You are on page 1of 10

JOURNAL OF MULTI-CRITERIA DECISION ANALYSIS

J. Multi-Crit. Decis. Anal. 14: 103–111 (2006)
Published online in Wiley InterScience
(www.interscience.wiley.com) DOI: 10.1002/mcda.405
AMulticriteria Decision Framework for Measuring Banks’
Soundness Around theWorld
CHRYSOVALANTISGAGANIS
a
, FOTIOSPASIOURAS
b,
* and CONSTANTINZOPOUNIDIS
a
a
Financial Engineering Laboratory, Department of Production Engineering and Management,Technical
University of Crete, University Campus, Chania 73100, Greece
b
School of Management, University of Bath, Claverton Down, Bath BA27AY, UK
ABSTRACT
In this paper, we use a sample of 894 banks from 79 countries to develop a multicriteria decision aid model, for the
classification of banks into three groups on the basis of their soundness. The model is developed with the UTilite´ s
Additives DIScriminantes (UTADIS) method, through a 10-fold cross-validation procedure using six financial and
four non-financial variables. The ratings of Fitch form the basis for assigning banks into the three groups. The
results indicate that the asset quality (as measured by loan loss provisions), capitalization, and the market where
banks operate are the most important criteria (in terms of weights) in classifying the banks. Profitability and
efficiency in expenses management are also important attributes, whereas size and listing in a stock exchange are the
least important ones. UTADIS achieves higher classification accuracies than discriminant analysis and ordinary
logistic regression which are used for benchmarking purposes. Copyright # 2007 John Wiley & Sons, Ltd.
KEY WORDS: multicriteria classification of banks’ soundness; UTADIS
1. INTRODUCTION
Over the last years, various studies have attempted
to develop models for assessing bank soundness
either by predicting bank failure or by examining
the credit ratings of banks. Studies falling within
the first category have used various methodologies
such as discriminant analysis (DA) (Sinkey, 1975;
Canbas et al., 2005), logit analysis (e.g. Rose and
Kolari, 1985; Pantalone and Platt, 1987; Canbas
et al., 2005), probit analysis (Cole and Gunther,
1998; Canbas et al., 2005), multidimensional
scaling approach (Mar-Molinero and Serrano-
Cinca, 2001), neural networks (Tam and Kiang,
1992), trait recognition (Kolari et al., 2002; Lanine
and Vander Vennet, 2006), multicriteria decision
aid (Kosmidou and Zopounidis, 2007) and neural
fuzzy systems (Tung et al., 2004). Many of these
studies have been successful in predicting bank-
ruptcy. However, one common drawback is that
they had concentrated on the classification of
banks into two groups, failed and non-failed.
Obviously, the classification of banks as ‘bad’ or
‘good’ reduces the usefulness of the model.
With respect to the second category, although
several studies have examined the ratings assigned
to bonds (Horrigan, 1966; Pinches and Mingo,
1973; Kaplan and Urwitz, 1979; Belkaoui, 1983;
Ederington et al., 1987; Kim, 1993; Huang et al.,
2004) or countries (e.g. Hammer et al., 2006; Yim
and Mitchell, 2005) there have been only a few
studies that examine the individual ratings as-
signed to banks such as the ones of Poon et al.
(1999), Poon and Firth (2005), Pasiouras et al.
(2006, 2007) and Demirguc-Kunt et al. (2006).
However, with the exception of Pasiouras et al.
(2007) who developed a classification model for
the rating of Asian banks, these studies are of a
more explanatory nature focusing on the determi-
nants of ratings rather than on the correct
classification of banks. More precisely, Poon
et al. (1999) examined the determinants of
Moody’s ratings, using logistic regression, while
Poon and Firth (2005) focused on the differences
between solicited and unsolicited ratings. Pa-
siouras et al. (2006) focused on the impact of
bank regulations and supervisory framework on
the ratings, while Demirguc-Kunt et al. (2006)
focused on the relation between compliance with
Basel core principles and bank soundness.
