Introduction into “Local Correlation” Modelling

Alex Langnau
Allianz SE and LMU Munich

May 28, 2018

arXiv:0909.3441v3 [q-fin.PR] 22 Sep 2009

Abstract
In this paper we provide evidence that financial option markets for equity indices
give rise to non-trivial dependency structures between its constituents. Thus, if the in-
dividual constituent distributions of an equity index are inferred from the single-stock
option markets and combined via a Gaussian copula, for example, one fails to explain
the steepness of the observed volatility skew of the index. Intuitively, index option
prices are encoding higher correlations in cases where the option is particularly sensi-
tive to stress scenarios of the market. As a result, more complex dependency structures
emerge than the ones described by Gaussian copulas or (state-independent) linear cor-
relation structures. In this paper we “decode” the index option market and extract
this correlation information in order to extend the multi-asset version of Dupire’s “lo-
cal volatility” model by making correlations a dynamic variable of the market. A
“local correlation” model (LCM) is introduced for the pricing of multi-asset deriva-
tives. LCM achieves consistency with both the constituent- and index option markets
by construction while preserving the efficiency and easy implementation of Dupire’s
model.

Keywords: implied correlation, local correlation, stochastic correlation, correlation

skew, index skew, basket options, multi-asset Dupire model, multi-asset local vol

1
1 Introduction
In 2008, the crisis in credit markets led to a spike in correlations that resulted in large losses
for several single-stock exotic trading books. As most market participants did not anticipate
the crises, one is tempted to also put these equity losses under the category of unpredictable
and unforeseen “bad” events. We do not fully subscribe to this view.

In this paper we argue that, at least to some extent, the losses could have been allevi-
ated if models had incorporated non-trivial correlation information, already encoded in the
index option market, properly into the pricing of multi-asset derivatives. This is because the
index volatility market already implies rising correlations in downturn scenarios.
Most investment banks utilise a Gaussian copula for the pricing of multi-asset exotic
European payoffs. For more involved payoffs a multi-asset version of Dupire’s local volatil-
ity model has become the market standard for the pricing and hedging of those products.
The model infers a local spot-dependent volatility surface from option prices for each as-
set and combines them via a (constant) correlation matrix between its Wiener increments.
However even when correlations are assumed time-dependent, the model does not incorpo-
rate all dependency information inherent in the derivatives market properly. In particular
with downward scenarios, index options imply correlations that can be substantially higher
compared to the “at-the-money” case. As a result the cost of short cross-gamma position
hedging could be priced in advance, more appropriately, by means of a better model.
In chapter 2 of this paper we take a snapshot of the options market and deduce the
distribution of the Dow Jones Euro Stoxx 50 (.STOXX50E) and DAX (.GDAXI) indices
together with all the composite distributions. In both cases we show that only half of the
volatility skew of the index is explained by the skew of the index constituents. We conclude
that (term) correlation effectively depends on the strike of the option which substantially
impacts the distribution of the index on the downside.
It is important that multi-asset derivative models incorporate these effects even if a
product is only sensitive to a subset of stocks in the index. This is because the picture
remains fundamentally the same: stocks correlate more strongly in a bear market, when
investment behaviour is driven by fear, than during bullish times. As a result portfolio
managers may find themselves less diversified, even when the composition of their portfolio
stays the same. As far as hedging of derivatives is concerned, an increase in correlations
has the unpleasant effect of increasing the cost of managing short cross-gamma positions.
This arises as a result of the non-linear behaviour of (multi-asset) options’ prices and man-
ifests itself in the fact that the option’s sensitivity to one asset increases with the value
of a second decreasing. During the credit crunch for example, the Nikkei often finished
down for the day creating a long delta position in EuroStoxx 50 for a hedging book in
London, even though the market there was closed. As a result, the trader would need
to immediately sell EuroStoxx 50 at the open. However EuroStoxx 50 would be likely to
open down as well because of the large positive correlation between the two assets. This
gives rise to systematic losses during periods when the correlation exceeds expectations.

The goal of this paper is to develop a model that incorporates higher correlations in down-
ward scenarios by “decoding” the state-dependent correlation information of the index option

2
market. One expects that such a model anticipates losses such as those described above at
the inception of the trade and adjusts the pricing accordingly. In addition one would expect
more accurate deltas in this case.
In order to make LCM a model that can be useful for practical applications, one needs
to provide a robust as well as efficient implementation for the underlying algorithms. This
is the strenght of Dupire’s local volatility model which is simple to implement, numerically
sufficiently efficient and provides consistency with the volatility skew of individual option
markets at the same time. The goal of this paper is to extend this methodology to the
multi-asset case by providing consistency with the index option market in addition to the
individual option markets. Note that, similar to the Dupire framework, LCM provides a
particular way of achieving consistency with the option markets, however this solution is
by no means unique. Chapter 7 provides alternative solutions to the same problem which
allows one to quantify ”residual exotic correlation risk”’.

Work in this direction has also been pursued by others. See for example [1],[2],[3],[4],[5].

Chapter 3 lays down the ground work for the “local correlation” model (LCM) that is
introduced in Chapter 4. A parameterisation of correlations is introduced that preserves the
positive semi-definiteness and displays some notion of strictly increasing correlation matrices.
Chapter 5 provides simulation results of LCM and provides explicit results for the strike
dependence of correlation. Chapter 6 deals with some implementation details and chapter 7
discusses residual correlation risk. Chapter 8 concudes our analysis.

