulting joint distri-
bution is consistent with distributional information provided by vanilla options
on all three FX rates.

3.1.1 Implied marginal distributions

We now consider the problem of recovering the implied marginal distributions


of the FX rates XT1,j (j = 2, 3) from market quotes of prices of vanilla options
written on these FX rates, which are available for a finite set of strikes.2 One
suitable method for constructing the marginal distributions such that they are
consistent with this market data is to assume a suitable parametric form such
as a mixture of two lognormal distributions.
This method is computationally simple, efficient and the optimisation ap-
pears to be relatively stable (see [Bahra, 1997]). It is often used as a bench-
mark with which to compare more sophisticated models (see, for example,
[Bliss et al, 2000], [Campa et al, 1997] and [Coutant et al, 1998])). The mix-
ture of lognormals model clearly reduces to the special case of a single lognor-
mal distribution (Black model) in the case that the mixture parameter is one
or zero; this is extremely useful in testing the numerical implementation of the
following copula model. Both density and distribution functions are analytical
functions of the cumulative Normal distribution function.
The mixture of lognormals distribution is calibrated to market data as fol-
lows. A vanilla call option on the FX rate Xt1,j , (j = 2, 3) with strike K gives
2 Of course, if we did know prices of vanilla options for all strikes this would completely

specify the marginal distributions of XT1,j (j = 2, 3) (see [Dupire, 1994]).

8
the holder a payoff of (XT1,j − K)+ at time T, hence by the usual arbitrage
pricing argument the value of this option at time zero is given by

C01,j = D0,T
1
EQ1 [(XT1,j − K)+ ]. (4)

This expectation may be written as an integral involving the probability density


function of XT1,j under Q1 .
If the distribution of XT1,j under Q1 is assumed to be a mixture of lognor-
mals, analytical formulae are obtained for prices of vanilla options in terms of
the standard cumulative Normal distribution function (see Appendix A for de-
tails). The density may be fitted to market quotes by performing a suitable
optimisation to find the density parameters which minimise the squared error
between model and market prices of the relevant vanilla options. Notice that
1,j
the forward FX rate M0,T is the mean of the marginal distribution of XT1,j under
Q1 . This is fitted exactly, reducing the number of free distribution parameters
by one.
Note that this method always results in proper distribution and density
functions which are both smooth and differentiable; this is important in the
following derivations and is not necessarily apparent in some methods.3 A
number of other methods may have been suggested with superior numerical
properties at the expense of more complex implementation.4 However, as noted
earlier it is possible to substitute an alternative to this method of obtaining the
marginal distributions without affecting the remainder of the methodology.

3.1.2 Partial Information on Joint Distribution

We now consider the partial information provided about the joint distribution
of XT1,2 and XT1,3 under Q1 which is given by the prices of vanilla options on
XT2,3 . The value in currency two of a vanilla call option on XT2,3 with strike K
at time zero is given by

C02,3 (K) = D0,T


2
EQ2 [(XT2,3 − K)+ ],
3 require that the density f 1,j (·) is both continuous and differentiable, with f 1,j (0) = 0,
R ∞ We 1,j (u) du = 1 and f 1,j (u) > 0 for all u ∈ R. In some “smoothing”-type methods where
0 f
call volatilities are interpolated and differentiated with respect to the strike to obtain the
distribution (following [Dupire, 1994]), lognormal tails are pasted onto the central part of the
distribution and it isR not always straightforward to smooth the density function across the
join whilst ensuring 0∞ f 1,j (u) du = 1.
4 The mixture of lognormals model has the disadvantage that the single optimal fit available

is not guaranteed to match market quotes to sufficient accuracy in all cases. More sophisti-
cated “smoothing”-type methods usually have a goodness-of-fit parameter allowing a closer
fit to market quotes at the cost of reduced smoothness. [Bliss et al, 2000] indicate that these
methods may offer a more stable fit by observing the effects of perturbing market quotes. Also
see [Malz, 1997], [Sherrick at al, 1996] and [Ait-Sahalia at al, 1998].

9
where Q2 is the equivalent martingale measure corresponding to the numeraire
2
Dt,T . We show that prices of such options provide information on the values of
integrals over specific regions of the joint distribution of (XT1,2 , XT1,3 ) under Q1 ,
via a change of measure.
The Radon-Nokodym derivative of Qj relative to Qi is given by
i,j
dQj ¯¯ Mt,T
¯
= i,j
,
dQi ¯Ft M0,T

so as a consequence of the Radon-Nikodym theorem


" #
dQ2 ¯¯
¯
2,3 2,3
EQ2 [(XT − K)+ ] = EQ1 (X − K)+
dQ1 ¯Ft T
1
D0,T
= X02,1 1,2 2,3
2 EQ [XT (XT − K)+ ].
1
D0,T

Therefore

C02,3 (K) = X02,1 D0,T


1
EQ1 [(XT1,3 − KXT1,2 )+ ]
Z ∞Z ∞
2,1 1 Q1
= X0 D0,T (u − Kv)fX 1,2
,X 1,3
(u, v) du dv, (5)
0 Kv T T

1
Q
where fX 1,2
,X 1,3
(·, ·) is the joint density of XT1,2 and XT1,3 under Q1 . If this joint
T T
density is known, the integral in Equation (5) may be evaluated numerically and
hence vanilla call option prices C02,3 (K) may be calculated for various strikes K
and compared against market quotes.
If the marginal distributions of XT1,3 and XT1,2 are already determined (as
described in Section 3.1.1), the model of the joint distribution is completed by
specifying a suitable model for the dependence structure of (XT1,3 , XT1,2 ) under
Q1 . The remaining practical problem is to choose the copula to match the partial
information given by the integrals (5) corresponding to the market prices of
vanilla options on XT2,3 for different strikes K.

3.2 Dependence Modeling with Copulas


A bivariate copula is a function which allows us to investigate the scale-invariant
dependence structure of two variables. It provides the link between the marginal
distribution functions and the joint distribution function. By changing the
copula it is possible to alter the dependence structure of the joint distribution
without altering the marginal distributions.

10
Here it is convenient to define a bivariate copula as a bivariate distribution
function whose one-dimensional margins are uniform on the interval [0,1]. As-
suming we know the margins, the Theorem below shows that we can obtain the
copula distribution from the joint distribution and vice versa.
There is a growing literature on copulas and their application in finance. For
formal definitions, theoretical properties and comparisons of common paramet-
ric copulas we refer the reader to [Nelson, 1998] and [Joe, 1997]. [Embrechts et al, 2001]
provides a survey of copulas and dependence concepts in relation to finance, with
particular focus on Marshall-Olkin, elliptical and Archimedean copulas and ap-
plications in insurance risk and market risk (VaR).
Recent work has recognised the potential application of copulas in derivative
pricing, especially in the field of credit derivatives. However, the copula is gener-
ally specified exogeneously. Usually either the copula is assumed to take a par-
ticular parametric form, which may have more desirable dependence character-
istics than the Normal copula for the particular application (such as the Student
t copula), or the copula is estimated using historical data. [Rosenberg, 2003]
performs the latter, applying non-parametric techniques to estimate the depen-
dence copula between two FX rates, while using implied volatilities to construct
the margins. However, the implied volatilities on the cross-FX rate XT2,3 do not
match market quotes. Practitioners usually need a model that is well calibrated
to current market prices. There is presently no literature regarding the practical
application of copulas to option pricing where the copula is calibrated to the
prices of vanilla options.

3.2.1 Sklar’s Theorem

Sklar’s theorem explains the role of copulas in describing the relationship be-
tween a multivariate joint distribution function and its univariate marginal dis-
tributions.

Theorem (Existence) Let H be a standard joint distribution function and let


F and G be the margins of H. Then there exists a copula C such that for all
x, y ∈ R̄,
H(x, y) = C(F (x), G(x)). (6)

Corollary If the marginal distribution functions F and G are strictly increasing,


then
C(u, v) = H(F −1 (u), G−1 (v)).

Since it is a prerequisite that the marginal densities of XT1,2 and XT1,3 under
Q1 are continuous, if we have in mind a model for the continuous joint density

11
we may calculate the unique copula associated with this joint density and vice
versa.

3.2.2 Parametric copulas

The most common practical example of a copula is probably the Normal copula,
which has distribution function

CN (u, v | ρ) = Φρ (Φ−1 (u), Φ−1 (v)),

where Φρ is the standard bivariate Normal distribution function with correla-


tion ρ and Φ is the standard univariate Normal distribution function. By the
Corollary above, this is the copula implied by a standard bivariate Normal joint
distribution function with standard univariate Normal marginal distributions.
It is clear that this elliptical copula exhibits both positive and negative depen-
dence (for positive and negative values of ρ respectively).
In the following we shall refer informally to the behaviour of the copula in
the extremes of the upper-right and lower-left quadrants of the domain as upper
and lower tail dependence respectively. Thus if we perturb the Normal copula to
increase upper tail dependence, we will observe that the probability of observing
very high values in both variables is higher than under the Normal copula.
The Normal copula defined above describes the dependence structure asso-
ciated with the Black model. We shall see below that in order to calibrate our
joint distribution to vanilla options on XT2,3 , we require a dependence model
with much greater flexibility. Specifically, the copula must exhibit both posi-
tive and negative overall dependence structures and it must be possible to alter
simultaneously the level of upper-tail and lower-tail dependence (see Section
3.3.3).
For most common parametric copulas in the literature it is not possible to
alter the overall level of dependence and the upper and lower tail dependence
characteristics simultaneously. This is a feature of all single parameter copulas.
A few two-parameter copulas have been studied in the literature, but these are
generally quite inflexible and it is difficult to construct a multi-parameter copula
directly which has the desired properties. However, we shall see that a copula
with all the required characteristics may be constructed by a transformation of
the Normal copula (see Section 3.3.3).