*Correspondence to: School of Management, University
of Bath, Claverton Down, Bath BA2 7AY, UK.
E-mail: f.pasiouras@bath.ac.uk
Copyright # 2007 John Wiley & Sons, Ltd.
While the development of a bankruptcy prediction
model or a model to replicate all the ratings of a
credit agency is beyond the scope of the present
paper, our work is nevertheless related to these two
strands of the literature. The objective of the present
study is the development of quantitative models for
the classification of banks into different groups on
the basis of their soundness. While our model is
based on the ratings of Fitch, to assess the soundness
of banks,
1
we do not attempt to replicate all the
ratings of Fitch, as the heterogeneous sample used in
our study, consisting of 894 banks from 79 countries,
could result in poor classifications.
2
We, therefore,
classify banks into three general groups. The first
group contains very strong or strong banks; the
second one contains adequate banks, while the third
group contains banks with weaknesses or serious
problems. While we acknowledge that this approach
might reduce the information provided by the
model, we believe that it does not reduce the
applicability of the model and the importance of
our study. For example, several studies have high-
lighted the importance of developing early warning
systems to identify troubled banks (e.g. Kolari et al.,
2002; Tung et al., 2004; Canbas et al., 2005; Lanine
and Vander Vennet, 2006). By focusing on non-
failed banks, and distinguishing between healthy,
adequate and troubled banks, our model can reduce
the expected cost of bank failure, either by minimiz-
ing the costs to the public or by taking actions to
prevent failure (Thomson, 1991). For example, Ravi
Kumar and Ravi (2007) mention, ‘As a bank or firm
becomes more and more insolvent, it gradually
enters a danger zone. Then, changes to its operations
and capital structure must be made in order to keep
it solvent’ (p. 1). Hence, the model developed in the
present study could be used in future applications to
provide an assessment of the soundness of banks not
rated by Fitch or other agencies (e.g. Moody’s) as
well as to monitor changes in the status of banks
from one year to another.
The model is developed with UTilite´ s Additives
DIScriminantes (UTADIS) multicriteria technique
following a 10-fold cross-validation procedure.
UTADIS is well suited for examining the sound-
ness of banks for several reasons. First, the groups
in our study are defined in an ordinal way, in
the sense that banks classified into the first group
are preferred to banks classified into the second
group, and so on. Traditional statistical classifica-
tion methods as well as popular machine learning
techniques (e.g. neural networks, rule induction
algorithms and decision trees) cannot cope with
this kind of information. On the other hand,
UTADIS is well suited to the study of ordinal
classification problems. Second, UTADIS is not
based on statistical assumptions that often cause
problems to the application of statistical methods,
3
such as the normality of the variables or the group
dispersion matrices (e.g. DA) and is not sensitive
to multicollinearity or outliers (e.g. logit analysis).
Third, it can easily incorporate qualitative data.
The rest of the paper is as follows: Section 2
presents the sample and the variables used in the
study, while Section 3 outlines the UTADIS
technique. Section 4 discusses the empirical results,
and Section 5 concludes the study.
2. SAMPLE AND VARIABLES
2.1. Sample
The data set consists of those banks that had
available financial and non-financial data and
1
Demirguc-Kunt et al. (2006) also used ratings to
measure bank soundness. However, they used the
ratings of Moody’s and focused on the relation between
compliance with Basel core principles and bank sound-
ness and did not develop a classification model.
2
One could argue that we could limit the heterogeneity
in the sample by decreasing the number of countries and
increasing the number of banks in the sample. However,
this is not possible since we have considered all the
banks rated by Fitch, with available data in Bankscope,
and by definition, we can only consider rated banks.
Hence, decreasing the number of countries is not an
alternative as this approach would limit the number of
banks in the sample resulting in an inefficient estimation
of the model.
