BIS Working Papers

No 796
Measuring contagion
risk in international
banking
by Stefan Avdjiev, Paolo Giudici and Alessandro Spelta

Monetary and Economic Department

July 2019

JEL classification: G01, C58, C63

Keywords: International banking, Contagion risk,

Multi-layer networks, Tensor decompositions
BIS Working Papers are written by members of the Monetary and Economic
Department of the Bank for International Settlements, and from time to time by other
economists, and are published by the Bank. The papers are on subjects of topical
interest and are technical in character. The views expressed in them are those of their
authors and not necessarily the views of the BIS.

This publication is available on the BIS website (www.bis.org).

© Bank for International Settlements 2019. All rights reserved. Brief excerpts may be
reproduced or translated provided the source is stated.

ISSN 1020-0959 (print)

ISSN 1682-7678 (online)
 Measuring contagion risk in international banking∗

S. Avdjiev†

Bank for International Settlements

‡§
P. Giudici
University of Pavia

¶
A. Spelta
University of Pavia

Center for Analysis, Decisions, and Society (CADS) - Human Technopole

Abstract

We propose a distress measure for national banking systems that incorporates not only banks’ CDS

spreads, but also how they interact with the rest of the global financial system via multiple linkage types.

The measure is based on a tensor decomposition method that extracts an adjacency matrix from a multi-

layer network, measured using banks’ foreign exposures obtained from the BIS international banking

statistics. Based on this adjacency matrix, we develop a new network centrality measure that can be

interpreted in terms of a banking system’s credit risk or funding risk.

JEL: G01, C58, C63

Keywords: International banking, Contagion risk, Multi-layer networks, Tensor decompositions

∗ The views expressed in this paper are those of the authors and do not necessarily reflect the views of the Bank for

International Settlements. We thank two anonymous journal referees, Inaki Aldasoro, Neil Esho, Leonardo Gambacorta, Hyun
Shin, Goetz von Peter, Philip Wooldridge for helpful comments and suggestions and as well as Pamela Pogliani, and Swapan
Pradhan, and Zuzana Filkova for excellent research support. A major part of the project was completed while Paolo Giudici
participated in the Research Fellowship programme of the Bank for International Settlements.
† Monetary and Economics Department E-mail: [email protected]
‡ Corresponding Author
§ Department of Economics and Management. E-mail: [email protected].
¶ Department of Economics and ManagementE-mail: [email protected]

1
1 Introduction

As a consequence of the very high degree of integration in the global financial system, financial stress that

originates in one country can quickly spill across borders into other countries. The Great Financial Crisis of

2007-09 and the subsequent European Sovereign Debt Crisis both illustrated that in a very stark manner.

The rapid growth of the global financial system over the past several decades has increased the importance

of properly measuring the risk of distress. This is true not only from a financial stability point of view, but

also from a macroeconomic perspective, as financial crises tend to have significant and persistent negative

effects on economic activity (see Acemoglu et al. (2012)). At the same time, the increased interconnectedness

and complexity of the global banking system have made that task extremely challenging.

The existing literature has employed three main approaches to measuring systemic risk in the global

financial system (Cerutti et al. (2012)). The first method uses information on (direct and indirect) aggregate

country-level bilateral exposures (Peek and Rosengren (2000), Kaminsky and Reinhart (2003), McGuire and

Tarashev (2008), Cetorelli and Goldberg (2012)). The second approach takes advantage of market data

on Credit Default Swap (CDS) spreads, bond spreads and equity prices in order to quantify systemic risk

(Huang et al. (2009), Acharya et al. (2010), Delatte et al. (2012)).1 The third approach utilizes simulation

techniques to assess the probability of various potential contagion paths (both within and across sectors),

conditional on an initial set of exposures (Árvai et al. (2009)).

In principle, the market price-based approach should be complementary to the bilateral exposure-based

approach (Cerutti et al. (2012)). Nevertheless, there are not many existing studies that have used the two

approaches in a complementary fashion, mainly due to the practical difficulties associated with matching

market-based data with data on bilateral exposures. Faced with this data constraint, researchers have usually

resorted to the “maximum entropy” assumption, under which exposures are assigned as uniformly as possible

across counterparties. Nevertheless, this assumption has multiple important shortcomings, especially in the

global context (Upper (2011) and Drehmann and Tarashev (2011)).

In this paper, we propose a methodology that merges the market price-based approach and the exposure-

based approach. Namely, the methodology that we utilize enhances the information contained in market

prices (i.e. CDS spreads) with the information contained in bilateral exposures (i.e. banks’ foreign claims,

broken down by counterparty country and sector). We assume that the global matrix of banks’ foreign

claims can be used to map the links in the global financial network (see also Tintchev (2016), De Souza et al.

1 A credit default swap is an agreement in which the seller commits to repay an obligation (e.g. a bond) underlying the

contract at par in the event of a default; to produce this guarantee, a regular premium is paid by the buyer (the creditor of
the reference loan), during a specified period. The premium is expressed in the form a “CDS spread”, a measure which is a
function of the expected loss for a unitary loan amount. For a given maturity, the expected loss is a function of the debtor’s
probability of default and loss given default. As a consequence, if an entity (e.g. a sovereign or a bank) is considered more
likely to default by the market, a higher protection fee (CDS spread) is charged.

2
(2016)).

In theory, banks’ CDS spreads should incorporate all available relevant information, including information

about banks’ (direct and indirect) exposures. In practice, however, this is not necessarily the case. It is

true that the majority of banks’ immediate (direct) exposures tend to be known to market participants (and

therefore incorporated in banks’ CDS spreads). For example, during the euro area crisis, it was well-known

that French banks had large exposures to borrowers in the euro area periphery countries (Greece, Ireland,

Italy, Portugal and Spain). Nevertheless, the available information about banks’ indirect (higher-order)

exposures was much less complete. For example, although US banks did not have large direct (first-order)

exposures to the euro area periphery countries, they were indirectly exposed to them through their exposures

to French banks. In turn, a number of other banking systems which were directly exposed neither to the

euro area, nor to French banks, were indirectly exposed to them through their exposures to US banks. It

is easy to see how market participants, who may have been able to quantify the direct exposures and even

the first-order of indirect exposures reasonably well, may have had very incomplete information about the

higher-order (indirect) exposures in the global network.

This motivates us to construct a measure that takes into account the higher-order exposures of banks.

To achieve this goal, a good predictor of a banking crisis should be based not only on the past CDS spreads

of a given banking system, which may vary in accordance with idiosyncratic factors, but also on how its

fluctuations are affected by contagion from other systems, in a multidimensional way. When constructing

our measure, we aim to achieve an improved prediction of CDS spreads and, ultimately, of crisis events.

As foreign exposures can be of different types, we represent them as a multi-layer financial network, which

can describe multidimensional interlinkages between economic agents (see Poledna et al. (2015), Montagna

and Kok (2016), Aldasoro and Alves (2016)). The modelling of multi-layer financial networks has been

recently employed in the domain of interbank lending. The papers in the existing literature are mostly

descriptive, rather than predictive, and aim to create an aggregate network representation by summing over

single networks. This approach may fail to adequately capture important non-linearities generated by the

multi-layer nature of the data (Gauvin and Panisson (2014)).

In this paper, we extend the above approach to international banking lending to multiple sectors (the

banking sector, the non-banks private sector and the official sector), using the rich dimensionality of the BIS

international banking statistics. More concretely, we consider a set of 23 countries. Financial relationships

between them can be represented by a multi-layer network, which contains the foreign claims of each country

vis-a-vis the other countries in three types of layers, which correspond to foreign claims on (i) the banking

sector, (ii) the official sector and (ii) the non-bank private financial sector.

We employ a tensor decomposition method that extracts an adjacency matrix from this multi-layer net-

3
work. In addition, we also develop a new network centrality measure that can be interpreted probabilistically

in terms of the credit risk or the funding risk (depending on the side of the balance sheet that is being an-

alyzed) of a given banking system, taking multiple type of foreign exposure into account. The multi-layer

methodology that we propose summarizes the relevant information in a non-linear, rather than linear, way.

As we demonstrate in the results section below, this non-linear approach has a superior predictive perfor-

mance than alternative linear measures constructed using unadjusted CDS spreads.

The rest of this paper is organized as follows. Our methodology is described in the next section. Section

3 contains a description of the data. Section 4 describes the key empirical findings. Section 5 concludes,

while offering several policy implications.

2 Methodology

The multi-layer network can be mapped into a tensor X ∈ RI×I×K where I represents countries and K the

type of exposures. Each element of the tensor, xkij represents the amount of funds claimed by country i

from country j, in type k, normalized by the total amount of foreign claims. Doing so, xkij represent the

percentage share of a particular claim. The tensor is composed by three slices X ∈R23×23 , each of which

represents the financial transactions between countries in each type of exposure.

