Quantile Connectedness: Modelling Tail Behaviour

in the Topology of Financial Networks∗

Tomohiro Ando
Melbourne Business School, University of Melbourne
Matthew Greenwood-Nimmo
Department of Economics, University of Melbourne
Centre for Applied Macroeconomic Analysis, Australian National University
Yongcheol Shin
Department of Economics and Related Studies, University of York

December 30, 2017

Abstract
We develop a new technique to estimate vector autoregressions by quantile regression.
A factor structure is used to remove cross-section correlation in the residuals such that
the system can be estimated on an equation-by-equation basis using existing quantile re-
gression toolboxes. We use our model to study credit risk spillovers among a panel of 18
sovereigns and their respective financial sectors between January 2006 and February 2012.
We show that idiosyncratic credit risk shocks do not propagate strongly at the median but
that powerful spillovers occur in both tails. Furthermore, rolling sample analysis reveals
marked time-varying tail-dependence. These important features of credit risk transmission
are obscured in models estimated using conventional conditional mean estimators.

JEL Codes: C31, C32, C58, E44, G01.

Keywords: Network Modelling; Quantile Vector Autoregression with Common Factors;
Quantile Connectedness; Financial–Sovereign Credit Risk Transmission; Tail-Dependence.

∗
Address for correspondence: M.J. Greenwood-Nimmo, 3.12 Faculty of Business and Economics, 111 Barry Street,
University of Melbourne, VIC 3053, Australia. Email: matthew.greenwood@unimelb.edu.au. Tel: +61 (0)3 8344 5354.
This work was supported by the Australian Research Council [grant number DE150100708]. We are grateful for many
helpful interactions with Jinseo Cho, David Harris, Jingong Huang, Christian Manicaro, Faek Menla Ali, Viet Nguyen,
Barry Rafferty, Chris Skeels, Artur Tarassow and Ben Wong and for the helpful comments of seminar participants at the
Universities of Melbourne and York. Greenwood-Nimmo acknowledges the hospitality of the University of York during a
sequence of short visits between 2015 and 2017. Shin acknowledges the hospitality of the University of Melbourne during
an extended visit in 2017. Computer programmes written for R are available on request. The usual disclaimer applies.

Declarations of interest: None.

Role of funding sources in the preparation of this manuscript: None.

1
1 Introduction

The topology of financial networks is central to the study of systemic risk. An adverse idiosyn-
cratic shock to one part of the financial system poses a threat to systemic stability if there are
linkages through which it can propagate to other parts of the system. Measuring the nature
and strength of financial market linkages is not only important for risk management strategies
but also to inform the policy response to systemic crises. Several techniques for the economet-
ric analysis of financial networks have been proposed in recent years, including those based on
Granger-causality and on innovation accounting (e.g. Billio, Getmansky, Lo and Pelizzon, 2012;
Diebold and Yilmaz, 2009, 2014; Alter and Beyer, 2014). A common feature of this literature
is the reliance on conventional conditional mean estimators such as ordinary least squares. The
result is an estimate of the average network structure which prevails when an average shock
affects the system. However, systemic shocks are likely to be much larger than average and it
need not be the case that large shocks propagate in the same way as smaller shocks. To ad-
dress this possibility, we develop a new framework which uses regression quantiles to investigate
whether the topology of a network changes with the size of the shocks that affect the system.
Our framework builds upon that of Diebold and Yilmaz (2009, 2014), where the m (m − 1)
bilateral interactions among an m-vector of variables, y t , are approximated by the h-step-ahead
forecast error variance decomposition (FEVD) of an underlying vector autoregression (VAR) in
y t . Consequently, the Diebold–Yilmaz framework answers the question ‘how much of the future
uncertainty associated with variable i can be attributed to shocks coming from variable j?’.
A major advantage of the Diebold–Yilmaz framework relative to the Granger-causal network
analysis of Billio et al. (2012) is that the resulting network is not only directed but also weighted
and therefore provides an estimate of the strength of bilateral spillover effects. It has proven
to be an influential technique, with applications including the connectedness among equity
markets (e.g. Diebold and Yilmaz, 2009), foreign exchange markets (e.g. Barunı́k, Kočenda and
Vácha, 2016; Greenwood-Nimmo, Nguyen and Rafferty, 2016a), systemically important financial
institutions (e.g. Demirer, Diebold, Liu and Yilmaz, 2017) and sovereign and corporate credit
spreads (e.g. Bostanci and Yilmaz, 2015; Greenwood-Nimmo, Huang and Nguyen, 2017).
Rolling sample analysis is typically used to track the evolution of the network over time,
with abrupt increases in connectedness typically being interpreted in relation to systemic shocks.
However, there is a tension between this interpretation and the fact that existing applications
of the Diebold–Yilmaz framework rely on a range of conventional estimators, including least

2
squares (e.g. Diebold and Yilmaz, 2009, 2014), the least absolute shrinkage and selection op-
erator, or LASSO (e.g. Greenwood-Nimmo et al., 2017) and elastic net regularisation (e.g.
Demirer et al., 2017). Each of these estimators evaluates the relationship between y t and

z t = y t−1 , y t−2 , . . . , y t−p at the mean of the conditional distribution of y t |z t . The parame-
ters of a VAR model estimated by least squares, for example, capture the dynamic relationships
among the variables in y t under the assumption that average shocks — which are precisely
equal to zero by definition — affect the system. The tension arises because systemic shocks are
likely to be larger than average. Consequently, studies in this literature must implicitly assume
that the relationship which prevails at the conditional mean can be generalised to the entire
conditional distribution. This is a strong assumption but it is vital if rolling sample analysis
based on conditional mean estimators is to provide a valid signal regarding the impact of large
systemic shocks.
We relax this assumption by developing a new approach based on the premise that if one
wishes to analyse the network structure associated with a large shock — for example a shock
in the ninety-fifth percentile of the size distribution of shocks — then one must set aside condi-
tional mean estimators and instead fit the VAR model at the ninety-fifth percentile by quantile
regression. Following Koenker and Bassett (1978), quantile regression can be used to evaluate
the dependence of y t on z t over the entire range of the conditional distribution of y t |z t . At
the time of writing, two approaches to the estimation of quantile VAR models have been pro-
posed by Cecchetti and Li (2008) and Schüler (2014), respectively.1 Cecchetti and Li develop
an equation-by-equation estimation framework for VAR models with cross-sectionally corre-
lated residuals, while Schüler develops a Bayesian framework for the analysis of structural VAR
models.
Our framework is distinct from these existing methods by virtue of our treatment of the
VAR residuals. We assume that the cross-sectional correlation of the VAR residuals is driven by
a finite number of common factors. This assumption serves two purposes. First, by purging the
common component from the VAR residuals, we are able to isolate the idiosyncratic shock to
each variable in the system. Not only does this align our framework with the large literature on
systematic vs. idiosyncratic variations in finance — an important consideration if one wishes to
use network models to analyse the propagation of idiosyncratic shocks, for example — but it also
1
The method of Cecchetti and Li (2008) has subsequently been applied by Linnemann and Winkler (2016)
and Zhu, Su, Guo and Ren (2016). A related paper by Chavleishvili and Manganelli (2016) develops a framework
for quantile impulse response analysis of a bivariate system with one endogenous and one exogenous variable.

3
addresses the likelihood that the failure to account for sources of common variation may generate
substantial biases in the analysis of networks. Specifically, if an omitted common factor affects
all variables in the VAR system, then the proportion of the forecast error variance which should
rightly be attributed to that factor must instead be attributed to one or more of the endogenous
variables in the system, thereby upwardly biasing the estimated spillover effects.2 The issue
of spurious Granger causality arising due to the omission of sources of common variation is
well-known but the impact of omitted common factors on Diebold–Yilmaz networks has not
been adequately explored to date (Greenwood-Nimmo, Nguyen and Shin, 2016b). Second, the
use of a factor structure simplifies the estimation problem substantially because it renders the
system of equations that comprise the VAR model cross-sectionally independent. Given that
the errors are uncorrelated across equations, the factor VAR system can be directly estimated
on an equation-by-equation basis using the standard quantile regression commands built in to
many statistical software packages (such as Roger Koenker’s quantreg package in R).
Our approach answers a modified version of the question addressed by the Diebold–Yilmaz
framework: ‘how much of the future uncertainty associated with variable i can be attributed
to idiosyncratic shocks coming from variable j as the shock size varies?’. Consequently, our
technique is ideally suited to the study of the propagation of idiosyncratic risk shocks and
contagion, the latter of which is often defined in relation to the difference in the way that
shocks propagate during rare events relative to normal times (e.g. Caporin, Pelizzon, Ravazzolo
and Rigobon, 2013). We introduce the term quantile connectedness to distinguish between our
quantile-based approach to network analysis and the established mean-based approach.
We apply our technique to study spillovers of idiosyncratic credit risk between sovereigns and
national financial sectors over the period January 2006 to February 2012, both within and across
borders. The study of credit risk transmission has become an important focus among policy
institutions, with a particular concern for the emergence of feedback loops and destabilising
spirals in credit markets (e.g. International Monetary Fund, 2013, pp. 65-6). Acharya, Drechsler
and Schnabl (2014) provide compelling evidence of just such an adverse feedback effect. They
demonstrate that the financial sector bailouts undertaken by many developed countries in 2008
amounted to a substantial transfer of private sector credit risk onto the public sector at a time
of rapid public debt growth. This combination ultimately proved untenable in several countries
2
In this paper, we focus on analysing direct spillovers of idiosyncratic credit risk having controlled for common
systematic factors. Although it is not our focus here, one may also be interested in analysing the indirect
propagation of shocks via the common factors. The Diebold–Yilmaz framework accounts for both direct and
indirect linkages but it does not allow them to be analysed separately.

4
and led to a resurgence of systemic risk driven by the emergence of an adverse feedback loop
between sovereign credit risk and financial sector credit risk. Variants of the Diebold–Yilmaz
technique have been applied to the analysis of the sovereign–financial credit risk network by
Alter and Beyer (2014) and Greenwood-Nimmo et al. (2017), although our paper represents the
first attempt to study quantile-variation in the structure of the credit risk network.
We follow the existing literature and use credit default swap (CDS) spreads to measure credit
risk.3 The existence of a factor structure in the cross-section of CDS spreads has been docu-
mented by Pan and Singleton (2008), Longstaff, Pan, Pedersen and Singleton (2011), Fender,
Hayo and Neuenkirch (2012) and Ang and Longstaff (2013), among others. To isolate the id-
iosyncratic variation in the CDS spreads, our model includes the following observed common
factors, which draw on those identified by Longstaff et al. (2011): (i) global stock market fac-
tors measured via the three Fama-French factors; (ii) US macroeconomic conditions proxied
by the five-year US Treasury yield; (iii) funding liquidity measured via the TED spread and
the Euribor-DeTBill spread; (iv) investor risk appetite captured by the S&P 500 variance risk
premium, the five-year US Treasury term premium and the US corporate investment grade
and high yield spreads; (v) a selection of ITRAXX credit default indices which capture pan-
European credit risk factors; and (vi) bilateral exchange rate fluctuations measured for each
currency against the US Dollar.
Our first finding is that the topology of the credit risk network varies significantly with
the shock size. We find that the effects of small idiosyncratic credit risk shocks in the central
70% of the conditional distribution are predominantly localised. Bilateral spillovers account
for no more than 9% of the five-days-ahead forecast error variance (FEV) over this range and
for just 4.34% at the conditional median. However, this is not true of large shocks. Large
adverse shocks in the right tail of the conditional distribution spread forcefully through the
financial system, with bilateral spillovers accounting for 22.87% and 83.32% of FEV at the
ninety-fifth and ninety-ninth conditional quantiles, respectively. This finding accords with the
existing evidence on increased financial market comovements under adverse conditions (e.g. Ang
and Bekaert, 2002). Interestingly, we also find evidence that large beneficial shocks propagate
strongly, with bilateral spillovers accounting for 21.07% and 76.57% of the FEV at the fifth
3
A CDS contract operates like an insurance agreement negotiated between two parties, one of whom holds
a risky bond and the other of whom agrees to absorb losses arising in the event that the bond issuer defaults.
The CDS spread defines the price that the bondholder must pay to the protection seller. Blanco, Brennan and
Marsh (2005) and Gyntelberg, Hördahl, Ters and Urban (2013) show that the CDS market is the leading forum
for credit risk price discovery, providing more timely signals of changes in the credit risk environment than bond
yield spreads.