3
 Ticker Wgt Ticker Wgt Ticker Wgt Ticker Wgt Ticker Wgt
AEGN.AS 2.67 AIRP.PA 0.50 ALVG.DE 0.86 ALSO.PA 0.38 ISPA.AS 1.76
GASI.MI 2.26 AXAF.PA 3.17 BBVA.MC 7.12 SAN.MC 15.50 BASF.DE 1.75
BAYG.DE 1.57 BNPP.PA 1.66 CARR.PA 1.16 SGOB.PA 0.76 CAGR.PA 1.93
DAIGn.DE 1.69 DBKGn.DE 1.18 DB1Gn.DE 0.37 DTEGn.DE 5.66 EONGn.DE 3.80
ENEI.MI 12.11 ENI.MI 4.63 FOR.BR 4.16 FTE.PA 3.61 GSZ.PA 2.53
DANO.PA 1.14 IBE.MC 6.54 ING.AS 3.95 ISP.MI 18.33 PHG.AS 1.85
OREP.PA 0.45 LVMH.PA 0.49 MUVGn.DE 0.38 NOK1V.HE 7.12 RENA.PA 0.38
REP.MC 1.50 RWEG.DE 0.77 SASY.PA 1.99 SAPG.DE 1.66 SCHN.PA 0.49
SIEGn.DE 1.55 SOGN.PA 1.10 TLIT.MI 19.43 TEF.MC 7.79 TOTF.PA 3.99
CRDI.MI 30.25 UNc.AS 2.95 SGEF.PA 0.94 VIV.PA 2.33 VOWG.DE 0.16

Table 1: Decomposition of EuroStoxx 50 on the 31 July 2009. The first column displays the
(Reuters) names of the assets followed by the weight in the index.

2 “Decoding” the Index Options Market

In this chapter we take a snapshot of the market as of 31 July 2009 and try to explain
the observed index option prices of EuroStoxx 50 and DAX in terms of their constituent
distributions. One finds that the skewness of the individual distributions fails to fully explain
the volatility skew of the index basket itself. Hence, the index volatility market encodes non-
trivial dependencies between its constituents. Let IT denote the value of the index at time
T . The price of a call option on the index with maturity T and strike K is given by C(T, K).
Note that

Z ∞
Q +
C(T, K) = DF E ((IT − K) ) = DF dI prob(T, I)(I − K)+ (1)
0

where DF is the discount factor to maturity and E Q denotes the expectation value under
the forward adjusted measure Q. Thus, Eq.1 relates option prices to the (forward-adjusted)
implied density prob(T, IT ) that is inferred by the option market with maturity T. Assume
that for a given maturity T all strikes are traded. In this case Eq.1 can easily be inverted
with the result
1 ∂ 2 C(T, K)
= prob(T, K). (2)
DF ∂K 2
Let us denote by Si (T ), i = 1, .., n, the price of constituent i at time T . The index is generally
defined as a weighted sum of its constituents e.g.
n
X
It = αi Si (t). (3)
i=1

Examples for basket weights are given in Table 1. For the DAX the weights are listed in
Table 2.
Theoretically the decomposition could change during the life of the option. However these
events are relatively rare so that we ignore potential effects for the pricing in this paper. Let

4
 Ticker Wgt Ticker Wgt Ticker Wgt
ADSG.DE 2.20 ALVG.DE 5.16 BASF.DE 10.46
BMWG.DE 3.66 BAYG.DE 9.41 BEIG.DE 0.93
CBKG.DE 8.20 DAIGn.DE 9.72 DBKGn.DE 7.04
DB1Gn.DE 2.12 DPWGn.DE 9.57 DTEGn.DE 33.91
EONGn.DE 18.76 FMEG.DE 2.14 FREGp.DE 0.92
HNRGn.DE 0.68 HNKGp.DE 2.03 SDFG.DE 1.40
LING.DE 1.92 LHAG.DE 5.21 MANG.DE 1.13
MRCG.DE 0.74 MEOG.DE 1.26 MUVGn.DE 2.25
RWEG.DE 4.71 SZGG.DE 0.43 SAPG.DE 9.63
SIEGn.DE 9.28 TKAG.DE 3.79 VOWG.DE 0.98

Table 2: Decomposition of DAX on the 31 July 2009. The first column displays the (Reuters)
names of the assets followed by the weight in the index.

us denote by Ci (K, T ) the options prices of the i-th constituent. Similar to before, these
prices are observed in the market so that individual distribution can be extracted according
to
1 ∂ 2 Ci (T, K)
= probi (T, K). (4)
DF ∂K 2
The integration of Eq.2 and Eq.4 yields for the cumulative distributions

1 ∂C(T, K)
P (IT < K) = 1 + (5)
DF ∂K
1 ∂Ci (T, K)
Pi (Si (T ) < K) = 1 + . (6)
DF ∂K
Let ω denote a n-dimensional standard normal correlated Gaussian variate with

E Q (ωi ωj ) = ρij i 6= j

E Q (ωi ) = 0 (7)

E Q (ωi2 ) = 1

for a given correlation matrix ρij . If one assumes a Gaussian copula one can construct the
joint distribution from the constituents by inversion of

ni ≡ N (ωi ) = Pi (Si (T ) < Ki ) i = 1, . . . , n. (8)

The resulting multivariate distribution is denoted by

prob(T, S1 (T ), S2 (T ), . . . , Sn (T )).

Hence, index option prices can be computed in two different ways:

1. by means of Eq.1 representing the market quotes of the index option

5
 Two−year Implied Volatility for .STOXX50E

0.34
●
0.32

●
0.30
0.28

●
volatility

0.26

●
0.24

●
0.22

● Market Implied Volatility

0.20

Gaussian Copula Implied Volatility

0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3

strikes relative to spot

Figure 1: Comparison of implied volatilities for the two year option market of EuroStoxx 50
between the market quotes and option prices obtained via multidimensional integration
assuming a Gaussian copula. Only (roughly) half the skew is accounted for by the skew
of the individual distributions.

