Journal of Empirical Finance 34 (2015) 34–44

Contents lists available at ScienceDirect

Journal of Empirical Finance

journal homepage: www.elsevier.com/locate/jempfin

Volatility co-movements: A time-scale decomposition analysis

Andrea Cipollini a,⁎, Iolanda Lo Cascio b, Silvia Muzzioli c
a
Department of Economics, Accounting and Statistics, University of Palermo, 90128 Palermo, RECent, CEFIN, University of Modena and Reggio Emilia, Italy
b
Department of Economics, Accounting and Statistics, University of Palermo, 90128 Palermo, Italy
c
Department of Economics, CEFIN, University of Modena and Reggio Emilia, 41121 Modena, Italy

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we are interested in detecting contagion from US to European stock market
Received 19 August 2014 volatilities in the period immediately after the Lehman Brothers collapse. The analysis is based
Received in revised form 4 August 2015 on a factor decomposition of the covariance matrix, in the time and frequency domain, using
Accepted 12 August 2015
wavelets. The analysis aims to disentangle two components of volatility contagion (anticipated
Available online 28 August 2015
and unanticipated by the market). Once we focus on standardized factor loadings, the results
show no evidence of contagion (from the US) in market expectations (coming from implied
JEL Classification: volatility) and evidence of unanticipated contagion (coming from the volatility risk premium)
C32
for almost any European country. Finally, the estimation of a three-factor model specification
C38
shows that a European common shock plays an important role in determining volatility
C58
G13 co-movements mainly in the tranquil period, while in the period of financial turmoil, the US
common shock is the main driver of volatility co-movements.
Keywords:
© 2015 Elsevier B.V. All rights reserved.
Implied volatility
Realized volatility
Volatility risk premium
Contagion
Heteroskedasticity bias
Wavelets

1. Introduction

The aim of this paper is to analyze whether, during the period immediately after the Lehman Brothers collapse (from
mid-September 2008 until the end of 2008), there was evidence of contagion between the volatilities in the US stock market and
the stock markets of the UK, Germany, France, the Netherlands, and Switzerland.
The definition of contagion adopted is an increase in the propagation of a country-specific shock (the US in our study) to other
markets in the transition from a period of non-crisis to a period of financial turmoil. For this purpose, we analyze co-movements by
conditioning on the information set not only in the time domain (see Forbes and Rigobon, 2002, based on a comparison of the
correlation of financial returns during crisis and non-crisis periods), but also in the frequency domain. The focus on frequency
bands, and, in particular, those associated with short-term horizons, is helpful for the analysis of contagion as a temporary
phenomenon. As a result, our main concern is the analysis of spillovers over the crisis window extending from mid-September
2008 to the end of December 2008, for frequency bands associated with a time horizon between 2–4, 4–8 and 8–16 days.
Contagion in financial markets has been investigated in the frequency domain by Bodard and Candelon (2009), who employ
Granger causality tests (applied to a pre- and post-crisis sub-sample period). Contagion has been also investigated in the time and
frequency domain, through wavelet analysis, by Rua and Nunes (2009) and Gallegati (2012), who employ the Continuous Wavelet
Transform and the Maximal Overlapping Transform, MODWT, respectively (see also Ranta, 2013, along these lines).

⁎ Corresponding author.
E-mail addresses: andrea.cipollini@unipa.it (A. Cipollini), iolanda.locascio@unipa.it (I.L. Cascio), silvia.muzzioli@unimore.it (S. Muzzioli).

http://dx.doi.org/10.1016/j.jempfin.2015.08.005
0927-5398/© 2015 Elsevier B.V. All rights reserved.
 A. Cipollini et al. / Journal of Empirical Finance 34 (2015) 34–44 35

Our approach improves the understanding of contagion compared to previous studies in at least three respects. First, we control for
heteroscedasticity bias, arising if an increase in the correlation between two variables from normal to crisis times can be ascribed
entirely to the outcome of an increase in volatility (and not associated with any increase in the propagation mechanism). In a
Monte Carlo simulation experiment, we show that, in the presence of series exhibiting long memory, heteroscedasticity bias arises
even if we move from high to low frequency bands for a given regime. In our study, we employ a structural form modeling approach,
based on a factor model specification along the lines of Dungey et al. (2005) and Dungey and Martin (2007) who focus on financial
returns.1 The factor model makes it possible to disentangle the role played by volatility and propagation (captured by standardized
factor loadings) from the US to European markets for different regimes and different frequency bands. As a result, we are able to
control for both sources of heteroscedasticity bias.
Second, unlike the previous wavelet-based studies of contagion (focusing on stock market returns), we investigate contagion
affecting a) realized volatilities, defined over the next month; b) “market fear” indices proxied by implied volatility, which is the
risk-neutral expectation of next-month volatility conditional on the information set available today; c) a forecast error, defined as
the difference between a) and b), interpreted as the volatility risk premium (see Buraschi et al., 2014). We argue that decomposing
contagion into different components of realized volatilities can shed light on contagion anticipated by the market (coming from
implied volatility) and contagion not anticipated by the market (coming from a volatility risk premium).
Our final contribution to previous wavelet-based studies is to account for an omitted variable bias by investigating the role played
not only by a US common shock but also by a European shock in shaping volatility co-movements during crisis and non-crisis periods
and across different frequency bands.
Our analysis is divided into two stages. First, following Percival and Walden (2000), we use Maximal Overlapping Discrete Wavelet
Transform, MODWT, to obtain the time series of wavelet coefficients of the realized volatilities (and of its components: implied
volatility and volatility risk premium) for different frequency bands. In the second stage, we explore the contribution of common
and idiosyncratic shocks to the variability of each volatility series for different frequency bands by fitting a factor model to the wavelet
coefficients retrieved in the first stage. The factor decomposition is obtained by maximum likelihood. Moreover, since the highest
frequency range considered by the wavelet decomposition is between 2 and 4 days, we are also able to overcome the problem of
asynchronous data, without losing any observations.2
The structure of the paper is as follows: Section 2 describes the empirical methodology, Section 3 presents the empirical evidence,
and Section 4 concludes.

