Download The Role of the Bank Balance-Sheet Channel for the Transmission of

Global Banking, Global Crises?
The Role of the Bank Balance-Sheet Channel for the Transmission of Financial Crises
By Rudiger Ahrend and Antoine Goujard1
May 2013
This paper examines whether shocks to leveraged creditors with cross border holdings have an incidence
on debtor countries’ risk of suffering financial turmoil. For this we use new measures of bank balancesheet shocks which capture well the contagion observed e.g. in the wake of the Mexican and Asian crises,
and confirm that contagion shocks observed in 2009/10 dwarfed those observed during previous financial
crises. Relying on a panel of 178 developed and emerging economies from 1984 to 2009, we find that bank
balance-sheet shocks indeed strongly increase the risk of a systemic banking crisis. Confirming these
results, bilateral bank flows significantly decrease when creditor banks’ assets are hit by negative shocks
from third-party countries. Moreover, certain forms of cross border lending are found to particularly
increase crisis risk. For example, short-term bank debt increases roll-over and funding risks, thereby
amplifying the impact of contagion shocks. In contrast, a high level of aggregate liquidity attenuates the
transmission of shocks to banks’ assets to debtor countries.
JEL Codes: E44, F34, F36, G01
Keywords: banking crises, contagion, financial stability, banking system, financial spillovers, balance
sheet.
1.
OECD Economics Department: [email protected], [email protected].
The authors are indebted to Romain Duval, Jørgen Elmeskov, Issam Hallak, Yannick Kalantzis, Jean-Luc
Schneider, Cyrille Schwellnus, Cedric Tille, as well as to seminar participants at the Geneva Graduate Institute of
International and Development Studies, at the Banque de France – Deutsche Bundesbank Workshop on Current
Macroeconomic Challenges and at the XX International Tor Vergata Conference on Money, Banking and Finance for
useful comments. The authors would like to thank Olga Tschekassin and Vera Zipperer for excellent research
assistance. The authors are also very grateful to the Bank of International Settlements, and in particular SwapanKumar Pradhan, for providing bilateral Locational Banking Statistics together with very helpful advice. All remaining
errors are those of the authors. The views expressed here are those of the authors, and do not necessarily reflect those
of the OECD or its member countries.
1
Introduction
Strong financial contagion was one of the key features of the 2007-09 financial crisis as localised
problems in specific segments of financial markets rapidly morphed into a crisis of global dimensions.
While the potential threat from financial contagion has been widely recognised, there is still no consensus
about the main channels through which financial contagion propagates financial instability across
countries. Even though balance sheets of leveraged financial institutions with cross-border holdings, such
as banks, have received increasing interest by commentators and economists alike (see e.g. Krugman,
2008), the empirical evidence on the importance of the bank balance-sheet channel remains sketchy at best.
To close this gap, we use newly-built indicators of bank balance-sheet contagion to explore whether the
bank balance-sheet channel has indeed played a systematic role in spreading financial crises during past
episodes of financial instability. We differentiate between bank balance-sheet shocks resulting from
economic developments in countries whose banks are actively internationally lending, and bank balancesheet shocks driven by developments in third-party borrowing countries (common-creditor shocks).
Finally, we explore possibilities to mitigate the impact of bank balance-sheet shocks on financial stability.
A large body of theoretical and empirical work has examined both the determinants of financial crises
and why financial crises tend to come in clusters.2 Beyond the possibility that simultaneous banking crises
in different countries may be caused by a common shock to fundamentals, the literature proposes financial
market perceptions, trade, and balance sheets of financial intermediaries as main channels of contagion.
This paper focuses on the balance sheet channel. Theoretical models of banking crises emphasise that large
and highly leveraged financial institutions, such as international banks, play a key role in propagating
financial turmoil (e.g. Greenwood et al., 2011 or Tirole, 2011 for a review). Such bank-driven contagion
arises, for example, when banks from a creditor country with a deteriorating risk profile decide to
withdraw international funding to comply with internal rules or prudential regulations such as capital
adequacy requirements. Bank balance-sheet driven contagion can also arise indirectly through the
international banking system when banks cut back on loans to a country in response to suffering losses on
loans to another country. This indirect financial contagion mechanism – often referred to as commoncreditor contagion - was first suggested by Calvo (1998) to introduce a causal link between the 1998
Russian crisis and the following crisis in Brazil.
While not directly addressing the link between bank balance-sheet contagion and financial crises,
Cetorelli and Goldberg (2011) examine how, during the 2007-09 crisis, adverse liquidity shocks to the
banking systems of the major developed economies were transmitted to emerging markets through a
contraction in cross-border lending. Similarly, McGuire and Tarashev (2008) provide some descriptive
evidence that international bank lending to emerging markets declined when bank health deteriorated over
the 1992-2007 period. Microeconomic studies also point to specific cases where bank balance-sheet
contagion likely played some role. Using differences in dollarisation of bank deposits in Pakistan prior to
the 1998 nuclear tests that led to runs on dollar deposits as a natural experiment, Khwaja and Mian (2008)
show that Pakistan banks more exposed to deposit runs contracted their lending to firms more strongly.
Similarly, Schnabl (2012) presents evidence that following the 1998 Russian default, domestically-owned
Peruvian banks dependent on international borrowing reduced lending to Peruvian firms, in contrast with
Peruvian banks without international borrowing exposure. However, in spite of an evolving literature, the
empirical evidence that international banks do indeed systematically transmit domestic and third-party
borrower’s shocks to their debtor countries remains lacking.
Filling this gap, we make three main contributions. First, we analyse the effect of bank balance-sheet
shocks on the likelihood of systemic banking crises and capital flow reversals. Using a large dataset of 178
developed and emerging economies over the 1983-2009 period and a difference-in-differences
2.
For a review, see e.g. Demirgüç-Kunt and Detragiache (2005) or Végh (forthcoming).
2
identification strategy, we show that bank balance-sheet contagion tends to be a general and powerful
mechanism of crisis transmission. Second, we show that bank balance-sheets not only transmit financial
turmoil resulting from shocks to the economies of the lending banks, but that there are also powerful
contagion effects to borrowing countries when foreign lending banks suffer from deteriorating credit
quality in third-party debtor countries. Finally, looking at factors that can affect the transmission of
banking crises, we show that larger roll-over risk due to a lower residual maturity of bank debt increases
exposure to bank balance-sheet shocks. Similarly, we find debtor countries to be more vulnerable to bank
balance-sheet shocks when their domestic banking systems is more leveraged, has lower liquidity reserves,
or a higher credit-to-deposits ratio.3 In contrast, vulnerability to bank balance-sheet shocks is lower in
situations of abundant global liquidity, underlining the importance of major central banks ensuring ample
international liquidity at times of financial turmoil.
The remainder of this paper is divided in four sections: Section 1 introduces our new measures of
bank balance-sheet shocks and describes our identification strategy. Section 2 describes the data and
presents some descriptive statistics. Section 3 presents the empirical analysis on the effect of bank balancesheet shocks on the likelihood of systemic banking crises. Finally, Section 4 examines how bank balancesheet shocks interact with other factors of financial vulnerability. In addition, it verifies that bank flows are
indeed a key channel through which bank balance-sheet shocks affect financial stability.
1. Identifying the effect of bank balance-sheet shocks on the likelihood of systemic banking crises
1.1 Measuring bank balance-sheet shocks
In essence, bank balance-sheet (BBS) shocks stem from deteriorations in the balance sheets of a
country's foreign bank creditors. BBS shocks can be direct or indirect. Direct BBS contagion arises, for
instance, when banks from a creditor country with a deteriorating risk profile decide to withdraw
international funding to comply with internal rules or prudential regulations such as capital adequacy
requirements. Indirect BBS contagion, in turn, arises when banks cut back on loans to a debtor country in
response to deteriorations in their loan book to another country. Krugman (2008) refers to such
international financial contagion as the “International Finance Multiplier”. In this section, we construct a
measure of aggregate BBS shocks for each debtor country. We then decompose these aggregate shocks
into indirect BBS shocks arising through third party countries (common-creditor shocks), and BBS shocks
arising through economic and financial developments in the countries where creditor banks are located
(referred to as lending-country spillovers).
A measure of overall bank balance-sheet shocks
Many international bank lenders aim for an internationally diversified portfolio of debt and equity
assets with a certain level of risk or leverage for each period - e.g. to conform to capital requirements.
When their asset portfolio is hit by a negative shock they could therefore (be forced to) cut credit to
borrowers from countries perceived as more risky. Consequently, borrowers may find their access to credit
restricted even in a situation where their country’s (and their own) credit risk has remained unchanged. We
construct an overall measure of BBS shocks from BIS locational statistics which have been collected since
1977. These bilateral data show the total amount that banks from each BIS reporting country have lent to
the financial and non-financial institutions of each other country (including to countries that are not
3.
A related literature investigates if lending by multinational banks can dampen financial shocks in debtor
countries (e.g. Peek and Rosengren, 2000a,b; De Haas and Van Lelyveld, 2006; 2010).
3
reporting to the BIS).4 More formally, for country d in period t, its level of contagion, i.e. the exposure of
its creditor banks to shocks, is captured by:
BBS Shocksdt 

rR , r  d
wbldrt 
rating rt
rating r ,t 1
(1)
with R the set of lending (reporting) countries. Wbldrt is the share (weight) of bank liabilities of country d
held by country r at time t. ratingr,t is the rating of country r according to Institutional Investor. These
ratings have been used, e.g., by Reinhart and Rogoff (2009), Eichengreen and Mody (2000) and Hallak
(2011), the latter also providing a detailed discussion of them. They are based on bi-annual data from a
survey in which institutional investors are asked to grade each country on a scale from 0 to 100, with 100
representing the highest level of creditworthiness. Grades are then aggregated by weighting them with the
actual investment exposure of the different institutional investors in a given country. These ratings are not
only available for a much larger sample than ratings from rating agencies, but by being based on the
perception of the main investors in each market are also likely to be better reflected in price developments.
This is a clear advantage for constructing BBS shocks as movements in prices would be expected to have
balance-sheet effects for banks and other leveraged financial intermediaries. For the purpose of this paper,
March and September ratings were averaged. Alternatively, we used only September ratings which did not
affect our main results. As a robustness check, we also computed a variant of Equation (1) replacing credit
ratings by indices of bank share prices, with qualitatively similar results (see Table A6 in the Appendix)
Bank balance-sheet shocks during the recent global financial crisis have dwarfed previously observed
levels of contagion (Figure 1). Looking at the average bank balance-sheet shocks across countries over the
last two decades shows that important episodes of global BBS contagion are mainly related to
developments in the large advanced economies. Also, average bank balance-sheet shocks have a cyclical
component in the sense of being particularly high in global recessions. Historically, bank balance-sheet
shocks have been relatively high during the global recession of the early 1990s, as well as in the early
2000s in a situation where economic weakness in the wake of the dot.com boom combined with fears that
large telecom companies may default on bonds issued to acquire UMTS licences.
The 1997-98 Asian crisis or the 1995 Mexican crisis did not lead to visible bank balance-sheet shocks
at the global level, even though at the regional level and for certain emerging economies outside the region
strong contagion effects were clearly visible at the time. Figure 2.A. shows the relative level of BBS
shocks faced by each country in 1995 in the wake of the Mexican crisis. It illustrates that Latin-American
countries indeed suffered from BBS shocks at the time. Similarly, Figure 2.B shows that Asian countries
were those most affected by BBS shocks during the Asian crisis. While there seems to have been some
contagion via the bank balance-sheet channel from Latin America to Asia, and vice versa, during these two
crises, emerging economies in Eastern Europe were left largely unaffected, as were most advanced
economies.
4.
These data are e.g. used by Kalemli-Ozcan et al. (2012) to examine the link between bilateral banking
integration and business-cycle synchronisation.
4
Figure 1. Financial contagion via the bank balance-sheet channel
6
Dot.com recession &
Telecom-bond scare
Early 1990s recession
5
Global financial
crisis
4
3
2
1
0
-1
-2
*
2011
2010
2009
2008
2007
2006
2005
2004
2003
2002
2001
2000
1999
1998
1997
1996
1995
1994
1993
1992
1991
1990
-3
Based on bank balance-sheet shocks calculated as in Equation (1), the average contagion shock is measured as the unweighted average of debtor countries’ contagion shocks. Each debtor country's contagion shock is computed as the weighted
sum of its creditors’ credit rating percentage changes.
Source: Authors’ calculations based on BIS bilateral locational banking statistics and Institutional Investor sovereign credit ratings.
Even though during the recent global financial crises all countries suffered from BBS shocks, the
regional focus was different from the earlier crises of the 1990s. This time, Latin American and Asian
economies were among the least affected whereas advanced and European emerging economies were those
exposed to the strongest BBS shocks (Figure 2.C). While large BBS shocks in European emerging
economies likely reflected large exposure to Western European economies, part of it may also have been
explained by the low degree of banking integration within the region. A relatively higher degree of intraregional banking integration probably amplified the Asian and Mexican crises in the respective region of
origin, but was an advantage in a crisis that originated and mainly affected the most advanced economies.
Though the systematic association between our measure of BBS shocks and widespread banking
crises is striking, caution is needed to give a causal interpretation to these descriptive statistics. The
previous measure of BBS shocks takes into account how changes in creditor banks' sovereign rating are
transmitted to debtor countries, and part of the observed pattern may be due to reverse causation: debtors
country shocks affecting creditor countries’ sovereign ratings. We deal with this issue by relying on a
difference-in-difference strategy that uses lagged measures of BBS shocks (see Section 3) and by
introducing BBS shocks that stem from third-party countries.