5
and first conditional quantiles, respectively. The finding that large shocks in both tails spillover
strongly is consistent with the existing literature on good and bad contagion (e.g. Londono,
2016). Crucially, when the model is estimated at the conditional mean by OLS, this quantile
variation is averaged out and bilateral spillovers are found to account for 11.49% of the FEV.
This result arises by definition because OLS is equivalent to an equally-weighted average of
quantile regression estimators over the entire conditional distribution. This cautions against
the use of network models estimated using conditional mean estimators to analyse the spillovers
associated with extreme events.
Our second result is that the adverse feedback loop documented by Acharya et al. (2014)
manifests as a marked intensification of the bidirectional feedback between each sovereign and
its domestic financial sector in the presence of large adverse shocks in the right tail of the
conditional distribution. However, we once again find that this behaviour is not unique to
adverse shocks — there is a similar intensification of bidirectional feedback in the left tail. This
leads us to conclude that the feedback loop described by Acharya et al. (2014) is associated
with a vicious circle which leads to the amplification of bad news but that the same feedback
effect can rapidly reduce credit spreads when beneficial shocks occur.
Our final result is derived from rolling sample estimation of our model. In this way, we
demonstrate that the dependence structure that exists among the cross-section of sovereigns
and financial institutions displays marked time-variation. The time-variation of bilateral credit
risk spillovers has already been demonstrated at the conditional mean (e.g. Greenwood-Nimmo
et al., 2017). However, we are the first to demonstrate that the time-variation in the net-
work topology observed in the tails of the conditional distribution does not closely resemble
the time-variation observed at either the conditional mean or median. This is an important
finding given the relevance of tail-dependence for financial supervision and risk management
(e.g. Betz, Hautsch, Peltonen and Schienle, 2016) and it suggests that the implications derived
from network models evaluated by conventional conditional mean estimators cannot necessarily
be generalised to the tails. We show that major adverse events are associated with an increase
in average connectedness but that their effects on the tails differs. Specifically, we find that
right-tail-dependence increases while left-tail-dependence decreases. This combination implies
that bad news leads to an increase in the propensity for the destabilising propagation of ad-
verse shocks coupled with a reduction in the stabilising propagation of beneficial shocks. By
contrast, stabilising policy interventions which reduce average connectedness tend to increase

6
left-tail-dependence while reducing right-tail-dependence. These findings lead us to develop a
new measure of relative tail-dependence which draws attention to the negative correlation of
left- and right-tail-dependence. We suggest that this negative correlation may arise from the
aggregate behaviour of market participants if the information revealed by a major event in
either tail causes a non-trivial proportion of market participants to focus disproportionately on
further events occurring in that tail while paying less attention to events in the other tail.
Aside from the work on empirical network modelling, we wish to highlight three strands of
literature to which our paper is related. First, our use of quantile regression to study extreme
events closely resembles a branch of the systemic risk literature which is well-represented by
Caporin et al. (2013) and Betz et al. (2016), both of which use quantile regression to study the
propagation of adverse shocks through the financial markets. Second, our concept of quantile
connectedness has a natural link to value-at-risk (VaR) and associated concepts such as CoVaR
(Adrian and Brunnermeier, 2016). VaR is widely used by investors to measure the potential
loss that they may endure on their positions due to adverse shocks over a defined horizon and
at a predetermined confidence level. To illustrate the conceptual link, assume that investors
maintain sufficient capital reserves to cover the VaR at the 95% confidence level. In this case,
losses up to the VaR can be absorbed within the capital buffer. However, the probability of
an adverse shock sufficiently severe to generate a loss in excess of the VaR is 5%. Investors
may find the losses caused by such large shocks untenable, raising the possibility of default
and insolvency. As a result, the transmission of risk among counterparties may be considerably
stronger in the case of large shocks than small shocks. This offers a partial explanation of the
quantile-variation that we document in the topology of the credit risk network. Finally, the
notion that the topology of a network may vary with the size of the shocks affecting the system
is related to Acemoglu, Ozdaglar and Tahbaz-Salehi’s (2015) insight that a phase change may
occur whereby dense financial networks are resilient to small shocks but can be vulnerable to
cascading failures in the event of a large adverse shock.
This paper proceeds as follows. In Section 2, we outline the quantile factor VAR model and
derive the associated generalised forecast error variance decomposition which is then used to
construct network statistics. We provide a detailed discussion of the construction and properties
of our dataset in Section 3 before presenting our estimation results in Section 4. We conclude
in Section 5.

7
2 Quantile Connectedness

In this section, we propose a new framework for the graphical analysis of VAR models estimated
at a given conditional quantile, τ ∈ (0, 1). To this end, we develop a framework for the equation-
by-equation estimation of a VAR model by quantile regression where a factor structure is used
to distinguish between the common and idiosyncratic components of the error process. We then
derive the associated h-step-ahead forecast error variance decomposition, which forms the basis
for network analysis in the tradition of Diebold and Yilmaz (2009, 2014).

2.1 The Quantile VAR Model

Consider a multi-country environment where countries are indexed by i = 1, 2, . . . , N and where

time periods are indexed by t = 1, 2, . . . T . Let y it = (∆sit , ∆fit )0 , where sit and fit denote the
sovereign and financial sector CDS spreads for the ith country, which will be discussed in detail
below. Moreover, let the m × 1 vector y t = (y 01t , y 02t , . . . , y 0N t )0 collect the credit spreads for all
N countries in the system, where m = 2N . Implementation of the Diebold–Yilmaz technique
requires us to specify a model capturing the dynamics of y t . To this end, consider the following
VAR model which expresses the sovereign and financial sector CDS spreads for the ith country
as a function of the lagged sovereign and financial sector CDS spreads for every country in the
system:
p
X
y it = µi + Φij y t−j + eit (1)
j=1

where µi is a vector of intercepts, Φij is the parameter matrix at the jth lag and the regression
2×1 2×m
residuals eit ∼ (0, Σi ), where Σi is a positive definite covariance matrix. By stacking (1) for
2×1 2×2
all countries in the system, we obtain the following VAR system for y t :

p
X
yt = µ + Φj y t−j + et (2)
j=1

0
where µ = (µ01 , µ02 , . . . , µ0N )0 is a vector of intercepts, Φj = Φ01j , Φ02j , . . . , Φ0N j is the jth
m×1 m×m
autoregressive parameter matrix and the residual process et = (e01t , e02t , . . . , e0N t )0 ∼ (0, Σ)
m×1
where Σ is positive definite.
m×m
The order of the VAR model (2), p, is estimated consistently using the Schwarz Informa-
tion Criterion and should be sufficient to yield serially uncorrelated residuals. Nonetheless, the
residuals will typically exhibit contemporaneous cross-section correlation and so Σ is likely to

8
be non-diagonal.4 Conditional mean estimation of unrestricted VARs of this type is straight-
forward, and can be achieved on an equation-by-equation basis using ordinary least squares.
Conditional quantile estimation is more challenging, however. Given that each equation in
(2) shares a common set of right hand side variables, the estimation problem has a seemingly
unrelated regressions (SUR) structure which has led several authors to pursue an equation-by-
equation quantile regression estimation strategy (e.g. Cecchetti and Li, 2008; Linnemann and
Winkler, 2016; Zhu et al., 2016). However, from equation (5) in Zellner and Ando (2010), it is
clear that this approach sets the off-diagonal elements of the covariance matrix of et to zero,
which amounts to the assumption of cross-section independence. The failure to adequately
account for the cross-section correlation among the regression residuals is likely to bias the
resulting parameter estimates.
We address this issue by modelling the cross-section correlation in the residuals as the result
of an f × 1 vector of common factors. In this way, we are able to separate the systematic and
idiosyncratic components of the error process, thereby aligning our approach with the large
literature on the distinction between systematic and idiosyncratic risks (see Feldhütter and
Nielsen, 2012, for a recent example which focuses on credit spreads). Furthermore, where one
wishes to analyse spillover effects between variables, it is important to focus on the idiosyncratic
variation having purged any systematic variation or else one is likely to obtain a biased estimate
of the spillover intensity, a phenomenon that we demonstrate in Section 4.1 below. Formally,
we assume that:
eit = λ0i f t + v it (3)

and, by extension, that:

et = Λf t + v t (4)

0
where f t is a vector of common factors, Λ = λ01 , λ02 , . . . , λ0N is a matrix of heterogeneous
f ×1 m×f
factor loadings and where v t = (v 01t , v 02t , . . . , v 0N t )0 ∼ (0, Ω) contains the idiosyncratic compo-
m×1
nents of et which are mutually uncorrelated such that Ω is diagonal. Combining (2) and (4)
m×m
4
When using VARs for macroeconomic analysis, it is common to transform the reduced form VAR (2) into a
structural counterpart with uncorrelated disturbance terms to which one can attach a structural interpretation.
This is typically achieved either by imposing Wold-causality as in Sims (1986), short-run exclusion restrictions
as in Blanchard and Watson (1986) or long-run restrictions as in Blanchard and Quah (1989). However, the
application of these traditional methods becomes increasingly challenging as the dimension of the model increases.

9
yields the following factor VAR model:

p
X
yt = µ + Φj y t−j + Λf t + v t (5)
j=1

The factors, f t , may be either observed or latent — we limit our attention to the case of observed
factors. This confers at least two benefits relative to unobserved factor approaches such as
principal components analysis (e.g. Bai, 2003, 2009) or the common correlated effects framework
of Pesaran (2006). First, it is often easier to achieve an economically meaningful interpretation
of observed factors than latent factors. Second, an observed factor structure is parsimonious and
can be feasibly implemented in relatively small datasets. By contrast, the reliable estimation
of principal components requires a relatively high-dimensional dataset. Likewise, the common
correlated effects approach can lead to a proliferation of estimated parameters because the
unobserved common factors are approximated by augmenting the model with p lags of the
cross-section averages of y t .
Given that the error terms in (5) are uncorrelated by construction, the model can be es-
timated by equation-by-equation quantile regression without loss of generality. We write the
quantile factor VAR (QFVAR) model evaluated at the τ th conditional quantile as follows:

p
X
y t = µ(τ ) + Φj(τ ) y t−j + Λ(τ ) f t + v t(τ ) (6)
j=1

where τ ∈ (0, 1) is a given quantile index. Following Koenker and Xiao (2006), we assume
that the optimal lag order for the conditional mean model remains valid at every conditional


quantile. Assuming that Qτ v t(τ ) |Ft−1 = 0, where Ft−1 denotes the information set available
at time t − 1, then:5

p
X
Qτ (y t |Ft−1 ) = µ(τ ) + Φj(τ ) y t−j + Λ(τ ) f t (7)
j=1

To illustrate the quantile regression procedure, it is useful to first re-write the ith equation of
(6) compactly as follows:
yit = β 0i(τ ) z t + vit(τ ) (8)

for i = 1, 2, . . . , m where z t is an (mp + f + 1)×1 vector containing all of the regressors including
5
Note that the residual covariance matrix, Ω, is the same for all τ . This assumption conforms with the
precedent in the frequentist literature (e.g. Cecchetti and Li, 2008), although a Bayesian technique for the
estimation of a VAR model with a quantile-dependent covariance matrix has been proposed by Schüler (2014).

10
the intercept and β i(τ ) contains the corresponding regression coefficients evaluated at the τ th
conditional quantile. We impose the usual assumption that the conditional quantile model is
correctly specified:



E ψτ vit(τ ) |z t = 0 (9)

R βi(τ ) zt
where ψτ (z) = τ − 1[z≤0] . This assumption implies that −∞ fyit |zt (t|z t )dt = τ , where
fyit |zt (t|z t ) is the density of yit conditional on z t . For a fixed value of τ , the single-step quantile
regression estimates are obtained as follows:

T
X  
min ξτ yit − β 0i(τ ) z t (10)
β i(τ )
t=1

where ξτ (z) is the check loss function defined as ξτ (z) = z(τ −1[z≤0] ) as in Koenker and Hallock
(2001). The solution to this minimisation, denoted β
b , is consistent and asymptotically
i(τ )

normal under the assumption that the quantile specification is correct and subject to a number
of mild regularity conditions (see Koenker, 2005, for further details).