12
3.3 Fitting a copula using the information available on the
joint distribution
3.3.1 Calculation of vanilla call prices under a given copula model

Suppose we have in mind a suitable model for the dependence structure of XT1,2
and XT1,3 under Q1 in terms of a specific copula. In order to assess the fit of this
copula, we need to be able to calculate the prices of vanilla options on XT2,3 .
Using the marginal distributions which have already been determined (as
described in Section 3.1.1), it is possible to recover the associated joint density
function as discussed below. From this joint density, prices of vanilla options
on XT2,3 may be calculated by numerical integration (see Section 3.3.2).
Q1
Denote the joint distribution function of XT1,2 and XT1,3 under Q1 by FX 1,2
,X 1,3
(·, ·)
T T
1 1
Q Q
and the associated marginal distribution functions by FX 1,2 (·) and F 1,3 (·) re-
X
T T

spectively. If the dependence structure between XT1,2 and XT1,2 under Q1 may
be modeled appropriately by a known copula with density function c(·, ·), then
using Sklar’s theorem (6), the joint density is given by
1 1 1 1 1
Q Q Q Q Q
fX 1,2
,X 1,3
(u, v) = fX 1,2 (u)f 1,3 (v)c(F 1,2 (u), F 1,3 (v)).
X X X
T T T T T T

If the copula density is known, we may now calculate the price of a vanilla
option with a given strike by evaluating the integral in Equation (5) numerically.
These model prices may then be compared against market quotes.

3.3.2 Reformulating Integration Region

We now consider the integration regions corresponding to the integral in Equa-


tion (5) for different strikes. These integrals correspond to the prices of vanilla
options on XT2,3 . We need to understand how the Normal copula must be modi-
fied such that the resulting joint distribution is well calibrated to market quotes
for a given set of strikes. We find that contour plots of copulas vs. Normal
marginals provide a very helpful clue as to how to accomplish this.
Q1
Suppose for the moment that the marginal distributions FX 1,j (·) are both
T
lognormal (j = 1, 2). Then for a given strike K, the integration region associated
with the integral in Equation (5) corresponds to the region of the copula density
below the line log(v) = log(u) + K when plotted against Normal marginals (see
Fig. 1). In the general case, of course, the integration region will not take such
a simple form. However, since the observed implied marginals will not be too
far removed from the single lognormal model, any intuition we can gain in this
simple setting should carry over to a more realistic model for the marginals.

13
u=v

2
0.02

0.04
0.06
0.08
0.1

1
0.12

0.14
0.16
0.18
0.2

0.22

log(v)

0
log(v)=log(u)+K

-1
-2

0.02

-2 -1 0 1 2

log(u)

Figure 1: Where the joint distribution of XT1,2 and XT1,2 under Q1 has lognormal mar-
gins (i.e. flat volatilities), the integration region associated with integral in Equation
(5) is the shaded region below the line log(v) = log(u)+K. For high strikes, integration
region is moved further to the bottom right. Contours shown correspond to a Normal
copula with positive correlation.

u=v
2
1

0.2
log(v)

0.2 0.22
0.2 0.2
0.18 log(u)+log(v)=K
0.2 0.16
0.12
0.14 0.02
0.1
0.08
0.06
-1

0.04
-2

0.02

-2 -1 0 1 2

log(u)

Figure 2: The reformulated integration region associated with integral in Equation


(7) is that above the line log(u) + log(v) = K. For high strikes, this region is moved
further to the upper right. Contours shown correspond to the same Normal copula
shown in Fig. 1.

14
Suppose now that we wish to modify the Normal copula to fit the strike-
dependent volatilities of vanilla call options on XT2,3 , but still retain a symmetric
structure, i.e. c(u, v) = c(v, u) for all u, v ∈ R. If the implied volatility of a
vanilla call with a high strike is higher than expected under the Normal copula
model, then if we wish our copula model to correctly price this option we must
increase the value of the integral (5). This may be accomplished by increasing
the mass of the copula density corresponding to high values of u and low values
of v (see Fig. 1). However, due to the symmetry of the copula this would also
increase the mass of the copula density corresponding to low values of u and
high values of v, thus also increasing call prices with low strikes simultaneously.
A symmetric structure is a feature of most copulas that are suitable for
practical applications. The application of non-symmetric copulas is a more
complex problem requiring much greater computational complexity. We note
that there is very little literature regarding the practical application f non-
symmetric copulas. It is therefore desirable to reformulate the model so that
symmetric copulas may be used.
If we specify the joint density in terms of 1/XT1,2 = XT2,1 and XT1,3 under Q1 ,
Z ∞Z ∞
Q1
C02,3 (K) = X02,1 D0,T
1
(u − K/v)fX 2,1
,X 1,3
(u, v) du dv, (7)
0 K/v T T

where
1 1 1 1 1
Q Q Q Q Q
fX 2,1
,X 1,3
(u, v) = fX 2,1 (u)f 1,3 (v)c(F 2,1 (u), F 1,3 (v))
X X X
(8)
T T T T T T

1 1 1 1
Q Q Q Q
and both fX 2,1 (·) and F 2,1 (·) may be found from f 1,2 (·) and F 1,2 (·) directly
X X X
T T T T
(see Appendix A). If the margins are lognormal, the region of the copula cor-
responding to the integration region of Equation (7) is illustrated in Fig. 2 for
a given strike. The values of integrals (7) for different values of K may now
be altered independently using a symmetric copula by shifting the mass of the
copula density along the line u = v. Thus relatively high call prices for high
strikes may be obtained by increasing upper tail dependence and so forth. By
controlling the upper and lower tail dependence of the copula, it is possible to
alter model call prices for different strikes such that they match market quotes.
Equations (7) and (8) allow us to calculate prices of vanilla options un-
der our copula model for any given strike. Of course, under the Black model
(corresponding to lognormal margins and a Normal copula with correlation ρ),
Equation (7) may be evaluated analytically. The implied volatility associated
with these call prices is given by

(σ (3) )2 = (σ (1) )2 + (σ (2) )2 + 2ρσ (1) σ (2) , (9)

15
where σ (1) and σ (2) are the (flat) implied volatilities of call options on XT1,2 and
XT1,3 respectively.5

3.3.3 Transformation of Normal Copula

The copula which we calibrate to the prices of vanilla options on XT2,3 will
not be unique because these prices only give partial information regarding the
dependence structure. However, we have greater confidence in our inferences if
our chosen copula model is close to the Normal copula, which is the dependence
model associated with the Black pricing formula. Therefore the copula we choose
to fit to market data is a perturbation of the Normal copula.
We construct a new symmetric copula by a suitable transformation of the
Normal copula, modifying the tail dependence structure such that the resulting
joint density is well calibrated to prices of vanilla options on XT2,3 . The following
Theorem gives sufficient conditions under which the transformed copula is itself
a copula (for proof see Appendix B).

Theorem [Genest, 2000] Let ϕ : [0, 1] → [0, 1] be a continuous, twice differen-


tiable concave function such that ϕ(0) = 0 and ϕ(1) = 1. Then

Cϕ (u, v) := ϕ−1 (C(ϕ(u), ϕ(v))) (10)

is a copula if C(·, ·) is a copula.

The density of the transformed copula Cϕ is obtained by taking the partial


derivative of (10) with respect to both arguments;

ϕ0 (u)ϕ0 (v)
µ
cϕ (u, v) = c(ϕ(u), ϕ(v))
ϕ0 (Cϕ (u, v))
ϕ00 (Cϕ (u, v)) ∂C

∂C
− 0 (ϕ(u), ϕ(v)) (ϕ(u), ϕ(v)) , (11)
[ϕ (Cϕ (u, v))]2 ∂u ∂v

where c(·, ·) is the density of the original copula C.


This transformation allows us to construct infinitely many new copulas from
any given copula. If our starting point is the Normal copula with parameter
ρ, then we may use the transformation function to modify the tail dependence
characteristics of the copula as required. Note that the identity transformation
ϕ(x) = x does not alter the original copula.
Observing the form of Equation (10), notice that the behaviour of ϕ(x) near
x = 1 controls the upper tail dependence of the transformed copula. To see
5 Note ρ has changed sign, c.f. Equation (2).

16
this, note that for u, v both near one, C(ϕ(u), ϕ(v)) will be near one, hence it
is the behaviour of ϕ(x) and ϕ−1 (x) for x close to one which controls the level
of Cϕ (u, v). Similarly, the behaviour of ϕ(x) near x = 0 controls the lower tail
dependence of the transformed copula.
Henceforth suppose the original copula C in (10) is the Normal copula. We
require that the second derivative ϕ00 (x) of the transformation function be near
zero around x = 21 or else the transformation will significantly alter the overall
dependence characteristics of the copula. Provided this is the case, experimen-
tation shows that the parameter ρ still controls the overall dependence charac-
teristics of the copula. This allows us to use the endpoints of the transformation
function to alter the tail dependence characteristics. Increasing the size of the
second derivative of the transformation near zero (one) is observed to increase
the lower (upper) tail dependence of the transformed copula.
Heuristically, how can we see this from Equation (11)? We are performing
a slight modification of the Normal copula, hence the optimal transformation
function should be close to ϕ(x) = x. For transformation functions with char-
acteristics described above, increasing the size of ϕ00 (·) < 0 near zero or one
will increase the second term of Equation (11), which will eventually dominate
since ϕ0 (·) will be close to one and the tails of cN (ϕ(u), ϕ(v)) decay exponen-
tially. This follows because the quantities ϕ(·), ϕ0 (·) and ∂C
∂u (·, ·) do not vary
N

00
as fast as ϕ (·). The mass of the copula density is shifted toward those regions
corresponding to lower or upper tail dependence.
If the copula C is a Normal copula then the transformed density function
(11) may be evaluated directly because the normal copula density cN and the
partial derivative ∂C ∂v are analytical (only requiring an approximation to the
N

standard Normal cumulative distribution, see Appendix B) and the distribution


function of the bivariate Normal copula CN (·, ·) may be approximated to high
accuracy (see, for example, [Genz, 2002]).
It remains to parameterise the transformation function ϕ such that its second
derivative may be modified in order to alter the upper and lower tail dependence
as required.