2. Multivariate analysis of realized volatility series and its components

A number of recent studies on volatility co-movements take into account long memory in implied and realized volatility series.3 In
order to account for long memory, the seminal work by Andersen et al. (2001) fits a VAR model to the differences of order d (e.g., the
fractional integration parameter) of the log of the currency markets realized volatilities. Bollerslev et al. (2013) adopt a co-fractional
VAR to investigate the relationship between implied volatility, realized volatility, and stock market return for the US. More recently,
Jung and Maderitsch (2014) suggest dealing with the strong persistence in log realized volatilities of the European, US, and Hong Kong
stock markets, using the Heterogeneous Autoregressive Model of Realized Volatility (HAR-RV) proposed by Corsi (2009). To the best
of our knowledge, the only studies analyzing implied volatility spillovers (distinguishing tranquil and turmoil periods) are the one by
Jiang et al. (2012), employing a VAR model and the one by Kenourgios (2014), employing a dynamic conditional correlation model.
All the studies cited above are based on fully parametric multivariate models specified within the time domain and they rely on the
choice of the lag length and/or the estimation of the fractional integration parameters. Unlike these studies, by using wavelet analysis,
we are able to circumvent the lag length specification and the estimation of the fractional integration parameter.4 Wavelet analysis is
particularly useful for decomposing fractional integrated series, given that the estimation results on co-movements based on wavelet
analysis are robust to the presence of long memory, as shown by Percival and Walden (2000).
In this study, focusing mainly on volatility co-movements and contagion, we intend to exploit an information set both in the time
and the frequency domain. Wavelet analysis is particularly suitable to exploring the evolution of co-movements over different time
periods and across different frequency bands (see the contagion study by Gallegati, 2012). Frequency domain approaches provide
an insightful representation of econometric data by means of decomposition into sinusoidal components at various frequencies,
with intensities varying across the frequency spectrum. The main shortcoming of Fourier analysis is related to the assumption of
intensities as constant over time. This makes Fourier methods ineffective in analyzing signals containing local irregularities, such as
spikes or discontinuities, that are a feature of financial time series. Wavelets can be a particularly useful tool when the signal is
localized in time as well as frequency.

1
Other studies of contagion within a structural form modelling framework are those based on structural VAR model identified through zero-exclusion restrictions on
lagged endogenous variables (as in Favero and Giavazzi, 2002); through zero-exclusion restrictions based on the use of country-specific regressors (see Pesaran and
Pick, 2007), through switches in second moments across time (see Caporale et al, 2005 and Dungey et al., 2010, where structural form country-specific shocks are
modeled as GARCH innovations).
2
Forbes and Rigobon (2002) use a moving average across two consecutive days to circumvent the problem of asynchronous data, halving the number of available
observations.
3
Baillie et al. (1996); Andersen and Bollerslev (1997); Comte and Renault (1998) provide evidence of long-run dependencies, described by a fractionally integrated
process, in GARCH, realized volatilities, and stochastic volatility models, respectively. More recently, empirical studies show that the volatility implied from option prices
exhibits properties well described by a fractionally integrated process (see, inter alia, Bandi and Perron, 2006).
4
The volatility series under investigation are fractionally integrated. Results are available upon request.
36 A. Cipollini et al. / Journal of Empirical Finance 34 (2015) 34–44

2.1. Wavelet decomposition

Our analysis is divided into two stages. In the first stage, we use Maximal Overlapping Discrete Wavelet Transform, MODWT (see
Percival and Walden, 2000; Whitcher et al., 2000), which decomposes a time series yt into time-scale orthogonal components:

yt ¼ D1t þ D2t þ … þ D jt þ … þ D Jt þ St ;

The details Djt reproduce the evolution over time of the original series for a particular level j of the decomposition (for j = 1,…,J),
associated with a given frequency range. The smooth component St captures the long-run trending behavior (for frequency bands
lower than those associated with level J) of the series yt. In particular, at level j and scale λj = 2j− 1, the time series of wavelet
coefficients, wt, are able to capture frequencies spanning cycles with periodicity between 2j and 2j + 1. As a result, the lowest scale
is associated with the highest frequency range and the highest scale (up to a maximum level of decomposition J) is associated with
the lowest frequency range.

2.2. Wavelet covariance

In the second stage of the analysis, unlike the studies employing wavelet analysis to study financial contagion, we do not compare
pre- and post-break pairwise (unadjusted) correlations of the wavelet coefficients of the two series for a given level of decomposition.
For this purpose, we employ ML to obtain a factor decomposition of the wavelet covariance matrix for n volatility series for a given
decomposition level. As shown by Percival and Walden (2000) (see also Whitcher et al., 2000), wavelet coefficients underlie the co-
variance between two fractionally integrated time series X and Y (with the orders of integration d1 and d2, respectively) for scale λj =
2j−1. In particular, the scale covariance, defined as γ(λj), is given by:

1 N−1X X Y
w w ð1Þ
N j t¼L −1 j;t j;t
j

where wj,t are the non-boundary and stationary wavelet coefficients for scale λj, obtained from a wavelet transform using a filter of
length LL; Nj = N − Lj + 1, Lj = (2j−1)(LL−1) + 1 is the filter length at level j.5 In this study, we use the Haar filter (i.e. LL equal
to 2), belonging to the Daubechies family of wavelets (see Daubechies, 1992).
Finally, given that contagion in international financial markets is fast, with the effect dying out after a few days (see, for instance,
Ait-Sahalia et al., 2010), our focus is on short-run co-movements. More specifically, given the use of daily data, we associate any
evidence of contagion with co-movements (during periods of crisis) related to the first three levels of decomposition, that is to
frequency bands corresponding to horizons between 2 and 4, 4 and 8, and 8 and 16 days, respectively.

2.3. A factor decomposition of the wavelet covariance matrix

We consider two common shocks (one associated with the US and another one to Europe) as the drivers of volatility co-
movements, during crisis and non-crisis periods. Moreover, we distinguish between Eurozone (Germany, France, and the
Netherlands) and non-Eurozone countries (the UK and Switzerland). In particular, for a given decomposition level j (e.g., frequency
band, the following three-factor model specification describes volatility co-movements between the US and the two non-Eurozone
countries:
    
volit ¼ δi þ δi;c Dt σ us þ σ us;c Dt ηus þ ðγ þ γc Dt Þηnz;t þ σ i þ σ i;c Dt ηi;t ð2Þ

where ηi,t is the idiosyncratic shock for country i at time t; ηus,t and ηnz are the US and common shock to the non-Eurozone countries at
time t, respectively.The coefficients δi, δi,c are restricted to zero when describing the dynamics of the US volatility series, which
remains:
 
volus;t ¼ σ us þ σ us;c Dt ηus;t

A time dummy variable, Dt, taking value 1 during the crisis period and zero otherwise, is used to model the regime shift from a
non-crisis to a crisis period.The size of the US shock and the associated (standardized) factor loadings to each of the non-Eurozone
volatility series during the non-crisis period are measured by σus and δi, respectively. The change in the volatility of the US common
shock during a regime shift is defined by σus,c; a change in the (standardized) factor loading of the US shock to each European volatility