2.2 Quantile Forecast Error Variance Decomposition

The functional form of the factor VAR model (5) is identical to a VARX(p, 0) model — that is,
a VAR model with p lags of a set of endogenous variables augmented with the contemporaneous
values of a set of exogenous variables — in the special case that the residual covariance matrix
is diagonal.6 Two approaches to innovation accounting with VARX models have been pursued
in the literature: (i) to conduct innovation accounting by conditioning on the values of the
exogenous variables; and (ii) to augment the VARX specification with an auxiliary marginal
model for the exogenous variables and then to undertake innovation accounting with respect
to the augmented system. Given that our interest is in modelling the propagation of the
idiosyncratic shocks as opposed to the effects of systemwide shocks, we pursue the former option.
In addition, as the computation of the FEVD for VARX models evaluated at the conditional
mean is well established (e.g. Garratt, Lee, Pesaran and Shin, 2006), we limit our attention to
the QFVAR case.
6
The model may be easily extended to include lags of the factors — i.e. one may estimate a VARX(p, q)
model with q > 0 — although this would substantially increase the number of parameters to be estimated.

11
 We start by re-writing the QFVAR model (6) as follows:

p
X
yt = Φj(τ ) y t−j + Λ∗(τ ) f ∗t + v t(τ ) (11)
j=1

 
0
where Λ∗(τ ) = µ(τ ) , Λ(τ ) , f ∗t = 1, f 0t . The Wold representation of (11) can be written as:

∞
X ∞
X
Qτ (y t |Ft−1 ) = B j(τ ) v t−j(τ ) + C j(τ ) f ∗t−j (12)
j=0 j=0

where B j(τ ) = Φ1(τ ) B j−1(τ ) +Φ2(τ ) B j−2(τ ) +. . . for j = 1, 2, . . . with B 0(τ ) = I m and B j(τ ) = 0
for ` < 0 and where C j(τ ) = B j(τ ) Λ∗(τ ) .
Following the precedent established by Cecchetti and Li (2008, p. 12) in the context of
multivariate forecasting with dynamic quantile regressions, we assume that the quantile index
is fixed throughout the forecast horizon.7 Under this assumption, (12) implies that the vector
of forecast errors associated with the prediction of y t+h conditional on the information at time
t − 1 and on the common factors is given by:

h
X
ut+h(τ ) = B `(τ ) v t+h−`(τ )
`=0

and the total forecast error variance matrix is:

h
X
B `(τ ) ΩB 0`(τ )


Cov ut+h(τ ) = (13)
`=0

Now, consider the covariance matrix of the forecast errors associated with predicting y t+h given
values of the shocks to the ith equation, vit(τ ) , vi,t+1(τ ) , . . . , vi,t+h(τ ) :

h
(i)
X


ut+h(τ ) = B `(τ ) v t+h−`(τ ) − E v t+h−`(τ ) |vi,t+h−`(τ ) (14)
`=0

Under the assumption that v t(τ ) ∼ i.i.d. (0, Ω) with Ω = diag (ω11 , ω22 , . . . , ωmm ), we have:

E v t+h−`(τ ) |vi,t+h−`(τ ) = ωii−1 Ωei vi,t+h−`(τ )

= ei vi,t+h−`(τ ) (15)
7
In principle, one could allow for the quantile index to vary across the forecast horizon, although it is not
clear how one could systematically determine the time-path of the quantile index from one horizon to the next.

12
where ei is an m × 1 selection vector with its ith element set to 1 and all other elements set to
zero and where ωii−1 Ωei = ei . Substituting this result into (14), we obtain:

h
(i)
X
B `(τ ) v t+h−`(τ ) − ωii−1 Ωei vi,t+h−`(τ )



ut+h(τ ) = (16)
`=0

and taking the unconditional expectation yields:

  h h
(i)
X X
Cov ut+h(τ ) = B `(τ ) ΩB 0`(τ ) − ωii−1 B `(τ ) Ωei e0i ΩB 0`(τ ) (17)
`=0 `=0

Therefore, the decline in the h-step-ahead forecast error variance of y t obtained as a result of
conditioning on future shocks to the ith equation is given by:

 
(i)
∆ih(τ ) = Cov ut+h(τ ) − ut+h(τ )
h
X
= ωii−1 B `(τ ) Ωei e0i ΩB 0`(τ ) (18)
`=0

Scaling the jth diagonal element of ∆ih(τ ) — that is, e0j ∆ih(τ ) ej — by the h-step-ahead forecast
error variance of the jth variable in y t , we obtain:
 
ωii−1
0 B
Ph 0 ΩB 0

`=0 je `(τ ) Ωe i e i `(τ ) ej
F EV D yjt ; uit(τ ) , h = Ph 0
0
`=0 ej B `(τ ) ΩB `(τ ) ej
 2
ωii−1 h`=0 e0j B `(τ ) Ωei
P
= Ph 0
(19)
0
`=0 ej B `(τ ) ΩB `(τ ) ej

for ` = 0, 1, . . . , h and i, j = 1, . . . , m, where ej selects the predicted variable and ei selects the


source innovation. Consequently, F EV D yjt ; uit(τ ) , h measures the proportion of the h-step-
ahead forecast error variance of the jth variable in y t accounted for by the ith idiosyncratic
innovation, vit(τ ) . Note that any variation in the FEVD over quantiles is due to the quantile-
dependence of the parameters of the Wold representation of the QFVAR model, the B `(τ ) ’s.
The quantile-variation in the parameters reflects changes in the dynamic relationships among
y t as shocks of different size affect the system. To facilitate the comparison of the FEVDs
across quantiles, rather than allowing the size of the shock to differ over quantiles, we scale the
ith shock to one standard deviation of the ith regression residual for all τ . By considering an
identical shock at every quantile, we are able to focus keenly on the quantile-variation in the

13
dependence structure captured by the parameters of the QFVAR model.

2.3 The Diebold–Yilmaz Approach to Network Analysis

Based on our definition of the quantile FEVD in (19), it is straightforward to generalise the
Diebold–Yilmaz framework for conditional mean network analysis to the conditional quantile
setting. The h-step-ahead m×m spillover matrix for y t evaluated at the τ th conditional quantile
may be written as follows:

 
(h) (h) (h)
 θ1←1,(τ ) θ1←2,(τ ) · · · θ1←m,(τ ) 
 (h) (h) (h)

 2←1,(τ ) θ2←2,(τ ) · · · θ2←m,(τ )
 θ 
(h)
= (20)

A(τ ) .. .. ..
.. 

 . . . . 

 
(h) (h) (h)
θm←1,(τ ) θm←2,(τ ) · · · θm←m,(τ )

(h)
(h)
where we define θj←i,(τ ) ≡ F EV D yjt ; uit(τ ) , h to simplify the notation and where θj←i,(τ )
measures the spillover of idiosyncratic shocks affecting variable i onto variable j. Note that
we need not apply the row sum normalisation used by Diebold and Yilmaz (2014) due to the
(h)
diagonality of the covariance matrix, Ω, which ensures that m
P
i=1 θj←i,(τ ) = 1, j = 1, 2, . . . , m

by construction.
(h)
Based on A(τ ) , we may define the following summary measures of the network topology at
the τ th conditional quantile:

(h) (h)
Oi←i,(τ ) = θi←i,(τ )
m
(h) (h)
X
Fi←•,(τ ) = θi←j,(τ )
j=1,j6=i
m
(h) (h)
X
T•←i,(τ ) = θj←i,(τ )
j=1,j6=i

(h) (h) (h)

Ni←i,(τ ) = T•←i,(τ ) − Fi←•,(τ ) (21)

The proportion of the h-step-ahead FEV of the i-th variable that can be attributed to shocks to
(h)
variable i itself is known as the own variance share, Oi←i,(τ ) . The from (or in) degree of variable
(h)
i, Fi←•,(τ ) , measures the total spillover from the system to variable i. As such, it measures the
dependence of variable i on external conditions. Likewise, the to (or out) degree of variable i,
(h)
T•←i,(τ ) , captures the total spillover from variable i to the system, which measures the influence

14
 (h)
of the ith node in the network. Ni←i,(τ ) is therefore a natural measure of the net directional
(h) (h)
connectedness of variable i. Note that Oi←i,(τ ) + Fi←•,(τ ) = 1, i = 1, 2, . . . , m by construction
(h)
but that T•←i,(τ ) can be greater than or less than one. Finally, the spillover index evaluated at
the τ th conditional quantile is given by:

m
(h) (h)
X
S(τ ) = m−1 Fi←•,(τ ) (22)
i=1

3 Credit Spread Data

Our model includes the following N = 18 countries: Austria† , Australia, Belgium† , China,
France† , Germany† , Greece† , Ireland† , Italy† , Japan, the Netherlands† , Norway∗ , Portugal† ,
Russia, Spain† , Sweden∗ , the U.K.∗ and the U.S.A. Eurozone members are marked with a dagger
symbol, while European countries which do not use the Euro are marked with an asterisk. Our
dataset is sampled at daily frequency over the period 03-Jan-2006 to 14-Feb-2012, with the
end date being determined by data availability. Specifically, the Greek sovereign CDS spread
exceeds 10,000bp on 15-Feb-2012, shortly before Greek sovereign CDS contracts started trading
points upfront due to the expectation of an imminent credit event. For each country, we include
two country-specific credit spreads, one for the sovereign and one for the financial sector. We
also include an array of common factors building on those identified by Longstaff et al. (2011).

3.1 Sovereign Credit Risk

We measure the change in sovereign credit risk using the first difference of the five-year sovereign
CDS spread, expressed in basis points. Following the CDS market conventions outlined by Bai
and Wei (2017), we work with US dollar denominated CDS in all cases except for the US,
where we employ Euro denominated CDS. In addition, we use CDS contracts with a complete
restructuring clause for every sovereign except Australia, where we use CDS contracts with a
modified restructuring clause as the data is more complete.

3.2 Financial Sector Credit Risk

We track changes in financial sector credit risk in the ith country using the first difference of
the synthetic sector-wide CDS spreads constructed by Greenwood-Nimmo et al. (2017). Taking
inspiration from the approach of ADS, Greenwood-Nimmo et al. (2017) construct a synthetic
credit spread for the ith country as an equally-weighted average of the CDS spreads for firms

15
which satisfy a variety of selection criteria. Among these criteria, firms must: (i) have USD
denominated five-year CDS spread data in the Markit database which conforms to the corporate
CDS market conventions documented by Bai and Wei (2017); (ii) be classified by Markit as
financials; (iii) be classified as either banking or insurance firms in Bureau van Dijk’s Osiris
database; (iv) be identified by Markit as operating in the ith country; and (v) hold assets of
USD10bn or more. The large majority of the firms included in the indices are publicly traded
although there are two notable exceptions: (i) in Austria, Raiffeisen Zentralbank is included in
the sample to ensure that the index is not based on data for a single firm; and (ii) in China,
data for four large state-sponsored banks is used as there is not enough CDS data for privately
held Chinese banks to construct a meaningful index. Similarly, rather than simply dropping
failed banks from the sample, Greenwood-Nimmo et al. (2017) include CDS data for several
institutions which became state-owned as a result of the crisis, such as the Irish Bank Resolution
Corporation.8

3.3 Observed Common Factors

As noted by Longstaff et al. (2011), a large number of global variables may exert a common
influence on credit spreads. The authors propose a parsimonious approach to the selection
of factors, focusing on market-determined variables on the grounds that they should impound
a wide array of information relevant to investors. We follow this precedent and include the
majority of the observable explanatory variables considered by Longstaff et al. that are reported
at daily frequency as well as several variables that they do not consider.9 Following Longstaff
et al. (2011), we start with a selection of US macroeconomic and financial indicators, which
proxy for global economic and financial conditions:10

(i) US stock market performance. To capture the key risk factors affecting US equity markets,
we include the Rm-Rf, SMB and HML factors developed by Fama and French (1993). Rm-
Rf, accounts for a market factor, while SMB and HML account for risk factors related to
firm size and book-to-market equity, respectively. The Fama-French factors are freely
8
For a detailed discussion of the construction and properties of the financial sector CDS spreads, the reader
is referred to Greenwood-Nimmo et al. (2017) and, in particular, to their Data Supplement.
9
Our use of daily data necessitates the exclusion of several of the explanatory variables used by Longstaff et
al. that are sampled at lower frequency, including bond and equity flows, for example.
10
The use of US data to approximate global factors is supported by the wealth of evidence that US macrofi-
nancial conditions exert a powerful and widespread influence on global economic and financial performance (e.g.
Dees, di Mauro, Pesaran and Smith, 2007; Chudik and Fratzscher, 2011; Helbling, Huidrom, Kose and Otrok,
2011; Longstaff et al., 2011; Pesaran and Chudik, 2013).