3.3.4 Calibration of transformation function

Given the correlation ρ of the Normal copula which defines the dependence
structure of XT2,1 and XT1,3 under Q1 and assuming the transformation function
ϕ has been parameterised and satisfies the conditions of the above Theorem,
the density of the transformed Normal copula may be computed, hence the joint
density of XT2,1 and XT1,3 under Q1 may be found using Equation (8). Vanilla
call option prices C02,3 (K) are computed under this model by evaluating the

17
integral in Equation (7). These prices can then be compared against market
quotes.
A non-linear optimisation may be performed to find the values of the trans-
formation parameters which give the closest match of model and market vanilla
option prices. A standard non-linear least squares algorithm will find the opti-
mal solution z∗ ∈ Rn which minimises the objective function
m
X
kr(z)k2 = ri (z)2 ,
i=1

where n ≤ m.
Let the residuals ri be given by

ri (z) = wi (σiobserved − σimodel (z)), i = 1, . . . , m

where σiobserved are the available market implied volatility quotes associated
with the prices of vanilla options on XT2,3 with strikes Ki , σimodel (z) are the cor-
responding implied volatilities calculated under the transformed copula model
with parameter vector z ∈ Rn and wi are given weights.
In order to modify the tail dependence structure of the Normal copula, it
is necessary to construct a transformation function with some flexibility, incor-
porating our prior intuition regarding the shape of a suitable transformation
function as discussed in Section 3.3.3. Therefore we choose to specify ϕ as a
cubic spline with k + 1 suitable predefined knot points pi ∈ [0, 1], such that the
constraints on ϕ in the Theorem of Section 3.3.3 are satisfied.
Since the size of the second derivative gives a good indicator of the increase
in tail dependence, it makes sense to parameterise this directly. Hence ϕ is
found by specifying the second derivative of the spline at each knot point, then
recovering the coefficients of the spline by making use of the requirement that
ϕ(0) = 0 and ϕ(1) = 1.
At each optimisation step, the first component of z = (z1 , . . . , zn )T may be
used to determine the correlation of the Normal copula ρ ∈ (−1, 1) (for example,
via ρ := sin(z1 )).
Suppose we choose knot points p = (p0 , p1 , . . . , pk )T with p0 = 0 and pk = 1.
Define the cubic spline between the j th and (j + 1)th nodes by
3
X
Sj (x) := ai,j (x − pj )i , (12)
i=0

for x ∈ [pj , pj+1 ] and coefficients ai,j to be determined.


For a well-defined optimisation we have n ≤ m, so ϕ may be constructed
using at most m−1 parameters. Since 21 Sj00 (x−pj ) = a2,j +3a3,j (x−pj ), setting

18
2
a2,j := −zj+1 < 0 (j = 1, . . . , k − 1) leads to a negative second derivative for all
x ∈ [0, 1] if the endpoints of the cubic spline

ϕ00 (0) = S000 (0) := 2a2,0 < 0, ϕ00 (1) = Sk00 (0) := 2a2,k < 0

are given. Solving for the other coefficients by incorporating the endpoints
ϕ(0) = 0 and ϕ(1) = 1 results in a cubic spline with negative curvature (see
Appendix C). However, the size of the predefined values of ϕ00 (0) and ϕ00 (1)
will have a material impact on the shape of ϕ.
In the current pricing problem we calibrate to prices of vanilla options for
m = 5 strikes. Since the objective of the transformation is to alter tail depen-
dence and not the general dependence structure (which we control by altering ρ),
we expect the second derivative of the optimal solution to be near zero around
x = 21 (see discussion in Section 3.3.3). Therefore we introduce the condition
ϕ00 ( 12 ) = 0, reducing the number of free parameters to be fitted by one. Let knot
points be given by p = (0, p1 , 21 , p2 , 1)T where p1 , p2 are given. Then using n = 5
parameters, the transformed copula is fully specified by setting ρ := sin(z1 ) and
constructing ϕ as a cubic spline with these knot points satisfying

a2,0 := −z22 , a2,1 := −z32 , a2,2 := 0, a2,3 := −z42 , a2,4 := −z52

with endpoints ϕ(0) = 0 and ϕ(1) = 1. For details see Appendix C.


At each optimisation step, we propose a parameter vector z which corre-
sponds to a value of ρ and a candidate transformation function given by the
cubic spline ϕ. We require that ϕ satisfy all of the conditions of the Theorem of
Section 3.3.3. Since the construction above gives rise to a cubic spline that takes
positive values and has negative curvature, it remains to check that ϕ(x) ≤ 1
for all x ∈ [0, 1]. But ϕ is concave and continuous therefore this condition must
hold unless ϕ0 (1) < 0. This is trivial to check. If this is the case, a large value
of the objective function may be returned to the optimisation loop so that the
current trial solution z will be discarded. In practice we find this situation does
not arise as long we choose the starting point of the optimisation z(0) such that
this condition holds.6
6 For instance, one suitable starting point is the vector z = (c, 0, . . . , 0)T for some constant

c ∈ R. This corresponds to a transformation function given by the identity transformation,


resulting in a Normal copula with correlation determined by the value of c (e.g. ρ = sin(c)).

19
4 Results of Calibration to European Call Op-
tions
4.1 Description of Dataset
Our model for the joint distribution is fitted to actual market quotes for the
implied volatilities of vanilla options written on exchange rates between the three
currencies JPY, EUR and USD, with strikes corresponding to standard values
of the option delta given by 10%, 25% and 50% and standard maturities 1M,
3M, 6M, 1Y and 2Y (see Appendix F). All option price data and corresponding
FX rates and discount factors were obtained at the close of 7 July 2001.

4.2 Comparison of implied volatilities of model prices of


vanilla options on XT2,3 against market quotes
Before presenting our findings on the calibration of our model to market data, it
is instructive to first consider the joint distribution model given by the Normal
copula with marginal distributions given by market data. Consider as exam-
ple the dataset corresponding to maturity 1M and the currency set-up where
(CCY1,CCY2,CCY3)=(USD,EUR,JPY). Implied volatilities of vanilla call op-
tions C02,3 (K) under the Normal copula model with various values of correlation
ρ are displayed in Fig. 3. As expected, it is observed that changing ρ alters the
overall dependence structure (as with the Black model) and therefore results
in a parallel shift in model implied volatilities with negligible change in slope
or curvature. Therefore it is not possible to find a close match to the market
quotes simply by varying ρ.
By contrast, Fig. 4 indicates that in this case the transformed Normal copula
offers essentially an exact fit to market quotes (with an objective function of
zero to 14 s.f. for this dataset), with transformation function ϕ(·) shown in Fig.
5. This transformation function is constructed by cubic splines as described in
Section 3.3.4, with knot points p = (0, 0.1, 0.5, 0.9, 1)T .
To see directly the effect of the transformation of the Normal copula, com-
pare the contour plot of the optimal transformed copula (Fig. 7) against that
of the optimal Normal copula (Fig. 8), which is calculated by finding the value
of the parameter ρ that minimises the same objective function described in Sec-
tion 3.3.4. Notice that the optimal transformed copula exhibits much higher
upper tail dependence as well as higher lower tail dependence than the Normal
copula. As discussed in Section 3.3.3, the resulting upper and lower tail depen-
dence characteristics of the transformed copula correspond to the graph of ϕ00 (·)
near one and zero respectively (Fig. 6). Experimentation shows that in order

20
0.18
0.16
Volatility

0.14

Actual Volatility
rho=0.2
rho=0.1
0.12

rho=0
rho=-0.1
rho=-0.2
rho=-0.3

0.0094 0.0096 0.0098 0.0100 0.0102 0.0104

Strike

Figure 3: Implied volatility of options on XT2,3 under the Normal copula model for
various values of correlation ρ. Here (CCY1,CCY2,CCY3)=(USD,EUR,JPY) and the
relevant marginal distributions have been calibrated to option prices with maturity
1M. Also shown are the actual market quotes for these implied volatilities.
0.18

Actual Volatility
Transformed Copula
Optimal Normal Copula
0.17
Volatility

0.16
0.15
0.14

0.0094 0.0096 0.0098 0.0100 0.0102 0.0104

Strike

Figure 4: Implied volatility of options on XT2,3 corresponding to the currency set-


up (USD,EUR,JPY) and maturity 1M, under the transformed Normal copula model
with optimal parameters. Also shown are the actual market quotes for these implied
volatilities and the implied volatilities under the optimal Normal copula model (which
minimises the same objective function as that used to assess the fit of the optimal
transformed Normal copula).