2.1 Probabilistic Matrix Decomposition

Our aim is to extract the relevant features from a multi-layer network of financial relationships. To exemplify,

we first show how this can be done in the simpler mono layer network setting.

We aim to summarise a network matrix, that contains a specific type k of financial claims between

countries, with two scores, that represent, respectively, hub score and authority score as in Kleinberg (1999),

associated to each banking system (a country, for brevity).

The economic intuition is that a country i that lends to many other countries with high authority score

receives a high hub score, ui . A country j that borrows from many other countries with a high hub score

receives a high authority score, vj .

We propose a model in which hub and authority scores are obtained from the transition probabilities of a

Markov chain between countries, based on the claim matrix. This has the advantage of a better interpretation

of the results, as probabilities are normalised by definition.

The network claim matrix Xk of a particular type k can be used to calculate conditional frequencies, as

4
follows:
xkij
hki|j = PI i = 1, ..., I
k
i=1 xij

xkij
akj|i = PI j = 1, ..., I,
k
j=1 xij

Without loss of generality, when xkij is equal to 0 for all 1 ≤ i, j ≤ I the values of hki|j (akj|i ) can be set to

1/I.

The previous conditional frequencies can be used to estimate the transition probabilities of a Markov

chain:

Pr Xτk = i|Yτk = j
 

Pr Yτk = j|Xτk = i ,
 

where Xτk and Yτk are random variables referring to the visit at any particular country as a hub or as an

authority at the Markov chain step τ . In other words, hki|j while estimates the probability that country i

borrows from to country j, akj|i estimates the probability that country j lends to i.

The conditional frequencies hki|j and akj|i can also be used to estimate the stationary distribution of

countries being hubs or authorities (borrowers or lenders). For a given τ step of the chain let:

 PJ
Pr Xτk = i = j=1 hki|j Pr Yτk = j
  

 PI
Pr Yτk = j = i=1 akj|i Pr Xτk = i
  

The stationary probabilities of each country being a hub or an authority are then:

uki = lim Pr Xτk = i

 
τ →∞

vjk = lim Pr [Yτ = j]

τ →∞

According to the previous results, to practically compute uki and vjk we need to solve the following system

of equations:

PI
uki = k k
j=1 hi|j vj i = 1, ..., I (1)

5
 PI
vjk = k k
i=1 aj|i ui j = 1, ..., I (2)

The system can be solved iterating between equations 1 and 2 until u(τ ) − u(τ −1) + v(τ ) − v(τ −1) < .

From an economic viewpoint, the hub score uki of country i is the weighted sum of the authority scores

vjk of the countries that i lends to. The weight associated with each borrower is the element of the transition

probability matrix hki|j between i and the borrower j. The authority score vjk of a country j is the weighted

sum of the hub scores uki of the countries that borrow from j. The weight associated with each lender is the

element of the transition probability matrix akj|i .

We are now ready to present how the claim matrix can be decomposed. As authority and hub scores are

stationary probability distributions, they can be employed to calculate the (joint) probability that an extra

unit claim is placed from country i to country j. Let mkij to represent the probability that, in sector k, the

banking system of country i has an extra unit of claim on country j and let Mk be the stochastic matrix of

all such probabilities, whose elements sum up to 1. Let uk and vk be, respectively, the hub and authority

scores. From Markov chain theory it follows that:

Mk = uk (vk )T .

and, therefore, mkij is the product between the hub score of i and the authority score of j.

Note that, when temporal data is available, such probability can be calculated at any given time point.

2.2 Probabilistic Tensor Decomposition

We now consider the more general multi-layer network, in which we aim to summarise a network tensor, that

contains financial claims between countries, of different types, with three scores that represent, respectively,

hub, authority scores associated to each country and a type score associated to each layer.

To achieve this aim, following Ng et al. (2011), we consider a Markov chain on the claim tensor, whose

joint stationary distribution will be the product of hub, authority and type scores.

While authority and hub scores are the same as those previously defined, type scores are specific to the

multi-layer framework. From an interpretational viewpoint, a financial relation type that connects countries

with high hub score with countries with high authority score receives a high type score, wk .

For a multi-layer network, the starting point of the construction of the decomposition is the computation

of the (bivariate) conditional frequencies H, A and R for hubs, authorities and types. They can be obtained

by normalizing the entries of the tensor X as follows:

6
 xijk
hi|jk = PI i = 1, ..., I
i=1 xijk

xijk
aj|ik = PI j = 1, ..., I
j=1 xijk

xijk
rk|ij = PJ k = 1, ..., K
k=1 xijk

As done before, when xijk is equal to 0 for all 1 ≤ i, j ≤ I the values of hi|jk (aj|ik ) are set to 1/I (1/K

for rk|ij ).

The above quantities can be used to estimate the conditional probabilities:

Pr [Xτ = i|Yτ = j, Zτ = k]

Pr [Yτ = j|Xτ = i, Zτ = k]

Pr [Zτ = k|Yτ = j, Xτ = i]

where Xτ , Yτ and Zτ are random variables referring to the visit at any particular country as a hub or as an

authority, and using any particular financial relation type, at the step τ of the Markov chain.

Note that hi|jk can now be interpreted as the probability of visiting the i-th country as a hub (or as an

authority) given that the j-th (or i-th) country as an authority (or as a hub) is currently visited and that

the k-th relation is used.

Similarly, aj|ik is the probability of visiting the j-th country as an authority given that the i-th country

as a hub is currently visited and that the k-th relation is used.

In addition, rk|ij can be interpreted as the probability of using the k-th relation given that the j-th

country as an authority is visited from the i-th country as a hub.

In analogy with the matrix decomposition setting, the conditional frequencies can also be used to derive

the stationary marginal probabilities, first computing:

PJ PK
Pr [Xτ = i] = j=1 k=1 hi|jk Pr [Yτ = j, Zτ = k]

PI PK
Pr [Yτ = j] = i=1 j=1 aj|ik Pr [Xτ = i, Zτ = k]

PI PI
Pr [Zτ = k] = i=1 j=1 rk|ij Pr [Yτ = j, Xτ = i] .

7
 and, then, taking their limiting distributions:

ui = lim Pr [Xτ = i]
τ →∞

vj = lim Pr [Yτ = j]
τ →∞

wk = lim Pr [Zτ = k] ,
τ →∞

which can be used as hubs, authority and type scores.

From a practical viewpoint, exploiting the result seen in the last subsection on the decomposition of

stationary bivariate probabilities into the product of univariate ones, the scores can be derived solving the

following system of equations:

PI PJ
ui = j=1 k=1 hi|jk vj wk i = 1, ..., I (3)

PI PK
vj = i=1 j=1 aj|ik ui wk j = 1, ..., I (4)

PI PI
wk = i=1 j=1 rk|ij ui vj k = 1, ..., K (5)

From an economic viewpoint, the hub score ui of country i is the weighted sum of the authority scores

vj of the borrowing countries that i points to. The weight associated with each borrower is the product of

the element of the transition probability tensor H between i and the borrower j times the type score wk of

the layer in which the transaction is performed.

The authority score vj of country j is the weighted sum of the hub scores ui of the lending countries that

point to j. The weight associated with each lender is the product of the element of the transition probability

tensor A times the type score wk of the layer in which the transaction is executed. Finally, the type score

of layer k is the sum, over all pairs of countries (i, j) involved in a transaction in layer k, of the product

between the hub score ui of lender i with the authority score vj of borrower j and with the element of the

transition probability tensor R between i and j.

The above system can be solved in a recursive way, with an algorithm that takes as inputs the three

tensors H, A and R, the two initial probability vectors u, v and the tolerance parameter . The algorithm

iterates through equations 3, 4 and 5 until u(τ ) − u(τ −1) + v(τ ) − v(τ −1) + w(τ ) − w(τ −1) < .

8
 Ng et al. (2011) demonstrated the existence and the uniqueness of a limiting probability distribution for

the above algorithm, which can be interpreted as the stationary distribution of the Markov chain on the

multi-layer network tensor.

As before, we can now focus on tensor decomposition. Let mij be the probability that the lending banking

system i has an extra unit of claim on country j, in any of the type sectors, and let M be the stochastic

matrix of all such probabilities.

From Markov chain theory it follows that:

M = u(v)T ,

and, therefore, as in the previous subsection, the probability mij can be obtained as the product between

the hub score of i and the authority score of j. However, differently from the single-layer case, M is not

sector specific but, rather, calculated integrating information from all relationship types.

Figure 1 below helps to further understand the meaning of hub, authority and type scores, the key

elements of the multi-layer network tensor decomposition.

FIGURE 1 ABOUT HERE

2.3 Contagion measures

Many summary measures of network models have been proposed in the literature (see e.g. Kramer et al.