16
 available from Ken French via http://mba.tuck.dartmouth.edu/pages/faculty/ken.
french/data_library.html.

(ii) US Treasury market conditions. We include the change in the five-year constant maturity
Treasury (CMT) yield to capture expectations regarding macroeconomic conditions in the
US and, by extension, in the world economy. In addition, given that investors regard
US Treasury debt as a safe haven asset, Longstaff et al. note that the CMT yield may
incorporate a flight-to-liquidity component. The CMT yield is published by the Federal
Reserve in its H.15 Statistical Release.

Next, in light of the evidence that variations in funding liquidity and in the risk appetite
of investors played an important role in the propagation of the global financial crisis (GFC)
(e.g. Brunnermeier and Pedersen, 2009; Chudik and Fratzscher, 2011; Greenwood-Nimmo et
al., 2016a; Pelizzon, Subrahmanyam, Tomio and Uno, 2016), we include the following factors:

(iii) The TED spread. The TED spread is the difference between the 3-month USD LIBOR
and the 3-month US Treasury bill yield. Variations in the TED spread reflect changes
in counterparty risk and liquidity in the US interbank market. Consequently, it is widely
used as an indicator of funding liquidity. The TED spread is available from the Federal
Reserve Economic Data Service (FRED) via https://fred.stlouisfed.org/.

(iv) The Euribor-DeTBill spread. To capture European-specific variations in funding liquidity,

we include the spread between the 3-month Euribor and the 3-month German Treasury
bill yield. The Euribor and the German yield data are available from Datastream.

(v) The variance risk premium (VRP). Bollerslev, Tauchen and Zhou (2009) define the VRP as
the difference between the one-month-ahead implied variance and a forecast of the realised
variance over the same period. Under this definition, the VRP is typically positive, with
higher values indicating a reduced risk appetite. We forecast the realised variance using
the same augmented version of Corsi’s (2009) heterogeneous autoregressive model used by
h i
(22)
Bekaert and Hoerova (2014). We compute the VRP as V RPt = V IXt2 −E RVt+1 , where
(22)
V IXt2 denotes the de-annualised squared VIX and RVt denotes the realised variance
for the S&P 500 measured over the next 22 trading days as the sum of squared five-minute
intraday returns. The VIX data is available from FRED, while we obtain the daily realised
variance from the Oxford Man Institute’s Realized Library (Heber, Lunde, Shephard and

17
 Sheppard, 2009, ver. 0.2).11

(vi) The Treasury term premium. The term premium measures the excess yield required to
induce investors to hold a long-term bond as opposed to a sequence of shorter-term bonds.
Consequently, it conveys valuable information on investors’ time preferences as well as their
expectations regarding the macroeconomic outlook. We include the 5-year Treasury term
premium derived from the five-factor no-arbitrage term structure model of Adrian, Crump
and Moench (2013) which is freely available from the Federal Reserve Bank of New York
via https://www.newyorkfed.org/research/data_indicators/term_premia.html.

(vii) US investment grade and high yield spreads. To capture changes in the required rate of
return on investment grade (IG) and high yield (HY) corporate bonds, we include both
the IG and HY spreads. We define the IG spread as the spread between the Bank of
America Merrill Lynch US corporate BBB and AAA effective yields and the HY spread
as the difference between the Bank of America Merrill Lynch US corporate BB and BBB
effective yields. The corporate bond yield data is available from FRED.

Longstaff et al. (2011) use cross-sectional averages of the credit spreads to proxy for regional
and global factors in a manner reminiscent of the way that unobserved factors are approximated
in the common correlated effects framework of Pesaran (2006). However, given our focus on
observable factors, we elect to include an array of tradeable credit spread indices instead:

(viii) ITRAXX indices to capture pan-European credit risk factors. To account for European
credit risk factors not captured elsewhere in our factor structure, we include five 5-year
ITRAXX indices to isolate European credit risk factors. Specifically, we include the
ITRAXX Europe index, the ITRAXX High Volatility index, ITRAXX Crossover index,
and the ITRAXX Senior and Subordinated Financials indices. ITRAXX data is available
via Datastream.12

Lastly, we control for currency fluctuations relative to the US dollar:

(ix) Bilateral spot exchange rate returns. To capture exchange rate fluctuations, we include
the daily log-return on the bilateral spot exchange rate for each currency in our sample in
11
Longstaff et al. (2011) also include the equity premium approximated at monthly frequency by the price-
earnings ratio for the S&P 100 index. We are obliged to exclude the equity premium because earnings per share
is unavailable at daily frequency. However, the variance risk premium should be highly correlated with the equity
premium and should impound much of its informational content.
12
We also experimented with the inclusion of North American and emerging markets CDX indices among our
factors but we found that they did not add substantially to the information content of our other observed factors.

18
 units of foreign currency per USD. The exchange rate data is obtained from Datastream.
By including every exchange rate in each equation of the factor VAR model, we are able
to control for portfolio adjustments affecting multiple currencies simultaneously.

3.4 Properties of the Dataset

Table 1 provides elementary summary statistics for the dataset. Preliminary analysis of the
autocorrelation structure in the data indicates that each series is stationary with relatively
limited serial correlation (results are available on request). The countries in our sample form
two natural groups, one composed of the GIIPS (Greece, Ireland, Italy, Portugal and Spain) and
Russia, which display high and volatile credit spreads, and the other composed of the remaining
countries in our sample, which display considerably lower and less volatile credit spreads. For
any given country, the sovereign CDS spread is typically lower and less volatile than the financial
sector CDS spread. In general, sovereign credit spreads should act as a lower bound on financial
sector credit spreads, not least because of the implicit guarantee that the sovereign extends to
the financial sector. However, due to their deep sovereign crises, this is not the case in Greece,
Italy, Portugal or Spain.

— Insert Table 1 here —

The credit spreads for all countries display pronounced excess kurtosis. This is also a feature
of our common factors and is a natural reflection of the severity of the shocks affecting global
financial markets over our sample period. The evidence of heavy tails in the data suggests
that estimation by quantile regression may be preferable to the use of simple conditional mean
estimators not just because it can illuminate tail relationships but also because it is more robust
in the presence of extreme observations.

4 Estimation Results

Before we proceed, we must first determine appropriate values for both the QFVAR lag order
and the forecast horizon used in the construction of our connectedness measures. We follow
the precedent of Koenker and Xiao (2006) and set the lag order at every conditional quantile
equal to the optimal lag order which is selected at the conditional mean by minimisation of the
Schwarz Information Criterion. This results in a first-order model which we verify is dynamically
stable — the largest eigenvalue of the companion matrix is just 0.39. Unfortunately, there is

19
no similar rule to select an optimal forecast horizon for connectedness analysis. For this reason,
we compare the properties of the adjacency matrix evaluated at the conditional mean for three
different forecast horizons, h ∈ {3, 5, 10} trading days. We select relatively short horizons in
light of the fact that the FEVDs obtained from a first order VAR model estimated on data
with a low degree of persistence are likely to rapidly converge to their long-run values. This
observation is borne out by our finding that the network statistics are largely invariant to the
choice of horizon within this range, with the spillover indices obtained under h = 3, h = 5
and h = 10 being identical to the first decimal place (a detailed elementwise comparison of the
adjacency matrices is available on request). A similar degree of invariance with respect to the
forecast horizon has been documented by Greenwood-Nimmo et al. (2016a) in the context of
a stationary first order VAR model applied to the analysis of risk spillovers among the G10
currencies. We therefore adopt a forecast horizon of five trading days without loss of generality,
although we shall return to the issue of horizon selection in the rolling sample context in Section
4.4 below.

4.1 Performance of the Observed Factors

In practice, a finite number of observed common factors is unlikely to completely eliminate the
cross-section correlation among the regression residuals but it should render it sufficiently weak
that the off-diagonal elements of the residual covariance matrix can be set to zero without loss
of generality. Consequently, the adequacy of our factors is ultimately an empirical question.
To evaluate the performance of our factors, we estimate two models at the conditional mean
— a simple VAR(1) model with no factors and our factor VAR(1) model — and compare the
residual cross-correlations in each case. Figure 1(a) reveals considerable correlation among
the residuals of the simple VAR(1) model, with more than half of the pairwise correlation
coefficients exceeding 0.2 in absolute value and more than 20% exceeding 0.4. Figure 1(b)
shows that the factors remove a great deal of this correlation. In the factor VAR(1) model,
more than three-quarters of the pairwise correlations are weaker than 0.2 in absolute value
and 91% are weaker than 0.4. This strongly supports the validity of our factors and suggests
that it is a reasonable simplification to treat the error terms of the factor VAR(1) model as
cross-sectionally uncorrelated.

— Insert Figure 1 here —

Next, we examine how the introduction of common factors affects the network statistics

20
obtained at the conditional mean. To this end, Figure 2 plots the distribution of bilateral
spillover effects — the spillover density to adopt the terminology of Greenwood-Nimmo et al.
(2016b) — for the models with and without factors. In the factor VAR case, each row of the
adjacency matrix defined in (20) sums to unity by virtue of the diagonality of the covariance
matrix, Ω. To obtain comparable values for the simple VAR model where the covariance matrix
is non-diagonal, we apply the row-sum normalisation suggested by Diebold and Yilmaz (2014).
For convenience, we multiply each element of the adjacency matrix by 100 so that the estimated
spillover effects can be interpreted as percentages rather than proportions. In principle, there-
fore, the spillover density has support [0, 100] although, in practice, the limiting cases of 0 and
100 will only arise from restricted VAR models where the parameter and covariance matrices
are sparse. Greenwood-Nimmo et al. (2016b) have shown that the spillover density of a network
constructed by the Diebold–Yilmaz method resembles a power law.13 This is a natural finding
given that the elements of the adjacency matrix are constructed from variance decompositions,
which are defined as ratios of quadratic forms. Consequently, for ease of interpretation, we
follow Acemoglu et al. (2012) and report empirical counter cumulative distribution functions
(CCDFs) on a logarithmic scale in Figure 2.

— Insert Figure 2 here —

The CCDF for the model without factors lies considerably to the right of the CCDF for the
model with factors, indicating that the omission of common factors leads to stronger estimated
spillover effects. The difference between the two CCDFs is profound, with the spillover index
defined in (22) obtained from the model without factors taking a value of 72.50% compared to
just 11.49% for the model with factors. The mechanism driving this result is straightforward.
If a common component which contributes to the h-step-ahead FEV of the system is omitted
in estimation, then the share of the FEV that should rightly be attributed to that common
factor must be attributed to one or more of the endogenous variables included in the model.
The omission of relevant common factors will therefore upwardly bias the estimated bilateral
spillover effects. It is important to distinguish between common and idiosyncratic sources of
variation if one wishes to measure true bilateral spillovers which are free of common components.
For this reason, we henceforth focus exclusively on factor VAR models.
13
Power law behaviour is pervasive in economic and financial networks — see Acemoglu, Carvalho, Ozdaglar
and Tahbaz-Salehi (2012) and Gabaix (2016), for example.