5
0.4
0.3
Emerging Latin America
Emerging Asia
Emerging Europe & South Africa
Advanced Economies
0.2
0.1
0.0
-0.1
-0.2
A.Mexican crisis
COL
CHL
MEX
GBR
URY
PRT
BRA
SVN
ARG
PHL
BEL
ISR
NLD
FRA
CHN
KOR
IDN
MYS
USA
LUX
THA
AUT
HKG
AUS
JPN
POL
CAN
IND
SGP
TUR
NOR
ESP
ITA
CHE
SVK
NZL
DEU
DNK
IRL
CZE
VNM
HUN
FIN
RUS
GRC
ZAF
SWE
ISL
EST
KAZ
-0.3
2.5
2.0
Emerging Latin America
Emerging Asia
Emerging Europe & South Africa
Advanced Economies
1.5
1.0
0.5
0.0
-0.5
-1.0
B. Asian crisis
HKG
SGP
THA
KOR
MYS
AUS
CHN
USA
IDN
IND
COL
PHL
VNM
CAN
NZL
CHL
GBR
MEX
BRA
BEL
URY
CHE
ISR
TUR
NLD
DEU
ARG
AUT
JPN
KAZ
LUX
DNK
ZAF
FRA
IRL
RUS
POL
SWE
CZE
ITA
ISL
ESP
SVK
FIN
HUN
GRC
NOR
PRT
SVN
EST
-1.5
1.5
Emerging Latin America
Emerging Asia
Emerging Europe & South Africa
Advanced Economies
1.0
0.5
0.0
-0.5
-1.0
C. 2009 crisis
-1.5
SVN
ISL
ZAF
NZL
SVK
PRT
JPN
MEX
ITA
AUS
NLD
CZE
KAZ
CHE
CAN
HUN
USA
FRA
FIN
EST
DEU
ESP
POL
IND
ARG
DNK
NOR
IRL
GBR
RUS
BRA
KOR
PHL
TUR
SWE
ISR
MYS
CHN
BEL
CHL
AUT
URY
HKG
LUX
IDN
GRC
THA
SGP
VNM
COL
relative to the country sample mean)
relative to the country sample mean)
relative to the country sample mean)
Figure 2. Bank balance-sheet shocks during the Mexican crisis, Asian crisis and the global financial crisis
Sample: The shocks are measured in 1995, 1998 and 2009 for Panel A, B and C, respectively. OECD countries, BRICS and selected
economies (Argentina, Colombia, Hong Kong SAR, Indonesia, Kazakhstan, Malaysia, Philippines, Singapore, Thailand, Uruguay,
Vietnam). Emerging Latin America refers to: ARG (Argentina), BRA (Brazil), CHL (Chile), COL (Colombia), MEX (Mexico), URY
(Uruguay). Emerging Asia refers to: CHN (China), HKG (Hong Kong SAR), IDN (Indonesia), IND (India), KOR (Korea), MYS
(Malaysia), PHL (Philippines), SGP (Singapore), THA (Thailand), VNM (Vietnam). Emerging Europe and South-Africa refers to: CZE
(Czech Republic), EST (Estonia), HUN (Hungary), KAZ (Kazakhstan), POL (Poland), RUS (Russia), SVK (Slovak Republic), SVN
(Slovenia), TUR (Turkey), ZAF (South Africa). The advanced economies refer to: AUS (Australia), AUT (Austria), BEL (Belgium),
CAN (Canada), CHE (Switzerland), DEU (Germany), DNK (Denmark), ESP (Spain), FIN (Finland), FRA (France), GBR (United
Kingdom), GRC (Greece), IRL (Ireland), ISL (Iceland), ISR (Israel), ITA (Italy), JPN (Japan), NLD (Netherlands), NOR (Norway), NZL
(New Zealand), PRT (Portugal), SWE (Sweden), and USA (United States).
Source: Authors’ calculations based on BIS bilateral locational banking statistics and Institutional Investor sovereign credit ratings.
6
“Common-creditor” shocks: Balance-sheet shocks stemming from third-party countries
Part of the empirical literature argues that banking crises have been transmitted from one country to
another through international banks with lending exposure to both countries. For example, Kaminsky et al.
(2003) present evidence that Thailand, Korea, Indonesia and Malaysia were particularly strongly affected
by common-creditor contagion during the Asian crisis. Therefore, apart from a general measure of bank
balance-sheet shocks, we propose two indices that capture, respectively, bank balance-sheet shocks
stemming from third party countries with which one happens to have common creditors (CCC), and bank
balance-sheet shocks stemming directly from economic developments in lending countries (LCS).
Common-creditor shocks and lending-country spillovers are fundamentally different. Common-creditor
shocks may lead to contagion when banks, being unable or unwilling to cushion borrowing countries
against such shocks, act as a conduit for transmitting shocks from third-party countries. By contrast,
lending-country spillovers represent (at least in large part) direct shocks to the countries of the main
international lending banks.
Consider the international lending of country r’s banking system. The quality of the loan portfolios of
country r’s banks can be affected either by a domestic shock or a foreign shock – i.e. a shock to countries
to which country r’s banks have lent money. The degree to which country r’s banks are affected by foreign
shocks is computed as the change in the rating of the countries to which its banks have been lending to,
weighted by the share of lending to the respective country. Common-creditor contagion (CCC) shocks are
calculated as the (aggregate) degree to which country d’s creditor banks have been affected by foreign
shocks (excluding shocks to country d), weighted by their lending to country d. More formally, for country
d in period t its level of common-creditor contagion (CCC), i.e. the exposure of its creditor banks to thirdcountry shocks, is captured by:
CCCShocksdt   ( wbldrt
rR

vV ,v  d
welrvt 
ratingvt
)
ratingv ,t 1
(2)
with R the set of lending (reporting) countries and V the set of borrowing countries, the latter also
including reporting countries. welrvt is the share (weight) of country v in the foreign loans extended by
country r at time t (excluding the assets of country r located in country d). Wbldrt is the share (weight) of
bank liabilities of country d held by country r at time t. This new measure of BBS shocks does not depend
on the direct relationship between debtor and creditor countries and is thus less likely to lead to problems
of reverse causation in the econometric analysis.
Lending-country spillovers: Balance-sheet contagion through shocks to creditor countries
The calculation of the strength with which creditor banks’ are affected by domestic and foreign
shocks for all BIS reporting countries allows to compute, in a second step, lending-country spillovers
(LCS) that are calculated as the (aggregate) degree to which country d’s creditor banks have been affected
by domestic shocks, weighted by their lending to country d. For each period, the degree to which country
r’s banks are affected by domestic shocks is measured by the change in country r’s own rating that is not
driven directly by foreign developments. The exposure of debtor country d to creditor banks’ domestic
shocks is measured through changes in the rating of its creditor countries. However, such rating changes do
not only reflect domestic developments in the creditor countries, but potentially also shocks to their
debtors. Therefore, direct shocks to creditor banks’ countries have been isolated from shocks to their
international debt portfolio. First, direct shocks were measured as the residuals from a regression of
changes in a creditor country’s credit rating on the aggregate change in the credit rating of the countries
included in its international bank assets portfolio.
7
ratingct     portfolio _ ratingct   ct
(3)
with portfolio_rating being computed as the sum of rating shocks weighted by asset exposure vis-à-vis all
countries indebted towards country c:
portfolio _ ratingct   weacvt 
vV
ratingvt
ratingv ,t 1
(4)
and where weacvt is the share (weight) of the external assets of country v in the portfolio of the banks’ of
the reporting creditor country c at time t. Alternatively, similar regressions are run that also feature country
and year fixed effects. The exposure of country d to direct creditor shocks at time t is then given by the
weighted sum of its creditors’ domestic shocks:

LCS Shocksdt   wbldrt   rt
(5)
rR
with wbldrt again the share (weight) of bank liabilities of country d held by country r at time t.
For both the Mexican and Asian crises, contagion to countries in the region occurred through LCS
spillovers and CCC shocks (Table 1), although common-creditor contagion was less widespread for the
Mexican crisis, with Brazil, Uruguay and Argentina being most strongly affected. For example, Kaminsky
et al. (2003) argue that Thailand, Korea, Indonesia and Malaysia were particularly strongly affected by
common-creditor contagion during the Asian crisis. And indeed, our CCC indicator confirms that – after
the financial centers in the region - Thailand, Korea, and Malaysia (plus Australia) faced the largest CCC
shocks during the Asian crisis. Similarly, Kaminsky (1999) argues that, outside Asia, Mexico was
particularly strongly affected by contagion during the Asian crisis. Again, the CCC indicator confirms that
Mexico was the emerging economy that faced the largest CCC shock outside Asia.
Table 1. Different types of bank balance-sheet contagion during the Asian and Mexican crises
Countries most affected by lending-country spillovers and common-creditor shocks in 1995 and 1998
Ranking by
(1)
shock size
1
2
3
4
5
6
7
8
9
10
1.
1995
Lending-Country
Common-Creditor
Spillovers
Shocks
Colombia
Brazil
Chile
United Kingdom
Mexico
Estonia
United Kingdom
Germany
Slovenia
Russia
Uruguay
Portugal
Brazil
Ireland
Philippines
Uruguay
Argentina
Japan
Korea
Belgium
1998
Lending-Country
Common-Creditor
Spillovers
Shocks
Hong Kong SAR
Singapore
Singapore
Hong Kong SAR
Thailand
Thailand
Korea
Malaysia
Malaysia
Australia
Australia
Korea
China
United States
India
China
Indonesia
United Kingdom
Colombia
Mexico
Each column shows the ten countries that were most affected by a given type of contagion in a given year. Countries are
ranked in decreasing order of contagion. Sample: OECD countries, BRICS and selected other economies (Argentina,
Colombia, Hong Kong SAR, Indonesia, Kazakhstan, Malaysia, Philippines, Singapore, Thailand, Uruguay, Vietnam).
Source: Authors’ calculations based on BIS bilateral locational banking statistics and Institutional investor sovereign credit ratings.
8
1.2. Cross-border transmission of systemic banking crises
In order to identify how bank balance-sheet shocks affect the likelihood of systemic banking crisis in
a country, the empirical analysis relies on a difference-in-differences specification. Countries have
different levels of exposure to BBS shocks, and similar BBS shocks do not have the same impact across
countries. We measure exposure to BBS shocks as (lagged) total bank debt, as a percent of GDP, using
consolidated bank debt towards BIS reporting banks from the BIS banking statistics as in the theoretical
models of Krugman (2008) and Devereux and Yetman (2010). Ex-ante reliance on international banking
strongly differs both across countries and over time, and a relevant comparison group can be used to
identify temporal variation in the likelihood of systemic banking crises that is not due to the strength of
BBS shocks. The effect of BBS shocks is captured by the interaction term between BBS shocks and the exante reliance of a country on cross-border bank funding. These specifications examine if countries that
were ex-ante more reliant on foreign bank funding were indeed more affected by BBS contagion shocks
than countries that were ex-ante less reliant on such funding.
More precisely, a baseline equation assumes that, conditional on the absence of crisis at time t-1, the
likelihood of a banking crisis at time t is determined by a linear probability model:
cit   Shocksit 1   Expit 1   Expit 1  Shocksit 1  X it 1  i   t   it ,
(6)
where cit is a dummy variable taking value one at the beginning of a systemic banking crisis. Shocksit-1 are
either bank balance-sheet shocks, common-creditor shocks or lending-country spillovers. Expit-1 is a
measure of exposure to bank balance-sheet shocks, for example the ratio of bank debt to GDP. Xit-1 is a row
vector containing the relevant financial account characteristics as well as additional explanatory variables
identified by the literature as key determinants of financial stability. γi and δt capture unobserved country
and time specific shocks. εit are idiosyncratic disturbances.
The functional form of the linear probability model is convenient for panel data sets because it leads
to estimators of crisis incidence that are free of country-specific heterogeneity, do not require specification
of initial conditions in the dynamic model, and provide the best linear approximations (in the mean-square
error sense) of the average marginal effects (Angrist, 2001; Stewart, 2007).5 All explanatory variables are
lagged in order to measure them prior to the crisis period and to mitigate concerns of endogeneity. As most
explanatory variables and financial factors are affected by systemic banking crises, they may be only
predetermined or weakly exogenous. Assuming that the explanatory variables are strictly exogenous would
imply, for example, the too-strong statement that the country-specific idiosyncratic shocks that coincide
with systemic banking crises are unrelated to future financial account developments. In particular, the
preferred empirical strategy relies on a dynamic framework rather than a fixed-effects specification.
Following Hyslop (1999), the preferred specification is a flexible linear first-difference equation:6
cit  Shocksit 1  Expit 1    Expit 1  Shocksit 1   X it 1   t   it
(7)
Specification (7) is less restrictive than a country fixed-effects model as it requires only that the
regressors are weakly exogenous. However, as argued by Griliches and Hausman (1986), specification (7)
5.
The main criticism of the linear probability model concerns its use for predictions which is not the purpose
of this paper.
6.
See Card (1990) for an early application and De Ree and Nillesen (2009) for an application in a panel of
countries. Falcetti and Tudela (2008) investigate the probability of banking crises in a related non-linear
dynamic framework including a lagged dependent variable. However, their implicit assumption is that the
start of a systemic banking crisis and its continuation are the results of the same stochastic process.
9
may lead to more severe attenuation bias than fixed-effect models in the presence of measurement error.
This is an important identification issue as systemic banking crises are rare events, and the dependent
variable has a limited variance. Hence, the empirical section reports the results of pooled OLS, fixedeffects, first-difference and first-difference instrumental variable specifications.7
The explanatory variables in Xit-1 are those typically suggested in the literature and are grouped into
three main categories:8

Macroeconomic and related country-specific characteristics, which include (log) per capita GDP,
the (log) population and the (log) openness to international trade measured as the sum of imports
and exports as a share of nominal GDP.

The country’s financial sector and international financial exposure. This is measured e.g. by the
development of the domestic banking sector, as proxied by the size of credit to the non-financial
sector.

Country specific balance-sheet characteristics of the financial account. These include the share of
debt in foreign liabilities, as well as quadratic functions of foreign currency reserves, other assets,
and of liabilities (all as a share of GDP).
Furthermore, Xit-1 include control variables for state and duration dependence. Indeed, after a systemic
banking crisis, some explanatory variables are likely to be affected by the crisis itself. Furthermore, the
probability that a crisis occurs in a country that already suffered financial turmoil in the past is typically
higher than for a country where no crisis occurred recently, which could confound the effects of
explanatory variables if not accounted for (Demirgüç-Kunt and Detragiache, 1998). These state and
duration dependence effects are approximated by interacting a quadratic function of the time elapsed since
the last systemic banking crisis with dummy variables for the number of previous crises. The final model
can be thought of as a flexible, discrete-time hazard model that allows for time-varying covariates, duration
and state dependence, and correlated unobserved heterogeneity. Allison (1982) shows that estimates from
models of this type converge to those obtained from continuous time-duration models.
We also test whether the impact of BBS shocks could be heterogeneous across countries depending on
the structure of their banking sector (see Section 4). Theoretical models of bank failures emphasise that
large and highly leveraged financial institutions, which result in part from the features and (lack of)
enforcement of financial regulation, play a key role in propagating financial turmoil (see e.g. Greenwood
et al., 2011, or Tirole, 2011, for a review). Therefore, we interact exposure-scaled BBS shocks with
measures of vulnerabilities of the domestic banking sector: its reliance on deposits to fund credit, its
leverage, and its level of liquid reserves over assets.
As a robustness check and to verify that bank flows are indeed a key channel through which BBS
shocks affect financial stability, we analyse the effect of BBS shocks on bank flows in a bilateral set-up.
Bilateral bank flows allow controlling for debtor-specific shocks and, in particular, cross-border lending
demand shocks. We rely on a specification close to Khwaja and Mian (2008) and Cetorelli and Goldberg
(2011). More precisely, bank flows Δln(Lcdt) are computed as the change in (end-of-year logarithm of)
7.
The empirical relevance of the first-difference and fixed-effect specifications is assessed using the
heteroskedacity- and autocorrelation-robust Hausman tests suggested by Wooldridge (2002). The tests
were implemented using a country block bootstrap process (Cameron and Trivedi, 2009). As a robustness
check, fixed-effects conditional logit models were estimated. The results were close to those of fixedeffects linear models (results not reported).
8.
The definition and construction of each variable is detailed in Section 2 and Table A1 of the Appendix.