3
Barniv and McDonald (1999) summarized some of the
problems related to discriminant, logit and probit that
were mentioned in previous studies. Logit and Probit are
sensitive to: (a) data properties, such as departure from
normality of financial variables (Frecka and Hopwood,
1983; Richardson and Davidson, 1984; Hopwood et al.,
1988); (b) overall small sample size (Noreen, 1988; Stone
and Rasp, 1991); (c) multicollinearity (Aldrich and
Nelson, 1984; Stone and Rasp, 1991). The basic
assumptions of discriminant analysis (DA) such as
normality, symmetry and equal covariance matrices
are also usually violated. Hopwood et al. (1988) pointed
out that DA is generally sensitive to departure from
normality and both logit and probit analyses are
sensitive to extreme non-normality.
C. GAGANIS ET AL. 104
Copyright # 2007 John Wiley & Sons, Ltd. J. Multi-Crit. Decis. Anal. 14: 103–111 (2006)
DOI: 10.1002/mcda
Fitch individual bank ratings in Bankscope
database.
4
The ratings correspond to October
2004, while the bank-specific characteristics corre-
spond to end of 2003 or March of 2004 depending
on the date of publication of the annual report.
The above selection criteria yielded a sample of
894 banks operating in 79 countries.
Table I presents the definitions of Fitch indivi-
dual bank ratings along with the coding used in the
present study. The ratings are based on A–E scale
and represent Fitch’s view on the likelihood that the
bank would fail, and therefore require support to
prevent it from defaulting. Fitch may also assign
the following intermediate ratings: A/B, B/C, C/D
and D/E. As mentioned earlier, the purpose of the
present study is not to explain or replicate the
ratings of Fitch, but rather to use them as the basis
for the development of a general model to assess the
soundness of banks. We, therefore, classify the
banks into three broad groups. The first consists of
banks with ratings A and B (i.e. very strong or
strong banks), the second with banks rated C (i.e.
adequate banks), and the third with banks rated D
and E (i.e. banks with weaknesses or serious
problems). Table II presents the number of banks
in sample by country and group.
2.2. Variables
Table III presents the variables used in the
models as criteria of banks’ soundness. Credit
agencies, auditors and bank regulators tend to
evaluate banks’ performance on the basis of
the CAMEL model that stands for the acronyms
of Capital, Asset quality, Management, Earnings
and Liquidity. We follow the same approach
and select financial variables that proxy for the
four of the five dimensions, as well as size.
Management has not been included in the analysis
due to its qualitative nature and the subjective
analysis that is required. Credit agencies point out
that during their rating they consider various
non-financial characteristics such as the environ-
ment in which banks operate, ownership and
franchise power. Consequently, we use additional
non-financial variables to proxy for these
characteristics.
Table I. Definitions of Fitch’s bank individual ratings
Fitch
rating
Coded in the
present study
Definition
A Group 1 A very strong bank. Characteristics may include outstanding profitability and balance sheet
integrity, franchise, management, operating environment or prospects
B Group 1 A strong bank. There are no major concerns regarding the bank. Characteristics may include
strong profitability and balance sheet integrity, franchise, management, operating environment or
prospects
C Group 2 An adequate bank, which, however, possesses one or more troublesome aspects. There may be
some concerns regarding its profitability and balance sheet integrity, franchise, management,
operating environment or prospects
D Group 3 A bank, which has weaknesses of internal and/or external origin. There are concerns regarding its
profitability and balance sheet integrity, franchise, management, operating environment or
prospects. Banks in emerging markets are necessarily faced with a greater number of potential
deficiencies of external origin
E Group 3 A bank with very serious problems, which either requires or is likely to require external support
Note: Fitch also uses the following intermediate assignments among the major five categories: A/B, B/C, C/D, D/E.
4
Using Bankscope has two main advantages. First, it
has information for a very large number of banks,
accounting for about 90% of total assets in each country
(Claessens et al., 2001). Second, and most important, the
financial information at the bank level is presented in
standardized formats, after adjusting for differences in
accounting and reporting standards. The data compiled
by Bankscope use financial statements and notes found
in audited annual reports. Each country in the Bank-
scope database has its own data template, thus allowing
for differences in the reporting and accounting conven-
tions. The data are then converted to a ‘global format’
using a globally standardized template derived from the
country-specific templates. The global format also
provides standard financial ratios, which can be
compared across banks and between countries. There-
fore, Bankscope is the most comprehensive database
that allows cross-country comparisons (Claessens et al.,
2001).