2. by means of multidimensional integration over its constituents, e.g.

Z ∞ ( n )+
X
C(K,
e T ) = DF dS1 ...dSn prob(T, S1 , S2 , . . . , Sn ) αk Sk − K (9)
0 k=1

If a Gaussian copula was to represent an adequate picture for combining individual spot
distributions one should find

C(K, T ) = C(K,
e T) ∀K, T . (10)

We have carried out this calculation, with some of the results depicted in Fig.1, Fig.2,
Fig.3. However instead of comparing option prices or distributions directly it is more mean-
ingful to compare implied volatilities instead.
Fig.1 shows that individual option distributions fail to explain the volatility skew of the
index for both EuroStoxx 50. Only roughly half of the skew can be attributed to the skew
of the individual assets themselves. Note that this inconsistency does not necessarily imply
dispersion arbitrage opportunities between the index option market and the options markets
of and the individual assets. It is more reasonable to conclude that the option market data

6
 Four−year Implied Volatility for .STOXX50E

0.34
0.32

●
0.30

●
0.28

●
volatility

●
0.26

●
0.24

●
0.22

● Market Implied Volatility

0.20

Gaussian Copula Implied Volatility

0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3

strikes relative to spot

Figure 2: Comparison of implied volatilities for the 4-year option market of EuroStoxx 50 and
results obtained by integration of individual distributions that were combined by a
Gaussian copula.

suggests a type of dependency between individual stocks that is more complex than the one
described by a Gaussian copula.
In order to study whether the effect is persistent in time Fig.2 displays the result of
the 4-year market. The qualitative picture remains the same: Deviations from normality of
individual log-returns cannot explain the non-normality of log-returns of the index.

The results presented so far could represent an anomaly of EuroStoxx 50. However Fig.3
displays the result of calculations for the DAX that are qualitatively similar.
Fig.1,2,3 show that individual option distributions fail to explain the volatility skew of the
index for both EuroStoxx 50 and DAX. Only roughly half of the skew can be attributed to
the skew of the individual assets themselves. Also, other choices for the correlation matrix
in Eq.7 fail to generate enough skew. Possible improvements by “scanning” the space of
correlation matrices would be marginal at best. Note that reasonable choices for correlation
matrices should result in basket implied volatilities that match the index’s implied volatility
curve around the most liquid point, which typically occurs at a strike of about 100%.
Thus, the option market data strongly suggests the behaviour of dependencies between
individual stocks in the index are more complex than the one described by a Gaussian copula.

7
 Two−year Implied Volatility for .GDAXI
0.34
0.32

●
0.30

●
0.28
volatility

●
0.26

●
0.24

●
0.22

● Market Implied Volatility

0.20

Gaussian Copula Implied Volatility

0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3

strikes relative to spot

Figure 3: Comparison of implied volatilities for the 4-year option market of DAX and result
obtained by integration of individual distributions that were combined by a Gaussian
copula.

8
3 One parameter family of correlation matrices
In this chapter we lay down the groundwork for the local correlation model (LCM) described
in the next chapter. We define a family of perturbations of a given (center) correlation matrix
that preserves all the relevant properties of the system, such as its positive semi-definiteness.
The construction of this family is motivated by Poisson jump processes (common jumps gen-
erally increases the correlation between assets while uncorrelated jumps cause de-correlation)
even though jumps do not play an explicit role in the LCM model. The important result of
this chapter lies in the fact that a one-parameter family of correlations can be defined that
allows one to continuously dial correlations in between the identity matrix and a correlation
matrix that is identical to one, i.e.ρij = 1.
Definition 3.1. Let V (n) be a n-dimensional vector space in R with a canonical basis given
by {ei : i = 1, . . . , n}. A linear operator that is represented by a n × n matrix C in the basis
{ei } is called a correlation matrix if
(i) C is positive semi-definite, i.e. ν T Cν ≥ 0 ∀ν ∈ V (n)
(ii) All diagonal elements of C are one , i.e. Cii = 1 i = 1, . . . , n
Remark 3.1. Since any projection of C onto the subspace of V (n) is positive semi-definite,
for any pair i, j ∈ {1, . . . , n} the 2 × 2 matrix M defined by

Mk(ĩ),k(j̃) = eTĩ Cej̃ ĩ, j̃ ∈ {i, j} , k(ĩ) = 1 + 1ĩ=j

is a correlation matrix as well. As detM ≥ 0 it follows that −1 ≤ Cĩj̃ = Mk(ĩ)k(j̃) ≤ 1 .

Hence from Definition 3.1 it follows that all off-diagonal elements of C must be between -1
and 1.
Lemma 3.1. Let V (n) be a n-dimensional vector space in R and ρ a correlation matrix .
Let u ∈ V (n) be an arbitrary vector and κ ∈ {0, 1}. The matrix defined by

 √ ρij +κui uj i 6= j ∈ {1, . . . , n}
 (1+u2i )(1+u2j )

ρ̂ij ≡ (11)


 1 i=j

is a correlation matrix.
Proof: In order to prove Lemma 3.1, it suffices to construct, for a fixed time T = 1, n
random variables XTi i = 1, ..., n with
0 < E Q [(XTi − E Q (XTi ))2 ] < ∞ and E Q [(XTi − E Q (XTi ))(XTj − E Q (XTi ))] < ∞
such that
E Q [(XTi − E Q (XTi ))(XTj − E Q (XTj ))]
q (12)
E Q [(XTi − E Q (XTi ))2 ]E Q [(XTj − E Q (XTj ))2 ]

9
yields Eq.11. Let (Ω, Ft , Q) be a probability space where the filtration Ft is generated by
a n-dimensional correlated Brownian motion ωt together with two Poisson process J1 =
(1) (2) (i)
Nt ν1 , J2 = Nt ν2 of (the same) intensity λ where Nt i = 1, 2 is the number of jumps of
Poisson process i that occurred between 0 and t and νi is the jump size. ωt is given by

d < ωti , ωtj >= ρij (t) dt (13)

If one defines
(i)
Xti ≡ ωti − Nt νi
It is straightforward to show that

E[XTi ] = −λT νi and E[(XTi )2 ] = T + (νi )2 λT (1 + λT ).