21
4.2 Conditional Mean vs. Conditional Median

To establish a point of reference for the conditional quantile analysis that follows, we first
compare the network structure evaluated at the conditional mean by OLS and at the conditional
median by LAD over the full sample. If the results from network models evaluated at the
conditional mean are to be generalised across the conditional distribution — as is currently the
norm, at least implicitly — then the results obtained under OLS and LAD should be similar.
However, unlike OLS, LAD belongs to the class of robust M-estimators and is therefore less
susceptible to the influence of outliers. Given the extreme credit spread movements recorded
in several countries during our sample — notably among the GIIPS and Russia — it is likely
that the OLS and LAD estimates will differ. The comparison of OLS and LAD also provides an
initial glimpse of the value of estimation by quantile regression. The OLS estimator is equivalent
to an equally weighted average of the quantile regression estimators for τ ∈ (0, 1), while the
LAD estimator is simply the quantile regression estimator at τ = 0.5. It follows, therefore, that
if we observe differences between the network under OLS and LAD, then there must also be
differences between the network under LAD and at other conditional quantiles — that is, the
network will display quantile variation. The more pronounced the differences between OLS and
LAD, the greater the quantile variation is likely to be.
Figures 3(a) and 3(b) plot the spillover density in both the OLS and LAD cases. Several
features of the spillover densities under OLS and LAD are noteworthy. First, the right tails
of both densities are similar, which indicates that the strongest spillovers in the system are
of comparable magnitude at the conditional mean and median. The three strongest bilateral
spillovers under OLS are from the Spanish sovereign to the Spanish financial sector (8.55%), the
Chinese sovereign to the Russian sovereign (8.12%) and the Russian sovereign to the Russian
financial sector (5.03%). By contrast, under LAD, the three strongest pairwise spillovers all arise
from the Spanish sovereign and affect the Italian sovereign (5.81%), the Spanish financial sector
(5.07%) and the German sovereign (4.51%). Although these values are small compared to the
spillover effects reported in much of the existing literature on Diebold–Yilmaz networks, recall
that we employ a factor structure to isolate uncorrelated idiosyncratic shocks. In this context,
a spillover which accounts for 5% of FEV at the five-days-ahead horizon in a model with thirty-
six endogenous variables represents a strong bilateral linkage. The evidence of strong spillovers
originating from Spain and affecting the local financial sector and other European sovereigns
reflects the integration of financial markets within the EU.

22
 — Insert Figure 3 here —

The similarity in the right tail of the spillover density does not extend to other parts of the
density, however. The CCDF is less curved and displays considerably greater left tail mass under
LAD than under OLS, implying a higher proportion of weak spillovers at the median than at the
mean. The weakest bilateral spillover at the conditional mean takes a value of 3.65 × 10−4 %. At
the conditional median, there are fifty-seven bilateral spillovers that are weaker than this and
the weakest spillover is two orders of magnitude smaller at just 4.37 × 10−6 %. To appreciate
the difference in the shape of the two spillover densities more clearly, consider an arbitrary
threshold — of the 1,260 bilateral spillovers that we study, 551 are weaker than 0.1% at the
conditional mean but this value rises to 947 at the conditional median.
The granular differences documented above accumulate substantially under aggregation.
Table 2 reports the to (weighted out-degree), from (weighted in-degree) and net statistics for
each node in the system. In every case, the OLS estimate is larger than the LAD estimate —
often substantially so. If one computes the ratio of the OLS estimate to the LAD estimate for
each of the reported weighted in- and out-degrees, the minimum, mean and maximum values
are 1.23, 6.14 and 124.15, respectively. In addition to these scale differences in the estimated
spillover effects, the ranking of the most influential nodes in the system — measured by the
weighted out-degree — also differs across estimators. Under OLS, the three most influential
nodes are the Spanish (35.83%), Italian (25.85%) and Austrian (24.00%) sovereigns. Meanwhile,
at the conditional median, the most influential nodes are the Spanish sovereign (29.21%), the
French financial sector (14.15%) and the Dutch sovereign (9.59%).

— Insert Table 2 here —

To provide an impression of how the differences surveyed above affect the network as a whole,
Figure 4 provides a visual comparison of the network topology under OLS and LAD. Sovereigns
are represented by white nodes and financial sectors by shaded nodes, while the strength of
bilateral spillovers is indicated by the relative thickness of the edges. The layout of both graphs
is identical and is determined using the force-directed algorithm of Fruchterman and Reingold
(1991) applied at the conditional mean. The networks display some similar features, notably
the centrality of the Spanish sovereign and the disposition of many of the strongest bilateral
spillovers. However, the excess connectedness of the network evaluated by OLS relative to LAD
is easily seen and is clearly reflected in the spillover index, which takes a value of 11.49% at the

23
conditional mean compared to 4.34% at the conditional median.

— Insert Figure 4 here —

The results obtained under OLS suggest that, with a few notable exceptions, idiosyncratic
credit risk shocks do not propagate strongly. The results obtained under LAD indicate that
credit risk spillovers are even weaker than under the OLS case. However, the large difference
between the spillover intensity under OLS and LAD suggests that stronger spillovers prevail at
some non-central location or locations in the conditional distribution. It is to this issue that we
now turn.

4.3 Quantile Variation: Network Topology and the Shock Size

Our interpretation of the quantile regression estimates below will be predicated on the distinc-
tion between large adverse shocks and large beneficial shocks. To see this, note that in the right
tail of the conditional distribution, the observed changes in the vector of credit spreads are
large conditional on the values taken by the explanatory variables — that is, credit spreads are
increasing sharply due to the effect of large adverse shocks. By contrast, in the left tail, credit
spreads are falling sharply conditional on the explanatory variables due to the impact of large
beneficial shocks. In light of the evidence that financial market comovements increase under
adverse conditions (e.g. Ang and Bekaert, 2002), it is natural to think that strong spillovers
should be observed in the right tail of the conditional distribution, where the largest adverse
shocks affect the system. In practice, the pattern of quantile variation in the spillover index
shown in Figure 5 reveals that strong spillovers occur in both the left and right tails of the
conditional distribution, indicating that spillover intensity increases with shock size for both
adverse (right tail) and beneficial (left tail) shocks. This is consistent with the literature on
good and bad contagion, which emphasises the transmission of unexpected events in both the
left and right tail (e.g. Londono, 2016).

— Insert Figure 5 here —

Over the central 70% of the conditional distribution, the spillover index never exceeds 9%
and the influence of idiosyncratic credit risk shocks is largely localised. By comparison, when
large idiosyncratic shocks affect the system, bilateral spillovers play a profound role in shaping
the evolution of sovereign and financial sector credit risk. At the 1st, 5th, 10th, 90th, 95th and

24
99th percentiles, the spillover index takes values of 76.57%, 21.07%, 10.04%, 12.58%, 22.87%
and 83.32%, respectively. It is interesting to note that the pattern of quantile-variation in the
spillover index is roughly symmetric. Given that the residual covariance matrix is quantile-
invariant, this effect arises because of similarities in the dynamic parameters of the QFVAR
model at quantiles τ = α and τ = 1 − α. In practice, this near-symmetry arises by chance over
the full sample and is not a general feature of our results — we will shortly demonstrate that
it breaks down in rolling sample analysis.
The increased influence of bilateral credit risk spillovers in both tails can be seen clearly
in Figures 3(c)–(f). Note that how the spillover density moves rightward and becomes more
peaked as τ → 1 and τ → 0. It is also readily apparent in Figure 6, which shows network
plots for τ = {0.05, 0.95} drawn on the same scale and with the same layout as Figure 4. The
increase in connectedness in the tails is marked, implying that large idiosyncratic credit risk
shocks propagate considerably more forcefully than weaker shocks.

— Insert Figure 6 here —

These large tail effects are not simply due to a lack of effective observations in the tails of
the conditional distribution. For any given τ , quantile regression makes use of every data point
with non-zero weight and our sample of 1,596 trading days is relatively sizable compared to the
dimensionality of the QFVAR system. Rather, our results are consistent with the hypothesis
that the informational content of large shocks is greater than that of small shocks, a point which
is well-established in the volatility literature (e.g. Dendramis, Kapetanios and Tzavalis, 2015).
When a large idiosyncratic shock affects a given sovereign or financial sector, investors gain
significant news which may lead to a reappraisal of the level of risk associated with other nodes
in the system. Consequently, the quantile variation documented in Figure 5 can be interpreted
like a regime-switching process where the transition between regimes of strongly beneficial news
at one extreme and strongly adverse news at the other occurs smoothly as the shock size varies.
An obvious question to ask as this stage is whether the ranking of nodes by centrality is
preserved across quantiles. That is, does the group of nodes that display the strongest outward
spillovers vary with the shock size? To this end, in Figure 7, we plot the weighted out-degree
rank of each node in the system, with the nodes ranked in decreasing order of influence. The
weighted out-degree captures the total strength of all outward spillovers originating from a given
node. As such, it represents a natural measure of the influence of a node.

25
 — Insert Figure 7 here —

The figure is organised with the GIIPS on the top row, the other European countries on
the next two rows and the non-European countries on the final row. Several features of Figure
7 are noteworthy. First, the weighted out-degree rank displays marked quantile variation in
the majority of cases. Consider the GIIPS to begin with. With the exception of Greece,
the GIIPS sovereigns are typically toward the top of the weighted out-degree ranking across
all quantiles. This is particularly apparent for Spain and Italy, where the combination of a
high and rising debt servicing cost with a substantial stock of outstanding debt generated a
significant risk of contagion — the scale of the losses generated by a Spanish or Italian default
would have posed a grave threat to global financial stability. Greece behaves quite differently
than the other GIIPS sovereigns, demonstrating a relatively high weighted out-degree rank in
the middle of the conditional distribution but a much lower rank in both tails. This suggests
that the propagation of large idiosyncratic Greek sovereign risk shocks was relatively mild. At
first sight, this appears to stand at odds with historical experience of the European debt crisis,
where Greek contagion was widely discussed.14 However, several studies have since shown no
Greek contagion at this time (e.g. Mink and de Haan, 2013; Pragidis, Aiellia, Chionis and
Schizas, 2015). A more nuanced interpretation of our finding is that although direct bilateral
spillovers of Greek sovereign risk may be weak, the crisis in Greece may have affected global
markets indirectly via its impact on global factors such as investor risk appetite.
A general feature of many of the European sovereigns — GIIPS and non-GIIPS alike — is
that the weighted out-degree rank is typically higher in the tails of the conditional distribution
than in the central region. The same is true of the of the US. The increased influence of
these sovereigns in the tails is a reflection of their centrality in the global financial crisis and the
European debt crisis. By contrast, the other non-European sovereigns stand apart. In Australia,
the weighted out-degree rank is low and relatively stable across conditional quantiles, reflecting
the country’s muted experience of both the global financial crisis and the European debt crisis.
In China, Japan and Russia, the weighted out-degree rank is considerably lower in the tails
than at the median, indicating that these countries were predominantly the recipients of large
external shocks during both crises as opposed to the source of large influential shocks.
Another result which emerges from Figure 7 is that, for a given country, the weighted out-
degree rank of the sovereign typically exceeds that of the financial sector. This effect can be
14
A good example may be found in Nouriel Roubini’s column in the Wall Street Journal on May 6, 2010
(https://www.forbes.com/2010/05/05/greece-bailout-imf-opinions-columnists-nouriel-roubini.html).

26
understood in relation to the credit risk transfers implicit in financial sector bailouts. Acharya
et al. (2014) demonstrate that if a sovereign elects to bail out the domestic financial sector in
order to maintain the stability of the financial system, a significant amount of private sector risk
is transferred onto the sovereign. Consequently, dysfunction in the ith financial sector is likely
to be felt disproportionately by the ith sovereign, not least because the sovereign guarantee
acts to partially insulate both domestic and foreign investors from shocks originating in the ith
financial sector. By contrast, in an environment where financial institutions hold internationally
diversified portfolios of debt instruments, dysfunction in a given sovereign debt market may
rapidly propagate to the financial sector both domestically and internationally. This effect is
likely to be particularly pronounced among the Eurozone member states, where financial market
integration, common monetary policy and a shared currency create an environment where shocks
may propagate forcefully and where many of the instruments of stabilisation policy cannot be
manipulated on a country-by-country basis.
The relationship between the credit risk of the ith sovereign and the ith financial sector is
central to the analysis of Acharya et al. (2014). Specifically, the authors document an adverse
feedback loop between sovereign credit risk and financial sector credit risk which emerges after
a financial sector bailout. In their analysis, the transfer of credit risk from the financial sector
to the sovereign associated with a financial sector bailout coupled with the fiscal burden of the
bailout leads to an increase in the sovereign’s credit risk. This undermines the value of the
sovereign’s implicit guarantee of the financial sector moving forward and reduces the value of
the sovereign debt portfolios held in the financial sector. This causes financial sector credit
risk to rise which, in turn, further exacerbates sovereign risk because it raises the likelihood of
further sovereign intervention in the financial sector and so the cycle continues.
In our model, the feedback between the ith sovereign and the ith financial sector is captured
(5) (5) (5)
by the sum of the bilateral spillovers between the two nodes, Tsi ↔fi ,(τ ) = Tsi ←fi ,(τ ) + Tfi ←si ,(τ ) .
(5)
Figure 8 reports the quantile variation in Tsi ↔fi ,(τ ) for all 18 countries. For most countries, the
feedback effect is negligible throughout the centre of the conditional distribution but intensifies
markedly in the tails. The mechanism described by Acharya et al. (2014) focuses on the feedback
associated with adverse shocks, in the right tail of the conditional distribution. Our results
indicate that the same feedback loop acts upon the arrival of large beneficial shocks. A sovereign
bailout is a natural example of such a beneficial shock. Suppose that the ith sovereign receives
a bailout which reduces its credit risk. This increases the ability of the ith sovereign to stabilise

27
the ith financial sector while simultaneously lowering the default risk associated with sovereign
bonds held by the ith financial sector. This leads to a reduction in financial sector credit
risk which lowers probability that the sovereign will be required to intervene in the financial
sector, further lowering sovereign risk and so on. This suggests that the same feedback loop
that promotes instability in the analysis of Acharya et al. (2014) can act to restore stability if
policymakers are able to generate large beneficial shocks.