21
to attain a good fit to market quotes, a great deal of curvature is required near
zero and one, whilst close to zero near 12 .
We now consider the results of the calibration to all currency set-ups and
all five maturities for which we have market implied volatility quotes. We find
that the transformed copula model offers a better fit to market data than the
Normal copula in all cases.
For the general case where the smile in implied volatilities is symmetric, the
calibration to market data is particularly good (see Appendix D). This is also
the case when implied volatilities of options on XT2,3 are skewed but increasing as
a function of strike. We find that for a given currency set-up, the behaviour of ϕ00
is usually stable across different maturities (see Fig. 13). This is to be expected,
since for each currency pair in our dataset the implied volatility smile has the
same qualitative features across maturities, hence the overall “correction” of
the Normal copula achieved by the calibrated transformation function ϕ should
be the same. In general the dependence structure implied by market quotes of
vanilla options on XT2,3 has greater upper and lower tail dependence than that
exhibited by a Normal copula.
For the scenarios where the implied volatilities of options on XT2,3 are heavily
skewed and decreasing, the fit is occasionally not as good. For the currency set-
up (EUR,USD,JPY) and maturities 1M and 3M, the market volatility quotes
that we calibrate to are almost flat for strikes above ATM rather than the usual
more symmetric shape and it is difficult to fit this skew. The fit of the model may
be improved by shifting the knot points appropriately; experimentation shows
this works extremely well for the 3M maturity. Alternatively, observe that the
ideal transformation function exhibits a great deal of curvature in its second
derivative near zero and one in all cases and a piecewise linear construction
is a poor approximation to this. The parameterisation of the transformation
function ϕ00 may be altered in order to allow such curvature and thus afford a
better fit in these cases (see Appendix C).
However, the fit to the currency set-up (EUR,JPY,USD) is found to be
extremely good for all maturities. Interchanging currencies two and three also
results in a model which may be used to price any quanto FX option with quanto
currency EUR. Therefore for any currency set-up and meturity we are able to
find a well-calibrated model for either the joint density fX 2,1 ,X 1,3 or fX 3,1 ,X 1,2
T T T T
and we are in a position to value a quanto FX option.

22
1.0
0.8
0.6
phi(x)

0.4
0.2
0.0

0.0 0.2 0.4 0.6 0.8 1.0

Figure 5: Optimal transformation function ϕ(·) used to match volatilies in dataset


corresponding to the currency triple (USD,EUR,JPY) and maturity 1M (Fig. 4).

0.0 0.2 0.4 0.6 0.8 1.0


0
-1
-2
phi’’(x)

-3
-4

Figure 6: Second derivative ϕ00 (·) of transformation function used to match volatilies
in dataset corresponding to the currency triple (USD,EUR,JPY) and maturity 1M
(Fig. 4).

23
2
0.02

1
0.02

log(v)
0.16 0.08 0.04
0.14 0.12 0.1

0
0.06

0.02
-1

0.02
-2

0.02 0.02

-2 -1 0 1 2

log(u)

Figure 7: Contour plot of optimal Normal copula corresponding to dataset corre-


sponding to maturity 1M and the currency set-up (USD,EUR,JPY), plotted against
Normal margins. Optimal value of ρ is found to be -0.1299.
2
1
log(v)

0.140.12
0.16 0.1
0.06
0.04
0.08
-1
-2

0.02 0.02

-2 -1 0 1 2

log(u)

Figure 8: Contour plot of optimal transformed Normal copula fitted to dataset cor-
responding to maturity 1M and the currency set-up (USD,EUR,JPY), plotted against
Normal margins. Although the overall level of association is similar to the optimal
Normal copula (see Fig. 7), much greater upper tail dependence is required in order to
fit the implied volatilities of vanilla options on XT2,3 with high strikes. The transformed
copula also exhibits increased lower tail dependence and a much higher central peak.

24
5 Quanto Option Pricing
5.1 Valuation of Quanto FX Options under a given Copula
Model
The value of a standard quanto FX call option paying out in currency one is
computed by evaluating the expectation

EQ1 [(XT2,3 − K))+ ],

as described in Section 2.1. This is straightforward once the distribution of XT2,3


under Q1 is known.
Recall XT2,3 = XT2,1 XT1,3 and so
Z ∞
Q1 1 Q1 ³z ´
fX 2,3 (z) = f 2,1 1,3 , x dx.
T 0 x XT ,XT x

Therefore the price of a standard quanto call option may be evaluated directly
by numerical integration:

C0quanto 1
= κD0,T EQ1 [(XT2,3 − K)+ ]
Z ∞Z ∞
1 1 Q1 ³z ´
= κD0,T fX 2,1 ,X 1,3 , x (z − K) dx dz. (13)
K 0 x T T x

It is useful to note that the double integral in (13) corresponds to a region


Q1
of the joint density fX 2,1
,X 1,3
(p, q) bounded below by the line p = K/q. If the
T T
margins are lognormal, this corresponds to the region of the contour plot of the
copula density above the line log(p) + log(q) = K when plotted against Normal
margins (c.f. Fig. 2). This is the same integration region as a vanilla call option
on XT2,3 with strike K. Once the implied marginals have been determined, the
value of (13) depends critically on the distribution of the copula mass along the
line u = v.
In our model we perturb the Normal copula to match information on the
dependence structure contained in vanilla option prices on XT2,3 . Given the im-
plied marginal distributions, the price of a quanto option is likely to be highly
dependent on the fit of the copula to these vanilla option prices. Any distortion
of the copula density other than along the line u = v (for instance any asym-
metry) is unlikely to have a material effect on the price of the quanto option.
Thus we have more confidence that although our fitted joint distribution model
is not the only copula that could be calibrated to market prices, any other cop-
ula calibrated to market data is unlikely to result in markedly different quanto
prices.

25
Prices of standard quanto call and put options with the same strike are
linked by the “put-call parity” relationship

C0quanto (K) − P0quanto (K) + κD0,T


1 1
K = κD0,T EQ1 [XT2,3 ],

where the right-hand side is independent of the strike K (but is still dependent
on the particular model of the joint distribution). It is a useful numerical
check to verify that put-call parity holds for calculated quanto prices under the
transformed copula model.

5.2 Comparison of Quanto Prices under Copula model vs.


Black approximation
We now compare the prices of quanto options calculated under the transformed
copula model (evaluating Equation (13)) with both quanto prices calculated
under the standard Black model (calibrated to at-the-money volatilities) and
prices calculated using the “adjusted” Black formula (defined in Section 2.2).
With the adjustment of the Black formula we attempt to allow for the smile in
implied volatilities on XT2,3 , but we ignore the effect of any smile behaviour in
XT1,2 and XT1,3 . This ad-hoc formula does not correspond to any joint distribution
of XT1,2 and XT1,3 , so the only comparison of these approaches available to us is
a simple comparison of prices. Note that all prices correspond to given values
of the conversion factor κ, but prices for other values of κ are easily obtained
since this is just a scaling factor.
From the results given in Appendix E it is clear that both the adjusted Black
formula and the transformed copula model give much higher quanto prices than
the simple Black model. This is to be expected. The adjusted Black formula
allows the practitioner to compensate somewhat for the increase in implied
volatility of XT2,3 for strikes away from at-the-money (assuming the lowest im-
plied volatility corresponds to the at-the-money strike, which is usually the
case). The transformed copula model allows for the smile in implied volatilities
of options on all three relevant currency pairs.
We now examine the difference between the prices of standard quanto call (or
put) options calculated via the adjusted Black formula and the corresponding
prices calculated under the transformed copula model. Although call and put
quanto prices are generally close, there are sometimes significant differences,
usually with strikes furthest from at-the-money. This difference can be as great
as 15% of the option value for standard strikes. In order to assist practitioners
familiar with the adjusted Black formula, we have also included in our results
the value of the adjusted implied volatility parameter σ̃ (3) that must be entered

26
into the adjusted Black formula to obtain the model quanto price (calculated
under our calibrated transformed Copula model). The adjustment required in
the size of this parameter is usually in the region of 0.1-0.3% (in standard units of
implied volatility) but is occasionally in the region of 1-1.3% for strikes furthest
away from at-the-money.
For the datasets used in this exercise, the transformed copula model usually
gives higher quanto call prices than the adjusted Black formula. The relative
difference commonly increases with the level of the strike of the quanto option.
Often the absolute difference in prices is of similar magnitude across strikes.
The only scenario where model call prices are slightly below adjusted Black
prices corresponds to the currency set-up (JPY,EUR,USD) with maturity 2Y.
There is no significant qualitative difference between the implied volatilities
of the relevant vanilla options corresponding to this maturity and the implied
volatilities for other maturities. In general, the implied volatilities corresponding
to the marginals of this dataset appear slightly more symmetric than those
corresponding to shorter maturities.
The transformed copula model usually leads to lower quanto put prices than
the adjusted Black formula. The greatest differences often occur for strikes away
from at-the-money, though there is not such a clear trend as that observed for
call prices. In a few scenarios, put prices under the transformed copula model
are slightly above those calculated using the adjusted Black formula for some
strikes, but on the whole these relative differences are small (see all results
corresponding to maturity 1M and also (JPY,USD,EUR) 2Y).
Using the results of these scenarios alone, it is not straightforward to deter-
mine exactly under what conditions the difference in model and adjusted Black
prices will be greatest because the adjusted Black formula does not correspond to
a valid model for the joint distribution of XT1,2 and XT1,3 . The transformed copula
model provides quanto prices as a result of a calibration to implied volatilities
on the three relevant FX rates and it is not possible to alter the implied volatil-
ities on one FX rate without adjusting the entire calibration. Further scenario
analysis is required to understand the exact relationship between adjusted Black
prices and model prices for different patterns of the relevant implied volatilities,
however this is beyond the scope of this paper.