(1) (2)
In the case where both jumps are triggered simultaneously, e.g. Nt = Nt , a direct
computation of Eq. 12 yields
ρij + λνi νj
ρ̂ij ≡ q (14)
(1 + λνi2 )(1 + λνj2 )

In the case where both jumps are independent one finds

ρij
ρ̂ij ≡ q (15)
(1 + λνi2 )(1 + λνj2 )

Equations 14 and 15 can be combined with the result

ρij + κui uj
ρ̂ij ≡ q (16)
(1 + u2i )(1 + u2j )
√
where κ = {0, 1} and ui ≡ λνi 

Remark 3.2. Equations 14 and 15 are consistent with intuition: In cases where sgn(νi ) =
sgn(νj ), correlated jumps increase term correlations between two random variables whereas
uncorrelated jumps tend to decrease correlations (in absolute terms).

Remark 3.3. One can reduce Equation 16 to a 1-parameter family of correlations with
parameter u ∈ R by introducing a (fixed) “principal mode” ξ ∈ V (n) to describe perturbations
ui = u ξi of the correlation matrix, e.g.
 2
 √ ρij +κu ξi ξj i 6= j ∈ {1, . . . , n}
 (1+ξi2 u2 )(1+ξj2 u2 )

ρ̂ij (u) ≡ (17)


 1 i=j

This motivates the following

10
Definition 3.2. The set
 
2
 ρij + κu ξi ξj 
Fκ (ρ, ξ) ≡ ρ̂ij (u) : u ∈ R; κ ∈ {0, 1}; ξ ∈ V (n) ; ρ̂ij (u) ≡ q i 6= j
 (1 + ξi2 u2 )(1 + ξj2 u2 ) 

(18)

where ρ̂ii = 1 is called a one-dimensional family of correlation matrices with center ρ(u) =
(ρ)ij i, j = 1, ...n and mode ξ.

Remark 3.4. In the special case of a flat mode, i.e. ξ = (1, 1, . . . , 1)T Equation 18 reduces
to
 ρ +κu2
ij
 (1+u2 )
 i 6= j ∈ {1, . . . , n}
ρ̂ij (u) ≡ (19)

1 i=j


This particular form will prove very useful for the construction of the LCM model because
of the analytic tractibility that goes along with it. This will turn out to be crucial if one
wants to construct a model that will be of practical use.

11
4 Local Correlation model (LCM)
The goal of this chapter is to extend the classic multi-asset Dupire local volatility model
by introducing the notion of local correlations into the dynamics between the assets. Local
correlations imply that, during an infinitesimal time-step, assets evolve according to a cor-
relation matrix that depends on the current state of the system and generally varies from
one time-step to another. The main result of this chapter is the derivation of a particular
functional form of the correlation matrix that achieves consistency with the index skew by
construction. This is crucial if one wants to avoid lengthy fitting procedures of correlation
parameters.

Assumption 4.1. Let (Ω, Ft , Q) be a probability space where the filtration Ft is generated
by a n-dimensional correlated Brownian motion Wt = (ωt1 , ..., ωtn ) e.g. Ft = F W where

d < ωti , ωtj >= ρij (t) dt (20)

ρij (t) is a correlation matrix with ρij (t) ≥ 0. There exist n assets Si (t) i = 1, ..., n that
evolve according to the following local volatility dynamics:

dSti
= µit dt + σi (t, Sti ) dωti i = 1, . . . , n (21)
Sti

where µit denotes a deterministic drift. The zero index defines a basket

n
X
St0 ≡ αi Sti (22)
i=1

together with
n
X
dSt0 = αi dSti (23)
i=1

for some given basket weights 0 ≤ αi < ∞. Locally there exists a Brownian motion ωt0

that matches the dynamics Eq.23 in a weak sense, e.g.

dSt0
0
= µ0t dt + σ0 (t, St0 )dωt0 (24)
St

12
 Furthermore for each asset i = {0, ..., n} there exists a market that trades plain vanilla
European options at all maturities T < T ∗ < ∞ and strikes K < K ∗ ≤ ∞. The set of call
options with price Cti (T, K) at time t of the i-th asset is denoted by

T < T ∗; K ≤ K ∗;
 i
Ct (T, K) : i = 0, ..., n (25)
where

∂Cti (T, K) ∂ 2 Cti (T, K)

, i = 0, ..., n (26)
∂T ∂K 2
exist and are continuous. For a single economy εi (i = 0, ..., n) consisting of a determin-
RT
stic bond B(t, T ) = exp(− t (rs ds)) together with Sti , i.e. εi = {B(t, T ), Sti }, there exists
an equivalent martingale measure (EMM) Qi .
Lemma 4.1. Under assumption 4.1, each single option market Eq.25 has no arbitrage if the
following statements are correct:

(i)
µit = rt − qti (27)

where rt denotes the (deterministic) instantaneous interest rate and qti the dividend yield.

