sustainability

Article
Financial Hazard Map: Financial Vulnerability
Predicted by a Random Forests Classification Model
Katsuyuki Tanaka, Takuji Kinkyo and Shigeyuki Hamori * ID

Graduate School of Economics, Kobe University, 2-1, Rokkodai, Nada-Ku, Kobe 657-8501, Japan;
[email protected] (K.T.); [email protected] (T.K.)
* Correspondence: [email protected]; Tel.: +81-78-803-6832




Received: 9 April 2018; Accepted: 8 May 2018; Published: 11 May 2018


Abstract: This study develops a systematic framework for assessing a country’s financial vulnerability
using a predictive classification model of random forests. We introduce a new indicator that
quantifies the potential loss in bank assets and measures a country’s overall vulnerability by
aggregating these indicators across the banking sector. We also visualize the degree of vulnerability
by creating a Financial Hazard Map that highlights countries and regions with underlying risks in
their banking sectors.

Keywords: financial hazard map; random forests; early warning system; bank failure

1. Introduction
The severe economic consequences of the global financial crisis of 2008–2009 highlighted the
importance of crisis prevention and sparked a renewed interest in early warning systems (EWSs).
An EWS aims to detect potential vulnerabilities in a financial system that could trigger a system-wide
crisis. A reliable EWS provides useful guidance for policy-makers to activate macro-prudential policy
in an effective and timely manner. The International Monetary Fund (IMF) [1] and the Committee
on the Global Financial System [2] provide a comprehensive discussion of the operational aspects of
the macro-prudential policy. After Frankel and Rose [3] and Kaminsky et al.’s [4] early contributions,
researchers have made considerable efforts to develop a consistently useful EWS for various types
of crises.
Kaminsky et al. [4] proposed the popular signaling approach, which Alessi and Detken [5] recently
used. This approach seeks to identify the threshold values for individual indicators that signal crises,
and thus trigger an early warning when the pre-defined threshold for the pre-selected indicator
is breached. A popular indicator common in these studies is the credit-to-GDP (Gross Domestic
Product) ratio, which is a key indicator signaling credit booms. However, this signaling approach has
a shortcoming given that, as a univariate approach, the decision would rely on only a single factor,
which can send a misleading signal.
Another conventional approach from the EWS literature is estimating the multivariate probit
and logistic regressions, which relate the probability of a crisis to a set of explanatory variables, such
as current account balance, real exchange rates, credit growth, and fiscal balance [6–10]. Despite its
popularity, the conventional approach has certain limitations. For one, researchers must pre-select
explanatory variables from a wide range of economic indicators based on some prior information. For
another, the logistic regression does not readily allow for non-linear or threshold effects of explanatory
variables. More generally, linear regressions often perform poorly in terms of prediction performance
relative to newer machine learning models [11]. Linear regressions may work well for small datasets
but they are not readily scalable to larger datasets.

Sustainability 2018, 10, 1530; doi:10.3390/su10051530 www.mdpi.com/journal/sustainability

Sustainability 2018, 10, 1530 2 of 18

Ghosh and Ghosh [12] and Frankel and Wei [13] employed a decision tree method that uses
a sequence of splitting rules to segment the space of explanatory variables. Hastie et al. [11] and
James et al. [14] provide details of tree methods, including decision trees and random forests. At
each node of a tree, the sample is split into two sub-branches according to the threshold value of an
explanatory variable. For classification trees, either the Gini index or the cross entropy is used to
evaluate the quality of a split. A smaller value of these indices indicates that the node is purer, and thus
contains more observations from a single class. The process is repeated until a stopping criterion is
reached, such as the minimum number of observations at each node. Each terminal node at the bottom
of the tree provides a class prediction for a given observation. Whereas linear logistic regressions
models require a handcrafted selection of explanatory variables to obtain reasonable early warning
performance, the decision tree systematically learns important variables, performs better in early
warning, and allows for non-linear effects. Although a decision tree is simple and provides explanatory
and intuitive decision rules, it suffers from high variance (i.e., a small change in the data can cause a
large change in the financial tree), so is likely to suffer over-fitting problems. This is largely owing to
the fact that the values of the thresholds depend heavily on the values of the training observations.
With an increased opportunity to gain access to larger datasets, exploring the significant scope
for economic modeling and analysis for a more flexible approach has become popular with data
scientists [15,16]. In this study, we take advantage of the advancements in predictive modeling
techniques of machine learning to build an EWS, and develop a systematic framework to assess and
visualize a country’s financial vulnerability. The main contributions of our study are three-fold. First,
our study differs from previous ones in that we used a novel machine-learning technique known as
random forests to construct an EWS to predict bank failures (random forests EWS). Random forests are
a variant of decision trees that significantly improve prediction accuracy by combining a large number
of trees using random input selection [17]. Second, we introduce a new indicator that quantifies the
expected potential loss in bank assets computed using the prediction of the random forests EWS.
To assess a country’s overall financial vulnerability, we aggregate individual banks’ expected potential
asset losses across the domestic banking sector. Finally, we visualize the degree of a country’s financial
vulnerability by creating a Financial Hazard Map that highlights countries and regions with significant
risks in their underlying banking sectors. Our work is similar to that of Tanaka et al. [18], but differs
by a few points. Our paper provides a financial analysis of the finance sector, whereas the interest of
Tanaka et al. focused on the industrial sector. Furthermore, we propose a novel indicator to assess the
overall financial vulnerability of each country.
We chose random forests (RF) for three reasons. First, RF can significantly improve prediction
accuracy by building a large number of decision trees on bootstrapped training samples—a technique
known as ensemble learning. Random forests also circumvent the over-fitting problem by adding
randomness to the tree building process, and thus reducing correlations among trees; hence, it performs
well with out-of-sample data. Second, random forests can better handle a large dataset as multiple
trees can be trained in parallel efficiently with a very simple hyper-parameter setting. The model can
be built by merely setting the number of trees. Finally, RF provide the importance measurement, which
can be used for certain levels of causality inference. Whereas various application areas use random
forests, including computer vision and bioinformatics, its application to economics remains limited.
Tanaka et al. [19] used random forests to predict bank failure in OECD member countries.
Another important feature of our study is the use of bank-level financial statements to predict
bank failure using the random forests EWS built from a large dataset of more than 15,000 banks globally.
As previous studies typically used macroeconomic indicators to predict currency and financial crises,
the recent literature indicates that the state of bank financial statements can explain differences in
performance across banks during financial crises [20,21]. Moreover, previous studies often defined
a crisis as an event in which the values of preselected indicators exceed predetermined thresholds.
Consequently, the prediction performance significantly depends on the choice of threshold. We define
the event of a bank failure as the change in a bank’s status from active to inactive (i.e., bankrupt,
Sustainability 2018, 10, 1530 3 of 18

in liquidation, or dissolved) based on the information provided by the Bureau Van Dijk Bankscope.
By doing so, we wanted to minimize arbitrariness, and thus reduce the possible bias in prediction.
The remainder of the paper proceeds as follows. Section 2 describes the methodology and data of
building the random forests EWS. Section 3 introduces a new indicator that quantifies the expected
potential losses in bank assets. We present the assessment of a country’s financial vulnerability and
visualize it by creating a Financial Hazard Map. Section 4 provides our conclusions.