5
Non-boundary coefficients are those not influenced by the end-effect problem. The wavelet transform and its variants, such as the MODWT, make use of circular
filtering. The series under investigation, x, is treated as a portion of a periodic sequence with period N. In other words, the transform considers xN−1, xN−2… as useful
surrogates for the unobserved x−1,x−2… A problem with the periodic extension can occur when there is significant discontinuity between the end of one replication
of the sample and the beginning of the next. In such cases, the coefficients produced by the transform are found to be remarkably high and the reconstructed details are
affected.
 A. Cipollini et al. / Journal of Empirical Finance 34 (2015) 34–44 37

during the crisis period is measured by δi,c. More specifically, the parameter δi,c measures the change in the standardized impact of the
US shock on the non-Eurozone stock market volatilities during the crisis period and it is the coefficient measuring contagion (if it is
positive and statistically significant).
The (non-standardized) factor loadings associated with the non-Eurozone common shock and the changes over the crisis period
are defined by γ and γc. The loadings of the idiosyncratic shocks for the non-Eurozone stock market volatilities are defined by σi and
the change in σi during a regime shift is defined by σi,c.
Eq. (2) makes it possible to control for heteroscedasticity bias when testing for the contagion from the US. For a given regime and a
given frequency band, the comparison between the role played by the US and the European common shocks in shaping volatility
co-movements is based on the associated non-standardized factor loading. As a result, conditional on a frequency band, while the
comparison over the tranquil period is between the coefficients γ and b, where b = δσus, the comparison over the crisis period is
between the coefficients (γ + γc) and (b + bc), where (b + bc) = (δ + δ,c)(σus + σus,c). In matrix notation, the factor model
specification can still be expressed in terms of Eq. (2′):
h i
1=2 1=2
volt ¼ ½Δ þ Δc Dt  Σ þ Σc Dt ηt ð2′Þ

" # " #
ηus;t
volus;t
ηnz;t
where: volt ¼ voluk;t ; ηt ¼
ηuk;t
volswi;t
ηswi;t

To (exactly) identify the structural form model, the matrix of factor loadings for the non-crisis period, Δ, and the one for the crisis
period, Δc, are defined as
2 3 2 3
1 0 0 0 1 0 0 0
Δ ¼ 4 δuk γ σ uk 0 5; Δc ¼ 4 δuk;c γc σ uk;c 0 5:
δswi γ 0 σ swi δswi;c γc 0 σ swi;c

The diagonal matrices, with the main diagonal elements given by the standard deviation (e.g., the unconditional volatilities) of the
four shocks are given by
2 3 2 3
σ us 0 0 0 σ us;c 0 0 0
6 0 1 0 07 6 0 1 0 07
¼6 7; Σ1=2 6 7:
1=2
Σ 4 0 ¼
05 4 0 0 1 05
c
0 1
0 0 0 1 0 0 0 1

For each regime (crisis or non-crisis) and a given level of decomposition (e.g., frequency band) we have six free parameters and
(given three endogenous variables) six unique variance–covariance equations (moment conditions), giving an exactly identified
factor model.
The analysis of volatility co-movements (for a given decomposition level j, e.g., frequency band) between the US and three
Eurozone countries (Germany, France and the Netherlands) is based on a factor model specification similar to the one described by
Eq. (2):
    
volit ¼ δi þ δi;c Dt σ us þ σ us;c Dt ηus þ ðγi þ γ ic Dt Þηez;t þ σ i þ σ i;c Dt ηi;t ð3Þ

The only difference with Eq. (2) is the presence of shock ηez,t which is common to the Eurozone countries: Germany, France and the
Netherlands. Using matrix notation we still have the factor model specification given by Eq. (2′), where:
2 3
2 3 ηus;t
volus;t 6 ηez;t 7
6 volger;t 7 6 7
volt ¼ 6 7 6
4 vol f ra;t 5; ηt ¼ 6 ηger;t 7
7
4 η f ra;t 5
volned;t
ηned;t

Exact identification is achieved by the following restrictions on the matrix of factor loadings for the non-crisis period, Δ, and on the
one for the crisis period, Δc:

2 3 2 3
1 0 0 0 0 1 0 0 0 0
6 δger 0 7 6 7
Δ¼6
γ ger σ ger 0 7; Δc ¼ 6 δger;c γ ger;c σ ger;c 0 0 7
4 δ f ra γ f ra 0 σ f ra 0 5 4 δ f ra;c γ f ra;c 0 σ f ra;c 0 5
δned γned 0 0 σ ned δned;c γ swi;c 0 0 σ ned;c
38 A. Cipollini et al. / Journal of Empirical Finance 34 (2015) 34–44

The diagonal matrices for the standard deviation (e.g, the unconditional volatilities) of the five shocks are given by:

2 3 2 3
σ us 0 0 0 0 σ us;c 0 0 0 0
6 0 1 0 0 07 6 0 1 0 0 07
6 7 6 7
¼6 07 ¼6 0 1 0 07
1=2 1=2
Σ 6 0 0 1 0 7 Σc 6 0 7;
4 0 0 0 1 05 4 0 0 0 1 05
0 0 0 0 1 0 0 0 0 1

As a result, for each regime (crisis or non-crisis) and a given level of decomposition (e.g., frequency band) we have 10 free
parameters and (given four endogenous variables) 10 unique variance–covariance equations (moment conditions), giving an exactly
identified factor model.
The estimation of the factor models (2) and (3) is carried out by maximizing the following Gaussian log-likelihood function fitted
to the vector of wavelet coefficients W jt 6:

T1 
X  XT2  
 ; ΔΣΔ0 þ
L W L W ; ½Δ þ Δ ½Σ þ Σ ½Δ þ Δ 0 ð4Þ
jt jt c c c
t¼1 t¼T 1 þ1

where L(.) is the multivariate Gaussian log-density at time t and for scale j, and the length of the sub-samples associated with the
tranquil and crisis period are T1, and T2–T1 respectively. Inference on the structural form parameters is based on White robust
standard error estimator, given by H−1GH−1, where G is the cross-product of the first derivatives and H is the Hessian.