(ii)
∂Cti (T,K) ∂Cti (T,K)
+ qti Cti (T, K) + µit K
σi2 (t, Sti ) = ∂T
i
∂ 2 Ct (T,K)
∂K
i = 0, ..., n (28)
1
2
K 2 ∂K 2

Lemma 4.1 states the Dupire-equation [7]. As there is some confusion in the literature about
the drift terms we repeat the proof in Appendix 1.
Remark 4.1. The lack of continuity condition of type Eq.26 generally leads to serious im-
pairments in the model’s ability to calibrate to the market. Mostly this occurs when market
participants define their own internal volatility surface representation that is based on param-
eterizations that include sharp cut-offs or by kludging together parameterizations of different
types at some point of the curve in a non-differentiable way.
The following lemma deals with cross-arbitrage (also called dispersion-arbitrage ) between
the individual option markets and the option market of the basket itself
Lemma 4.2. Under assumption 4.1 there is no-dispersion arbitrage between individual op-
tion markets and the option market of the basket if and only if
n
X
St0 µ0t = αi µit Sti (29)
i=1

13
 and for any t < T ∗ and for any (St1 , ..., Stn ) ∈ Rn,+ one has
n
!2 n
X X
αi Sti σ02 (t, St0 ) = αi αj Sti Stj σi (t, Sti ) σj (t, Stj ) ρij (t) (30)
i=1 i,j=1

As the basket weights are assumed greater or equal to zero in Eq.22, the maximal covari-
ance is achieved in the case when the correlation matrix is identical one, i.e. ρij = 1 . The
lower bound is given in the case when the correlation matrix equals the identity matrix due
to the non-negative correlations assumption made in Assumption 4.1. This gives rise to the
following

Corollary 4.1. Under assumption 4.1 a necessary condition for no-dispersion arbitrage
between individual option markets and the option market of the basket is
n n
!2 n
2
X X X
αi Sti σi (t, Sti ) αi Sti σ02 (t, St0 ) ≤ αi αj Sti Stj σi (t, Sti ) σj (t, Stj )


≤
i=1 i=1 i,j=1
(31)

Note that Eq.31 could be violated by the market since all variables are infered from it. If
this happens it would indicate arbitrage opportunities provided the form of Eq.21 is correct.
Proof: Note that from Ito’s lemma one obtains
RT
(ST0 − K)+ = Ct0 (T, K) + t
dCu0 (T, K)
RT ∂Cu0 (T,K)
= Ct0 (T, K) + t ∂Su0
(dSu0
− µ0u Su0 du)
 
1 ∂ 2 Cu0 (T,K) ∂Cu0 (T,K) 0 0
RT 0 0 ∂Cu (T,K)
+ t 2 ∂2S0
d < Su , Su > + ∂u
+ ∂Su0
µu S u du
u

RT ∂Cu0 (T,K)
= Cu0 (T, K) + t ∂Su0
(dSu0 − µ0u Su0 du)
2
where d < Su0 , Su0 >= du (σ0 (u, Su0 )Su0 ) which follows from Eq.24.

The cancellation of the last two terms is due to the Black-Scholes partial differential equation
but can also be viewed as a direct consequence of the martingale representation theorem.
The last line describes the hedge-replication strategy of the option payout in terms of delta-
hedging with the basket spot. In order to prove the absence of dispersion arbitrage one needs
to re-write this equation as hedge-integral in terms of its components itself. If one inserts
Eq.23 one achieves this goal as long as the drift terms still cancel so that discounted option
prices remain martingales. This is indeed the case as long as the quadratic variation of the
basket obeys
n

2 X
dt σ0 (t, St0 )St0 =d< St0 , St0 >= dt αi αj Sti Stj σi (t, Sti ) σj (t, Stj ) ρij (t)
i,j=1

14
 Hence the above hedging strategy can be re-written in terms of an arbitrage-free replica-
tion strategy that hedges into the individual components instead:

T n
∂Ct0 (T, K) X
Z
(ST0 +
− K) = Ct0 (T, K) + ( ωi (dSti − µit Sti dt))
t ∂St0 i=1

Theorem 4.1. (LCM) Let Fκ (ρ0 , ξ) be a one-dimensional family of correlation matrices

with center ρ0 and mode ξ = (1, ..., 1)T and ρ̂ij (u) ∈ F (ρ0 , ξ). Let

n
X
cov0 ≡ αi αj Sti Stj σi (t, Sti )σj (t, Stj )(ρ0 )ij (32)
i,j=1

n
X
cov1 ≡ αi αj Sti Stj σi (t, Sti )σj (t, Stj ) (33)
i,j=1

Under assumption 4.1 and the assumption of Eq.31 the option market of Eq.25 has no arbi-
trage if Eq.27,28,29 hold and the correlation ρij in Eq.30 is set to

ρij (t) = ρ̂ij (u∗ )

where

r 
cov0 −σ02 (t,St0 )(St0 )2



 − cov1 −σ02 (t,St0 )(St0 )2
if cov0 − σ02 (t, St0 )(St0 )2 < 0

u∗ = (34)

 r 
 cov0 −σ02 (t,St0 )(St0 )2
−

2 0 0 2 else
σ0 (t,St )(St )

κ is set to one when cov0 − σ02 (t, St0 )(St0 )2 < 0 and zero else.

Proof: The no-arbitrage hypothesis of individual markets εi follows directly from Lemma
4.1. In order to proof the absence of dispersion-arbitrage, because of Lemma 4.2, it suffices
to construct a correlation ρ̂ij that satisfies Eq.30 explicitly. If one inserts Eq.19 into Eq.30
one observes the following cases: If cov0 < σ02 (t, St0 )(St0 )2 the center correlations are not large
enough to satisfy Eq.30 and one needs to set κ = 1 in Eq.19 as ρ̂ij → 1 for u → ∞. Note
that the covariance matrix defined by

