arXiv:2009.09770v1 [q-fin.

ST] 21 Sep 2020

Wolfgang Karl Härdle and Elena Silyakova*
Implied Basket Correlation Dynamics
Abstract: Equity basket correlation can be estimated both using the physi-
cal measure from stock prices, and also using the risk neutral measure from
option prices. The difference between the two estimates motivates a so-called
“dispersion strategy”. We study the performance of this strategy on the Ger-
man market and propose several profitability improvement schemes based on
implied correlation (IC) forecasts. Modelling IC conceals several challenges.
Firstly the number of correlation coefficients would grow with the size of the
basket. Secondly, IC is not constant over maturities and strikes. Finally, IC
changes over time. We reduce the dimensionality of the problem by assuming
equicorrelation. The IC surface (ICS) is then approximated from the implied
volatilities of stocks and the implied volatility of the basket. To analyze the
dynamics of the ICS we employ a dynamic semiparametric factor model.

Keywords: correlation risk, dimension reduction, dispersion strategy, dynamic

factor models

Wolfgang Karl Härdle: Ladislaus von Bortkiewicz Chair of Statistics, Humboldt-Universität

zu Berlin, Unter den Linden 6, 10099 Berlin, Germany, Sim Kee Boon Institute for Financial
Economics, Singapore Management University Administration Building, 81 Victoria Street,
188065 Singapore. Email: haerdle@[Link].
*Corresponding Author: Elena Silyakova: Ladislaus von Bortkiewicz Chair of Statis-
tics, Humboldt-Universität zu Berlin, Unter den Linden 6, 10099 Berlin, Germany. Email:
silyakova@[Link]

1 Introduction
Equity basket correlation is an important risk factor. It characterizes the
strength of linear dependence between assets and thus measures the degree
of portfolio diversification. It is an input for many pricing models, plays a
key role in portfolio optimization and risk management. The concept of a
time-varying correlation is frequently used in studies that describe the joint
dynamics of assets, Bollerslev et al. (1988), Engle (2002). However, the idea
of considering the correlation as an asset, on its own, is relatively new and has
recently gained popularity together with the emergence of such derivative in-
struments as variance, volatility, correlation swaps and trading strategies with
them, Demeterfi et al. (1999), Bossu et al. (2005). In this context being able
2 Wolfgang Karl Härdle and Elena Silyakova

to predict correlation patterns might help to reveal profitable trading oppor-

tunities. One of the most common ways of obtaining a correlation exposure is
to replicate it with variance swaps. In this paper we study the behaviour of a
particular vehicle for trading correlation known as a “dispersion strategy”, in
which one sells a stock index volatility and buys individual volatilities, Allen
and Granger (2005). We propose several ways of improving the profitability
of the strategy by extracting information from a dynamic model of implied
correlation.
Unlike asset prices, correlations are not directly observed in the market
and need to be estimated in the context of a particular model. Obtaining a
well-conditioned and invertible estimate of an empirical correlation matrix is
often a complicated task, in particular when the dimensionality of basket el-
ements N is higher than the time series length T . Here some work has been
done in the field of random matrix theory (RMT), in which the case “large N ,
small T ” is studied in an asymptotic setting, Bai (1999), Laloux et al. (1999),
Plerou et al. (2002). A further segment of research has moved in the direction
of developing various regularization methods for sample covariance and cor-
relation matrices, such as a shrinkage technique proposed in Ledoit and Wolf
(2003), regularization via thresholding in Bickel and Levina (2008a), bending
in Bickel and Levina (2008b), factor models in Fan et al. (2008) and many
others. There are some studies that propose a dynamic model for returns cor-
relation such as a DCC model by Engle (2002), and in a high-dimensional
setting, Engle et al. (2008). The common feature of all these studies is that
the empirical correlation matrix is estimated under the physical measure from
the time series of asset returns. Alternatively, instead of relying on historical
data, one can infer correlation from the current snapshot of the option market.
Option prices reflect the expectations of market participants about the future
price (volatility) and disclose their perceptions of market risk, Bakshi et al.
(2000), Britten-Jones and Neuberger (2000). Some recent studies have shown
that the implied volatility (IV), that equates the model option price and the
one taken from the market, contains incremental information beyond the his-
torical estimate and outperforms it in forecasting future volatility, Christensen
and Prabhala (1998), Fleming (1998), Blair et al. (2001). Yet only a few pa-
pers have studied the predictive content of the correlation, implied by option
prices. Some work has been done for foreign exchange (FX) options, Campa
and Chang (1998), Lopez and Walter (2000), which showed that correlations
implied from FX options are useful for forecasting future currency correla-
tions. Skintzi and Refenes (2005) investigated the average correlation implied
by equity options and introduced the Implied Correlation index (ICX). They
showed that ICX, computed from current option prices, is a useful proxy for
 Implied Basket Correlation Dynamics 3

the future realized correlation. Driessen et al. (2009) investigate the power
of options implied correlation to explain the future realized correlation and
conclude that its predictive power is quite high.
Here we model the implied correlation (IC), which is an object of very high
dimensionality. Similarly to the IV, every day one recovers an IC surface. We
model the IC with a dynamic semiparametric factor model (DSFM), Fengler
et al. (2007), Park et al. (2009) and Song et al. (2014), and find that it yields
a low dimensional representation as a linear combination of a small number
of time-invariant basis functions (surfaces), whose time evolution is driven by
a series of coefficients; technical aspects are also described in Sperlich et al.
(1999). We produce an IC forecast and use it in several hedging schemes
for a dispersion strategy. For the empirical analysis we chose the German
market represented by the DAX portfolio over the 2-years sample period from
20100802 to 20120801 (dates are written as YYYYMMDD). Backtesting shows
that the hedge allows to the reduction of potential losses and increases the
average profitability of the strategy.
The paper is structured as follows. In Section 2 we introduce the notions
of realized, model-implied and model-free implied volatility and correlation
and describe the basic setup of a dispersion strategy with variance swaps.
The DSFM model for IC is introduced in Section 3 starting with general de-
scription in Section 3.1, followed by the description of the functional principal
component analysis (FPCA) approach to find the basis functions in Section
3.2 and the estimation procedure for both factors and factor loadings in Sec-
tion 3.3. Section 4 presents the dataset taken for the empirical study, followed
by a description of the estimation results in Section 5. Here, first, we inter-
pret obtained factors and factor loadings and propose a time series model for
low-dimensional factors in 5.1. Finally in Section 5.2 we propose and compare
alternative dispersion strategy setups: a no hedge, a naïve approach and an
advanced hedge. Section 6 concludes.
4 Wolfgang Karl Härdle and Elena Silyakova

2 Correlation trading
2.1 Average basket correlation

In a basket of N assets, correlation ρi,j measures linear dependence between

the i-th and the j -th asset return, i, j ∈ {1, . . . , N }. Standard statistical anal-
ysis yields that the basket variance σB 2 can be decomposed as:

X XX
2
σB = wi2 σi2 + wi wj σi σj ρij , (1)
i i j6=i

where σi2 denotes the variance of the i-th asset return and wi its weight
in the basket. Now, assuming that ρij is constant for every pair (i, j), one can
imply the equicorrelation ρ from (1):
2 − P 2 2
σB i wi σi
ρ= P P . (2)
i j6=i wi wj σi σj

Later we call ρ a basket correlation or simply a correlation. The corre-

sponding correlation matrix has all the off-diagonal elements equal to ρ and
thus offers several advantages. Firstly, plugging ρi,j = ρ into (1) reproduces
the basket variance σB 2 . Secondly, if − 1
N −1 < ρ < 1 then the correlation ma-
trix is positive semi-definite, Härdle and Simar (2012). This property becomes
particularly important if N is large. A closer look also reveals that (2) is in
fact a nonlinear weighted average over all ρi,j in the basket:
XX
ρ= ci,j ρi,j (3)
i j6=i

with weights ci,j defined by:

wi wj σi σj
ci,j = P P . (4)
i j6=i wi wj σi σj

Bourgoin (2001) showed that if a correlation matrix is positive semi-

definite, for sufficiently large baskets it holds that 0 ≤ ρ ≤ 1. Using this
property, maximum and minimum variances of a basket, σB,min2 2
and σB,max
respectively, are defined as follows:
X
2
σB,min = wi2 σi2 , (5)
i
X XX
2
σB,max = wi2 σi2 + wi wj σi σj . (6)
i i j6=i
 Implied Basket Correlation Dynamics 5

