Dynamic Effects of Network Exposure On Equity Markets: Mardi Dungey Moses Kangogo Vladimir Volkov

Eurasian Economic Review (2022) 12:569–629
https://doi.org/10.1007/s40822-022-00210-y
ORIGINAL PAPER
Dynamic effects of network exposure on equity markets
Mardi Dungey1 · Moses Kangogo1 · Vladimir Volkov1
Received: 20 September 2021 / Revised: 16 January 2022 / Accepted: 1 February 2022 /

Published online: 28 February 2022
© Crown 2022
Abstract
We investigate the contribution of network exposure to both shock transmission and
absorption. Our data sample comprises 45 economies for the period 1998–2018 to
which we apply spatial econometric estimation technique. Our empirical findings
show that both network intensity and interconnectedness in the financial system
have impact on increasing network exposure. We also demonstrate how to estimate
network intensity in the financial system. Our results indicate that an increased net-
work intensity parameter is associated to period when the financial system is under
stress. The results show high exposure of the financial system to vulnerabilities. The
results suggest the changing market conditions increase the exposures to the finan-
cial system. Thus, effective ways to monitor the financial system should be imple-
mented by the policy makers to reduce the chances of financial instabilities.
Keywords Financial markets · Financial networks · Financial stability
JEL Classification G15 · G10 · G01 · C21
1 Introduction
The occurrence of the 2007–2009 global financial crisis still raises concerns to pol-
icy makers, regulators and academic researchers. The focus has been to find ways
and mechanisms to develop measures to predict distress in financial institution so as
Mardi Dungey passed away just after we completed the draft of this paper.
The authors received no financial support for the research, authorship, and/or publication of this
article. The authors are grateful for comments from participants at INFINITI Conference on
International Finance 2018 held at University of Sydney, particularly Samuel Vigne.
* Moses Kangogo
moses.kangogo@utas.edu.au
1
Tasmanian School of Business and Economics, University of Tasmania, Private Bag 84, Hobart,
TAS 7001, Australia
13
Vol.:(0123456789)
570 Eurasian Economic Review (2022) 12:569–629
to limit further destabilization of the global economy. Until recently, there has been
growing research focusing on how to predict systemic risks to minimize the recur-
rence of financial crises, while the importance of understanding how network expo-
sure contributes to the spread of financial distress in the financial system has been
largely underestimated. Motivated with increasing uncertainty of the stability of the
financial markets, this paper aims at investigating the effect of network exposure to
common factors.
With the advances in both technology and globalization in recent years, there has
been an increase in size, complexity, interconnectedness and concentration of the
financial system. These factors combined make the financial system more vulner-
able to a collapse. To ensure stability in the entire financial system, it is important to
study how the global financial system is interconnected. In a nutshell, interconnect-
edness (the key source of systemic risk) helps to assess systemic risk and financial
stability (Billio et al., 2010; Cai et al., 2014). To enable effective monitoring condi-
tions, there is need to quantify and measure these financial linkages.
This paper focuses on identifying the effect of interconnectedness in the finan-
cial system. This is achieved through examining the cross-border financial linkages
using real data. The paper aims to identify the impact of interconnectedness in either
absorbing or spreading risks in the financial system. It focuses on the cross-border
exposure, aiming at identifying the role played by interconnectedness in propagat-
ing shocks across countries, in the period of 1999–2018. The sample period covers
different crisis periods which include the global financial crisis of 2007–2009, the
European sovereign debt crisis of 2010 and the Chinese stock market crash of 2015.
Different approaches have been used to measure financial networks, systemic
risks and interconnectedness. Recent studies on measuring systemic risk show that
it was not only the size of the institution that led to the distress in the financial sys-
tem but also the interconnectedness between institutions (Billio et al., 2012; Dungey
et al., 2012).1 Contagion rapidly spread the crisis through financial linkages, lead-
ing to severe disruption of financial stability in both the United States and glob-
ally. The recent financial crisis emphasized the importance of using interconnected-
ness (defined as a set of relationships and interactions among the financial markets
participants) as a proxy in measuring systemic risk. Hautsch et al. (2014) note that
both size and interconnectedness of financial institutions determine its systemic
relevance.
Although currently there is ongoing research focusing on the interconnectedness
of institutions (Diebold & Yilmaz, 2009, 2012, 2013; Glasserman and Young, 2015;
Tonzer, 2015), there is still need to understand whether an increase of financial link-
ages affect the stability (either weakening or strengthening) of the financial system
and the impact of financial interconnectedness to both common factors and conta-
gion. It is also important to assess the resilient nature of financial networks to shocks
(both exogenous and endogenous).
Different studies have extended the variance decomposition model proposed
by Diebold and Yilmaz (2012) to measure connectedness in across different mar-
kets. For example, Giudici et al. (2021) provide a methodology to build a minimum
1
Interconnectedness address the Too Interconnected To Fail (TITF) paradigm.
13
Eurasian Economic Review (2022) 12:569–629 571
variance currency basket aimed at assessing contagion spillovers among foreign

exchange markets using the variance decomposition model. Giudici and Pagnottoni
(2019) extent variance decomposition model to examine the relationships of five
major Bitcoin exchange platforms. Giudici and Pagnottoni (2020) also extent the
variance decomposition model to a generalized vector error correction framework
to investigate return connectedness across eight of the major exchanges of Bitcoin.
Other emerging research have proposed different approaches to predicting sys-
temic risks across different markets. For example, Resta et al. (2020) apply techni-
cal analysis on bitcoin market and find trading on daily data is more profitable than
going intraday. Peralta and Zareei (2016) use Markowitz framework to construct a
network-based investment strategy. Spelta et al. (2021) develop a novel methodology
to detect the emergence of crisis and provide early-warning market signals to policy
makers. Spelta (2017) introduces a tensor decomposition technique to empirically
extract complex relationships from prices’ time series. Spelta et al. (2022) propose
a dynamical systems theory for non-linear time series forecasting and investment
strategy development which correctly make predictions at long time horizons.
Our paper differs from these approaches in measuring dynamic exposures in
the financial system. To estimate the network exposure in the financial system, this
paper uses spatial economic approach proposed by Anselin (1988). This is innova-
tive way of estimating the intensity of the exposures to vulnerabilities. The network
intensity parameter measures the network exposure in the financial system. We
use data on equity because of its ability to accurately reflect market conditions and
sentiments.
Our work focusing on the dynamic effects of the network exposure on the finan-
cial market has the following contributions. First our paper contributes to the existing
literature on interconnectedness by using the spatial econometric approach. Previous
research has mainly focused on the advance economies however, little attention has
focused on emerging economies. This paper aims at identifying network exposure
on both advanced and emerging economies. To the best of our knowledge, this is a
major piece of work, that utilizes spatial econometric techniques to estimate the net-
work exposure in the financial system.
The key findings of this paper highlight the role of network exposure in increas-
ing vulnerability, with both interconnectedness and network intensity playing key
roles in monitoring these exposures. We find high network intensity coefficient to
be associated with extreme events. This suggests that high network intensity param-
eter relates to crisis period. We also find that both interconnectedness and network
intensity increase exposures in the financial system. Cautions must be taken by pol-
icy makers in regulating and monitoring financial system to avoid re-occurrences of
crises.
The remainder of this paper is organized as follows. Section 2 reviews related
literature and develops key hypotheses which form the basis for empirical testing. It
also introduces the spatial econometric concept. Section 3 discusses the mechanism
underlying network exposure to common factors. Section 4 presents the results and
the effect of network exposures. Section 5 discusses various methods of estimat-
ing the network intensity parameter. Section 6 presents empirical evidence of the
13
network intensity parameter. Section 7 outlines the implications of our results, and
Section 8 concludes.
2 Literature review
Recent studies have shown that a major contributor to the transmission of shocks
during crisis such as the GFC (2007–2009) was not only the institution size, but
also the institution interconnectedness. Different interpretations of interconnected-
ness have resulted in the development of various measurement methods. All these
measures aim to assess the role and impact of interconnectedness during financial
distress.
Recent findings indicate that interconnectedness acts as a channel through which
shocks and losses spread to other financial institutions (Glasserman & Young, 2015,
2016). Other findings show that interconnectedness acts as a ‘double edge’ by being
able to absorb shocks up to certain point and also transmitting them to the financial
system after a given threshold is reached (Acemoglu et al., 2015; Cohen-Cole et al.,
2012; Gai & Kapadia, 2010; Tonzer, 2015). Gai and Kapadia (2010) suggested that
an increase in connectivity may lower the chance of contagion, but conditional on
a default by a given node, an increase in interconnectedness may trigger defaults to
other nodes, making the financial system more sensitive to defaults.
With increasing growth in the cross-border financial activities, interconnect-
edness poses threats to the financial system via increased vulnerability to shocks
spreading globally. Minoiu and Sharma (2014) supported the fact that a high degree
of interconnectedness triggered the breakdown of financial system during the
2007–2009 GFC. This implies that the more interconnected an institution, the higher
the likelihood of risk amplification to the entire system threatening the stability of
the economy.
Markose et al. (2012) referred to institutions that are ‘too interconnected’ as
‘super-spreaders’ of shocks in the financial system. This means that interconnected-
ness of financial institutions tends to spread shocks extensively across links, caus-
ing instability in the financial system. Gai and Kapadia (2010) showed that degree
of interconnectedness has an impact on contagion by acting as a channel through
which contagion spreads and shocks amplify. Greater interconnectedness aids in
lowering the likelihood of contagion but increases shocks transmission when the
financial system experiences difficulties.2 Using spatial modeling, Tonzer (2015)
assessed whether cross-border linkages have any impact on the stability of inter-
connected institutions. Though interconnectedness is beneficial in stable conditions,
Tonzer (2015) showed that interconnection of financial institution to foreign entities
provided a channel for propagating shocks when the system experienced difficulties.
Assessing interconnectedness in financial institutions could serve as an early
warning indicator for distress in financial systems. Econometric measures based on
Granger causality and principal components analysis proposed by Billio et al. (2012)
2
See Acemoglu et al. (2015), Gai and Kapadia (2010) and Glasserman and Young (2015) for more
details.
13
measure interconnectedness. These measures show that an increase in links in a

financial institution before a crisis signal an early warning. In addition, Minoiu et al.
(2015) focused on determining whether interconnectedness in financial institution
is a possible source of systemic risk that could serve as an early warning of crisis.
Their findings suggested that interconnectedness has early warning indicator proper-
ties for a crisis. Diebold and Yilmaz (2009, 2012) proposed new measures of inter-
connectedness by measuring risk and management in financial institutions based on
variance decomposition. Their results showed that global financial interconnected-
ness is time-varying, implying that network exposure within financial institutions
varies over time.
Minoiu et al. (2015) showed that an increase in linkages within a country and
a decrease in cross-border linkages are associated with a high chance of financial
crisis. This is consistent with the results of Peltonen et al. (2019) which indicated
that more interbank linkages increase the chance of banking crises. The increase
in cross-border transactions led to more interactions and relationships between dif-
ferent markets, leading to the formation of international ‘robust-yet-fragile’ finan-
cial networks. A financial network is ‘robust-yet-fragile’ when it serves as a shock
absorber (promoting financial stability) up to a certain point, beyond which it ampli-
fies shocks (leading to financial instability) in the whole financial system. While
financial markets benefit through the formation of more robust and stronger inter-
connections beyond a certain point, they also create potential channels of shock
transmission.
Several studies have explored roles that financial networks play in good and
bad periods. The first strand of literature relates interconnectedness to risk diver-
sification, cross-border investment opportunities and availability of different finan-
cial products in the market. A financial network is robust when it absorbs shocks,
enhancing the stability and health of a financial system (Allen & Gale, 2000). Hav-
ing more interconnections implies more risk-sharing and diversification, so shocks
hitting the network will be shared among the various interconnected institutions
building a resilient financial system (Glasserman & Young, 2016). This mechanism
is supported by Vitali et al. (2016), who found that an increase in interconnections
makes the financial system more resilient and increases shock diversification. Other
studies show that formation of these interlinkages helps absorb shocks to a certain
point before contributing to their spread (Acemoglu et al., 2015; Cohen-Cole et al.,
2012; Gai & Kapadia, 2010; Tonzer, 2015). Kubelec and Sá (2012) argued that
financial interconnectedness increases due to countries becoming more open, there-
fore causing the entire network to collapse. While Gai and Kapadia (2010) deter-
mined that stronger connectivity in the financial sector would improve absorption
of shocks, they also suggested that conditional on the default of an institution, an
increase in network connectivity propagates shocks from one institution to others.
The second strand of literature focuses on how interconnectedness enhances the
channels through which shocks spread and intensify to the broader financial system
(Battiston & Caldarelli, 2013; Glasserman & Young, 2015, 2016). That is, financial
networks tend to be fragile when they amplify shocks rather than contain them. This
may destabilize the entire financial system by increasing systemic risk, leading to
financial instability. Many studies show the extent to which interconnectedness could
13
have a negative effect on financial stability. For instance, Battiston and Caldarelli
(2013) demonstrated that although individual institutions benefit from increased
interlinkages, this could be a channel through which contagion and distress spread
to the entire financial system. Battiston et al. (2012) found that an increase in finan-
cial interconnections increased credit exposure which increases systemic risk. Amini
et al. (2016) argued that institutions with more interconnections contribute more to
financial instability. Further, Acemoglu et al. (2015) stated that more interconnec-
tions can make the financial system more fragile due to increased shock propaga-
tion when shocks are either large or coincidental. These findings are supported by
Markose et al. (2012), who referred to institutions that are ‘too interconnected’ as
‘super-spreaders’ of shocks. Minoiu and Reyes (2013) analyzed the global banking
network using 184 countries and reported that connectivity in the banking network
tends to increase particularly when the market is under stress. This aligns with the
findings of Glasserman and Young (2015), which asserted that interconnectedness
among different markets were key contributors to the GFC of 2007–2009. Yellen
(2013) regarded interconnectedness as a financial stability concern after the occur-
rence of the global crisis while Sun and Chan-Lau (2017) argued that interconnect-
edness was the source of systemic risk.
The third strand of literature shows that specific institutions or markets play a key
role in spreading shocks in the network. For instance, by investigating the patterns
of international trade and financial integration, Schiavo et al. (2010) demonstrated a
cause of the global crisis was shocks spreading from advanced economies to other
markets, leading to network-wide distress. Kubelec and Sá (2012) found the US and
UK to be the key players in the global financial network, with high interconnections
compared to the rest of the world.
2.1 Spatial econometric concept
Introduction of spatial econometric techniques into financial application can be use-

ful in modeling spillovers. The spatial econometric technique has been used recently
in finance. For example, Eder and Keiler (2015) used it to model contagion risk
among financial institutions; Fernandez (2011) employed it to measure risk pre-
mium propagation among firms; Asgharian et al. (2013) used it to investigate stock
market co-movements while Catania and Billé (2017) applied it to advancements in
score-driven models typically used in time series econometrics.
Spatial dependence parameter (network intensity parameter) captures the strength
of the spatial dependent units; thus, it is a key component in investigating the struc-
ture of the spatial autoregressive process. Network intensity parameter falls within
the range of 0 and 1, where 0 (1) is the minimum (maximum) estimate. A robust-
yet-fragile financial system is associated with high (greater than 0 and approaching
1) network intensity. Financial systems benefit from high network intensity through
risk-sharing and diversification. Conversely, increasing intensity beyond certain lim-
its will increase the rate at which shocks propagate, leading to financial instability
(Eder & Keiler, 2015). As a consequence, the financial system benefits from relative
high network intensity (especially when there is no shock hitting the system), which
13
allows for effective absorption of shocks rather than their amplification them to the
entire system (Affinito & Pozzolo, 2017). Restricting network intensity parameters
may improve stability leading to a more robust financial system (Gofman, 2017).
In addition, since estimation of network intensity depends on the connection
matrix, the structure of the connection matrix will have an impact on the estimation
of network intensity. Let d ∈ [d, d] be the degree of network connectivity, where d
and d are the minimum and maximum degree of connectivity respectively; then, a
robust-yet-fragile network is associated with degree of connectivity close to d . The
network is robust in the sense that the risk-sharing and diversification are higher
in the absence of large shocks and fragile when large shocks hit the network. High
connectivity with shocks hitting the financial system imply more shocks being
propagated in the network causing a fragile financial system (Glasserman & Young,
2015).
Consider a full or complete network with equivalent row-normalized weighted
connection matrix, an example represented as:
⎛ 0 0.25 0.25 0.25 0.25⎞