A MULTICRITERIA DECISION FRAMEWORK 105
Copyright # 2007 John Wiley & Sons, Ltd. J. Multi-Crit. Decis. Anal. 14: 103–111 (2006)
DOI: 10.1002/mcda
T
a
b
l
e
I
I
.
B
a
n
k
s
b
y
c
o
u
n
t
r
y
a
n
d
r
i
s
k
g
r
o
u
p
G
r
o
u
p
1
G
r
o
u
p
2
G
r
o
u
p
3
G
r
o
u
p
1
G
r
o
u
p
2
G
r
o
u
p
3
G
r
o
u
p
1
G
r
o
u
p
2
G
r
o
u
p
3
A
r
g
e
n
t
i
n
a
0
0
4
H
o
n
g
K
o
n
g
1
0
8
0
P
h
i
l
i
p
p
i
n
e
s
0
6
8
A
u
s
t
r
a
l
i
a
8
1
0
H
u
n
g
a
r
y
0
1
1
P
o
l
a
n
d
0
2
4
A
u
s
t
r
i
a
2
0
0
I
c
e
l
a
n
d
0
2
0
P
o
r
t
u
g
a
l
6
3
0
A
z
e
r
b
a
i
j
a
n
0
0
1
I
n
d
i
a
0
6
2
0
Q
a
t
a
r
1
2
0
B
a
h
r
a
i
n
1
1
0
I
n
d
o
n
e
s
i
a
0
5
7
R
o
m
a
n
i
a
0
0
3
B
a
n
g
l
a
d
e
s
h
0
0
3
I
r
e
l
a
n
d
6
0
0
R
u
s
s
i
a
n
F
e
d
e
r
a
t
i
o
n
0
2
2
3
B
e
l
a
r
u
s
0
0
2
I
s
r
a
e
l
0
2
1
S
a
u
d
i
A
r
a
b
i
a
7
0
1
B
e
l
g
i
u
m
7
0
0
I
t
a
l
y
1
1
1
5
1
S
i
n
g
a
p
o
r
e
4
0
0
B
e
n
i
n
0
1
0
J
a
p
a
n
1
6
1
7
S
l
o
v
e
n
i
a
1
5
0
B
r
a
z
i
l
0
9
1
J
o
r
d
a
n
2
1
0
S
o
u
t
h
A
f
r
i
c
a
5
3
0
B
u
l
g
a
r
i
a
0
2
1
K
a
z
a
k
h
s
t
a
n
0
1
5
S
p
a
i
n
4
4
4
1
C
a
n
a
d
a
6
1
0
K
e
n
y
a
0
0
4
S
r
i
L
a
n
k
a
0
0
5
C
h
i
l
e
3
1
0
K
o
r
e
a
(
R
e
p
u
b
l
i
c
o
f
)
2
7
2
S
w
e
d
e
n
7
0
0
C
h
i
n
a
0
0
1
4
K
u
w
a
i
t
2
6
0
S
w
i
t
z
e
r
l
a
n
d
5
1
0
C
r
o
a
t
i
a
0
1
0
L
a
t
v
i
a
0
1
1
T
a
i
w
a
n
5
6
1
0
C
y
p
r
u
s
0
3
1
L
e
b
a
n
o
n
0
2
0
T
h
a
i
l
a
n
d
0
4
6
C
z
e
c
h
R
e
p
u
b
l
i
c
0
2
0
L
i
t
h
u
a
n
i
a
0
1
3
T
u
n
i
s
i
a
0
0
2
D
e
n
m
a
r
k
5
0
0
L
u
x
e
m
b
o
u
r
g
2
0
0
T
u
r
k
e
y
0
5
1
9
D
o
m
i
n
i
c
a
n
R
e
p
u
b
l
i
c
0
0
2
M
a
l
a
y
s
i
a
2
5
4
U
k
r
a
i
n
e
0
0
4
E
g
y
p
t
0
1
1
M
a
l
t
a
0
1
1
U
n
i
t
e
d
A
r
a
b
E
m
i
r
a
t
e
s
1
5
0
E
l
S
a
l
v
a
d
o
r
0
1
4
M
e
x
i
c
o
2
6
1
U
n
i
t
e
d
K
i
n
g
d
o
m
4
2
4
0
E
s
t
o
n
i
a
1
0
0
N
e
t
h
e
r
l
a
n
d
s
9
2
1
U
S
A
2
1
9
2
2
4
F
i
n
l
a
n
d
3
0
0
N
e
w
Z
e
a
l
a
n
d
3
0
0
V
e
n
e
z
u
e
l
a
0
0
8
F
r
a
n
c
e
1
3
9
0
N
o
r
w
a
y
8
0
0
V
i
e
t
n
a
m
0
0
4
G
e
o
r
g
i
a
(
R
e
p
.