10
consolidated liabilities of country d (debtor) towards the banks of country c (creditor). The relationship
between yearly flows and shocks is given by:
 ln( Lcd ,t )    Shockscd ,t 1   d ,t   cd   cd ,t
,
(8)
where Shockscd,t-1 represents a creditor-specific shock, more precisely a shock stemming from third-party
countries. γd,t captures any time-varying shocks that are specific to debtor countries (e.g. demand-side
shocks or valuation effects) as well as the influence of the overall macroeconomic environment
(e.g. changes in world interest rates). δcd takes into account the time-invariant factors that could affect the
level of bank flows of the country pair. Alternatively, some specifications only control for creditor fixed
effects. εcd,t are idiosyncratic disturbances. The empirical specifications for consolidated flows allow to
examine if consolidated lending flows towards a given debtor country decline when the balance sheets of
its creditor banks deteriorate due to rating changes in third-party debtor countries to which these banks are
exposed. In this bilateral set-up, shocks to international bank portfolios are approximated by:
Shockscd ,t   v ,v  d wblcdv ,t
ratingv ,t
ratingv ,t 1
,
(9)
where wblcdvt represents the share of all cross-border loans by the banks of creditor country c that have
been extended to country v (excluding the debtor country d), and ratingvt is the credit rating of debtor
country, v.
2. Data and descriptive statistics
2.1 Data
We use a panel data set of 178 countries over the period 1984-2009. This sample corresponds to the
178 countries covered in the 2009 update of Lane and Milesi-Ferretti (2007) for which consolidated
banking statistics are available as of December 1983. Our main dependent variable, the occurrence of
systemic banking crises, is taken from Laeven and Valencia (2010), a data set that has been widely used
and is considered as one of the main sources of information on systemic banking crises. Laeven and
Valencia define episodes of systemic banking crises as periods displaying both substantial financial
distress in the banking system (as indicated by significant bank runs, losses in the banking system, and
bank liquidations) and notable banking policy interventions in response to losses in the banking system. As
Laeven and Valencia (2010) define the end of a systemic banking crisis as the year preceding two
consecutive years of real GDP and credit growth, observations corresponding to an ongoing banking crisis
and the two subsequent years are not included in the regressions.9 We observe 129 systemic banking crises
in 114 countries over the 1984-2009 period. In average, each year 7.7% of the countries in our sample are
in a crisis episode, including roughly 2.6% of the countries that are entering into a crisis episode.10 These
crisis episodes have large economic costs. Laeven and Valencia (2010) calculate that they are associated
with an average decrease in GDP of 28 percents over the first three years. As the coded end of a crisis
episode could also capture a temporary drop in crisis intensity that resurges a few years later, we test the
sensitivity of our results by using two alternative dependent variables. First, we drop the 15 borderline
9.
When the first two years of a crisis record real GDP and real credit growth, the first year of the crisis is also
the end year.
10 .
See Laeven and Valencia (2010, 2012) and Appendix Tables A4 and Figure A1 for additional descriptive
statistics.
11
systemic banking crises from Laeven and Valencia (2010). 11 Second, we bring into play the Reinhart and
Rogoff (2011) dataset on banking crises which is available for a smaller sample of 70 countries, covering
not only systemic banking crises but also isolated banking failures.12 As individual banking failures are
more likely to be the consequence of mismanagement at the concerned institution rather than a sign of
financial contagion, we use the Reinhart and Rogoff dataset to improve the quality of measurement by
restricting our original sample to the crises that have the same starting date as in Reinhart and Rogoff. We
find similar results using the two definitions of Laeven and Valencia, or when restricting the sample of
crises as described.13
Our main explanatory variables, the measures of BBS shocks and the indicators of domestic
exposures to those shocks are constructed from the BIS locational and consolidated banking statistics and
Institutional Investor’s sovereign ratings (as described in Section 1.1). Our preferred measure of exposure
to BBS shocks is the ratio of consolidated debt to foreign banks over GDP (hereafter debt to foreign banks
over GDP). Consolidated data provide internationally comparable measures of national banking systems’
exposure to other individual countries: banks headquartered in a particular reporting country provide
information on their foreign claims on borrowers in up to 200 vis-à-vis countries. This has two main
advantages over more traditional measures of exposure to foreign banking system such as locational bank
lending, which is recorded on balance-of-payment principles. First, in contrast to locational bank data,
consolidated data do not track flows from the global banks to their destination country even when they are
intermediated by financial centers.14 Using locational bank flows would therefore overestimate both the
contribution of financial centers to BBS shocks, as well as their exposure to them, while underestimating
those figures for the other countries.15 Second, consolidated debt to foreign banks offers a break-down by
residual maturity that allows to measure funding risk. Short-term debt to foreign banks is measured as debt
with remaining maturity up to and including one year, i.e. both debt with original maturity of up to one
year and debt with original maturity of more than one year but falling due within the next 12 months.
Deposits or claims of banks that are receivable on demand are also considered as short-term debt. By
contrast, all other claims of foreign banks are considered as long-term exposure to foreign banks.
Our control variables are constructed from several data sources.16 The size and composition of the
financial account and GDP in current US dollars are taken from the 2009 update of Lane and MilesiFerretti (2007). 2008 data are taken from the IMF IFS update of March 2012. Population, openness to trade
and export to GDP ratio are taken from the World Bank World Development Indicators March 2011
11 .
Laeven and Valencia (2010) define borderline cases of banking crises as almost meeting their definition of
a systemic event. In particular, 10 of the 15 borderline cases examined here took place in 2008.
12 .
A simple comparison of Reinhart and Rogoff (2011) and Laeven and Valencia (2010) definitions indicates
that, for the same sample of countries over the 1983-2009 period, the downgrades of credit rating
associated with a banking crisis and a systemic banking crisis are 2.7% and 4.2%, respectively, while GDP
growth rates are +0.9% and -0.1%, respectively. In 81% of the cases of ongoing systemic banking crises,
the country is also considered as in banking crisis by Reinhart and Rogoff (2011). However, only 48% of
the ongoing banking crises identified in Reinhart and Rogoff correspond to systemic events.
13 .
Results are available upon request.
14 .
The difference between consolidated and locational statistics varies strongly across countries, but can be
substantial. For example, at the end of 2009, the difference between locational and consolidated debt was
around 10% in France or Germany, but reached 100% for the United Kingdom and 115% in the main
offshore banking centers.
15 .
Despite this difference, the results were qualitatively similar when we replace our preferred measure of
exposure by the ratio of locational debt to foreign banks over GDP. The two variables have a linear
correlation of 0.98.
16 .
Appendix Table A1 displays the definitions of the control variables.
12
update (WDI). The measures of domestic banking sector vulnerability are taken from the World Bank WDI
and Beck et al. (2009). Export prospects are measured as the export-weighted sum of trading partners’ real
growth, and the reliance of the domestic economy on export prospects is quantified by the interaction of
the export prospects term with the export to GDP ratio (lagged by three years). More precisely, export
weights are computed as average exports over the period 1990-2009 from UNCTAD data, while real GDP
growth is based on (by order of importance): OECD, IMF WEO, IMF IFS, or proxied by real industrial
production growth from the IMF IFS to extend the sample coverage. The availability of the control
variables restricts our initial sample of 178 countries to an unbalanced panel of 146 countries covering the
1984-2009 period. This dataset covers countries at a wide range of levels of economic development. 17
2.2. Descriptive statistics
Our identification strategy relies on the fact that before the occurrence of BBS shocks, economies
have ex-ante different degrees of exposure to those external funding shocks. We test for two main factors
of exposure: the debt to foreign banks as a share of GDP, and the short-term debt to foreign banks as a
share of GDP (short-term debt being defined as debt with residual maturity below 1 year). Table 2 provides
summary information on these indicators. These measures of exposure to BBS shocks show important
cross-sectional variation: a simple analysis of variance shows e.g. that nearly 63% of the variation in the
foreign bank debt variable is cross-sectional (Panel A). The pair-wise correlations between the foreign
bank and short-term bank debt variables are particularly high (0.98). This indicates that those measures
capture similar vulnerabilities, though short-term debt to foreign banks appears much more volatile within
countries than the overall debt to foreign banks. By contrast, the credit-to-deposits ratio, the leverage of the
banking system and the banking system liquidity reserves present low correlations with the debt to foreign
banks variables that characterise the exposure to external BBS shocks (Panel B). These three variables
present mostly cross-sectional variation. Finally, countries tend to experience a similar number of systemic
banking crises over the sample period, with important variation in their timing. While banking crises
mainly affect emerging economies during the 1980s and 1990s, the 2008-2009 global financial crisis led to
systemic banking crises mostly in advanced economies. Similarly, given the aggregation at the creditor
country level, bank balance-sheet shocks are strongly correlated across countries, and present most of their
variation over time. Furthermore, bank balance-sheet shocks do not appear contemporaneously correlated
with the measures of vulnerability to those shocks. For example, the correlation between the foreign bank
debt variable and BBS shocks is close to zero (0.01).
17 .
Appendix Table A2 provides descriptive statistics about the control variables and Table A3 lists the
countries included in the empirical analysis.
13
Table 2. Summary statistics 1984-2009
Panel A.
Variable
Mean
Standard deviation
Share of between country variance
Year of first observation
# Observations
# Countries
Panel B.
Onset of systemic banking crises
Bank balance-sheet shocks
Debt to foreign banks / GDP
Short-term debt to ext. banks / GDP
Bank deposits / credit
Bank assets / capital
Bank liquid reserves / assets
Onset of
systemic
banking
crises
(1)
0.26
0.16
0.02
1984
4806
178
Bank
balancesheet
shocks
(7)
0.00
0.02
0.02
1984
4355
173
1.00
0.01
0.01
0.00
0.13
-0.04
-0.09
1.00
0.01
0.02
0.03
0.03
-0.00
Summary Statistics
Debt to
Short-term
Bank
foreign
debt to ext.
deposits
banks /
banks /
/ credit
GDP
GDP
(2)
(3)
(4)
0.48
0.22
1.46
3.45
1.27
1.14
0.63
0.35
0.84
1984
1984
1984
3980
3585
4234
168
168
173
Pair-wise linear correlations
1.00
0.98
0.26
0.05
-0.22
1.00
0.30
0.05
-0.20
1.00
-0.04
0.40
(5)
12.67
5.56
0.50
2000
904
103
Bank
liquid
reserves /
assets
(6)
0.18
0.18
0.80
1984
1074
112
1.00
-0.28
1.00
Bank
assets /
capital
Note: The indicators cover at most the period 1983-2009. The ratio of bank assets over capital is only available after 2000.
Observations are country-year in Panels A and B.
Source: BIS banking statistics, Lane and Milesi-Ferretti (2007), WB WDI (2011) and Beck et al. (2009).
3. Main empirical findings
3.1. Bank balance-sheet shocks and likelihood of banking crisis
Table 3 reports the results of our baseline specification (6) and our preferred first-differenced
specification (7). We only report the estimates of the difference-in-differences specification and the
estimates of the effect of the control variables are reported in the Appendix. In order to ensure that BBS
shocks do not just proxy for trade integration and trade shocks, all specifications reported in Table 3
control both for lagged de-facto trade openness, the growth prospects of trade partners (a proxy for export
market growth), and the likely impact of export market growth on the economy. As trade integration is
strongly correlated with regional integration, these variables also proxy for unobserved regional shocks
that could amplify financial vulnerabilities.
The OLS, fixed effects and first difference estimates indicate that BBS shocks increase the likelihood
of systemic banking crises (Columns 1 to 5). In our preferred first difference specification, Column 5, for a
country with debt to foreign banks accounting for 50% of its GDP (e.g. Italy in 2007), a median-sized
negative shock to its creditors’ creditworthiness would increase crisis risk by 12% (0.4 percentage points)
through the financial-contagion channel. More specifically, the estimates would imply that, when in 2008
the rating of Ireland’s creditors was downgraded this increased the likelihood of a banking crisis in Ireland
by roughly 50% (1.5 percentage points) due to a high level of Irish exposure to foreign bank debt.
14
Table 3. Bank balance-sheet shocks and probability of banking crises
Dependent variable:
Estimator:
Debt to foreign banks / GDP
Bank balance-sheet shocks
Bank balance-sheet shocks
x Debt to foreign banks / GDP
Country fixed effects
Controls for duration(1)
Instruments(2)
Hansen (P-val)
Kleibergen-Paap (P-val)
Observations
Countries
Note:
1.
2.
Pooled OLS
(1)
0.006
(0.007)
-0.306
(0.454)
0.195
(0.372)
Start of a banking crisis
Fixed effects
First differences
(2)
(3)
(4)
(5)
0.020
0.008
0.034
0.032
(0.027)
(0.024)
(0.042)
(0.042)
0.077
0.010
0.189
0.182
(0.505)
(0.501)
(0.311)
(0.309)
0.432
0.532**
0.960***
0.951***
(0.296)
(0.269)
(0.124)
(0.122)
No
No
No
Yes
No
No
Yes
Yes
No
Yes
No
No
Yes
Yes
No
2,419
146
2,419
146
2,419
146
2,204
146
2,204
146
GMM-2S
(6)
0.401**
(0.167)
0.132
(0.340)
0.917***
(0.118)
Yes
Yes
Yes
0.24 (0.88)
7.17 (0.07)
2,055
146
Baseline controls consist of year fixed effects, the export weighted growth of main trading partners, its interaction with threeyear lagged exports over GDP, (log) GDP per capita, (log) population, domestic credit growth, the lagged level of domestic
credit over GDP, a linear and quadratic term for total assets (excluding reserves) over GDP, foreign reserves (excluding gold)
over GDP, the share of debt in external liabilities, and total liabilities over GDP. All regressors are lagged at least one period.
Standard errors are clustered at the country level. * denotes significance at the 10% level, ** at 5%, *** at 1%.
The controls for duration consist of a quadratic function in the number of years since the last crisis interacted, respectively,
with a dummy variable for having, or not having experienced a crisis since 1970.
The specification in Column 6 instruments the exposure to contagion by the first difference of bank debt over GDP in t-3 and
t-4, the bank debt over GDP in t-3 interacted with bank balance-sheet shocks, and the average of bank debt over GDP
interacted with a measure of bank balance-sheet shocks keeping the lending share of each creditor constant.
Our identification strategy hinges on the fact that countries are affected by BBS shocks independently
of their exposure to those shocks. In other words, we rely on the fact that the ex-ante reliance of borrower
countries on foreign banks does not systematically vary with the size of the shocks. Our assumption would
be violated if debtor countries anticipating shocks to creditor banks were either reducing their overall
foreign borrowing or changing creditor banks. However, the results are robust to instrumenting
(Column 6), implying that the size or structure of the exposure to BBS shocks is not driven by anticipations
of financial turmoil. We introduce two instruments for the interaction between a country’s level of
borrowing from foreign banks and the shock exposure of a country’s creditor banks. These instruments do
not use the short-run variation in debt to foreign banks as a share of GDP that may lead to some
endogeneity. The first one is the interaction between a country’s average level of borrowing from foreign
banks, a proxy for the time-invariant foreign bank dependence of each country, and our measure of BBS
shocks. The second one is the interaction between a country’s average level of borrowing from foreign
banks, and our measure of BBS shocks keeping the share of each creditor country constant over time.18
Furthermore, bank debt over GDP is instrumented by its lagged first difference in t-3 and t-4.19 In this
specification, the Hansen J statistics (p-value above 0.88) suggests that the instruments are jointly valid,
and the value of the Kleibergen-Paap statistics (7.2) shows that the instruments are unlikely to generate
weak instrument bias.20 The estimate of the instrumented specification estimated by two step-Generalized
Method of Moments (GMM-2S) is almost identical to the non-instrumented specification. Furthermore, we
18 .
More precisely, the weights, wbldrt, of equation 1 are replaced by their average over the 1984-2009 period.
19 .