⎜0.25 0 0.25 0.25 0.25⎟
W = ⎜0.00 0.50 0 0.00 0.50⎟
⎜ ⎟
⎜0.25 0.25 0.25 0 0.25⎟
⎝0.25 0.25 0.25 0.25 0 ⎠
A shock hitting the above network will be proportionally shared among the nodes
in the network depending on the weights of the edges. The shock will either be
equally shared or proportionally shared across the entire system depending on the
size of the shock. Thus, a robust network exists when shocks are equally distrib-
uted to all institutions since the network is more resilient to small shocks (Hüser,
2015). Conversely, when weights are beyond a certain threshold, the risk-sharing
effect is endangered by larger shocks being amplified in the financial system rather
than being contained. This implies that shocks will affect nodes that are strongly
interconnected.
Generally, financial institutions benefit from high connectivity in the absence of
shocks, while high connectivity can lead to financial instability when large shocks
hit the system. This argument is supported by various studies. For example, Hal-
dane (2009) asserted that high connectivity in the financial system leads to greater
risk-sharing and diversification; above certain connectivity thresholds, it will propa-
gate shocks to the entire system. In addition, Vitali et al. (2016) argued that high
connectivity beyond certain threshold leads not only to large systemic events, but
also to more frequent occurrence of distress events. Acemoglu et al. (2015) also
found that highly connected institutions are more resilient to small shocks that pose
a high chance of contagion in the presence of large shocks. This is also supported
by Silva et al. (2016), who identified a high potential for a default to be triggered,
especially in a dense interconnected network since risk-sharing effects vanish when
large shocks hit this network. Schiavo et al. (2010) suggested that the structure of the
connection determines how the financial system responds to shocks.
13
Estimates of network intensity are dependent on the interaction of the endoge-

nous spatial lag (Wy); thus, increasing interactions (a denser weighting matrix) leads
to greater amplification of shocks rather than sharing shocks across these networks.
This is supported by various studies. For example, using European CDS spread data,
Blasques et al. (2016) showed that high time-varying spatial coefficients are associ-
ated with credit riskiness, which leads to fragility and potential collapse of financial
system. Battiston et al. (2012) found that financial systems can be more resilient
when the financial accelerator is low; when it is at a maximum, adverse effects are
inflicted via spreading shocks. This suggests that when shocks hit the financial sys-
tem, a high network intensity estimate signifies a higher probability that the financial
system will be fragile (Vitali et al., 2016).
2.2 Motivation and hypotheses development
Financial integration is a process through which either financial markets, countries

or regions become interconnected in different ways. This process includes cross-
border lending and borrowing and is an important phenomenon in financial mar-
kets. Increasing integration is often associated with a more complex financial sector.
Financial integration is beneficial to markets in terms of efficient capital allocation,
higher investment and growth opportunities and risk-sharing. Risk-sharing improves
the resilience of the global financial system (González-Páramo, 2010).3
Financial integration also serves to spread shocks to the entire financial system.
According to Schiavo et al. (2010), it is through integration that advanced econo-
mies become more interconnected with other markets, thereby spreading shocks
to these markets. This leads to global distress and increased cross-border exposure
threatening financial system stability. Hüser (2015) showed that an increase in the
integration of the interbank network poses an increased risk of contagion, and as a
consequence increases systemic risk. Asgharian et al. (2013) argued that cause of
the Asian crisis was trade integration (measured by cross-border flows of imports
and exports) between Asian countries, especially those emanating from Thailand
and spreading rapidly to its neighbors (Indonesia, Malaysia and even Korea). In
this context, we consider financial integration a contributory factor to increased net-
work intensity. As countries engage in cross-border activities, financial integration
expands potentially making financial markets more volatile.
Based on these motivation and related literature, we outline four hypotheses:
Hypothesis 1: Network intensity increases during periods of stress. This
hypothesis tests whether network intensity estimate changes over time. It will
determine whether it tends to increase or decrease when the market is under
stress. We would expect network intensity to increase when shocks hit the net-
work.
3
See González-Páramo (2010) for more details on the benefit of financial integration in the global finan-
cial system.
13
Hypothesis 2: Degree of connectivity affects estimation of network intensity. This

aims to investigate whether greater interconnectedness among different markets
influences estimation of network intensity parameter. This will provide an insight
into how the connection matrix makes the financial network robust-yet-fragile.
Hypothesis 3: Financial integration affects the estimation of network inten-
sity. This tests whether financial integration can explain why network intensity
increases or decreases during different periods. With increasing cross-border
activities, markets have become more integrated forming a possible channel
through which shocks can spread in financial systems.
Hypothesis 4: Advanced economies have a greater impact in spreading shocks.
This tests whether developed economies have a greater impact in the estimation
of network intensity. This test is in line with Schiavo et al. (2010), who found
that advanced economies were the key spreaders of shocks during the GFC.
3 Financial network and exposure to common factors
We now consider the impact of exposure to common factors in financial networks.

This involves examination of how the structural model (which incorporates both
systematic and idiosyncratic shocks) behaves in the presence of network exposure.
The starting point focus on the structural model, capturing exposures to common
factors as considered in Billio et al. (2015). According to Sharpe (1964) and Lintner
(1965), the traditional capital asset pricing model (CAPM) is given by:
rit − rft = 𝛼it + 𝛽it (rmt − rft ) + 𝜀it (1)
where rit is the return on stock i at time t, rmt is the market return at time t, rft is the
risk-free rate and 𝜀it is a vector capturing the idiosyncratic shocks of stock i at time t.
𝛼it and 𝛽it are the parameters of the model.
The traditional CAPM model can be extended to the Fama–French three-factor
linear model. Considering the pricing perspective, the Fama–French three-factor for
a set of risk asset returns, rit at time t is given by:
rit − rft = 𝛼it + 𝛽it (rmt − rft ) + hit HMLit + sit SMBit + 𝜀it (2)
where HML is the book-to-market factor of stock i at time t, SMBit is the size factor
of stock i at time t. hit and sit are additional parameters of the model; 𝜀it is a vector
capturing the idiosyncratic shocks of stock i at time t.
The main focus of this approach is on both network exposures (endogenous) and
exposures to common factors (structural exposure, which is exogenous). Equation
(2) can be rewritten in a structural form as:
S(rit − 𝔼 [rit ]) = 𝛽M ritM + 𝛽HML ritHML + 𝛽SMB ritSMB + 𝜂it (3)
rit − 𝔼 [rit ] = S−1 (𝛽M ritM + 𝛽HML ritHML + 𝛽SMB ritSMB + 𝜂it ) (4)
13
The spatial matrix S in Eq. (4) captures the contemporaneous relations associated
with interconnections between different assets, while 𝜂t is the structural idiosyn-
cratic risk at time t.
Our aim is to construct a structural model that contains contemporaneous rela-
tionships driven by links across assets, and systematic and idiosyncratic shocks.
Thus, the spatial matrix S can be parametrized as S = In − 𝜌W , where In is n × n
identity matrix, |𝜌| < 1 is the spatial dependence parameter (network intensity
parameter) indicating the strength of the network exposure. It monitors the network
impact while W represents relationships across assets.4
Equation (4) into a spatial autoregressive framework (SAR) as:
(In − 𝜌W)(rit − 𝔼 [rt ]) = 𝛽M ritM + 𝛽HML ritHML + 𝛽SMB ritSMB + 𝜂it (5)
rit − 𝔼 [rit ] = (In − 𝜌W)−1 (𝛽M ritM + 𝛽HML ritHML + 𝛽SMB ritSMB + 𝜂it ) (6)
If we let 𝛽M ritM + 𝛽HML ritHML + 𝛽SMB ritSMB = Z , using a geometric series expansion to
the first degree,5 the above model can be represented as:
∑
∞
∑
∞
rit − 𝔼 [rit ] = Z + 𝜂it + 𝜌j W j Z + 𝜌j W j 𝜂it
⏟⏟⏟ ⏟⏟ ⏟ j=1 j=1 (7)
i ii ⏟⏞⏞⏟⏞⏞⏟ ⏟⏞⏞⏞⏟⏞⏞⏞⏟
iii iv
where
i. structural exposure to common factors

ii. structural impact to idiosyncratic component
iii. network exposure to common factors
iv. network impact to idiosyncratic component
Equation (7) captures the impact of exposures (both structural and network expo-
sures) of both systematic and idiosyncratic shocks. Therefore, we conclude that both
idiosyncratic and systematic components are influenced by the presence of intercon-
nections across assets/institutions.
4 Dataset and effect of network exposure
The empirical analysis of this paper uses different datasets. We use daily return con-
structed from daily equity market indices obtained from Thompson Reuter’s Data-
stream. We also used the liability, market value and 90 days treasury bill (T-Bill)
4
See Anselin (1988) for more details on spatial econometrics. For simplicity, we refer to 𝜌 as the net-
work intensity parameter.
5
By geometric expansion, we have (In − 𝜌W)−1 = In + 𝜌W + 𝜌2 W 2 + ..., where 𝜌W represent the influ-
ence of neighbors on each unit while 𝜌2 W 2 second neighborhood influences each unit and so on.
13
rates data. Other datasets that include foreign exchange (FX), interest rate (IR),
S&P 500 volatility index—US (VIX), Euro STOXX 50 volatility index—Europe
(VSTOXX) and trade were also considered in the second part of the analysis.
We chose our sample of markets based on the availability of: (i) closing values,
(ii) closing hours, and (iii) changes in closing prices, listed by region in Table 1. Our
analysis of equity return spillovers is based on local currencies to avoid blurring
the extent of market co-movements with fluctuations in the foreign exchange market
(Mink, 2015).
The daily return (rt ) for all markets are calculated as the log differences of the
total daily equity market indices of a given economy at time t. This can be expressed
as:
rt = ln(Pt ∕Pt−1 ) × 100 (8)
where rt is the return at time t, Pt is the closing stock price of a given financial insti-
tution at time t, Pt−1 is the lagged price and ln is the natural logarithm.
We study 42 stock markets in three categories: developed, emerging, and frontier.
We extended the previous research that primarily focused on a few developed or
emerging markets (e.g. the G7 [Canada, France, Germany, Italy, Japan, the UK and
the US] stock markets investigated by Apostolakis and Papadopoulos (2014), the 10
developed and 11 emerging markets in Asia studied by Yarovaya et al. (2016), and
Asian markets examined by Narayan et al. (2014) and did not consider all possible
interconnectedness across different stock markets).
Table 2 presents the descriptive statistic of the daily returns for each market.
The mean returns are positive for all economies with standard deviation ranging
from 0.0096 to 0.0237. The kurtosis results suggest that the daily return would be
‘peaked’ and have ‘fat-tailed’ distribution. Unit root tests revealed the usual charac-
teristics of stationary returns in each series. The analysis was conducted using de-
meaned returns (as the mean is usually extremely close to 0 and, as we are focused
on variance decompositions, this assumption is innocuous). Analysis of the com-
plete network, consisting of 42 nodes, formed the initial benchmark for the study.
To construct our network, we used the data with its recorded local closing
time date. The choice of time-zone treatment can have dramatic effects; no sin-
gle choice is dominant due to the complications of wanting to test for two-way
causality. Other researchers have used the dates as provided with the data (Wang
et al., 2018), averaged data over consecutive days (Forbes & Rigobon, 2002) or
used time-matched data series (Kleimeier et al., 2008). Although the last of these
is arguably the most appropriate, it is difficult to obtain these data for the mar-
kets examined here and to control for problems associated with out-of-local trad-
ing time liquidity effects (most markets have different price-impact effects dur-
ing local and non-local trading). The averaging procedure used by Forbes and
Rigobon (2002) introduced a moving average bias into the problem, and, with
Granger-causality testing, created additional problems with the performance of
the statistic. Further, it is debated whether the use of lagged or non-lagged sam-
ples introduces or reduces noise in the process. Sensitivity analysis to different
choices of date-lagging produced important differences; the most pronounced of
13
Table 1 List of country-specific stock indices and their corresponding Thomson Reuters Datastream
codes
Country code Country name Stock index Datastream code
European
AT Austria ATX-AUSTRIAN TRADED INDEX ATXINDX
BE Belgium BEL 20 BGBEL20
CZ Czech Republic PRAGUE SE PX CZPXIDX
DK Denmark OMX COPENHAGEN (OMXC20) DKKFXIN
FI Finland OMX HELSINKI 25 (OMXH25) HEX25IN
FR France FRANCE CAC 40 FRCAC40
DE Germany DAX 30 PERFORMANCE DAXINDX
GR Greece ATHEX COMPOSITE GRAGENL
HU Hungary BUDAPEST (BUX) BUXINDX
IE Ireland IRELAND SE OVERALL (ISEQ) ISEQUIT
IT Italy FTSE MIB INDEX FTSEMIB
NL Netherlands AMSTERDAM MIDKAP AMSMKAP
PL Poland WARSAW GENERAL INDEX 20 POLWG20
PT Portugal PORTUGAL PSI-20 POPSI20
ES Spain IBEX 35 IBEX35I
SE Sweden OMX STOCKHOLM 30 (OMXS30) SWEDOMX
CH Switzerland SWISS MARKET (SMI) SWISSMI
TR Turkey BIST NATIONAL 100 TRKISTB
UK United Kingdom FTSE ALL SHARE FTALLSH
Americas
CA Canada S&P/TSX COMPOSITE TTOCOMP
US United States S&P 500 COMPOSITE S&PCOMP
AR Argentina ARGENTINA MERVAL ARGMERV
BR Brazil BRAZIL BOVESPA BRBOVES
CL Chile CHILE SANTIAGO SE GENERAL (IGPA) IGPAGEN
MX Mexico MEXICO IPC (BOLSA) MXIPC35
Asia
AU Australia S&P/ASX 200 ASX200I
CN China SHANGHAI SE COMPOSITE CHSCOMP
JP Japan NIKKEI 225 STOCK AVERAGE JAPDOWA
IN India S&P BSE NATIONAL 200 IBOM200
ID Indonesia IDX COMPOSITE JAKCOMP
HK Hong Kong HANG SENG HNGKNGI
MY Malaysia DJGL MALAYSIA DJTM MALAYSIA DJMALYL
NZ New Zealand S&P/NZX 50 NZ50CAP
PK Pakistan KARACHI SE 100 PKSE100
PH Philippines PHILIPPINE SE I(PSEi) PSECOMP
SG Singapore STRAITS TIMES INDEX SNGPORI
KR South Korea KOREA SE KOSPI 200 KOR200I
LK Sri Lanka COLOMBO SE ALL SHARE SRALLSH
TW Taiwan TAIWAN SE WEIGHED TAIEX TAIWGHT
13
Table 1 (continued)
Country code Country name Stock index Datastream code
TH Thailand BANGKOK S.E.T. BNGKSET

Africa
EG Egypt MSCI EGYPT MSEGYTL
ZA South Africa FTSE/JSE ALL SHARE JSEOVER
these is that when the US data are lagged, there is virtually no evidence of trans-
mission from the US to Asia, which seems at odds with our understanding of
international financial markets and the transmission of shocks. Consequently, this
chapter uses the convention of actual day dating in its analysis.
We first examine the evolution of the unweighted and weighted networks over
the sample period and augmented this analysis with scenarios based on alterna-
tive clustering of markets, as per the Asian Development Bank member countries
and the role of regional groupings, including the Association of Southeast Asian
Nations (ASEAN) with other regions across the globe.
The sample period considered is January 1999–December 2017 because our
focus is to observe the dynamics of the network exposure in the twenty-first cen-
tury. By taking advantage of the long horizon with a large number of observations
(4956), we subdivided the sample into four phases, as represented in Table 3:
Phase 1 is the pre-crisis (1 January 1999–14 September 2008) period, Phase 2 is
the GFC (15 September 2008–31 March 2010) period, Phase 3 is the European
debt crisis (EDC) (1 October 2010–21 November 2013) and phase three is the
most recent period (22 November 2014–29 December 2017). We followed Dun-
gey et al. (2015) and Dungey and Renault (2018) when choosing these dates.
We use the BIS database to obtain liabilities data to construct the weighting
matrix. The BIS bilateral locational banking statistics provided a comprehensive
cross-border data set of international banking transactions. This included aggre-
gate international cross-border claims and liabilities of a set of both reporting and
non-reporting countries. We use cross-border liabilities of reporting countries,
measured on a quarterly basis from 1999Q1–2017Q4 to construct the connection
matrix. The weighting matrix is obtained using the combined Granger causality and
DY approach (see Chowdhury et al., 2019). Each country is represented by direct
liability towards all the other countries in all financial sectors (central banks, banks,
non-bank financial institutions and non-financial sectors). We consider 42 (mature
and emerging markets) countries in our sample (see Table 4), for which the data
were complete and reliable. We also use different specifications of the connection
matrix in our empirical analysis. Particularly, we randomly generated sparse (fewer
interconnections) and denser (more interconnections) matrices for the markets in our
sample.6
6
Sparse connection matrix will be used to investigate the effect of network exposures with decreased
interconnectedness while denser connection matrix will be used to investigate the effect of network expo-
sure with increased interconnectedness.
13
Table 2 Descriptive statistics of daily return for each market

Country Mean Min Max Std. dev Kurtosis Skewness ADF test No. obs.
AT 0.0002 − 0.0974 0.1277 0.0135 10.7384 − 0.2140 − 71.1573** 5739