— Insert Figure 8 here —

4.4 Time-Varying Tail-Dependence

To this point, our analysis has focused exclusively on full-sample statistics. We have demon-
strated substantial quantile-variation in the topology of the credit risk network and have shown
that bilateral spillovers of idiosyncratic credit risk are an order of magnitude stronger in the
tails than they are at the conditional median. This implies that network models estimated
at the conditional mean are unlikely to adequately capture the extent of dependence observed
when large shocks occur. As noted by Betz et al. (2016), it is tail-dependence that should be of
the greatest interest for surveillance and regulatory purposes. Unlike the existing literature on
Diebold–Yilmaz networks, by estimating our QFVAR model on a rolling sample basis, we can
directly study time-variation not only in average connectedness but also in the extent of left-
and right-tail-dependence.
Before we proceed, we must first choose an appropriate window length for our rolling samples.
Existing studies in the Diebold–Yilmaz network literature which work with daily data have used
a variety of window lengths, including 100 days (e.g. Diebold and Yilmaz, 2014), 150 days (e.g.
Demirer et al., 2017), 200 days (e.g. Barunı́k et al., 2016) and 250 days (e.g. Greenwood-Nimmo
et al., 2016a, 2017). In the absence of a firm precedent, we follow Greenwood-Nimmo et al.
(2016a) and evaluate the sensitivity of our results to a set of three candidate window lengths,
w ∈ {200, 250, 300} trading days — we do not consider shorter windows to ensure that we do not
encounter small-sample issues in estimation. In addition, we take this opportunity to further
explore the robustness of our estimation results to the choice of forecast horizon, h ∈ {3, 5, 10}.
Figure 9 reports the spillover index under the nine possible combinations of window length
and forecast horizon at the conditional mean and median as well as the fifth and ninety-fifth
conditional quantiles. First, consider panels (a) and (b). At the conditional mean and median,
the choice of forecast horizon has little discernible effect on the spillover index, a result which

28
reinforces our findings over the full sample. The choice of window length affects the level of the
spillover index, with shorter windows yielding somewhat higher values. Nonetheless, it is the
dynamics of the spillover index which are of primary concern and they are largely unaffected
by the choice of window length. The correlation among the spillover indices obtained under the
nine different combinations of w and h is close to one in all cases. The same basic features are
also apparent at the fifth and ninety-fifth percentiles, although with greater noise. Critically,
however, the correlation across different combinations of window length and forecast horizon
remains substantial in the tails. We therefore conclude that the choice of window length does
not exert an undue influence on our results and we proceed with w = 250 and h = 5 trading
days without loss of generality. This leaves us with 1,347 rolling samples.

— Insert Figure 9 here —

With the forecast horizon and window length set, Figure 10(a) re-plots the spillover index
evaluated at the conditional mean and median — both of which are different measures of the
average connectedness of the system — on the same axes. To assist the reader, the dates of
several important events marked by vertical dashed lines. Figure 10(b) plots the spillover index
evaluated at the fifth conditional quantile as a measure of left-tail-dependence. This captures
the propensity for beneficial shocks that reduce credit risk to propagate through the system. All
else equal, stronger left-tail-dependence is stabilising. Figure 10(c) plots the the spillover index
at the ninety-fifth conditional quantile as a measure of right-tail-dependence, the propensity
for destabilising adverse shocks to spread. Lastly, Figure 10(d) plots the linear combination
(5) (5)
RT D = S0.95 − S0.05 . The time-variation in RT D clearly demonstrates that the near-symmetry
of Figure 5 in the full-sample setting is not preserved over rolling samples. Of greater interest,
however, is the natural interpretation of RT D as a measure of relative tail-dependence, with
positive (negative) values indicating stronger (weaker) dependence in the right tail than in the
left tail. We interpret increases (decreases) in RT D as evidence of rising (falling) financial
fragility. The correlations between the the four different spillover indices and our RTD measure
are reported in Table 3.

— Insert Figure 10 & Table 3 here —

The spillover indices evaluated at the conditional mean and median display broadly similar
behaviour. Both increase abruptly as a result of major adverse shocks such as the freezing of
redemptions in selected investment funds by BNP Paribas in August 2007 and the bankruptcy

29
of Lehman Brothers in September 2008. Similarly, beneficial shocks such as the announcement
of the Troubled Asset Relief Program (TARP) and the GIIPS sovereign bailouts cause both
indices to fall gradually. However, the spillover index evaluated at the conditional mean falls
much more markedly than its counterpart evaluated at the conditional median. As a result,
Table 3 reveals that the two are only moderately correlated (0.57).
A natural question to ask at this stage is whether the spillover indices evaluated at the
conditional mean and median share common dynamics with the spillover indices evaluated
in the tails of the conditional distribution. The correlations in Table 3 provide a striking
answer. Both are positively correlated with our measure of right-tail-dependence (0.79 and
0.55, respectively) but are essentially uncorrelated with our measure of left-tail-dependence
(-0.11 and 0.07, respectively). Furthermore, our measures of left- and right-tail-dependence
are mutually negatively correlated (-0.30). It is the last result which is most interesting as
it suggests that changes in right-tail-dependence coincide with oppositely-signed changes in
left-tail-dependence. For example, a period of growing fragility associated with an increased
propensity for adverse shocks to propagate is also likely to be a period where the propagation
of beneficial shocks becomes weaker and vice-versa. This result may arise from the aggregate
behaviour of market participants if the information revealed by a major event in either tail
causes a non-trivial proportion of market participants to focus disproportionately on further
events occurring in that tail while paying less attention to events in the other tail.
The negative association between left- and right-tail-dependence is easily seen in Figures
10(b)-(d). Consider the revelation of major losses at the Bear Stearns High-Grade Structured
Credit Fund and the Bear Stearns High-Grade Structured Credit Enhanced Leveraged Fund in
July 2007 as an example of a large adverse shock which led market participants to re-evaluate the
level of risk associated with mortgage-backed securities. At this time, we observe a prolonged
downward drift in left-tail-dependence, indicating a reduced propensity for spillovers of good
news. At the same time, there is a sharp and sustained increase in right-tail-dependence,
indicating a substantial increase in the sensitivity of market participants to adverse shocks.
This combination naturally generates a marked increase in RT D.
Now, consider the announcement of TARP as an example of a major beneficial shock. TARP
represented a major government intervention into the financial markets, providing funds for the
purchase of toxic assets and equity from troubled financial institutions. The introduction of
TARP is associated with a mild reduction in right-tail-dependence but with a strong and sus-

30
tained increase in left-tail-dependence. Recall that we have controlled for a raft of factors
including liquidity conditions in the US so this result is not simply a manifestation of the im-
provement in the economic outlook and in funding liquidity in particular brought about by
TARP. Rather, it suggests that major policy interventions can generate a pronounced intensi-
fication of stabilising beneficial spillovers.
As a final exercise, to rule out the possibility that the behaviour of the relative tail-
dependence documented above is simply an artefact of our choice to work with the fifth and
ninety-fifth conditional quantiles, we plot two alternative measures of relative tail-dependence
in Figure 11. Specifically, alongside our benchmark 5% RT D, we plot the 10% and 1% RT Ds
(5) (5) (5) (5)
defined as RT D10% = S0.90 − S0.10 and RT D1% = S0.99 − S0.01 , respectively. The dynamics of
the three RT D measures are remarkably similar, indicating that our results are robust to the
precise definition of the left- and right-tail-dependence measures.

— Insert Figure 11 here —

5 Concluding Remarks

In this paper, we develop a new technique for the econometric analysis of financial networks
where the topology of the network is allowed to vary with the size of the shocks that affect the
system. Our approach is based on a novel framework for the estimation of vector autoregressions
by quantile regression. We employ a factor structure to isolate the idiosyncratic component of
the error process from the systematic component. Not only does this align our model with the
long literature on systematic and idiosyncratic risk but it also simplifies the estimation problem
as it renders the system of equations cross-sectionally independent. As a result, we are able
to estimate the model on an equation-by-equation basis using the standard quantile regression
toolboxes built into many statistical software packages. Our approach is therefore considerably
easier to implement than the existing frameworks for the estimation of quantile VAR models
associated with Cecchetti and Li (2008) and Schüler (2014).
We apply our technique to study the transmission of credit risk among a panel of eighteen
sovereigns and their respective financial sectors between January 2006 and February 2012. We
document marked quantile variation in the topology of the network. We show that idiosyncratic
credit risk shocks do not propagate strongly at the conditional mean or median but that powerful
spillovers occur in both tails of the conditional distribution. In addition, by studying the

31
bidirectional feedback between each sovereign and its domestic financial sector, we find that the
adverse feedback loop between sovereign credit risk and financial sector credit risk documented
by Acharya et al. (2014) manifests as a marked intensification of feedback in the right tail,
where large adverse shocks occur. Interestingly, however, we note a similar intensification
in the left tail, which indicates that the same feedback loop can act in a stabilising manner
in the presence of large beneficial shocks, such as sovereign bailouts. Finally, we use rolling
sample analysis to document time-variation in the degree of tail-dependence. This reveals an
interesting phenomenon — the level of left-tail-dependence is negatively correlated with the level
of right-tail-dependence. Specifically, our results indicate that major adverse (beneficial) events
are associated with a reduction (intensification) in stabilising left-tail spillovers coupled with
a simultaneous intensification (reduction) in destabilising right-tail spillovers. Furthermore,
although the dynamic evolution of spillover activity in the right tail is broadly similar to that
observed at the conditional mean and median, this is not true of spillover activity in the left
tail. Consequently, the evolution of relative tail-dependence is obscured when network models
are estimated at the conditional mean by OLS, as is common in the literature. We therefore
conclude that our framework for the analysis of quantile connectedness represents a valuable
addition to the existing literature on empirical network modelling.

32
References
Acemoglu, Daron, Asuman Ozdaglar, and Alireza Tahbaz-Salehi, “Systemic Risk and
Stability in Financial Networks,” American Economic Review, 2015, 105, 564–608.

, Vasco M. Carvalho, Asuman Ozdaglar, and Alireza Tahbaz-Salehi, “The Network

Origins of Aggregate Fluctuations,” Econometrica, 2012, 80, 1977–2016.

Acharya, Viral V., Itamar Drechsler, and Philipp Schnabl, “A Pyrrhic Victory? Bank
Bailouts and Sovereign Credit Risk,” Journal of Finance, 2014, 69, 2689–2739.

Adrian, Tobias and Markus K. Brunnermeier, “CoVaR,” American Economic Review,

2016, 106, 1705–1741.

, Richard K. Crump, and Emanuel Moench, “Pricing the Term Structure with Linear
Regressions,” Journal of Financial Economics, 2013, 110, 110–138.

Alter, Adrian and Andreas Beyer, “The Dynamics of Spillover Effects during the European
Sovereign Debt Turmoil,” Journal of Banking and Finance, 2014, 42, 134–153.

Ang, Andrew and Francis A. Longstaff, “Systemic Sovereign Credit Risk: Lessons from
the U.S. and Europe,” Journal of Monetary Economics, 2013, 60, 493–510.

and Geert Bekaert, “International Asset Allocation with Regime Shifts,” Review of Fi-
nancial Studies, 2002, 15, 1137–1187.

Bai, Jennie and Shang-Jin Wei, “Property Rights Gaps and CDS Spreads: When is there a
Strong Transfer Risk from the Sovereigns to the Corporates?,” Quarterly Journal of Finance,
2017, In press.