The main reason to choose these instruments is that they are nearly uncorrelated (0.00) with the level of
BBS shocks interacted with the bank debt to GDP ratio.
20 .
We also estimated the model using a Limited Information Maximum Likelihood (LIML) Estimator that is
less likely to suffer from small sample bias than two-step GMM (GMM-2S) and obtained similar results
(Angrist and Pischke, 2009).
15
tested more directly the validity of our identifying assumption. The exposure to bank balance-sheet shocks
appears uncorrelated to the strength of the shocks and to the reallocation of international sources of
funding.21 This absence of reverse causation supports our preferred first-difference specification (Column
5) and is consistent with the similarity between the first difference and the GMM estimates (Columns 5-6).
Hale (2011) and Minoiu and Reyes (2011) provide complementary evidence that our assumption is likely
to hold. Neither the strength of the network of bank flows nor the degree of interconnectedness of countries
appears affected before the onset of a systemic banking crisis. This suggests that large financial shocks,
such as BBS shocks, are indeed weakly exogenous to the likelihood of systemic banking crises of
individual countries.
The obtained estimates suggest that the financial-contagion channel dominates the trade channel:
while estimates for the former are statistically and economically significant, estimates for the latter reported in the Appendix - consistently remain statistically insignificant, even though they show the
expected negative sign.22 To further assess if bank balance-sheet shocks proxy for regional or trade shocks,
we construct three placebo shocks. The first two placebo tests assume that creditor country c’s assets in
recipient country d are inversely proportional to the distance between country c and country d, or its
square. 23 The third placebo test explores if the banking channel could be confounded by the trade channel.
It assumes that creditor country c’s assets in country d are proportional to its exports as measured by the
UNCTAD dataset. For each of the three tests, regional shocks to the debtor countries are constructed as an
average of their neighbouring countries’ credit rating changes weighted by geographical closeness or
bilateral exports according to Equation 1. These regional and trade shocks are then interacted with a
country's reliance on international bank financing (as measured by the ratio of BIS bank debt over GDP)
and introduced jointly with the previously measured bank balance-sheet shocks.
Table 4 reports the results of the three placebo tests. Column 1 presents the benchmark estimates of
the impact of contagion shocks and foreign bank financing on the risk of suffering a systemic banking
crisis. Columns 2 to 4 reproduce the specification of Column 1 adding one by one the placebo shocks and
their interaction with reliance on international bank financing. Column 5 introduces jointly all the placebo
shocks and their interactions with reliance on international bank financing. The estimate of the effect of
BBS shocks on the likelihood of systemic banking crises remains stable and highly significant across the
specifications in Columns 2 to 5. This further confirms that our variable of BBS shocks is not driven by
some forms of regional shocks. Finally, Column 6 uses a less parametric strategy to control for regional
shocks. It controls for a whole range of region-year specific shocks by adding a full set of continental
dummy variables interacted with yearly dummies.24 The coefficient of the impact of BBS shocks declines
slightly but remains significant at the 1% level. Noticeably, in Columns 2 to 6, while the magnitude of the
interaction term between BBS shocks and exposure to them remains highly significant, the magnitude of
the estimate for BBS shocks decreases from 0.18 to estimates ranging from 0.03 to 0.09. This means that
while BBS shocks may partly proxy for unobservable regional shocks the initial difference-in-differences
strategy takes already into account this potential bias. In other words, the magnitude of regional shocks is
21 .
Results not reported. We compute a simple proxy for the reallocation of creditor countries, the linear
correlation between the end of period shares of creditor countries in year t and t-1. In the notations of
Equation 1, we compute the correlations between wbldrt and wbldrt-1, for each debtor country, d, and year, t.
This index takes higher values if the pattern of foreign borrowing is more stable over time.
22.
The trade channel becomes significant at the 10% level if the interaction term is included without the noninteracted measure of exposure to trade shocks as in Ilzetzki and Vegh (2008). Appendix Table A5
provides the estimates for the control variables of Table 1.
23.
Geodesic distances between capitals are from the CEPII database (Mayer and Zignano, 2011).
24.
We use Mayer and Zignano (2011) classification in five continents.
16
not systematically correlated with our measure of vulnerability to BBS shocks, namely the bank debt to
GDP ratio.
Table 4. Falsification tests: bank balance-sheet shocks versus regional and trade shocks
Dependent variable:
Estimator:
(1)
Debt to foreign banks / GDP
Bank balance-sheet shocks
Bank balance-sheet shocks
x Debt to foreign banks / GDP
Placebo (distance)
Placebo (distance squared)
Placebo (trade)
Continent * year dummies
Observations
Countries
Note:
Start of a banking crisis
First differences linear probability model
(2)
(3)
(4)
(5)
(6)
0.032
(0.042)
0.182
(0.309)
0.951***
(0.122)
0.029
(0.045)
0.040
(0.321)
1.436***
(0.222)
0.034
(0.043)
0.078
(0.322)
1.027***
(0.145)
0.037
(0.044)
0.060
(0.330)
0.892***
(0.326)
0.046
(0.043)
0.095
(0.330)
0.663*
(0.347)
0.034
(0.040)
0.030
(0.321)
0.646***
(0.163)
No
No
No
No
2,204
146
Yes
No
No
No
2,148
146
No
Yes
No
No
2,148
146
No
No
Yes
No
2,139
145
Yes
Yes
Yes
No
2,139
145
No
No
No
Yes
2,204
146
Baseline controls consist of year fixed effects, the export weighted growth of main trading partners, its interaction with threeyear lagged exports over GDP, (log) GDP per capita, (log) population, domestic credit growth, the lagged level of domestic
credit over GDP, a linear and quadratic term for total assets (excluding reserves) over GDP, foreign reserves (excluding gold)
over GDP, the share of debt in external liabilities, and total liabilities over GDP. The controls for duration consist of quadratic
function in the number of years since the last crisis interacted, respectively, with a dummy variable for having, or not having
experienced a crisis since 1970. All regressors are lagged at least one period. Standard errors are clustered at the country
level. * denotes significance at the 10% level, ** at 5%, *** at 1 %.
As it has been argued that Emerging Market Economies are more vulnerable to external shocks than
developed countries (McGuire and Tarashev, 2008), we replicated each estimation presented in Table 3
reducing the sample to Emerging Market Economies. We used the IMF WEO definition of Emerging
Market Economies to separate Emerging and Developed Economies (Duttagupta et al., 2011). The results
from this estimation led to similar estimates as those presented in Table 1, and are therefore only reported
in Appendix Table A7. A likely reason for the similarity of the point estimates is that the exposure to bank
balance-sheet shocks is not systematically correlated with countries’ level of development.
A last concern is that the constructed BBS contagion shocks could capture risks to financial stability
that are transmitted through other channels than the banking system. To examine whether the degradation
of creditor’s bank balance-sheets could be transmitted by other financial channels than liabilities towards
banks, we replace debt to foreign banks by, respectively, a country's external debt, equity and FDI
liabilities. Underlying data for these alternative measures of exposure are taken from the 2009 update of
the External Wealth of Nation (Lane and Milesi-Ferreti, 2007). Table 5 consequently assesses if BBS
shocks are transmitted by other financial channels than global banking. Column 1 reports the baseline
estimate. In Columns 2 to 4 we add interaction terms between BBS shocks and a country's external debt,
equity and FDI liabilities as a share of GDP. In Column 5 we introduce the four interaction terms jointly in
the crisis regression. The four specifications confirm that deteriorations in creditor banks’ balance sheets
are indeed transmitted through the bank debt channel, and only through that channel. The point estimate
for the BBS shock is stable, while the estimated effects of the three shocks constructed with other forms of
international financial integration are close to zero and economically and statistically insignificant. Finally,
in Column 6 we report a further robustness test. Even though – following McGuire and Tarashev (2008)
and Takàts (2010) - consolidated bank lending is conceptually the relevant metric of exposure to external
shocks, we calculate debt to foreign banks also on a locational basis. Adding an interaction term between
17
this variable and BBS shocks to the regression hardly affects our main point estimates for consolidated
lending from foreign banks despite the high correlation between the two measures.25
Table 5. Bank debt-to-GDP ratio and other forms of exposure to bank balance-sheet shocks
Dependent variable:
Estimator:
(1)
Debt to foreign banks / GDP
Bank balance-sheet shocks
Bank balance-sheet shocks
x Debt to foreign banks / GDP
Debt Shocks
Equity Shocks
FDI Shocks
Locational bank shocks
Observations
Countries
Note:
0.032
(0.042)
0.182
(0.309)
0.951***
(0.122)
Start of a banking crisis
First differences linear probability model
(4)
(5)
(2)
(3)
0.040
(0.039)
-0.118
(0.314)
-1.236***
(0.279)
0.045
(0.039)
-0.143
(0.315)
-1.094***
(0.295)
0.033
(0.042)
-0.159
(0.314)
-1.061***
(0.319)
0.041
(0.036)
-0.236
(0.360)
-1.224***
(0.288)
Yes
Yes
Yes
Yes
2,204
146
2,201
146
Yes
Yes
(6)
0.028
(0.042)
-0.152
(0.310)
-1.070***
(0.318)
Yes
2,204
146
2,204
146
2,201
146
2,204
146
Baseline controls consist of year fixed effects, the export weighted growth of main trading partners, its interaction with threeyear lagged exports over GDP, (log) GDP per capita, (log) population, domestic credit growth, the lagged level of domestic
credit over GDP, a linear and quadratic term for total assets (excluding reserves) over GDP, foreign reserves (excluding gold)
over GDP, the share of debt in external liabilities, and total liabilities over GDP. The controls for duration consist of quadratic
function in the number of years since the last crisis interacted, respectively, with a dummy variable for having, or not having
experienced a crisis since 1970. All regressors are lagged at least one period. Standard errors are clustered at the country
level. * denotes significance at the 10% level, ** at 5%, *** at 1 %.
3.2. Banking-debt maturity and external funding risk
External funding risk may not only depend on the size of foreign bank debt, but also on its maturity
structure. A sudden inability to refinance external funding positions may force borrowers to liquidate assets
not only earlier than planned, but typically in distressed market conditions, and resulting losses may render
them insolvent. For a given level of external debt, refinancing needs rise with a shorter maturity structure
of the outstanding debt. Consequently, short-term borrowing from foreign banks may pose external
funding risk beyond the size of total debt to foreign banks. In order to test for this mechanism, we interact
bank balance-sheet shocks with exposure to short-term and long-term debt to foreign banks:
cit   Shocksit 1  1Exp(short )it 1   2 Exp(long )it 1 
 1Exp( short )it 1  Shocksit 1   2 Exp(long )it 1  Shocksit 1  X it 1  i   t   it ,
(10)
where Exp(short)it-1, and Exp(long)it-1 are the debt exposures to foreign bank with, respectively,
remaining maturity below and above one year as a share of GDP.
However, short-term debt could be a coincident indicator rather than a cause of pending financial
instability, as the short-term exposure to foreign bank debt may also reflect lenders’ concerns about the
borrower country's deteriorating financial situation. The maturity structure of the debt may therefore be
25 .
The linear correlation between consolidated bank debt over GDP and locational bank debt over GDP is
0.98.
18
endogenous to the likelihood of systemic banking crisis, and to avoid endogeneity problems in the
econometric analysis we use two alternative instrumental variable strategies.
Table 6 reports estimation results for Equation 10, which is similar to Equation 7 with short-term debt
to foreign banks as a share of GDP being an additional factor of exposure to BBS-driven contagion. These
specifications examine if countries with higher external funding needs due to a debt structure that is biased
towards short-term bank debt have been more affected by BBS-driven contagion shocks than countries exante less reliant on short-term funding by foreign-owned banks. We take particular care of the likely
endogeneity of this ratio described above. Our baseline specification includes the ratio of short-term bank
debt lagged one year, and we use two sets of instruments to assess its robustness. First, we use further lags
of the same ratio and the average country value interacted with bank balance-sheet shocks.26 Second, we
use long-term debt obligations that mature over time. More specifically, we use the ratios of debt with
remaining maturity between one and two year over GDP at the end of year t-3 as instrument for short-term
debt in t-1. The latter instruments are close to the ones used at the bank level by Benmelech and Dvir
(2011). A simple rationale for this instrument is that the amount of debt scheduled to mature in t-1 and
originated in t-3 or before when information about the deterioration of the debtor countries deteriorating
situation were arguably unavailable.
Table 6. Maturity of bank debt, bank balance-sheet shocks and probability of banking crises
Dependent variable:
Start of a banking crisis
OLS
(1)
0.036
(0.038)
First differences
GMM-2S
(2)
-0.036
(0.165)
GMM-2S
(3)
0.418
(0.660)
0.210
(0.311)
1.810***
(0.363)
0.319
(0.324)
1.558***
(0.253)
0.313
(0.459)
1.763***
(0.534)
No
Yes
2.04 (0.36)
3.90 (0.27)
2,018
146
Yes
0.05 (0.98)
2.02 (0.57)
1,617
141
Estimator:
Short-term debt to foreign banks / GDP
Long-term debt to foreign banks / GDP
Bank balance-sheet shocks
Bank balance-sheet shocks
x Short-term debt to ext. banks / GDP
Bank balance-sheet shocks
x Long-term debt to ext. banks / GDP
Instruments1
Hansen (P-val)
Kleibergen-Paap (P-val)
Observations
Countries
Note:
1.
2,172
146
OLS
(4)
0.036
(0.031)
0.052
(0.128)
0.237
(0.323)
1.527
(1.156)
0.050
(0.174)
Yes
2,172
146
First differences
GMM-2S
GMM-2S
(5)
(6)
0.063
-0.066
(0.380)
(0.128)
0.185
0.169
(0.136)
(0.118)
0.316
0.602
(0.333)
(0.404)
1.645**
1.056*
(0.765)
(0.552)
0.059
0.149
(0.231)
(0.242)
Yes
3.49 (0.32)
5.49 (0.24)
1,926
146
Yes
3.63 (0.82)
11.49 (0.18)
1,611
142
Baseline controls consist of year fixed effects, the export weighted growth of main trading partners, its interaction with threeyear lagged exports over GDP, (log) GDP per capita, (log) population, domestic credit growth, the lagged level of domestic
credit over GDP, a linear and quadratic term for total assets (excluding reserves) over GDP, foreign reserves (excluding gold)
over GDP, the share of debt in external liabilities, and total liabilities over GDP. The controls for duration consist of quadratic
function in the number of years since the last crisis interacted, respectively, with a dummy variable for having, or not having
experienced a crisis since 1970. All regressors are lagged at least one period. Standard errors are clustered at the country
level. * denotes significance at the 10% level, ** at 5%, *** at 1 %.
The specification in Columns 2 and 5 instruments the lagged share of short-term debt to foreign banks in GDP by its lags in t3 and t-4. The specification in Columns 2 and 5 instruments the the lagged share of long-term debt to foreign banks in GDP
by the long term maturing debt to foreign banks with residual maturity between 1 and 2 years in t-2 and t-3.
Columns 1 to 3 concentrate on short-term debt, excluding bank debt with residual maturity above one
year. Short-term bank debt appears to propagate BBS shocks. Column 1, the point estimates of the
26 .
This arguably uses only the fixed part of short-term debt that may be due to regular roll-over of maturing
debt.