6 Conclusion
This paper presents a new methodology for the valuation of quanto FX options,
in which we separate the modelling of the dependence structure of the relevant
FX rates from the modelling of the implied marginal distributions. Our model

27
for the dependence structure is a copula obtained by a suitable perturbation of
the Normal copula. This allows us to modify the upper and lower tail depen-
dence characteristics appropriately to allow calibration to the smile in implied
volatilities on all three relevant FX rates. Most ad-hoc approaches to this prob-
lem do not capture all the relevant information contained in market quotes,
indeed many do not correspond to a proper model for the joint distribution of
the relevant FX rates.
Although our model for the dependence structure (the transformed Normal
copula) is not the only model which may be calibrated to market prices, it
is generally close to the dependence model corresponding to the Black pricing
formula. We adjust the dependence structure in such a way that even though
the calibrated dependence model is not unique, we have a degree of confidence
that any other model calibrated to all market data will give similar values of
the quanto option.
The resulting quanto prices evaluated under a number of real scenarios are
generally close to prices calculated using the ad-hoc adjusted Black formula,
which is certainly a much better proxy for the model price than the standard
Black formula. However, we find the adjusted Black model often gives lower
quanto call prices and higher quanto put prices than those calculated under the
transformed copula model. The relative difference in quanto prices is occasion-
ally large (10-15%) for standard strikes furthest away from at-the-money.
The quanto FX option pricing problem we have focused on in this paper is
a particular application of what is a general method for perturbing a Normal
copula (the dependence model associated with the Black model) to modify the
tail dependence characteristics of the overall dependence structure. A similar
methodology may be applied to any two-asset or hybrid option pricing problem
if vanilla option quotes exist to inform the implied dependence structure. Where
such option quotes are unavailable (as is the case with equity spread options),
historical data may be used to estimate the copula directly (see for example
[Rosenberg, 2003]). However the method outlined in this paper could easily be
adapted so that the transformed Normal copula is estimated from historical
data, allowing construction of a smoother copula that places less reliance on the
quality of the historical data and the estimation methods.
This methodology may also have other applications in the statistical mod-
elling of real-world dependence where the Normal copula does not provide a
sufficiently realistic model for the dependence structure, for example in scenario
analysis or stress-testing.

28
References
[Ait-Sahalia at al, 1998] Ait-Sahalia, Y. & Lo, A.W., 1998, “Nonparametric Estima-
tion of State-Price Densities Implicit in Financial Asset Prices,”
Journal of Finance.

[Bahra, 1997] Bahra, B., 1997, “Implied Risk-Neutral Probability Density Func-
tions from Option Prices: Theory and Application,” Bank of Eng-
land Working Paper.

[Bliss et al, 2000] Bliss, R.R. & Panigirtzoglou, N., 2000, “Testing the stability of
implied probability density functions,” Bank of England Working
Paper.

[Burden et al, 1997] Burden, R.L. & Faires, J.L., 1997, Numerical Analysis, Sixth
Edition, Brooks/Cole.

[Campa et al, 1997] Campa, J.M., Chang, P.H.K. & Reider, R.L., 1997, “Implied Ex-
change Rate Distributions: Evidence from OTC Option Markets,”
Working Paper.

[Coutant et al, 1998] Coutant, S., Jondeau, E. & Rockinger, M., 1998, “Reading In-
terest Rate and Bond Futures Options Smiles: How PIBOR and
Notional Operators Appreciated the 1997 French snap election,”
Banque de France Working Paper.

[Dupire, 1994] Dupire, B., 1994, “Pricing with a Smile,” Risk Magazine 7 (Jan-
uary).

[Embrechts et al, 2001] Embrechts, P., Lindsog, F. & McNeil, A., 2001, “Modelling
Dependence with Copulas and Applications to Risk Management,”
ETHZ Working Paper.

[Genest, 2000] Genest, 2000, “Distributions with Given Marginals and Statistical
Modelling,” Conference Proceedings, Barcelona, July 17-20, 2000.

[Genz, 2002] Genz, A., 2002, “Function for computing bivariate normal proba-
bilities,” Washington State University Working Paper.

[Joe, 1997] Joe, 1997, Multivariate Models and Dependence Concepts, Chap-
man & Hall.

[Malz, 1997] Malz, M., 1997, “Estimating the Probability Distribution of the
Future Exchange Rate from Option Prices,” Journal of Derivatives.

[Moré, 1977] Moré, J.J., 1977, “The Levenberg-Marquardt algorithm: imple-


mentation and theory,” Numerical Analysis, pp. 105-116, Lecture
Notes in Mathematics 630, Springer-Verlag.

[Nelson, 1998] Nelson, R.B., 1998, An Introduction to Copulas, Lecture Notes In


Statistics, Springer-Verlag.

[Rosenberg, 2003] Rosenberg, J.V., 2003, “Non-parametric Pricing of Multivariate


Contingent Claims,” Journal of Derivatives.

29
[Schönbucher at al, 2001] Schönbucher, P.J. & Schubert, D., 2001, “Copula-
Dependent Default Risk in Intensity Models,” Working Paper,
University of Bonn.

[Sherrick at al, 1996] Sherrick, B.J., Garcia, P. & Tirupattur, V., 1996, “Recovering
Probabalistic Information From Option Markets: Tests of Distri-
butional Assumptions,” Journal of Futures Markets.

30
A Calculation of Implied Marginal Density and
Distribution Functions
As outlined in Section 3.1.1, initial calibration of marginal distributions to implied
volatility quotes is by a weighted non-linear least squares optimisation, similarly to
[Bahra, 1997].
i
Fix the maturity T. Suppose the probability density function f Q i,j (·) of the FX
XT
rate XTi,j under the measure Qi is given by a mixture of lognormal distributions
i
f Q i,j (x) = θi,j L(α1i,j , β1i,j | x) + (1 − θi,j )L(α2i,j , β2i,j | x), (14)
XT


where αki,j = µi,j i,j 2 i,j
k − (σk ) T /2, βk = σk
i,j
T , L(·) is the standard lognormal density
function
à !
i,j i,j 1 −(log x − αki,j )2
L(αk , βk | x) = √ exp , k = 1, 2
xβki,j 2π 2(βki,j )2

and θi,j , µi,j i,j i,j


1 , µ2 , σ1 and σ2i,j are parameters to be determined from market quotes
of vanilla option prices as follows.
The price of a vanilla FX call option on XTi,j is given by direct evaluation of the
expectation given in Equation (4):
i,j i,j
+(β1 )2 /2
C0i,j = i
D0,T [θi,j {eα1 Φ(di,j i,j
1,1 ) − KΦ(d2,1 }
i,j i,j
+(β2 )2 /2
+(1 − θi,j ){eα2 Φ(di,j i,j
1,2 ) − KΦ(d2,2 }], (15)

where
− log K + αki,j + (βki,j )2
di,j
1,k = , di,j i,j i,j
2,k = d1,k − βk , k = 1, 2
βki,j
and Φ(·) is the standard Normal cumulative distribution function. Put prices are
obtained similarly.
Values of parameters θi,j , µi,j i,j i,j
1 , µ2 , σ1 and σ2i,j are found by a standard least-
squares optimisation procedure that minimises the weighted squared error between
market quotes for implied volatilities of vanilla options and the corresponding values
i,j
under this model. Since we know the FX forward M0,T is the mean of the distribution
of the FX rate XTi,j under the measure Qi ,
i,j i,j i,j i,j
i,j +(β1 )2 /2 +(β2 )2 /2
M0,T = θi,j eα1 + (1 − θi,j )eα2 , (16)

therefore β2i,j may be found immediately given the other parameters. A computational
requirement of subsequent joint density calculations is that the forward rate implied by
the marginal distribution is exact (at least to around 8 s.f.); the following optimisation
procedure is also found to be far more stable if the forward is not allowed to be a free
parameter (c.f. [Bahra, 1997]).
The distribution function of XTi,j under Qi may be found from the density function
(14) by direct calculation:

log x − α1i,j log x − α2i,j


µ ¶ µ ¶
Qi i,j i,j
F i,j (x) = θ Φ + (1 − θ )Φ .
XT β1i,j β2i,j

31
Given the distribution of XTi,j under Qi it is possible to calculate the distribution
and density functions of the inverse FX rate XTj,i under the corresponding measure
Qj , so it is only necessary to perform a single calibration of model parameters for each
currency pair XTi,j , i > j under Qi .
Let us assume the density of XTi,j under Qi is given by a mixture of lognormal
distributions as outlined above. By no arbitrage

P0j,i (K) = KX0j,i C0i,j (1/K),

hence by differentiating both sides with respect to K we see that


i,j i,j i,j i,j
+(β1 )2 /2 +(β2 )2 /2
j θi,j eα1 Φ(di,j
1,1 ) + (1 − θ
i,j α2
)e Φ(di,j
1,2 )
F Q j,i (K) = i,j i,j i,j i,j ,
XT α1 +(β1 )2 /2 α2 +(β2 )2 /2
θi,j e + (1 − θi,j )e

which is the distribution function of a mixture of two lognormal distributions.