2
σB,min is achieved when ρ = 0 that is when the assets in a basket are
fully diversified. In the case of no diversification, one observes the maximal
2
possible basket variance σB,max corresponding to ρ = 1.
Further we can rewrite ρ by substituting (5) and (6) to (2):
2 − σ2
σB B,min
ρ= 2 2 (7)
σB,max − σB,min
and obtain an additional interpretation as a measure for the degree of
diversification, Skintzi and Refenes (2005). In fact (7) shows how far σB2 is from
2 2 2
its minimal value σB,min relative to the possible value range σB,max − σB,min ,
or in other words, how far the basket is from the perfect diversification. High
ρ is the sign of a poorly diversified portfolio, which is typical for the market
downturn, when asset prices simultaneously drop driving σB 2 up. It means

diversification benefits disappear in times when they are needed most. To

hedge against correlation risk investors look for derivative securities that offer
higher payoffs (premia) when the correlation decreases.
If a basket is constructed from the constituents of an equity index with
weights equal to index weights, then the corresponding basket correlation
would serve as a benchmark for a sector, an industry or a whole market
average correlation. Figure 1 shows an example of the DAX correlation to-
gether with the volatility of DAX and some of its components. Firstly, we see
that the correlation and the volatility vary over time. Secondly, the volatil-
ity of the basket (DAX) is smaller than almost any individual volatility of
its constituents, which illustrates the impact of the diversification effect on
the portfolio risk. Finally, there is a clear linear dependence of the correla-
tion of the basket and its volatility. However the strength of this dependence
changes when the volatility exceeds a certain threshold. We investigate this
phenomenon and propose a dataset correction scheme in Section 4.

2.2 Implied versus realized correlation

Based on (2) we conclude that the exposure to the basket correlation ρ can be
2 and its constituents, σ 2 .
achieved by exposures to the variances of a basket σB i
Such trades can be realized via a combination of variance swaps. A variance
swap is an over-the-counter contract opened at t, which at t + τ pays the
2
difference between the variance cumulated over the life time of the swap σt+τ
2
and the fixed pre-defined strike σ˜t (τ ):
n o
2
σt+τ − σ̃t2 (τ ) Nvar , (8)
6 Wolfgang Karl Härdle and Elena Silyakova

Volatilities of DAX and constituents, % 80 0.8 0.8

70 0.7 0.7

60 0.6 0.6

Correlation of DAX
50 0.5 0.5

40 0.4 0.4

30 0.3 0.3

20 0.2 0.2

10 0.1 0.1

0 0 0
Mar10 Sep10 Apr11 Oct11 May12 Nov12 0 10 20 30 40 50 60 70 80

Fig. 1. Left panel: DAX correlation (2) - dashed, DAX volatility (11) - solid black, volatility
of DAX constituents Adidas, BMW, Siemens, Daimler, [Link], Lufthansa volatilities (11)
- color, the stock market fall 2011 - shaded area. Right panel: scatter plot DAX volatility
vs. correlation. Estimation period - from 20100104 to 20121228, estimation window - 3
months.

where Nvar is the notional amount. Here and later t and τ are given in
fractions of a year.
The strike of a variance swap is the risk-neutral expectation at t of the
integrated variance from t to t + τ . It is also known as the model-free implied
variance (MFIV), where “model-free” indicates that the expectation does not
depend on the specification of the underlying price process, Britten-Jones and
Neuberger (2000). MFIV can be approximated by a function of current option
prices, Breeden and Litzenberger (1978), Carr and Madan (1998), Britten-
Jones and Neuberger (2000), which has the following form
 t+τ 
Z
σ˜t 2 (τ ) = EQ 
t σ 2 (s)ds =
t
 
2erτ ZSt P (K, τ )dK
 Z∞ 
Ct (K, τ )dK 
t
+ , (9)
τ 
 K2 K2 

0 St

where EQ t expected value at t under the risk-neutral measure Q, Pt (K, τ )

{Ct (K, τ )} price at t of put {call} with exercise price K and time to maturity τ ,
St price of the asset in t, r the annualized continuously compounded risk-free
interest rate.
 Implied Basket Correlation Dynamics 7

MFIV can be opposed to the implied variance σ bt2 (κ, τ ), the square of the
implied volatility (IV), which is obtained by solving

Vt (σ
b, κ, τ ) − V̆t (κ, τ ) = 0, (10)

where Vt is the theoretical (model) option price, V̆t option price taken from
K
the market, κ = moneyness of the option. IV, in comparison to MFIV,
St erτ
is a function of both κ and τ , meaning that at every t one recovers a cloud of
points, which can be approximated by a surface, Cont and Da Fonseca (2002),
Fengler et al. (2007).
The floating leg of the variance swap, the realized variance (RV) of an
asset from t to t + τ , can be computed from the time series of daily asset
returns in different ways, depending on the contract specification. Here we
use the most common following form

252(t+τ )  2
X Si
2
σt+τ = τ −1 log . (11)
i=252t
Si−1

In Carr and Wu (2009) σt+τ 2 − σ̃t2 (τ ) is referred to as the variance risk

premium (VRP), which is shown to be strongly negative for major US stock
indexes over the sample period from January 1996 to December 2003. The neg-
ative sign indicates that investors are willing to pay extra to hedge themselves
against possible future market turmoil. Bakshi et al. (2003), who investigated
the S&P100 index and its largest constituents from 1991 to 1995, also found
significant negative difference between realized and option implied volatilities
for the average of 25 stocks and stressed that this difference is less pronounced
than for the index. Driessen et al. (2009) study each S&P100 constituent in-
dividually. Their t-test for H0 , that the sample means of RV and MFIV are
equal, was not rejected for the majority of stocks in the sample from January
1996 to December 2003.
We check the same hypothesis on the German market for the sample pe-
riod from 20100104 to 20121228 using the dataset described in Section 4. Table
4 in Appendix A.1 summarizes the results of a t-test for the null hypothesis
that RV and MFIV are on average equal against the alternative RV<MFIV.
H0 is strongly rejected for the DAX index. For the DAX constituents the re-
jection rate decreases with the options’ maturity τ : with τ = 0.25 (3 months)
and τ = 0.5 (6 months) the H0 cannot be rejected at a 5% significance level
for 8 out of 30 DAX constituents, with τ = 1 (1 year) for 13 constituents.
Table 4 in the Appendix A.1 reports the t-test results for these 13 stocks.
In addition, Table 5 in the Appendix A.2 report sample averages of RV and
8 Wolfgang Karl Härdle and Elena Silyakova

1.4 1.4

1.2 1.2
MFIC (red), RC (blue)

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
Mar10 Sep10 Apr11 Oct11 May12 Nov12 0 0.2 0.4 0.6 0.8 1 1.2 1.4

Fig. 2. Left panel: DAX ρt,τ - blue, ρ̃t (τ ) - red, right panel: scatter plot of DAX ρt,τ (hori-
zontal axis) vs ρ̃t (τ ) (vertical axis), for t + 0.25 from 20100802 till 20120801.