19
interaction terms of BBS shocks and is almost two times larger for short-term bank debt than for overall
bank debt (Tables 1-3). Column 2 treats the possible endogeneity of the maturity structure of foreign bank
debt by using further lags of the debt to GDP ratio, while the last specification in Column 3 uses the longterm debt obligations that matured in t-1 as a source of exogenous variation to instrument short term debt.
The results of the instrumented specifications are economically close to and statistically indistinguishable
from the baseline estimate of Column 1.27 To further confirm that short-term bank debt is the main factor
of risk when countries are hit by bank balance-sheet shocks, the specifications in Columns 4-6 introduce
jointly the ratio of short-term debt over GDP and the ratio of long-term debt over GDP. Foreign short-term
bank debt appears much more risky than long-term foreign bank debt in case of BBS shocks, as the
increased external funding risk further amplifies contagion shocks, while the interaction term between
long-term bank debt over GDP and BBS shocks appears close to zero and statistically insignificant.
These results partly confirm macroeconomic level correlations, which have shown that higher shortterm debt was associated with a higher probability of large capital-account reversals (Rodrik and Velasco,
1999, and Radelet and Sachs, 1998). By contrast, Bleakley and Cowan (2010) and Benmelech and Dvir
(2011) find no evidence of an additional negative effect of maturity mismatches and short-term debt using
microeconomic firm and bank data, respectively. More precisely, Benmelech and Dvir (2011) find that
banks’ exposure to short-term debt does not predict bank failures during the 1997-1998 Asian financial
crisis. Our estimates show that higher short-term borrowing from foreign banks increases crisis risk when
countries are hit by contagion shocks. Besides the larger sample of external bank balance-sheet shocks
considered in the estimations in this paper, one important difference is that Benmelech and Dvir consider
the Asian crisis as a uniform shock over five countries, while the empirical specifications presented here
rely on country-specific external funding shocks that are determined ex-ante by the international funding
structure of each countries’ banking system. However, estimated effects are in line with findings of
Cetorelli and Goldberg (2011) who, using the last financial crisis as a natural experiment, show that
emerging markets that were exposed to creditors with higher short-term US dollar funding needs suffered
larger contractions in cross-border lending by foreign banks during the crisis.
3.3. Lending-country spillovers and common-creditor contagion
International banks may transmit domestic shocks as well as third-party debtor country shocks to their
debtor countries. Table 7 reports the estimates when decomposing BBS shocks into lending-country
spillovers and common-creditor contagion shocks. Both types of shocks may be expected to increase the
likelihood of banking crises in the borrowing country. Indeed, larger exposure to CCC and LCS shocks are
found to increase the risk of systemic banking crises (Table 5, Columns 1-3). A median-sized negative
LCS and CCC shock to creditor countries’ perceived creditworthiness would increase the likelihood of a
banking crisis by, respectively, 14% (0.5 percentage points) and 75% (2.3 percentage points) in an OECD
country with a median level of exposure to international banking (debt to foreign banks of 53% of GDP;
based on Columns 2 and 3). A possible concern is that if some region relies on a similar set of creditor
banks, common-creditor shocks may partly proxy for regional shocks. For example, if Latin American
countries rely predominantly on foreign bank flows channelled through US banks and US banks are
overexposed to Latin American assets, our measure of CCC would be correlated with regional shocks.
While the baseline specification of Table 7 controls for trade shocks, we account for regional shocks using
region interacted with year dummies in Columns 4 to 6. The CCC shock related estimates are not sensitive
to the inclusion of these controls for regional shocks, providing further evidence that our CCC shocks are
genuine and not simply proxies for regional shocks.
27 .
As short-term debt is measured with error (a residual share of foreign bank debt has unknown maturity), it
is possible that the attenuation bias, due to measurement error, and endogenity bias partly cancel each other
out in the non instrumented specification.
20
Table 7. Common-creditor contagion, lending-country spillovers and probability of banking crises
Dependent variable:
Estimator:
Debt to foreign banks / GDP
Lending-country spillovers
Lending-country spillovers
x Debt to foreign banks / GDP
Common-creditor contagion shocks
(1)
0.044
(0.037)
0.132
(0.110)
0.908***
(0.163)
Common-creditor contagion shocks
x Debt to foreign banks / GDP
Continent x year dummy variables
Observations
Countries
Note:
No
2,148
146
Start of a banking crisis
First differences linear probability model
(2)
(3)
(4)
(5)
0.032
0.047
0.040
0.035
(0.038)
(0.037)
(0.039)
(0.038)
0.135
-0.034
(0.111)
(0.108)
0.735***
0.502***
(0.183)
(0.141)
2.211*
2.344*
2.338*
(1.180)
(1.325)
(1.293)
1.306***
0.904***
0.983***
(0.196)
(0.235)
(0.187)
No
2,204
146
No
2,204
146
Yes
2,148
146
Yes
2,148
146
(6)
0.043
(0.040)
-0.011
(0.112)
0.319*
(0.170)
2.634*
(1.382)
0.889**
(0.387)
Yes
2,148
146
Baseline controls consist of year fixed effects, the export weighted growth of main trading partners, its interaction with threeyear lagged exports over GDP, (log) GDP per capita, (log) population, domestic credit growth, the lagged level of domestic
credit over GDP, a linear and quadratic term for total assets (excluding reserves) over GDP, foreign reserves (excluding gold)
over GDP, the share of debt in external liabilities, and total liabilities over GDP. The controls for duration consist of quadratic
function in the number of years since the last crisis interacted, respectively, with a dummy variable for having, or not having
experienced a crisis since 1970. All regressors are lagged at least one period. Standard errors are clustered at the country
level. * denotes significance at the 10% level, ** at 5%, *** at 1 %.
The measured effects of common-creditor shocks on the likelihood of systemic banking crises are
similar to those of BBS shocks (Table 7 Column 2 and Table 3 Column 5). As common-creditor shocks are
much less likely to be endogenous to a particular debtor-creditor relationship than bank balance-sheet
shocks, the similarity of the point estimates between the two different types of bank shocks provides
additional empirical support for the causal interpretation of the impact of BBS shocks on the likelihood of
systemic banking crises. Furthermore, the impact of CCC shocks appears somewhat larger than that of
LCS shocks (Table 7, Columns 3 and 6).28 The larger impact of CCC shocks may partly be driven by a lack
of precise information about debtor countries that could amplify the (perceived) need for balance-sheet
adjustment of the affected creditor banks. For example, Calvo (1998) proposes a “lemons” model in which
- in the wake of a shock - investors trying to sell their assets differentiate depending on the amount of
information they have about them, accepting larger price discounts for assets in debtor countries about
which they have less information.
Overall, our main empirical results provide strong evidence that financial contagion is an important
and general mechanism to explain the occurrence of banking crises, with BBS-intermediated contagion
shocks being amplified by exposure to borrowing from foreign banks. Though our preferred sample
include all available countries and years, we checked that they were not driven by a subset of particular
countries such as international financial centres or the 2008-2009 global financial crisis. The main results
are nearly identical when we restrict our sample to developing and emerging economies or when we omit
the 2007-2009 global financial crisis (Appendix Tables Table A.2). Our main results are also consistent
with those of Kaminsky and Reinhart (2001) for the Asian (1997-98) and the Mexican (1994-95), or with
the recent study by Cetorelli and Goldberg (2011) for the 2007-09 global financial crisis, although these
two papers only focus on a few selected events and on domestic credit shortages rather than banking crises.
Strong contagion effects through the bank balance-sheet channel are also in line with results from
microeconometric analysis. Khwaja and Mian (2008) find that liquidity shocks to banks are transmitted to
28 .
The mean and the standard deviation of CCC and LCS shocks are roughly similar.
21
their borrowing firms, with small unconnected firms in particular being unable to hedge and facing
particularly strong financial distress. Hale (2011) uses bank-level data on syndicated loans between 1980
and 2009 to show that banks at the periphery of the banking network (in the sense of being the ultimate
debtors among banks and not intermediaries) are most strongly affected by banking crises as their access to
bank lending gets curtailed and they often cannot roll-over their maturing debt.
4. Factors shaping the impact of bank balance-sheet shocks
4.1. Banking sector vulnerabilities and bank balance-sheet shocks
We next turn to a formal investigation of how vulnerabilities in the banking sector shape the impact of
BBS shocks on the likelihood of systemic banking crises. For example, countries with more leveraged
banking systems may be more vulnerable when hit by a BBS shock. We therefore introduce an interaction
term between banking sector leverage (measured as the ratio of total assets to bank capital and reserves)
and the exposure-scaled BBS shocks into Equation (6):
cit   Shocksit 1   Expit 1   Levit 1   1   2 Levit 1   Expit 1  Shocksit 1  X it 1  i   t   it
(11) where the exposure-scaled BBS shocks correspond to the interaction between the ratio of bank debt to
GDP and the BBS shocks.
More generally, domestic banking sector vulnerabilities other than high leverage may also affect the
vulnerability to exposure-scaled BBS shocks. We therefore replace banking sector leverage by the liquidity
of the banking sector and its reliance on domestic deposits in Equation (11). Liquidity of the banking
sector is defined as banks' domestic currency holdings and deposits with the monetary authorities as a
share of bank assets, and the reliance on domestic deposits as deposits as a share of bank credit.
Table 8 reports results for the interaction terms between exposure-scaled BBS shocks and measures of
vulnerability of debtor countries’ banking sector as specified in Equation (11). The point estimates indicate
that countries which rely more on deposit-funded credit appear less vulnerable to BBS shocks (Column 1),
likely because the latter affect deposits less than alternative sources of funding, such as cross-border
lending. 29 This is in line with bank-level evidence that domestic banks which relied more on deposits were
less affected than the affiliates of multinational banks during the global financial crisis (de Haas and van
Lelyveld, 2011).
Other forms of vulnerabilities also display the expected signs. The higher degree of bank leverage
indeed increases the likelihood of suffering a systemic banking crisis when countries are hit by BBS
shocks (Column 2).30 The point estimates suggest that, for a median exposure-scaled BBS shock during the
global financial crisis, a country's likelihood of systemic banking increases by 1.7 percentage points when
the leverage factor as measured by the asset-to-capital ratio goes from 11 to 15 (11 being the sample
median and 15 the sample third quartile). Similarly, countries where the banking system has higher
liquidity reserves appears less prone to banking crisis in the case of BBS shocks (Column 3). Despite the
limited sample size, the point estimates are relatively robust to the joint introduction of different interaction
terms in a unique regression (Columns 4-6). This suggests that the interaction terms of exposure-scaled
29.
The credit-to-deposits ratio is taken from Beck et al. (2009).
30 .
Bank leverage ratios and liquid reserves to asset ratio are taken from the World Bank WDI. De facto bank
leverage ratios are only available after 2000. Therefore, the magnitude of the estimates cannot be directly
compared to that of other specifications fitted over the 1983-2009 period.
22
BBS shocks with, respectively, banking sector leverage, liquidity, and reliance on deposits indeed capture
different forms of vulnerability to BBS shocks, rather than countries’ overall financial health.31
Table 8. Banking sector’s vulnerabilities, bank balance-sheet shocks and probability of banking crises
Dependent variable:
Estimator:
OLS
(1)
Debt to foreign banks / GDP
Bank balance-sheet shocks
Bank balance-sheet shocks
x Debt to foreign banks / GDP
x --- x Bank deposits / credit
0.035
(0.041)
0.098
(0.312)
1.751***
(0.351)
-0.337***
(0.125)
x --- x Bank assets / capital
Start of a banking crisis
First differences linear probability model
OLS
OLS
OLS
OLS
(2)
(3)
(4)
(5)
0.124
(0.135)
1.108**
(0.476)
-2.354*
(1.267)
0.170***
(0.060)
x --- x Bank liquidity reserves /
Note:
0.192
(0.138)
0.939*
(0.473)
-1.336
(0.857)
-0.907***
(0.237)
-0.242***
(0.055)
-7.893**
(3.803)
assets
Observations
Countries
-0.071
(0.064)
0.681*
(0.372)
1.208***
(0.201)
2,195
146
504
90
601
97
504
90
-0.066
(0.066)
0.617
(0.378)
1.497**
(0.754)
-0.107
(0.232)
OLS
(6)
0.073
(0.125)
1.825***
(0.635)
-1.708
(1.489)
-7.656**
(3.777)
-0.146**
(0.070)
-4.971
(8.105)
601
97
367
70
Baseline controls consist of year fixed effects, the export weighted growth of main trading partners, its interaction with threeyear lagged exports over GDP, (log) GDP per capita, (log) population, domestic credit growth, the lagged level of domestic
credit over GDP, a linear and quadratic term for total assets (excluding reserves) over GDP, foreign reserves (excluding gold)
over GDP, the share of debt in external liabilities, and total liabilities over GDP. The controls for duration consist of quadratic
function in the number of years since the last crisis interacted, respectively, with a dummy variable for having, or not having
experienced a crisis since 1970. All regressors are lagged at least one period. Standard errors are clustered at the country
level. * denotes significance at the 10% level, ** at 5%, *** at 1 %.
4.2. Global liquidity and bank balance-sheet shocks
Finally, the effect of BBS shocks may depend on the global context. BBS shocks may be magnified in
situations of tight global liquidity, as creditor banks may not be able to raise money on the interbank
market and alternative sources of funding may be unavailable for debtor countries. Indeed, simulations of
stylised models of bank balance-sheet contagion show that liquidity injections in the interbank market can
partly offset the effect of BBS shocks (Tressel, 2010). This is in line with cross-country time-series
evidence that, over the 1993-2000 period, a weakening in bank balance sheets led to a stronger reduction in
credit growth when monetary policy was tight (Nier and Zicchino, 2008). Furthermore, when domestic
monetary policy is tight, global banks may reduce the lending activity of their foreign affiliates, and the
parent banks may increase their borrowing through internal capital markets. Cetorelli and Goldberg (2008)
provide empirical evidence that inflows into US parent banks are larger when US domestic monetary
policy is tighter. Van den Heuvel (2012) shows that US states with more leveraged banks (i.e. that have a
lower capital-to-asset ratio) are more affected by federal monetary policy. Therefore, we allow the impact
of exposure-scaled BBS shocks to depend on aggregate liquidity conditions, replacing leverage Levit-1 by
global liquidity conditions in Equation (9).
31 .
We also estimated interaction term with de jure banking regulations. More stringent de jure capital
requirements according to Ahrend et al. (2011) are associated with a lower likelihood of systemic banking
crises when countries are hit by BBS contagion shocks. By contrast, the overall banking regulation did not
seem to affect the strength of BBS shocks (Ahrend and Goujard, 2011, 2012).
23
Table 9 reports the estimates of the interaction terms between exposure-scaled BBS shocks and
different measures of global liquidity conditions. The empirical specification of Equation 9 measures the
effect of global liquidity conditions on the likelihood of systemic banking crises when countries are hit by
BBS shocks, while controlling for the direct impact of global macroeconomic conditions by including time
fixed effects. Global liquidity conditions are measured using real US interest rates, a proxy of global real
interest rate, or a proxy for real global money supply based on M2, alternatively. The two latter global
variables are constructed as the GDP weighted average of the US, the Euro-zone and Japan.