(2009)). Among them we refer to the Alpha-centrality (see e.g. Bonacich and Lloyd (2001), Ide et al. (2013)),

which can be converted into a probability measure. The alpha-centrality measures the total number of paths

from a node, exponentially attenuated by their length. It can be interpreted as a cumulative probability of

an infection spreading through the network.

More formally, let c be a vector of positive quantities, let A be an adjacency matrix, and let β be a non

negative parameter. The alpha-centrality vector S is the steady state solution of the recursive equation:

Sµ = C+ β A Sµ−1 ,

with µ indicating the number of iterations. An equivalent representation can be given in cumulative

terms, as follows:

Sµ = I+βA+β 2 A2 + ... + β µ Aµ C,

9
 where the (i, j) element of the powered matrix Aµ represents the number of paths that, from node i, can

reach node j in µ steps.

Note that, to achieve convergence, β k must be equal to:

α
β=
λ1 (A)

where λ1 the largest eigenvalue of M and α can be interpreted as an “average” contagion probability

(0 ≤ α ≤ 1.).

To derive the Alpha-centrality measure in our context, we need some preliminary assumptions.

The adjacency matrix A can be set equal to M, the tensor decomposition that contains the probabilities

that an extra claim is placed from i to j. In this way, the adjacency matrix describes “exposure” to contagion.

Concerning the rate of contagion α, we can assume that:

αo = wo , αb = wb , αp = wp

which implies that, the higher the importance of a type of financial relationship, the higher the probability

that it can be used to channel funds between countries, the higher a contagion could come from that sector.

Note that, consequently:

αk
βk =
λ1 (M)

Further assumptions concerns the channels of contagion.

As in Tintchev (2016) we assume that contagion can come both from the lending side (credit risk) and

from the borrowing side (funding risk). In our context, cik is the CDS spread of country i in type k, and ck
PK
is the vector of all spreads of type k. We then take c = k1 ∗ k=1 ck .

In addition, the transmission of contagion to a banking sector has the possibility to reverberate into the

system through higher order connections (as discussed in Battiston et al. (2012) and Hale et al. (2016)).

Contagion transmission to the official and the non bank private sectors, instead, does not have this feature,

and stops at the first order iteration.

Following the stated assumption, we obtain a multivariate spread, respectively on the borrowing (SB)

and on the lending side (SL), as:

P∆
SB= c + cb µ=1 β
b,µ
Mµ + cb β o M + cb β p M

10
 P∆
µ
SL= c + cb µ=1 β
b,µ
MT + co β o MT + cp β p MT ,

where b, o, p indicate the bank, official and non bank sectors and where ∆ indicates the network diameter

(the longest of the shortest path between any two nodes).

From an interpretational viewpoint, both solutions, SB and SL, can be interpreted as network-modified,

or multivariate, CDS spreads. They are based on a diffusive non conservative process that distributes

through the network as an infection, or, in other words, as a random walk which replicates at each step on

the network.

It is important to note that the notion of contagion differs from that of interdependence. The seminal work

of Forbes and Rigobon (2002) highlights the difference between the two concepts by examining cross-country

stock market co-movements. In particular, they empirically identify cases of interdependencies between

economic quantities without contagion. For instance, during the 1997 East Asian crisis, the 1994 Mexican

peso collapse, and the 1987 U.S. stock market crash, there was no contagion, but only interdependence. In

our case, while co-movements in national banking systems’ exposures capture the notion of interdependence,

co-movements in CDS spreads capture the notion of contagion. Indeed, our measure of centrality assesses

to what extent interdependence between banking sectors exposure drives contagion by letting CDS spreads

interact in a financial network environment. Thus, besides being able to identify systemically important

national banking systems, our approach can quantify the potential capital losses arising from contagion

effects. In addition, it allows for an intuitive visualization of both interdependence and contagion effects.

Interpreting network spreads as modified CDS spreads, we can also compute the expected losses of each

country. Let E be the total amount of foreign claims of the system. The expected credit loss for a country

i, on the lending side, can be obtained as:

X
ELL
i =E MTij SLi . (6)
j

The expected funding loss for a country i, on the borrowing side, can be obtained as:

X
ELB
i =E Mij SBi (7)
j

The above expected losses can be calculated in any given time point, and provide a very relevant in-

dicator measure of the overall risk of each banking system, both in terms of funding risk and credit risk.

The technique can also be easily adapted to a stress test scenario, making alternative assumptions on the

11
distribution of banking claims between countries.

We finally remark that, if we work separately on each sector using the probabilistic matrix decomposition

approach, we can separately consider each sector’s exposure matrix. We can then calculate multivariate

spreads in a similar fashion, however with three different measures, on the borrowing and the lending side.

In formulae:

SBo = Co + Co β o Mo SLo = Co + Co β o Mo,T

b P∆ P∆
µ
SBb = Cb + C µ=1 β b,µ Mb,µ SLb = Cb + Cb µ=1 β b,µ Mb,T

SBp = Cp + Cp β p Mp SLp = Cp + Cp β p Mp,T

The resulting disaggregated expected loss can be found using equations (6-7) using each sector’s total

exposure.

2.4 Predictive analysis

Any proposed systemic risk measure should be validated in a predictive performance setting. This is the

aim of this subsection.

Specifically, we investigate whether the inclusion of contagion helps in predicting the next period CDS

spreads. We compare the predictions obtained using only the information on the past CDS spreads values

with those obtained using both past CDS spreads and past multi-layer values, merged in the multivariate

CDS spread derived in the previous subsection.

In more detail, when we rely only on past CDS values we use the average of the last n1 values to forecast

the CDS spread in next period as:

Pt
\=− x=t−n1 CDSx
CDS .
n1

When we also employ multi-layer network information, we embed the stochastic matrices Mt , obtained

in different time periods, in a three dimensional tensor M ∈ RI×I×T where I represents countries, T the

time periods and accordingly, the generic element mijt represents the probability of having a connection

from the borrowing country i to the lending country j at time t.

We then follow Spelta (2017) in predicting the next period slice of M which, together with the average

of the last n1 values of the CDS spreads, are used to forecast the CDS. More formally, we employ a tensor

decomposition technique called CP decomposition (see e.g. Harshman (1970) and Carroll and Chang (1970)).

The decomposition aims at writing the tensor M as the outer product of two spatial vectors a and b, that

contain the importance of each country in terms of borrowing and lending probabilities, and one temporal

vector g, containing the temporal profile of each slice Mt .

12
 The vectors a, b and g can be found solving the following minimization problem:

2
min kM − a ◦ b ◦ gk
a>0,b>0,g>0

This is equivalent to minimizing the difference between each of the modes M(i) and their respective approx-

imation in terms of factors:

2
min M(1) − a (b g) F
a>0

2
min M(2) − b (a g) F
(8)
b>0

2
min M(3) − g (a b) F
g>0

Since financial flows are always non negative, a non-negative tensor factorization method can be employed

to solve (8). Specifically, the Block Coordinate Descent Method for Regularized Multiconvex Optimization

(Xu and Yin (2013)) and the Matlab Tensor Toolbox (Bader and Kolda (2004)) can be used to solve (8).

We let the reader to refer to the Appendix for a more complete description of tensor calculus.

The next step consists in generating the adjacency matrix of the predicted network. A simple average,

applied to the last n1 observations of the temporal profile vector g, extracts a scalar ĝ representing the

forecast of the next period value of such vector. The adjacency matrix containing the predicted probability

of a link between all country pairs can then be obtained as a linear combination of the two spatial vectors a

and b and of the forecast ĝ of the temporal profile vector. In matrix terms: M̂ = ĝ ∗ a ◦ b or, elementwise

m̂ij = ĝai bj .

The predicted spreads are finally found as:

d Ĉ + Ĉb β M̂ + Ĉb P∆ β µ M̂µ + Ĉp β M̂,

SB= µ=1

for the borrowing side, and:

µ
c Ĉ + Ĉo β M̂T + Ĉb P∆ β µ M̂T

SL= µ=1 + Ĉp β M̂T ,
Pt k
1 x=t−n1 Cx
for the lending side. Note that, for simplicity, we set β = 3 and Ĉk = n1 where k = {o, b, p}.

13
3 Data

We obtain data on banks’ foreign claims using the BIS Consolidated Banking Statistics (CBS). The BIS CBS

measure banks’ country risk exposures. They capture the worldwide consolidated claims of internationally

active banks headquartered in BIS reporting countries. The consolidated statistics include the claims of

banks’ foreign affiliates, but exclude intragroup positions, in accordance with the consolidation approach

followed by banking supervisors.

There are two main BIS CBS datasets. The CBS on an immediate counterparty basis (CBS/IC) provide

information on banks’ direct counterparties without accounting for any associated credit risk transfers. By

contrast, the CBS on an ultimate risk basis (CBS/UR) account for credit risk transfers (BIS (2015a)).