BE 0.0003 − 0.0829 0.1087 0.0111 9.9523 0.0112 − 70.3872** 5739
CZ 0.0002 − 0.1494 0.1316 0.0132 14.8753 − 0.1718 − 69.4570** 5739
DK 0.0004 − 0.1091 0.0986 0.0113 9.7303 − 0.2713 − 71.1096** 5739
FI 0.0004 − 0.0897 0.0973 0.0152 6.5627 − 0.1113 − 73.8211** 5739
FR 0.0003 − 0.0904 0.1118 0.0143 7.8396 0.0829 − 77.0792** 5739
DE 0.0004 − 0.0849 0.1140 0.0147 7.5954 − 0.0077 − 76.5194** 5739
GR 0.0000 − 0.1902 0.1327 0.0189 10.6342 − 0.1824 − 68.5693** 5739
HU 0.0007 − 0.1650 0.1459 0.0166 13.5187 − 0.2345 − 71.7532** 5739
IE 0.0003 − 0.1303 0.1022 0.0129 11.3909 − 0.5090 − 71.4974** 5739
IT 0.0001 − 0.1272 0.1161 0.0149 7.8602 − 0.0388 − 76.8145** 5739
NL 0.0003 − 0.0950 0.0830 0.0124 7.4651 − 0.4448 − 68.6160** 5739
PL 0.0003 − 0.1320 0.1469 0.0168 7.4476 0.0083 − 74.4086** 5739
PT 0.0001 − 0.0986 0.1073 0.0118 9.9042 − 0.2103 − 68.3531** 5739
ES 0.0003 − 0.1235 0.1443 0.0146 8.9329 0.0124 − 74.0971** 5739
SE 0.0004 − 0.0842 0.1165 0.0147 7.2704 0.1721 − 77.0466** 5739
CH 0.0003 − 0.0867 0.1139 0.0118 9.5391 − 0.0438 − 73.2515** 5739
TR 0.0013 − 0.1768 0.1856 0.0237 9.3668 0.3075 − 73.9692** 5739
GB 0.0002 − 0.0834 0.0921 0.0108 9.3767 − 0.0988 − 75.9863** 5739
AR 0.0006 − 0.1684 0.1775 0.0224 9.3782 0.1776 − 70.0926** 5739
BR 0.0006 − 0.1312 0.2796 0.0188 18.5465 0.7545 − 72.1713** 5739
CL 0.0002 − 0.0710 0.1475 0.0106 14.8593 0.4589 − 63.2493** 5739
MX 0.0006 − 0.1334 0.1292 0.0146 10.6563 0.2751 − 68.8651** 5739
CA 0.0003 − 0.0932 0.0982 0.0105 12.5107 − 0.5306 − 74.9890** 5739
US 0.0003 − 0.0903 0.1158 0.0117 11.6484 − 0.0660 − 80.9848** 5739
AU 0.0002 − 0.0834 0.0589 0.0096 8.5885 − 0.3593 − 77.1427** 5739
CN 0.0004 − 0.1639 0.3099 0.0175 25.8061 0.7424 − 73.9442** 5739
IN 0.0005 − 0.1187 0.1631 0.0149 10.0010 − 0.1242 − 69.1956** 5739
ID 0.0005 − 0.1195 0.1403 0.0151 11.8284 0.0327 − 65.6407** 5739
JP 0.0001 − 0.1141 0.1415 0.0149 8.7971 − 0.1237 − 78.7296** 5739
HK 0.0003 − 0.1370 0.1882 0.0160 14.7057 0.3911 − 76.4475** 5739
MY 0.0002 − 0.2071 0.2344 0.0127 65.9316 1.8058 − 69.8174** 5739
NZ 0.0001 − 0.1507 0.1179 0.0104 15.9717 − 0.3705 − 76.3666** 5739
PK 0.0007 − 0.1238 0.1361 0.0147 9.9650 − 0.1841 − 69.2419** 5739
PH 0.0002 − 0.1278 0.1769 0.0148 15.0100 0.5099 − 66.7017** 5739
SG 0.0001 − 0.0936 0.1160 0.0126 9.9460 0.1695 − 71.7193** 5739
KR 0.0003 − 0.1196 0.1572 0.0180 9.3491 0.2213 − 72.4785** 5739
LK 0.0004 − 0.1297 0.2007 0.0105 42.2018 0.9596 − 61.2101** 5739
TH 0.0001 − 0.1484 0.1202 0.0152 10.9150 0.2835 − 70.1885** 5739
TW 0.0001 − 0.0980 0.0961 0.0149 6.0186 0.0605 − 74.3705** 5739
13
Table 2 (continued)
Country Mean Min Max Std. dev Kurtosis Skewness ADF test No. obs.
ZA 0.0005 − 0.1278 0.0816 0.0118 9.5149 − 0.4721 − 69.8633** 5739
EG 0.0007 − 0.1552 0.1385 0.0162 10.0429 0.0050 − 68.2147** 5739
The sample period is January 1995–December 2016. The augmented Dickey–Fuller (ADF) statistic tests
for unit root
**Statistical significance at a 5% level
Table 3 Phases of the sample

Phase Period Representing Number of
observa-
tions
All phases 01.01.1999–29.12.2017 Entire period 4956

Phase 1 01.01.1999–14.09.2008 Pre-crisis 2531
Phase 2 15.09.2008–31.03.2010 Global financial crisis 403
Phase 3 01.04.2010–21.11.2013 European debt crisis 951
Phase 4 22.11.2013–29.12.2017 Recent period 1071
Fig. 1 Average liability for the countries included in this study: 1999Q1-2017Q4.
The source of the data is BIS database with our own plotting
13
Table 4 Structural betas for countries in our sample

Country All phases Phase 1 Phase 2 Phase 3 Phase 4
Austria 0.9090 0.8274 0.9694 1.0169 0.9924

Belgium 0.0422 0.0134 0.8190 0.9975 0.9809
Czech Republic 0.5767 0.5701 0.8952 0.2423 0.3125
Denmark 0.0271 0.0145 0.9037 0.1605 0.5133
Finland 0.5594 0.5234 0.9620 0.5400 0.2193
France 0.0352 0.0676 0.9956 1.0077 0.9762
Germany 0.8195 0.5882 0.9779 1.0063 0.9054
Greece 0.0033 0.0032 0.0024 0.0570 0.0083
Hungary 0.6966 0.6444 0.8784 0.7639 0.3745
Ireland 0.3212 0.5090 0.0875 0.7248 0.9697
Italy 0.1926 0.0663 0.9905 0.9991 0.9333
Netherlands 0.8997 0.9320 0.8402 0.8932 0.9464
Norway 0.8838 0.7685 0.9929 0.9766 0.8878
Poland 0.0510 0.0316 0.9387 0.9882 0.9883
Portugal 0.9515 0.8766 0.9958 0.9775 0.9806
Romania 0.3757 0.3440 1.0007 0.9377 0.9911
Spain 0.9295 0.8406 0.9951 0.9459 0.9805
Sweden 0.1073 0.0308 0.9594 0.9977 0.9584
Switzerland 0.1837 0.0420 0.9967 0.9978 0.9968
Turkey 0.1006 0.1005 0.9740 0.9856 0.9654
United Kingdom 0.4765 0.1825 0.9918 0.9810 0.9833
Argentina 0.4963 0.2596 0.6725 0.9745 0.9940
Brazil 0.7767 0.5681 0.9534 0.7994 0.8971
Chile 0.9786 0.9900 1.0264 0.9822 0.8356
Mexico 0.7463 0.7059 0.9603 0.7001 0.9993
Canada 0.8965 0.9078 0.6805 0.9878 0.9823
United States 0.2102 0.0383 0.9776 1.0006 0.9968
Australia 0.4424 0.0911 0.8300 0.9899 0.9881
China 0.0302 0.0040 0.8010 1.0001 0.9122
India 0.7825 0.6903 0.9681 0.9860 1.0012
Indonesia 0.8583 0.8606 0.7730 0.9007 0.8730
Japan 0.0373 0.0076 0.8250 0.8125 0.9451
Hong Kong 0.0195 0.0167 0.9617 0.3539 0.8560
Malaysia 0.8413 0.8380 0.7939 0.9454 0.9828
New Zealand 0.0448 0.0055 0.9888 0.9819 0.9933
Pakistan 0.0033 0.0025 0.1978 0.8878 0.9124
Philippines 0.6169 0.7677 0.0608 0.9821 0.9436
Singapore 0.9806 0.9819 0.8075 0.8190 0.8837
South Korea 0.0004 0.0002 0.7882 0.7543 0.7782
Sri Lanka 0.8996 − 0.0014 − 0.0256 0.9792 0.9869
Thailand 0.2956 0.2913 0.9771 0.1056 0.9502
Taiwan 0.2038 0.1772 0.5884 0.8109 0.6233
South Africa 0.9806 0.9791 0.9834 0.9785 0.9839
Egypt 0.8001 0.7772 0.2363 0.9202 0.9857
Israel 0.0816 0.0744 0.9789 1.0065 0.9684
The period covered in the sample is 1 January 1999–31 December 2017. All coefficients are at a 5% level
of significance
13
Figure 1 displays the average cross-border liability flow, measured in US billion

dollars, for the countries in our sample. Liability defined as what a country or com-
pany owes to others (including loans, bonds and other debts), plays a significant role
in propagating shocks in the financial system. Having unsecured lending and bor-
rowing could increase cross-border liability within the financial system. There was
a change in the average liability between the entire period of our sample. As shown
in Fig. 1, the average liability drastically changed during crises, implying that on
average, countries in our sample paid more than required by a liability. This may
be a contributory factor to the collapse of the financial system, because the failure
of these countries to pay liabilities could lead to defaults for their counter-parties.
Therefore, liability within countries serves as a major contributory factor to crises.
The higher the liability, the greater the chance of system exposure to distress. These
findings concur with those of Gai and Kapadia (2010), who found that liability from
defaulting banks led to the spread of contagion, which in turn increased the vulner-
ability of interconnected institutions.
4.1 Effect of financial network to common factor model
This subsection investigates whether network exposure affects common factors.

This enables greater understanding of the importance of financial networks in both
spreading and absorbing financial shocks in the system. Our investigation focuses on
the individual countries in our sample. Following Billio et al. (2015), we estimated
both systematic and idiosyncratic components in the structural model. Our analysis
relied on estimating these parameters because the Fama–French factors available in
Kenneth French’s data library were limited to few countries. Since the idiosyncratic
component is unobservable and model-dependent, we used indirect estimation pro-
posed by Campbell et al. (2001) to estimate it.
4.1.1 Estimating beta and idiosyncratic volatilities
To estimate idiosyncratic volatility for an individual stock in our sample, we

assumed the return of each country i to be driven by a common factor and country-
specific shock 𝜀it . To be precise, we followed Sharpe (1964) and Lintner (1965),
who assumed a single factor return generating process and estimated the market
model using Eq. (1).
In this analysis, we computed the excess returns of individual countries as the log
return on the global market index (rmt) minus absolute change of 90 days’ T-Bill rates
(rft), which we considered the risk-free rate. The 90 days’ T-Bill rates and market value
data were obtained from Thompson Reuter’s Datastream for January 1999–December
2017.
The return on the global market index (rmt ) was computed as the value of
weighted excess return of each country over the 90 days T-Bill rates rft of each
country:
13
∑
n
rmt = 𝜔it ∗ rit (9)
i=1
where 𝜔it is the ratio of country’s i market value to the total market value of the
entire market m in time t and n is the total number of countries.
The beta and residual estimates for each market were obtained by running the
regression for each market index in the sample using Eq. (1).
Following Bali and Cakici (2008), we estimated country-specific idiosyncratic
volatility as the standard deviation of the residuals of each individual country given
by:
√
IVOLit = var(𝜀it ) (10)
4.1.2 Effect of financial network on betas
The beta estimate for each country was estimated by running separate regressions.
Table 4 displays the structural betas for individual countries in our sample. We esti-
mated betas using Eq. (1) for all phases categorized in Table 3. Our results clearly
show that beta coefficients are different in all phases. In most countries, the struc-
tural betas were lower in the pre-crisis period, while increased in the GFC with the
exception of some emerging markets (Indonesia, Malaysia, Philippines, Singapore,
Sri Lanka and Egypt) in which estimates decreased during the global financial crisis
period. The structural betas remain high in Phase 3, associated with the European
debt crisis. Country-specific betas changed with the introduction of many factors.
For instance, using Fama–French three-factor model might result in different esti-
mates. Since our focus was the effect of network exposure on structural betas, we
did not focus on discussing each country’s specific betas.
Using beta coefficients obtained from the regression model, we investigated how
they changed with the increase in network exposure. First, we examined the effect of
the connection matrix (W) on the structural beta. Figure 2a displays how the struc-
tural beta (𝛽̄i ) changed with the interaction of the connection matrix, given by:
∑
n
𝛽i∗ = 𝛽̄i + W 𝛽̄i (11)
i=1
where 𝛽i∗ is the new (augmented) beta obtained from the interaction with the connec-
tion matrix while 𝛽̄i is the country-specific structural beta. The results show that the
structural beta changes with the interaction with the connection matrix. The connec-
tion matrix is based on liability linkages obtained using the combined Granger cau-
sality and DY measure (Chowdhury et al., 2019). This suggests that the increased
interconnection between various market participants leads to change in the structural
beta of a given country. The results also revealed the role of the weighting matrix in
spreading shocks in a financial system. For example, the connection matrix increased
the values of structural betas by more than 50% for countries whose beta values
were small. Countries with low betas included Greece, Philippines and Poland. The
13
Fig. 2 Structural betas for all the phases under different scenarios. The figures exhibit how the struc-
tural beta for all phases changed under different scenarios. We used country’ abbreviations from BIS.
a Displays changes in betas with and without the presence of the connection matrix. This was obtained
by multiplying the betas with the corresponding liability weighted matrix. b Shows how betas changed
using sparse weighting matrix. c Shows how betas changed across different network intensity parameters.
d Shows the effect of network exposure on betas. The period covered in the sample is 1 January 1999–31
December 2017
size of the augmented betas varied depending on the strength of the connections a
country has with others. In Fig. 2a, we noted that beta values of countries (includ-
ing Denmark, Greece, Sri Lanka, Malaysia and the Philippines) tend to be 0 but are
greatly influenced by the weighting matrix. The beta values increased depending on
the level of shock one country receives from others. These results depict the role of
interconnections in spreading risk. This observation is consistent with Glasserman
and Young (2016), who showed that countries with high connectivity tend to suffer
more when shocks hit the financial system.
For countries (including Greece, Sri Lanka and the Philippines) with lower beta
estimates due to stronger links with other countries, we observed that the structural
beta was amplified. This implies that these countries are strongly affected by other
countries, leading to amplification of shocks.
13
It is important to understand whether using a different weighting matrix has a dif-

ferent impact on structural betas. Figure 2b shows how betas changed using a sparse
matrix.7 We randomly generated a sparse matrix. From the results, we noted that the
contribution of the weighting matrix changed depending on the strength of connec-
tions between countries. This is depicted using orange bars in Figures 2a and b. For
example, the Philippines had stronger links in Fig. 2a, leading to greater change in
exposure, while in Fig. 2b, it had weak links, leading to a smaller change in struc-
tural betas.
The impact of the different network intensity parameters on the structural betas
can be estimated as:
∑
n
𝛽i∗∗ = 𝛽̄i + 𝜌𝛽̄i (12)
i=1
where 𝛽i∗∗ is the new beta obtained from the interaction with the different network
intensity parameters.
Figure 2c shows how structural betas changed across the different network inten-
sity parameters (we assumed network intensity parameters to take quartiles values
[i.e. 0.25, 0.5 and 0.75]). It is clear from these that structural betas tend to increase
across different network intensity parameters. This is an indicator that as the network
intensity increases, the level of risk in the financial system also tends to increase.
Unlike the connection matrix, whose effect is severe to all countries with stronger
connections including countries whose beta values are small, the network inten-
sity parameters have more influence on the countries whose beta values are large.
Countries with small values of beta (Greece, Sri Lanka and the Philippines) are
less affected by network intensity parameters. Conversely, countries with high beta
estimates (e.g. South Korea) are more affected by larger network intensity param-
eters. This explains the role of network intensity parameters in spreading and reduc-
ing shocks. Our finding indicates that network intensity parameters have a greater
impact in spreading risk than absorbing it.
Next, we investigated the effects of network exposure on structural betas. This
involved combining the connection matrix and network intensity parameters. This is
because both coefficients have a great impact on betas. We used the following equa-
tion to gauge the role of network exposure on structural betas:
∑
n
𝛽i∗∗∗ = 𝛽̄i + 𝜌W 𝛽̄i (13)
i=1
where 𝛽i∗∗∗ is the new beta obtained from the interaction with the changing network
exposure.
Figure 2d shows how the structural betas change across the changing network
exposure. These results show that both the weighting matrix and the network inten-
sity parameters have effects on the structural beta.
7
We can define sparse matrix as a connection matrix with few interconnected institutions/assets.
13
Fig. 3 Change in structural betas due to network exposures. The figure reports contribution of connection matrix, network intensity parameters and network exposures on
structural betas. Blue represents % change of betas due to the connection matrix, orange due to network intensity coefficient and yellow due to the network exposures. The
period covered in the sample is 1 January 1999–31 December 2017
589
13
To obtain a clear insight into how structural beta changed, we calculated the per-
centage change in betas in 𝛽i∗, 𝛽i∗∗ and 𝛽i∗∗∗. Figure 3 summarizes the contribution
of connection matrix, network intensity parameters and network exposure to struc-
tural betas. The height of the bar represents the percentage change of betas. These
results revealed stronger connections between markets increase network exposure.
This is consistent with Silva et al. (2016) who found that high connectivity trig-
gers a greater probability of default in the financial system. It also shows that net-
work intensity parameter has an amplifying effect on shocks. It is clear from the bar
size that both the weighing matrix and the network intensity parameters have differ-
ent effects. Although the contribution of the connection matrix is similar, it varies
depending on the strength and number of the connections. Thus, our main finding
from Fig. 3 is that both network intensity parameters and connection matrix are key
ingredients in either increasing or decreasing network exposure.
4.1.3 Network exposure to common factors
To investigate the effect of network exposure on common factors, we used slope