2. Materials and Methods

In this study, we considered the task of building an EWS as a classification problem to identify
a bank’s status (i.e., active or inactive) based on the underlying financial conditions. Drawing on
insights from the extensive literature on corporate bankruptcy predictions, we used information about
individual banks’ financial statements as predictors to build models. Altman [22] provides an early
contribution to the literature. In contrast to existing studies that are more concerned with identifying
the key predictors of bankruptcy, we prioritized improving the prediction accuracy. To this end, we
used random forests that tend to perform better in terms of prediction accuracy than conventional
methods, such as logistic regressions, which have been widely used in previous studies.

2.1. Major Features of Random Forests

Random forests are a variant of decision trees, which overcome the over-fitting problem by
building multiple trees and combing the results of these trees [17], effectively forming forests. Each
tree in a random forest is built using randomly selected data samples and/or randomly selected input
variables from the original data to split each node. After generating a large number of trees, the model
votes for the most popular class. A single-tree classifier tends to have only marginally better accuracy
than a random choice of class. However, by combining a large number of trees using random input
selection, random forests can produce a powerful model.
Breiman et al. [17] constructed such trees using the Gini index criterion, which measures the best
split criterion based on the impurity of each node. The algorithm aims to select the optimal splitting
variable and the corresponding threshold value by making each node as pure as possible. Suppose Mn
is the number of pieces of information reaching node n and Mni is the number of data points belonging
to class Ci , the Gini index, GIn , of node n is obtained using Equation (1):

K  2 Mni
GIn = 1 − ∑ pin , where pin =
Mn
(1)
i =1

A smaller Gini index value for node n represents greater purity, which implies that the node
contains more observations from a single class. Hence, a decreasing Gini index is an important criterion
when splitting a node.
In comparison to single-tree modeling, random forests have several desirable features [17,23].
First, random forests perform better in terms of classification accuracy by building a large number
of trees instead of only a single tree. Each tree is built using randomly selected data samples and
randomly selected input variables from the original data to split each node. After generating a large
number of trees, they vote for the most popular class. A single-tree classifier tends to have only
marginally better accuracy than a random choice of class. However, by combining a large number of
trees using random input selection, random forests can improve accuracy. Second, random forests
provide better generalization abilities and are robust to over-fitting. Hence, RF may have better
out-of-sample accuracy when using a random selection of input variables to split each node and
combining the results of multiple trees yields error rates that compare favorably to alternative methods
and are more robust with respect to noise. Third, random forests can better handle large datasets as
multiple trees can be efficiently trained in parallel. Finally, random forests provide a measure for the
relative contribution of each variable to generate a prediction. These variable importance measures
Sustainability 2018, 10, 1530 4 of 18

help identify the variables that are important for distinguishing between active and inactive banks,
Sustainability
and thus for2018, 10, x FORbank
predicting PEER failure.
REVIEW 4 of 19

2.2.Data
2.2. Data

WeWesourced
sourcedour ourdata
datafor
forbank
bankfinancial
financialstatement
statementindicators
indicatorsfromfromBankscope.
Bankscope.The Theadvantage
advantageof of
usingthis
using thisdata
datasource
sourceisisthat
thatititprovides
providesaabroad
broadcoverage
coverageofofbanks
bankswith withstandardized
standardizeddata dataformats
formats
acrosscountries.
across countries.We Weusedused4848indicators
indicatorsderived
derivedfrom
fromthe theSummary
SummaryAnalytics
Analyticscategory
categoryclassified
classifiedinto
into
fourgroups:
four groups: profitability
profitability ratio, capitalization, loan loan quality,
quality, and
andfunding.
funding Our sampleB).
(Appendix included 23,455
Our sample
commercial banks, saving banks, and cooperatives incorporated in 198
included 23,455 commercial banks, saving banks, and cooperatives incorporated in 198 countries and countries and regions. The
trainingThe
regions. set included
training setannual observations
included of the latest available
annual observations financial
of the latest statements
available financialfor each bankfor
statements up
to 2014.
each bankWe updefined
to 2014.aWe bank failurea event
defined as the change
bank failure event asinthe a bank’s
changestatus from active
in a bank’s statusto inactive
from active(i.e.,
to
bankrupt, in liquidation, or dissolved) as reported by Bankscope. We
inactive (i.e., bankrupt, in liquidation, or dissolved) as reported by Bankscope. We assumed that the assumed that the latest
available
latest financial
available statements
financial statements for for
active banks
active bankshad hadsound
soundfinancial
financialstatus
statusandand inactive banks
banks hadhad
unsoundfinancial
unsound financialstatus.
status.WeWethen then systematically
systematically identified
identified patterns
patterns distinguishing
distinguishing the the differences
differences by
by random
random forests. forests.
AsAsthere
therewerewerefewer
fewerinactive
inactivebanks
banks(7294
(7294banks),
banks),we weselected
selectedthe thelargest
largest7294
7294active
activebanks
banksin in
terms of total assets to match the number of inactive banks. We also selected
terms of total assets to match the number of inactive banks. We also selected the smallest 7294 active the smallest 7294 active
banksto
banks tobuild
buildaamoremoreflexible
flexiblemodel
modeltotoprevent
preventaabias biastoward
towardlarger
largerbanks.
banks. To Toavoid
avoidmodel
modelbiasbias
createdby
created byananimbalanced
imbalancedtraining
trainingset,set,
wewe evened
evened outoutthethe sample
sample sizessizes of active
of active andand inactive
inactive banksbanks
by
by doubling the sample size of inactive banks by duplicating each observation.
doubling the sample size of inactive banks by duplicating each observation. In addition, we eliminated In addition, we
eliminated
variables variables
if more thanif50%
moreofthan 50% ofwere
its values its values
missing were missing
(7294 (7294 × and
× 2 biggest 2 biggest and active
smallest smallest active
banks +
banks
7294 + 7294banks).
inactive inactiveThus,
banks).
weThus, we eliminated
eliminated 6 variables 6 variables
and used and used 42 for
42 variables variables for experiments.
experiments. We used
Werandom
the used theforests
random and forests
caret and caret packages
packages in the R package
in the R software software to package
train and to train and evaluate
evaluate our models. our
models.
Figure Figure 1 illustrates
1 illustrates the modelthe modelprocess
building building forprocess
the randomfor the random
forests EWS.forests
AppendixEWS.CAppendix
reports theC
reports the classification accuracy of
classification accuracy of the random forests EWS.the random forests EWS.

Figure1.1.Model
Figure Modelbuilding
buildingprocess
processof
ofrandom
randomforests
forestsearly
earlywarning
warningsystem
system(EWS).
(EWS).