B Copula calculations
The Normal copula has distribution function

CN (u, v | ρ) = Φρ (Φ−1 (u), Φ−1 (v)),

where Φρ is the standard bivariate Normal distribution function with correlation ρ


and Φ is the standard univariate Normal distribution function (for which standard
approximations are available).
Let x = Φ−1 (u) and y = Φ−1 (v). By standard transformation of variables,
µ 2
x + y2

1
cN (u, v | ρ) = (1 − ρ2 )−1/2 exp − (x2
+ y 2
− 2ρxy) . (17)
2 2(1 − ρ2 )

The partial derivative of the Normal copula with respect to one of its arguments is
required in evaluation of the transformed Normal copula density (11). By integration
of (17) we obtain à !
∂C Φ−1 (p) − ρ Φ−1 (q)
(p, q) = Φ .
∂q
p
1 − ρ2
We transform the Normal copula CN in Section 3.3.3 via

Cϕ (u, v) := ϕ−1 (C(ϕ(u), ϕ(v))), (18)

where ϕ : [0, 1] → [0, 1] is a continuous, twice differentiable concave function such that
ϕ(0) = 0 and ϕ(1) = 1. This has density given by the partial derivative of (18) with
respect to both its arguments,
µ
ϕ0 (u)ϕ0 (v)
cϕ (u, v) = c(ϕ(u), ϕ(v))
ϕ0 (Cϕ (u, v))

ϕ00 (Cϕ (u, v)) ∂C ∂C
− 0 (ϕ(u), ϕ(v)) (ϕ(u), ϕ(v)) , (19)
[ϕ (Cϕ (u, v))]2 ∂u ∂v

where c(·, ·) is the density of the original copula C.

32
Following [Genest, 2000], to show the transformation (18) results in a copula, note
that if ϕ is concave then
ϕ00 (ϕ−1 (C(u, v))) ∂C ∂C
(u, v) (u, v) ≤ 0,
[ϕ0 (ϕ−1 (C(u, v)))]2 ∂u ∂v
hence
ϕ00 (ϕ−1 (C(u, v))) ∂C ∂C
c(u, v) − (u, v) (u, v) ≥ 0. (20)
[ϕ0 (ϕ−1 (C(u, v)))]2 ∂u ∂v
Integrating the transformed density (19) between 0 < u1 ≤ u2 < 1 and 0 < v1 ≤ v2 <
1, we obtain
Cϕ (u2 , v2 ) − Cϕ (u1 , v2 ) − Cϕ (u2 , v1 ) + Cϕ (u1 , v1 ) ≥ 0. (21)
Since ϕ(0) = 0 and ϕ(1) = 1, Equation (21) holds for every 0 ≤ u1 ≤ u2 ≤ 1 and
0 ≤ v1 ≤ v2 ≤ 1 with Cϕ (u, 0) = Cϕ (0, v) = 0, Cϕ (u, 1) = u and Cϕ (1, v) = v. Thus
Cϕ is a copula.
Note that from the definition (18), if we assume that Cϕ is a copula, then the
density of the copula (19) is only positive if (20) holds. Thus Equation (20) is actually
a necessary and sufficient condition for Cϕ to be a copula.

C Calibration procedure
The transformed Normal copula allows enough flexibility to fit the required depen-
dence structure implied by market quotes of implied volatilities of vanilla options on
XT2,3 . In our implementation of the calibration procedure outlined in Section 3.3.4,
the transformation function ϕ(·) is constructed from each trial parameter z ∈ Rn us-
ing cubic splines with knot points p = (0, p1 , 21 , p3 , 1)T , where p1 and p3 are given
constants (we have found taking p = (0, 0.1, 0.5, 0.9, 1.0)T allows an effective calibra-
tion to market data in most scenarios). We calibrate our model to a set of implied
volatilities of vanilla options on XT2,3 corresponding to an increasing set of strikes
K = (K0 , K1 , K2 , K3 , K4 )T , where K2 is the at-the-money strike. In our implemen-
tation of the calibration procedure, we have chosen weights w = (2, 10, 20, 10, 2)T in
order to obtain an especially close fit to implied volatilities of options struck near at-
the-money, where options are most liquid. The actual optimisation is performed by a
standard non-linear least-squares algorithm.
For each strike K, the double integral (7) corresponding to each vanilla call option
is estimated by truncating the region of integration at upper and lower quantiles of each
marginal distribution corresponding to a given p-value. Computed implied volatilities
of vanilla options on XT2,3 under the Black model are observed to agree with theoretical
values (as given in Equation (9)) to at least four significant figures when the integral
is evaluated using simple numerical quadrature with a grid of 500 × 500 points and
a p-value of 10−6 . The precision of the approximation is increased by reducing the
p-value and increasing the definition of the grid. Computed values of vanilla option
prices also satisfy put-call parity to at least four significant figures. Clearly there are
more efficient quadrature methods which might be used to approximate this integral,
but this simple method is quite sufficient to allow us to examine the properties of the
copula transformation model.
Suppose we have fixed all knot points and optimisation weights. Defining the cubic
spline according to Equation (12), Section 3.3.4, we see that the second derivative of
this candidate transformation function is piecewise linear.

33
If a2,j < 0 has been determined from the proposed optimisation solution (see
Section 3.3.4), then the resulting cubic spline ϕ(x) is concave since it has negative
second derivative for all x. Similar to the standard construction of cubic splines, by
matching the level and first and second derivatives of the splines where they join at
each knot point we find a number of relations between spline coefficients. If we insist
ϕ(x) passes through (0,0) and (1,1), we find that the system of equation relating
coefficients a2,j to a0,j forms a tridiagonal linear system. This system may be solved
efficiently by Crout factorisation (see, for example [Burden et al, 1997]) and it is then
trivial to recover the remaining spline coefficients.
Constructing ϕ00 (·) as a piecewise linear function does not give such an excellent fit
to data in all cases and does not necessarily lead to smooth distortion of the Normal
copula (see Fig. 8). The following alternative parameterisation may resolve these
issues and does not require specification of any knot points other than that at x =
1
2
, which is fixed at zero. If the second derivative of the transformation function
is constructed by two cubic polynomials defined on [0, 21 ] and [ 12 , 1], with endpoints
of both fixed at ( 21 , 0), then imposing the conditions ϕ(3) ( 12 ) = ϕ(4) ( 12 ) = 0 forces
the transformation function to be smooth and close to zero near 21 . Integrating, the
transformation function ϕ(·) is a peicewise polynomial of order five. The unknown
coefficients may be found by matching the level and the first derivative of ϕ(·) at 12
and incorporating the endpoint conditions ϕ(0) = 0 and ϕ(1) = 1.
To assess differences in quanto prices after the model has been calibrated to market
data, we need to guage the precision of the integration procedures used to calculate
quanto prices using the transformed copula model. Under the Black model quanto
prices are analytical, so it is straightforward to compare these prices with results
calculated by evaluating the double integral of Equation (13) under the corresponding
Normal copula model (with lognormal margins). Testing across several scenarios and
examining quanto options with a range of strikes, we find that the largest relative
difference between the calculated and analytical values is -0.06%. Prices calculated
under the copula model will always be below correct values (at least if the number
of evaluation points used in the numerical integration is not excessive), because in
evaluating (13) we have truncated the double integral at extreme values corresponding
to a given p-value (similar to truncation of integrals in the calibration step above).

34
D Plots

3M 6M
0.17

0.17
Actual Volatility Actual Volatility
Transformed Copula Transformed Copula
0.16

0.16
0.15

0.15
Volatility

Volatility
0.14

0.14
0.13

0.13
0.12

0.12
0.009 0.010 0.011 0.012 0.013 0.009 0.010 0.011 0.012 0.013

Strike Strike

1Y 2Y
0.17

0.17

Actual Volatility Actual Volatility


Transformed Copula Transformed Copula
0.16

0.16
0.15

0.15
Volatility

Volatility
0.14

0.14
0.13

0.13
0.12

0.12

0.009 0.010 0.011 0.012 0.013 0.009 0.010 0.011 0.012 0.013

Strike Strike

Figure 9: Black volatility of options on XT2,3 for the currency triple {USD,EUR,JPY}
under the optimal transformed Normal copula model, compared against market quotes.

35
1M 3M
0.15

0.15
Actual Volatility Actual Volatility
Transformed Copula Transformed Copula
0.14

0.14
0.13

0.13
Volatility

Volatility
0.12

0.12
0.11

0.11
0.10

0.10
1.0 1.1 1.2 1.3 1.4 1.0 1.1 1.2 1.3 1.4

Strike Strike

6M 1Y
0.15

0.15

Actual Volatility Actual Volatility


Transformed Copula Transformed Copula
0.14

0.14
0.13

0.13
Volatility

Volatility
0.12

0.12
0.11

0.11
0.10

0.10

1.0 1.1 1.2 1.3 1.4 1.0 1.1 1.2 1.3 1.4

Strike Strike

2Y
0.15

Actual Volatility
Transformed Copula
0.14
0.13
Volatility

0.12
0.11
0.10

1.0 1.1 1.2 1.3 1.4

Strike

Figure 10: Black volatility of options on XT2,3 for the currency triple {JPY,EUR,USD}
under the optimal transformed Normal copula model, compared against market quotes.

36
1M 3M
0.14

0.14
Actual Volatility Actual Volatility
Transformed Copula Transformed Copula
0.13

0.13
0.12

0.12
Volatility

Volatility
0.11

0.11
0.10

0.10
0.09

0.09
90 100 110 120 130 90 100 110 120 130

Strike Strike

6M 1Y
0.14

0.14

Actual Volatility Actual Volatility


Transformed Copula Transformed Copula
0.13

0.13
0.12

0.12
Volatility

Volatility
0.11

0.11
0.10

0.10
0.09

0.09

90 100 110 120 130 90 100 110 120 130

Strike Strike

2Y
0.14

Actual Volatility
Transformed Copula
0.13
0.12
Volatility

0.11
0.10
0.09

90 100 110 120 130

Strike

Figure 11: Black volatility of options on XT2,3 for the currency triple {EUR,JPY,USD}
under the optimal transformed Normal copula model, compared against market quotes.