15
 n
X (ρ0 )ij + κu2
cov(u ) ≡2
ωi ωj Sti Stj σi (t, Sti ) σj (t, Stj ) (35)
i,j=1
1 + u2

is a strictly increasing function in u2 for κ = 1 and a solution must exist in this case due
to the bound in Eq.31. The result is given by the first line of Eq.34.
Similarly if the center correlation is too large implying that cov0 > σ02 (t, St0 )(St0 )2 one
needs to set κ = 0 in this case as ρ̂ij → 0 for u → ∞ and i 6= j. cov(u2 ) is a strictly
decreasing function and hence invertible in this area with a solution sure to exist due to
the lower bound of Eq.31. The result is given by the second line of Eq.34. We note here
that in the case where non-flat modes are defined (see Eq.17) similar arguments apply for
the existence of solutions to Eq.30 provided ξi > 0. The explicit solution may require the
application of a simple numerical root-finding procedure however .
Remark 4.2. For simplicity we restricted ourselves to a situation where ρij ≥ 0 since this
is generally fulfilled in the equity market. However assume that, instead of the bounds given
by Eq.31, more general bounds are obeyed by the market

n n
!2 n
X X X
αi αj Sti Stj σi (t, Sti ) σj (t, Stj )ρdown
ij ≤ αi Sti σ02 (t, St0 ) ≤ αi αj Sti Stj σi (t, Sti ) σj (t, Stj )
i,j=1 i=1 i,j=1
(36)
where ρdown
ij is a correlation matrix that may contain negative correlations entries. In
order to accomodate this case also Eq. 17 needs to be generalized according to

+u ξi ξj ρup
 down 2
ρij
 √ ij
 (1+ξ u2 )(1+ξ2 u2 )
 2
i 6= j ∈ {1, . . . , n}
i j
ρ̂ij (u) ≡ (37)


1 i=j


This can easily be achieved by correlating not only the Bronian motions by ρdown but also
the changes in the counting processes dNi by means of ρup in the proof of Lemma 3.1.
The corresponding equation for u∗ are quite similar to before and given by
r 
cov0 −σ02 (t,St0 )(St0 )2



 − cov1 −σ02 (t,St0 )(St0 )2
if cov0 − σ02 (t, St0 )(St0 )2 < 0

u∗ = (38)

r 
cov −σ 2 (t,S 0 )(S 0 )2
− −1 20 0t 0t 2



cov0 −σ0 (t,St )(St )
else

where

n
X
cov−1 ≡ αi αj Sti Stj σi (t, Sti )σj (t, Stj )(ρdown )ij (39)
i,j=1

16
 Two−year Volatility Skew of .STOXX50E in the LCM

0.34
●
0.32
0.30

●
0.28

●
volatility

0.26

●
0.24

●
0.22

Market Implied Volatility

Standard Multi−Asset LocalVol Model
LCM Model
0.20

● Local Vol .STOXX50E Single Asset

0.4 0.6 0.8 1.0 1.2

strikes relative to spot

Figure 4: The calculation compares the market implied volatility quotes for 2-year options with
the (standard) multi-asset local volatility model as well as LCM. Whereas LCM tracks
the market distribution quite closely, multi-asset local vol deviates from the market
substantially. In order to identify possible sources of numerical errors, we included
EuroStoxx 50 in the same simulation as a “single” asset as well.

5 Simulation Results of LCM

In chapter 2 we presented the results of the Gaussian copula calculation and demonstrated
that the volatility skew of the components of an index do not suffice to explain the traded
skew of the index itself. For the results presented here, the choice of correlation matrix was
made similar to other market-participants even though the qualitative results are indepen-
dent of the choice of correlations.
In this chapter we show the results of a similar calculation using the multi-asset local
volatility model described in Eq.21, when the basket consistency equation 30 is not enforced.
As one can see in Fig. 4 the results are qualitatively similar to the copula calculation of
chapter 2: A constant correlation cannot explain the steepness of the volatility skew of the
index.
In addition Fig. 4 presents results using the LCM model. LCM closely tracks the distri-
bution of the index.
LCM can provide information for the average (off-diagonal) correlations for a given op-
tion. The results are presented in Table 3. The results show that for a two-year put option
struck at 70% of its spot the correlation is 15% percentage points higher than for a corre-

17
 Strike in terms of spot average correlation in percent
0.7 58.3
0.8 56.1
0.9 53.8
1.0 50.6
1.1 46.9
1.2 43.5

Table 3: Average instantaneous correlations of two-year options between the constituents of Eu-
roStoxx 50 along the simulated paths in LCM. Correlations are significantly higher on
the downside.

Strike in terms of spot average correlation in percent

0.7 55.0
0.8 51.2
0.9 47.7
1 44.5
1.1 41.9
1.2 39.5

Table 4: Average instantaneous correlations of four-year options between the constituents of DAX
along the simulated paths in LCM. Correlations are significantly higher on the downside.

sponding call option struck at 120% of spot. To our knowledge this strike-dependence of
correlations is substantially higher than what a trading desk typically accounts for in taking
reserves for correlation risk (typically 5% points for liquid instruments) . Hence, one expects
mispricings in cases where an exotic payout is convex. The corresponding results for 4-year
options are given in table 4.