MFIV and their differences for all 30 DAX constituents. The latter are found
to be negative for most of the stocks and for the DAX index.
Driessen et al. (2009) interpret their t-test results as indirect evidence that
a negative correlation risk premium (CRP) exists. To identify the existence
of CRP in the DAX dataset we compute the model free implied correlation
(MFIC) ρ̃t (τ ) from the MFIVs of DAX and its constituents and the realized
correlation (RC) ρt+τ from the corresponding RV by applying (2):
2 P 2 2
σ̃t,DAX (τ ) − i wi σ̃t,i (τ )
ρ̃t (τ ) = P P , (12)
i j6=i wi wj σ̃t,i (τ )σ̃t,j (τ )

2 P 2 2
σt+τ,DAX − i wi σt+τ,i
ρt+τ = P P . (13)
i j6=i wi wj σt+τ,i σt+τ,j

Figure 2 plots the MFIC and the RC of DAX computed over the 3-
month window and with 3 month maturity respectively (τ = 0.25). The H0 :
RC=MFIC of the t-test is strongly rejected. Using this finding and taking into
account results in the literature we would expect the ρt+τ − ρ̃t (τ ) (CRP) to
be negative most of the time. One of the ways of exploiting this observation
is to make a bet on the market correlation by entering a dispersion strategy.
 Implied Basket Correlation Dynamics 9

2.3 Dispersion strategy with variance swaps

We study one of the variations of the dispersion strategy, which consists of

selling the variance of the basket (DAX) and buying variances of basket con-
stituents.
The dispersion strategy can be implemented by taking a short position in
the variance swap (8) on an index and long positions in variance swaps on its
constituents with notional amounts proportional to index weights. The payoff
of a dispersion strategy at t + τ is then defined by

n o N
X n o
2 2
Dt+τ = − σt+τ,B − σ̃t,B (τ ) + 2
wi2 σt+τ,i 2
− σ̃t,i (τ ) . (14)
i=1

Then we apply (2) and rewrite (14) in the following form:

XX XX
Dt+τ = ρ̃t (τ ) wi wj σ̃t,i (τ )σ̃t,j (τ ) − ρt+τ wi wj σt+τ,i σt+τ,j . (15)
i j6=i i j6=i

Based on empirical findings described in Section 2.2 we assume σ̃t,i (τ ) ≈

σt+τ,i for each constituent stock and simplify the payoff (15), as follows
XX
Dt+τ ≈ wi wj σ̃t,i (τ )σ̃t,j (τ ) {ρ̃t (τ ) − ρt+τ } , (16)
i j6=i

which illustrates that by entering the dispersion strategy one obtains ex-
posure to ρt+τ − ρ̃t (τ ), where the floating leg ρt+τ is computed with (11) and
(2) at expiry, and the fixed leg ρ̃t (τ ) is a function of variance swap strikes (9).
Test results described in Section 2.2 suggest that we should, on average, ex-
pect ρt+τ − ρ̃t (τ ) < 0. It also means the dispersion strategy with payoff Dt+τ
on average would have a profit. However, as one can see in Figure 2, there
might be days when ρt+τ − ρ̃t (τ ) ≥ 0. In order to hedge against these potential
losses one needs a forecast of the floating leg of the dispersion strategy.
Another possible modification of the dispersion trading strategy does not
involve trading on the OTC market and can be implemented with standard-
ized market instruments, puts and calls. The strategy consists in selling index
option straddles and purchasing straddles in options on index components.
The forecast of the implied correlation surface can provide the insight into the
relative cost of index options compared to the price of options on individual
stocks that comprise the index. In comparison to the single historical or im-
plied volatility forecast, usually used for this purpose, the correlation surface
can provide information for trading options on the whole maturity spectrum.
10 Wolfgang Karl Härdle and Elena Silyakova

Which means one can buy straddles with different strikes, depending on the
implied correlation forecast.

3 Modeling and forecasting correlation

dynamics
To determine the amount of hedge for Dt+τ we model the implied correlation
(IC) and use the forecast to approximate the floating leg of the dispersion
strategy ρt+τ . By applying (2) to IV of a basket σ bt,B (κ, τ ) and its N con-
stituents σt,i (κ, τ ), i ∈ {1, . . . , N }, every t we obtain the IC surface (ICS):
b

2 (κ, τ ) − P 2 b 2 (κ, τ )
σ
bt,B i wi σt,i
ρbt (κ, τ ) = P P . (17)
i j6=i wi wj σ
bt,i (κ, τ )σ
bt,j (κ, τ )
Figure 3 displays ρbt (κ, τ ) in different trading days: 20111209, 20120710.
Due to the specific option data structure, every day one observes a “cloud of
strings” that visually resembles a surface and can be recovered by applying
nonparametric smoothing. One can clearly see that surfaces have shape sim-
ilarities, but vary in levels, slopes and curvatures. Thus they may be treated
as daily realizations of a random function. In addition one can observe that
the strings do not have fixed spacial locations. In order to model the dynam-
ics of such a complicated multi-dimensional object we apply the DSFM that
reduces the dimensionality of the problem and allows the ICS to be studied
in a conventional time-series context.

3.1 Model Characterization

At every day t one observes ICs ρb(κt,j , τt,j ), t = 1, . . . , T , j = 1, . . . , Jt , where

j is the index of observations and Jt the total number of observations at day
t. Prior to introducing the model we exclude the case of a fully undiversi-
fied basket, with ρb = 1, from the analysis and apply a variance stabilizing
transformation. Fisher’s Z-transformation (Härdle and Simar (2012)) gives:

def 1 1+u
T (u) = log (18)
2 1−u
def
with Yt,j = T {ρb(κt,j , τt,j )}.
Our aim is to model the dynamics of {(Yt,j , Xt,j ), 1 ≤ t ≤ T, 1 ≤ j ≤ Jt },
where Xt,j = (κt,j , τt,j ). The technique we employ allows us to reduce the
 Implied Basket Correlation Dynamics 11

Fig. 3. ICS implied by prices of DAX options traded on the 20111209, 20120710, surfaces
recovered by the Nadaraya-Watson smoothing

dimensionality and to simultaneously study the dynamics of Yt by approx-

imation through an L-dimensional object with L << J . The DSFM, first
introduced by Fengler et al. (2007) in an application to IV surface dynamics,
and then extended by Park et al. (2009) and Song et al. (2014) has these
desired properties.
The basic idea is to approximate E(Yt |Xt ) by the sum of L+1 smooth basis
def
functions m = {m0 , . . . , mL }⊤ (factor loadings) weighted by time dependent
def
coefficients Zt = (1, Zt,1 , . . . , Zt,L )⊤ (factors):

L
X
Yt,j = m0 (Xt,j ) + Zt,l ml (Xt,j ) + εt,j . (19)
l=1

In representation (19) m are chosen data driven and do not have a par-
ticular (parametric) form.
Here two important remarks are appropriate. Firstly, the unknown basis
functions m must be estimated. Fengler et al. (2007) estimate both m and Zt
iteratively using kernel smoothing techniques, Park et al. (2009) approximates
m by tensor B-splines basis functions weighted by a coefficients matrix. Here
we employ a functional principal component analysis (FPCA) approach that
will be described in Section 3.2. The nonparametric estimation procedure that
we use is introduced in Section 3.3; the basics of this technique can be found
in Song et al. (2014).
12 Wolfgang Karl Härdle and Elena Silyakova

The second issue is the estimation of the latent factors Zt . Having the
data-driven basis m
b l in hand, we can estimate daily factors by the ordinary
least squares (OLS) method. Afterwards one fits the econometric model to Zbt ,
as it was done by Cont and Da Fonseca (2002) and Hafner (2004), who fitted
AR(1) to every Zt,l , l ∈ {1, . . . , L}, or in Fengler et al. (2007) who considered
a multivariate VAR(2) process.

3.2 Correlation surface with FPCA

We approximate the ICS by the sum of orthogonal functions. By doing so

we involve the FPCA theory by looking at the ICS as a stationary random
function f : R2 → R.
Let J = [κmin , κmax ] × [τmin , τmax ] the range of possible values of κt,j
and τt,j . We introduce (ρt ), t ∈ {1, . . . , T }, the sample of i.i.d. smooth ran-
dom functions (surfaces). Every ρt is a smooth map ρt : J → R and sat-
R
isfies J E(ρ2t ) < ∞. Also for every ρt we assume a well-defined mean func-
tion µ(u) = E {ρt (u)} and an existence of a covariance function ψ(u, v) =
E [{ρt (u) − µ(u)} {ρt (v) − µ(v)}]. With φ(u, v) = E {ρt (u)ρt (v)} the covariance
function can be expressed as