coefficient betas as the systematic risk of specific country’s market portfolios. This
is useful to examine the effect of network exposure on both systematic and idio-
syncratic volatility. Bali and Cakici (2010) used beta coefficients as the systematic
risk of a country’s market portfolio to determine whether country-specific risks are
priced into the intertemporal capital asset pricing model (ICAPM). They found that
country-specific risks are significantly priced into the ICAPM framework.
Table 5 reports country-specific idiosyncratic volatility in all sample periods,
estimated in Eq. (10). As noted by Hueng and Yau (2013), these estimates may
vary depending on the data used because they are model-dependent. Notably, the
country-specific idiosyncratic volatility of emerging markets as greater than those of
developed economies is consistent with the findings of Bali and Cakici (2008, 2010)
and Hueng and Yau (2013). We also observed, on average, that Turkey had high idi-
osyncratic volatility in the full sample period and also in Phase 1. These results are
consistent with Bali and Cakici (2010) and Hueng and Yau (2013), who determined
that Turkey had a higher estimate than other countries in their sample. Interestingly,
most countries had greater idiosyncratic estimates in the crisis period. This could be
explained by higher uncertainty in the market during the GFC.
Therefore, we will not discuss country-specific idiosyncratic volatilities because
our aim is to determine the role of network exposure on both systematic and idi-
osyncratic components. By assuming the first order neighborhood in Eq. (7), we
investigated the effect of network exposure on the structural model. We used the net-
work intensity parameter to capture the strength of the network exposure (network
intensity). This coefficient lies between 0 and 1, where close to 0 implies lower net-
work intensity and close to 1 signifies higher network intensity. The existence of net-
work exposure is captured by the weighting matrix which was row-normalized. The
weighting matrix takes the values between 0 and 1 as representing exposure from
other markets, where values close to 0 imply less exposure, while close to 1 implies
high exposure. We used the weighting matrix constructed from combined Granger
13
Table 5 Estimates of Country All phases Phase 1 Phase 2 Phase 3 Phase 4

idiosyncratic volatilities
Austria 0.0179 0.0185 0.0275 0.0151 0.0110
Belgium 0.0359 0.0281 0.0324 0.0114 0.0098
Czech Republic 0.0201 0.0235 0.0273 0.0131 0.0083
Denmark 0.0280 0.0279 0.0266 0.0148 0.0127
Finland 0.0180 0.0198 0.0225 0.0158 0.0110
France 0.0352 0.0288 0.0217 0.0145 0.0110
Germany 0.0219 0.0234 0.0225 0.0128 0.0146
Greece 0.0219 0.0552 0.0224 0.0214 0.0240
Hungary 0.0559 0.0701 0.0643 0.0201 0.0156
Ireland 0.0296 0.0261 0.0441 0.0259 0.0115
Italy 0.0352 0.0283 0.0237 0.0180 0.0156
Netherlands 0.0176 0.0141 0.0332 0.0219 0.0108
Norway 0.0201 0.0221 0.0284 0.0123 0.0115
Poland 0.0983 0.1125 0.0285 0.0155 0.0131
Portugal 0.0147 0.0142 0.0168 0.0166 0.0124
Romania 0.4326 0.5782 0.0620 0.0239 0.0084
Spain 0.0177 0.0174 0.0212 0.0211 0.0119
Sweden 0.0493 0.0356 0.0238 0.0131 0.0157
Switzerland 0.0651 0.0404 0.0184 0.0104 0.0102
Turkey 2.4059 3.3873 0.0321 0.0169 0.0205
United Kingdom 0.0429 0.0336 0.0190 0.0134 0.0110
Argentina 0.1950 0.1980 0.0595 0.0290 0.0417
Brazil 0.0450 0.0491 0.0562 0.0265 0.0179
Chile 0.0122 0.0106 0.0118 0.0116 0.0143
Mexico 0.0283 0.0366 0.0217 0.0143 0.0089
Canada 0.0219 0.0150 0.0541 0.0107 0.0098
United States 0.0734 0.0410 0.0416 0.0121 0.0081
Australia 0.0525 0.0312 0.0602 0.0138 0.0089
China 0.0672 0.0398 0.0363 0.0155 0.0222
India 0.1183 0.1525 0.0767 0.0234 0.0087
Indonesia 0.0418 0.0546 0.0392 0.0180 0.0148
Israel 0.0980 0.1307 0.0194 0.0116 0.0068
Japan 0.0487 0.0324 0.0358 0.0380 0.0151
Hong Kong 0.0572 0.0686 0.0252 0.0128 0.0106
Malaysia 0.0149 0.0173 0.0253 0.0066 0.0055
New Zealand 0.0696 0.0345 0.0235 0.0177 0.0110
Pakistan 0.0714 0.0851 0.0702 0.0216 0.0136
Philippines 0.0854 0.0844 0.0737 0.0208 0.0158
Singapore 0.0227 0.0295 0.0224 0.0087 0.0093
South Korea 0.0334 0.0367 0.0359 0.0141 0.0092
Sri Lanka 0.0562 0.0125 0.0132 0.0521 0.0201
Thailand 0.0627 0.0685 0.0300 0.0463 0.0131
Taiwan 0.0250 0.0315 0.0241 0.0115 0.0081
South Africa 0.0216 0.0216 0.0275 0.0243 0.0162
Egypt 0.0632 0.0534 0.0877 0.0409 0.0245
The table reports the idiosyncratic volatilities for all countries in

all phases. The period covered in the sample is 1 January 1999–31
December 2017
13
causality and DY approach by using the cross-border liabilities. Based on simple

continuity, we assumed that the network intensity parameters exert a similar effect
on each country. To be precise, we assume this network intensity parameter to be
0.5, which is related to the mean estimate obtained in Sect. 6. Other studies, includ-
ing Blasques et al. (2016), found the estimate to be higher (approximately 0.7). Fig-
ure 4 shows how the structural model responds to network exposure across the entire
sample period. The blue bar represents the structural systematic component, and the
yellow bar is the idiosyncratic component. The orange bar is the absolute contribu-
tion of network exposure to change in the systematic component while the purple
bar is the absolute contribution of network exposure to change in the idiosyncratic
component.
Turkey, Romania and Argentina had greater values of idiosyncratic volatility than
other economies with smaller estimates. We observed that the systematic component
was predominant in most economies compared to the idiosyncratic volatility. The
effect of network exposure on systematic component was higher (represented by the
size of the orange bars in Fig. 4) than on idiosyncratic component, with the excep-
tion of Turkey, Argentina and Romania.
Figure 5 shows the contribution of network exposure on the structural model in
different periods. On average, the idiosyncratic volatility of Turkey was still domi-
nant in Phase 1 but reduced in all other phases. We can relate Turkey’s high idi-
osyncratic volatility to the banking crisis that led to capital flight and recession in
the economy at the end of 2000. This demonstrates that Turkey’s banking crisis was
largely idiosyncratic even though it could have been triggered by other external fac-
tors. Higher idiosyncratic volatility also explains the ability of Turkey’s investors
(who are mostly foreign) to diversify their portfolios. Turkey’s idiosyncratic volatil-
ity seemed to diminish in Phase 2. Surprisingly, we expected it to increase due to
the GFC. This may have happened due to Turkey’s restructure of its financial system
after the banking crisis in 2001.
The results in Phase 1 for all other countries shows that the network contribution
to the idiosyncratic component is almost irrelevant. A possible suggestion is that
the network exposure has a diversifying effect on the idiosyncratic component. The
network contribution to the systematic component was large; thus, it has an amplify-
ing effect on the systematic component. The results in all other phases indicate that
network exposure has a greater impact on the systematic component and less impact
on idiosyncratic volatility. These results support the notion that network exposure
contributes to spreading and diversifying risks.
In general, our contribution highlights the distinction between the spreading and
sharing of sharing. From Fig. 5, it can be observed that the presence of network
exposure increases systematic risk and reduces idiosyncratic risk. Institutions with
increased unsecured borrowing and lending have a higher chance of receiving and
spreading shocks to other institutions. Figure 5 also reveals the changing nature of
interconnections in the different phases. This is because we used a constant net-
work intensity parameter while changing the connection matrix. Billio et al. (2015)
reported similar results in which the presence of network effect increased the sys-
tematic component and decreased the idiosyncratic component. With the increase
in network intensity parameters (from 0.5 to 0.75), the financial system became
13
Fig. 4 Network exposures on systematic and idiosyncratic shocks in the entire period. The figure reports both the structural and network contribution to systematic risk
and idiosyncratic volatilities for entire period. The period covered in the sample is 1 January 1999–31 December 2017
593
13
Fig. 5 Network exposures on systematic and idiosyncratic shocks in all the phases. The figures show
the contribution of network exposure to both systematic and idiosyncratic volatility in each phase. The
period covered in the sample is 1 January 1999–31 December 2017
more vulnerable to shocks and benefited more from diversification. If the network
intensity parameter is decreased from 0.5 to 0.25, there will be a decrease in shock
spreading and a reduced diversification effect. These results are consistent with
recent studies that showed that the presence of interconnection increases vulnerabil-
ity in the financial system while also helping to diversify risks.8 The results also
indicate that network exposure had greater impact in Phases 2, 3 and 4 than Phase
1. This is attributable to the GFC in Phase 2, EDC in Phase 3 and Chinese market
crash in Phase 4.
We also investigated the effect of network connectivity on the structural model.
For our case, we used a sparse matrix to determine whether a decrease in network
connectivity increased or decreased network exposure. We randomly generated the
sparse matrix. As shown in Fig. 6, fewer interconnections in the financial system led
to a reduction in the size of the bars. This suggests that having fewer interconnec-
tions reduces the magnitude of the spread of absorption of shocks .
8
See Acemoglu et al. (2015), Tonzer (2015) and Gai and Kapadia (2010).
13
Fig. 6 Effect of network exposures with decreased interconnectedness. The figures show how the con-
tribution of network exposure to both systematic and idiosyncratic volatility changed with reduced inter-
connections between markets. This involved using a sparse weighting matrix. The period covered in the
sample is 1 January 1999–31 December 2017
We also used dense matrix to determine its effect on the structural model. Fig-
ure 7 shows the contribution of using denser weighting matrix in all sample peri-
ods. This weighting matrix was based on trade linkages. Trade linkages are denser
because countries are more interlinked through bilateral trade. This figure shows
that more interconnection in the financial system increases the systematic risk and
reduces the idiosyncratic component. This implies that using a denser weighting
matrix increases the network exposure and amplify shocks as well as diversify some
shocks. Overall, our approach shows that an increase in network exposure signifi-
cantly increases systematic shocks through risk spreading, and reduces idiosyncratic
risk through diversification. Further, these results signify that a large degree of con-
nectivity does not necessarily dampen risk exposure, but amplifies shocks in the
financial system. This is consistent with Amini et al. (2016), who affirmed that an
increase in network connection may lead to systemic instability.
13
596
13
Fig. 7 Effect of network exposures with increased interconnectedness. The figure shows how the contribution of network exposure to both systematic and idiosyncratic
volatility changed with increased interconnections between markets. This involved using a denser connection matrix. The period covered in the sample is 1 January 1999–
31 December 2017
5 Optimal value of the network intensity parameter
The next question to address is the estimation of the network intensity parameter
(𝜌), which captures the strength of the network exposure. This coefficient is also
important, as it plays a key role in monitoring network exposure (Sect. 4). Differ-
ent estimation methods have been proposed to estimate network intensity parameter,
including ordinary least squares (OLS), maximum likelihood estimation (MLE),
method of moments (MoM), two-stage least squares (2SLS) and the generalized
method of moments (GMM).
5.1 Static network intensity parameter
We introduce the spatial autoregressive (SAR) model to measure network intensity

parameter. We begin by considering a simple first-order (pure) SAR model, where
‘pure’ refers to the absence of exogenous regressors (Xn ) as proposed by Anselin
(1988) and is defined as:
y = 𝜌Wy + 𝜀 (14)
where n is the number of observations, y = (y1 , y2 , … , yn )� is a vector of observa-
tions on the dependent variable, W is n × n exogenous connection matrices, 𝜌 is a
scalar representing the network intensity coefficient and 𝜀 = (𝜀1 , 𝜀2 , … , 𝜀n )� as a
vector of residuals assumed to be independent and identically distributed.
By rearranging Eq. (14), the error term yields:
𝜀 = y − 𝜌Wy (15)
LeSage and Pace (2009) describe the vector Wy as spatial lag representing a linear
combination of the neighboring values to each observation. This is supported by
Lee (2007), who shows that the influence in the neighboring asset is due to spatial
effects.
Our first estimate of 𝜌 is based on OLS. However, the OLS estimate of 𝜌 is con-
sidered biased and inconsistent. Following Anselin (1988), the OLS estimate of 𝜌 is
denoted by p̂ and given by,
p̂ = (y� W � Wy)−1 y� W � y (16)

An estimate of 𝜌 is unbiased if E(̂p) = p. We prove below that E(̂p) = p.
[ ]
E(̂p) = E (y� W � Wy)−1 y� W � (𝜌Wy + 𝜀)
[ ] (17)
= 𝜌 + E (y� W � Wy)−1 y� W � 𝜀
Therefore, the OLS estimate is biased since the second term in Eq. (17) does not
equal 0, E(̂p) ≠ p. To show that the OLS estimate is inconsistent, Anselin (1988)
demonstrated that the probability limit (plim) for the term y′ W ′ 𝜀 is not 0 except in
13
trivial cases when 𝜀 = 0.9 Since the OLS estimate is biased and inconsistent, we had
to consider alternatives to estimate the network intensity parameter.
We first consider the MLE method because of its simplicity. According to LeSage
and Pace (2009), the MLE for 𝜌 requires identification of the value of the SAR coef-
ficient that maximizes the likelihood function, given by:
{ }
1 1 �
L(𝜎 2 ;𝜀) = exp − (y − 𝜌Wy) (y − 𝜌Wy)
2Π𝜎 2(n∕2) 2𝜎 2
{ } (18)
1 1 �
L(y|𝜌, 𝜎 2 ) = (J)exp − (y − 𝜌Wy) (y − 𝜌Wy)
2Π𝜎 2(n∕2) 2𝜎 2
The Jacobian function can be obtained through the differentiation of Eq. (15) with
respect to the dependent variable y yielding:
𝜕𝜀
J= = |In − 𝜌W| (19)
𝜕y
where In is n × n identity matrix. Substituting Eq. (19) for Eq. (18) gives:
{ }
1 1
L(y|𝜌, 𝜎 2 ) = |In − 𝜌W|exp − (y − 𝜌Wy) �
(y − 𝜌Wy) (20)
2Π𝜎 2(n∕2) 2𝜎 2
Lee (2002) asserts that deriving the eigenvalue (𝜆) of the connection matrix W sim-
plifies the computation problem. The natural logarithm of Eq. (20) is:
n 1
ln L = − ln(2Π𝜎 2 ) + ln |In − 𝜌W| − 2 (y − 𝜌Wy)� (y − 𝜌Wy)
2 2𝜎
(21)
n n 1
= − ln(2Π) − ln(𝜎 2 ) + ln |In − 𝜌W| − 2 (y − 𝜌Wy)� (y − 𝜌Wy)
2 2 2𝜎
The natural logarithm can be further restructured by eliminating the residual param-
eter, 𝜎 2. This is achieved by substituting with the error term, given by:
1
𝜎̂ 2 = (y − 𝜌Wy)� (y − 𝜌Wy) (22)
n
This yields to:
n n
ln L = − ln(2Π) − ln(y − 𝜌Wy)� (y − 𝜌Wy) + ln |In − 𝜌W| (23)
2 2
The parameter space of 𝜌 requires that the determinants of In − 𝜌W to be strictly
positive. A univariate optimization problem can be used to maximize the above
expression with respect to 𝜌. This implies that the optimal search of 𝜌 estimates take
feasible values within the range:
1∕𝜆min < 𝜌 > 1∕𝜆max (24)
9
Anselin (1988) shows that plimN −1 (y� W � 𝜀) = plimN −1 𝜀� W(I − pW)−1 𝜀.
13
where 𝜆min is the minimum eigenvalue of the standardized matrix W while 𝜆max is the
largest eigenvalue of the same matrix.
Equation (14) can be extended to investigate how network intensity changes
with the presence of exogenous variables. The mixed-SAR model (see LeSage &
Pace, 2009) can be written as:
y = 𝜌Wy + 𝛽X + 𝜀
(25)
𝜀 ∼ N(0, 𝜎 2 In )
where X = (X1 , X2 , … , Xn ) a vector (n × k , and where k is the number of variables)