18
6 Implementation details of LCM
In this chapter we briefly outline how Monte Carlo code needs to be structured to implement
LCM.
Consider a simulation of nPaths, nAssets and nDates. Traditionally Monte Carlo code
generates one path at a time for each asset in turn. However this will not work in this case
as the choice of correlation matrix depends on the state defined by all the assets. Hence the
structure should look as follows:

begin
call PrepareCholesky(numberOf States, shif t, centerCorrelation)
for ipath := 1 to nP aths
currentP ath = initialSpots
for idate := 1 to nDates
randoms = GenerateRandomNumbers()
StateV ar = CalculateCorrelationState(currentP ath)
CorrelatedRandoms = CorrelateRandoms(randoms, CholeskyDecomp[StateV ar])
for iasset := 1 to nAssets
P athArray[ipath][iasset][idate] = GeneratePathArray(CorrelatedRandoms)
end
currentP ath = P athArray[ipath]
end
end
call CalcPayout(P athArray)
end
In order to correlate random numbers one needs to perform a Cholesky decomposition.
In the case where correlations are constant the latter needs to be computed just once, prior
to the simulation of the paths. This is not true for LCM as every set of paths will generally
require a different matrix to be decomposed. In order not to adversely impact performance
it is useful to prepare a table of correlation matrices beforehand together with their Cholesky
decompositions. The table implements a discrete version of F (ρbase , ξ = (1, ..., 1)T ) for ρ̂ij (u∗ )
from Eq. 19 for about 50-100 values for u∗ , along the lines of:

begin P repareCholesky(numberOf States, shif t, centerCorrelation) :

// This function prepares a list of an odd number of correlation
// matrices together with its Cholesky decomposition and stores the
// results together with the state variable. The index k = k* = (numberOfStates - 1) / 2
// corresponds to centerCorrelation. For k > k* one sets κ = 1 in Eq.34
// whereas for k < k* it is set to zero in this case.
if numberOf States = even
numberOf States := numberOf States + 1
end
m := (numberOf States − 1)/2

19
 correl[m] = centerCorrelation
for k := m + 1 to numberStates
l := k − m
Sum Over i,j
if i = j
correl[i][j] = 1
else
u = shif t + l ∗ shif t
correl[k][i][j] = correl[k][j][i] = (centerCorrelation[i][j] + u2 )/(1 + u2 )
StateV ariable[k] = u
cholesky[k] = CholeskyDecompose(correl[k])
correl[m − l][i][j] = correl[m − l][j][i] = centerCorrelation[i][j]/(1 + u2 )
StateV ariable[m − l] = −u
cholesky[m − l] = CholeskyDecompose(correl[m − l])
end
end
end

Once the simulation has started the function CalculateCorrelationState(currentP ath)

simply calculates the state variable according to Eq.34 and looks up the closest position in the
pre-calculated table from P repareCholesky. Note that all current path values together with
each constituent’s local volatility and the local volatility of the index need to be available to
compute Eq.34.

20
7 Residual (exotic) Correlation Risks in the Pricing of
Multi-asset Derivatives: The “Chewing-Gum” Ef-
fect
LCM allows one to compute single stock delta and vega exposures for multi-asset derivatives.
In addition it produces sensitivities when the basket volatility skew and thus the distribution
of the basket is changing. As the basket volatility depends on the correlation, vega hedging
of the index component of an exotic, to some extent, hedges the correlation exposure as well.

In a sense the idea is to “lock in” the projection of the exotic instrument onto the bas-
ket “subspace” and therefore to hedge some of the correlation exposure along with it.
However a hedge into the index only hedges overall (correlation) movements of the market.
Mathematically, the specification of the distributions of the sum together with its components
does not suffice to fully describe the joint distribution. We give a simple example in footnote
[8].
In order to demonstrate this point, consider a basket Bt of two equally weighted stocks
S1 (t), S2 (t) with S1 (0) = 100, S2 (0) = 100 , e.g.
1
Bt = (S1 (t) + S2 (t))
2
and consider a “chewing-gum” move where S1 (t) goes to 80 and S2 (t) to 120. As the
basket level stays the same in such moves, neither the individual distributions nor the basket
distribution change in this case. However an option that pays the worst of the two assets is
affected in this case.

In this chapter we try to get an estimate of this residual risk by finding alternative
solutions that provide consistency with the basket skew. Any price sensitivity of worst-of
options will be attributed to exotic correlation risk in this case.
The LCM model was introduced for a family of correlations ρ̂ij ∈ F (ρ, ξ) centered around
ρ with a flat mode ξ = (1, ..., 1)T . The merits of this choice lie in the fact that ρ̂(u) can easily
be inverted analytically in this case with the result given by Eq.34. However the model can
easily be extended to non-flat perturbation provided ξi > 0 ∀ i = 1, ..., n . This is feasible as
long as one is willing to deal with a slightly more involved condition for the inversion of ρ̂(u)
(see remark at the end of the proof of theorem 4.1).
In this paper we pursue a simpler approach by simply varying the center-correlation instead.
Recall that the center-correlation is merely a starting point around which a table of corre-
lation matrices is constructed together with their Cholesky decompositions. The individual
selection of correlation matrices is done locally according to Eq. 34.

In the following we present the results for two-year put options on the worst-of all 50 assets
of EuroStoxx 50, which pays at maturity

S1 (T ) S50 (T ) +
(K − M in( , ..., ))
S1 (0) S50 (0)

21
 Strike in terms of spot worst-of put price: set1 worst-of put price: set2
0.6 0.306 0.463
0.7 0.391 0.560
0.8 0.482 0.657
0.9 0.575 0.754

Table 5: Price of worst-of put

for two different choices for the center-correlation:

Set 1 presents results based on F (ρ, (1, ..., 1)T ) where ρ was chosen close to where we
believe the “market” was trading whereas set 2 is based on F (diag(1, ..., 1), (1, ..., 1)T ) with
the identity matrix chosen as center.
We have also computed the average correlation which comes out roughly the same in
both cases. Henceforth despite the fact that all individual as well as basket distributions
are the same, the value of worst-of puts substantially differ between the two sets. In a
sense basket options are mainly sensitive to overall shifts of the correlation matrix only and
hence to the first principal component of the correlation matrix itself. This is different for
worst(best)- of call and put options which show significant sensitivity to “higher” vibrations
such as chewing-gum moves representing higher principal components of the correlation
matrix. This is why liquid prices for worst(best)-of options would be very useful in providing
further important clues about the joint dynamics of the system. In other words, the volatility
skew for the basket can provide only little insight into the pricing of the worst-of (best-of)
options, particularly when the number of basket members is large. This makes them ideal
candidates to augment the correlation information provided by the basket.