In our empirical exercise, we use the CBS/UR in order to construct a quarterly panel of 23 bank national-

ities (home countries) and 23 borrowing countries for the period Q1 2005 to Q4 20152 . These 23 nationalities

include the major internationally active banks, and account for more than 95% of the aggregate stock of

global foreign claims at end-Q4 2014 (see BIS (2015b)). The main variables we focus on are bilateral foreign

claims, broken down by reporting national banking systems and borrowing countries. Foreign claims (FC)

are defined as the sum of (i) cross-border claims (XBC) and (ii) local claims (LC), booked by banks’ affiliates

located in the borrowing country (i.e. FC = XBC + LC).

Furthermore, we define three distinct layers of foreign exposures by taking advantage of the fact that

the CBS/UR contain breakdowns not only by lending banking systems and borrowing countries, but also

by borrowing sectors. More concretely, foreign claims in the CBS/UR dataset can be divided into claims

on the official sector, claims on other (non-related) banks and claims on the non-bank private sector. This

counterparty sector breakdown allows us to examine the financial relationships between lending banks and

borrowers from different countries and sectors in the context of the multi-layer network framework outlined

in Section 2.2. More specifically, each layer in the network we examine is defined by the foreign exposures of

a given national banking system to one of the three counterparty sectors (listed above) in a given borrowing

country and corresponds to a weighted adjacency matrix. All networks have 23 nodes, corresponding to the

full set of reporting national banking systems.

We construct three sets of CDS spreads series – one for each of the three counterparty sectors in the BIS

CBS/UR foreign claims dataset. First, for the public sector of each borrowing country, we use the respective

five-year sovereign on-the-run CDS spread, obtained from Markit. Second, for the banking sector of each

2 The 23 national banking systems for which we obtain data from the BIS CBS are: ’AU - Australia’ ’AT - Austria’
’BE - Belgium’ ’CA - Canada’ ’CL - Chile’ ’TW - Chinese Taipei’ ’FR - France’ ’FI - Finland’ ’DE -
Germany’ ’GR - Greece’ ’IN - India’ ’IE - Ireland’ ’IT - Italy’ ’JP - Japan’ ’NL - Netherlands’ ’PT
- Portugal’ ’SG - Singapore’ ’ES - Spain’ ’SE - Sweden’ ’CH - Switzerland’ ’TR - Turkey’ ’GB - United
Kingdom’ ’US - United States’.

14
borrowing country, we construct an equally-weighted average of senior five-year CDS spreads (obtained from

Markit) for a sample of domestic financial institutions (following a methodology similar to the one used

in Avdjiev et al. (2012)). Finally, for the non bank private sector of each borrowing country, we use the

option-adjusted spread (OAS) between the yield on a US dollar-denominated non bank financial bond index

for the respective country and the corresponding yield on US-Treasuries of a similar maturity.3

FIGURE 2 ABOUT HERE

Some summary network statistics are reported in Figure 2. The Network density (fraction of existing

links over all possible links) is reported in the left panel. It is quite high (relative to its counterparts in usual

financial networks) and increasing over time. Across the different borrowing sectors, the sovereign sector is

the most sparse one, whereas the bank and the non bank private sector layers have approximately the same

number of links. The percentage of reciprocated links (displayed in the right panel) exhibits similar patterns

both, over time and across sectors.

4 Empirical findings

4.1 Tensor decomposition

As discussed above, the proposed methodology distinguishes between Systemically Important Lender coun-

tries (hubs) and Systemically Important Borrower countries (authorities).

FIGURE 3 ABOUT HERE

Figure 3 presents the hub and the authority scores for each countries’ banking sector obtained from the

proposed tensor decomposition approach together with the type score associated to each network layer. The

figure reveals that the United States and the United Kingdom are clearly the main Systemically Important

Borrower (left panel). Meanwhile, the Systemically Important Lenders distribution (central panel) is much

more evenly spread across bank nationalities, with Germany, Japan and France joining the United States

and the United Kingdom as the main Systemically Important Lenders. Thus, the latter two countries have

a dual role - they are systemically important both as hubs and as authorities.

The right panel of Figure 3 also reveals that the non bank private sector layer contributes more than the

other two sectors to the systemic importance of countries. The relative importance of the other two sectors

3 We obtain the spreads the non bank private sector from two sources. Spreads for Australia, Austria, Belgium, Canada,

France, Germany, Italy, Japan, Netherlands, Spain, Sweden, Switzerland, United Kingdom and the United States are obtained
using the Barclays Global Aggregate Corporate Index. Spreads for Chile, Chinese Taipei, India, Singapore and Turkey are
obtained using JPMorgan’s Corporate Emerging Markets Bond Index (CEMBI).

15
has been reversed after the global financial crisis: whereas the bank sector was more systemically important

than the official sector before the crisis, the opposite became true during the post-crisis period.

4.2 Contagion measures

We now derive the network modified CDS spreads separately for each country. Figure 4 plots, in the upper

panels, the network based PD on the borrowing side.

FIGURE 4 ABOUT HERE

For the sake of clarity, we present the results only for two groups of countries: the largest five national

banking systems (the United States, the United Kingdom, Japan, Germany, France) in the upper right panel

and the euro area periphery countries (Greece, Portugal, Italy, Spain, Ireland) in the upper left panel.

Figure 4 suggests that, for most countries, there are non-negligible contagion effects, as approximated by

the difference between the two measures (reported in the bottom panels). For example, Greece, Portugal,

Italy and, to a lesser extent, Spain and Ireland have considerable differences. On the other hand, countries

such as the United States, the United Kingdom, Japan and France have a low univariate PD (reported in

the central right panel), but, when contagion is taken into account (bottom right panel), their overall PD

reaches levels similar to those of countries such as Italy and Portugal.

The evolution of the contagion measure is consistent with the observed timing of crisis events. For

example, the contagion measure for Greece spikes up in 2010 and remains at elevated levels until the second

half of 2012 (after Draghi’s “whatever it takes speech” in July 2012). It spikes again in late 2014/early 2015,

on the eve of the January 2015 parliamentary elections and June 2015 referendum in Greece. The CDSs for

Italy and Portugal exhibit similar, slightly delayed, dynamics, in line with the timing of the occurrence of

their respective fiscal shocks.

The prominent role that Greece played in the euro area crisis is well-known and extensively documented.

The BIS CBS allow us to quantify the foreign exposures of various national banking systems to residents

of Greece and track their evolution over time. As Table 1 clearly illustrates, foreign bank’s exposures to

Greece, which were very high on the eve of the crisis ($310 billion as of end-Q2 2008), declined very rapidly

and steadily in the next few years (reaching $22 billion as of end-Q4 2015). The table further illustrates that

all major global banking systems (US, UK, Japan, France, and Germany), as well as all euro area periphery

banking systems (Italy, Spain, Ireland and Portugal) drastically reduced their exposures to Greece.4

TABLE 1 ABOUT HERE

4 We would like to thank an anonymous referee for the suggestion to include a table that tracks the evolution of foreign
banks’ claims on Greece.

16
 The multivariate CDS of Ireland and Spain are also fairly high. Nevertheless their dynamics are somewhat

different from those of the CDS of Greece, Italy and Portugal. This is most likely related to the initial

contagion trigger along the bank-sovereign nexus for the two groups of countries. In the case of the former

group, the original shock hit the banking system and was later transmitted to the sovereigns (which were

initially in good shape); conversely, for the second group, the initial shock came from the sovereign and was

later transmitted to the banking system (whose balance sheet was relatively healthy prior to that).

Comparing the time dynamics of low- and high-spread countries, it is notable that the multivariate CDS

for the former peak at the time of the global financial crisis, whereas the multivariate CDS of the latter spike

during the Euro area sovereign debt crisis. This is in line with the conventional wisdom that the largest

banking systems in the world were more severely impacted by the global financial crisis, whereas the banking

systems of the Euro area periphery were hit harder during the European sovereign crisis.

FIGURE 5 ABOUT HERE

Figure 5 reports the results on the lending side. It reveals that the overall dynamics of the multivariate

CDS on the lending side for the top five banking systems are roughly similar to those on the borrowing side.

Once again, there are two peaks in the contagion measure for this set of countries: one around the global

financial crisis and one around the European debt crisis, with the former being higher.

Nevertheless, there are important differences among individual countries. For example, the United States

has the highest contagion effect on the borrowing side. By contrast, on the lending side, the leading role

is taken by United Kingdom, France and Germany, followed by the United States and Japan. The above

differences between the borrowing and the lending side of the top five systems can be explained by their

different roles. The United States is by far the largest borrowing country in the world. Meanwhile, even

though it is also a relatively large lender, its share of global lending is not as large as those of United

Kingdom, France and Germany. When it comes to the euro area periphery countries, the dynamic on the

lending side is quite similar to that of the borrowing side. One notable exception is Spain, for which the

contagion measure tends to be higher on the lending side. This is likely due to the fact that the share of

Spain as a lender is higher than its share as a borrower.