of observations on the exogenous variables having 𝛽 = (𝛽1 , 𝛽2 , … , 𝛽k ) coefficients.
Rewriting Eq. (25) repeatedly yields:
y = 𝜌Wy + 𝛽X + 𝜀
= 𝜌Wy(𝜌Wy + 𝛽X + 𝜀) + 𝛽X + 𝜀
= 𝜌Wy(𝜌Wy(𝜌Wy + 𝛽X + 𝜀) + 𝛽X + 𝜀) + 𝛽X + 𝜀 (26)
∑
∞
= [𝜌W]n (𝛽X + 𝜀)
n=1
Equation (26) clearly shows the effect of the weighting matrix in spreading shocks
from one entity to the other until it diminishes leading to a steady-state.
Equation (25) can also be written in a more compact way as:
(I − 𝜌W)y = 𝛽X + 𝜀 (27)
which provides a structure to the contemporaneous relationship based on the spatial
proximity in association with the SAR model. Thus, the model includes contempo-
raneous relationships, driven by interconnections across different assets (markets),
exogenous regressors and asset (market) specific shocks. Equation (27) is conveni-
ently expressed in compact because our focus is to estimate the network intensity
parameter, 𝜌 which captures the endogenous effect of network exposure.
The general idea is to first construct a univariate optimization problem for the
parameter 𝜌 . Following Anselin (1988) and LeSage and Pace (2009), this is done
by maximizing the full likelihood function of the dependent variable with respect
to the unknown parameters. This is given by:
{ }
1 1
L(y|𝜌, 𝛽, 𝜎 2 ) = |I − 𝜌W|exp − (y − 𝜌Wy − X𝛽)�
(y − 𝜌Wy − X𝛽)
2Π𝜎 2(n∕2) n 2𝜎 2
(28)
The natural logarithm function in Eq. (28) can be specified as:
n 1
ln L = − ln(2Π𝜎 2 ) + ln |In − 𝜌W| − 2 e� e
2 2𝜎
n n 1
= − ln(2Π) − ln(𝜎 2 ) + ln |In − 𝜌W| − 2 e� e
2 2 2𝜎
where
13
e = (y − 𝜌Wy − X𝛽)
(29)
𝜌 ∈ (min(𝜔)−1 , max(𝜔)−1 )
where 𝜔 is the eigenvalue constructed from matrix W. The value of 𝜌 is assumed to
be bounded between 0 and 1. Next, we estimate each parameter in Eq. (29). This is
done by solving the first-order derivatives of Eq. (29) with respect to the individual
parameters.
• Estimate of 𝛽 by differentiating Eq. (29) with respect to 𝛽 . We obtain:
𝜕ln(L)
=0
𝜕𝛽
( )
1
𝜕ln(L) 𝜕 2𝜎 2
(y − 𝜌Wy − X𝛽)� (y − 𝜌Wy − X𝛽)
=
𝜕𝛽 𝜕𝛽
( )
𝜕 1
(y − 𝜌Wy − X𝛽)� (y − 𝜌Wy − X𝛽) (30)
2𝜎 2
0=
𝜕𝛽
1
0 = 2 (X � (y − 𝜌Wy) − X � X𝛽)
2𝜎
𝛽 = (X � X)−1 X � (In − 𝜌W)y
From Eq. (30), the estimate of 𝛽 is:
𝛽̂ = (X � X)−1 X � (In − 𝜌W)y (31)
For simplicity, this can be written as:
𝛽̂ = (X � X)−1 X � y − 𝜌(X � X)−1 X � Wy (32)

• Estimate of 𝜎 2 we differentiate Eq. (29) with respect to 𝜎 2 to yield:
𝜕ln(L)
=0
𝜕𝜎 2 ( )
n 2 ) + 1 (y − 𝜌Wy − X𝛽)� (y − 𝜌Wy − X𝛽)
𝜕ln(L) 𝜕 − 2
ln(𝜎 2𝜎 2
2
=
𝜕𝜎 𝜕𝜎 2
0=− 2 +
n 1
(y − 𝜌Wy − X𝛽)� (y − 𝜌Wy − X𝛽) (33)
2𝜎 2(𝜎 2 )2
1
0 = −n + 2 (y − 𝜌Wy − X𝛽)� (y − 𝜌Wy − X𝛽)
𝜎
𝜎2 =
n
Thus, the estimate of 𝜎 2 is given by:
13

𝜎̂2 = (34)
n
• Estimate of 𝜌 unlike 𝛽 and 𝜎 2 , which have closed form solutions, 𝜌 needs
to be estimated using optimization problem that maximizes Eq. (29) with
respect to 𝜌 . By replacing estimates of 𝛽 and 𝜎 2 in Equation (29) and letting
𝛿̂0 = (X � X)−1 X � y , 𝛿̂d = (X � X)−1 X � Wy in Eq. (32), we have:
y = X 𝛿̂0 + ê 0 and Wy = X 𝛿̂d + ê d (35)
which can be estimated by OLS. Thus, Eq. (32) can be rewritten as:
𝛽̂ = (X � X)−1 X � y − 𝜌(X � X)−1 X � Wy

(36)
= 𝛿̂0 − 𝜌𝛿̂d
The error term from Eq. (35) can be given by: ê 0 = y − X 𝛿̂0 and ê d = Wy − X 𝛿̂d .
Substituting to 𝜎 2 yields
(e0 − 𝜌ed )� (e0 − 𝜌ed )
𝜎2 = (37)
n
Using the results of 𝛽 and 𝜎 2, Eq. (29) becomes:
( (e − 𝜌e )� (e − 𝜌e ) )
n n 0 d 0 d
ln L = − ln(2Π) − ln
2 2 n
1
+ ln |In − 𝜌W| −
2( )
n n n
= − ln(2Π) − ln (e0 − 𝜌ed )� (e0 − 𝜌ed ) − ln(n)
2 2 2
1
+ ln |In − 𝜌W| −
2
which can be written as:
( )
n
= c − ln (e0 − 𝜌ed )� (e0 − 𝜌ed ) + ln |In − 𝜌W|
2
n n 1 (38)
c = − ln(2Π) − ln(n) −
2 2 2
Thus, to obtain the estimates of 𝜌, we need to simplify the log-likelihood with
respect to the scalar 𝜌 and optimize the following equation:
� �
⎛c − n ln (e0 − 𝜌1 ed )� (e0 − 𝜌1 ed ) + ln �In − 𝜌1 W�⎞
⎛f (𝜌1 )⎞ ⎜ 2 � � ⎟
⎜f (𝜌n )⎟ ⎜c − n ln (e0 − 𝜌2 ed )� (e0 − 𝜌2 ed ) + ln �In − 𝜌2 W�⎟
⎜ ⋮ ⎟=⎜ 2 ⎟ (39)
⎜ ⎟ ⎜ � ⋮ � ⎟
⎝ f (𝜌r ) ⎠ ⎜ ⎟
⎝ c − 2 ln (e0 − 𝜌r ed ) (e0 − 𝜌r ed ) + ln �In − 𝜌r W� ⎠
n �
13
5.2 Dynamic network intensity parameter
The static SAR model specified in Eq. (25) can be further extended to a dynamic
SAR. This allows estimation of a time-varying network intensity parameter ( 𝜌t ).
This is useful in understanding how the spatial parameter changes over time. As
pointed out by Blasques et al. (2016), time-varying network intensity parameters
indicate how the spillover changes over time. We considered the case when we
have constant volatility.
A time-varying SAR model with constant disturbances is defined as:
yt = 𝜌t Wyt + 𝛽Xt + 𝜀t
(40)
𝜀t ∼ N(0, Σ), t = 1, 2, … , T
where Σ is constant over time.

Assuming constant disturbances allows us to investigate how the spatial
parameter changes over a specific point in time. This is an important aspect of
examining financial systems in which the volatility of returns is known to vary
considerably between non-crisis and crisis periods (Blasques et al., 2016; Catania
& Billé, 2017). The diagonal elements represent the time-conditional variances of
the cross-sectional independent innovation at any given point in time. We impose
diagonality assumption as the standard constant conditional correlation (CCC)
and dynamic conditional correlation (DCC) model proposed by Engle (2002).
The generalized log-likelihood function of the constant ( Lct ) variance models
becomes:
n n ∑ T
1
ln Lct = − ln(2Π) − ln(Σ) + ln |In − 𝜌t W| − e�t Σ−1 et
2 2 t=1
2
where
et = (yt − 𝜌t Wyt − Xt 𝛽)
(41)
𝜌t ∈ (min(𝜔)−1 , max(𝜔)−1 )
Allowing for time-varying variance in the shocks led to the following:

yt = 𝜌t Wyt + 𝛽Xt + 𝜀t
where
𝜀t ∼ N(0, Σt ), t = 1, 2, … , T
(42)
Σt ∼ diag(𝜎12 , 𝜎22 , ⋯ , 𝜎t2 )
The generalized log-likelihood function of the time-varying ( Lvt ) variance models

becomes:
n n ∑ T
1
ln Lvt = − ln(2Π) − ln(Σt ) + ln |In − 𝜌t W| − e�t Σ−1
t et
2 2 t=1
2
13
Table 6 Network intensity estimates and their robust standard errors

All phases Phase 1 Phase 2 Phase 3 Phase 4
Whole sample 0.5072 (0.0043) 0.4727 (0.0059) 0.6134 (0.0142) 0.5495 (0.0099) 0.5110 (0.0091)
Advanced 0.5292 (0.0042) 0.5014 (0.0059) 0.5919 (0.0144) 0.5664 (0.0097) 0.5378 (0.0089)
Emerging 0.3125 (0.0040) 0.2804 (0.0054) 0.4232 (0.0147) 0.3379 (0.0095) 0.3236 (0.0084)
Europe 0.5281 (0.0043) 0.5018 (0.0059) 0.6196 (0.0147) 0.5651 (0.0099) 0.5232 (0.0092)
All America 0.3497 (0.0040) 0.3327 (0.0056) 0.4447 (0.0149) 0.3758 (0.0098) 0.3321 (0.0081)
Asia 0.3515 (0.0041) 0.3132 (0.0055) 0.4466 (0.0148) 0.4068 (0.0097) 0.3569 (0.0084)
The table reports the estimated network intensity and their robust standard errors in parentheses for the
static SAR model. The period covered in the sample is 1 January 1999–31 December 2017
where
et = (yt − 𝜌t Wyt − Xt 𝛽)
(43)
𝜌t ∈ (min(𝜔)−1 , max(𝜔)−1 )
6 Empirical analysis
We use the MLE method to estimate the values of 𝜌. MLE is preferred over OLS
because of the limitations of OLS as discussed in Sect. 5.1. Despite the known limi-
tations of OLS in estimating 𝜌, Elhorst (2010) stated that the OLS estimate of 𝜌
could serve as a guide of the expected true value. The initial OLS estimate of 𝜌 for
our data was 0.5327. Therefore, it is expected that the optimal value of the estimate
of 𝜌 will be within this range. Next, we estimate 𝜌 using the MLE and allow the
search to be within the range 1∕𝜆min– 1∕𝜆max . The y vector is the average return of
each country in our sample over the entire sample period, while we constrained W
to lie within the interval {0,1} through the process of row standardization; using the
row-normalized contiguity matrix of weights ensures that each row of the matrix
sums to unity. The row-normalized matrix represents the portion of total liability
that the source country/institution shares among its target nodes.
By estimating static network intensity parameter using a pure SAR model
(Eq. 21), we ensure consistency with the extant spatial literature (Asgharian et al.,
2013; Fernandez, 2011). The static network intensity parameter (which captures the
endogenous effect of network exposure) is estimated at each phase. We begin the
estimation by excluding additional explanatory variables. Table 6 contains the static
network intensity estimates with their corresponding standard errors in parentheses.
The estimate for the whole sample was 0.5072 with a small standard error of 0.0043.
This is close enough to the estimate obtained using OLS (0.5327) and represents the
potential impact of the parameter on the entire network. Comparing estimates in the
different phases, the network intensity parameter was higher in Phase 2 (0.6134).
From Fig. 8, we observed a more than 30% increase in the estimate from Phase 1
13
Fig. 8 Change in network intensity estimates in subsequent phases. The figure reports the percentage
change of network intensity estimates in the subsequent phases. The entire period covered in the sample
is 1 January 1999–31 December 2017
to 2. It only increased by 16% in Phase 3 which is approximately half the increase

reported for Phase 2.
From the static network intensity results in Table 6 and Fig. 8, it is evident that
network intensity increased drastically in Phase 2 (corresponding to the GFC), fol-
lowed by in phase 3 (associated with the EDC) compared to the other two phases.
This is consistent with the finding of Blasques et al. (2016), who used the CDS data
of big players in Europe to show that network intensity is higher when the finan-
cial system is under stress, suggesting higher spillover in the financial system. Our
results also reveal that the GFC, which is associated with large network intensity,
had a severe impact on the entire financial system compared to the EDC. The GFC
spread throughout the entire network while the EDC severely affected only Euro-
pean countries.
Overall, we can conclude that increases in network intensity estimate could be
associated with periods when the financial system is under stress. This is in line with
Hypothesis 1, which states that higher network intensity could be associated with
crises. For instance, higher estimates in Phase 2 corresponded to GFC; in Phase 3,
they could be associated with the EDC. Finally, in Phase 4, they could correspond to
the Chinese market crash of 2015 (Alter & Beyer, 2014; Black et al., 2016; Yu et al.,
2017).
The increase in the estimate during times of stress could have an economic
effect. Large network intensity estimates are assumed to signify higher propagation
strength of a shock to the entire system. This is because of high interconnectedness,
which increases network exposure, and thus, may increase fragility in the financial
system (Minoiu & Reyes, 2013). We can relate this to the findings of Tonzer (2015)
and argue that high network intensity is associated with increased cross-border
13
Fig. 9 Network intensity estimates for the entire period. The figures display the network effect for the
whole sample period without regressors. The light areas are the 95% confident intervals with the hori-
zontal line representing average estimate in the whole period. Figure 9a presents the dynamic estimates
obtained with an assumption of constant (𝜌 = 0.5) initial value while Fig. 9b displays the estimates
obtained using changing initial value (it follows 𝜌t = 0.5 + 0.4 ∗ cos(2𝜋t∕200)). Figure 9c presents net-
work intensity estimates using sparse matrix. The period covered in the sample is 1 January 1999–31
December 2017
exposures.10 Our findings are also similar to those of Cao et al. (2017), who found
that cross-border linkages tend to increase during crises. This could be a signal of
greater propagation of shock when institutions are under distress.
To capture the dynamics of the network intensity parameter, we conduct an esti-
mation using a 251-day (one-year horizon) rolling window. We investigate how
network intensity changed over time. A one-year period is assumed to be adequate
to capture any significant change in economies. Before estimating the param-
eter, it would be interesting to determine if the estimates differed using a con-
stant initial value of 𝜌 and a changing initial value of 𝜌 at each point in time (the
initial value will be used as the starting values in search of the real values in our
10
Tonzer (2015) argues that the foreign exposure during the GFC increased in the banking sector, lead-
ing to risk spreading through the interconnected links in the financial system.
13
optimization problem). To proxy the initial changing values of 𝜌, we assume the

pattern of the network intensity parameter to be same as those used by Blasques
et al. (2016).11 Specifically, we assumed the following: constant (𝜌t = 0.5), sine
(𝜌t = 0.5 + 0.4 ∗ cos(2𝜋t∕200), fast sine (𝜌t = 0.5 + 0.4 ∗ cos(2𝜋t∕20) and step
(𝜌t = 0.9 − 0.5 ∗ (k > 500)). We estimate the network intensity parameters by both
excluding and including the additional explanatory regressors.
The evolution of the network intensity parameter for the whole sample with the
95% confidence intervals are presented in Fig. 9a–c. In terms of whether using the
different initialization of 𝜌 within the range of 0 and 1 leads to different estimates,
our results indicate that dynamic network intensity estimates are identical when
using any specification of 𝜌 within (0,1). Next, we investigate whether using varying
initial values of 𝜌 would result in different estimates of the network intensity coeffi-
cient. The results in Fig. 9a and b show that both initializations of 𝜌 produce similar
plots. The implication here is that dynamic network intensity estimates does not nec-
essarily depend on the initial value of 𝜌.
Figure 9a and b clearly show that the network intensity parameter changes over
time. This is consistent with other studies, such as Forbes and Rigobon (2002),
who stated that spillover is time variant. We also observed that estimates oscillated
between 0.2 and 0.8, which signifies a higher variation of propagation chance of
shock hitting specific nodes in the network. There is a notable repetition of simi-
lar trends of network intensity estimates in that estimates were lower before a crisis
and increased during the crisis. This is an indicator of higher propagation of shocks
to the system. This finding supports Hypothesis 1, which associates high network
intensity to periods of stress, which could be due to increased interconnectedness
resulting in fragility of the entire financial system.
In general, higher network intensity estimates in Fig. 9a and b coincide with past
major events in the financial sector that include:
• the dot-com bubble in 2002