2.4. Monte Carlo simulation for the two-factor model decomposition of the wavelet covariance matrix

We ran a Monte Carlo simulation experiment, based on 500 replications, to show that the focus on the standardized loading
coefficients makes it possible to control for heteroscedasticity bias arising both across time and across different levels of
decomposition. In particular, the DGP created artificially for a pair of time series (each of length equal to 1000) yt is given by

h i
1=2 1=2
yt ¼ ½Δ þ Δc Dt  Σ þ Σc Dt ηt ð5Þ

hη i h i h 1 0 i 1=2 h σ us 0 i 1=2 h σ us;c 0 i

where ηt ¼ 1t ; Δ ¼ 1 0
; Δc ¼ ;Σ ¼ ; Σc ¼
η2t δi σ i δi;c σ i;c 0 1 0 1
Dt is a dummy variable taking 1 only for the last 50 observations and zero otherwise. Each component of the vector ηt = [η1,t, η2,t]′
is a realization of a Gaussian fractional white noise r.v., where the fractional integration parameter d (for both shocks) is set equal
either to 0, or to 0.2, or to 0.35, or to 0.45. The matrix for the standard deviation of the shocks during the calm period is Σ, set equal
to the identity matrix, while the (diagonal) matrix associated with the turmoil period, Σc, has the first row, first column element
h i
1 0
set equal to either 1, or to 4, or to 9. Moreover, Δ ¼ and Δcis the identity matrix. As a result, the DGP does not show any
0:5 1
contagion effect during the crisis period from the first to the second series and the only coefficient affected (by an increase set
equal to 1, or to 4 or to 9) is the standard deviation of η1. Then, we employ MODWT to decompose, up to level j = 5, the two artificially
generated series and we estimate by ML the unadjusted factor models:

yt ¼ ½B þ Bc Dt ηt ð6Þ

h i h i
b11 0 b 0
where B ¼ and Bc ¼ 11;c .
b12 b22 b12;c b22;c
A (upward) heteroscedasticity bias in the unadjusted factor model would occur in the presence of evidence of a positive value of
b12,c. By way of comparison, we also estimate by ML an adjusted factor model given by Eq. (5) to show that the bias is considerably
reduced.
From Table 1 we can observe that when we use the unadjusted factor model, the point estimate of the increase in the coefficient
measuring the impact of the first series on the second series during the regime change is strongly biased upwards, and this bias holds
across different levels of decomposition (from 1 to 5). Mean values of δi,c (across the 500 replications) are close to zero when we turn
our focus to the adjusted factor model specification. Moreover, the higher the fractional integration parameter (the higher the long
memory), the greater the heteroscedasticity bias associated with contagion. Consequently, standardization helps to considerably
reduce the heteroscedasticity bias and this result holds across different levels of decomposition (from 1 to 5).

6
If the focus is on the US, UK, and Swiss volatility co-movements, then the column vector W jt has three components. If the focus is on the US, France, Germany, and
Netherlands volatility co-movements, then the column vector W jt has four components.
 A. Cipollini et al. / Journal of Empirical Finance 34 (2015) 34–44 39

Table 1
Monte Carlo simulation of the two-factor decomposition of the wavelet covariance matrix.

d=0 d = 0.2 d = 0.35 d = 0.45

Level decomposition j = 1
σ1c = 1 0.347 0.325 0.312 0.306
0.000 0.000 0.000 0.001
σ1c = 4 1.399 1.273 1.160 1.075
0.000 0.000 0.000 0.000
σ1c = 9 3.147 2.848 2.567 2.350
0.000 0.000 0.000 0.000

Level decomposition j = 2
σ1c = 1 0.237 0.242 0.250 0.257
−0.002 −0.001 0.000 0.000
σ1c = 4 0.965 0.957 0.935 0.910
0.000 0.000 0.000 0.000
σ1c = 9 2.174 2.142 2.068 1.987
0.000 0.000 0.000 0.000

Level decomposition j = 3
σ1c = 1 0.160 0.181 0.204 0.225
−0.004 −0.003 −0.002 −0.001
σ1c = 4 0.656 0.724 0.771 0.797
−0.002 −0.001 −0.001 −0.001
σ1c = 9 1.475 1.620 1.704 1.737
−0.001 −0.001 0.000 0.000

Level decomposition j = 4
σ1c = 1 0.107 0.137 0.169 0.198
0.005 0.007 0.009 0.009
σ1c = 4 0.432 0.537 0.628 0.691
0.000 0.001 0.001 0.001
σ1c = 9 0.965 1.194 1.378 1.495
0.000 0.000 0.000 0.000

Level decomposition j = 5
σ1c = 1 0.060 0.087 0.121 0.152
0.000 0.002 0.004 0.005
σ1c = 4 0.254 0.359 0.463 0.544
−0.001 −0.001 −0.002 −0.002
σ1c = 9 0.570 0.802 1.020 1.181
−0.002 −0.002 −0.003 −0.003

Note: The DGP for a vector of two series y1 and y2 considered postulates the absence of any contagion effect. The top figure of each entry is the mean value (across 500
replications) of the point estimate of b1,2c coefficient of the unadjusted factor model specification given by Eq. (6)?. The bottom figure of each entry is the mean value
(across 500 replications) of the point estimate of δic coefficient of the adjusted factor model specification given by Eq. (5). Furthermore, d is the fractional integration
parameter and σ1c is the coefficient measuring the increase (over the last 50 observations) in the volatility of y1 relative to a value equal to one during the first 950
observations.

3. Data and empirical evidence

3.1. Descriptive statistics

The time series of realized volatilities (and their components) for the US, the UK, Germany, France, the Netherlands, and
Switzerland are observed at daily frequency from 1/2/2000 till 31/12/2008. The annualized implied volatilities indices (in percentage
values), IV(t), are the risk-neutral expectations for next-month volatility. The realized volatility over the next month, RV(t + 1), is the
square root of the annualized realized variance between t and t + 1 of the stock market index return (obtained from the sum of
squared log returns in the 21 days trading occurring between t and t + 1). The volatility risk premium (see Buraschi et al, 2014),
VRP(t), is given by the difference between the implied volatility, IV(t) and the realized volatility, RV(t + 1), representing the risk
adjustment to be made to the risk-neutral expectation about volatility (IV(t)), given the risk aversion of agents. Since the volatility
risk premium represents compensation for providing volatility insurance, it can be considered as a proxy for risk aversion (see
Bekaert and Horeova, 2014; Muzzioli, 2013a,b).
Descriptive statistics for pre- and post-break (that is over the 1/2/2000–14/9/2008 sub-sample and over the 15/9/2008–31/12/2008
sub-sample, respectively) are presented in Table 2. It may be observed that, after the Lehman Brothers collapse, the realized volatility
experiencing the largest increase in the mean and in the standard deviation is the one for the US (the increase is equal to 288% in the
mean and 111% in the standard deviation). Likewise, the implied volatility index displaying the largest increase in mean (138%) and
standard deviation (185%) is the US one. Among European countries, the ones with the highest changes in implied and realized
volatilities from the pre-break to the post-break period are the UK, France, and the Netherlands. The volatility risk premia are all
positive on average during the non-crisis period, ranging from a low of 2.657 for Switzerland to a high of 3.863 for the Netherlands
on a percentage annualized basis. This shows that on average, selling volatility was highly profitable during the non-crisis period.
40 A. Cipollini et al. / Journal of Empirical Finance 34 (2015) 34–44