Table 9. Global liquidity, bank balance-sheet shocks and probability of banking crises
Dependent variable:
Estimator:
(1)
Debt to foreign banks / GDP
Bank balance-sheet shocks
Bank balance-sheet shocks
x Debt to foreign banks / GDP
x --- x US real interest rate
0.057
(0.036)
-0.005
(0.311)
2.268***
(0.416)
0.647***
(0.135)
x --- x World real interest rate
Start of a banking crisis
First differences linear probability model
(2)
(3)
(4)
(5)
0.047
(0.038)
-0.035
(0.316)
2.073***
(0.458)
Note:
1.
0.073*
(0.042)
-0.099
(0.351)
2.023***
(0.427)
0.612***
(0.159)
0.676***
(0.233)
x --- x Global liquidity
Controls for Risk aversion (VXO)1
Observations
Countries
0.036
(0.035)
-0.041
(0.303)
0.832**
(0.368)
0.069*
(0.041)
-0.128
(0.351)
1.952***
(0.487)
0.750***
(0.241)
-0.883**
(0.375)
No
2,204
146
No
2,204
146
(6)
0.038
(0.036)
0.000
(0.342)
0.825**
(0.383)
No
2,079
146
-0.854**
(0.383)
Yes
2,025
146
Yes
2,025
146
Yes
1,900
146
Baseline controls consist of year fixed effects, the export weighted growth of main trading partners, its interaction with threeyear lagged exports over GDP, (log) GDP per capita, (log) population, domestic credit growth, the lagged level of domestic
credit over GDP, a linear and quadratic term for total assets (excluding reserves) over GDP, foreign reserves (excluding gold)
over GDP, the share of debt in external liabilities, and total liabilities over GDP. The controls for duration consist of quadratic
function in the number of years since the last crisis interacted, respectively, with a dummy variable for having, or not having
experienced a crisis since 1970. All regressors are lagged at least one period. Standard errors are clustered at the country
level. * denotes significance at the 10% level, ** at 5%, *** at 1 %.
Specifications (4) to (6) control for the overall risk aversion by including an additional term: Bank balance-sheet shocks x
(Debt to foreign banks / GDP) x VXO.
All point estimates of Table 9 have the expected signs and are statistically significant, implying that
global liquidity reduces the impact of external BBS shocks on the likelihood of banking crises. For
example, a decrease of the US real interest rate by 1 percentage point reduces the effect of BBS shocks on
the likelihood of systemic banking crises by 25 percent. A possible concern regarding this result is that
periods of global liquidity may be correlated with periods of lower risk aversion. Therefore, an interaction
term between BBS shocks and an indicator of market risk aversion, the VXO index, is introduced in the
regressions. The estimated effects of global liquidity are robust to the inclusion of this additional control
variable (Columns 4-6). 32 This is in line with cross-country time-series evidence that, over the 1993-2000
period, a weakening in bank balance-sheets led to a stronger reduction in credit growth when monetary
policy was tight (Nier and Zicchino, 2008). More generally, Cetorelli and Goldberg (2008) provide
empirical evidence that US domestic liquidity shocks can be partly smoothed when banks have access to
alternative funding, while Van den Heuvel (2012) shows that US states with more leveraged banks
(i.e. that have a lower capital-to-assets ratio) are more affected by federal monetary policy.
32.
The VXO represents the option-implied volatility of the S&P 100 stock price futures. The VXO indicator is
available starting in January 1986, compared with January 1990, for the VIX. The correlation between
these two indicators, however, is almost 0.99. As discussed by Bekaert et al. (2010), the VXO volatility
index can be interpreted as reflecting both an uncertainty component, that captures the actual expected
stock market volatility, and a component reflecting risk aversion.
24
4.3. Bank balance-sheet shocks and bank flows
The obvious channel through which BBS shocks could affect the likelihood of banking crisis is
through reversals and sudden-stops in lending from foreign banks. To further confirm the relevance of the
BBS channel, we therefore use bilateral data on international cross-border lending to directly test whether
BBS shocks affect bank capital flows. We consider bilateral annual bank flows from 30 creditor countries
towards 214 partner countries over the 1984-2010 period. The bilateral data on international bank lending
are from the Bank for International Settlements’ (BIS) Consolidated International Banking Statistics. This
database contains information on positions of banks from BIS reporting countries with respect to
counterparties around the world, with data aggregated across all reporting banks from the source countries.
The lending variable corresponds to international claims, which capture the sum of cross-border lending
and local claims by foreign affiliates.33 Thus, on-lending through subsidiaries is captured as exposure of
the original bank creditor (e.g. Japanese lending to Thailand through a Thai or Hong Kong subsidiary is
counted as Japanese lending). Similar data were used by Van Rijckeghem and Weder (2003), Mc Guire
and Tarashev (2008) and Cetorelli and Goldberg (2011) to study the effect of the 2007-09 global financial
crisis. In total, there are observations for 3,289 country pairs. Bilateral bank lending increased in average
by 7% each year over our sample period, with large variation in growth rates across countries and over
time (standard deviation of 73%). Common-creditor shocks in the bilateral setting are defined according to
Equation 9. Both bank balance-sheet shocks and bilateral bank flows have most of their variation over
time. Country-pair fixed effects explain only 11% of the variation in bank balance-sheet shocks, and 6% of
the variation in bilateral bank lending.
Table 10 reports the point estimates of Specification 8 for bank flows that control for creditor-year
fixed effects. As the main explanatory variable is the shock to creditor banks’ assets and as bank flows of
different creditor banks towards the same counterparty country could be correlated, the standard errors are
clustered at the debtor times creditor-year level. The first three columns introduce common-creditor
contagion shocks controlling for debtor-year fixed effects and either creditor country fixed effects or
creditor-debtor fixed effects. In addition, Columns 4-6 control for linear and quadratic trends, as well as
local shocks to creditor countries such as domestic GDP growth. Though these controls may be partly
determined by the strength of common-creditor contagion shocks, the robustness of the estimates for the
impact of BBS shocks on cross-border bank flows to these different specifications indicates that the
measured effect of BBS shocks is unlikely to arise from creditor countries’ domestic vulnerabilities.
In all the specifications of Table 10, the shocks on creditor banks appear with the expected negative
sign: a deterioration in creditors’ balance sheets leads to a decrease in bank flows towards their debtor
countries.34 Hence, countries that rely more on international funding appear vulnerable when creditor
banks are hit by negative shocks to their international asset holdings. Taken at face value, the estimates
imply that when the average rating of a creditor country's portfolio is downgraded by 1%, this implies –
ceteris paribus - a reduction in its bank lending to each debtor country by 2.7 to 3.7 percentage points. The
inclusion of linear and quadratic trends or GDP growth in creditor countries does not appear to
significantly reduce the effect of bank balance-sheet shocks on international lending (Columns 4-6).
Therefore, it appears highly unlikely that these bank-driven shocks proxy for real trade shocks or other
creditor-country specific shocks. Consequently, the decrease in bilateral cross-border lending appears to be
driven by financial shocks in third-party debtor countries. These results are in line with Van Rijckeghem
33 .
More precisely, the data corresponds to the BIS Table 9b, Foreign claims by nationality of reporting banks,
immediate borrower basis, from the BIS Consolidated banking statistics. Bank flows are computed as the
first difference in (log) consolidated lending.
34.
As a robustness check, we also performed the same analysis reducing the sample of debtor countries to
emerging market economies; the results reported in Appendix Table A8 are stronger and qualitatively
similar.
25
and Weder (2003) who find that exposure to Mexico and Thailand partly predicts bank flows to third-party
countries after the Mexican and Asian crises.35 More recent analysis based on microeconomic bank data
finds that international bank linkages contributed to transmit the 1998 Russian crisis to Peruvian banks
(Schnabl, 2012).
Table 10. Bank balance-sheet shocks and cross-border bank lending flows
Dependent variable:
Estimator:
(1)
Common-creditor Contagion shock
Creditor fixed effects
Creditor x debtor fixed effects
Creditor linear time trend
Creditor quadratic time trend
Creditor domestic GDP growth
# observations
# creditor*year
# debtor countries
Yearly change in (log) creditor’s assets in country d
Debtor-year fixed-effects model
(2)
(3)
(4)
(5)
(6)
-2.714***
(0.969)
-3.709***
(0.949)
-3.418***
(0.936)
-3.472***
(0.933)
-2.482**
(0.996)
-2.492**
(1.000)
No
No
No
No
No
41,303
431
211
Yes
No
No
No
No
41,303
431
211
No
Yes
No
No
No
41,303
431
211
No
Yes
Yes
No
No
41,303
431
211
No
Yes
Yes
Yes
No
41,303
431
211
No
Yes
Yes
Yes
Yes
41,303
431
211
Note:
Observations are yearly debtor-creditor flows. Standard errors are two-way clustered at the debtor and creditor*year
levels. The explanatory variable is lagged one period. * denotes significant estimate at the 10% level, ** at 5%, *** at 1%.
1. All specifications control for debtor-year fixed effects.
Conclusion
This paper investigates the effects of financial shocks transmitted by international banks on the
likelihood of systemic banking crises and proposes a new empirical strategy to identify contagion in the
context of global banking. Our empirical analysis indicates that contagion through the balance sheets of
international banks strongly affects the occurrence of systemic banking crises. One of the main
mechanisms of propagation of bank balance-sheet shocks is through cross-border lending: international
banks negatively hit by shocks to some of their debtor countries appear to reduce lending towards thirdparty countries.
The propagation of financial shocks through bank balance-sheets implies that high levels of
international banking activity – and in particular when they take the form of short-term loans - pose serious
risks to global financial stability when financial turmoil arises in some part of the global financial system.
Banking-driven financial contagion appears to have played a significant role not only during the 2007-09
global financial crisis, but also historically during periods of financial turmoil, and across different types of
countries. To further support the claim that global banking facilitates the global spreading of financial
turmoil, we provide extensive robustness checks that lend additional credibility to the causal interpretation
of our results. The size of effects is also economically important.
Finally, we show that the potential impact of bank balance-sheet shocks on a country depends both on
the structural features of its financial sector and the global macro-economic environment. The effect of
BBS shocks on the likelihood of systemic banking crisis is more pronounced for countries where the
banking sector is more highly leveraged, and has lower liquidity reserves. Similarly, bank balance-sheet
shocks tend to increase the likelihood of systemic banking crises more in countries with a greater reliance
on external finance, in particular from foreign banks. In contrast, financial crisis risk decreases when a
larger share of bank credit is funded by domestic deposits. More plentiful global liquidity also reduces the
35.
The specification of Van Rijckeghem and Weder (2003) assumes that there is an origin country for each
financial crisis while our measure of bank balance-sheet shocks is agnostic about this.
26
effect of bank balance-sheet shocks, underlining the importance of major central banks ensuring ample
international liquidity at times of financial turmoil.
More research is needed to fully understand how shocks to the balance sheets of financial
intermediaries affect debtor countries. In particular, it will be useful to confirm the findings of this paper
using data that allow for identification of specific banks rather than relying on aggregate country and
banking sector characteristics. Likewise, it would be interesting to learn more about the transmission of
shocks through the balance sheets of non-bank financial institutions.
27
BIBLIOGRAPHY
Ahrend, R., J. Arnold and F. Murtin (2011), "Have More Strictly Regulated Banking Systems Fared Better
During the Recent Financial Crisis?," Applied Economics Letters, Vol. 18(5), pp. 399-403.
Ahrend, R. and A. Goujard (2011), "Drivers of Systemic Banking Crises: The Role of Bank Balance-Sheet
Contagion and Financial Account Structure," OECD Economics Department Working Papers,
No. 902, OECD Publishing.
Ahrend, R. and A. Goujard (2012), "How do structural policies affect financial crisis risk? Evidence from
past crises across OECD and emerging economies", OECD Economics Department Working Papers,
forthcoming , OECD Publishing.
Allison, P., (1982), "Discrete-time methods for the analysis of event histories", Sociological Methodology
XII, pp. 61–98.
Angrist, J. D., (2001), "Estimations of Limited Dependent Variable Models with Dummy Endogenous
Regressors: Simple Strategies for Empirical Practice", Journal of Business and Economic Statistics,
19(1), 2-16.
Angrist, J. D. and J. Pischke, (2009), Mostly Harmless Econometrics: An Empiricist's Companion,
Princeton University Press.
Beck, T. and A. Demirgüç-Kunt and R. Levine, (2009), "Financial Institutions and Markets across
Countries and over Time", World Bank Policy Research Working Paper 4943.
Benmelech E. and E. Dvir, (2011), "Does Short-Term Debt Increase Vulnerability to Crisis? Evidence
from the East Asian Financial Crisis", NBER Working Papers 17468.
Bleakley, H. and C. Kevin, (2010), "Maturity mismatch and financial crises: Evidence from emerging
market corporations," Journal of Development Economics, Elsevier, vol. 93(2), pp. 189-205.
Broner, F. A., G. Gelos and C. M. Reinhart (2006), "When in Peril, Retrench: Testing the Portfolio
Channel of Contagion," Journal of International Economics, Elsevier, Vol. 69(1), pp. 20-230, June.
Cameron, A. C. and P. K. Trivedi, (2009), Microeconometrics using Stata, Stata Press.
Card, D., (1990), "Strikes and Wages: A Test of an Asymmetric Information Model", The Quarterly
Journal of Economics, 105(3), 625-659.
Cerutti, E., S. Claessens and P. McGuire (2011), "Systemic Risks in Global Banking: What Available Data
can tell us and What More Data are Needed?" IMF Working Papers 11/222, International Monetary
Fund.
Cetorelli, N. and L. S. Goldberg (2008), "Banking Globalization, Monetary Transmission, and the Lending
Channel," NBER Working Papers 14101, National Bureau of Economic Research.
Cetorelli, N. and L. S. Goldberg (2011), "Global Banks and International Shock Transmission: Evidence
from the Crisis," IMF Economic Review, Palgrave Macmillan, Vol. 59(1), pp. 41-76, April.
28
de Haas, R. and I. van Lelyveld I. (2006), "Foreign Banks and Credit Stability in Central and Eastern
Europe. A Panel Data Analysis", Journal of Banking and Finance, Vol. 30, Issue 7, pp. 1927-1952.
de Haas, R. and N. van Horen (2011), "Running for the Exit: International Banks and Crisis Transmission,"
DNB Working Papers 279, Netherlands Central Bank, Research Department.
de Haas, R. and I. van Lelyveld (2011), "Multinational Banks and the Global Financial Crisis. Weathering
the Perfect Storm?" DNB Working Papers 322, Netherlands Central Bank, Research Department.
Demirgüç-Kunt, A. and E. Detragiache, (1998), The determinants of Banking Crises in Developing and
Developed Countries, IMF staff paper, 45 (1), 81-109.
Demirgüç-Kunt, A. and E. Detragiache, (2005), "Cross-Country Empirical Studies of Systemic Bank
Distress: A Survey", IMF Working Paper, No. 0596.
De Ree, J. and E. Nillesen (2009), "Aiding violence or peace? The impact of foreign aid on the risk of civil
conflict in sub-Saharan Africa", Journal of Development Economics, 88, 301–313.
Devereux, M. B. and J. Yetman, (2010), "Leverage Constraints and the International Transmission of
Shocks," Journal of Money, Credit and Banking, vol. 42(s1), pp. 71-105.
Duttagupta, R., J. Bluedorn, J. Guajardo and P. Topalova, (2011), "International Capital Flows: Reliable or
Fickle? ", IMF World Economic Outlook, Chapter 4.