o
f
)
0
0
2
O
m
a
n
0
4
0
T
o
t
a
l
4
6
6
2
1
4
2
1
4
G
e
r
m
a
n
y
7
1
9
2
P
a
k
i
s
t
a
n
0
0
4
G
r
e
e
c
e
3
2
0
P
a
n
a
m
a
0
4
1
C. GAGANIS ET AL. 106
Copyright # 2007 John Wiley & Sons, Ltd. J. Multi-Crit. Decis. Anal. 14: 103–111 (2006)
DOI: 10.1002/mcda
2.2.1. Financial variables. CAP is the equity to
total assets ratio that serves as a measure of capital
strength.
5
PROVIS is the loan loss provisions to
net interest revenue ratio that is a measure of asset
quality. ROAA is the return on average assets that
is a classical measure of profitability. EXPENSES
is the cost-to-income ratio
6
that reveals the
efficiency in managing expenses. LIQ corresponds
to the liquid assets
7
to customer and short-term
funding ratio that shows the percentage of
customer and short-term funding that could be
met if they were withdrawn suddenly. SIZE is the
logarithm of total assets and is a measure of size.
2.2.2. Non-financial variables. SUBS is the number
of subsidiaries that is used as a proxy for the
diversification of business and franchise. OWNERS
is the number of institutional shareholders, and
LIST is a dummy variable indicating whether the
bank is listed on a stock exchange (LIST ¼ 1) or not
(LIST ¼ 0). Both variables are used as proxies of
corporate governance and ownership. Our last
variable is a dummy variable indicating whether
the banks operate in a developed (MARKET ¼ 1)
or developing (MARKET ¼ 0) market.
3. METHODOLOGY
The most common approach to address multi-
criteria classification problems is to develop a
criteria aggregation model based on absolute
judgements, which provides a rule for the classi-
fication of the alternatives on the basis of their
comparison with some reference profiles (cut-off
points) that distinguish the classes. Following this
approach, the objective of the UTADIS method is
to develop a classification model in the form of an
additive value function.
VðxÞ ¼
X
n
i¼1
w
i
v
i
ðx
i
Þ
where x ¼ ðx
1
; x
2
; . . . ; x
n
Þ is the vector of decision
attributes (financial ratios), w
1
; w
2
; . . . ; w
n
are
attributes’ weights defined such that w
i
50 and
P
i
w
i
¼ 1, and v
1
; v
2
; . . . ; v
n
are the attributes’
marginal value functions normalized in [0, 1].
Each marginal value function, v
i
, is used to assess
the partial performance of each bank in attribute
x
i
in an increasing 0–1 scale.
Given the global values of the banks as defined
by the estimated additive value function, their
classification into q groups C
1
; C
2
; . . . ; C
q
can be
performed in a straightforward way through the
introduction of qÀ1 value cut-off points
t
1
; t
2
; . . . ; t
qÀ1
, such that
Vðx
i
Þ5t
1
, bank i belongs to group C
1
t
2
4Vðx
i
Þ5t
1
, bank i belongs to group C
2
.
.
.
.
.
.