Table 2
Descriptive statistics for pre- and post-break period.

US UK GER SWI FRA NED

Implied volatilities
Mean 19.866 20.052 24.466 19.118 22.746 23.543
47.469 44.854 46.563 43.606 44.727 49.437
Std dev 6.752 7.971 10.021 7.759 8.802 10.334
19.272 17.237 19.604 17.597 17.040 18.812
Min 9.890 9.099 11.650 9.239 9.242 10.121
18.810 19.510 20.480 18.823 20.410 21.520
Max 45.080 57.137 62.630 53.037 61.463 65.669
80.860 75.540 83.230 84.896 78.050 81.220

Realized volatilities
Mean 16.675 16.242 21.527 16.460 19.889 19.679
64.785 55.794 55.634 50.283 60.749 64.516
Std dev 7.337 8.567 11.577 9.006 10.054 12.011
15.484 14.737 17.040 16.337 14.696 17.157
Min 5.496 5.122 6.634 5.182 7.391 6.077
28.146 28.791 23.493 24.067 30.234 25.915
Max 46.290 52.742 67.318 55.561 59.444 67.425
84.783 79.481 80.732 78.082 84.820 89.386

Volatility risk premia

Mean 3.699 3.818 2.939 2.657 2.857 3.863
−17.315 −10.939 −9.071 −6.676 −16.021 −15.097
Std dev 4.441 5.529 6.373 6.208 6.091 7.098
19.743 21.517 20.498 24.121 19.662 23.346
Min −16.539 −24.713 −34.146 −33.911 −27.139 −31.131
−49.103 −45.541 −44.853 −44.141 −50.963 −54.207
Max 19.404 21.135 20.143 18.169 20.869 24.359
25.295 23.327 22.520 33.269 15.168 21.941

Note: The values at the top of each entry are the descriptive statistics for the pre-break sub-sample period (1/2/2000–14/9/2008). The values at the bottom of each entry
are the descriptive statistics for the post-break sub-sample period (15/9/2008–31/12/2008 sub-sample).

On the other hand, after the Lehman Brothers collapse, those who were short in volatility suffered a huge loss, given that in this period
the volatility risk premium is negative and much higher in absolute terms than in the non-crisis period. This is the case especially for
the US, and among the European countries under investigation for France and the Netherlands.7

3.2. Volatility co-movements: The role of the US and the European shock

We first assess the relative role played by the European and the US common shock in driving European volatility co-movements,
and we focus on the non-standardized coefficients measuring the loadings of the European common shock (distinguishing between
Eurozone and non-Eurozone countries) and of the US shock to European volatilities. The non-standardized coefficients measuring the
loading of the US shock to European volatilities are given by the coefficients b obtained by estimating a factor model given by Eq. (6)
where the first addend (δi + δi,cDt)(σus + σus,cDt) in Eqs. (2) and (3) is replaced by (bi + bi,cDt).. The coefficients γ's are the
non-standardized coefficients measuring the loading of the European shock to European volatilities (as in Eqs. 2 and 3). We compare
the absolute values of the coefficients b and γ over the tranquil period (e.g., bi and γi) and their increase over the period of financial
turmoil (e.g., bi,c and γi,c), for each country under investigation: the UK, Germany, Switzerland, France, the Netherlands, (i = uk, ger,
swi, fr, ned, respectively). Although increments in volatility levels could be the main driver of an increase (during a switch from the
non-crisis to the crisis period) in the factor loadings of both common shocks, here we highlight only the relative importance
(in terms of absolute values) of the two common shocks in each period and for each series (realized volatility, implied volatility,
and the volatility risk premium). The results are shown in Table 3 for the US volatility, Table 4 for the non-Eurozone countries, and
Table 5 for the Eurozone countries.
As for implied volatility, during the non-crisis and crisis periods, while the European common shock is the most important driver for
all countries over a 2–4 day horizon, the US shock plays the most important role in driving implied volatility co-movements for a
frequency band associated over a 4–16 day horizon. Most of the realized volatility co-movements, during the non-crisis period and
for any of the three frequency bands considered, are driven by the European region-specific shock. On the other hand, during the crisis
period, the US shock plays the most important role in driving implied volatility co-movements. The European region-specific shock
plays a dominant role in driving volatility risk premia co-movements during the non-crisis period and for any of the frequency band
considered. During the crisis period, the European region-specific shock prevails over the US one only for the frequency band
associated with a horizon between 2 and 4 days. For the 4–8 day horizon, the US shock is prominent for Eurozone countries.
For the 8–16 day horizon, the US shock plays the most important role for any country (except the UK).

7
A number of studies (see Buraschi et al., 2014; Carr and Wu, 2009) find (on average) a negative value for the volatility risk premium, since they compute VRP(t), as
the difference between RV(t + 1) and IV(t).
 A. Cipollini et al. / Journal of Empirical Finance 34 (2015) 34–44 41

Table 3
Volatility of the US shock.

Implied volatility Realized volatility Volatility risk premium

j=1 j=2 j=3 j=1 j=2 j=3 j=1 j=2 j=3

σus 0.720 (0.089) 0.854 (0.090) 1.069 (0.097) 0.480 (0.021) 0.625 (0.028) 0.977 (0.048) 0.802 (0.043) 0.927 (0.036) 1.200 (0.032)
σus,c 2.255 (0.310) 2.347 (0.265) 2.537 (0.306) 1.255 (0.215) 1.670 (0.193) 2.474 (0.222) 2.394 (0.279) 2.667 (0.283) 3.157 (0.242)

Note: The columns with the label j are those for the different levels of decomposition (e.g., frequency band). The point estimates (standard errors in parentheses) of σus
and σus,c are obtained from the estimation (through ML) of the factor model given by Eq. (2) or by Eq. (3).