ψ(u, v) = φ(u, v) − µ(u)µ(v), (20)

which can be also interpreted as a covariance coefficient of two points on
the surface with coordinates u and v ∈ J . Since (20) is a symmetric positive
definite function we can use it as a nucleus of the integral transform, performed
by the linear operator. Define the covariance operator Γ:
Z
(Γf )(u) = ψ(u, v)f (v)dv (21)
J

that transforms f into (Γf ). Γ is a symmetric positive operator with or-

thonormal eigenfunctions {γj }∞ j=1 , γj : J → R, and associated eigenvalues
∞
{λj }j=1 with λ1 ≥ λ2 ≥ . . . ≥ 0. Now we can express (20) in terms of eigen-
functions and eigenvalues of the covariance operator Γ by applying Mercer’s
theorem, e.g. Indritz (1963):
∞
X
ψ(u, v) = λj γj (u)γj (v). (22)
j=1

Taking eigenfunctions {γj }∞ j=1 as a basis, we represent ρt (u) − µ(u) as a

R
generalized Fourier series with coefficients given by ζtj = J {ρt (u) − µ(u)} γj (u)du
 Implied Basket Correlation Dynamics 13

2 ) = λ and
called the j -th principal component score with E(ζtj ) = 0, E(ζtj j
E(ζtj ζik ) = 0 for j 6= k, Ramsay and Silverman (2010). Thus one may rewrite
ρt (u) − µ(u) in the Karhunen-Loève form:
∞
X
ρt (u) − µ(u) = ζtj γj (u). (23)
j=1

Here ζtj indicates how strong the influence of the j -th basis function on
the shape of the t-th surface is. The higher the score, the closer the shape of
ρt resembles the shape of the j -th eigenfunction.
In practice one needs to take L eigenfunctions to replace the infinite sum
in (23) by the finite sum of L basis functions, corresponding to the highest
eigenvalues. One calls {γj }Lj=1 the empirical orthonormal basis, Ramsay and
Silverman (2010). In the next Section we discuss the estimation procedure for
{γj }L
j=1 as well as criteria for the L selection.

3.3 Estimation Algorithm

In model (19) both Zt and m must be estimated. We do that in two steps.

At the first step we estimate the covariance operator introduced in Sec-
tion (3.2) and take µ
b as mb 0 and γ bl as mb l , l ∈ {1, . . . , L} .
The covariance function (20) is estimated as described in Yao et al. (2005)
and Hall et al. (2006). The procedure consists in least-squares fitting of two
local linear models, for µ
b and ψb.
Given u ∈ J we choose (b aµ , bbµ ) = (aµ , bµ ) to minimize

Jt
T X
X
{Yt,j − aµ − bµ (u − Xt,j )}2 Khµ (Xt,j − u) , (24)
t=1 j=1

and take µ b(u) = b aφ , bbφ,1 , bbφ,2 ) =

aµ . Then, given u, v ∈ J we choose (b
(aφ , bφ,1 , bφ,2 ) to minimize

T
X X
{Yt,j Yt,k − aφ − bφ,1 (u − Xt,j ) − bφ,2 (v − Xt,k )}2 (25)
t=1 j,k:1≤j6=k≤Jt


×Khφ (Xt,j − u) Khφ Xt,k − v ,

b v) = b
and take φ(u, aφ .
Here Kh denotes the two-dimensional product kernel, Kh (q̄) = kh1 (q̄1 ) ×
kh2 (q̄2 ), h = (h1 , h2 )⊤ , based on one-dimensional kh (q̄) = h−1 k(h−1 q̄). For our
14 Wolfgang Karl Härdle and Elena Silyakova

Fig. 4. Mean function µ

b(u) of the DAX ICS with corresponding data points, estimated from
20100802 to 20110801 with hµ = (hµ,1 , hµ,2 )⊤ = (0.12, 0.17)⊤

application we selected the quartic kernel, where k(q̄) = 15/16(1− q̄ 2)2 for |q̄| <
1 and 0 otherwise. For both (24) and (25) kernel bandwidths hµ = (hµ,1 , hµ,2 )⊤
and hφ = (hφ,1 , hφ,2 )⊤ are to be selected. The procedure is described in Ap-
pendix A.5. Figure 4 shows an example of µ b(u) estimated using the dataset
described in Section 4 for a sub-sample from 20100802 to 20110801.
Finally, having estimates µ b v), we compute ψ(u,
b(u) and φ(u, b v) using (20)
and take its L eigenfunctions corresponding to the largest eigenvalues as m b l,
l ∈ {1, . . . , L}. Parameter L is chosen in such a way that the selected eigen-
functions explain the large share of variability in the original data. It is also
necessary to mention that ψ(u, b v) is a matrix of a very large dimensionality.
To obtain its consistent estimator, suitable for further spectral decomposition,
various matrix regularization techniques can be used., e.g. banding as in Bickel
and Levina (2008b), thresholding in Bickel and Levina (2008a), eigenvalues
cleaning as in Laloux et al. (1999) and factor models described in Fan et al.
(2008). We use the latter in this step.
In the second step using m bt = (1, Z
b we obtain the estimates Z bt,1 , . . . , Z
bt,L )⊤
as minimizers of the following least squares criterion:

X Jt n
T X o2
Zbt = arg min Yt,j − Zt⊤ m(X
b t,j ) . (26)
Zt
t=1 j=1
 Implied Basket Correlation Dynamics 15

Min. Max. Mean Median Stdd. Skewn. Kurt

IC κ 0.8000 1.2000 0.9825 0.9825 0.0986 0.0690 2.0661
τ 0.0274 0.9671 0.2442 0.1753 0.1979 1.3717 4.3941
bt (κ, τ )
ρ 0.0587 0.9998 0.6150 0.6290 0.1566 -0.2739 2.6115
MFIC ρ̃t (0.083) 0.3895 0.4860 0.6061 0.6193 0.0834 0.0696 0.1957
ρ̃t (0.25) 0.4446 0.9795 0.6549 0.6573 0.0850 0.0613 0.1631
ρ̃t (0.5) 0.4997 1.4730 0.7037 0.6953 0.0866 1.8188 0.1305
ρ̃t (1) 0.5611 1.0851 0.7496 0.7422 0.0905 0.7764 0.6788
RC ρt+0.083 0.1754 0.8955 0.5373 0.5013 0.1331 0.5221 -0.2154
ρt+0.25 0.2774 0.8149 0.5566 0.5363 0.1192 0.2489 -0.8083
ρt+0.5 0.3794 0.7343 0.5759 0.5713 0.1053 -0.0243 -1.4012
ρt+1 0.4312 0.6581 0.5924 0.6050 0.0522 -1.2443 0.9875

Table 1. Summary statistics: IC data computed from the DAX index and constituents op-
tions over the period from 20090803 to 20120801 including the 1 year estimation period (3
years, 770 trading days, 135 obs./day). MFIC computed from daily variance swaps rates.
RC computed from daily stock returns from 20100802 to 20120801 (2 years, 515 trading
days). The figures are given after filtering and data preparation.

4 Data
We study the dispersion strategy over the two year sample period from
20100802 to 20120801 on the German market represented by the DAX bas-
ket. The basket is composed of 23 stocks, constituents of DAX, with the most
liquidly traded options and weights proportional to the current market capi-
talization. To model the dynamics of the IC and to construct the dispersion
trade we operate with three main variables representing different correlation
estimates. MFIC, RC, and IC. The datasets are described in Table 1.
The MFIC dataset contains daily series of MFICs with maturities 0.083,
0.25, 0.5 and 1 years computed via (12) from variance swap rates given by
Bloomberg as a discrete approximation of (9).
The RC dataset contains daily series of RCs computed with (11) and (13)
from the Bloomberg end-of-day stock prices over estimation windows 0.083,
0.25, 0.5 and 1 years.
The IC dataset is constructed using out-of-the-money (OTM) DAX and
single stock options from the EUREX database. To estimate the DSFM model
and produce forecasts for the sample period the dataset covers one additional
year from 20090803 to 20100730. The dataset is transaction-based, meaning
every trade is registered with the date it occurred, expiry date, underlying
16 Wolfgang Karl Härdle and Elena Silyakova