37
1M 3M

0.14

0.14
Actual Volatility Actual Volatility
Transformed Copula Call Transformed Copula Call
0.13 Transformed Copula Put Transformed Copula Put

0.13
0.12

0.12
Volatility

Volatility
0.11

0.11
0.10

0.10
0.09

0.09
0.008 0.009 0.010 0.011 0.008 0.009 0.010 0.011

Strike Strike

6M 1Y
0.14

0.14
Actual Volatility Actual Volatility
Transformed Copula Call Transformed Copula Call
Transformed Copula Put Transformed Copula Put
0.13

0.13
0.12

0.12
Volatility

Volatility
0.11

0.11
0.10

0.10
0.09

0.09

0.008 0.009 0.010 0.011 0.008 0.009 0.010 0.011

Strike Strike

2Y
0.14

Actual Volatility
Transformed Copula Call
Transformed Copula Put
0.13
0.12
Volatility

0.11
0.10
0.09

0.008 0.009 0.010 0.011

Strike

Figure 12: Black volatility of call and put options on XT2,3 for the currency triple
{EUR,USD,JPY} under the optimal transformed Normal copula model, compared
against market quotes. Some numerical error is observed in the calculation of model
implied volatilities for high strikes (call and put implied volatilities are not equal)
when |ϕ00 (0)| is very large. The fit may be improved somewhat by altering the knot
points p1 and p3 .

38
{USD,EUR,JPY} {JPY,EUR,USD}
x x
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
0

0
-2
-10

-4
phi’’(x)

phi’’(x)
-20

-6
1M
1M
3M
3M
6M
6M
1Y
-8
1Y
-30

2Y
2Y
-10

{EUR,JPY,USD}
x
0.0 0.2 0.4 0.6 0.8 1.0
0
-2
phi’’(x)

-4

1M
3M
6M
1Y
-6

2Y
-8

Figure 13: Second derivatives of optimal transformation function ϕ(·) for each cur-
rency triple.

39
E Quanto Prices
Here ∆σ̃ (3) (Call) denotes the required adjustment to the σ̃ (3) parameter that results
in equality between the quanto call price calculated with the adjusted Black formula
and that given under the transformed copula model.

Strike 1.1248 1.1491 1.1762 1.2045 1.2304


BS Call 5.51 3.46 1.72 0.65 0.21
Adj BS Call 5.56 3.48 1.72 0.71 0.30
Model Call 5.58 3.48 1.74 0.72 0.31
Call Rel. Err. 0.37% 0.17% 1.08% 1.58% 5.01%
∆σ̃ (3) (Call) 0.31% 0.06% 0.14% 0.10% 0.21%
BS Put 0.19 0.61 1.64 3.45 5.65
Adj BS Put 0.23 0.62 1.63 3.50 5.73
Model Put 0.24 0.61 1.63 3.49 5.73
Put Rel. Err. 1.30% -1.82% 0.09% -0.17% -0.03%
∆σ̃ (3) (Put) 0.05% -0.10% 0.01% -0.05% -0.02%

Table 1: (JPY,EUR,USD) 1M quanto prices.

Strike 1.0927 1.1331 1.1779 1.2265 1.2718


BS Call 9.16 5.79 2.95 1.13 0.37
Adj BS Call 9.25 5.82 2.95 1.22 0.53
Model Call 9.31 5.86 3.01 1.27 0.59
Call Rel. Err. 0.64% 0.63% 1.94% 3.66% 10.28%
∆σ̃ (3) (Call) 0.53% 0.20% 0.25% 0.23% 0.45%
BS Put 0.31 1.02 2.71 5.80 9.62
Adj BS Put 0.41 1.06 2.71 5.89 9.78
Model Put 0.40 1.03 2.71 5.88 9.77
Put Rel. Err. -0.78% -2.48% -0.14% -0.27% -0.03%
∆σ̃ (3) (Put) -0.03% -0.14% -0.02% -0.08% -0.02%

Table 2: (JPY,EUR,USD) 3M quanto prices.

Strike 1.0632 1.1189 1.1792 1.2485 1.3139


BS Call 12.53 7.95 4.18 1.62 0.53
Adj BS Call 12.67 8.01 4.18 1.74 0.75
Model Call 12.80 8.10 4.31 1.84 0.85
Call Rel. Err. 1.03% 1.12% 2.87% 5.38% 11.61%
∆σ̃ (3) (Call) 0.74% 0.30% 0.35% 0.33% 0.50%
BS Put 0.43 1.42 3.69 8.06 13.52
Adj BS Put 0.57 1.48 3.69 8.18 13.73
Model Put 0.55 1.43 3.67 8.13 13.68
Put Rel. Err. -2.55% -3.85% -0.60% -0.58% -0.35%
∆σ̃ (3) (Put) -0.08% -0.20% -0.04% -0.13% -0.19%

Table 3: (JPY,EUR,USD) 6M quanto prices.

40
Strike 1.0233 1.0998 1.1778 1.2777 1.3732
BS Call 16.86 10.74 5.99 2.34 0.77
Adj BS Call 17.06 10.84 5.99 2.52 1.11
Model Call 17.30 11.03 6.22 2.72 1.30
Call Rel. Err. 1.37% 1.72% 3.62% 7.34% 14.64%
∆σ̃ (3) (Call) 1.05% 0.52% 0.49% 0.49% 0.68%
BS Put 0.59 1.97 4.86 10.98 18.78
Adj BS Put 0.79 2.06 4.86 11.17 19.12
Model Put 0.74 1.96 4.80 11.07 19.02
Put Rel. Err. -7.36% -5.08% -1.34% -0.82% -0.53%
∆σ̃ (3) (Put) -0.26% -0.28% -0.14% -0.23% -0.38%

Table 4: (JPY,EUR,USD) 1Y quanto prices.

Strike 0.9693 1.0733 1.1622 1.3064 1.4466


BS Call 21.58 13.75 8.58 3.37 1.13
Adj BS Call 21.85 13.88 8.58 3.63 1.62
Model Call 21.86 13.80 8.56 3.61 1.65
Call Rel. Err. 0.05% -0.58% -0.19% -0.61% 2.12%
∆σ̃ (3) (Call) 0.04% -0.16% -0.03% -0.04% 0.09%
BS Put 0.80 2.71 5.87 14.17 25.07
Adj BS Put 1.06 2.84 5.87 14.43 25.55
Model Put 1.11 2.79 5.89 14.44 25.61
Put Rel. Err. 3.85% -1.75% 0.24% 0.05% 0.24%
∆σ̃ (3) (Put) 0.15% -0.10% 0.02% 0.01% 0.16%

Table 5: (JPY,EUR,USD) 2Y quanto prices.

Strike 115.59 117.53 119.73 121.94 123.99


BS Call 0.0430 0.0271 0.0134 0.0052 0.0017
Adj BS Call 0.0432 0.0270 0.0134 0.0059 0.0028
Model Call 0.0433 0.0272 0.0134 0.0059 0.0031
Call Rel. Err. 0.23% 0.55% 0.56% 0.32% 8.11%
∆σ̃ (3) (Call) 0.17% 0.14% 0.06% 0.02% 0.31%
BS Put 0.0014 0.0045 0.0123 0.0257 0.0423
Adj BS Put 0.0016 0.0044 0.0123 0.0265 0.0434
Model Put 0.0015 0.0044 0.0122 0.0263 0.0435
Put Rel. Err. -4.79% -0.53% -0.79% -0.58% 0.18%
∆σ̃ (3) (Put) -0.13% -0.02% -0.07% -0.13% 0.10%

Table 6: (EUR,USD,JPY) 1M quanto prices.

Strike 111.60 115.00 118.91 122.80 126.54


BS Call 0.0767 0.0488 0.0243 0.0098 0.0032
Adj BS Call 0.0775 0.0489 0.0243 0.0109 0.0054
Model Call 0.0778 0.0494 0.0247 0.0113 0.0060
Call Rel. Err. 0.48% 0.99% 1.65% 2.87% 9.33%
∆σ̃ (3) (Call) 0.35% 0.28% 0.18% 0.16% 0.39%
BS Put 0.0023 0.0077 0.0215 0.0450 0.0750
Adj BS Put 0.0031 0.0078 0.0215 0.0462 0.0772
Model Put 0.0028 0.0076 0.0212 0.0458 0.0771
Put Rel. Err. -12.63% -3.06% -1.47% -0.87% -0.21%
∆σ̃ (3) (Put) -0.34% -0.13% -0.13% -0.20% -0.12%

Table 7: (EUR,USD,JPY) 3M quanto prices.

41
Strike 107.47 112.22 117.78 123.18 128.62
BS Call 0.1087 0.0697 0.0350 0.0146 0.0049
Adj BS Call 0.1104 0.0703 0.0350 0.0162 0.0082
Model Call 0.1112 0.0714 0.0359 0.0172 0.0090
Call Rel. Err. 0.76% 1.53% 2.46% 5.62% 8.66%
∆σ̃ (3) (Call) 0.55% 0.44% 0.27% 0.33% 0.38%
BS Put 0.0032 0.0106 0.0302 0.0627 0.1062
Adj BS Put 0.0048 0.0112 0.0302 0.0642 0.1095
Model Put 0.0042 0.0108 0.0297 0.0637 0.1088
Put Rel. Err. -14.43% -3.42% -1.92% -0.78% -0.63%
∆σ̃ (3) (Put) -0.42% -0.15% -0.18% -0.17% -0.34%

Table 8: (EUR,USD,JPY) 6M quanto prices.