• the second war in Iraq in 2003
• the GFC between 2007 and 2009
• the EDC in May 2010
• the rapid fall of prices of gold in early 2013
• Chinese stock market turbulence in 2015.
These results imply that network intensity tends to increase during times of stress,
which could be associated with an increase in interconnectedness in the financial
system (Geraci & Gnabo, 2018). Additionally, the results indicate financial conta-
gion inthe financial system (Hung, 2021). These findings are supported by Blasques
et al. (2016), who related high network intensity to increased spillover in the finan-
cial system. These findings are similar to Cao et al. (2017), who reported that cross-
border linkages tend to increase during crisis periods. Conversely, larger network
intensity—especially during times of stress—are associated with increased cross-
border lending, which results in transmission of stronger shocks between markets.
11
See Appendix A.4 for more details on the patterns of the network intensity parameter.
13
Previous research by Tonzer (2015) and Sun and Chan-Lau (2017) supports this rea-
soning. They found that foreign exposures play a significant role in spreading risk.
This implies that countries are exposed more to more risks due to large exposure
from trading partners. The high network intensity across markets signifies the strong
exposure of a shock to the entire financial system (Forbes & Rigobon, 2002).
To check Hypothesis 2, we use an alternative sparse matrix to estimate network
intensity parameter. The sparse matrix provides us with an idea of how the degree
of interconnection in a financial system affects estimation of network intensity
parameter. We use a randomly generated sparse network matrix with the exception
of main diagonal taking zeros. Figure 9c shows the dynamic estimates of the net-
work intensity parameter obtained using sparse matrix with 95% confidence inter-
vals. From this figure, we observe that network intensity estimates shift downwards
when sparse weighting matrix is used. The dynamics of network intensity in Fig. 9c
differ from those in Fig. 9a. The results also indicate that interconnectedness among
financial markets changes the patterns of network intensity over time. A more stable
economy with higher network connectivity would be beneficial in shock absorption,
leading to a more resilient financial system. The converse is also true. Further, these
results imply that as the degree of connectivity increases (decreases), then network
intensity parameter shifts upwards (downwards). As a result, the degree of intercon-
nection plays a key role in the estimation of network intensity parameters. These
findings are consistent with previous studies. For instance, Silva et al. (2016) found
that shocks spread from highly interconnected networks, leading to financial distress
in the entire financial system. This is also supported by Amini et al. (2016), who
showed that financial market with larger connections is associated with spreading
shocks in the network, leading to financial instability. As Minoiu and Reyes (2013)
described the financial network as volatile, we expect this to have an impact on the
estimation of network intensity parameters.
To obtain a clearer picture of the evolution of network intensity, we obtain the
values of the estimates at each phase. Figure 10 displays the dynamic network
intensity coefficients at each phase. From these, we draw a similar conclusion those
drawn from the static estimates. On average, these estimates are lower in Phase
1 than in other phases, and increase in Phase 2. The network intensity parameter
remains higher in Phase 3 (after the crisis and during the EDC). This suggests that
network intensity tends to be higher when the market is under stress. Blasques et al.
(2016), who used European CDS data, arrived at a similar conclusion.
6.1 Impact of exogenous factors on the estimation
Next, we investigate the marginal effects of the explanatory variables on the esti-
mation of network intensity (Eq. 28). The beta coefficient of the model represents
the exogenous exposure to the common factors, while the network intensity param-
eter captures the endogenous effect of the network exposure in the model. All these
external regressors are country-specific. They include volatility index, FX and IR.
Volatility index captures the change in risk appetite, which gauges the overall mar-
ket sentiment. It is measured using the implied volatility of the world index. We
13
Fig. 10 Network intensity estimates in different phases. The plots display the network effect for the
whole sample period in each phase. The light areas are the 95% confident intervals while the horizontal
line is the is the average estimate in the entire period. The period covered in the sample is 1 January
1999–31 December 2017
considered implied volatilities of two major stock indices, VIX and VSTOXX,
because of the unavailability of individual country data.
Figure 11 shows the trend of the implied volatility of these stock indices. We use
these two major implied volatility indices to investigate the impact on network inten-
sity estimation. It can be observed in Fig. 11 that the implied volatility depicts simi-
lar patterns. For example, during the GFC of 2007–2009, the two indices reached
their peak over the whole period. We also observe a comparable shift in the mag-
nitude of volatility at different points in time. High spikes in the implied volatility
were associated with periods when financial markets were under stress. For instance,
the spikes in 2002 are associated with the dot-com bubble, those in 2003 are associ-
ated with war in Iraq (Degiannakis et al., 2018), 2007–2009 spikes were associated
with the GFC, 2010 spikes with the EDC, 2013 with the rapid fall of prices of gold
and 2015 with turbulence in the Chinese stock market.
The country-specific regressors include IR and FX. Local market returns measure
the growth of the economy of any country; this measures the stability of a country’s
economic outlook. IR affects the cost of borrowing; higher IRs are associated with
13
Fig. 11 Volatility estimates. The figures display the implied volatilities for VIX and VSTOXX indices
during the entire period. The period covered in the sample is 1 January 1999–31 December 2017
increases in borrowing costs. In addition, IR measures financial integration because

it reflects capital movement between countries (Asgharian et al., 2013). We use the
absolute changes of 90 days’ T-Bills as a proxy for IRs. FX involves trading curren-
cies across the global market. This may affect international trade and capital flows,
thereby affecting the economy. Research has found that sudden change in exchange
rates have implications for the entire financial system (Flood & Garber, 1984; Krug-
man, 1979; Salant & Henderson, 1978). This fluctuation in exchange rate is associ-
ated with currency crisis (Frankel and Rose, 1996). Following the approach used in
Asgharian et al. (2013), we use exchange rate volatility, which is computed as the
standard deviation of daily log changes in bilateral exchange rates. All these vari-
ables are indicators of financial integration and thus may either directly or indirectly
have an effect on estimating the network intensity parameter. Daily data for these
explanatory variables were obtained from Thompson Reuters’ Datastream. The sam-
ple consists of 45 countries and the period is January 1999–December 2017; we
exclude the weekends.
The estimates shown in Table 7 represent the mean value of the network inten-
sity parameter in the whole sample period. Both liability and trade weighting matrix
were used to estimate the static network intensity parameter. These results show that
the estimate is greater (0.5149) when additional regressors are excluded than when
they are included (0.1933).
13
Table 7 Comparison of network intensity estimates with and without regressors

Base IR FX VIX All regr. VSTOXX
Liability matrix
Est. Par. 0.5149 0.5117 0.5166 0.1055 0.1933 0.0978
(0.0042) (0.0042) (0.0042) (0.0027) (0.0039) (0.0025)
Trade matrix
Est. Par. 0.5698 0.5669 0.5718 0.1781 0.2000 0.1617
(0.0041) (0.0042) (0.0041) (0.0038) (0.0040) (0.0036)
The table reports the comparison in network intensity parameter with and without regressors. These esti-
mates are obtained using a dynamic SAR model and they represent the mean values of the whole sample
period. The estimates represent the average value of the estimate in the whole sample period. Base are
the estimates obtained without including external regressors, All regr. are estimates obtained with inclu-
sion of all other regressors, excluding the VSTOXX index. The other estimates are based on individual
regressors
This is an approximately 62% change (decrease) in the estimate when all addi-
tional regressors are included in the estimation. Therefore, the presence of explan-
atory variables has a discernible effect on the estimation of the network intensity
parameter. This suggests that each additional regressors may have either positive or
negative effects on the estimation. We observe that the introduction of each variable
separately has an effect on network intensity estimate.
Although there is no significant difference between the network intensity
estimates with and without inclusion of IR and FX, it is worth discussing their
impact on the estimation. IRs across countries may fluctuate due to the FX.
Higher IR fluctuations may have a greater impact on network intensity estimates.
This is because increases in IR volatility increases uncertainty, which create a
channel through which shocks can spread in financial markets (Edwards et al.,
1998). This may lead to an increase or decrease in network intensity parameter.
Our results reveal that the IR fluctuations lead to decrease in network intensity
of 0.6%. FX rates differ from country to country, and this may affect the borrow-
ing rates of each country (Bruno & Shin, 2014). A more volatile exchange rate
increases currency risk premium, and thereby effecting financial market co-move-
ments (Asgharian et al., 2013). This suggests that FX may have a greater impact
on estimation. From the static results, we observe that, on average, the volatility
in exchange rate leads to an increase in network intensity parameter of 0.3%.
Volatility (the amount of uncertainty regarding change in each stock market
index) has a greater impact on network intensity estimation. As displayed in
Fig. 11, fluctuations in the implied volatility index, especially during periods of
stress, cause shifts in network intensity. High fluctuations in volatility result, on
average, in a decrease in network intensity parameter by approximately 80% in the
case of VIX and 81% for VSTOXX. Both VIX and VSTOXX have almost simi-
lar effects on network intensity parameter. Based on these results, we conclude
that implied volatility has a major impact on the estimation of the network inten-
sity parameter and would have discernible effects on the financial system. This is
supported by Antonakakis et al. (2013), who showed that implied volatility, for
13
Fig. 12 Network intensity estimates with and without regressors. The figures display network effect for
the entire sample period with the addition of external regressors. The horizontal line is the average esti-
mate in the entire period. a Displays network intensity estimates without external regressors. b–d display
network intensity estimates with IR, VIX and FX respectively being the external regressors. e displays
network intensity estimates with all external regressors
instance VIX, dampens returns, which could result in lower network estimates.
Therefore, among the external regressors, the volatility of stock market index has
a greater impact on the estimation of the network intensity parameter, resulting in
a 62% decrease in the estimate.
Figure 12 shows dynamic network intensity parameter, including and excluding
explanatory regressors. The estimates were obtained using the dynamic SAR model
specification using a 251 rolling window size.
Figure 12a displays a time-varying trend of estimate without the external regres-
sors. The horizontal line represents the average estimate of the whole sample period.
Figure 12a–d display similar patterns, while Fig. 12c and e show varying patterns.
This is because while the IR and FX fluctuations either increased or decreased
estimates, implied volatility (we used VIX as a proxy of implied volatility since
VSTOXX provided almost identical results) of stock index had discernible effect on
the estimation of network intensity parameter. Higher volatility changes the trend
13
of network intensity estimates. These effects are clearly observed during the global
crisis, for which the trends of Fig. 12a and c of Fig. 12 differ.
Although we can observe a similar trend in Fig. 12 (which excluded the explana-
tory variables), reduction of the estimated values (previously presented in Table 7)
can be observed. The estimates fluctuate between 0.3 and 0.75. This led to the same
conclusion as previously discussed. The spikes in the estimates are associated with
periods when the market was under stress. This is in line with Hypothesis 1. For
example, the spike before 2002 is associated with the dot-com bubble, 2007–2009
with the GFC, post-2010 with the European debt crisis and 2015 with the Chinese
market crash.
We now introduce additional regressors to support Hypothesis 3. Our preliminary
findings showed that external factors (integration measures) have an effect on esti-
mation. This may suggest that the financial market is highly integrated in terms of
cross-border activities.
As stated in Hypothesis 2, increased interconnectedness between different mar-
kets results in increased network intensity estimates. Let us relate the horizontal
line (mean value) to the period of financial system stability (robust network inten-
sity estimate) while periods when there are spikes above the line correspond when
the financial system is under stress (fragile network intensity estimate). Figure 12
depicts a robust-yet-fragile network intensity estimate. A robust-yet-fragile network
would diversify small shocks, while propagating large shocks to the entire financial
system, leading to distress (Acemoglu et al., 2015; Gai & Kapadia, 2010; Tonzer,
2015). Although the network intensity was robust-yet-fragile between 2002 and 2006
(as shown in Fig. 12), the financial system benefited from risk-sharing effect. During
the GFC, shocks were amplified in the financial system, causing financial instability.
6.2 Developed versus emerging markets
There is increasing involvement of emerging markets in enhancing financial growth

and stability. Therefore, it is important to estimate the network intensity parameter
for these markets and compare them with those of developed economies. The coun-
tries in our sample were classified as developed (54%) or emerging (46%) based on
IMF 2017 classification.12
In Table 6, we observe that developed economies have higher network inten-
sity estimates than emerging markets. This could be due to their high inter-linkage
with other markets. Schiavo et al. (2010) stated that developed economies are more
interconnected to the other countries, and thus, spread shocks to other economies.
Although developed markets have higher estimates than the emerging markets,
emerging markets experience high fluctuations in different periods (see Fig. 8). We
observe that emerging markets experienced more than a 50% change in the estimate
in Phase 2. This implies that they are largely exposed to and affected by external
12
Most countries in our sample are developed economies.
13
Fig. 13 Network intensity estimates for developed and emerging markets
shocks, originating from developed economies, especially in times of GFC (Aizen-

man et al., 2016a).
Figure 13 displays the dynamics of network intensity parameters for both devel-
oped and emerging markets. Figure 13 suggests that the dynamics of estimates for
both economies differ. While the estimates of the developed markets are identical
to those of the entire sample, estimates of emerging markets have different patterns.
Emerging markets exhibit higher fluctuations than developed markets. This is in line
with Hypothesis 4. From these results, it can be theorized that developed markets
with stable economies tend to experience high network exposure, but less fluctua-
tion. Conversely, network exposure for emerging economies varies more. Harvey
(1995) suggested that this could be due to segmentation from global markets.
Our results concur with those of Aizenman et al. (2016a)—that developed mar-
kets have greater influence with higher network intensity than do emerging markets.
This greater influence can be associated to developed economies being the key con-
tributors to the GFC. That is, when a shock hit these economies through their net-
work links, it had a greater effect on the entire economy. This finding is supported
by Schiavo et al. (2010), who showed that one of the contributors of the GFC was
shocks from developed markets spreading to other markets. Kubelec and Sá (2012)
suggested that shocks from the US and UK (being big players in developed mar-
kets) propagated to the entire financial system during the global crisis. We identified
developed economies as having many interconnections with other markets, making
them more prone to risk in terms of spreading shocks to other markets. This is sup-
ported by Amini et al. (2016), who argued that institutions with large connections
have a higher chance of affecting the stability of the entire system due to their link
structures. Aizenman et al. (2016b) also argued that emerging markets were more
13
Fig. 14 Ratio of the regional

representation of the financial
markets
resilient during and after the GFC. Thus, developed markets played a significant
role in propagating shocks to the entire financial system. This is a clear indication
that policy makers should be more concerned when network intensity estimates are
greater for developed markets.
Although emerging markets have lower network intensity estimates, they may
also have the greatest influence in propagating shocks. We observed that emerging
markets had having greater spikes especially when the market was under stress. This
finding shows that emerging markets are not immune during the GFC. This suggests
that emerging markets serve as hubs through which shocks from developed econo-
mies spread to the entire financial system. Aizenman et al. (2016a) also found that
emerging markets were also exposed to external shocks during the GFC, especially
to shocks originating in developed markets. These findings indicate that emerging
markets increasingly play a role in the world economy by engaging in cross-border
relationships with developed and emerging economies (Bekaert & Harvey, 2017).
In terms of integration, we observe that developed economies with high network
intensity have more cross-border activities, making the estimate higher than those
of emerging markets. These results suggest that developed economies that are more
stable benefit from higher network intensity. However, in the presence of shocks,
these economies might have a great impact on financial stability. These findings are
similar to those of Chevallier et al. (2018), who showed that developed markets play
a dominant role in propagating shocks to the entire system while emerging markets
are becoming more integrated with other markets. This means they can transmit
shocks to other economies. These results are also supported by Schiavo et al. (2010),
who found that developed economies tend to be more integrated and more clustered,
resulting in larger estimates of network intensity parameters.
13
Finally, developed economies network intensity estimates exhibit a robust-yet-

fragile feature. The financial system could benefit more from risk-sharing and diver-
sification when small shocks hit the system. There is also a danger of large shocks
being amplified throughout in the entire system, making the financial system more
susceptible to collapse.
Generally, the findings suggest that developed markets are dominant in terms of
high propagation of shocks to other markets compared to emerging markets (Arnold
et al., 2013). This could be a result of large cross-border transactions to other mar-
kets. Although emerging markets are less dominant, they still contribute to global
propagation of shocks.
6.3 Region specific network intensity
We investigate the evolution of network intensity parameters in different regions.