ticker, exercise price (strike) and settlement price. To obtain IV from option
prices via (10) we distinguish between index and single stock options. For
index options, which have the European type of option payoff, the Black-
Sholes (BS) model is used. To account for dividends and early execution in
options on single stocks (American payoff) we use binomial trees, Cox et al.
(1979), and bisection algorithm. Other necessary model parameters, such as
stock prices, index levels, dividend amounts for constituent stocks, interest
rates and stock market capitalization are taken from the Bloomberg database.
As a risk free rate proxy we take daily values of EURIBOR (Euro Interbank
Offered Rate) with 1 week up to 1 year maturities and use linear interpolation
to obtain values for required option τ . We use the most liquid segment of
data with κ ranging from 0.8 to 1.2 and τ from 10 days to 1 year. Options
outside of this range are excluded from the data set due to the poor data
quality, which does not allow to recover implied volatility surfaces for the
DAX and all constituents and to compute implied correlation on a daily basis.
Figure 3 in Section 3 shows an example of the ICS plotted using the entire
available option data, including options outside of the τ -range from 0.8 to 1.2,
for two selected “rich with data” days (20111209 and 2012071). As one can
see, some correlations observed in Figure 3 are more extreme in comparison
to the values in Table 1. The plots show the nature of the implied correlation
estimate, which is not necessarily observed in a range from 0 to 1. Those days
reveal the possibility of a so-called “volatility arbitrage”. Having in mind the
empirical findings described in Section 2.2, stating that the VRP of an index
is much more pronounced then of constituents, one might take a short position
in a too-expensive delta-hedged index option, when the implied correlation is
considerably higher than 1.
Options from original EUREX dataset are not given on a regular (κ, τ )-
grid, required in (17). In the τ -dimension, maturities are standardized by
market regulation, so every t one can find several τt , similarly for the index
and for all constituents. However, in κ-dimension one needs to interpolate. At
every t we use the original (κt , τt ) grid of the index and linearly interpolate
IVs of all constituents to obtain values corresponding to this grid. To avoid
computational problems with a highly skewed empirical distribution of (κt , τt ),
we transform the initial space [0.8, 1.2] × [0.03, 1] to [0, 1]2 using an empirical
distribution function. Also, we remove options with extremely high IVs (larger
than 50%) considering them the misprints in trade registration and finally use
(17) to obtain IC, which produces, on average, 135 observations per day.
Figure 1 shows there is a linear dependence between basket correlation
and volatility. We check this finding in the RC dataset for different estimation
windows and in IC dataset for different maturities. The RC data allows for
 Implied Basket Correlation Dynamics 17

Fig. 5. Factor loadings m

b 0, m
b 1, m b 3 estimated from 20090803 to 20100730
b 2, m

the identification of a breakpoint, a threshold, after which the strength of

the dependence changes, Appendix A.3. This phenomenon is persistent over
different estimation windows. The IC dataset does not show any clear change
in correlation/volatility dependence. Since the IC is used to obtain a forecast
of a floating leg of the dispersion strategy, which is RC, we propose making a
regime dependent correction of the IC dataset as described in Appendix A.4.

5 Empirical results
5.1 Estimation Results and Factor Modeling

Using the IC dataset described in Section 4 we estimate the DSFM model

for three non-overlapping sub-samples 20090803 - 20100730 (the 1st year),
20100802 - 20110729 (the 2nd year), 20110802 - 20120801 (the 3rd year), and
for the entire sample 20090803 - 20120801. All sub-samples include particularly
volatile periods caused by the stock market falls in May 2010, “Flash Crash
2010”, and a more pronounced drop in August 2011.
An example of an estimation over the 1st sample year common factor
loadings m b 0, m b 1, mb 2, m bt,1 , Z
b 3 and the daily time series of factors Z bt,2 , Z
bt,3 is
given in Figure 5 and Figure 6. Now the modeling task is simplified to the
low-dimensional analysis of factor series. We fit the VAR model of order p
for Zbt,1 , Zbt,2 , Zbt,3 . Before proposing a proper VAR specification, we check if
Zbt has characteristics that violate assumptions for linear multiple time series
models. We perform the augmented Dickey-Fuller (ADF) test to check each
18 Wolfgang Karl Härdle and Elena Silyakova

1 1 1
0.5 0.5 0.5
t,1

t,2

t,3
d

d
0 0 0
Z

Z
−0.5 −0.5 −0.5
−1 −1 −1
Aug09 Nov09 Mar10 Jul10 Aug09 Nov09 Mar10 Jul10 Aug09 Nov09 Mar10 Jul10
t t t

1 1 1

0.5 0.5 0.5

ACF

0 0 0

−0.5 −0.5 −0.5

0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Lag Lag Lag

bt,1 , Z
Fig. 6. Driving factors of the DAX ICS Z bt,2 , Z
bt,3 and ACF up to the 20th lag from
20090803 to 20100730

Zbt,1 , Zbt,2 , Zbt,3 for stationarity, Appendix A.6. For Zbt,2 in sub-sample 20100802
- 20110729 we cannot reject the hypothesis of a unit root, so we use its first
differences instead. Then we define the appropriate number of lags, or order
p, by computing Akaike’s information criterion (AIC), Schwarz’s Bayesian in-
formation criterion (SBIC), and the Hannan and Quinn information criterion
(HQIC) values, Appendix A.7. An * appearing next to the test statistics in-
dicates the optimal lag. Except for the sub-sample 20110802 - 20120801, the
test statistics suggest p = 2, so we make a choice in favor of this specification.
The estimation results are summarized in Appendix A.8. We also conducted
a portmanteau (Q) test for the null hypothesis that a series of residuals ex-
hibits no autocorrelation. The test does not indicate the presence of a serial
correlation.
Based on the results, we can distinguish the influence of each factor on
the time evolution of the ICS. The first factor can be interpreted as level, the
second as maturity and the third as a moneyness effect. The relative size of
the largest eigenvalues of (20) suggest that m b 1 is capable of capturing the
biggest share of the surface variability. The variation captured by the second
mb 2 has a smaller influence, since it is only responsible for the surface shape
transformation in the τ dimension. Finally, since the variation of the ICS in
the κ dimension is relatively small, the m b 3 has a smaller impact, which is also
b
reflected in the Zt,3 series.
The forecast of Zbt,1 , Zbt,2 , Zbt,3 modeled with VAR(2) together with esti-
mated fixed m b 0, m
b 1, m
b 2, m
b 3 give a forecast of the ICS.
 Implied Basket Correlation Dynamics 19

τ Min. Max. Mean. Median Stdd. Skew. Kurt.

0.083 -108.04 72.30 -1.14 -0.71 8.00 -6.61 100.49
0.25 -255.48 49.53 -1.20 -0.41 11.49 -17.58 372.33
0.5 -216.04 32.78 -0.74 -0.30 9.37 -18.66 425.86
1 -64.84 76.59 -0.01 -0.38 7.47 2.74 46.85

Table 2. Performance of naïve hedge, summary statistics for εh

t+τ from 20100101 to
20120801

5.2 Backtesting the dispersion strategy

Here we show that using the correlation forecast one can improve the original
dispersion strategy (14) and test it empirically over the 2-years sample period
20100801 - 20120802. We compare the payoff of the strategy without hedging
with the naïve hedging strategy and propose its improvement, the advanced
strategy.
To obtain the value of the naïve hedge position to be held over ∆t days
from t + τ − ∆t till t + τ we make a ∆t-days ahead DSFM forecast ρbt+τ (1, t + τ )
and use it as ρt+τ in (14). Thus the size of the position is defined by
XX
h
Dt+τ = wi wj σ̃t,i (τ )σ̃t,j (τ ) {ρ̃t (τ ) − ρbt+τ (1, t + τ )} . (27)
i j6=i

The corresponding relative hedging error is given by

h
Dt+τ − Dt+τ ρbt+τ (1, t + τ ) − ρt+τ
εht+τ = =− , (28)
Dt+τ ρ̃t (τ ) − ρt+τ
where εht+τ < 0(> 0) means that the hedge (27) under-(over-)estimates
the actual position (14). Table 2 gives summary statistics for the (28) over
the studied sample period for 3 trades with four different maturities: 0.083,
0.25, 0.5 and 1 years. The statistic includes 515 trades originated every day
and expired over the given 2 year sample period, ∆t is one day.
The improved version of the strategy uses the DSFM forecast ρbt+τ (1, t+τ )
as a trigger which defines whether one should hedge or not. If ρbt+τ (1, t + τ ) ≥
ρ̃t (τ ) (DSFM predicts loss in dispersion strategy), takes an offsetting (with
negative sign) position in (27); if ρbt+τ (1, t + τ ) < ρ̃t (τ ) (DSFM predicts gain
20 Wolfgang Karl Härdle and Elena Silyakova

Strategy τ Min. Max. Mean. Stdd.