3.3.Results
Results

3.1.
3.1.Variable
VariableImportance
ImportanceMeasures
Measures
AAuseful
usefulproperty
propertyof
ofrandom
randomforests
forestsisisthat
thatititprovides
providesvariable
variableimportance
importancemeasures
measuresthat
thathelp
help
identify
identify the most important variables for distinguishing between active and inactive banks. Hence,
the most important variables for distinguishing between active and inactive banks. Hence,
RF
RFshould
shouldprovide
providesome
someclues
cluesto
tothe
theunderlying
underlyingcausescausesof
ofbank
bankfailures.
failures.For
Forclassification
classificationtrees,
trees,we
we
obtained the variable importance measures from each variable’s contribution to the reduction in the
Gini index. The Gini index is a common measure of the degree of inequality in income distribution.
The smaller the value of the index, the more equal the society.
Sustainability 2018, 10, x FOR PEER REVIEW 5 of 19

In our random forests algorithm, the Gini index is the measure for the purity of each node. A
Sustainability 2018, 10, 1530 5 of 18
smaller value of the Gini index represents a purer node, which implies that the node contains more
observations from a single class. The goal of the algorithm was to make each node as pure as possible
by selecting
obtained the optimal
the variable splittingmeasures
importance variable from
and theeachcorresponding thresholdtovalue.
variable’s contribution Therefore,
the reduction we
in the
calculated
Gini index.variable
The Giniimportance by summing
index is a common measurethe total
of thereduction
degree ofin the Gini in
inequality index
income by splits over a
distribution.
given
The variable,
smaller the averaged
value of theover all bagged
index, the moretrees.
equal the society.
Figure
In our 2 illustrates
random the variable
forests algorithm, importance
the Ginimeasures
index is theas the mean decrease
measure in the Gini
for the purity index
of each for
node.
each
A variable.
smaller value Considering
of the Gini indexthis represents
model, we identified
a purer node, the
whichfollowing indicators
implies that the node as contains
the top morefour
predictors: interest
observations from aexpense/average
single class. The interest-bearing
goal of the algorithm liabilities,
was to interest
make income
each node on asloan/average gross
pure as possible
loans,
by interestthe
selecting expense
optimal on splitting
customervariable
deposits/average customer deposits,
and the corresponding and interest
threshold value.income/average
Therefore, we
earning assets.
calculated variableTheimportance
importancebymeasure
summing forthe
thetotal
firstreduction
indicatorin was
theby farindex
Gini the largest.
by splitsThese
overtop four
a given
indicatorsaveraged
variable, fall into over
the profitability ratio category. In contrast, the importance measures for the other
all bagged trees.
categories
Figureof2 illustrates
indicators,thethatvariable
is, capitalization,
importanceloan quality,
measures as and liquidity
the mean are much
decrease in thesmaller.
Gini indexThe
results indicate that bank profitability has the most important impact
for each variable. Considering this model, we identified the following indicators as the top four on the probability of bank
failure.
predictors: interest expense/average interest-bearing liabilities, interest income on loan/average gross
loans,We showexpense
interest the experimental
on customer result of a single decision
deposits/average tree deposits,
customer in Figureand3 forinterest
comparison. Though a
income/average
single tree
earning selects
assets. similar
The criteriameasure
importance to distinguish
for thebetween active and
first indicator wasinactive
by far the banks, it does
largest. not top
These perform
four
as well as fall
indicators random forests.
into the This is ratio
profitability due to the factInthat
category. the random
contrast, forests model
the importance measures produces
for thea other
more
flexible model
categories as it produces
of indicators, that is,multiple trees toloan
capitalization, analyze different
quality, patternsare
and liquidity in the
much data, whereas
smaller. Thearesults
single
tree onlythat
indicate produces one set of rules
bank profitability for amost
has the classification
importantdecision.
impact on the probability of bank failure.

Figure
Figure 2. Variable importance
2. Variable importance measures
measures of
of random
random forests.
forests.

We show the experimental result of a single decision tree in Figure 3 for comparison. Though a
single tree selects similar criteria to distinguish between active and inactive banks, it does not perform
as well as random forests. This is due to the fact that the random forests model produces a more
flexible model as it produces multiple trees to analyze different patterns in the data, whereas a single
tree only produces one set of rules for a classification decision.
Sustainability 2018, 10, 1530 6 of 18
Sustainability 2018, 10, x FOR PEER REVIEW 6 of 19

Figure 3. Plot of a tree model.

Figure 3. Plot of a tree model.
3.2. New Indicator for the Expected Potential Asset Loss
3.2. New Indicator for the Expected Potential Asset Loss
We introduce a new indicator to assess the degree of financial vulnerability using the prediction
We introduce a new indicator to assess the degree of financial vulnerability using the prediction
of the random forests EWS. We used the 2014 financial statement data to predict the probability of
of the random forests EWS. We used the 2014 financial statement data to predict the probability of
bank failure. We define the expected potential asset loss of a bank as follows:
bank failure. We define the expected potential asset loss of a bank as follows:
EPALi,j i,j==PPi,ji,j ×
EPAL Total Assets
× Total Assetsi,ji,j,, (2)
(2)
where EPAL
where EPALi,j denotes
denotes thethe expected
expected potential
potential loss loss in
in bank
bank ii in
in country
country j,j, PPi,j denotes the probability of
i,j i,j denotes the probability of
failure for bank i in country j given by the random forests EWS prediction,
failure for bank i in country j given by the random forests EWS prediction, and Total Assets and Total Assetsi,j denotes
i,j denotes
the value of total assets of bank i in country j. To measure a country’s overall
the value of total assets of bank i in country j. To measure a country’s overall financial vulnerability, financial vulnerability,
we aggregate
we aggregatethethe value
value of EPAL
of EPAL i,j across the domestic banking sector. Given that we used
i,j across the domestic banking sector. Given that we used consolidated
financial statement data, all the expectedall
consolidated financial statement data, the expected
potential loss of potential
multinationalloss banks
of multinational
was counted banks was
as losses
counted
in as losses
the country where in the
theheadquarters
country where the banks
of these headquarters of these
were located. Webanks were located.
acknowledge that thisWe is
acknowledge that this is the limitation of our work and consider overcoming
the limitation of our work and consider overcoming this limitation as our future task. Hence, this limitation as our
the
future task. Hence,
country-level expectedthe potential
country-level loss inexpected potential
the domestic loss insector
banking the domestic
denoted bankingby EPALsector denoted
j is given by:
by EPALj is given by:
j =∑
EPAL
EPAL j= EPALi,j
∑ii EPAL i,j.
. (3)
(3)
gauge the
To gauge the impact
impact of
of the
the expected
expected potential
potential asset loss on the
the domestic
domestic banking sector and
activities, we calculated
economic activities, calculated the
the share
share of EPAL
EPALjj in the total assets of the domestic banking sector
and in nominal GDP.
GDP. Table
Table11summarizes
summarizesthe theresults.
results.
The left column of the table ranks 50 countries in terms of their share in banking banking sector
sector assets.
assets.
The ranking indicates that Suriname, Grenada, Denmark, Gabo, and Guatemala are the five most
vulnerable countries
vulnerable countriesininthe
the sense
sense thatthat
the the impact
impact of theofexpected
the expected potential
potential loss on loss on the domestic
the domestic banking
banking
sector sector
can can be relatively
be relatively large.these
large. Thus, Thus,countries
these countries
have ahave a relatively
relatively highofrisk
high risk of a system-
a system-wide
wide banking
banking crisis.crisis.
The right column of the table ranks countries in terms of the share in nominal GDP. The ranking
indicates that the Palestinian Territories, Luxembourg, Cyprus, Denmark, and France are the five
most vulnerable countries in the sense that the impact of the expected potential asset loss on domestic
Sustainability 2018, 10, 1530 7 of 18

economic activities could be relatively large. This is particularly the case if the assets of banks with
high probability of failure consist primarily of domestic loans and investments.

Table 1. Top 50 countries in terms of the shares of expected potential asset loss based on 2014
financial statements.