For the sake of a summary comparison, Tables 2-4 present some descriptive statistics for each country

in the dataset. In particular, Table 2 refers to the original CDS values while Tables 3 and 4 refer to the

multivariate cases (the borrowing and the lending cases respectively). In both tables we report the mean

of the observed (univariate) market CDS spreads, and the mean of the estimated (multivariate) spreads,

on both the borrowing and the lending sides, along with its maximum, standard deviation, skewness and

kurtosis.

17
 TABLE 2 ABOUT HERE

TABLE 3 ABOUT HERE

TABLE 4 ABOUT HERE

From the tables note that contagion increases the univariate spreads of the smaller peripheral countries

(Greece,Portugal,Ireland) about three times, similarly on both the borrowing and the lending sides. Larger

peripherals, that is Italy and Spain, increase instead by four times, indicating a larger vulnerability to

contagion. Core countries have much smaller univariate CDS spreads, whose mean does not exceed 10 basis

points. However, contagion sensibly increases the spreads, and especially for the most exposed countries.

Note, in particular, the US, whose borrowing modified spread increases at 80 basis points versus 51 on the

lending side, indicating a much higher liquidity, rather than credit risk. Conversely, France and, to a lesser

extent, Japan and Germany, are instead more influenced on the lending side, due to their relevant credit risk

exposure.

FIGURE 6 ABOUT HERE

Figure 6 reports the expected losses on the borrowing side. It is dominated by the United States, the

largest borrower. The impact of the very large outstanding amount for the United States dominates the

impact of its relatively low CDS spread. It is worth pointing out that, in terms of total expected losses, the

peak associated with the sovereign crisis of 2011 is larger than the peak associated with the Global Financial

Crisis of 2008. This is especially true for the top five banking systems, which account for the overwhelming

majority of the expected losses during crisis period. At the same time, the ordering of peripheral European

countries is only roughly correlated with their relative sizes as borrowing destinations. This is due to the

relatively more powerful offsetting impact of the market CDS effect, which is particularly strong for Greece

and for Ireland in late 2010/early 2011. For that group of countries, the two (Global Financial Crisis and

euro area crisis) peaks are comparable.

FIGURE 7 ABOUT HERE

Figure 7 reports the expected losses results, on the lending side. Their overall distribution is more evenly

spread among countries than the respective distribution on the borrowing side. A more detailed analysis, at

the country level, reveals that the EL dynamics vary quite a bit across the euro area periphery countries,

depending on the geographical distribution of their lending exposures. Finally, it is also worth pointing out

that the peaks on the borrowing tend to be higher than those on the lending side.

FIGURE 8 ABOUT HERE

18
 Figure 8 reports the expected losses, cumulated over all the considered time, on both the borrowing and

the lending side. This allows to identify the countries that, throughout the considered period, are associated

with the highest expected losses. The largest losses are expected by the United States, distantly followed

by the United Kingdom, with the rest of the countries far behind. By contrast, on the lending side, losses

are much more evenly distributed. The greatest expected losses are observed for the two largest European

countries France and Germany, followed by Japan, the United Kingdom and the United States.

The results we discussed up to this point were all based on claims on all sectors. We now focus on a

sector-wise comparison, in terms of cumulated losses over time.

FIGURE 9 ABOUT HERE

Figure 9 compares the overall expected losses in the official sector. Note that, in the borrowing side,

the United States expected losses are much larger than those of the United Kingdom. It is notable that

those of Germany are larger than those of France and those of Italy are larger than those of Spain. On the

lending side we have the opposite relationships, for each one of the above pairs. This in line with the relative

importance of those countries as lenders and borrowers, respectively.

FIGURE 10 ABOUT HERE

Figure 10 compares the expected losses, cumulated over time, in the banking sector. It reveals the

dominant role that the United States and United Kingdom play on the borrowing side. Conversely, on the

lending side, we find a prominent role for Spain, France, Germany and Japan. In addition, the Netherlands

and Switzerland also appear relevant. These results are consistent with the relative international presence

of the respective banking systems.

FIGURE 11 ABOUT HERE

Figure 11 compares the cumulative expected losses in the non bank private sector. Once again, the United

States and the United Kingdom are the most relevant on the borrowing side. When it comes to lending, the

United Kingdom is considerably larger than the United States and France is greater than Germany. Japan

also appears quite relevant on the lending side.

Finally, we examine whether non-linear, higher order effects are present.

FIGURE 12 ABOUT HERE

Figure 12 compares the expected losses obtained with the tensor decomposition methods presented in

Section 2.2, applied to all three sectors simultaneously, with the sum of the expected losses, calculated

19
separately for each sector using the matrix decomposition of Section 2.1. The difference between the two, if

positive, indicates the presence of non-linear, higher order contagion effects that make the overall expected

losses higher than the sum of the sector-wise losses. Figure 12 reveals that this is indeed the case during the

main crisis periods not only for many individual countries (upper two panels), but also at the global level

(lower two panels).

4.3 Predictive analysis

To evaluate the predictive power of our proposed method, in Figure 13 we present the results from an out-

of-sample predictive analysis. Specifically, the aim is to predict, at time t − 1, the one-step ahead value of

the CDS spread (at time t).

FIGURE 13 ABOUT HERE

To achieve this goal we compare: i) a model based only on the average of the last n1 observed past spread

values; ii) a model based on the same observed past spreads, modified by network contagion, as described in

Section 2.4. For both models, we compare the predictive Root Mean Squared Error (RMSE), for different

n1 lengths: N = 1, . . . , 8 quarters. We also compare their binarised versions, in which we set a performance

indicator equal to 1 when the RMSE of i) is higher than that of ii) and zero otherwise.

As illustrated by Figure 13, our proposed method is clearly superior on the eve of and during crisis

periods (as indicated by its considerably lower RMSE), in line with the results obtained when examining the

non-linear effects in the preceding sub-section. Moreover, as suggested by the above figure, there appears

to be a structural break in Q1-2013. Indeed, between that quarter and Q1-2014 our measure performs

slightly worse than the alternative model (based on the average of past CDS spread values). The most likely

explanation for this is related to the behaviour of the Greeks sovereign CDS during the same period. After

being suspended in late 2011, this instrument was re-introduced into the market with a lower value (1/3

of the original 2011 value). This mechanically reduces the model-estimated level of stress. In contrast, as

previously discussed, when the second wave of the Greek crisis escalated in 2014, our measure once again

outperforms the one generated by the alternative model.

In addition to the above RMSE-based comparison, which treats predictions errors in a symmetric fashion,

we also evaluate the predictive performance of our model using an alternative, asymmetric statistical measure

of goodness of fit.5 Namely, we focus on the Mean Signed Deviation (MSD), which is defined as the average

of the predictions errors, split conditional on their signs. This measure is not symmetric and, therefore,

could be utilised to distinguish between predictions, depending on whether they are underestimating or
5 We would like to thank an anonymous referee for suggesting this asymmetric comparison.

20
overestimating future CDS spreads. The results from this alternative comparison are summarized in Figure

14.

FIGURE 14 ABOUT HERE

The results reveal that our measures remain close to the actual CDS values up until Q1-2011, while the

forecasts obtained using a model based only on past CDS spreads tends to under-predict. In the middle of

the sample, all forecast tend to under-predict CDS spreads, but the network model performs better, having

a MSD closer to zero up to Q1-2013. Subsequently, the model based only on past CDS spreads perform

better, but only because the multivariate measure over-estimates future CDS spreads. Nevertheless, towards

the end of the sample, our proposed model outperforms the simple model once again.

5 Summary and policy implications

In this paper, we propose a technique aimed at summarising multi-layer networks into a single adjacency

matrix and a resulting methodology for transforming univariate, market based, CDS spreads into multivariate

CDS that take into account contagion effects arising from changing default probabilities of other countries

and from changes in flow compositions. We apply our measure to a multi-layer international banking network

derived from the BIS international banking statistics. We show that the resulting measure predicts CDS

spreads better than an alternative measure based on (unadjusted) past values of CDS spreads. This is the

case, especially during crisis times, when the non-linear network effects tend to be more important.

We believe our methodology can be rather useful for policymakers, as it gives an early warning measure

of a national banking system’s distress levels, which incorporates information on its foreign exposures. Our

measure can also be extended to any multi-layer financial network, such as an interbank network.

Furthermore, the methodology we propose could potentially be utilised in a bottom-up stress test. More

precisely, our proposed methodology could generate estimates of an institution’s expected losses, while

incorporating all relevant information on (direct and indirect) exposures, linkages and contagion probabilities.