3.3. Contagion results

The analysis of contagion from the US is carried out by taking into account heteroscedasticity bias arising from an increase in the
volatility of the US shock once we move from a non-crisis to a crisis period and, for a given regime, once we move to lower frequency
bands (see Table 3). First, we contrast the results obtained from the estimation of the δ's coefficients in the factor model given by
Eqs. (2) and (3) with the non-standardized coefficients b's obtained by estimating Eq.(6). For the five countries under investigation,
it may be observed in Tables 4 and 5 that while the estimation of the unadjusted factor (including b's) shows evidence of contagion
(during a regime change) from US to European markets at any level of decomposition and for any series (realized volatility, implied
volatility, and the volatility risk premia), the results from the factor model (controlling for heteroscedasticity bias) given by
Eqs. (2) and (3) are markedly different.
As for implied volatility, there is no evidence of contagion (except in the case of France at the third decomposition level). Based on
our results, it may be said that there is no “expected” volatility contagion. Overall, the empirical findings related to the factor-adjusted
model specification contrast with those of Jiang et al. (2012) and of Kenourgios (2014) who focus only on first differences in implied
volatilities (not on the levels).
Furthermore, the estimation of the factor-adjusted model shows evidence of contagion (see Tables 4 and 5) only for the realized
volatility series in the second and third level of decomposition. Overall, the empirical findings related to the factor-adjusted model
contrast with the study by Jung and Maderitsch (2014) who focus on log realized volatilities. The same authors find that once the
increase (over a crisis period) in the standard deviation of the log realized volatilities is controlled for, there is no evidence of
contagious effect between the Hong Kong, European, and US stock markets.
Finally, we observe that there is evidence of contagion between the US volatility risk premium and European (except Germany)
volatility risk premia, but only at the third level of decomposition. Consequently, we can say that, based on our results, we find
evidence of unexpected contagion (for a frequency band associated with a horizon between 8 and 16 days).
Summing up, our results point to contagion in realized volatilities in the 4–16 day time horizon; contagion is not in the risk-neutral
markets expectations (except France in the 8–16 day time horizon) and comes unexpectedly in all countries (except Germany in the
8–16 day time horizon).

4. Conclusions

In this paper, we analyzed whether, during the period immediately after the Lehman Brothers collapse, there was evidence of
contagion between the volatility of the US stock market and the volatilities of European stock markets (the UK, Germany, France,
the Netherlands, and Switzerland).
We used wavelet analysis, transforming the data into different time-scale components, in order to investigate contagion exploiting
information both in the time and the frequency domain.
Unlike previous studies on contagion using wavelet analysis, the focus of which was on pairwise wavelet (cross) correlation and/or
wavelet coherence across stock market index returns, we used factor model decomposition of the wavelet covariance matrix,
controlling for heteroscedasticity bias (see Forbes and Rigobon, 2002). The factor decomposition makes it possible to disentangle
the role played by an increase in the volatility of the US shock from the one measuring pure transmission in the transition from a
non-crisis to a crisis period and across different frequency bands. In a Monte Carlo simulation experiment, we highlight the potential
shortcomings and the misunderstanding that such a bias may give rise to, even if we move across different frequency bands for a given
regime.
In addition to controlling for heteroscedasticity bias, our approach improves the understanding of co-movements across different
regimes and frequency bands with respect to previous studies in at least two additional aspects. First, unlike previous studies, we do
not analyze contagion in stock index returns, but in the realized volatilities and their components (risk-neutral expectations about
volatility, i.e., implied volatility and an adjustment given by the risk aversion of investors, i.e., the volatility risk premium). In this
way, we shed light on the two different components of contagion: contagion anticipated by the market (coming from implied
volatility) and contagion not anticipated by the market (coming from the adjustment to risk aversion). Furthermore, by employing
a three-factor model specification (to control for an omitted variable bias), we are able to estimate the contribution of the US shock
and a European common shock in shaping volatility co-movements during a crisis and non-crisis regime and for different frequency
bands.
Overall, the results show no evidence of contagion (from the US) in market expectations (coming from implied volatility), except
in the case of France, and evidence of unanticipated contagion (coming from the volatility risk premium), except in the case of
 42
Table 4
Common factor loadings for volatility series (non-Eurozone countries).

A. Cipollini et al. / Journal of Empirical Finance 34 (2015) 34–44

Implied volatility Realized volatility Volatility risk premium

j=1 j=2 j=3 j=1 j=2 j=3 j=1 j=2 j=3

buk 0.348 (0.014) 0.549 (0.015) 0.809 (0.019) 0.166 (0.010) 0.302 (0.013) 0.585 (0.020) 0.359 (0.017) 0.559 (0.020) 0.845 (0.027)
buk,c 1.010 (0.303) 1.042 (0.297) 2.316 (0.337) 0.294 (0.131) 0.819 (0.153) 1.619 (0.234) 0.926 (0.349) 1.203 (0.349) 2.825 (0.408)
bswi 0.301 (0.011) 0.486 (0.013) 0.759 (0.017) 0.150 (0.011) 0.300 (0.014) 0.606 (0.021) 0.321 (0.016) 0.497 (0.019) 0.780 (0.027)
bswi,c 0.508 (0.246) 1.126 (0.335) 1.725 (0.390) 0.573 (0.134) 1.254 (0.180) 2.277 (0.283) 0.665 (0.296) 1.707 (0.406) 3.256 (0.539)
δuk 0.484 (0.053) 0.643 (0.026) 0.756 (0.046) 0.345 (0.025) 0.483 (0.025) 0.599 (0.024) 0.447 (0.042) 0.603 (0.024) 0.704 (0.027)
δuk,c −0.028 (0.136) −0.151 (0.108) 0.110 (0.083) 0.095 (0.083) 0.238 (0.093) 0.288 (0.059) −0.045 (0.126) −0.113 (0.094) 0.138 (0.083)
δswi 0.418 (0.091) 0.570 (0.039) 0.710 (0.016) 0.311 (0.044) 0.480 (0.043) 0.621 (0.038) 0.401 (0.078) 0.537 (0.048) 0.650 (0.033)
δswi,c −0.146 (0.130) −0.071 (0.128) −0.021 (0.090) 0.165 (0.114) 0.274 (0.113) 0.309 (0.066) −0.092 (0.130) 0.077 (0.116) 0.276 (0.123)
γ 0.453 (0.020) −0.471 (0.019) 0.538 (0.021) 0.365 (0.025) −0.464 (0.025) −0.696 (0.030) 0.575 (0.028) −0.675 (0.031) −0.931 (0.039)
γc 0.990 (0.241) −1.206 (0.324) 0.922 (0.351) 0.826 (0.120) −0.798 (0.098) 2.334 (0.138) 1.260 (0.251) 2.851 (0.425) 3.120 (0.291)