Strike 101.15 107.65 115.47 122.55 130.42


BS Call 0.1514 0.0980 0.0494 0.0224 0.0077
Adj BS Call 0.1543 0.0993 0.0494 0.0246 0.0125
Model Call 0.1571 0.1023 0.0521 0.0275 0.0149
Call Rel. Err. 1.77% 2.93% 5.30% 10.48% 16.13%
∆σ̃ (3) (Call) 1.25% 0.88% 0.61% 0.70% 0.78%
BS Put 0.0041 0.0142 0.0420 0.0842 0.1464
Adj BS Put 0.0071 0.0155 0.0420 0.0865 0.1512
Model Put 0.0059 0.0146 0.0409 0.0854 0.1497
Put Rel. Err. -18.61% -6.45% -2.79% -1.21% -1.02%
∆σ̃ (3) (Put) -0.55% -0.28% -0.25% -0.25% -0.52%

Table 9: (EUR,USD,JPY) 1Y quanto prices.

Strike 90.28 99.20 110.54 118.73 130.46


BS Call 0.2116 0.1385 0.0691 0.0374 0.0133
Adj BS Call 0.2157 0.1405 0.0691 0.0400 0.0205
Model Call 0.2195 0.1447 0.0731 0.0443 0.0238
Call Rel. Err. 1.74% 2.93% 5.46% 9.82% 13.90%
∆σ̃ (3) (Call) 1.33% 0.95% 0.66% 0.74% 0.73%
BS Put 0.0053 0.0193 0.0605 0.1086 0.1989
Adj BS Put 0.0095 0.0212 0.0605 0.1112 0.2062
Model Put 0.0090 0.0212 0.0602 0.1113 0.2052
Put Rel. Err. -5.14% -0.28% -0.43% 0.08% -0.47%
∆σ̃ (3) (Put) -0.18% -0.01% -0.04% 0.01% -0.22%

Table 10: (EUR,USD,JPY) 2Y quanto prices.

Strike 0.009304 0.009548 0.009824 0.010112 0.010378


BS Call 4.6E-04 2.9E-04 1.4E-04 5.5E-05 1.8E-05
Adj BS Call 4.6E-04 2.9E-04 1.4E-04 6.5E-05 3.1E-05
Model Call 4.6E-04 2.9E-04 1.5E-04 6.6E-05 3.2E-05
Call Rel. Err. 0.25% 0.44% 0.95% 1.85% 2.77%
∆σ̃ (3) (Call) 0.28% 0.17% 0.14% 0.15% 0.16%
BS Put 1.6E-05 5.2E-05 1.4E-04 3.0E-04 4.8E-04
Adj BS Put 1.6E-05 4.9E-05 1.4E-04 3.1E-04 5.0E-04
Model Put 1.6E-05 4.9E-05 1.4E-04 3.1E-04 5.0E-04
Put Rel. Err. 0.34% 0.30% 0.19% 0.04% -0.05%
∆σ̃ (3) (Put) 0.01% 0.02% 0.03% 0.01% -0.04%

Table 11: (USD,EUR,JPY) 1M quanto prices.

42
Strike 0.009076 0.009468 0.009905 0.010382 0.010829
BS Call 7.4E-04 4.6E-04 2.4E-04 9.1E-05 2.9E-05
Adj BS Call 7.4E-04 4.6E-04 2.4E-04 1.0E-04 5.5E-05
Model Call 7.5E-04 4.7E-04 2.4E-04 1.1E-04 5.9E-05
Call Rel. Err. 0.67% 1.09% 2.21% 4.54% 6.99%
∆σ̃ (3) (Call) 0.01% 0.16% 0.23% 0.30% 0.37%
BS Put 2.6E-05 8.4E-05 2.2E-04 4.8E-04 7.9E-04
Adj BS Put 3.2E-05 8.3E-05 2.2E-04 4.9E-04 8.2E-04
Model Put 3.2E-05 8.2E-05 2.2E-04 4.9E-04 8.2E-04
Put Rel. Err. -0.79% -0.17% 0.04% -0.06% -0.14%
∆σ̃ (3) (Put) -0.06% -0.05% -0.08% -0.25% -0.68%

Table 12: (USD,EUR,JPY) 3M quanto prices.

Strike 0.008916 0.009440 0.010011 0.010672 0.011299


BS Call 9.8E-04 6.2E-04 3.3E-04 1.3E-04 4.1E-05
Adj BS Call 9.9E-04 6.2E-04 3.3E-04 1.4E-04 7.7E-05
Model Call 1.0E-03 6.3E-04 3.4E-04 1.5E-04 8.7E-05
Call Rel. Err. 1.18% 1.53% 3.33% 6.24% 10.86%
∆σ̃ (3) (Call) 1.01% 0.52% 0.48% 0.47% 0.63%
BS Put 3.5E-05 1.1E-04 3.0E-04 6.5E-04 1.1E-03
Adj BS Put 4.8E-05 1.2E-04 3.0E-04 6.6E-04 1.1E-03
Model Put 4.7E-05 1.1E-04 2.9E-04 6.6E-04 1.1E-03
Put Rel. Err. -2.32% -2.82% -0.55% -0.49% -0.31%
∆σ̃ (3) (Put) -0.10% -0.17% -0.07% -0.16% -0.24%

Table 13: (USD,EUR,JPY) 6M quanto prices.

Strike 0.008735 0.009463 0.010212 0.011178 0.012111


BS Call 1.3E-03 8.5E-04 4.7E-04 1.8E-04 6.1E-05
Adj BS Call 1.3E-03 8.5E-04 4.7E-04 2.1E-04 1.1E-04
Model Call 1.4E-03 8.7E-04 4.9E-04 2.2E-04 1.2E-04
Call Rel. Err. 1.43% 1.96% 3.82% 7.60% 12.14%
∆σ̃ (3) (Call) 1.16% 0.65% 0.56% 0.58% 0.70%
BS Put 4.9E-05 1.6E-04 4.0E-04 9.0E-04 1.5E-03
Adj BS Put 7.0E-05 1.7E-04 4.0E-04 9.2E-04 1.6E-03
Model Put 6.8E-05 1.6E-04 3.9E-04 9.2E-04 1.6E-03
Put Rel. Err. -3.04% -2.83% -0.73% -0.51% -0.41%
∆σ̃ (3) (Put) -0.13% -0.18% -0.09% -0.16% -0.31%

Table 14: (USD,EUR,JPY) 1Y quanto prices.

Strike 0.008625 0.009686 0.010604 0.012112 0.013601


BS Call 1.8E-03 1.2E-03 7.2E-04 2.8E-04 9.4E-05
Adj BS Call 1.9E-03 1.2E-03 7.2E-04 3.0E-04 1.5E-04
Model Call 1.9E-03 1.2E-03 7.5E-04 3.3E-04 1.8E-04
Call Rel. Err. 1.79% 1.95% 3.96% 7.86% 14.74%
∆σ̃ (3) (Call) 1.36% 0.61% 0.64% 0.60% 0.83%
BS Put 7.2E-05 2.4E-04 5.2E-04 1.3E-03 2.2E-03
Adj BS Put 1.2E-04 2.6E-04 5.2E-04 1.3E-03 2.3E-03
Model Put 1.1E-04 2.5E-04 5.2E-04 1.3E-03 2.3E-03
Put Rel. Err. -3.14% -6.20% -1.69% -1.00% -0.56%
∆σ̃ (3) (Put) -0.15% -0.41% -0.19% -0.30% -0.43%

Table 15: (USD,EUR,JPY) 2Y quanto prices.

43
F Market Data
The following market data used to calibrate the copula model was obtained at the
close of 7 July 2001. X0EUR,JPY = 0.009781 and X0EUR,USD = 1.175226.

T EUR USD JPY


1
12 0.996243 0.996661 0.999948
1
4 0.988734 0.990054 0.999805
1
2 0.978343 0.980503 0.999566
1 0.958074 0.959798 0.998965
2 0.916243 0.911070 0.997404

i
Table 16: Discount factors D0,T corresponding to currency i and maturity T.

T (Puts) Delta (Calls)


10% 20% 50% 20% 10%
1
12 0.1289 0.1230 0.1215 0.1260 0.1340
1
4 0.1277 0.1210 0.1190 0.1235 0.1319
1
2 0.1272 0.1203 0.1180 0.1223 0.1305
1 0.1269 0.1196 0.1170 0.1216 0.1303
2 0.1249 0.1176 0.1150 0.1196 0.1283

Table 17: EUR/USD implied volatility quotes corresponding to standard values of


Black delta and maturity T.

T (Puts) Delta (Calls)


10% 20% 50% 20% 10%
1
12 0.1467 0.1441 0.1480 0.1606 0.1747
1
4 0.1466 0.1372 0.1385 0.1487 0.1667
1
2 0.1444 0.1329 0.1320 0.1414 0.1597
1 0.1443 0.1320 0.1300 0.1380 0.1557
2 0.1508 0.1358 0.1308 0.1358 0.1508

Table 18: EUR/JPY implied volatility quotes corresponding to standard values of


Black delta and maturity T.

T (Puts) Delta (Calls)


10% 20% 50% 20% 10%
1
12 0.1109 0.1015 0.0950 0.0945 0.0990
1
4 0.1156 0.1045 0.0985 0.0995 0.1066
1
2 0.1178 0.1055 0.1000 0.1024 0.1122
1 0.1186 0.1065 0.1010 0.1048 0.1166
2 0.1235 0.1105 0.1060 0.1105 0.1235

Table 19: USD/JPY implied volatility quotes corresponding to standard values of


Black delta and maturity T.

44

S-ar putea să vă placă și