Dt+τ 0.083 -1502.58 1080.23 87.09 356.94
(no hedge) 0.25 -1531.94 1282.31 101.92 440.54
0.5 -1270.90 1301.28 136.91 456.75
1 -872.76 760.92 134.26 299.01
h
Dt+τ − Dt+τ 0.083 -3237.72 617.40 15.35 203.09
(naïve hedge) 0.25 -1726.53 413.28 35.90 110.14
0.5 -1301.47 344.91 41.13 91.91
1 -914.27 327.03 79.62 93.14
adv
Dt+τ 0.083 -1375.99 1011.38 100.93 256.50
(advanced hedge) 0.25 -1137.79 1282.31 195.09 248.41
0.5 -760.85 1301.28 231.35 281.66
1 -367.89 623.38 123.04 190.80

h
Table 3. Summary statistics for payoffs Dt+τ (no hedge), Dt+τ − Dt+τ (naïve hedge),
adv (advanced hedge) from 20100101 to 20120801, ∆t is one day, best results (highest
Dt+τ
min, max, mean and smallest stdd.) are given in italic

in dispersion strategy), do not hedge. Thus we can write the payoff of the
advanced strategy at t + τ as follows:
(
h
Dt+τ − Dt+τ , if ρbt+τ (1, t + τ ) ≥ ρ̃t (τ )
adv
Dt+τ = (29)
Dt+τ , if ρbt+τ (1, t + τ ) < ρ̃t (τ ).
Since a variance swap contract costs nothing to initiate (we ignore trans-
actions costs), the presented series of daily payoffs correspond to daily P&L
of the hypothetical trade where swaps expire daily over the whole period from
20100801 till 20120802. We compare the cash flows from three strategies. As
one can see in Table 3, the advanced strategy outperforms the other two by
having the smallest maximal losses, highest maximal gains (τ = 0.25, 0.5) and
the highest (second highest for τ = 1) average payoff over the studied sample
period.
In this simplified setting the financing costs are not taken into account for
both strategies.

6 Conclusions
In this study we investigated the implied correlation (IC) of the DAX index
basket and introduced a hedging approach for the dispersion trading strategy
 Implied Basket Correlation Dynamics 21

using the IC forecast. We apply the dynamic semiparametric factor model

(DSFM) to the IC dataset from January 2010 to August 2012, recover four
basis functions and three time series of factors and use them to forecast the
IC. The advanced dispersion strategy we employ using the IC forecast shows
the smallest maximal losses, the highest maximal gains and the highest av-
erage payoff over the studied sample period in comparison to the alternative
strategies. So, we conclude that our modeling approach can be of potential
use in equity dispersion trading.
The choice of DSFM as a model for the IC surface (ICS) dynamics is
motivated by the degenerated dataset design, which has to be modeled non-
parametrically. On the other hand we were driven by the necessity to reduce
the dimensionality of the problem and facilitate the forecasting. DSFM satis-
fies both requirements. It captures the form of the ICS by its nonparametric
part well, and allows a simple parametric model for dynamics to be used. At
the later modeling stage we fit the three-dimensional VAR(2) model, which is
a good choice to carry out the forecasting exercise. In addition, we found that
it is possible to separate the influence of each recovered basis function on the
ICS shape. The functions allow their interpretation as level, moneyness and
maturity effects. The strength of these effects is defined by the time series of
corresponding factors, which can be characterized as drivers of the correlation
risk. An interesting task would be to study the presence, size and magnitude
of the correlation risk premia, captured by these factors. We consider this
findings to be important topics for further research.

Acknowledgement: The authors gratefully acknowledge financial support

from the Deutsche Forschungsgemeinschaft through CRC 649 “Economic
Risk”.
Thanks go to L. Udvarhelyi for editorial assistance.

References
Allen, P. and Granger, N. (2005). Correlation vehicles. techniques for trading equity correla-
tion., Technical report, JPMorgan European Equity Derivatives Strategy.
Bai, Z. D. (1999). Methodologies in spectral analysis of large dimensional random matrices,
Statistica Sinica 9: 611–677.
Bakshi, G., Cao, C. and Chen, Z. (2000). Do call and underlying prices always move in the
same direction?, Review of Financial Studies 13(3): 549–584.
Bakshi, G., Kapadia, N. and Madan, D. (2003). Stock return characteristics, skew laws, and
differential pricing of individual equity options, Review of Financial Studies 16: 101–
22 Wolfgang Karl Härdle and Elena Silyakova

143.
Bickel, P. J. and Levina, E. (2008a). Covariance regularization by thresholding, Ann.
Statist. 36(6): 2577–2604.
Bickel, P. J. and Levina, E. (2008b). Regularized estimation of large covariance matrices,
Ann. Statist. 36(1): 199–227.
Blair, B. J., Poon, S.-H. and Taylor, S. J. (2001). Forecasting S&P100 volatility: the in-
cremental information content of implied volatilities and high-frequency index returns,
Journal of Econometrics 105(1): 5 – 26.
Bollerslev, T., Engle, R. F. and Wooldridge, J. M. (1988). A capital asset pricing model
with time-varying covariances, Journal of Political Economy 96(1): 116–31.
Bossu, S., Strasser, S. and Guichard, R. (2005). Just What You Need To Know About
Variance Swaps, Technical report, JPMorgan Equity Derivatives report.
Bourgoin, F. (2001). Stress-testing correlations: An application to portfolio risk manage-
ment, in C. Dunis, A. Timmermann and J. Moody (eds), Developments in forecast
combination and portfolio choice, New York: Wiley.
Breeden, D. T. and Litzenberger, R. H. (1978). Prices of state-contingent claims implicit in
option prices, Journal of Business 51(4): 621–51.
Britten-Jones, M. and Neuberger, A. J. (2000). Option prices, implied price processes, and
stochastic volatility, Journal of Finance 55(2): 839–866.
Campa, J. M. and Chang, P. (1998). The forecasting ability of correlations implied in for-
eign exchange options, Journal of International Money and Finance 17(6): 855 – 880.
Carr, P. and Madan, D. (1998). Towards a theory of volatility trading, Reprinted in Option
Pricing, Interest Rates, and Risk Management, Musiella, Jouini, Cvitanic, University
Press, pp. 417–427.
Carr, P. and Wu, L. (2009). Variance risk premiums, Review of Financial Studies
22(3)(3): 1311–1341.
Christensen, B. and Prabhala, N. (1998). The relation between implied and realized volatil-
ity, Journal of Financial Economics 50(2): 125 – 150.
Cont, R. and Da Fonseca, J. (2002). Dynamics of implied volatility surfaces, Quantitative
Finance 2: 45–60.
Cox, J. C., Ross, S. A. and Rubinstein, M. (1979). Option pricing: A simplified approach,
Journal of Financial Economics 7: 229–263.
Demeterfi, K., Derman, E., Kamal, M. and Zou, J. (1999). More than You Ever Wanted to
Know about Volatility Swaps, Technical report, Goldman Sachs Quantitative Strategies
Research Notes.
Driessen, J., Maenhout, P. J. and Vilkov, G. (2009). The price of correlation risk: Evidence
from equity options, Journal of Finance 9: 1377–1406.
Engle, R. (2002). Dynamic conditional correlation: A simple class of multivariate generalized
autoregressive conditional heteroskedasticity models, Journal of Business & Economic
Statistics 20(3): 339–50.
Engle, R. F., Shephard, N. and Sheppard, K. (2008). Fitting vast dimensional time-varying
covariance models, Economics Series Working Papers 403, University of Oxford, De-
partment of Economics.
URL: [Link]
Fan, J., Fan, Y. and Lv, J. (2008). High dimensional covariance matrix estimation using a
factor model, Journal of Econometrics 147: 186–197.
 Implied Basket Correlation Dynamics 23