Note: The columns with the label j are those for the different levels of decomposition (e.g., frequency band). Coefficients measuring the variation in the crisis regime are in bold. The point estimates (standard errors in parentheses)
of the non-standardized US common factor loadings coefficients bi and bi,c are obtained from the ML estimation of a factor model described by Eq. (6). The point estimates (standard errors in parentheses) of the non-standardized
non-Eurozone common factor loadings coefficients γ and γc (restricted to be the same for UK and Switzerland to achieve exact identification) are obtained from the ML estimation of a factor model described by Eq. (2). The
contribution of the US and of the non-Eurozone common shock to volatility co-movements (during the non-crisis period) is measured by bi and γ, respectively. The contribution of the US and European common shock to volatility
co-movements (during the crisis period) is measured by bi + bi,c and γ + γc, respectively.
The standardized factor loadings coefficients δ's are obtained from the ML estimation of the factor model given by Eq. (2). The coefficients measuring contagion from the US (controlling for heteroscedasticity bias) are given by δi,c.
 A. Cipollini et al. / Journal of Empirical Finance 34 (2015) 34–44 43

Table 5
Common factor loadings for volatility series (Eurozone countries).

Implied volatility Realized volatility Volatility risk premium

j=1 j=2 j=3 j=1 j=2 j=3 j=1 j=2 j=3

bger 0.403 (0.014) 0.624 (0.016) 0.900 (0.020) 0.247 (0.013) 0.339 (0.016) 0.626 (0.024) 0.458 (0.019) 0.657 (0.022) 0.977 (0.031)
bger,c 0.741 (0.280) 1.849 (0.338) 2.403 (0.399) 0.335 (0.121) 0.810 (0.158) 1.644 (0.268) 1.021 (0.326) 2.101 (0.387) 2.965 (0.479)
bfra 0.292 (0.014) 0.474 (0.017) 0.725 (0.020) 0.225 (0.012) 0.355 (0.015) 0.633 (0.022) 0.316 (0.018) 0.502 (0.022) 0.770 (0.029)
bfra,c 1.107 (0.356) 1.719 (0.361) 2.541 (0.369) 0.473 (0.149) 1.073 (0.187) 2.088 (0.297) 0.773 (0.395) 1.697 (0.389) 3.151 (0.445)
bned 0.374 (0.015) 0.599 (0.017) 0.915 (0.022) 0.214 (0.012) 0.364 (0.016) 0.685 (0.024) 0.371 (0.019) 0.571 (0.023) 0.878 (0.032)
bned,c 0.966 (0.228) 1.476 (0.264) 2.142 (0.314) 0.506 (0.155) 1.357 (0.207) 2.496 (0.359) 0.961 (0.306) 2.024 (0.339) 3.269 (0.426)
δger 0.560 (0.035) 0.731 (0.021) 0.842 (0.031) 0.514 (0.044) 0.543 (0.028) 0.641 (0.043) 0.572 (0.043) 0.709 (0.034) 0.814 (0.033)
δger,c −0.176 (0.116) 0.034 (0.117) 0.074 (0.096) 0.122 (0.017) 0.241 (0.110) 0.243 (0.092) −0.108 (0.115) 0.058 (0.103) 0.091 (0.093)
δfra 0.406 (0.048) 0.556 (0.027) 0.678 (0.027) 0.468 (0.049) 0.569 (0.039) 0.648 (0.025) 0.394 (0.031) 0.542 (0.031) 0.642 (0.032)
δfra,c 0.064 (0.158) 0.123 (0.134) 0.228 (0.102) 0.015 (0.099) 0.157 (0.104) 0.236 (0.072) −0.053 (0.140) 0.070 (0.106) 0.258 (0.094)
δned 0.519 (0.093) 0.702 (0.041) 0.856 (0.026) 0.445 (0.042) 0.582 (0.037) 0.701 (0.030) 0.463 (0.054) 0.616 (0.034) 0.732 (0.032)
δned,c −0.069 (0.119) −0.060 (0.086) −0.008 (0.081) 0.141 (0.094) 0.274 (0.081) 0.310 (0.053) −0.046 (0.107) 0.106 (0.080) 0.220 (0.072)
γger 0.510 (0.029) 0.552 (0.027) 0.642 (0.027) 0.461 (0.039) 0.565 (0.037) 0.847 (0.047) 0.688 (0.036) 0.824 (0.041) 1.115 (0.051)
γger,c 1.502 (0.396) 1.518 (0.315) 1.389 (0.328) 0.506 (0.194) 0.577 (0.138) 0.557 (0.161) 1.389 (0.410) 1.186 (0.372) 1.061 (0.296)
γfra 0.428 (0.024) 0.506 (0.022) 0.582 (0.024) 0.506 (0.030) 0.620 (0.031) 0.893 (0.039) 0.635 (0.032) 0.797 (0.037) 1.068 (0.045)
γfra,c 1.375 (0.416) 1.733 (0.337) 1.534 (0.326) 1.043 (0.194) 0.998 (0.123) 0.987 (0.108) 2.034 (0.439) 2.001 (0.289) 1.698 (0.248)
γned 0.545 (0.032) 0.554 (0.032) 0.639 (0.034) 0.498 (0.027) 0.627 (0.029) 0.938 (0.039) 0.733 (0.044) 0.869 (0.042) 1.209 (0.055)
γned,c 0.823 (0.217) 0.821 (0.202) 0.587 (0.209) 0.884 (0.183) 0.568 (0.124) 0.403 (0.142) 1.314 (0.219) 0.825 (0.267) 0.261 (0.222)