Share of Expected Potential Asset Loss in Domestic Banking Sector Share of Expected Potential Asset Loss in Nominal GDP
SURINAME 37.92% PALESTINIAN TERRITORIES 137.53%
GRENADA 31.00% LUXEMBOURG 135.54%
DENMARK 28.14% CYPRUS 100.92%
GABON 28.06% DENMARK 81.80%
GUATEMALA 27.33% FRANCE 71.26%
VENEZUELA 25.65% VENEZUELA 67.12%
SENEGAL 25.30% PORTUGAL 58.03%
NEPAL 24.41% LEBANON 56.36%
UZBEKISTAN 24.41% JORDAN 56.25%
KYRGYZSTAN 24.31% BAHRAIN 52.90%
CAMEROON 24.04% SPAIN 50.81%
LESOTHO 23.54% MAURITIUS 50.56%
EL SALVADOR 23.25% UNITED KINGDOM 41.92%
DOMINICA 23.00% SWITZERLAND 39.83%
ROMANIA 22.44% AUSTRIA 38.16%
THAILAND 21.48% HONG KONG 37.43%
HUNGARY 20.99% GERMANY 37.28%
MONTENEGRO 20.74% NETHERLANDS 36.94%
ETHIOPIA 20.51% BAHAMAS 34.16%
ARGENTINA 20.38% SAINT KITTS AND NEVIS 32.75%
LIBERIA 19.95% THAILAND 31.86%
TUNISIA 19.81% BELGIUM 31.19%
LIECHTENSTEIN 19.72% FINLAND 30.61%
BOSNIA AND HERZEGOVINA 19.66% SAN MARINO 30.59%
COTE D'IVOIRE 19.01% NEW ZEALAND 28.96%
PARAGUAY 18.75% PANAMA 28.10%
SERBIA 18.50% GRENADA 27.02%
MADAGASCAR 18.22% ITALY 25.76%
BOLIVIA 18.17% GREECE 23.50%
PORTUGAL 18.17% MOROCCO 21.80%
REPUBLIC OF KOREA 18.09% AUSTRALIA 21.56%
SAINT KITTS AND NEVIS 17.99% CHILE 20.94%
HONDURAS 17.96% IRELAND 20.55%
BELIZE 17.81% NEPAL 19.99%
PERU 17.30% CROATIA 19.43%
AUSTRIA 17.30% ICELAND 18.82%
JORDAN 17.27% CAPE VERDE 18.66%
MAURITIUS 17.13% GUATEMALA 17.61%
UKRAINE 17.05% REPUBLIC OF KOREA 17.19%
CAPE VERDE 17.03% EL SALVADOR 17.05%
TURKMENISTAN 16.96% ANTIGUA AND BARBUDA 17.03%
NEW ZEALAND 16.70% SWEDEN 16.97%
MALI 16.15% CANADA 16.74%
CROATIA 15.97% HUNGARY 16.69%
CHILE 15.60% DOMINICA 16.46%
BRUNEI DARUSSALAM 15.33% HONDURAS 16.04%
ARMENIA 15.30% BARBADOS 15.89%
DOMINICAN REPUBLIC 15.04% SURINAME 15.52%
BAHRAIN 14.89% VIETNAM 15.34%
ECUADOR 14.84% TUNISIA 15.23%

Interestingly, many Organisation for Economic Co-operation and Development (OECD) countries,
including European countries, at the top of the list. This may indicate that these countries have not
recovered fully from the major financial crises, notably, the global financial crisis of 2008–2009 and the
European debt crisis of 2010–2013, or new financial risks may be looming. Given the relatively large
size of their domestic banking sectors, these countries can be the epicenter of cross-border financial
spillovers by withdrawing overseas loans and investments in the face of financial difficulties.
In Figure 4, we create a scatter plot of the vulnerability measures reported in Table 1.
The combination of higher values of these measures in a particular country implies greater financial
vulnerability. The figure clearly indicates that Denmark and Venezuela stand out in terms of both
measures, signaling significant risks. The bold horizontal and vertical lines indicate the medians
of these measures; the shadows indicate the first quartiles. The medians of the share of EPALj in
the banking sector assets and nominal GDP are 11.60% and 7.85%, respectively. The level of these
Sustainability 2018, 10, 1530 8 of 18

medians indicates the overall vulnerability of the global banking sector, with a significant increase in
signaling financial risks. Notably, our vulnerability measure raises a red flag for potential trouble, but
it does not identify the causes or the likely outcomes of the trouble. However, we believe that these
measures are useful for spotting vulnerabilities, and thus encouraging regulators and investors to take
preemptive actions.
Sustainability 2018, 10, x FOR PEER REVIEW 9 of 19

Figure
Figure 4.
4. Scatter
Scatter plot
plot of
of country
country vulnerabilities.
vulnerabilities.

In Appendix A, we summarize the predicted bank failures for each country for 2014. The table
In Appendix A, we summarize the predicted bank failures for each country for 2014. The table
shows the number of banks and the sum of assets held by the banks for each category of predicted
shows the number of banks and the sum of assets held by the banks for each category of predicted
probability of failures with a 10-percentage-point interval.
probability of failures with a 10-percentage-point interval.
3.3. Financial Hazard Map
3.3. Financial Hazard Map
Finally, we visualized the degree of a country’s financial vulnerability by creating a Financial
Finally, we visualized the degree of a country’s financial vulnerability by creating a Financial
Hazard Map. Corresponding to each definition of vulnerability in Table 1, we present two types of
Hazard Map. Corresponding to each definition of vulnerability in Table 1, we present two types of
maps. Figure 5 shows the share of EPALj in the assets of domestic banking sectors. The areas that are
maps. Figure 5 shows the share of EPALj in the assets of domestic banking sectors. The areas that are
darker red indicate a higher degree of vulnerability in terms of the impact of the expected potential
darker red indicate a higher degree of vulnerability in terms of the impact of the expected potential
asset loss on the domestic banking sector. Figure 6 shows the share of EPALj in nominal GDP. Darker
asset loss on the domestic banking sector. Figure 6 shows the share of EPALj in nominal GDP. Darker
red areas indicate a higher degree of vulnerability in terms of the impact of the expected potential
red areas indicate a higher degree of vulnerability in terms of the impact of the expected potential
asset loss on domestic economic activities.
asset loss on domestic economic activities.
The Financial Hazard Map highlights countries and regions with significant vulnerability in
The Financial Hazard Map highlights countries and regions with significant vulnerability in their
their underlying banking sector. The darker red areas correspond to the top 50 countries listed in
underlying banking sector. The darker red areas correspond to the top 50 countries listed in Table 1.
Table 1. The map provides a clear and understandable assessment of financial vulnerability in
The map provides a clear and understandable assessment of financial vulnerability in particular
particular countries and regions. It also shows the geographical distribution of financial risk and the
countries and regions. It also shows the geographical distribution of financial risk and the danger of
danger of potential contagion for neighbors of high vulnerability areas.
potential contagion for neighbors of high vulnerability areas.
Sustainability 2018, 10, 1530
Sustainability 2018, 10, x FOR PEER REVIEW 10 of 19
9 of 18

Figure 5. Impact of expected potential asset loss on the banking sector.

Figure 5. Impact of expected potential asset loss on the banking sector.
Sustainability 2018, 10, 1530 10 of 18

Sustainability 2018, 10, x FOR PEER REVIEW 11 of 19

Figure 6. Impact of expected potential asset loss on economic activities.