As noted in EBA (2017), in the context of banking supervision, a stress test should take into account that

risks at the institution-wide level may not be well reflected by simple aggregation of stress tests on portfolios,

individual risk areas or business units of the group. This opens room for possible vulnerabilities coming from

contagion effects, as discussed in Greenwood et al. (2015), in the context of a simple fire sale model for the

banking sector, based on the interbank networks. These concepts can be extended from banks to national

banking systems, with the tensor approach overcoming the problem of a simple linear aggregation of risks (or

exposures), and our proposed centrality measure becoming useful to investigate how changes in exposures

21
and/or in CDS spreads affect the soundness of national banking systems.

22
Appendix

Formally, following Kolda and Bader (2009), an N -th order tensor is an element of the tensor product of

N vector spaces, each of which has its own coordinate system. In a tensor representation, vectors (tensors

of order one) are denoted by boldface lowercase letters, e.g., x. Matrices (tensors of order two) are denoted

by boldface capital letters, e.g., X. Higher-order tensors (order three or higher) are denoted by boldface

Euler script letters, e.g., X . Scalars are denoted by lowercase letters, e.g., x. The i-th entry of a vector X is

denoted by xi ; the element (i, j) of a matrix X is denoted by xij , and the element (i, j, k) of a third-order

tensor X is denoted by xijk . The n-th element in a sequence is denoted by a superscript in parentheses, e.g.,

Xn denotes the n-th matrix in a sequence. Fibers are the higher-order analogue of matrix rows and columns,

defined by fixing every index but one. A matrix column is a mode-1 fiber and a matrix row is a mode-2

fiber. Third-order tensors have column, row, and tube fibers, denoted by x:jk , xi:k , and xij: , respectively.

Slices are two-dimensional sections of a tensor, defined by fixing all but two indices: Xi:: , X:j: , and X::k ,

denotes horizontal, lateral, and frontal slides of a third-order tensor X . Alternatively, the k-th frontal slice of

a third-order tensor, X::k , may be denoted more compactly as Xk . The norm of a tensor X ∈ RI1 ×I2 ×···×IN

is the square root of the sum of the squares of all its elements,

v
u I1 I2 IN
uX X X
k X k= t ... x2i1 ,i2 ,...iN , (9)
i1 =1 i2 =1 iN =1

in analogy with the matrix Frobenius norm. The inner product of two same-sized tensors X , Y ∈ RI1 ×I2 ×···×IN

is the sum of the products of their entries, i.e.,

I1 X
X I2 IN
X
hX , Yi = ... xi1 ,i2 ,...iN yi1 ,i2 ,...iN (10)
i1 =1 i2 =1 iN =1

An N-way tensor X ∈ RI1 ×I2 ×···×IN is of rank one if it can be written as the outer product of N vectors,

i.e.,

X =a(1) ◦ a(2) ◦ ... ◦ a(N ) , (11)

where the symbol “◦” represents the vector outer product. Matricization, also known as unfolding or flat-

tening, is the process of reordering the elements of an N-way array into a matrix. The mode-matricization

of a tensor X ∈ RI1 × I2 × ··· × IN is denoted by X(n) and arranges the mode-n fibers to be the columns of

the resulting matrix. A tensor element (i1 , i2 , ..., iN ) maps to a matrix element (in , j), where

PN Qk−1
j =1+ k=1,k6=n (ik − 1) Jk with Jk = m=1,m6=n Im (12)

23
Several matrix products are important in the following sections, so we briefly define them here. The Kronecker

product of matrices X ∈ RI×J and Y ∈ RK×L is denoted by X ⊗ Y: the result is a matrix of size (IK)×(JL).

The Khatri–Rao product is the “matching column-wise” Kronecker product. Given two matrices X ∈ RI×J

and Y ∈ RK×L , their Khatri–Rao product is denoted by X Y; the result is a matrix of size (IJ) × K.

The Hadamard product is the elementwise matrix product, and it is denoted by “∗”.

24
References

D. Acemoglu, V. M. Carvalho, A. Ozdaglar, A. Tahbaz-Salehi, 2012. The network origins of aggregate

fluctuations. Econometrica, 80(5):1977–2016.

V. Acharya, T. Cooley, M. Richardson, I. Walter, 2010. Manufacturing tail risk: A perspective on the

financial crisis of 2007–2009. Found. Trends Fin., 4(4):247–325.

I. Aldasoro, I. Alves, 2016. Multiplex interbank networks and systemic importance: An application to

european data. J. Financ. Stabil.

Z. Árvai, K. Driessen, I. Ötker, 2009. Regional financial interlinkages and financial contagion Within Europe.

International Monetary Fund.

S. Avdjiev, Z. Kuti, E. Takats, 2012. The euro area crisis and cross-border bank lending to emerging markets.

BIS Q. Rev., pages 37–47.

B. Bader, T. Kolda, 2004. Matlab tensor classes for fast algorithm prototyping. Technical report, Sandia

National Laboratories.

S. Battiston, M. Puliga, R. Kaushik, P. Tasca, G. Caldarelli, 2012. Debtrank: Too central to fail? financial

networks, the fed and systemic risk. Sci. rep., 2:srep00541.

BIS, 2015a. Guidelines for reporting the bis international banking statistics.

BIS, 2015b. Introduction to bis statistics. BIS Q. Rev., pages 37–39.

P. Bonacich, P. Lloyd, 2001. Eigenvector-like measures of centrality for asymmetric relations. Soc. net.,

23(3):191–201.

J. Carroll, J. Chang, 1970. Analysis of individual differences in multidimensional scaling via an n-way

generalization of “eckart-young” decomposition. Psychometrika, 35(3):283–319.

E. Cerutti, S. Claessens, P. McGuire, 2012. Systemic risks in global banking: What available data can tell

us and what more data are needed? Technical report, National Bureau of Economic Research.

N. Cetorelli, L. Goldberg, 2012. Follow the money: Quantifying domestic effects of foreign bank shocks in

the great recession. Am. Econ. Rev., 102(3):213–18.

S. R. De Souza, T. C. Silva, B. M. Tabak, M. Solange, 2016. Evaluating systemic risk using bank default

probabilities in financial networks. J. Econ. Dyn. Control, 66:54–75.

25
A. Delatte, M. Gex, A. López-Villavicencio, 2012. Has the cds market influenced the borrowing cost of

european countries during the sovereign crisis? J. Int. Money Financ., 31(3):481–497.

M. Drehmann, N. Tarashev, 2011. Systemic importance: some simple indicators.

EBA, 2017. Guidelines on stress testing and supervisory stress testing.

K. Forbes, R. Rigobon, 2002. No contagion, only interdependence: measuring stock market comovements.

J. Financ., 57(5):2223–2261.

L. Gauvin, a. C. C. Panisson, A., 2014. Detecting the community structure and activity patterns of temporal

networks: a non-negative tensor factorization approach. Pone, 9(1):e86028.

R. Greenwood, A. Landier, D. Thesmar, 2015. Vulnerable banks. J. Financ. Econ., 115(3):471–485.

G. Hale, T. Kapan, C. Minoiu, 2016. Crisis transmission in the global banking network. International

Monetary Fund.

R. Harshman, 1970. Foundations of the parafac procedure: models and conditions for an” explanatory”

multimodal factor analysis.

X. Huang, H. Zhou, H. Zhu, 2009. A framework for assessing the systemic risk of major financial institutions.

J. Bank. Fin., 33(11):2036–2049.

K. Ide, R. Zamami, A. Namatame, 2013. Diffusion centrality in interconnected networks. Proc. Comp. Sci.,

24:227–238.

G. Kaminsky, C. Reinhart, 2003. The center and the periphery: the globalization of financial turmoil.

Technical report, National Bureau of Economic Research.

J. M. Kleinberg, 1999. Authoritative sources in a hyperlinked environment. JACM, 46(5):604–632.

T. Kolda, B. Bader, 2009. Tensor decompositions and applications. SIAM review, 51(3):455–500.

M. Kramer, U. Eden, S. Cash, E. Kolaczyk, 2009. Network inference with confidence from multivariate time

series. Phys. Rev. E, 79(6):061916.

P. McGuire, N. Tarashev, 2008. Global monitoring with the bis international banking statistics.

M. Montagna, C. Kok, 2016. Multi-layered interbank model for assessing systemic risk.

M. K. Ng, X. Li, Y. Ye, 2011. Multirank: co-ranking for objects and relations in multi-relational data. In

Proc. ACM SIGKDD, pages 1217–1225. ACM.

26
J. Peek, E. S. Rosengren, 2000. Collateral damage: Effects of the japanese bank crisis on real activity in the

united states. Am. Econ. Rev., 90(1):30–45.

S. Poledna, J. L. Molina-Borboa, S. Martı́nez-Jaramillo, M. Van Der Leij, S. Thurner, 2015. The multi-layer

network nature of systemic risk and its implications for the costs of financial crises. J. Financ. Stabil.,

20:70–81.