Note: The columns with the label j are those for the different levels of decomposition (e.g., frequency band). Coefficients measuring the variation in the crisis regime are
in bold. The point estimates (standard errors in parentheses) of the non-standardized US common factor loadings coefficients bi and bi,c are obtained from the ML
estimation of a factor model described by Eq. (6). The point estimates (standard errors in parentheses) of the non-standardized Eurozone common factor loadings
coefficients γ and γc are obtained from the ML estimation of a factor model described by Eq. (3). The contribution of the US and of the non-Eurozone common
shock to volatility co-movements (during the non-crisis period) is measured by bi and γ, respectively. The contribution of the US and European common shock to
volatility co-movements (during the crisis period) is measured by bi + bi,c and γ + γc, respectively.
The standardized factor loadings coefficients δ's are obtained from the ML estimation of the factor model given by Eq. (3). The coefficients measuring contagion from the
US (controlling for a heteroscedasticity bias) are given by δi,c,

Germany. As a result, the empirical evidence suggests that the driver of contagion (from the US) in “real market movements” (e.g.,
realized volatilities) has to be addressed (for a frequency band associated with a horizon between 8 and 16 days) to the risk premium
component of “market fear” (volatility risk premium). Moreover, the European common shock is an important driver of European
volatility co-movements mainly during the tranquil period, except for market expectations (implied volatility), where the US
common shock plays a significant role. In the crisis period, the US common shock is more important than the European one, mainly
for Eurozone countries for both the unexpected and expected components of realized volatility.

Acknowledgements

The authors wish to thank two anonymous referees. The usual disclaimer applies. S. Muzzioli gratefully acknowledges financial
support from the Fondazione Cassa di Risparmio di Modena, for the project “Volatility modeling and forecasting with option prices:
the proposal of a volatility index for the Italian market” and MIUR. A. Cipollini gratefully acknowledges financial support from
MIUR, for the PRIN project “Forecasting economic and financial time series: understanding the complexity and modeling structural
change.”

References

Ait-Sahalia, J., Cacho-Diaz, J., Laeven, R.J.A., 2010. Modeling financial contagion using mutually exciting jump processes. NBER Working Paper No. 15850.
Andersen, T.G., Bollerslev, T., 1997. Intraday periodicity and volatility persistence in financial markets. J. Empir. Finance 4, 115–158.
Andersen, T.G., Bollerslev, T., Diebold, F.X., Labys, P., 2001. The distribution of exchange rate volatility. J. Am. Stat. Assoc. 96, 42–55.
Baillie, R.T., Bollerslev, T., Mikkelsen, H.O., 1996. Fractionally integrated generalized autoregressive conditional heteroskedasticity. J. Econ. 74, 3–30.
Bandi, F.M., Perron, B., 2006. Long memory and the relation between implied and realized volatility. J. Financ. Econ. 4, 636–670.
Bekaert, G., Horeova, M., 2014. The VIX, the variance premium and stock market volatility. ECB working paper, 1675.
Bodard, V., Candelon, B., 2009. Evidence of interdependence and contagion using a frequency domain framework. Emerg. Mark. Rev. 10 (2), 140–150.
Bollerslev, T., Osterrieder, D., Sizova, N., Tauchen, G., 2013. Risk and return: long-run relationships, fractional cointegration, and return predictability. J. Financ. Econ.
108, 409–424.
Buraschi, A., Trojani, F., Vedolin, A., 2014. When uncertainty blows in the orchard: comovement and equilibrium volatility risk premia. J. Financ. 69 (1), 101–137.
Caporale, G., Cipollini, A., Spagnolo, N., 2005. Testing for contagion: a conditional correlation analysis. J. Empir. Finance 12, 476–489.
Carr, P., Wu, L.R., 2009. Variance risk premiums. Rev. Financ. Stud. 22 (3), 1311–1341.
Comte, F., Renault, E., 1998. Long memory in continuous-time stochastic volatility models. Math. Financ. 8, 291–323.
Corsi, F., 2009. A simple approximate long memory model of realized volatility. J. Financ. Econ. 7 (2), 174–196.
Daubechies, I., 1992. Ten lectures on wavelets. Society for Industrial and Applied Mathematics, Philadelphia.
Dungey, M., Martin, V.L., 2007. Unravelling financial market linkages during crises. J. Appl. Econ. 22, 89–119.
Dungey, M., Fry, R., Gonzalez-Hermosillo, B., Martin, V.L., 2005. Empirical modelling of contagion: a review of methodologies. Quant. Finan. 5 (1), 9–24.
Dungey, M., Milunovich, G., Thorp, V., 2010. Unobservable shocks as carriers of contagion. J. Bank. Financ. 34, 1008–1021.
Favero, C.A., Giavazzi, F., 2002. Is the International Propagation of Financial shocks nonlinear evidence from the ERM. J. Int. Econ. 57 (1), 231–246.
Forbes, K., Rigobon, R., 2002. No contagion, only interdependence: measuring stock market co-movements. J. Financ. 57 (5), 2223–2261.
Gallegati, M., 2012. A wavelet-based approach to test for financial contagion. Comput. Stat. Data Anal. 56, 3491–3497.
44 A. Cipollini et al. / Journal of Empirical Finance 34 (2015) 34–44

Jiang, G.J., Konstantinidi, E., Skiadopoulos, G., 2012. Volatility spillovers and the effect of news announcements. J. Bank. Financ. 36 (8), 2260–2273.
Jung, R.C., Maderitsch, R., 2014. Structural breaks in volatility spillovers between international financial markets: contagion or mere interdependence? J. Bank. Financ.
47, 331–342.
Kenourgios, D., 2014. On financial contagion and implied market volatility. Int. Rev. Financ. Anal. 34, 21–30.
Muzzioli, S., 2013a. The forecasting performance of corridor implied volatility in the Italian market. Comput. Econ. 41 (3), 359–386.
Muzzioli, S., 2013b. The information content of option based forecasts of volatility: evidence from the Italian stock market. Q. J. Financ. 3 (1) (1350005 (46 pages)).
Percival, D.B., Walden, A.T., 2000. Wavelet Methods for Time Series Analysis. Cambridge University Press.
Pesaran, H., Pick, A., 2007. Econometric issues in the analysis of contagion. J. Econ. Dyn. Control. 31 (4), 1245–1277.
Ranta, M., 2013. Contagion among major world markets: a wavelet approach. Int. J. Manag. Financ. 9 (2), 133–150.
Rua, A., Nunes, L.C., 2009. International co-movement of stock market returns: a wavelet analysis. J. Empir. Financ. 16 (4), 632–639.
Whitcher, B., Guttorp, P., Percival, D.B., 2000. Wavelet analysis of covariance with application to atmospheric time series. J. Geophys. Res. 105 (D11), 14941–14962.