Figure 6. Impact of expected potential asset loss on economic activities.
Sustainability 2018, 10, 1530 11 of 18

4. Conclusions
We developed a systematic framework for assessing and visualizing a country’s financial
vulnerability. We employed a novel machine-learning approach known as random forests to construct
an EWS to predict bank failures and introduced a new indicator that quantifies the expected potential
loss in bank assets computed based on the random forests EWS prediction. We assessed the financial
vulnerability of each country by aggregating individual banks’ indicators across the banking sector.
To gauge the impact of expected potential asset loss, we calculated the shares in the banking sector
assets and nominal GDP. We identified countries and regions with high vulnerability in terms of these
shares. Furthermore, we visualized the degree of a country’s financial vulnerability by creating a
Financial Hazard Map. We demonstrated the usefulness of the Financial Hazard Map in spotting
vulnerable countries and regions and understanding the geographical distribution of risk.
We hope that the Financial Hazard Map will prove useful for both policy-makers and private
investors in detecting potential risk, and thereby prompting precautionary actions. Our framework
of assessing financial vulnerability is simple, and therefore readily applicable to other types of risk
analysis. A future task may be to develop a dynamic framework that allows for an assessment of
contagion risks between banks and countries potentially in trouble, taking account of country-specific
macroeconomic and institutional factors.

Author Contributions: S.H. and T.K. conceived and designed the experiments; K.T. performed the experiments;
S.H., T.K. and K.T. analyzed the data; K.T. contributed reagents/materials/analysis tools; K.T. wrote the paper.
Acknowledgments: We are grateful to four anonymous referees for their helpful comments and suggestions.
This work is supported by JSPS KAKENHI Grant Number 17K18564 and 17H00983.
Conflicts of Interest: The authors declare no conflict of interest. The founding sponsors had no role in the design
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the
decision to publish the results.
Sustainability 2018, 10, 1530 12 of 18
Sustainability 2018, 10, x FOR PEER REVIEW 13 of 19

Appendix A
Appendix A
Numbers and assets of banks for each category of bank failure probabilities.
Numbers and assets of banks for each category of bank failure probabilities.

Number of Banks Total Asset(millions of US dollar) Nominal GDP

0%-10% 10%-20% 20%-30% 30%-40% 40%-50% 50%-60% 60%-70% 70%-80% 80%-90% Country Total 0%-10% 10%-20% 20%-30% 30%-40% 40%-50% 50%-60% 60%-70% 70%-80% 80%-90% Country Total (billions of US dollar)
AFGHANISTAN 3 2 2 3 1 11 1,832 1,485 467 572 341 4,697 20.44
ALBANIA 2 8 2 1 13 1,361 8,371 932 385 11,049 13.30
ALGERIA 5 4 6 1 1 17 61,900 6,990 43,402 278 413 112,983 213.52
ANDORRA 3 1 4 11,106 1,566 12,672
ANGOLA 8 2 4 1 15 31,675 8,428 12,464 1,965 54,532 126.78
ANGUILLA 2 2 644 644
ANTIGUA AND BARBUDA 8 1 1 10 8,012 113 416 8,541 1.25
ARGENTINA 23 25 12 5 65 15,688 70,364 68,799 6,694 161,545 544.73
ARMENIA 8 3 2 1 2 16 3,531 731 716 206 671 5,855 11.64
ARUBA 1 1 2 580 1,037 1,617
AUSTRALIA 16 7 14 1 1 1 40 1,309,679 1,357,747 127,694 16,161 1,749 47,393 2,860,423 1441.95
AUSTRIA 184 43 29 21 11 1 289 283,139 239,865 344,623 58,462 35,246 3,782 965,117 437.58
AZERBAIJAN 21 5 1 1 28 10,969 14,605 1,623 367 27,564 75.25
BAHAMAS 23 1 3 3 30 31,715 160 3,656 4,655 40,186 8.51
BAHRAIN 4 5 2 1 12 14,232 104,028 729 1,248 120,237 33.84
BANGLADESH 24 13 3 40 59,439 25,723 3,566 88,728 183.82
BARBADOS 2 1 1 2 6 3,020 257 200 1,713 5,190 4.35
BELARUS 17 2 7 26 10,101 15,299 10,705 36,105 76.14
BELGIUM 19 9 5 4 5 42 652,514 329,070 273,204 3,069 15,149 1,273,006 532.39
BELIZE 4 1 1 6 621 386 358 1,365 1.70
BENIN 8 1 9 2,765 1,431 4,196 9.59
BERMUDA 2 2 4 21,317 1,775 23,092
BHUTAN 3 3 1,202 1,202 1.99
BOLIVIA 1 8 1 2 12 53 12,050 3,128 1,382 16,613 33.24
BOSNIA AND HERZEGOVINA 4 11 4 2 5 26 1,831 6,158 3,118 1,225 1,656 13,988 18.52
BOTSWANA 8 1 2 11 5,830 418 1,168 7,416 15.88
BRAZIL 83 36 17 5 141 1,376,576 599,875 29,506 74,468 2,080,425 2417.16
BRUNEI DARUSSALAM 1 1 2,706 2,706 17.10
BULGARIA 10 10 2 1 23 22,808 21,205 1,468 2,729 48,210 56.72
BURKINA FASO 5 3 1 9 1,158 3,029 611 4,798 12.48
BURUNDI 4 1 1 6 259 224 187 670 2.90
CAMBODIA 5 11 4 1 1 1 23 4,297 7,969 873 99 332 353 13,923 16.78
CAMEROON 5 1 3 3 12 959 433 2,957 2,936 7,285 31.63
CANADA 53 18 10 3 1 85 2,846,839 1,051,419 39,746 3,701 10 3,941,715 1783.78
CAPE VERDE 1 3 2 6 259 998 788 2,045 1.87
CAYMAN ISLANDS 19 1 20 22,196 12,376 34,572
CENTRAL AFRICAN REPUBLIC 2 2 258 258 1.73
CHAD 4 1 5 814 319 1,133 13.94
CHILE 11 6 10 6 2 35 73,641 192,501 54,875 5,914 20,391 347,322 258.68
CHINA 111 34 13 29 3 190 20,605,051 725,637 217,344 1,504,762 25,481 23,078,275 10430.71
COLOMBIA 10 10 4 3 27 123,640 94,505 15,177 551 233,873 377.87
COMOROS 1 1 0 0 0.70
CONGO 5 5 3,147 3,147 13.55
COSTA RICA 25 13 5 3 46 17,819 28,841 1,065 213 47,938 49.60
COTE D'IVOIRE 7 3 6 1 17 1,640 2,658 5,176 376 9,850 33.74
CROATIA 8 10 11 3 1 33 5,984 46,496 15,954 749 366 69,549 57.17
CUBA 5 1 6 10,490 5,073 15,563
Sustainability 2018, 10, 1530 13 of 18