Eichengreen, B. and A. Mody (2000), "What Explains Changing Spreads on Emerging Market Debt?,"
NBER Chapters, in: Capital Flows and the Emerging Economies: Theory, Evidence, and
Controversies, pp. 107-136.
Falcetti, E. and M. Tudela, (2008), "What do Twins Share? A Joint Probit Estimation of Banking and
Currency Crises", Economica, 75, 199-221.
Greenwood, R., A., Landier and D. Thesmar, (2011), "Vulnerable Banks", Harvard Business School
Working paper.
Griliches, Z. and J. A. Hausman, (1986), "Errors in variables in panel data", Journal of Econometrics,
31(1), 93-118.
Hale G., (2011), "Bank Relationships, Business Cycles, and Financial Crises," NBER Working Papers
17356, National Bureau of Economic Research, Inc.
Hallak, I., (2011), "External Debt to the Private Sector and the Price of Bank Loans", University of Bocconi
Working Paper.
Hyslop, D. R., (1999), State Dependence, Serial Correlation and Heterogeneity in Intertemporal Labor
Force Participation of Married Women, Econometrica, 67(6), 1255-1294.
Ilzetzki, E. and C. A. Végh, 2008. "Procyclical Fiscal Policy in Developing Countries: Truth or Fiction?,"
NBER Working Papers 14191, National Bureau of Economic Research, Inc.
Jiménez G., S. Ongena, J-L. Peydró and J. Saurina, (2010), "Credit Supply: Identifying Balance-sheet
Channels with Loan Applications and Granted Loans," Banco de España Working Papers 1030,
Banco de España.
29
Kalemli-Ozcan, S., E. Papaioannou and J. L. Peydro (2012), "Financial Regulation, Financial
Globalization and the Synchronization of Economic Activity", Journal of Finance, forthcoming.
Kaminsky, G. L. and C. M. Reinhart (2001), "Bank Lending and Contagion: Evidence from the Asian
Crisis," NBER Chapters, in: Regional and Global Capital Flows: Macroeconomics Causes and
Consequences, NBER-EASE Vol. 10, pp. 73-116.
Kaminsky, G., C. M. Reinhart and C. A. Vegh (2003), "The Unholy Trinity of Financial Contagion,"
Journal of Economic Perspectives, vol. 17(4), pp. 51-74.
Khwaja, A. and A. Mian (2008), "Tracing the Impact of Bank Liquidity Shocks: Evidence from an
Emerging Market", American Economic Review, vol. 98(4), pp. 1413-42.
Krugman, P. (2008), "The International Finance Multiplier", Princeton University mimeo.
Laeven, L. and F. Valencia, (2008), “Systemic Banking Crises: A New Database”, IMF Working Paper,
No. 08224.
Laeven, L. and F. Valencia, (2010), Resolution of Banking Crises: The Good, the Bad, and the Ugly, IMF
Working Paper, No. 10146.
Laeven, L. and F. Valencia, (2012), Systemic Banking Crises Database: An Update, IMF Working Paper,
No. 12163.
Lane, P. R. and G. M. Milesi-Ferretti (2007), "External Wealth of Nations Mark II: Revised and Extended
Estimates of Foreign Assets and Liabilities 1970-2004", Journal of International Economics, 73,
pp. 223-250.
Mayer, T. and S. Zignago, (2011), "Notes on CEPII’s distances measures: The GeoDist database," CEPII
Working Papers 2011-25.
McGuire, P. and Tarashev N., (2008), "Bank Health and Lending to Emerging Markets," BIS Quarterly
Review, Bank for International Settlements, December.
Minoiu, C. and J. A. Reyes, (2011), "A network analysis of global banking:1978-2009," IMF Working
Papers 11/74, International Monetary Fund.
Nier E., and L. Zicchino (2008), "Bank Losses, Monetary Policy and Financial Stability-Evidence on the
Interplay from Panel Data", IMF Working Papers 08/232, International Monetary Fund.
Peek, J. and E. S. Rosengren, (2000a), "Collateral Damage: Effects of the Japanese Bank Crisis on Real
Activity in the United States," American Economic Review, American Economic Association, vol.
90(1), pp. 30-45.
Peek, J. and E. S. Rosengren, (2000b), "Implications of the globalization of the banking sector: the Latin
American experience," New England Economic Review, Federal Reserve Bank of Boston, issue Sep,
pages 45-62.
Radelet, S. and J. D. Sachs (1998). "The East Asian Financial Crisis: Diagnosis, Remedies, Prospects",
Vol. 1998, No. 1, Brookings Papers on Economic Activity, pp. 1-74.
30
Reinhart, C. M. and K. S. Rogoff, (2009), This Time Is Different: Eight Centuries of Financial Folly,
Princeton University Press.
Reinhart, C. M. and K. S. Rogoff, (2011), "From Financial Crash to Debt Crisis." American Economic
Review, 101(5): 1676–1706.
Rodrik, D., Velasco, A., (1999), “Short-term capital flows”, NBER Working Paper No. W7364.
Schnabl, P. (2012), "The International Transmission of Bank Liquidity Shocks: Evidence from an
Emerging Market", Journal of Finance, 67(3).
Stewart M. B. (2007), "The interrelated dynamics of unemployment and low-wage employment," Journal
of Applied Econometrics, John Wiley & Sons, Ltd., vol. 22(3), pages 511-531.
Takàts, E. (2010), "Was it Credit Supply? Cross-border Bank Lending to Emerging Market Economies
during the Financial Crisis", BIS Quarterly Review, June 2010, pp. 49-56.
Tirole J. (2011), "Illiquidity and All Its Friends," Journal of Economic Literature, American Economic
Association, Vol.49(2), pp. 287-325, June.
Tressel T. (2010), "Financial Contagion through Bank Deleveraging: Stylized Facts and Simulations
Applied to the Financial Crisis," IMF Working Papers 10/236, International Monetary Fund.
Van den Heuvel, S. J. (2012). "Banking Conditions and the Effects of Monetary Policy: Evidence from
U.S. States," The B.E. Journal of Macroeconomics: Advances.
Van Rijckeghem, C. and B. Weder (2001), "Sources of Contagion: is it Finance or Trade?" Journal of
International Economics, Elsevier, Vol. 54(2), pp. 293-308, August.
Van Rijckeghem, C. and B. Weder, (2003), "Spillovers through banking centers: a panel data analysis of
bank flows," Journal of International Money and Finance, Elsevier, vol. 22(4), pp. 483-509.
Vegh, C. (forthcoming), Open Economy Macroeconomics in Developing Countries, The MIT Press.
Wooldridge, J. M., (2002), Econometric Analysis of Cross Section and Panel Data, The MIT Press.
31
Table A1. Description of the main variables used in the empirical analysis
Measure
Financial crisis start
Share of debt in external
liabilities and financial
account structure
Assets, reserves and
liabilities as a share of GDP
GDP per capita
Population
Credit and credit growth,
credit-to-GDP ratio and
credit-to-deposit ratio
Bank debt
Short-term bank debt
Bilateral bank positions
Openness to trade
Capital adequacy
requirements and overall
banking regulation
Credit rating
Bank balance-sheet shocks
(BBS shocks)
Growth of real GDP in
country trading partners
US real interest rates
World real interest rates
Definition and source
Systemic banking crises are defined according to Laeven and Valencia (2008,
2010) with data taken from the 2010 update.
The share of debt liabilities in total external liabilities and financial account
variables are taken from Lane and Milesi-Ferretti (2007), “External Wealth of
Nations” dataset, 1970-2007 (updated in 2009).
The ratio of assets, liabilities and reserves over GDP are taken from Lane and
Milesi-Ferretti (2007), “External Wealth of Nations” dataset, 1970-2007.
GDP in current US dollars, taken from Lane and Milesi-Ferretti (2007),
“External Wealth of Nations” dataset, 1970-2007. Population as defined below.
The primary source is the WB WDI. Missing values are filled with data from the
IMF IFS and, subsequently, from the IMF WEO when available.
Private credit by deposit money banks and other financial institutions over
GDP. All data are taken from Beck et al. (2009). Data updated in November
2010.
Bank debt is measured by the debt liabilities towards BIS reporting banks (BIS
Table 9, Variable A). The BIS consolidated banking statistics (on the
immediate borrower basis) report banks' on-balance-sheet financial claims on
the rest of the world and provides a measure of the risk exposures of lenders'
national banking systems.
Short-term bank debt is measured by the debt liabilities towards BIS reporting
banks with residual maturity below one year (BIS Table 9, Variable B). The BIS
data report the residual maturity (not the original maturity) of bank debt,.
Locational and consolidated bank positions come from the banking statistics of
the Bank of International Settlements (BIS).
The sum of imports and exports divided by GDP from the WB WDI. Missing
values are completed with data from the IMF-IFS.
Banking regulations such as capital adequacy requirements and overall
banking regulation are computed as the average of the World bank surveys
(1998-2006) and aggregated according to Ahrend et al. (2011). More precisely,
the indicator of de jure capital requirements includes the regulation of minimum
capital to asset ratios, variation of capital to asset ratios according to individual
banks’ credit market or operational risks, the application of simple leverage
ratios, and the acceptability of subordinated debt and of revaluation gains as
capital. Overall banking regulation takes into account requirements related to
liquidity and diversification, capital, accounting and provisioning, external
auditing and information disclosure, entry rules and ownership structure, exit
rules and disciplining devices, depositor protection and the authority of the
banking supervisor.
Based on ratings from “Institutional Investor”. Published each March and
September, these ratings are based on a survey of international bankers, who
assign a numerical value ranging from 0 to 100 (with 100 indicating zero
probability of default).
This index is calculated based on changes in credit ratings (see above) and
bilateral assets position of the BIS reporting banks on a locational basis. The
locational banking statistics gather quarterly data on international financial
claims and liabilities of banks in the BIS reporting countries.
Export-weighted average of GDP growth of a country’s trading partners. Export
weights are computed as average exports over the period 1990-2009 from
UNCTAD data. Real GDP growth is based on (by order of importance): OECD,
IMF WEO, IMF IFS, or proxied by real industrial production growth from the
IMF IFS to extend the sample coverage.
The real U.S. federal funds rate is computed using the Effective Federal Funds
Rate taken from the FED website. The one-year-ahead expected inflation rate
used to construct the ex ante real rate for the United States corresponds to the
forecasts of the change in the GDP deflator from the Survey of Professional
Forecasters, published by the Federal Reserve Bank of Philadelphia.
Global real interest rates are proxied by a GDP weighted average of the real
32
Overall liquidity
VXO index
Bank capital to assets
ratio (%)
Bank liquid reserves to
bank assets ratio (%)
European Central Bank financing rate (the Bundesbank base rate prior to
1999) and the real U.S. federal funds rate computed as above. The one-yearahead expected inflation rate used to construct the ex ante real rate for Europe
is calculated using the one-year-ahead forecast of consumer price inflation
from the European Central Bank or the realised lagged core inflation rate from
OECD statistics when the forecasted rate was not available.
Global liquidity is proxied by a GDP weighted average of the monetary
aggregate M2 for the U.S., Europe and Japan, all taken form Datastream.
Global risk aversion is taken from the Chicago Board of Options Exchange
Volatility Index, VXO. The VXO index is a measure of implied volatility
calculated using 30-day S&P 100 index at-the-money options.
Bank capital to assets is the ratio of bank capital and reserves to total assets.
Capital and reserves include funds contributed by owners, retained earnings,
general and special reserves, provisions, and valuation adjustments. Capital
includes tier 1 capital (paid-up shares and common stock), which is a common
feature in all countries' banking systems, and total regulatory capital, which
includes several specified types of subordinated debt instruments that need not
be repaid if the funds are required to maintain minimum capital levels (these
comprise tier 2 and tier 3 capital). Total assets include all nonfinancial and
financial assets. It is taken from the WB WDI.
Ratio of bank liquid reserves to bank assets is the ratio of domestic currency
holdings and deposits with the monetary authorities to claims on other
governments, nonfinancial public enterprises, the private sector, and other
banking institutions. It is taken from the WB WDI.
33
Table A2. Descriptive Statistics
1
Whole data
Including
Excluding
ongoing
ongoing
banking crises banking crises
(1)
(2)
Mean (s.d.)
Mean (s.d.)
Dependent variable
Start of a systemic banking crisis
Ongoing systemic banking crisis
Bank balance-sheet Shocks
Bank debt to GDP in t-1
Short-term bank debt to GDP in t-1
BBS shocks
CCC shocks
LCS shocks
Baseline control variables
(log) GDP per capita in t-1
(log) population in t-1
Credit growth in t-1
(log) credit in t-2
(log) Trade openness in t-1
exports/GDP in t-3
exports/GDP in t-3
x growth of partner countries in t-1
Growth of partner countries in t-1
Debt in total Liabilities in t-1
External assets / GDP in t-1
(excluding reserves)
Reserves / GDP in t-1
Liabilities / GDP in t-1
Year
# Observations
# Countries
Estimation sample
2
Whole
sample
Starting
crises
(3)
Mean (s.d.)
(4)
Mean (s.d.)
0.026
0.077
0.029
0.029
0.029
0.029
1
1
0.476
(3.478)
0.218
(1.274)
0.003
(0.012)
0.001
(0.013)
-0.001
(0.016)
0.389
(2.204)
0.188
(0.764)
0.004
(0.012)
0.002
(0.013)
-0.001
(0.016)
0.253
(0.629)
0.14
(0.380)
0.004
(0.012)
0.003
(0.013
-0.001
(0.016)
0.507
(1.389)
0.259
(0.682)
0.006
(0.008)
0.004
(0.008)
-0.007
(0.011)
-6.285
(1.585)
15.513
(1.960)
0.04
(0.176)
-1.31
(1.018)
0.829
(0.484)
0.372
(0.248)
0.013
(0.012)
0.034
(0.019)
0.72
(0.214)
0.925
(5.069)
0.131
(0.169)
1.485
(5.044)
1996
(7.790)
-6.226
(1.597)
15.436
(1.993)
0.045
(0.153)
-1.303
(1.015)
0.846
(0.490)
0.381
(0.252)
0.013
(0.012)
0.035
(0.019)
0.709
(0.218)
1.000
(5.352)
0.135
(0.176)
1.497
(5.226)
1996.11
(7.944)
-6.274
(1.540)
15.64
(1.875)
0.044
(0.159)
-1.42
(0.986)
0.823
(0.481)
0.37
(0.242)
0.014
(0.012)
0.037
(0.015)
0.683
(0.215)
0.883
(5.617)
0.136
(0.144)
1.477
(5.488)
1998.007
(7.009)
-5.876
(1.796)
16.253
(1.452)
0.066
(0.116)
-1.178
(1.092)
0.773
(0.488)
0.352
(0.238)
0.014
(0.012)
0.037
(0.016)
0.744
(0.175)
2.642
(14.553)
0.087
(0.081)
3.19
(14.337)
1997.606
(7.448)
4806
178
4329
178
2420
146
71
64
Note: 1. The whole sample refers to the sample of observations for which banking crises are observed and countries included in Lane
and Milesi-ferreti (2007).
2. The estimation sample includes observations for which the baseline control variables are non-missing. Table A2 lists the countries
included in this sample by geographical region.