From Fig. 14, all America (include both North and Latin America) represent 13%
of the total sample, Asia represents 35% and Europe represents 48%. The liability
weighting matrix for each region was obtained using the combined DY and Granger
causality approach. These analyses aid in investigating the extent to which network
intensity parameter differs among these regions. This is in line with Hypothesis 3.
Hypothesis 3 aims to investigate whether regional integration has an impact on the
estimation of network intensity parameters.
Network intensity estimates in Table 6 differ for each region. On average, Amer-
ica has the lowest (0.3497) estimate with low standard error than all other regions.
The estimate is 31% lower than the original average estimate of 0.5072. The net-
work intensity estimate of Asia is also lower (0.3515) than the original estimate,
representing a 30% decrease. Europe has the highest estimate (0.5281) in the whole
period, representing a 4% increase in estimate as compared from the original esti-
mate. From Fig. 8, these estimates changed (increased or decreased) from one phase
to another. For instance, these estimates increased when moving from Phase 1 to
Phase 2 by more than 24% in all regions. Asia experienced a significant increase in
estimates (40% change). The estimates slightly changed (decreased) moving from
Phase 2 to 3. They also changed (decreased) moving from Phase 3 to 4.
With the increasing development in the financial sector and globalization,
there has been a high degree of both regional and global integration, which
may be depicted in the results (Chevallier et al., 2018). The increase of cross-
border transactions has led countries to become more interconnected, leading to
increased financial integration between these markets. An increase in integra-
tion is be associated with increased network exposure, which tends to increase
network intensity estimates (Hüser, 2015). Higher network intensity estimates in
Europe signify greater integration in the European market.
Figure 15 shows the dynamic network estimates for each region. Figure 15a
presents the network intensity estimates for all regions. The horizontal line is the
average network intensity parameter in the whole sample. According to Fig. 15a,
Europe has higher estimates than other regions. The estimates for the European
13
616
13
Fig. 15 Network intensity estimates for each region

market fluctuated above the mean, while other regions, they fluctuated above and
below the mean.
All America and Europe produced different patterns, suggesting that the two
regions have different exposures. This could be due to different banking systems
across regions, making cross-border banks from large countries (mostly the US)
pose the ‘too big to fail’ problem. Propagation of shocks led to financial disloca-
tion and tensions especially in the Euro areas (Belke & Gros, 2016). Additionally,
network exposure for the US and Europe would differ because the equity returns
of these markets react differently to shocks. Previous literature has shown that the
US had problems in banking and sovereign debt, thereby establishing a diabolical
loop (Chan-Lau et al., 2015; Dufrénot & Keddad, 2014).
For all America, we observe that network intensity estimates were higher at
the beginning of the sample period and continued to decline until 2007, when the
estimate fluctuate upwards. There were spikes in mid-2008, implying that propa-
gation strength increased suddenly. The estimates fluctuated at around 0.4 before
increasing to 0.6 in 2014. The presence of Latin American countries (Argentina,
Brazil, Chile and Mexico) affected network intensity estimates. These emerging
markets exhibited higher fluctuations in return over time.
For Europe, the opposite was true. Network intensity estimates were lower at
the beginning of the sample period and then increased. Before 2012, there were
spikes (the propagation strengths are higher) in network intensity estimates. They
are associated with the onset of the EDC in 2010 (Mink & De Haan, 2013). The
estimates then fluctuated in an increasing trend in 2012. This suggests that Euro-
pean markets became more interconnected, increasing their exposures in the
financial system.
The propagation strength was higher during crisis periods and remained higher
during the sovereign debt crisis. This could indicate a larger impact of shock propa-
gation, especially when a shock hits the financial system.
Conversely, the Asian region depicted a different pattern from other regions.
There was a spike in 2002, and it remained higher until 2003 before dropping then
fluctuates again. The estimate declined at the beginning of the crisis period before
spiking in 2007. Thereafter, there is a declining trend of network intensity until hit
its lowest point in 2014. The estimate remained low after 2013. These results are
similar to the static estimates in Table 6. These results are consistent to those of Gui-
marães-Filho and Hong (2016) ,who argued that Asian markets are more exposed to
shocks from other region, thereby increasing their exposure during crises.
Overall, the results from the regional network intensity show that exposures are
high especially during crisis periods. This suggests that the fragility of the financial
system tends to increase during time of stress (Sun & Chan-Lau, 2017). This leads
to financial instability.
Our findings indicate the possibility that regional network intensity estimates
have an implication for policies that affect economic growth and stability. A high
network intensity estimate may imply higher propagation effects from shocks to
the financial system, leading to financial instability (Sun & Chan-Lau, 2017). This
aligns with Tonzer (2015), who showed that regional integration might be beneficial
13
to stable economies. Therefore, by having higher network intensity, a region might

benefit from diversification of shocks.
The estimates of network intensity using alternative weighting matrix are dis-
cussed in Appendix A.1. The comparison of MLE with other approaches is dis-
cussed in Appendix A.3.
7 Implication of empirical study
From our empirical results, we highlight why the network intensity estimate is
important, especially to the financial system. We do so in an attempt to answer the
questions posed below.
7.1 Do high network intensity estimates signify spillover in the financial system?
The changing nature of network intensity parameters raises the question of whether
high or low network intensity estimates are associated with return spillover in the
markets. Transmission of shocks across the financial system through different chan-
nels is known to cause financial distress. This transmission of shocks can be a result
of many factors, not limited to the increasing growth of cross-border activities in the
financial system.
With increasing cross-border activities over recent years, there has been a ten-
dency for increased exposures throughout the financial system. This is depicted from
our results, in which we observed that the network exposure tends to increase when
the financial system is under stress. This can signify a spillover in the financial sys-
tem. Moreover, increasing interactions between different markets imply high expo-
sure to these markets in terms of risk, thereby posing a threat to the stability of the
financial system. Network intensity is also affected by financial integration through
cross-border flows. This in itself creates a channel of increasing spillovers in the
financial system.
In general, network intensity parameters capture the strength of exposure, which
relates to spillovers. We conclude, based on our results, that network intensity
evolves over time and during important events. When network intensity is high, it
implies that spillover is increasing in the financial system. It is worth noting that
with increasing cross-border financial activities, financial institutions have become
more interconnected. This has resulted in high exposure of the financial system to
shocks. These results confirm that a high network intensity parameter is associated
with high interconnectedness in the financial system.
7.2 Does network intensity respond to different market conditions?
Ever changing market conditions have led to greater complexity in the financial sec-
tor. Recent studies have revealed increased co-movements of cross-border activities.
This means that with favorable market conditions the financial market has become
13
more integrated. A natural question to ask is whether the different market conditions
increase the chance of vulnerability in the financial sector.
For example, FX volatility may have a positive influence on network intensity
estimate. This is reflected in our results, in which we observed high network inten-
sity corresponding to periods of high volatility in FX rates.
Implied volatility used to capture overall market riskiness is expected to have a
positive influence on network intensity estimates. With increasing uncertainty in the
market, financial sectors are at a higher risk of failure. Our results show the consid-
erable impact of implied volatility on network intensity estimates. All this suggests
that with changing market conditions, there is an increased possibility of high net-
work intensity, and thus, a possibility of stress in the financial system.
8 Conclusion
This paper investigates the dynamics of the network intensity parameter that moni-
tors network exposure. To be specific, this paper produces two empirical findings on
how network exposure contribute to increasing vulnerability in the financial system.
We examine the impact of network exposure on common factors. Our findings show
that both the network intensity coefficient and interconnectedness increase exposure
to common factors.
We also extended our work to estimate a dynamic network intensity parameter
to determine whether a high network intensity is associated with period of extreme
events. Our findings suggest that a high network intensity coefficient is associated
with extreme events that are related to period of distress in the financial system. The
size of the network intensity coefficient serves as an indicator of stress events and
could be useful in monitoring the financial system, ultimately promoting financial
stability.
These findings highlight the importance of network exposure by showing the
extent to which financial systems are exposed to shocks from existing linkages.
Overall, the results suggest the changing market conditions increases the exposures
to the financial system. Caution must be taken to monitor these exposures in order to
reduce the transmission impact of these shocks.
These findings have important policy implications. Caution must be taken to
monitor these exposures in order to reduce the transmission impact of these shocks.
Improving regulations of the financial system can help reduce the adverse effects of
network exposures in the financial system.
13
13
◂ Fig. 16 Network intensity estimates based on trade and liability matrix. The figures compare the network
effect using both liability and trade connection matrices for the whole sample period
Appendix
Network intensity using alternative weighting matrix
Here, we investigate the role of weighting matrix in estimating network intensity

parameters. We consider the first difference of trade data to construct an alternative
weighting matrix. This matrix is then used to estimate network intensity coefficients.
Our data comprise of quarterly export and import data from the international mon-
etary fund (IMF), world economic outlook (WEO) database for the selected econo-
mies. Both indirect and direct trade linkages acted as channels through which shock
was transmitted in the financial system. Trade linkages represent high trade expo-
sures in the financial system. For instance, Kali and Reyes (2010) reported that a
shock is amplified in the system when financial institutions are more integrated in
terms of trade linkages. Asgharian et al. (2013) found that linkages through bilateral
trade capture dependencies between stock markets. This is due to feedback effects
among financial markets. This suggests that trade concentration is one of the impor-
tant channels through which shocks spread in the financial system.
In Fig. 16b, we present the network intensity estimates based on trade weight-
ing matrix. The horizontal red line represents the mean value (0.5698) in the whole
sample period while the shaded area is the 95% confident interval. These results
show that higher network intensity estimates were associated with periods when
financial systems were under stress. For instance, higher network intensity in 2002
corresponded with the dot-com bubbles while there was sharp decrease in the esti-
mate in 2003. The estimate then fluctuate upwards and decreased just before the
global crisis. There is a sharp increase in the estimate during the GFC. The esti-
mates remain higher after the global financial crisis. A network intensity above 65%
at the end of the sample period signifies a high risk of collapse when a shock hit the
financial system. As can be observed in Fig. 16b, network intensity increases when
a shock hit the financial system and decreased in normal periods. This supports our
conclusion that a sharp increase in network intensity signals that the financial sys-
tem is in distress. This observation would help regulators and policy makers monitor
these financial institutions.
Comparison of estimates using both liability and trade weights
Figure 16a depicts the role of the weighting matrix in estimating the network inten-
sity parameter. The horizontal line is the mean value of the estimate obtained using
both trade and liability weighting matrix in the whole sample. These results reveal
that the trade weighting matrix contributed to the upwards shift of the weighting
13
matrix from 0.5149 to 0.5698. This also implies that the mean value of the estimate
is 0.5698 using trade weight matrix and 0.5149 when the liability weighting matrix
is used.
The connection matrix play an important role in the estimation of the network
intensity parameter. From Fig. 16a, it is clear that although the weighting matrix
results in estimates with almost similar trends, their sizes differ in both cases. The
network intensity estimates based on trade weights were higher than those obtained
using the liability weighting matrix. This suggests that shocks through trade link-
ages would be more sensitive to the economy compared to those from liability
linkages. This supports the fact that the strength of trade linkages increase due to
bilateral trade among different markets. The patterns of network intensity estimate
also differed at different points in time. For example, the network intensity obtained
using trade weights increased before the dot-com bubbles, while network intensity
estimated using liability data decreased.
Comparison of MLE with other approaches
Although we used the dynamic MLE approach to estimate the network intensity
parameter, we also explored a state-space approach. A state-space model describes
the dynamics of a latent state and how the data relate to this state. A general SAR
model can be represented in a state-space form with observation and state evolution
respectively as:
Observation equation:
yt = C𝜌t Wyt + DXt + vt , vt ∼ N(0, Vt ) (44)
State equation:
pt = c1 + Apt−1 + BXt−1 + wt−1 , wt−1 ∼ N(0, Wt−1 ) and t = 1, 2, … , T
(45)
where At , Bt and Dt are the input variables and Ct is the state loading matrix. vt and
wt are measurement and state space process errors respectively, Xt is the exogenous
variables. The observation equation can be expressed in matrix form as
[ ] [ ][ ] [ ]
Xt 0 0 Xt 0
yt
=
Dt Ct 𝜌t Wyt
+
vt (46)
while the state equation can be represented as