Sustainability 2018, 10, x FOR PEER REVIEW 14 of 19

Number of Banks Total Asset(millions of US dollar) Nominal GDP

0%-10% 10%-20% 20%-30% 30%-40% 40%-50% 50%-60% 60%-70% 70%-80% 80%-90% Country Total 0%-10% 10%-20% 20%-30% 30%-40% 40%-50% 50%-60% 60%-70% 70%-80% 80%-90% Country Total (billions of US dollar)
CURACAO 13 1 2 16 11,647 3,703 1,463 16,813
CYPRUS 11 6 3 1 21 60,348 62,969 53,753 529 177,599 23.11
CZECH REPUBLIC 9 8 2 1 20 93,674 107,793 3,958 271 205,696 205.27
DEMOCRATIC PEOPLE'S REPUB 1 1 8,955 8,955
DEMOCRATIC REPUBLIC OF CON 5 8 1 14 913 3,109 131 4,153 35.92
DENMARK 34 25 10 8 77 55,066 99,672 134,603 716,900 1,006,241 346.12
DJIBOUTI 4 1 5 972 410 1,382 1.59
DOMINICA 1 1 375 375 0.52
DOMINICAN REPUBLIC 28 17 14 2 61 10,873 16,973 5,942 119 33,907 64.06
ECUADOR 6 9 4 2 21 6,285 24,036 5,940 1,936 38,197 100.92
EGYPT 16 4 1 2 23 155,170 11,839 1,634 47,083 215,726 301.39
EL SALVADOR 2 5 3 2 2 14 836 9,680 2,651 2,702 2,585 18,454 25.16
EQUATORIAL GUINEA 3 3 1,763 1,763 15.53
ERITREA 2 2 1,854 1,854 4.05
ESTONIA 5 1 1 1 8 22,246 233 311 311 23,101 26.51
ETHIOPIA 1 7 1 1 10 207 4,923 12,424 399 17,953 55.51
FEDERATED STATES OF MICRON 1 1 129 129 0.32
FIJI 1 1 405 405 4.53
FINLAND 38 3 2 2 45 574,425 152,017 349 65,636 792,427 272.77
FRANCE 128 51 28 18 2 3 230 6,953,615 4,961,192 3,770,293 51,662 785 1,484 15,739,031 2833.69
GABON 1 2 2 1 1 7 131 1,489 5,965 270 1,928 9,783 18.21
GAMBIA 7 1 8 346 104 450 0.82
GEORGIA 14 1 15 10,845 270 11,115 16.52
GERMANY 1,280 239 66 31 12 1,628 8,755,293 1,235,557 3,092,779 309,316 65,984 13,458,929 3874.44
GHANA 14 6 2 2 24 5,258 3,647 1,117 1,322 11,344 38.62
GIBRALTAR 2 2 5,474 5,474
GREECE 4 5 9 10,781 433,608 444,389 235.95
GRENADA 1 1 1 3 204 278 313 795 0.91
GUATEMALA 8 2 10 3 1 24 11,344 799 22,113 3,469 173 37,898 58.83
GUINEA 6 1 7 653 410 1,063 6.70
GUINEA BISSAU 1 1 114 114 1.11
GUYANA 2 1 3 1,152 270 1,422 3.08
HAITI 4 1 5 3,655 109 3,764 8.71
HONDURAS 7 3 6 1 1 1 19 7,548 2,522 5,150 1,526 372 306 17,424 19.51
HONG KONG 27 5 3 35 1,895,556 158,181 3,129 2,056,866 291.23
HUNGARY 8 5 4 5 4 1 27 3,295 73,216 9,724 20,421 2,624 695 109,975 138.35
ICELAND 14 2 1 17 159,422 249 8,070 167,741 17.16
INDIA 54 15 8 3 1 81 1,615,932 355,212 19,492 6,558 3,413 2,000,607 2042.56
INDONESIA 60 18 3 1 82 374,962 59,895 2,662 329 437,848 890.60
IRAQ 6 1 4 1 1 13 4,219 412 3,114 268 479 8,492 223.51
IRELAND 10 3 1 1 15 225,999 209,285 15,146 16,649 467,079 250.81
ISLAMIC REPUBLIC OF IRAN 1 1 7,332 7,332 416.49
ISRAEL 10 2 12 364,103 6,268 370,371 305.67
ITALY 399 93 21 9 1 1 1 525 2,007,276 1,315,729 354,360 346,926 269 274 1,265 4,026,099 2141.94
JAMAICA 6 1 7 9,088 256 9,344 13.89
JAPAN 555 33 5 593 15,230,263 193,546 25,979 15,449,788 4596.16
JORDAN 8 2 1 11 29,743 50,691 36,421 116,855 35.88
KAZAKHSTAN 18 9 6 1 34 44,640 46,100 17,985 372 109,097 217.87
KENYA 32 3 1 36 37,487 365 289 38,141 60.94
Sustainability 2018, 10, 1530 14 of 18

Sustainability 2018, 10, x FOR PEER REVIEW 15 of 19

Number of Banks Total Asset(millions of US dollar) Nominal GDP

0%-10% 10%-20% 20%-30% 30%-40% 40%-50% 50%-60% 60%-70% 70%-80% 80%-90% Country Total 0%-10% 10%-20% 20%-30% 30%-40% 40%-50% 50%-60% 60%-70% 70%-80% 80%-90% Country Total (billions of US dollar)
KIRIBATI 1 1 2 27 122 149 0.19
KOSOVO 3 2 5 1,517 773 2,290 7.40
KUWAIT 3 1 1 1 6 104,380 14,363 17,917 13,303 149,963 171.96
KYRGYZSTAN 5 2 2 1 10 432 483 330 223 1,468 7.47
LAO PEOPLE'S DEMOCRATIC RE 5 2 2 9 4,233 496 304 5,033 11.69
LATVIA 7 9 2 1 1 20 9,643 23,203 781 304 278 34,209 31.34
LEBANON 34 8 5 1 48 117,629 72,592 37,338 206 227,765 49.94
LESOTHO 2 1 1 4 148 311 604 1,063 2.22
LIBERIA 3 1 4 73 279 352 2.01
LIBYA 6 4 10 56,416 7,620 64,036 44.42
LIECHTENSTEIN 1 1 2 669 361 1,030
LITHUANIA 4 3 1 2 10 24,641 2,587 2,146 300 29,674 48.47
LUXEMBOURG 25 23 15 6 2 71 209,156 363,201 82,838 3,923 3,959 663,077 64.98
MACAO 6 2 1 1 10 25,935 31,416 112 150 57,613 55.52
MACEDONIA (FYROM) 8 5 2 1 16 3,641 2,533 927 318 7,419 11.33
MADAGASCAR 3 2 1 6 950 628 401 1,979 10.67
MALAWI 4 1 2 1 8 1,018 222 137 103 1,480 6.06
MALAYSIA 20 14 2 36 526,240 174,995 789 702,024 338.11
MALDIVES 1 1 1,179 1,179 3.06
MALI 6 3 1 10 916 2,374 644 3,934 14.45
MALTA 9 3 1 13 24,699 2,021 536 27,256 10.74
MAURITANIA 6 2 1 1 10 1,127 236 163 140 1,666 5.30
MAURITIUS 5 6 1 5 17 15,305 10,605 3,231 8,132 37,273 12.63
MEXICO 24 18 11 8 3 1 65 145,916 252,836 34,409 36,873 391 359 470,784 1297.85
MONGOLIA 7 1 1 9 1,648 2,150 781 4,579 12.20
MONTENEGRO 3 1 2 2 1 9 1,231 706 472 630 285 3,324 4.60
MOROCCO 8 3 4 1 16 84,532 95,839 23,250 3,049 206,670 110.01
MOZAMBIQUE 8 3 1 1 13 8,633 824 319 375 10,151 16.86
MYANMAR 11 1 1 13 551,525 179 5,522 557,226 65.75
NAMIBIA 3 2 1 6 3,523 4,881 1,883 10,287 13.19
NEPAL 8 7 4 7 3 1 30 3,599 3,834 2,311 4,889 1,069 477 16,179 19.76
NETHERLANDS 14 15 4 1 34 1,879,558 1,122,609 51,090 766 3,054,023 880.72
NEW ZEALAND 3 9 7 3 2 24 1,527 264,776 74,001 2,868 159 343,331 197.93
NICARAGUA 6 6 5,813 5,813 11.81
NIGER 5 1 1 7 691 431 429 1,551 8.26
NIGERIA 12 6 1 1 20 127,573 23,153 675 1,483 152,884 574.00
NORWAY 70 31 19 9 1 130 404,326 95,215 35,785 61,185 193 596,704 500.52
OMAN 2 2 2 6 37,348 11,462 17,832 66,642 77.77
PAKISTAN 18 3 1 22 110,300 3,922 9 114,231 243.38
PALESTINIAN TERRITORIES 1 1 1 3 2,425 669 279 3,373 0.25
PANAMA 52 11 11 4 78 41,083 48,724 11,305 3,995 105,107 49.17
PAPUA NEW GUINEA 4 4 10,752 10,752 16.65
PARAGUAY 3 5 5 5 18 3,696 7,057 3,807 4,037 18,597 30.88
PERU 6 4 1 9 20 18,411 66,660 690 22,673 108,434 202.90
PHILIPPINES 33 11 10 3 57 128,842 45,889 50,488 19,347 244,566 284.78
POLAND 21 12 4 3 5 2 47 169,073 172,667 28,596 4,302 7,693 5,467 387,798 544.86
PORTUGAL 62 36 12 3 2 115 147,780 301,485 15,495 191,662 79,598 736,020 230.48
QATAR 5 2 7 60,552 181,884 242,436 210.11
REPUBLIC OF KOREA 8 6 4 1 5 24 402,300 382,133 532,250 1,031 22,478 1,340,192 1410.38
REPUBLIC OF MOLDOVA 7 3 1 2 13 2,648 311 26 875 3,860 7.98
Sustainability 2018, 10, 1530 15 of 18
Sustainability 2018, 10, x FOR PEER REVIEW 16 of 19