34
Table A3. List of countries included in the estimation sample
Africa
America
Asia & Pacific
Europe
Algeria
Angola
Benin
Botswana
Burkina Faso
Burundi
Cameroon
Cape Verde
Central Afr. Rep.
Chad
Côte d'Ivoire
Egypt
Equatorial Guinea
Ethiopia
Gabon
Ghana
Guinea-Bissau
Kenya
Lesotho
Libya
Madagascar
Malawi
Mali
Mauritania
Mauritius
Morocco
Mozambique
Niger
Nigeria
Republic of Congo
Rwanda
Senegal
Seychelles
Sierra Leone
South Africa
Sudan
Swaziland
Tanzania
The Gambia
Togo
Tunisia
Uganda
Zambia
Argentina
Belize
Bolivia
Brazil
Canada
Chile
Colombia
Costa Rica
Dominica
Dominican Republic
Ecuador
El Salvador
Grenada
Guatemala
Guyana
Haiti
Honduras
Jamaica
Mexico
Panama
Paraguay
Peru
St. Lucia
St. Vincent & Grenadines
Trinidad and Tobago
United States
Uruguay
Venezuela
Armenia
Bahrain
Bangladesh
Bhutan
Brunei Darussalam
Cambodia
Georgia
Hong Kong SAR
India
Indonesia
Iran (Islamic Rep.)
Israel
Japan
Jordan
Kazakhstan
Kuwait
Kyrgyz Republic
Laos (PDR)
Malaysia
Mongolia
Myanmar
Nepal
Oman
Pakistan
Philippines
Qatar
Republic of Yemen
Russia
Saudi Arabia
Singapore
Sri Lanka
Syria (Arab Rep.)
Thailand
Vietnam
Australia
Fiji
New Zealand
Papua New Guinea
Solomon Islands
Tonga
Vanuatu
Albania
Austria
Belgium
Bulgaria
Croatia
Cyprus
Czech Republic
Denmark
Estonia
Finland
Macedonia (FYROM)
France
Germany
Greece
Hungary
Iceland
Ireland
Italy
Latvia
Lithuania
Luxembourg
Malta
Moldova
Netherlands
Norway
Poland
Portugal
Romania
Slovak Republic
Slovenia
Spain
Sweden
Switzerland
Turkey
Note: The table includes the 146 countries used in the main empirical analysis. The classification by continent is taken from the CEPII
distance dataset (Mayer and Zignano, 2011).
35
Table A4. Systemic banking crises, 1983-2009 period
Whole sample
Emerging Economies
Advanced Economies
(1)
(2)
(3)
A. Crises occuring over the 1983-2009 period
# of crises
129
74
20
Starting year
1995.08
1994.28
2003.08
(s.d.)
(7.30)
(6.06)
(7.39)
Duration
2.74
2.72
2.74
(s.d.)
(1.51)
(1.49)
(1.51)
B. Crises occuring over the 1983-2009 period excluding ongoing crises in 2009
# of crises
106
68
5
Starting year
1992.29
1993.07
1991.6
(s.d.)
(4.60)
(4.67)
(3.29)
Duration
2.87
2.78
3.8
(s.d.)
(1.63)
(1.54)
(1.79)
Output losses1
0.28
0.28
0.31
(s.d.)
(0.34)
(0.33)
(0.29)
# observations
88
57
5
output losses1
# countries
194
98
22
Note: 1. Output losses are taken from Laeven and Valencia (2010). They are computed as the cumulative difference between actual
and trend real GDP, expressed as a percentage of trend real GDP for the period [T, T+3] where T is the starting year of the crisis.
Trend real GDP is computed by applying an HP filter (λ=100) to the GDP series over [T-20, T-1]. No output losses are reported for
crises in transition economies that took place during the period of transition to market economies.
Source: Laeven and Valencia (2010), IMF IFS for the number of countries, and IMF WEO Outlook 2011 for the classification of
Emerging and Advanced Economies (Duttagupta et al. 2011).
Figure A.1 Share of countries starting and in an ongoing systemic banking crisis
Emerging Economies
0.30
0.30
0.25
0.25
Share of countries
Share of countries
Whole sample
0.20
0.15
0.10
0.20
0.15
0.10
0.05
0.05
0.00
0.00
1970
1980
Start of banking crisis
1990
2000
2010
Ongoing ba nking crisis
1970
1980
Start of banking crisis
1990
2000
2010
Ongoing ba nking crisis
Source: Laeven and Valencia (2010), IMF IFS for the number of countries, and IMF WEO Outlook 2011 for the classification of
Emerging Economies (Duttagupta et al. 2011).
36
Table A5. Bank balance-sheet shocks, control variables, and probability of banking crises
Dependent variable:
Estimator:
Debt to foreign banks / GDP
Bank balance-sheet shocks
Bank balance-sheet shocks
x Debt to foreign banks / GDP
External debt /
External liabilities
Domestic credit growth
(log) Credit over GDPt-2
(log) GDP per capita
(log) Population
Trade openness
Growth of trading partners
Growth of trading partners
x Exports/GDPt-3
Liabilities / GDP
(Liabilities / GDP)2
Assets / GDP excl. reserves
(Assets / GDP)2 excl. reserves
Reserves / GDP
(Reserves / GDP)2
Country fixed effects
Controls for duration(1)
Instruments(2)
Hansen (P-val)
Kleibergen-Paap (P-val)
Observations
Countries
Note:
1.
2.
Pooled OLS
(1)
0.006
(0.007)
-0.306
(0.454)
0.195
(0.372)
Start of a banking crisis
Fixed effects
First differences
(2)
(3)
(4)
(5)
0.020
0.008
0.034
0.032
(0.027)
(0.024)
(0.042)
(0.042)
0.077
0.010
0.189
0.182
(0.505)
(0.501)
(0.311)
(0.309)
0.432
0.532**
0.960***
0.951***
(0.296)
(0.269)
(0.124)
(0.122)
GMM-2S
(6)
0.401**
(0.167)
0.132
(0.340)
0.917***
(0.118)
0.046**
(0.018)
0.024
(0.017)
0.005
(0.005)
0.003
(0.004)
0.005***
(0.002)
-0.005
(0.015)
-0.482
(0.398)
0.702
(0.718)
-0.007
(0.007)
0.003***
(0.001)
0.009
(0.007)
-0.003***
(0.001)
-0.131**
(0.065)
0.091
(0.065)
0.121***
(0.046)
0.032
(0.021)
0.027*
(0.015)
0.030
(0.019)
-0.057
(0.066)
-0.002
(0.027)
-0.288
(0.568)
0.798
(0.986)
-0.008
(0.012)
0.003***
(0.001)
0.018
(0.014)
-0.003***
(0.001)
-0.202
(0.125)
0.222*
(0.133)
0.085*
(0.048)
0.030
(0.021)
0.011
(0.014)
0.008
(0.018)
-0.104
(0.073)
0.021
(0.028)
-0.218
(0.516)
1.033
(0.896)
-0.012
(0.012)
0.003***
(0.001)
0.018
(0.014)
-0.003***
(0.001)
-0.208*
(0.114)
0.233*
(0.124)
0.159***
(0.061)
0.032
(0.023)
-0.003
(0.030)
-0.019
(0.032)
-0.313
(0.256)
0.036
(0.038)
-0.398
(0.537)
-0.524
(0.941)
-0.208
(0.184)
0.188
(0.184)
0.004***
(0.001)
0.233
(0.184)
-0.004***
(0.001)
0.000
(0.000)
0.149**
(0.061)
0.034
(0.024)
-0.001
(0.029)
-0.018
(0.032)
-0.352
(0.268)
0.037
(0.038)
-0.385
(0.536)
-0.527
(0.938)
-0.200
(0.183)
0.180
(0.184)
0.004***
(0.001)
0.224
(0.183)
-0.004***
(0.001)
0.000
(0.000)
0.098
(0.067)
0.014
(0.025)
-0.028
(0.032)
0.014
(0.039)
-0.193
(0.287)
0.022
(0.039)
-0.396
(0.550)
-0.533
(1.013)
0.004
(0.019)
-0.003
(0.003)
-0.014
(0.024)
0.003
(0.003)
-0.268
(0.196)
0.103
(0.162)
No
No
No
Yes
No
No
Yes
Yes
No
Yes
No
No
Yes
Yes
No
2,419
146
2,419
146
2,419
146
2,204
146
2,204
146
Yes
Yes
Yes
0.24 (0.88)
7.17 (0.07)
2,055
146
Baseline controls consist of year fixed effects, the export weighted growth of main trading partners, its interaction with threeyear lagged exports over GDP, (log) GDP per capita, (log) population, domestic credit growth, the lagged level of domestic
credit over GDP, a linear and quadratic term for total assets (excluding reserves) over GDP, foreign reserves (excluding gold)
over GDP, the share of debt in external liabilities, and total liabilities over GDP. All regressors are lagged at least one period.
Standard errors are clustered at the country level. * denotes significance at the 10% level, ** at 5%, *** at 1%.
The controls for duration consist of a quadratic function in the number of years since the last crisis interacted, respectively,
with a dummy variable for having, or not having experienced a crisis since 1970.
The specification in Column 6 instruments the exposure to contagion by the first difference of bank debt over GDP in t-3 and
t-4, the bank debt over GDP in t-3 interacted with bank balance-sheet shocks, and the average of bank debt over GDP
interacted with a measure of bank balance-sheet shocks keeping the lending share of each creditor constant.
37
Table A6. Bank balance sheet shocks based on bank indices and probability of banking crises
Dependent variable:
Estimator:
Bank debt / GDP
Bank Balance Sheet Shocks
Bank Balance Sheet Shocks
x Bank debt / GDP
Country fixed effects
Controls for duration(1)
Instruments(2)
Hansen (P-val)
Kleibergen-Paap (P-val)
Observations
Countries
Start of a banking crisis
Pooled
OLS
(1)
0.003
(0.019)
-0.007
(0.031)
0.093**
(0.042)
(2)
-0.047
(0.036)
-0.018
(0.030)
0.143**
(0.063)
(3)
-0.043
(0.031)
0.010
(0.031)
0.144**
(0.063)
(4)
0.013
(0.028)
-0.003
(0.027)
0.116***
(0.037)
(5)
0.008
(0.028)
-0.002
(0.027)
0.115***
(0.036)
(6)
0.256
(0.359)
-0.008
(0.036)
-0.068*
(0.040)
No
No
No
Yes
No
No
Yes
Yes
No
Yes
No
No
Yes
Yes
No
1,643
147
1,643
147
1,643
147
1,460
147
1,460
147
Yes
Yes
Yes
5.15 (0.27)
5.16 (0.40)
1,415
147
Fixed effects
Note:
First differences
GMM-2S
All regressors are lagged one period. Standard errors are clustered at the country level. * denotes a significant estimate at the
10% level, ** at 5%, *** at 1%.
1.
Baseline controls include the export weighted growth of the main trading partners, its interaction with the three-year lagged of
export over GDP, (log) GDP per capita, (log) population, domestic credit growth, the lagged level of domestic credit over
GDP, a linear and quadratic terms for total assets (excluding reserves) over GDP, external reserves (excluding gold) over
GDP, the share of debt in external liabilities, and total liabilities over GDP. The controls for duration include a quadratic
function in the number of years since the last crisis interacted, respectively, with a dummy variable for having, or not having
experienced a crisis since 1970.
2.
The specification in Column 6 instruments the exposure to contagion by the first difference of bank debt over GDP in t-3 and
t-4, the bank debt over GDP in t-3 interacted with Bank Balance Sheet Shocks, and the average of bank debt over GDP
interacted with a measure of Bank Balance Sheet Shocks keeping the lending share of each creditor constant.
Source: FTSE equity index of the banking sector 1994-2009, downloaded from Datastream.
38
Table A7. Bank balance-sheet shocks and probability of banking crises in Emerging Economies
Dependent variable:
Estimator:
Bank debt / GDP
Bank balance-sheet Shocks
Bank balance-sheet Shocks
x Bank debt / GDP
Start of a banking crisis
Pooled
OLS
(1)
0.006
(0.007)
-0.262
(0.704)
0.799
(0.635)
(2)
0.025
(0.028)
-0.074
(0.760)
0.999*
(0.523)
(3)
0.006
(0.032)
-0.196
(0.723)
1.244**
(0.528)
(4)
0.043
(0.057)
0.066
(0.459)
1.019***
(0.373)
(5)
0.045
(0.057)
0.052
(0.450)
0.965***
(0.342)
No
No
No
Yes
No
No
Yes
Yes
No
Yes
No
No
Yes
Yes
No
Country fixed effects
Controls for duration(1)
Instruments(2)
Hansen (P-val)
Kleibergen-Paap (P-val)
Fixed effects
First differences
GMM-2S
(6)
0.225**
(0.107)
0.043
(0.459)
0.795***
(0.275)
Yes
Yes
Yes
0.36 (0.84)
5.65 (0.13)
1,509
1,509
1,509
1,375
1,375
1,287
Observations
85
85
85
85
85
85
Countries
Note:
All regressors are lagged one period. Standard errors are clustered at the country level. * denotes a significant estimate
at the 10% level, ** at 5%, *** at 1%. Emerging Economies are defined according to Duttagupta et al., 2011.
1.
2.
Baseline controls include the export weighted growth of the main trading partners, its interaction with the three-year lagged of
export over GDP, (log) GDP per capita, (log) population, domestic credit growth, the lagged level of domestic credit over
GDP, a linear and quadratic terms for total assets (excluding reserves) over GDP, foreign reserves (excluding gold) over
GDP, the share of debt in external liabilities, and total liabilities over GDP. The controls for duration include a quadratic
function in the number of years since the last crisis interacted, respectively, with a dummy variable for having, or not having
experienced a crisis since 1970.
The specification in Column 6 instruments the exposure to contagion by the first difference of bank debt over GDP in t-3 and
t-4, the bank debt over GDP in t-3 interacted with Bank balance-sheet Shocks, and the average of bank debt over GDP
interacted with a measure of Bank balance-sheet Shocks keeping the lending share of each creditor constant.
Table A8. Bank balance-sheet shocks and cross-border bank lending flows to Emerging Economies
Dependent variable:
Estimator:
(1)
Common-creditor Contagion shock
Creditor fixed effects
Creditor*debtor fixed effects
Creditor linear time trend
Creditor quadratic time trend
Creditor domestic GDP growth
# observations
# creditor*year
# debtor countries
Yearly change in (log) creditor’s assets in country d
Debtor-year fixed-effects model
(2)
(3)
(4)
(5)
(6)
-3.046***
(0.994)
-4.195***
(0.986)
-4.140***
(0.972)
-4.209***
(0.970)
-3.106***
(1.050)
-3.108***
(1.053)
No
No
No
No
No
Yes
No
No
No
No
No
Yes
No
No
No
No
Yes
Yes
No
No
No
Yes
Yes
Yes
No
No
Yes
Yes
Yes
Yes
24,614
416
97
24,614
416
97
24,614
416
97
24,614
416
97
24,614
416
97
24,614
416
97
Note:
Observations are yearly debtor-creditor flows. Standard errors are two-way clustered at the debtor and creditor*year
levels.The explanatory variable is lagged one period. * denotes significant estimate at the 10% level, ** at 5%, *** at 1%. Emerging
Economies are defined according to Duttagupta et al., 2011.
1. All specifications control for debtor-year fixed effects.
39