Fengler, M. R., Härdle, W. K. and Mammen, E. (2007). A semiparametric factor model for
implied volatility surface dynamics, Journal of Financial Econometrics 5(2): 189–218.
Fleming, J. (1998). The quality of market volatility forecasts implied by S&P100 index
option prices, Journal of Empirical Finance 5(4): 317 – 345.
Hafner, R. (2004). Stochastic Implied Volatility, Springer Verlag, Heidelberg.
Hall, P., Müller, H.-G. and Wang, J.-L. (2006). Properties of principal component methods
for functional and longitudinal data analysis, Ann. Statist. 34(3): 1493–1517.
Härdle, W., Müller, M., Sperlich, S. and Werwatz, A. (2004). Nonparametric and Semipara-
metric Models, Springer Verlag, Heidelberg.
Härdle, W. and Simar, L. (2012). Applied Multivariate Statistical Analysis, 4rd edn,
Springer Verlag, Heidelberg.
Indritz, J. (1963). Methods in analysis, Macmillan New York.
Laloux, L., Cizeau, P., Bouchaud, J.-P. and Potters, M. (1999). Noise dressing of financial
correlation matrices, Physical Review Letters 83: 1467–1470.
Ledoit, O. and Wolf, M. (2003). Improved estimation of the covariance matrix of stock
returns with an application to portfolio selection, Journal of Empirical Finance
10(5): 603–621.
Lopez, J. and Walter, C. (2000). Is implied correlation worth calculating? Evidence from
foreign exchange options and historical data, Journal of Derivatives 7(3): 65 – 81.
Muggeo, V. M. R. (2003). Estimating regression models with unknown break-points, Statis-
tics in Medicine 22(19): 3055–3071.
Park, B., Mammen, E., Härdle, W. and Borak, S. (2009). Dynamic semiparametric factor
models, J. Am. Statist. Assoc. 104(485): 284–298.
Plerou, V., Gopikrishnan, P., Rosenow, B., Amaral, L. A. N., Guhr, T. and Stanley, H. E.
(2002). Random matrix approach to cross correlations in financial data, Phys. Rev. E
65: 066126.
URL: [Link]
Ramsay, J. and Silverman, B. W. (2010). Functional Data Analysis (Springer Series in
Statistics), 2nd edn, Springer Verlag, Heidelberg.
Skintzi, V. and Refenes, A. (2005). Implied correlation index: A new measure of diversifica-
tion, Journal of Futures Markets 25(2): 171–197.
Song, S., Härdle, W. K. and Ritov, Y. (2014). High dimensional nonstationary time series
modelling with generalized dynamic semiparametric factor model, Econometrics Journal
17: 1–32.
Sperlich, S., Linton, O. B. and Härdle, W. K. (1999). Integration and backfitting methods
in additive models-finite sample properties and comparison, Test 8(2): 419–458.
Yao, F., Müller, H.-G. and Wang, J.-L. (2005). Functional data analysis for sparse longitu-
dinal data, J. Am. Statist. Assoc. 100: 577–590.
 24 Wolfgang Karl Härdle and Elena Silyakova

A Appendix
A.1 Realized versus model free implied volatility

τ = 0.25 τ = 0.5 τ =1
BASF 0.00000 0.00000 0.15298*
Commerzbank 0.00000 0.00005 0.66073*
Continental AG 0.99998* 0.99998* 0.99998*
Deutsche Bank 0.00001 0.00000 0.06985*
Deutsche Börse 0.95674* 0.95064* 0.96357*
Fresenius Medical Care 0.45640* 0.27716* 0.81540*
Henkel 0.90404* 0.99997* 0.99680*
K+S 0.00000 0.00000 0.99981*
Lanxess 0.27625* 0.05776* 0.99989*
Linde 0.76162* 0.87214* 1.00000*
RWE 0.40725* 0.21673* 0.05305*
ThyssenKrupp 0.01733 0.00272 0.69073*
Volkswagen 0.99998* 0.99998* 0.99998*
Table 4. The results of t-test for the equality of sample averages with H0 : RV = M F IV
against the alternative RV < M F IV of DAX constituents for which the H0 is not rejected
at 5% significance level at least for one τ . * denotes test results, which cannot reject H0 .
The test was performed for the volatilities of the DAX index and its 30 constituent stocks
computed over the time period 20100104 - 20121228 for 3 different maturities/estimation
windows: τ = 0.25, 0.5, 1. The test results are presented for the subsample of 13 DAX
constituents, for which H0 cannot be rejected at least for one τ .
 Implied Basket Correlation Dynamics 25

A.2 Variance risk premium

τ = 0.25 τ = 0.5 τ =1
σ σ̃ σ − σ̃ σ σ̃ σ − σ̃ σ σ̃ σ − σ̃
Adidas 27.56 30.06 -2.50 28.35 31.07 -2.71 29.72 31.45 -1.73
Allianz 29.38 31.93 -2.55 30.44 33.03 -2.58 32.72 33.94 -1.22
BASF 28.95 31.08 -2.13 29.73 31.62 -1.89 31.32 31.58 -0.26
Bayer 27.39 30.74 -3.35 27.93 31.23 -3.31 28.85 31.17 -2.31
Beiersdorf 33.90 37.09 -3.20 34.53 37.96 -3.43 35.95 37.94 -1.99
BMW 19.57 23.95 -4.38 19.79 24.17 -4.38 20.26 23.90 -3.64
Commerzbank 46.47 52.77 -6.30 47.40 52.54 -5.14 51.02 51.70 -0.68
Continental AG 41.45 39.82 1.63* 43.84 40.55 3.29* 48.28 41.60 6.69*
Daimler 34.18 37.55 -3.36 34.93 38.45 -3.52 36.93 38.85 -1.93
Deutsche Bank 39.26 43.56 -4.30 39.76 43.93 -4.18 42.52 43.51 -1.00
Deutsche Börse 30.03 30.40 -0.37 31.09 31.10 0.00* 32.90 31.48 1.43*
Deutsche Post 31.61 33.63 -2.02 32.05 34.18 -2.13 33.13 34.67 -1.54
Deutsche
25.50 27.54 -2.04 26.20 28.26 -2.06 27.66 29.12 -1.46
Telekom
[Link] 23.62 26.26 -2.64 24.11 26.53 -2.41 24.67 27.48 -2.81
Fresenius Med-
28.12 29.21 -1.09 28.67 29.68 -1.01 30.19 30.04 0.16*
ical Care
Fresenius SE 19.01 22.44 -3.43 19.21 23.17 -3.96 19.89 23.62 -3.73
HeidelbergCement21.16 23.56 -2.40 21.61 23.69 -2.08 22.77 24.08 -1.31
Henkel 39.04 39.00 0.05* 40.96 39.49 1.46* 45.03 39.99 5.04*
Infineon 23.15 26.08 -2.94 23.57 26.70 -3.13 24.30 27.05 -2.75
K+S 40.78 43.75 -2.98 41.75 45.03 -3.27 45.88 45.08 0.80*
Lanxess 29.99 30.84 -0.85 30.86 31.78 -0.92 33.51 32.52 1.00*
Linde 39.90 39.92 -0.02 41.02 40.55 0.47* 42.83 41.07 1.76*
Lufthansa 22.18 25.18 -2.99 22.72 26.32 -3.61 23.47 26.77 -3.30
Merck 23.83 26.06 -2.23 24.43 26.40 -1.97 25.51 26.39 -0.89
Munich Re 23.35 26.54 -3.19 23.88 27.91 -4.03 25.28 29.10 -3.82
RWE 27.43 29.15 -1.72 27.93 29.71 -1.78 28.75 30.38 -1.63
SAP 21.45 23.67 -2.22 21.73 25.14 -3.42 22.00 26.82 -4.82
Siemens 25.94 28.53 -2.59 26.82 29.78 -2.95 28.63 30.35 -1.72
ThyssenKrupp 36.32 38.37 -2.04 36.78 38.71 -1.93 38.80 38.90 -0.10
Volkswagen 37.91 36.31 1.61* 39.48 36.90 2.59* 41.94 36.91 5.03*
DAX Index 21.72 25.38 -3.66 22.30 26.67 -4.37 23.40 27.81 -4.41
√ √ √ √
Table 5. Mean of the RV (σ) and M F IV (σ̃) and their difference RV − M F IV
(σ − σ̃), for DAX index and its 30 constituent stocks computed over the time period
20100104 - 20121228 for 3 different maturities/estimation windows: τ = 0.25, 0.5, 1).
σ − σ̃ ≥ 0 are marked with *.
26 Wolfgang Karl Härdle and Elena Silyakova