Bai, Jushan, “Inferential Theory for Factor Models of Large Dimensions,” Econometrica, 2003,
71, 135–171.

, “Panel Data Models with Interactive Fixed Effects,” Econometrica, 2009, 77, 1229–1279.

Barunı́k, Jozef, Evžen Kočenda, and Lukáš Vácha, “Asymmetric Connectedness on the
U.S. Stock Market: Bad and Good Volatility Spillovers,” Journal of Financial Markets, 2016,
27, 55–78.

Bekaert, Geert and Marie Hoerova, “The VIX, the Variance Premium and Stock Market
Volatility,” Journal of Econometrics, 2014, 183, 181–192.

Betz, Frank, Nikolaus Hautsch, Tuomas A. Peltonen, and Melanie Schienle, “Sys-
temic Risk Spillovers in the European Banking and Sovereign Network,” Journal of Financial
Stability, 2016, 25, 206–224.

Billio, Monica, Mila Getmansky, Andrew Lo, and Loriana Pelizzon, “Econometric
Measures of Connectedness and Systemic Risk in the Finance and Insurance Sectors,” Journal
of Financial Economics, 2012, 104, 535–559.

Blanchard, Olivier J. and Danny Quah, “The Dynamic Effects of Aggregate Demand and
Supply Disturbances,” American Economic Review, 1989, 79, 655–673.

and Mark W. Watson, “Are all Business Cycles Alike?,” in R.J. Gordon, ed., The Amer-
ican Business Cycle, Chicago: University of Chicago Press, 1986, pp. 123–160.

33
Blanco, Roberto, Simon Brennan, and Ian W. Marsh, “An Empirical Analysis of the
Dynamic Relationship between Investment-Grade Bonds and Credit Default Swaps,” Journal
of Finance, 2005, 60, 2255–2281.

Bollerslev, Tim, George Tauchen, and Hao Zhou, “Expected Stock Returns and Variance
Risk Premia,” Review of Financial Studies, 2009, 22, 4463–4492.

Bostanci, Görkem and Kamil Yilmaz, “How Connected is the Global Sovereign Credit
Risk Network?,” Working Paper 1515, Koç University, Istanbul August 2015.

Brunnermeier, Markus K. and Lasse H. Pedersen, “Market Liquidity and Funding Liq-
uidity,” Review of Financial Studies, 2009, 22, 2201–2238.

Caporin, Massimiliano, Loriana Pelizzon, Francesco Ravazzolo, and Roberto

Rigobon, “Measuring Sovereign Contagion in Europe,” Working Paper 18741, NBER, Cam-
bridge (MA) January 2013.

Cecchetti, Stephen G. and Hong Li, “Measuring the Impact of Asset Price Booms Using
Quantile Vector Autoregressions,” Mimeo: Brandeis University February 2008.

Chavleishvili, Sulkhan and Simone Manganelli, “Quantile Impulse Response Functions,”

Mimeo: Goethe University Frankfurt January 2016.

Chudik, Alexander and Marcel Fratzscher, “Identifying the Global Transmission of the
2007–2009 Financial Crisis in a GVAR Model,” European Economic Review, 2011, 55, 325–
339.

Corsi, Fulvio, “A Simple Approximate Long Memory Model of Realized Volatility,” Journal
of Financial Econometrics, 2009, 7, 174–196.

Dees, Stephane, Filippo di Mauro, M. Hashem Pesaran, and L. Vanessa Smith,

“Exploring The International Linkages Of The Euro Area: A Global VAR Analysis,” Journal
of Applied Econometrics, 2007, 22 (1), 1–38.

Demirer, Mert, Francis X. Diebold, Laura Liu, and Kamil Yilmaz, “Estimating Global
Bank Network Connectedness,” Journal of Applied Econometrics, 2017, in press.

Dendramis, Yiannis, George Kapetanios, and Elias Tzavalis, “Shifts in Volatility Driven
by Large Stock Market Shocks,” Journal of Economic Dynamics and Control, 2015, 55, 130–
147.

Diebold, Francis X. and Kamil Yilmaz, “Measuring Financial Asset Return and Volatility
Spillovers, with Application to Global Equity Markets,” The Economic Journal, 2009, 119,
158–171.

and , “On the Network Topology of Variance Decompositions: Measuring the Connected-
ness of Financial Firms,” Journal of Econometrics, 2014, 182, 119–134.

Fama, Eugene F. and Kenneth R. French, “Common Risk Factors in the Returns on
Stocks and Bonds,” Journal of Financial Economics, 1993, 33, 3–56.

Feldhütter, Peter and Mads Stenbo Nielsen, “Systematic and Idiosyncratic Default Risk
in Synthetic Credit Markets,” Journal of Financial Econometrics, 2012, 10, 292–324.

Fender, Ingo, Bernd Hayo, and Matthias Neuenkirch, “Daily Pricing of Emerging Mar-
ket Sovereign CDS Before and During the Global Financial Crisis,” Journal of Banking and
Finance, 2012, 36, 2786–2794.

34
Fruchterman, Thomas M.J. and Edward M. Reingold, “Graph Drawing by Force-
Directed Placement,” Software – Practice and Experience, 1991, 21, 1129–1164.

Gabaix, Xavier, “Power Laws in Economics: An Introduction,” Journal of Economic Per-

spectives, 2016, 30, 185–206.

Garratt, Anthony, Kevin Lee, M. Hashem Pesaran, and Yongcheol Shin, Global and
National Macroeconometric Modelling: A Long-Run Structural Approach, Oxford: Oxford
University Press, 2006.

Greenwood-Nimmo, Matthew J., Jingong Huang, and Viet H. Nguyen, “Financial

Sector Bailouts, Sovereign Bailouts and the Transfer of Credit Risk,” Unpublished Manuscript
March 2017.

, Viet H. Nguyen, and Barry Rafferty, “Risk and Return Spillovers among the G10
Currencies,” Journal of Financial Markets, 2016, 31, 43–62.

, , and Yongcheol Shin, “What’s Mine is Yours: Sovereign Risk Transmission during the
European Debt Crisis,” Mimeo, University of Melbourne May 2016.

Gyntelberg, Jacob, Peter Hördahl, Kristyna Ters, and Jörg Urban, “Intraday Dy-
namics of Euro Area Sovereign CDS and Bonds,” Working Paper 423, Bank for International
Settlements, Basel September 2013.

Heber, Gerd, Asger Lunde, Neil Shephard, and Kevin Sheppard, “Oxford-Man Insti-
tute’s Realized Library,” Dataset, Oxford-Man Institute, University of Oxford 2009.

Helbling, Thomas, Raju Huidrom, M. Ayhan Kose, and Christopher Otrok, “Do
Credit Shocks Matter? A Global Perspective,” European Economic Review, 2011, 55, 340–
353.

International Monetary Fund, “Old Risks, New Challenges,” Global Financial Stability
Report, April 2013, IMF, Washington DC April 2013.

Koenker, Roger, Quantile Regression, Cambridge: Cambridge University Press, 2005.

and Gilbert Bassett, “Regression Quantiles,” Econometrica, 1978, 46, 33–50.

and Kevin F. Hallock, “Quantile Regression,” Journal of Economic Perspectives, 2001,

15, 143–156.

and Zhijie Xiao, “Quantile Autoregression,” Journal of the American Statistical Associa-
tion, 2006, 101, 980–90.

Linnemann, Ludger and Roland Winkler, “Estimating Nonlinear Effects of Fiscal Policy
using Quantile Regression Methods,” Oxford Economic Papers, 2016, 68, 1120–1145.

Londono, Juan M., “Bad Bad Contagion,” International Finance Discussion Papers 1178,
Board of Governors of the Federal Reserve System, Washington D.C. September 2016.

Longstaff, Francis A., Jun Pan, Lasse H. Pedersen, and Kenneth J. Singleton, “How
Sovereign Is Sovereign Credit Risk?,” American Economic Journal: Macroeconomics, 2011,
3, 75–103.

Mink, Mark and Jakob de Haan, “Contagion during the Greek Sovereign Debt Crisis,”
Journal of International Money and Finance, 2013, 34, 102–113.

35
Pan, Jun and Kenneth J. Singleton, “Default and Recovery Implicit in the Term Structure
of Sovereign CDS Spreads,” Journal of Finance, 2008, 63, 2345–2384.

Pelizzon, Loriana, Marti G. Subrahmanyam, Davide Tomio, and Jun Uno, “Sovereign
Credit Risk, Liquidity, and European Central Bank Intervention: Deus Ex Machina?,” Jour-
nal of Financial Economics, 2016, 122, 86–115.

Pesaran, M. Hashem, “Estimation and Inference in Large Heterogeneous Panels with a

Multifactor Error Structure,” Econometrica, 2006, 74, 967–1012.

and Alexander Chudik, “Econometric Analysis of High Dimensional VARs Featuring a

Dominant Unit,” Econometric Reviews, 2013, 32, 592–649.

Pragidis, Ioannis C., G.P. Aiellia, Dionysios Chionis, and Panagiotis Schizas, “Con-
tagion Effects during Financial Crisis: Evidence from the Greek Sovereign Bonds Market,”
Journal of Financial Stability, 2015, 18, 127–138.

Schüler, Yves S., “Asymmetric Effects of Uncertainty over the Business Cycle: A Quantile
Structural Vector Autoregressive Approach,” Working Paper 2014-02, University of Konstanz
January 2014.

Sims, Christopher A., “Are Forecasting Models Usable for Policy Analysis?,” Minneapolis
Federal Reserve Bank Quarterly Review, 1986, Winter, 2–16.

Zellner, Arnold and Tomohiro Ando, “A Direct Monte Carlo Approach for Bayesian Anal-
ysis of the Seemingly Unrelated Regression Model,” Journal of Econometrics, 2010, 159,
33–45.

Zhu, Huiming, Xianfang Su, Yawei Guo, and Yinghua Ren, “The Asymmetric Effects
of Oil Price Shocks on the Chinese Stock Market: Evidence from a Quantile Impulse Response
Perspective,” Sustainability, 2016, 8, 1–19.

36
 Sovereign CDS Financial CDS Observed Factors
Mean S.D. Skew Kurt Mean S.D. Skew Kurt Mean S.D. Skew Kurt
European Countries Factors
Austria 0.105 4.102 0.877 19.095 0.177 6.694 1.190 21.021 Rm-Rf 2.475 152.971 -0.089 7.223
Belgium 0.141 5.285 -0.676 20.709 0.148 7.481 9.342 287.651 SMB 1.182 65.265 0.034 4.097
France 0.115 3.202 -0.293 17.066 0.141 5.200 -0.204 13.751 HML -0.484 81.816 0.688 8.164
Germany 0.051 1.751 0.211 12.120 0.118 10.167 -1.806 68.381 US 5y CMT Yield -0.222 7.006 -0.113 3.169
Greece 5.720 196.789 -0.735 58.131 1.428 32.204 -0.385 16.559 TED Spread -0.011 7.759 0.547 41.879
Ireland 0.357 12.760 -0.493 28.233 0.565 49.298 -1.137 56.276 Euribor-DeTBill 0.056 7.547 0.471 26.162
Italy 0.245 8.303 0.430 19.354 0.207 6.767 0.169 12.675 Equity VRP 0.008 19.349 -0.225 35.710
Netherlands 0.059 2.114 1.178 23.699 0.137 9.848 1.544 84.734 Term Premium 0.010 4.715 -0.240 3.204
Norway 0.020 1.403 0.496 27.955 0.077 3.187 -0.328 46.948 IG Spread 0.085 6.922 0.444 299.608
Portugal 0.733 16.821 -0.248 30.262 0.664 15.979 0.404 31.613 HY Spread -0.005 7.642 0.231 8.454
Spain 0.236 8.115 -0.182 13.801 0.271 8.085 -0.665 18.275 ITRAXX Europe 0.061 4.674 0.027 8.000
Sweden 0.035 2.094 0.365 15.740 0.087 3.131 0.626 22.296 ITRAXX High Vol. 0.080 8.318 0.283 13.336
UK 0.047 2.345 -0.258 10.737 0.106 5.547 -0.253 26.297 ITRAXX Crossover 0.194 18.008 -0.060 9.420
ITRAXX Sen. Fin. 0.130 6.643 -0.162 10.796

37
ITRAXX Sub. Fin. 0.217 11.388 0.094 12.391
Non-European Countries AUD/USD -2.358 102.044 0.777 9.658
Australia 0.040 2.575 0.502 17.005 0.118 5.886 -4.337 111.245 CNY/USD -1.552 10.257 0.071 6.969
China 0.065 4.809 0.103 42.408 0.126 16.692 -0.846 199.949 EUR/USD -0.651 69.107 -0.044 1.841
Japan 0.077 2.557 2.135 38.132 0.084 4.190 1.402 18.931 GBP/USD 0.590 65.966 0.472 2.986
Russia 0.093 16.530 2.225 57.542 0.241 31.436 -0.442 64.587 JPY/USD -2.552 71.902 0.131 6.189
US 0.025 1.523 0.645 11.995 0.214 24.257 -0.380 148.344 NOK/USD -1.016 91.735 0.113 2.667
RUR/USD 0.287 61.365 1.149 13.403
SEK/USD -1.086 94.605 0.028 2.449

Notes: Descriptive statistics are reported over a period of 1,596 trading days from 03-Jan-2006 to 14-Feb-2012. With the exception of the Fama-French factors, the equity VRP and the
exchange rate data, every series is expressed as a first difference and is measured in basis points. Descriptive statistics for the Fama-French factors are reported in basis points for the
levels of the series. The equity VRP is expressed as a first difference and is measured in squared percentage units. The foreign exchange rates are bilateral spot rates expressed in foreign
currency units per US dollar. For each exchange rate, we compute the daily log-return in basis points.