22
8 Conclusion
In this paper we show that the index option’s market is inconsistent with the option market
of the individual constituents if the dependency between the assets was to be described by
a deterministic correlation matrix only or by a Gaussian copula. The data strongly suggests
higher correlations on the downside and gives rise to a correlation skew which is computed
explicitly in this paper. Roughly only half of the skew of the index can be attributed to the
skew of the individual components, indicating that state-dependent correlation dynamics
could play a crucial role in explaining the distribution of the index.
This paper generalizes Dupire’s local volatility model to the multi-asset case in a numeri-
cally efficient way and provides consistency not only with the individual option’s markets but
also with the volatility market of the index itself by making correlations a dynamic variable.
We also discuss the non-uniqueness of the solution and introduce the ”‘chewing-gum”’
effect in this context where single stocks move against each other while leaving the distri-
bution of the basket as well as all individual distributions unchanged. However worst-of
options show price sensitivity in such moves which makes them ideal candidates to further
“complete” the market and augment our knowledge of the joint dynamics of the system.

23
9 Appendix 1
In the first step consider n + 1 separateR T economies εi for i = 0, ..., n each one consisting of
a deterministic bond B(t, T ) = exp(− t (rs ds)) together with St , e.g. εi = {B(t, T ), Sti }.
i

Because of the specific form of Eq.21, that is the fact that volatility depends on the spot
only, each εi is complete. The existence of an equivalent martingale measure states that the
dividend-adjusted processes S̃ti = Sti /e−qt are martingales under the cash numeraire B(t, T )
yielding Eq.27 [6]. The specific form of Eq.28 has been derived in the literature already (see
for example [7]). However this paper ignores drift terms. In several instances Eq.28 has
been published incorrectly. The second term in the numerator qti Cti (T, K) is mistakenly
replaced by µit Cti (T, K). We quickly re-derive this equation at the end of this Appendix. In
the second step one considers the economy ε = {B(t, T ), St0 , ..., , Stn } that trades the option
market specified in Eq.25. In order to avoid “spot” arbitrage between the index and its
components taking appropriate (Qi ) expectation values on both sides of Eq.22 immediately
yields Eq.29. One is left to show that Eq.30 avoids dispersion arbitrage between the basket
option market and the individual option markets. This is done in Chapter 4 .

Re-derivation of Dupire’s formula:

If one defines

ξt ≡ (St − K)+ βt
Rt
where βt ≡ exp(− 0
rs ds). The application of Tanaka’s formula yields
 
1 + 2 2 +
dξt = βt 1St >K dSt + δ(St − K) St σt dt − rt (St − K)
2
Stochastic integration on both sides with subsequent expectation taking yields:

Z T  
+ 1 2 2 +
C0 (T, K) − (St − K) = dt βt µt E[St 1St >K ] + K E[δ(St − K)σt ] − rt E[St − K) ]
0 2
Differentiation on both sides with respect to T gives:
 
∂C0 (T, K) 1 2 2 +
= βT µT E[ST 1ST >K ] + K E[δ(ST − K)σT ] − rT E[ST − K) ]
∂T 2
Note that plain vanilla options obey

C0 (T, K) = βT (E[ST 1ST >K ] − KE[1ST >K ])

The second expectation is just a call spread, e.g. E[1ST >K ] = − ∂C0∂K (T,K)
. Similarly the
expectation of the Dirac delta function can be inferred from option prices according to
2
E[δ[ST − K)+ ] = ∂ C∂02(T,K)
K
. If one inserts this back into the above equation and uses the fact
that
E[δ(ST − K)σT2 ] = E[δ(ST − K)]E[σT2 kST = K]
this yields Eq.28, as for a state-dependent local vol the expectation drops, e.g. E[σT2 kST =
K] = σT2 (ST = K)

24
10 Acknowledgement
I would like to thank Romain Barc for many inspiring discussions on the subject. I am great-
ful to Alexander Kabanov, Daniel Cangemi and Thilo Meyer-Brandis for useful comments. I
also would like to thank Michael Demuth, Thomas Quillet and Syed Uzair Aqeel for valuable
suggestions on the paper.

References
[1] Option pricing when correlations are stochastic: an analytical framework, Jose Da
Fonseca, ESILV and Zeliade Systems, September 2006.

[2] Explaining the index skew via stochastic correlation models, Matthias R. Fengler, Hel-
mut Herwartz , Christian Menn , Christian Werner.

[3] Option-Implied Correlations and the Price of Correlation Risk, Joost Driessen, Pascal
Maenhout, Grigory Vilkov, March 2005.

[4] The Shape and Term Structure of the Index Option Smirk:Why Multifactor Stochastic
Volatility Models Work so Well, Peter Christoersen Steven Heston Kris Jacobs.

[5] Towards a Generalization of Dupire’s Equation for Several Assets, P. Amster , P. de

Napoli and J. P. Zubelli, April 2006.

[6] ”Martingales and Stochastic integrals in the theory of continuous trading”. Stochastic
Processes and their Applications, Harrison, J. Michael; Pliska, Stanley R. (1981).

[7] Pricing and hedging with smiles, Bruno Dupire,April 1993.

[8] Let ω1 , ω2 , ω3 be three independent Brownian Motions. Consider the set 1 of two vari-
ables X1 = ω1 +ω2 +ω3 and X2 = −ω1 as well as set 2 containing X3 = ω1 +ω2 +ω3 and
X4 = − √12 ω1 − √12 ω2 . Even though individually the variables share the same individual
distributions between the two sets together with the distribution of the sum, the joint
distributions differ between the two sets.

25