[ ] [ ][ ] [ ]
Xt 0 0 wt−1 0
pt
=
Bt−1 At−1 pt−1
+
wt−1 (47)
Figure 17 shows the network intensity estimates obtained using a Kalman filter.
These estimates support our previous findings in Fig. 9a, which showed that the
network intensity remained higher during periods of stress. This explains the high
exposure of financial markets during difficult times. Caution is required to correctly
13
Fig. 17 Network intensity estimates using Kalman filter
monitor markets during periods of stress, which correspond with increased fragility
in the financial system (Sun & Chan-Lau, 2017).
Dynamics of network intensity parameter
We simulate 4956 daily data (the estimate choice is based on the number of sample
size in our analysis) using different specifications of the spatial coefficient. These
patterns are similar to Brownlees and Engle (2016) and follow:
• constant: 𝜌t = 0.5
• sine: 𝜌t == 0.5 + 0.4 ∗ cos(2 ∗ pi ∗ k∕200)
• fast sine: 𝜌t = 0.5 + 0.4 ∗ cos(2 ∗ pi ∗ k∕20)
• step: 𝜌t = 0.9 − 0.5 ∗ (k > 500)
All these specifications give different patterns of the network intensity parameter.
Using these different specifications will help to investigate whether the initial val-
ues of the network intensity parameter matter in the estimation. Figure 18 shows
these forms of network intensity parameters. The results show that network intensity
parameters have different forms of changes. The constant shows a constant trend,
Sine shows exhibit gradual change, and the fast sine has rapid change while step
changes in different steps.
13
Fig. 18 Simulated network intensity dynamics
13
Funding Open Access funding enabled and organized by CAUL and its Member Institutions. The authors
have not disclosed any funding.
Declarations
Conflict of interest On behalf of all authors, the corresponding author states that there is no conflict of
interest.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License,
which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as
you give appropriate credit to the original author(s) and the source, provide a link to the Creative Com-
mons licence, and indicate if changes were made. The images or other third party material in this article
are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is
not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission
directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licen
ses/by/4.0/.
References
Acemoglu, D., Ozdaglar, A., & Tahbaz-Salehi, A. (2015). Systemic risk and stability in financial net-
works. American Economic Review, 105(2), 564–608.
Affinito, M., & Pozzolo, A. F. (2017). The interbank network across the global financial crisis: Evidence
from Italy. Journal of Banking & Finance, 80, 90–107.
Aizenman, J., Jinjarak, Y., Lee, M., & Park, D. (2016). Developing countries’ financial vulnerability
to the Eurozone crisis: An event study of equity and bond markets. Journal of Economic Policy
Reform, 19(1), 1–19.
Aizenman, J., Jinjarak, Y., & Park, D. (2016). Fundamentals and sovereign risk of emerging markets.
Pacific Economic Review, 21(2), 151–177.
Allen, F., & Gale, D. (2000). Financial contagion. Journal of Political Economy, 108(1), 1–33.
Alter, A., & Beyer, A. (2014). The dynamics of spillover effects during the European sovereign debt tur-
moil. Journal of Banking & Finance, 42, 134–153.
Amini, H., Cont, R., & Minca, A. (2016). Resilience to contagion in financial networks. Mathematical
Finance, 26(2), 329–365.
Anselin, L. (1988). Spatial econometrics: Methods and models. Boston: Kluwer Academic Publishers.
Antonakakis, N., Chatziantoniou, I., & Filis, G. (2013). Dynamic co-movements of stock market returns,
implied volatility and policy uncertainty. Economics Letters, 120(1), 87–92.
Apostolakis, G., & Papadopoulos, A. P. (2014). Financial stress spillovers in advanced economies. Jour-
nal of International Financial Markets, Institutions and Money, 32, 128–149.
Arnold, M., Stahlberg, S., & Wied, D. (2013). Modeling different kinds of spatial dependence in stock
returns. Empirical Economics, 44(2), 761–774.
Asgharian, H., Hess, W., & Liu, L. (2013). A spatial analysis of international stock market linkages. Jour-
nal of Banking & Finance, 37(12), 4738–4754.
Bali, T. G., & Cakici, N. (2008). Idiosyncratic volatility and the cross section of expected returns. Journal
of Financial and Quantitative Analysis, 43(1), 29–58.
Bali, T. G., & Cakici, N. (2010). World market risk, country-specific risk and expected returns in interna-
tional stock markets. Journal of Banking & Finance, 34(6), 1152–1165.
Battiston, S., & Caldarelli, G. (2013). Systemic risk in financial networks. Journal of Financial Manage-
ment, Markets and Institutions, 1(2), 129–154.
Battiston, S., Gatti, D. D., Gallegati, M., Greenwald, B., & Stiglitz, J. E. (2012). Liaisons dangereuses:
Increasing connectivity, risk sharing and systemic risk. Journal of Economic Dynamics and Con-
trol, 36(8), 1121–1141.
13
Bekaert, G., & Harvey, C. R. (2017). Emerging equity markets in a globalising world. Available at SSRN
2344817.
Belke, A., & Gros, D. (2016). On the shock-absorbing properties of a banking union: Europe compared
with the United States. Comparative Economic Studies, 58(3), 359–386.
Billio, M., Caporin, M., Panzica, R., & Pelizzon, L. (2015). Network connectivity and systematic risk.
Working Paper.
Billio, M., Getmansky, M., Lo, A. W., & Pelizzon, L. (2010). Measuring systemic risk in the finance and
insurance sectors. MIT Sloan School Working Paper.
Billio, M., Getmansky, M., Lo, A. W., & Pelizzon, L. (2012). Econometric measures of connectedness
and systemic risk in the finance and insurance sectors. Journal of Financial Economics, 104(3),
535–559.
Black, L., Correa, R., Huang, X., & Zhou, H. (2016). The systemic risk of European banks during the
financial and sovereign debt crises. Journal of Banking & Finance, 63, 107–125.
Blasques, F., Koopman, S. J., Lucas, A., & Schaumburg, J. (2016). Spillover dynamics for systemic risk
measurement using spatial financial time series models. Journal of Econometrics, 195(2), 211–223.
Brownlees, C., & Engle, R. F. (2016). SRISK: A conditional capital shortfall measure of systemic risk.
Review of Financial Studies, 30(1), 48–79.
Bruno, V., & Shin, H. S. (2014). Cross-border banking and global liquidity. The Review of Economic
Studies, 82(2), 535–564.
Cai, J., Saunders, A., & Steffen, S. (2014). Syndication, interconnectedness and systemic risk. ESMT
European School of Management and Technology.
Campbell, J. Y., Lettau, M., Malkiel, B. G., & Yexiao, X. (2001). Have individual stocks become more
volatile? An empirical exploration of idiosyncratic risk. The Journal of Finance, 56(1), 1–43.
Cao, Y., Gregory-Smith, I., & Montagnoli, A. (2017). Transmission of liquidity shocks: Evidence on
cross-border bank ownership linkages. Journal of International Financial Markets, Institutions and
Money, 53, 158–178.
Catania, L., & Billé, A. G. (2017). Dynamic spatial autoregressive models with autoregressive and heter-
oskedastic disturbances. Journal of Applied Econometrics, 32(6), 1178–1196.
Chan-Lau, J. A., Liu, E. X., & Schmittmann, J. M. (2015). Equity returns in the banking sector in the
wake of the great recession and the European sovereign debt crisis. Journal of Financial Stability,
16, 164–172.
Chevallier, J., Nguyen, D. K., Siverskog, J., & Uddin, G. S. (2018). Market integration and financial link-
ages among stock markets in Pacific Basin countries. Journal of Empirical Finance, 46, 77–92.
Chowdhury, B., Dungey, M., Kangogo, M., Abu Sayeed, M., & Volkov, V. (2019). The changing network
of financial market linkages: The Asian experience. International Review of Financial Analysis, 64,
71–92.
Cohen-Cole, E., Patacchini, E., & Zenou, Y. (2012). Systemic risk and network formation in the inter-
bank market. CAREFIN Research Paper No. 25/2010.
Degiannakis, S., Filis, G., & Hassani, H. (2018). Forecasting global stock market implied volatility indi-
ces. Journal of Empirical Finance, 46, 111–129.
Diebold, F. X., & Yilmaz, K. (2013). Measuring the dynamics of global business cycle connectedness.
PIER Working Paper.
Diebold, F. X., & Yilmaz, K. (2009). Measuring financial asset return and volatility spillovers, with
application to global equity markets. The Economic Journal, 119(534), 158–171.
Diebold, F. X., & Yilmaz, K. (2012). Better to give than to receive: Predictive directional measurement of
volatility spillovers. International Journal of Forecasting, 28(1), 57–66.
Dufrénot, G., & Keddad, B. (2014). Spillover effects of the 2008 global financial crisis on the volatility
of the Indian equity markets: Coupling or uncoupling? A study on sector-based data. International
Review of Financial Analysis, 33, 17–32.
Dungey, M., Luciani, M., & Veredas, D. (2012). Ranking systemically important financial institutions.
Tinbergen Institute Discussion Paper.
Dungey, M., Milunovich, G., Thorp, S., & Yang, M. (2015). Endogenous crisis dating and contagion
using smooth transition structural GARCH. Journal of Banking & Finance, 58, 71–79.
Dungey, M., & Renault, E. (2018). Identifying contagion. Journal of Applied Econometrics, 33(2),
227–250.
Eder, A., & Keiler, S. (2015). CDS spreads and contagion amongst systemically important financial insti-
tutions: A spatial econometric approach. International Journal of Finance & Economics, 20(4),
291–309.
13
Edwards, S., et al. (1998). Interest rate volatility, contagion and convergence: An empirical investigation
of the cases of Argentina, Chile and Mexico. Journal of Applied Economics, 1(1), 55–86.
Elhorst, J. P. (2010). Applied spatial econometrics: Raising the bar. Spatial Economic Analysis, 5(1),
9–28.
Engle, R. (2002). Dynamic conditional correlation: A simple class of multivariate generalised autore-
gressive conditional heteroskedasticity models. Journal of Business & Economic Statistics, 20(3),
339–350.
Fernandez, V. (2011). Spatial linkages in international financial markets. Quantitative Finance, 11(2),
237–245.
Flood, R. P., & Garber, P. M. (1984). Collapsing exchange-rate regimes: Some linear examples. Journal
of International Economics, 17(1–2), 1–13.
Forbes, K. J., & Rigobon, R. (2002). No contagion, only interdependence: Measuring stock market
comovements. The Journal of Finance, 57(5), 2223–2261.
Frankel, J. A., & Rose, A. K. (1996). Currency crashes in emerging markets: An empirical treatment.
Journal of international Economics, 41(3–4), 351–366.
Gai, P., & Kapadia, S. (2010). Contagion in financial networks. Proceedings of the Royal Society A:
Mathematical, Physical and Engineering Sciences, 466(2120), 2401–2423.
Geraci, M. V., & Gnabo, J.-Y. (2018). Measuring interconnectedness between financial institutions with
Bayesian time-varying vector autoregressions. Journal of Financial and Quantitative Analysis,
53(3), 1371–1390.
Giudici, P., Leach, T., & Pagnottoni, P. (2021). Libra or librae? Basket based stablecoins to mitigate for-
eign exchange volatility spillovers. Finance Research Letters, 44, 102054.
Giudici, P., & Pagnottoni, P. (2019). High frequency price change spillovers in bitcoin markets. Risks,
7(4), 111.
Giudici, P., & Pagnottoni, P. (2020). Vector error correction models to measure connectedness of bitcoin
exchange markets. Applied Stochastic Models in Business and Industry, 36(1), 95–109.
Glasserman, P., & Young, H. P. (2015). How likely is contagion in financial networks? Journal of Bank-
ing & Finance, 50, 383–399.
Glasserman, P., & Young, H. P. (2016). Contagion in financial networks. Journal of Economic Literature,
54(3), 779–831.
Gofman, M. (2017). Efficiency and stability of a financial architecture with too-interconnected-to-fail
institutions. Journal of Financial Economics, 124(1), 113–146.
González-Páramo, J. M. (2010). Globalisation, international financial integration and the financial cri-
sis-the future of european and international financial market regulation and supervision. Discurso
en el Institute of International and European Affairs.
Guimarães-Filho, R., & Hong, G. H. (2016). Dynamic connectedness of Asian equity markets. IMF
Working Paper WP/16/57.
Haldane, A. (2009). Rethinking the financial network. Speech delivered at the Financial Students Asso-
ciation, 28 April 2009, Amsterdam.
Harvey, C. R. (1995). Predictable risk and returns in emerging markets. The Review of Financial Studies,
8(3), 773–816.
Hautsch, N., Schaumburg, J., & Schienle, M. (2014). Financial network systemic risk contributions.
Review of Finance, 19(2), 685–738.
Hueng, C. J., & Yau, R. (2013). Country-specific idiosyncratic risk and global equity index returns. Inter-
national Review of Economics & Finance, 25, 326–337.
Hung, N. T. (2021). Financial connectedness of gcc emerging stock markets. Eurasian Economic Review,
11(4), 753–773.
Hüser, A.-C. (2015). Too interconnected to fail: A survey of the interbank networks literature. SAFE
Working Paper No. 91.
Kali, R., & Reyes, J. (2010). Financial contagion on the international trade network. Economic Inquiry,
48(4), 1072–1101.
Kleimeier, S., Lehnert, T., & Verschoor, W. F. C. (2008). Measuring financial contagion using time-
aligned data: The importance of the speed of transmission of shocks. Oxford Bulletin of Economics
and Statistics, 70(4), 493–508.
Krugman, P. (1979). A model of balance-of-payments crises. Journal of Money, Credit and Banking,
11(3), 311–325.
Kubelec, C., & Sá, F. (2012). The geographical composition of national external balance sheets: 1980–
2005. International Journal of Central Banking, 8(2), 143–189.
13
Lee, L.-F. (2002). Consistency and efficiency of least squares estimation for mixed regressive, spatial
autoregressive models. Econometric Theory, 18(2), 252–277.
Lee, L. (2007). GMM and 2 SLS estimation of mixed regressive, spatial autoregressive models. Journal
of Econometrics, 137(2), 489–514.
LeSage, J., & Pace, R. K. (2009). Introduction to spatial econometrics. Chapman and Hall/CRC.
Lintner, J. (1965). The valuation of in stock portfolios and capital budgets. The Review of Economics and
Statistics, 47(1), 13–37.
Markose, S., Giansante, S., & Shaghaghi, A. R. (2012). Too interconnected to fail’ financial network of
US CDS market: Topological fragility and systemic risk. Journal of Economic Behavior & Organi-
zation, 83(3), 627–646.
Mink, M. (2015). Measuring stock market contagion: Local or common currency returns? Emerging
Markets Review, 22, 18–24.
Mink, M., & De Haan, J. (2013). Contagion during the Greek sovereign debt crisis. Journal of Interna-
tional Money and Finance, 34, 102–113.
Minoiu, C., & Sharma, S. (2014). Financial networks key to understanding systemic risk. IMF Survey
Magazine.
Minoiu, C., Kang, C., Subrahmanian, V. S., & Berea, A. (2015). Does financial connectedness predict
crises? Quantitative Finance, 15(4), 607–624.
Minoiu, C., & Reyes, J. A. (2013). A network analysis of global banking: 1978 − 2010. Journal of Finan-
cial Stability, 9(2), 168–184.
Narayan, P. K., Narayan, S., & Prabheesh, K. P. (2014). Stock returns, mutual fund flows and spillover
shocks. Pacific-Basin Finance Journal, 29, 146–162.
Peltonen, T. A., Rancan, M., & Sarlin, P. (2019). Interconnectedness of the banking sector as a vulner-
ability to crises. International Journal of Finance & Economics, 24(2), 963–990.
Peralta, G., & Zareei, A. (2016). A network approach to portfolio selection. Journal of Empirical
Finance, 38, 157–180.
Resta, M., Pagnottoni, P., & Giuli, M. E. D. (2020). Technical analysis on the bitcoin market: Trading
opportunities or investors’ pitfall? Risks, 8(2), 44.
Salant, S. W., & Henderson, D. W. (1978). Market anticipations of government policies and the price of
gold. Journal of Political Economy, 86(4), 627–648.
Schiavo, S., Reyes, J., & Fagiolo, G. (2010). International trade and financial integration: A weighted
network analysis. Quantitative Finance, 10(4), 389–399.
Sharpe, W. F. (1964). Capital asset prices: A theory of market equilibrium under conditions of risk. The
Journal of Finance, 19(3), 425–442.
Silva, T. C., de Souza, S. R. S., & Tabak, B. M. (2016). Network structure analysis of the Brazilian inter-
bank market. Emerging Markets Review, 26, 130–152.
Spelta, A. (2017). Financial market predictability with tensor decomposition and links forecast. Applied
Network Science, 2(1), 7.
Spelta, A., Flori, A., Pecora, N., & Pammolli, F. (2021). Financial crises: Uncovering self-organized pat-
terns and predicting stock markets instability. Journal of Business Research, 129, 736–756.
Spelta, A., Pecora, N., & Pagnottoni, P. (2022). Chaos based portfolio selection: A nonlinear dynamics
approach. Expert Systems with Applications, 188, 116055.
Sun, A. J., & Chan-Lau, J. A. (2017). Financial networks and interconnectedness in an advanced emerg-
ing market economy. Quantitative Finance, 17(12), 1833–1858.
Tonzer, L. (2015). Cross-border interbank networks, banking risk and contagion. Journal of Financial
Stability, 18, 19–32.
Vitali, S., Battiston, S., & Gallegati, M. (2016). Financial fragility and distress propagation in a network
of regions. Journal of Economic Dynamics and Control, 62, 56–75.
Wang, G.-J., Xie, C., & Stanley, H. E. (2018). Correlation structure and evolution of world stock markets:
Evidence from Pearson and partial correlation-based networks. Computational Economics, 51(3),
607–635.
Yarovaya, L., Brzeszczyński, J., & Lau, C. K. M. (2016). Intra and inter-regional return and volatility
spillovers across emerging and developed markets: Evidence from stock indices and stock index
futures. International Review of Financial Analysis, 43, 96–114.
Yellen, J. L. (2013). Interconnectedness and systemic risk: Lessons from the financial crisis and policy
implications: A speech at the American economic association/American finance association joint
luncheon, San Diego. American Finance Association.
13
Yu, H., Fang, L., Sun, B., & Du, D. (2017). Risk contribution of the Chinese stock market to developed
markets in the post-crisis period. Emerging Markets Review, 34, 87–97.
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published
maps and institutional affiliations.
13

Dynamic Effects of Network Exposure On Equity Markets: Mardi Dungey Moses Kangogo Vladimir Volkov

Uploaded by

Copyright:

Available Formats

Dynamic Effects of Network Exposure On Equity Markets: Mardi Dungey Moses Kangogo Vladimir Volkov

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Dynamic Effects of Network Exposure On Equity Markets: Mardi Dungey Moses Kangogo Vladimir Volkov

Uploaded by

Copyright:

Available Formats

Eurasian Economic Review (2022) 12:569–629

Dynamic effects of network exposure on equity markets

Mardi Dungey1 · Moses Kangogo1 · Vladimir Volkov1

Received: 20 September 2021 / Revised: 16 January 2022 / Accepted: 1 February 2022 /

Keywords Financial markets · Financial networks · Financial stability

JEL Classification G15 · G10 · G01 · C21

variance currency basket aimed at assessing contagion spillovers among foreign

measure interconnectedness. These measures show that an increase in links in a

2.1 Spatial econometric concept

Introduction of spatial econometric techniques into financial application can be use-

⎛ 0 0.25 0.25 0.25 0.25⎞

Estimates of network intensity are dependent on the interaction of the endoge-

2.2 Motivation and hypotheses development

Financial integration is a process through which either financial markets, countries

Hypothesis 2: Degree of connectivity affects estimation of network intensity. This

3 Financial network and exposure to common factors

We now consider the impact of exposure to common factors in financial networks.

S(rit − 𝔼 [rit ]) = 𝛽M ritM + 𝛽HML ritHML + 𝛽SMB ritSMB + 𝜂it (3)

i. structural exposure to common factors

4 Dataset and effect of network exposure

TH Thailand BANGKOK S.E.T. BNGKSET

Table 2 Descriptive statistics of daily return for each market

AT 0.0002 − 0.0974 0.1277 0.0135 10.7384 − 0.2140 − 71.1573** 5739

Table 3 Phases of the sample

All phases 01.01.1999–29.12.2017 Entire period 4956

Table 4 Structural betas for countries in our sample

Austria 0.9090 0.8274 0.9694 1.0169 0.9924

Figure 1 displays the average cross-border liability flow, measured in US billion

4.1 Effect of financial network to common factor model

This subsection investigates whether network exposure affects common factors.

4.1.1 Estimating beta and idiosyncratic volatilities

To estimate idiosyncratic volatility for an individual stock in our sample, we

4.1.2 Effect of financial network on betas

It is important to understand whether using a different weighting matrix has a dif-

4.1.3 Network exposure to common factors

To investigate the effect of network exposure on common factors, we used slope

Table 5 Estimates of Country All phases Phase 1 Phase 2 Phase 3 Phase 4

The table reports the idiosyncratic volatilities for all countries in

causality and DY approach by using the cross-border liabilities. Based on simple

5 Optimal value of the network intensity parameter

5.1 Static network intensity parameter

We introduce the spatial autoregressive (SAR) model to measure network intensity

p̂ = (y� W � Wy)−1 y� W � y (16)

where X = (X1 , X2 , … , Xn ) a vector (n × k , and where k is the number of variables)

• Estimate of 𝛽 by differentiating Eq. (29) with respect to 𝛽 . We obtain:

From Eq. (30), the estimate of 𝛽 is:

𝛽̂ = (X � X)−1 X � (In − 𝜌W)y (31)

For simplicity, this can be written as:

𝛽̂ = (X � X)−1 X � y − 𝜌(X � X)−1 X � Wy (32)

(y − 𝜌Wy − X𝛽)� (y − 𝜌Wy − X𝛽)

y = X 𝛿̂0 + ê 0 and Wy = X 𝛿̂d + ê d (35)

𝛽̂ = (X � X)−1 X � y − 𝜌(X � X)−1 X � Wy

5.2 Dynamic network intensity parameter

where Σ is constant over time.

Allowing for time-varying variance in the shocks led to the following:

The generalized log-likelihood function of the time-varying ( Lvt ) variance models

Table 6 Network intensity estimates and their robust standard errors

to 2. It only increased by 16% in Phase 3 which is approximately half the increase

optimization problem). To proxy the initial changing values of 𝜌, we assume the

2.1 Spatial econometric concept

2.2 Motivation and hypotheses development

3 Financial network and exposure to common factors

4 Dataset and effect of network exposure

4.1 Effect of financial network to common factor model

4.1.1 Estimating beta and idiosyncratic volatilities

4.1.2 Effect of financial network on betas

4.1.3 Network exposure to common factors

5 Optimal value of the network intensity parameter

5.1 Static network intensity parameter

5.2 Dynamic network intensity parameter

6.1 Impact of exogenous factors on the estimation

6.2 Developed versus emerging markets

6.3 Region specific network intensity

7 Implication of empirical study

7.2 Does network intensity respond to different market conditions?