Table 1: Descriptive Statistics

 Sovereign Credit Risk Financial Sector Credit Risk
To From Net To From Net
ols lad ols lad ols lad ols lad ols lad ols lad
Australia 6.90 1.29 19.52 9.06 -12.62 -7.77 8.14 3.25 16.99 5.18 -8.85 -1.92
Austria 24.00 5.21 9.83 6.00 14.18 -0.80 10.47 3.29 13.21 3.32 -2.73 -0.03
Belgium 11.79 3.10 14.52 6.19 -2.72 -3.08 7.80 2.78 4.13 0.77 3.67 2.02
China 19.06 4.77 13.68 7.89 5.38 -3.12 10.30 3.18 11.46 0.45 -1.16 2.73
France 17.04 5.71 9.03 6.85 8.02 -1.14 18.33 14.15 12.89 7.66 5.45 6.49
Germany 11.91 3.48 13.85 10.87 -1.93 -7.39 4.93 1.14 7.14 0.85 -2.21 0.29
Greece 3.82 3.05 7.30 0.06 -3.47 2.99 5.59 1.24 8.20 2.39 -2.61 -1.15
Ireland 10.11 7.16 6.46 3.03 3.65 4.13 2.11 1.10 7.94 0.42 -5.83 0.68
Italy 25.85 7.07 11.54 8.70 14.32 -1.63 18.17 4.32 10.41 5.38 7.76 -1.06

38
Japan 7.18 2.29 13.85 6.65 -6.67 -4.36 4.77 1.20 11.59 5.23 -6.83 -4.02
Netherlands 20.39 9.59 7.37 5.63 13.02 3.96 3.50 0.48 7.62 0.65 -4.12 -0.17
Norway 7.67 1.32 11.00 2.45 -3.33 -1.13 5.87 0.50 5.77 1.06 0.11 -0.56
Portugal 11.99 9.33 10.27 1.85 1.72 7.48 9.55 3.90 11.71 5.81 -2.16 -1.92
Russia 20.42 7.49 19.68 3.09 0.74 4.40 9.64 1.08 17.18 1.48 -7.54 -0.40
Spain 35.83 29.21 8.90 3.04 26.93 26.17 4.50 0.72 18.39 8.23 -13.89 -7.51
Sweden 10.83 3.17 12.53 3.97 -1.70 -0.80 9.15 1.39 10.52 3.74 -1.37 -2.35
UK 14.47 3.39 12.63 7.20 1.84 -3.81 7.66 2.60 14.07 3.68 -6.41 -1.08
US 10.23 2.40 14.36 7.07 -4.12 -4.67 3.64 0.95 8.14 0.42 -4.50 0.53
Notes: Values are reported in percent.

Table 2: Topology of the Credit Risk Network at the Conditional Mean and Median
 Mean Med. 5% 95% RTD
Mean 1.00 0.56 -0.11 0.79 0.58
Med. 0.56 1.00 0.07 0.55 0.32
5% -0.11 0.07 1.00 -0.30 -0.78
95% 0.79 0.55 -0.30 1.00 0.83
RTD 0.58 0.32 -0.78 0.83 1.00

Table 3: Correlation among Spillover Indices Evaluated at Selected Quantiles

(a) Simple VAR(1) Model (b) Factor VAR(1) Model

Notes: The histograms show the distribution of the absolute pairwise correlations between the residuals of the
simple VAR(1) model and our factor VAR(1) model evaluated at the conditional mean by OLS.

Figure 1: Comparison of Absolute Residual Correlations, with and without Factors

Notes: The spillover density for the model with factors is shown as a dashed gray line, while the black circles
show the spillover density for the model without factors.

Figure 2: Spillover Density at the Conditional Mean, with and without Factors

39
 (a) OLS (b) τ = 0.50

(c) τ = 0.05 (d) τ = 0.95

(e) τ = 0.01 (f) τ = 0.99

Notes: The figure reports the empirical counter cumulative distribution function (CCDF) of the m (m − 1) off-
diagonal elements of the adjacency matrix on a logarithmic scale. The CCDF under OLS is shown as a dashed line
in every panel for comparison.

Figure 3: Quantile Variation in the Shape of the Spillover Density

40
41
(a) Conditional Mean (Ordinary Least Squares Estimator) (b) Conditional Median (Least Absolute Deviations Estimator)

Notes: Countries are identified by their respective alpha-2 ISO codes, with white nodes corresponding to sovereigns and shaded nodes to national financial sectors. Edges are drawn as
curves and the direction of pairwise spillovers is anticlockwise (i.e. the strong spillover between the Spanish sovereign and the Spanish financial sector goes from the sovereign to the financial
sector). The thickness of an edge is proportional to its weight. The layout of the nodes is determined by applying the force-directed algorithm of Fruchterman and Reingold (1991) to the
network evaluated at the conditional mean.

Figure 4: Network Visualisations for the Model Evaluated at the Conditional Mean and at the Conditional Median
Notes: The figure reports the value of the spillover index defined in (22) evaluated at the τ th conditional
quantile (plotted as a circle) relative to the value at the conditional mean (shown by the dashed line).

Figure 5: Variation in the DY Spillover Index over the Conditional Distribution

42
43
(a) τ = 0.05 (b) τ = 0.95

Notes: Countries are identified by their respective alpha-2 ISO codes, with white nodes corresponding to sovereigns and shaded nodes to national financial sectors. Edges are drawn as
curves and the direction of pairwise spillovers is anticlockwise. The thickness of an edge is proportional to its weight. The layout of the nodes is determined by applying the force-directed
algorithm of Fruchterman and Reingold (1991) to the network evaluated at the conditional mean.

Figure 6: Network Visualisations for the Model Evaluated by Quantile Regression at τ = 0.05 and τ = 0.95
44
Notes: The horizontal axis shows the quantile index multiplied by 100 while the vertical axis records the weighted out-degree rank from 1 (strongest) to 36 (weakest).

Figure 7: Weighted Out-Degree Rank of the Financial Sector (Fine Line) and the Sovereign (Heavy Line), by Country
45
(5) (5) (5)
Notes: The horizontal axis shows the quantile index multiplied by 100 while the vertical axis records the strength of the sovereign–financial feedback effect, Ts ↔f ,(τ ) = Ts ←f ,(τ ) +Tf ←s ,(τ ) ,
i i i i i i
measured in percent.

Figure 8: Total Feedback between the Financial Sector and the Sovereign by Country
 (a) Conditional Mean

(b) Conditional Median

(c) Fifth Conditional Quantile

(d) Ninety-Fifth Conditional Quantile

Notes: In each panel, results for our baseline setting with w = 250 and h = 5 trading days are shown as a
heavy black line. Results for each other combination of w ∈ {200, 250, 300} and h ∈ {3, 5, 10} trading days
are shown as fine gray lines. Letters a–i in the common-sample correlation heatmaps refer to the following
combinations: (a) w = 200, h = 3; (b) w = 200, h = 5; (c) w = 200, h = 10; (d) w = 250, h = 3; (e) w = 250,
h = 5; (f) w = 250, h = 10; (g) w = 300, h = 3; (h) w = 300, h = 5; and (i) w = 300, h = 10.

Figure 9: Robustness to the Choice of Rolling Window and Forecast Horizon

46
 AB D E F H I J K L N O P

C G M

(a) Conditional Mean and Median

AB D E F H I J K L N O P

C G M

(b) Fifth Conditional Quantile

AB D E F H I J K L N O P

C G M

(c) Ninety-Fifth Conditional Quantile

AB D E F H I J K L N O P

C G M

(d) Relative Tail-Dependence

Notes: The figure reports the value of the spillover index defined in (22) evaluated at the mean (shown as a fine line
in panel (a)) and at the 5th, 50th and 95th conditional quantiles (shown as heavy lines in panels (a)–(c)) as well as the
difference between the spillover index at the 95th and 5th conditional quantiles (shown as a heavy line in panel (d)). The
results are obtained from rolling regressions with a window length of 250 trading days. The date shown corresponds to the
last day of each rolling window. We suppress four rolling samples where the system exhibits instability at the conditional
mean. The following events are marked: A: S&P and Moody’s downgrade bonds backed by subprime loans (01-Jun-07);
B: Bear Stearns announces hedge fund losses (17-Jul-07); C: BNP Paribas halts redemptions on three investment funds
(09-Aug-07) D: UK Treasury announces liquidity support for Northern Rock (14-Sep-07); E: Bear Stearns is acquired by
JP Morgan (24-Mar-08); F: Lehman Brothers files for bankruptcy (15-Sep-08); G: the Troubled Asset Relief Program is
announced (14-Oct-08); H: the Fed begins purchasing mortgage-based securities issues by Fannie Mae and Freddie Mac
(05-Jan-09); I: signing of the American Recovery and Reinvestment Act (17-Feb-09); J: Greek parliament is dissolved
(08-Sep-09); K: report on the falsification of Greek data released (12-Jan-10); L: Greece requests aid (23-Apr-10); M: the
European Financial Stability Facility is announced (09-May-10); N: Ireland requests aid (22-Nov-10); O: Portugal requests
aid (06-Apr-11); and P: second Greek bailout (22-Jul-11).

Figure 10: Time-Varying Dependence at Selected Quantiles

 AB D E F H I J K L N O P

C G M

(5) (5)
(a) 10% Relative Tail-Dependence, S0.90 − S0.10

AB D E F H I J K L N O P

C G M

(5) (5)
(b) 5% Relative Tail-Dependence, S0.95 − S0.05

AB D E F H I J K L N O P

C G M

(5) (5)
(c) 1% Relative Tail-Dependence, S0.99 − S0.01

Notes: The results are obtained from rolling regressions with a window length of 250 trading days. The date shown
corresponds to the last day of each rolling window. We suppress four rolling samples where the system exhibits instability
at the conditional mean. The following events are marked: A: S&P and Moody’s downgrade bonds backed by subprime
loans (01-Jun-07); B: Bear Stearns announces hedge fund losses (17-Jul-07); C: BNP Paribas halts redemptions on three
investment funds (09-Aug-07) D: UK Treasury announces liquidity support for Northern Rock (14-Sep-07); E: Bear Stearns
is acquired by JP Morgan (24-Mar-08); F: Lehman Brothers files for bankruptcy (15-Sep-08); G: the Troubled Asset Relief
Program is announced (14-Oct-08); H: the Fed begins purchasing mortgage-based securities issues by Fannie Mae and
Freddie Mac (05-Jan-09); I: signing of the American Recovery and Reinvestment Act (17-Feb-09); J: Greek parliament is
dissolved (08-Sep-09); K: report on the falsification of Greek data released (12-Jan-10); L: Greece requests aid (23-Apr-10);
M: the European Financial Stability Facility is announced (09-May-10); N: Ireland requests aid (22-Nov-10); O: Portugal
requests aid (06-Apr-11); and P: second Greek bailout (22-Jul-11).

Figure 11: Alternative Measures of Relative Tail-Dependence