A.3 Dependence of index volatility and equicorrelation

































































































Fig. 7. DAX σB,t,τ (solid line) vs ρt,τ (dashed line), scatter plot σB,t,τ vs ρt,τ , for t, τ
from 20100104 till 20121228, estimated with (11) and (2) over 1 month (τ = 0.083), 3
months (τ = 0.25) and 6 months (τ = 0.5) window. Shaded area: Aug 2011 market fall.
Switch point for two regression line is defined as described in Section 4.
 Implied Basket Correlation Dynamics 27

































































































Fig. 8. DAX σ bt,B (1, τ ) (solid line) vs ρbt (1, τ ) (dashed line), scatter plot σ
bt,B (1, τ ) and
ρbt (1, τ ), for t, τ from 20100104 till 20121228, estimated from IVs with (17) for option with
1 month (τ = 0.083), 3 months (τ = 0.25) and 6 months (τ = 0.5) maturity. Shaded area:
Aug 2011 market fall.
28 Wolfgang Karl Härdle and Elena Silyakova

A.4 Switch point selection for correlation regimes

τ σB,t+τ ρt+τ Slope 1 Slope 2

0.083 20.24 0.5917 0.0361 0.0085
0.25 20.34 0.5728 0.0336 0.0093
0.5 22.42 0.6008 0.0286 0.0094
Average 21.00 0.5884 0.0328 0.0091
Table 6. Segmented linear regression of ρt+τ on σB,t+τ with one break point, τ =
0083, 0.25, 0.5 for t + τ , from 20100104 till 20120801

The dependence of ρ and σB observed in RV and RC is not pronounced in

case of ATM IV and IC (Appendix A.3). Therefore we propose a market regime
correction scheme for the IC dataset. The idea is to find a breakpoint between
two segments of a piecewise linear regression of ρt+τ on σB,t+τ . Using the
procedure described in Muggeo (2003) we fit a segmented linear regression
with one break point. Based on results summarized in Table 6 we make a
following state dependent correction: if σ
bB,t (1, τ ) > 21 (high volatility regime),
then ρbt (κ, τ ) = 0.0091σ
bB,t (κ, τ )

A.5 Smoothing Parameters Selection

For both (24) and (25) kernel bandwidths hµ = (hµ,1 , hµ,2 )⊤ and hφ = (hφ,1 , hφ,2 )⊤
are to be selected. As suggested in Härdle et al. (2004) we use the penalizing
function approach to select optimal hopt
µ , minimizing mean integrated squared
error (MISE):

T Jt
( L
)2  
1X 1 X X Wh∗ ,t,j (Xt,j )
Yt,j − Zbt,l m
b l (Xt,j ) wh∗ ,t (Xt,j )ΞAIC ,
T t=1 Jt j=1 l=1
T Jt
(30)
with the Akaike (1970) Information Criterion (AIC) as penalizing function
ΞAIC (q) = exp(2q) and Wh∗ ,t,j (Xt,j ) defined by

Kh (0)
Wh∗ ,t,j (Xt,j ) = P Jt
, (31)
Jt−1 k=1 Kh Xt,k − Xt,j
for every Xt,j , 1 ≤ t ≤ T , 1 ≤ j ≤ Jt .
 Implied Basket Correlation Dynamics 29

Since the distribution of the observations is very uneven, we are using

def
the weighted version of the criterion with weights wh∗ ,t (ū) = p−1 h∗ ,t (ū), where
ph ,t (ū) is the average design density. For every Xt,j , 1 ≤ t ≤ T , 1 ≤ j ≤ Jt it
∗

is defined by:

Jt
X

ph∗ ,t (Xt,j ) = Jt−1 Kh Xt,k − Xt,j , (32)
k=1

The bandwidth hopt ⊤

µAIC = (hµ1 , hµ2 ) corresponding to the minimal crite-
rion 30 is taken as optimal. The bandwidth h∗ of the weighting function is
constant and does not depend of choice of hµ .

A.6 Testing Factor Time Series for Stationarity

bt,1
Z Zbt,2 Zbt,3
20090803 - 20100730 (the 1st year) -2.991 (1) -6.982 (1) -5.710 (3)
20100802 - 20110729 (the 2nd year) -1.666* (3) -3.090 (2) -4.480 (1)
20110802 - 20120801 (the 3rd year) -3.511 (2) -3.796 (3) -3.480 (2)
20090803 - 20120801 (the entire sample) -4.025 (1) -6.912 (3) -8.979 (1)

Table 7. Augmented Dickey-Fuller (ADF) test on Z bt,1 , Z

bt,2 , Z
bt,3 . Number of lags included
in the ADF regression (in brackets) is chosen by starting with 3 lags and subsequently delet-
ing lag terms, until the last one is significant at 5% level. Test statistics that does not reject
the hypothesis of a unit root at 5% level are denoted by *.
30 Wolfgang Karl Härdle and Elena Silyakova

A.7 Determining the Number of Lags for VAR Model

AIC HQIC SBIC

20090803 - 20100730 1 1.923 2.061 2.162
(the 1st year) 2 1.839* 1.975* 2.152*
3 1.856 2.052 2.304
4 1.882 2.060 2.389
20100802 - 20110729 1 -2.868 -2.800 -2.699
(the 2nd year) 2 -3.075* -2.932* -2.755*
3 -3.068 -2.898 -2.645
4 -3.051 -2.854 -2.525
20110802 - 20120801 1 -0.118 -0.051 0.048
(the 3rd year) 2 -0.355 -0.238* -0.064*
3 0.361* -0.193 0.055
4 -0.360 -0.144 0.179
20090803 - 20120801 1 0.745 0.773 0.818
(the entire sample) 2 0.384* 0.461* 0.539*
3 0.397 0.467 0.579
4 0.412 0.475 0.621

Table 8. Akaike’s information criterion (AIC), Schwarz’s Bayesian information criterion

(SBIC), and the Hannan and Quinn information criterion (HQIC) for defining the optimal
lag order p of a VAR model for DAX and S&P100 ICS factors Z bt,1 , Z
bt,2 , Z
bt,3 . * appearing
next to the test statistics indicates the optimal lag at 5% significance level.
 Implied Basket Correlation Dynamics 31

A.8 Parameters of VAR model for DAX ICS factors

20090803 - 20120801 20090803 - 20100730

(the entire sample) (the 1st year)
Z1,t Z2,t Z3,t Z1,t Z2,t Z3,t
Z1,t−1 0.645* -0.012 -0.019 0.630* -0.032 0.029
Z1,t−2 0.310* 0.008 0.029 0.276* 0.013 -0.060*
Z2,t−1 -0.104* 0.259* 0.156 -0.036 0.047 0.036
Z2,t−2 0.057* 0.406* -0.014 -0.039 0.339 -0.104*
Z3,t−1 -0.07 0.140* 0.471* -0.091 -0.494* 0.525*
Z3,t−2 0.149* 0.118* 0.251* 0.046 0.181 -0.208*
c 0.004 0.006 -0.003 -0.004 -0.001 0.001
20100802 - 20110729 20110802 - 20120801
(the 2nd year) (the 3rd year)
Z1,t Z2,t Z3,t Z1,t Z2,t Z3,t
Z1,t−1 0.809* 0.048 -0.202* 0.339* 0.191* -0.001
Z1,t−2 0.254* -0.029 0.112* 0.351* 0.041 -0.036
Z2,t−1 0.188* 0.687* -0.223* 0.355* 0.264* 0.132*
Z2,t−2 0.091 0.262* 0.018 0.084 0.302* -0.045
Z3,t−1 0.453* 0.051 0.162* -0.197 -0.008 0.623*
Z3,t−2 0.118 0.118 0.275* 0.044 0.240* 0.204*
c -0.004 0.001 -0.002 0.003 -0.001 -0.002

Table 9. The estimated parameters for VAR(2) model for DAX ICS factors. * marks esti-
mates which are not significant at 5% level.