Number of Banks Total Asset(millions of US dollar) Nominal GDP

0%-10% 10%-20% 20%-30% 30%-40% 40%-50% 50%-60% 60%-70% 70%-80% 80%-90% Country Total 0%-10% 10%-20% 20%-30% 30%-40% 40%-50% 50%-60% 60%-70% 70%-80% 80%-90% Country Total (billions of US dollar)
ROMANIA 6 9 7 3 2 1 28 14,088 33,901 26,593 13,572 4,948 334 93,436 199.37
RUSSIAN FEDERATION 132 45 344 283 24 3 831 1,029,291 177,322 124,970 58,073 10,672 238 1,400,566 2029.62
RWANDA 5 1 2 8 1,257 285 409 1,951 7.89
SAINT KITTS AND NEVIS 1 1 2 195 1,356 1,551 0.85
SAINT LUCIA 4 1 5 1,323 1,379 2,702 1.40
SAINT VINCENT AND THE GRENA 1 1 301 301 0.73
SAMOA 3 3 309 309 0.82
SAN MARINO 2 3 5 309 4,542 4,851 1.85
SAO TOME AND PRINCIPE 3 3 37 37 0.34
SAUDI ARABIA 8 8 432,362 432,362 753.83
SENEGAL 3 2 5 1 1 12 820 778 3,461 699 1,365 7,123 15.36
SERBIA 12 7 6 3 2 1 31 10,296 5,159 11,822 1,549 1,883 319 31,028 44.21
SEYCHELLES 4 2 6 646 580 1,226 1.35
SIERRA LEONE 10 1 11 709 71 780 4.75
SINGAPORE 17 3 1 21 928,556 30,223 310 959,089 306.36
SINT MAARTEN 1 1 623 623
SLOVAKIA 9 3 2 14 45,180 19,143 4,102 68,425 100.33
SLOVENIA 6 10 1 2 19 6,542 36,068 1,300 860 44,770 49.57
SOUTH AFRICA 10 4 1 1 16 315,712 649 3,651 27 320,039 350.14
SOUTH SUDAN 5 5 568 568 13.90
SPAIN 80 32 19 6 4 1 142 1,970,542 1,219,637 1,939,048 12,757 12,150 13,194 5,167,328 1383.54
SRI LANKA 10 4 4 18 39,683 10,147 1,533 51,363 75.05
SUDAN 8 1 9 2,808 182 2,990 74.36
SURINAME 1 1 2 4 90 1,237 805 2,132 5.21
SWAZILAND 1 3 4 310 840 1,150 4.42
SWEDEN 74 7 1 3 85 661,030 304,789 563 5,211 971,593 571.10
SWITZERLAND 237 67 11 11 1 327 1,584,496 1,135,554 56,579 10,637 271 2,787,537 701.23
SYRIAN ARAB REPUBLIC 10 3 13 27,997 1,780 29,777
TAIWAN 39 4 1 1 1 46 1,451,115 227,072 25,829 4,330 382 1,708,728 530.04
TAJIKISTAN 5 1 1 7 1,395 60 340 1,795 9.24
THAILAND 5 6 9 2 1 23 32,781 265,736 284,832 7,877 8,435 599,661 404.32
TOGO 3 2 1 6 509 1,436 194 2,139 4.61
TONGA 1 1 109 109 0.43
TRINIDAD AND TOBAGO 7 1 2 10 21,232 3,900 6,166 31,298 27.27
TUNISIA 6 6 3 2 17 6,111 13,238 11,328 5,922 36,599 47.61
TURKEY 23 6 4 2 35 630,271 167,915 6,137 146 804,469 798.33
TURKMENISTAN 1 1 2 5,002 2,167 7,169 46.22
TUVALU 1 1 20 20 0.04
UGANDA 16 1 2 19 5,463 266 194 5,923 27.52
UKRAINE 19 65 58 6 3 151 15,676 28,102 38,077 1,086 65 83,006 132.34
UNITED ARAB EMIRATES 8 7 4 19 215,505 183,895 73,201 472,601 399.45
UNITED KINGDOM 63 34 25 18 6 146 6,330,619 353,267 2,458,011 1,174,218 1,734 10,317,849 2991.69
UNITED REPUBLIC OF TANZANIA 25 2 3 2 1 33 13,290 504 733 255 69 14,851 48.09
UNITED STATES OF AMERICA 5,857 778 208 76 24 6 2 1 1 6,953 12,991,678 1,654,967 267,778 1,374,887 83,455 1,808 8,177 690 113 16,383,553 17348.08
URUGUAY 13 10 4 27 24,220 11,196 3,709 39,125 57.47
UZBEKISTAN 9 4 3 2 1 19 4,568 3,898 4,236 493 5,248 18,443 63.10
VANUATU 1 1 2 252 141 393 0.82
VENEZUELA 15 1 20 2 38 37,641 2,310 523,950 90,935 654,836 250.28
VIETNAM 16 21 4 3 3 1 48 66,397 151,559 4,899 5,606 5,326 123 233,910 185.90
YEMEN 5 1 6 3,441 2,227 5,668 43.23
ZAMBIA 9 5 1 1 16 4,183 2,050 360 289 6,882 27.14
ZIMBABWE 9 1 3 1 14 2,733 159 1,019 287 4,198 14.20
Sustainability 2018, 10, 1530 16 of 18

Appendix B
List of variables used to build the random forests EWS.
List of incorporated bank-level variables