A. Spelta, 2017. Financial market predictability with tensor decomposition and links forecast. Appl. Netw.

Sci., 2(1):7.

K. Tintchev, 2016. Connected to whom? international bank borrowing during the global crisis.

C. Upper, 2011. Simulation methods to assess the danger of contagion in interbank markets. J. Financ.

Stab., 7(3):111–125.

Y. Xu, W. Yin, 2013. A block coordinate descent method for regularized multiconvex optimization with

applications to nonnegative tensor factorization and completion. SIAM J., 6(3):1758–1789.

27
Tables and Figures

Table 1: Exposures to Greece and PIIGS countries: banking sector foreing claims to all-sectors Greece on quarterly base.

Mean Univar Cds Max Univar Cds Std Univar Cds Skew Univar Cds Kurt Univar Cds
GR 0.0373 0.0902 0.0738 2.1279 5.8092
PT 0.0030 0.0113 0.0031 1.3229 3.9996
IT 0.0016 0.0044 0.0013 0.7881 2.6014
ES 0.0017 0.0042 0.0014 0.7659 2.4496
IE 0.0025 0.0103 0.0027 1.3948 3.8835
US 0.0006 0.0009 0.0004 0.2854 2.2760
GB 0.0006 0.0014 0.0004 0.0317 1.8118
FR 0.0007 0.0022 0.0006 0.9970 3.4068
JP 0.0006 0.0014 0.0004 0.1537 2.5473
DE 0.0005 0.0013 0.0004 0.5605 2.9022

Table 2: Aggregate Statistics CDS spreads, univariate

Mean Mult. Borr. Cds Max Mult. Borr. Cds Std Mult. Borr. Cds Skew Mult. Borr. Cds Kurt Mult. Borr. Cds
GR 0.1120 0.2706 0.2212 2.1280 5.8093
PT 0.0091 0.0342 0.0093 1.3268 4.0103
IT 0.0062 0.0153 0.0038 0.6501 2.5821
ES 0.0062 0.0137 0.0041 0.6338 2.4424
IE 0.0080 0.0314 0.0083 1.3724 3.8503
US 0.0072 0.0112 0.0043 1.5216 5.6945
GB 0.0053 0.0097 0.0029 0.8027 3.5346
FR 0.0041 0.0089 0.0022 0.3871 2.3965
JP 0.0025 0.0045 0.0012 0.1146 2.9704
DE 0.0039 0.0071 0.0019 0.5929 3.1892

Table 3: Aggregate Statistics CDS spreads, multivariate borrowing

28
 Mean Mult. Lend. Cds Max Mult. Lend. Cds Std Mult. Lend. Cds Skew Mult. Lend. Cds Kurt Mult. Lend. Cds
GR 0.1119 0.2705 0.2213 2.1279 5.8091
PT 0.0091 0.0342 0.0093 1.3278 4.0138
IT 0.0056 0.0143 0.0037 0.7090 2.6848
ES 0.0062 0.0148 0.0045 0.8280 2.7027
IE 0.0080 0.0315 0.0083 1.3637 3.8342
US 0.0054 0.0093 0.0030 0.5002 2.4662
GB 0.0066 0.0117 0.0040 1.0253 3.8756
FR 0.0060 0.0124 0.0038 0.9761 3.2264
JP 0.0042 0.0075 0.0022 0.5665 3.0674
DE 0.0030 0.0052 0.0013 -0.2392 2.5492

Table 4: Aggregate Statistics CDS spreads, multivariate lending

C4 L1 C3
r411 r131

a411 h131
C1
a512 h122

C5 L2 C2

Figure 1: Multi-layer network example: the network has two layers L1 and L2 , and five countries. In L1 country C1 borrows
from country C4 with probability a411 , moreover C1 lends to C3 with probability h131 . In L2 country C1 borrows from country
C1 with probability a512 and lends with probability h122 to C2 . The authority score of C1 is uC1 = a411 vC4 wL1 + a512 vC5 wL2 .
The hub score of C1 is vC1 = h131 vC3 wL1 + h122 vC2 wL2 . Now consider layer L1 . The type score of L1 is: wL1 = r411 vC4 uC1 +
r131 vC1 uC3 .

Number of Links (%) Reciprocated Links (%)

0.55
Official Sector Official Sector
BankSector BankSector
0.7 Non bank Private Sector Non bank Private Sector
0.5

0.65 0.45

0.6 0.4

0.55 0.35

0.5 0.3

0.25
0.45

0.2
0.4

0.15
0.35

0.1
Q4-05 Q4-06 Q4-07 Q4-08 Q4-09 Q4-10 Q4-11 Q4-12 Q4-13 Q4-14 Q4-15 Q4-05 Q4-06 Q4-07 Q4-08 Q4-09 Q4-10 Q4-11 Q4-12 Q4-13 Q4-14 Q4-15

Figure 2: Network statistics: the left panel shows the network density (the fraction of existing links over all possible links).
The right panel displays the percentage of reciprocated links.

29
 Authority Hub Type
0.35

0.16
0.5
0.3
AU
AT 0.14
BE
0.45
CA
CL 0.25
TW 0.12 Official Sector
FR BankSector
0.4
FI
DE Non bank Private Sector
0.2 0.1
GR
IN
0.35
IE
IT 0.08
JP 0.15
NL
0.3
PT
0.06
SG
ES
0.1
SE
0.25
CH 0.04
TR
GB
US 0.05
0.02 0.2

0 0 0.15
Q4-05 Q4-07 Q1-10 Q2-12 Q3-14 Q4-05 Q4-07 Q1-10 Q2-12 Q3-14 Q4-05 Q4-07 Q1-10 Q2-12 Q3-14

Figure 3: Hub, Authority and Type Scores: the left panel shows the authority score associated to each country, the central
panel displays the hub score whereas the right panel reports the type score associated with each network layer.

Multivariate Borrowing PD Multivariate Borrowing PD

1 GR 0.2 US
PT GB
IT FR
0.5 ES 0.1 JP
IE DE
0 0
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15 Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15

Univariate PD Univariate PD
1 GR 0.03 US
PT GB
IT
0.02 FR
0.5 JP
ES
0.01
IE DE
0 0
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15 Q4−05Q4−06Q4−07Q4−08Q4−09Q4−10Q4−11Q4−12Q4−13Q4−14Q4−15

Difference GR Difference
0.4 0.2 US
PT
GB
IT
FR
ES
0.2 0.1 JP
IE
DE

0 0
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15 Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15

Figure 4: Modified Borrowing PD: for the sake of clarity we present the results only for the euro area periphery countries in
the left panels and for the largest five national banking systems in the right panels. In particular, for the two groups the upper
panels show the modified borrowing PD; the central panels display the original PD an the bottom panels report the difference
between the two measures.

30
 Multivariate Lending PD Multivariate Lending PD
1 GR 0.2 US
PT GB
IT FR
0.5 ES
0.1 JP
IE DE
0 0
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15 Q4−05Q4−06Q4−07Q4−08Q4−09Q4−10Q4−11Q4−12Q4−13Q4−14Q4−15

Univariate PD Univariate PD
1 GR 0.03 US
PT GB
IT
0.02 FR
0.5 JP
ES
0.01 DE
IE
0 0
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15 Q4−05Q4−06Q4−07Q4−08Q4−09Q4−10Q4−11Q4−12Q4−13Q4−14Q4−15

Difference Difference
GR US
0.4 0.2
PT GB
IT FR
0.2 ES 0.1 JP
IE DE

0 0
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15 Q4−05Q4−06Q4−07Q4−08Q4−09Q4−10Q4−11Q4−12Q4−13Q4−14Q4−15

Figure 5: Modified Lending PD: for the sake of clarity we present the results only for the euro area periphery countries in the
left panels and for the largest five national banking systems in the right panels. In particular, for the two groups the upper
panels show the modified lending PD; the central panels display the original PD an the bottom panels report the difference
between the two measures.

4
x 10 Borrowing Expected Loss 5
x 10 Borrowing Expected Loss
15 12
GR US
PT 10 GB
IT FR
10 ES 8 JP
IE DE
6

5 4

0 0
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15 Q4−05Q4−06Q4−07Q4−08Q4−09Q4−10Q4−11Q4−12Q4−13Q4−14Q4−15

6
x 10 Sum Borrowing Exp. Loss
2.5
Total
US
2 GB
FR
1.5 JP
DE

0.5

0
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15

Figure 6: Borrowing EL: the upper panels report the borrowing expected losses for the two groups, namely for the euro area
periphery countries on the left and for the largest five national banking systems on the right. The bottom panel displays the
total borrowing expected losses together with the contribution of the largest five national banking systems.