• Interest Income on Loans/Average Gross Loans

• Interest Expense on Customer Deposits/Average Customer Deposits
• Interest Income/Average Earning Assets
• Interest Expense/Average Interest-bearing Liabilities
• Net Interest Income/Average Earning Assets
• Net Int. Increase Less Loan Impairment Charges/Average Earning Assets
• Net Interest Increase Less Preferred Stock Dividend/Average Earning Assets
• Non-Interest Income/Gross Revenues
• Non-Interest Expense/Gross Revenues
• Non-Interest Expense/Average Assets
• Pre-impairment Operating Profit/Average Equity
• Pre-impairment Operating Profit/Average Total Assets
• Loans and securities impairment charges/Pre-impairment Operating Profit
• Operating Profit/Average Equity
• Operating Profit/Average Total Assets
• Taxes/Pre-tax Profit
• Pre-Impairment Operating Profit/Risk Weighted Assets
• Operating Profit/Risk Weighted Assets
• Net Income/Average Total Equity
• Net Income/Average Total Assets
• Fitch Comprehensive Income/Average Total Equity
• Fitch Comprehensive Income/Average Total Assets
• Net Income/Risk Weighted Assets
• Fitch Comprehensive Income/Risk Weighted Assets
• Tangible Common Equity/Tangible Assets
• Tier 1 Regulatory Capital Ratio
• Total Regulatory Capital Ratio
• Equity/Total Assets
• Cash Dividends Paid and Declared/Net Income
• Cash Dividend Paid and Declared/Fitch Comprehensive Income
• Net Income–Cash Dividends/Total Equity
• Growth of Total Assets
• Growth of Gross Loans
• Impaired Loans (NPLs)/Gross Loans
• Reserves for Impaired Loans/Gross loans
• Reserves for Impaired Loans/Impaired Loans
• Impaired Loans less Reserves for Impaired Loans/Equity
• Loan Impairment Charges/Average Gross Loans
• Net Charge-offs/Average Gross Loans
• Impaired Loans + Foreclosed Assets/Gross Loans + Foreclosed Assets
• Loans/Customer Deposits
• Customer Deposits/Total Funding excluding Derivatives
Sustainability 2018, 10, 1530 17 of 18

List of eliminated bank-level variables (variables with missing values more than 50% of observations)
• Net Income/Average Total Assets + Average Managed Securitized Assets
• Fitch Core Capital/Weighted Risks
• Fitch Eligible Capital/Weighted Risks
• Core Tier 1 Regulatory Capital Ratio
• Cash Dividends and Share Repurchase/Net Income
• Interbank Assets/Interbank Liabilities

Appendix C
Evaluating classification accuracy
To evaluate the classification accuracy of the random forests EWS, we used K-fold cross-validation,
which is a standard resampling technique used for estimating model performance. The basic idea
involves using parts of the sample data to fit the model (training set) and the remaining part to estimate
the prediction error of the model (validation set). First, we randomly split the observations into K
folds of roughly equal size. Then, we treated one of the K folds as a validation set and fit the model
on the remaining K–1 folds. We calculated the prediction error of the observations in the validation
set. We repeated the process K times to obtain K different estimates of the prediction error. The K-fold
cross-validation estimate of the prediction error is the average of these values. Since the typical choice
of K is either 5 or 10, we used a 10-fold cross-validation. Using the same setup, we compared the
performance of the random forests EWS with that of conventional EWSs based on a logistic regression
and a decision tree. The random forests model produced an accuracy rate of 93.64%, whereas the
logistic regression and decision tree produced accuracy rates of 65.73% and 73.75%, respectively. The
result clearly indicates that random forests can build more reliable EWSs than conventional methods.
We also conducted a historical back test to evaluate the classification accuracy. More specifically,
we predicted the bank status (active or inactive) in 2013 and 2014 based on banks’ financial statements
in 2013, excluding those that became inactive before 2013. We evaluated performance in terms of
accuracy by comparing the predicted bank status with the actual status in 2013 and 2014. The random
forests model produced an accuracy rate of 85.27%, whereas the logistic regression and decision tree
produce accuracy rates of 53.44% and 56.43%, respectively. Once again, the result clearly indicates that
the random forest EWS outperforms the conventional EWSs.

References
1. International Monetary Fund (IMF). Macroprudential Policy: An Organizing Framework; International Monetary
Fund: Washington, DC, USA, March 2011.
2. Committee on the Global Financial System. Operationalising the Selection and Application of Macroprudential
Instruments; CGFS Paper No. 48; Bank for International Settlements: Basel, Switzerland, December 2012.
3. Frankel, J.A.; Rose, A.K. Currency crashes in emerging markets: An empirical treatment. J. Int. Econ. 1996,
41, 351–366. [CrossRef]
4. Kaminsky, G.L.; Lizondo, S.; Reinhart, C.M. The leading indicators of currency crises. IMF Staff Pap. 1998, 45,
1–48. [CrossRef]
5. Alessi, L.; Detken, C. Identifying Excessive Credit Growth and Leverage; ECB Working Paper No. 1723; European
Central Bank: Frankfurt, Germany, August 2014.
6. Bussiere, M.; Fratzscher, M. Towards a new early warning system of financial crises. J. Int. Money Financ.
2006, 25, 953–973. [CrossRef]
7. Rose, A.K.; Spiegel, M.M. Cross-country causes and consequences of the crisis: An update. Eur. Econ. Rev.
2011, 55, 309–324. [CrossRef]
8. Frankel, J.; Saravelos, G. Can leading indicators assess country vulnerability? Evidence from the 2008–2009
global financial crisis. J. Int. Econ. 2012, 87, 216–231. [CrossRef]
9. Duca, M.L.; Peltonen, T.A. Assessing systemic risks and predicting systemic events. J. Bank. Financ. 2013, 37,
2183–2195. [CrossRef]
Sustainability 2018, 10, 1530 18 of 18

10. Mathias, D.; Mikael, J. Evaluating early warning indicators of banking crisis: Satisfying policy requirements.
Int. J. Forecast. 2014, 30, 759–780.
11. Hastie, T.; Tibshirani, R.; Friedman, J. The Elements of Statistical Learning; Springer: New York, NY, USA, 2008.
12. Ghosh, S.R.; Ghosh, A.R. Structural vulnerabilities and currency crises. IMF Staff Pap. 2002, 50, 481–506.
[CrossRef]
13. Frankel, J.; Wei, S.-J. Managing macroeconomic crises: Policy lessons. In Managing Economic Volatility and
Crises: A Practitioner’s Guide; Aizenman, J., Pinto, B., Eds.; Cambridge University Press: Cambridge, UK,
2005; pp. 315–405.
14. James, G.; Witten, D.; Hastie, T.; Tibshirani, R. An Introduction to Statistical Learning; Springer: New York, NY,
USA, 2013.
15. Varian, H.R. Big data: New tricks for econometrics. J. Econ. Perspect. 2014, 28, 3–28. [CrossRef]
16. Einav, L.; Levin, J. The data revolution and economic analysis. Innov. Policy Econ. 2014, 14, 1–24. [CrossRef]
17. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
18. Tanaka, K.; Higashide, T.; Kinkyo, T.; Hamori, S. Forecasting the Vulnerability of Industrial Economic
Activities: Predicting the Bankruptcy of Companies. J. Manag. Inf. Decis. Sci. 2017, 20, 1–24.
19. Tanaka, K.; Kinkyo, T.; Hamori, S. Random forests-based early warning system for bank failures. Econ. Lett.
2016, 148, 118–121. [CrossRef]
20. Berger, A.N.; Bouwman, C.H.S. How does capital affect bank performance during financial crises? J. Financ. Econ.
2013, 109, 146–176. [CrossRef]
21. Vazquez, F.; Federico, P. Bank funding structures and risk: Evidence from the global financial crisis. J. Bank. Financ.
2015, 61, 1–14. [CrossRef]
22. Altman, E.I. Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. J. Financ.
1968, 23, 189–209. [CrossRef]
23. Kuhn, M.; Johnson, K. Applied Predictive Modeling; Springer: New York, NY, USA, 2013.

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).