31
 4
x 10 Lending Expected Loss 5
x 10 Lending Expected Loss
15 5
GR US
PT GB
IT 4 FR
10 ES JP
IE 3 DE

2
5
1

0 0
Q4−05Q4−06Q4−07Q4−08Q4−09Q4−10Q4−11Q4−12Q4−13Q4−14Q4−15 Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15

6
x 10 Sum Lending Exp. Loss
2.5
Total
US
2 GB
FR
1.5 JP
DE

0.5

0
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15

Figure 7: Lending EL: the upper panels report the lending expected losses for the two groups, namely for the euro area periphery
countries on the left and for the largest five national banking systems on the right. The bottom panel displays the total lending
expected losses together with the contribution of the largest five national banking systems.

105 Aggregate EL borrowing

16

14

12

10

0
AU AT BE CA CL TW FR FI DE GR IN IE IT JP NL PT SG ES SE CH TR GB US

105 Aggregate EL lending

0
AU AT BE CA CL TW FR FI DE GR IN IE IT JP NL PT SG ES SE CH TR GB US

Figure 8: Borrowing vs Lending Expected Losses all sectors: the all sectors’ borrowing expected losses cumulated over time are
shown in the upper panel while the all sectors’ lending expected losses are reported in the lower panel.

32
 Aggregate EL borrowing - type
105 Official
12

10

0
AU AT BE CA CL TW FR FI DE GR IN IE IT JP NL PT SG ES SE CH TR GB US

Aggregate EL lending - type

104 Official

15

10

0
AU AT BE CA CL TW FR FI DE GR IN IE IT JP NL PT SG ES SE CH TR GB US

Figure 9: Borrowing vs Lending Expected Losses official sector: the official sector’s borrowing expected losses cumulated over
time are shown in the upper panel while the official sector’s lending expected losses are reported in the lower panel.

Aggregate EL borrowing - type

105 Bank

0
AU AT BE CA CL TW FR FI DE GR IN IE IT JP NL PT SG ES SE CH TR GB US

Aggregate EL lending - type

105 Bank
2.5

1.5

0.5

0
AU AT BE CA CL TW FR FI DE GR IN IE IT JP NL PT SG ES SE CH TR GB US

Figure 10: Borrowing vs Lending Expected Losses banking sector: the banking sector’s borrowing expected losses cumulated
over time are shown in the upper panel while the banking sector’s lending expected losses are reported in the lower panel.

33
 Aggregate EL borrowing - type
106 Non Bank

0
AU AT BE CA CL TW FR FI DE GR IN IE IT JP NL PT SG ES SE CH TR GB US

Aggregate EL lending - type

105 Non Bank

0
AU AT BE CA CL TW FR FI DE GR IN IE IT JP NL PT SG ES SE CH TR GB US

Figure 11: Borrowing vs Lending Expected Losses non bank private sector: the non bank private sector’s borrowing expected
losses cumulated over time are shown in the upper panel while the non bank private sector’s lending expected losses are reported
in the lower panel.

34
 x 10
4 Expected Loss Borrowing − Difference 4
x 10 Expected Loss Borrowing − Difference
2 4
GR US
1.5 PT GB
IT 3 FR
1 ES JP
IE 2 DE
0.5
1
0

−0.5 0

−1 −1
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15 Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15

4
x 10 Expected loss Lending − Difference 4
x 10 Expected loss Lending − Difference
1.5 6
GR US
PT GB
1 IT FR
4
ES JP
0.5 IE DE
2
0

0
−0.5

−1 −2
Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15 Q4−05 Q4−06 Q4−07 Q4−08 Q4−09 Q4−10 Q4−11 Q4−12 Q4−13 Q4−14 Q4−15

Agg. Expected Loss Borrowing - Difference

14000
Total
12000
US
10000 GB
FR
8000 JP
DE
6000

4000

2000

Q4-05 Q4-06 Q4-07 Q4-08 Q4-09 Q4-10 Q4-11 Q4-12 Q4-13 Q4-14 Q4-15

Agg. Expected loss Lending - Difference

14000
Total
12000 US
GB
10000 FR
JP
8000 DE

6000

4000

2000

0
Q4-05 Q4-06 Q4-07 Q4-08 Q4-09 Q4-10 Q4-11 Q4-12 Q4-13 Q4-14 Q4-15

Figure 12: Non Linear Effects: the plots show the difference in terms of borrowing and lending expected losses between the tensor
decomposition methods and the sum of the expected losses, computed separately for each sector using the matrix decomposition
technique. The four upper panels show the difference disaggregated for the two groups, the two lower panels report the results
for the complete economy together with the contribution of the five main advanced countries.

35
 CDS spreads - Multivar. Borrowing CDS spreads
8
7 0.05
6
5 0
4
3
2 -0.05
1
Q1-08 Q1-09 Q1-10 Q1-11 Q1-12 Q1-13 Q1-14 Q1-15 Q1-16
CDS spreads - Multivar. Lending CDS spreads
8
7 0.05
6
5 0
4
3
2 -0.05
1
Q1-08 Q1-09 Q1-10 Q1-11 Q1-12 Q1-13 Q1-14 Q1-15 Q1-16
CDS spreads - Multivar. Borrowing CDS spreads
1
8
7
6
5
0.5
4
3
2
1
0
Q1-08 Q1-09 Q1-10 Q1-11 Q1-12 Q1-13 Q1-14 Q1-15 Q1-16
CDS spreads - Multivar. Lending CDS spreads
1
8
7
6
5
0.5
4
3
2
1
0
Q1-08 Q1-09 Q1-10 Q1-11 Q1-12 Q1-13 Q1-14 Q1-15 Q1-16

Figure 13: Predictive performance: the differences between the RMSE obtained using a model based only on past spreads
values and a model based on network contagion, are shown in the first two panels (for the borrowing and for the lending side
respectively). Positive values imply the network contagion model is superior in predicting one-step ahead value of the CDS
spread. The lower panels show the binarised version of the results, the yellow colour indicates a superior performance of the
network model.

36
 MSD - Multivar. Borrowing CDS
8
7 0.02
6
5 0
4
3
2 -0.02
1
Q1-08 Q1-09 Q1-10 Q1-11 Q1-12 Q1-13 Q1-14 Q1-15 Q1-16

MSD - Multivar. Lending CDS

8
7 0.02
6
5 0
4
3
2 -0.02
1
Q1-08 Q1-09 Q1-10 Q1-11 Q1-12 Q1-13 Q1-14 Q1-15 Q1-16

MSD - CDS spread

8 0.03
7
6 0.02
5
4 0.01
3
2 0
1
Q1-08 Q1-09 Q1-10 Q1-11 Q1-12 Q1-13 Q1-14 Q1-15 Q1-16

Figure 14: Predictive performance: MSD of the network based model obtained using the borrowing and lending multivariate
CDS spreads (upper and central panels) and MSD obtained using a model based only on past market CDS spreads values (lower
panel). Positive values imply underestimating the future CDS values while negative values imply an overestimation.

37
Previous volumes in this series

795 Unintended consequences of unemployment Yavuz Arslan, Ahmet Degerli and

July 2019 insurance benefits: the role of banks Gazi Kabaş

794 Why have interest rates fallen far below the Magali Marx, Benoît Mojon and
July 2019 return on capital? François R Velde

793 Global real rates: a secular approach Pierre-Olivier Gourinchas and

June 2019 Hélène Rey

792 Is the financial system sufficiently resilient: Paul Tucker

June 2019 a research programme and policy agenda

791 Has globalization changed the inflation Kristin J Forbes

June 2019 process?

790 Are international banks different? Evidence Ata Can Bertay, Asli Demirgüç-Kunt
June 2019 on bank performance and strategy and Harry Huizinga

789 Inflation and deflationary biases in inflation Michael J Lamla, Damjan Pfajfar and
June 2019 expectations Lea Rendell

788 Do SVARs with sign restrictions not identify Jef Boeckx, Maarten Dossche,
June 2019 unconventional monetary policy shocks? Alessandro Galesi, Boris Hofmann
and Gert Peersman

787 Industry heterogeneity and exchange rate Camila Casas

May 2019 pass-through

786 Estimating the effect of exchange rate Thierry Mayer and Walter Steingress
May 2019 changes on total exports

785 Effects of a mandatory local currency pricing Renzo Castellares and Hiroshi Toma
May 2019 law on the exchange rate pass-through

784 Import prices and invoice currency: evidence Fernando Giuliano and Emiliano
May 2019 from Chile Luttini

783 Dominant currency debt Egemen Eren and Semyon Malamud

May 2019

782 How does the interaction of macroprudential Előd Takáts and Judit Temesvary
May 2019 and monetary policies affect cross-border
bank lending?

781 New information and inflation expectations Serafin Frache and Rodrigo Lluberas
April 2019 among firms

All volumes are available on our website www.bis.org.