Vðx
i
Þ5t
qÀ1
,bank i belongs to group C
q
The estimation of the additive value function
and the cut-off thresholds is performed through
linear programming techniques. The objective of
the method is to develop the additive value model
so that the above classification rules can reproduce
the predetermined grouping of the banks as
accurately as possible. Therefore, a linear pro-
gramming formulation is employed to minimize
Table III. List of criteria (variables)
Financial
CAP Equity/total assets
PROVIS Loan loss provisions/net interest revenue
ROAA Return on average assets
EXPENSES Cost-to-income ratio
LIQ Liquid assets/customer and short-term funding
SIZE Logarithm of total assets
Non-financial
SUBS The number of subsidiaries
OWNERS The number of institutional shareholders
LIST Dummy variable indicating whether the bank
is listed (LIST ¼1) in a stock exchange or not
(LIST ¼ 0)
MARKET Dummy variable indicating whether the bank
is operating in a developed (MARKET ¼ 1) or
developing (MARKET ¼ 0) market
5
Probably, the employment of a risk-weighted ratio
such as the Tier 1 ratio would be more appropriate.
However, due to too many missing values for Tier 1, we
rely on EQAS that is considered one of the basic ratios
whose use dates back to the 1900s, and is still being used
in many recent studies in banking.
6
Cost refers to overheads that are the expenses for
running business, such as staff salaries and benefits, rent
expenses, equipment expenses and other administrative
expenses.
7
These are generally short-term assets that can be easily
converted into cash, such as cash itself, deposits with the
central bank, treasury bills, other government securities
and interbank deposits among others.
A MULTICRITERIA DECISION FRAMEWORK 107
Copyright # 2007 John Wiley & Sons, Ltd. J. Multi-Crit. Decis. Anal. 14: 103–111 (2006)
DOI: 10.1002/mcda
the sum of all violations of the above classification
rules for all the banks in the training sample.
Detailed description of the mathematical program-
ming formulation can be found in the works of
Zopounidis and Doumpos (1999) and Doumpos
and Zopounidis (2002).
4. EMPIRICAL RESULTS
Table IV shows descriptive statistics while distin-
guishing between the three groups of banks, as
well as between developed and developing mar-
kets. We also present the results of a Kruskal–
Table IV. Descriptive statistics and Kruskal–Wallis test (total sample): (A) Continuous variables and (B)
Categorical variables (no.)
1 2 3
Mean Std. Deviation Mean Std. Deviation Mean Std. Deviation Kruskal–Wallis
chi-square
(A)
SIZE
Total 4.388 0.669 3.843 0.689 3.535 0.865 188.992
***
Developed 4.398 0.678 4.010 0.764 4.462 0.885 26.279
***
Developing 4.221 0.506 3.650 0.529 3.328 0.713 50.538
***
PROVIS
Total 13.664 15.556 23.103 21.500 32.198 39.085 85.516
***
Developed 13.568 15.889 25.612 22.730 52.776 55.357 61.67
***
Developing 15.172 8.877 20.190 19.691 27.612 32.936 7.58
**
CAP
Total 8.159 3.690 9.297 4.874 9.443 6.099 6.51
**
Developed 8.019 3.654 7.970 4.612 5.543 3.962 27.919
***
Developing 10.353 3.615 10.838 4.735 10.312 6.159 4.127
ROAA
Total 1.189 0.740 1.124 1.043 1.231 1.609 3.652
Developed 1.152 0.715 0.589 0.669 0.061 0.772 121.639
***
Developing 1.763 0.899 1.746 1.056 1.492 1.632 11.268
***
EXPENSES
Total 57.054 13.241 57.105 15.577 57.891 18.599 0.551
Developed 57.625 12.921 62.432 14.597 57.708 17.537 11.503
***
Developing 48.131 15.148 50.918 14.406 57.932 18.876 12.365
***
LIQ
Total 19.588 19.473 28.024 21.329 30.632 19.763 82.258
***
Developed 18.666 19.198 25.675 23.702 15.611 17.039 11.591
***
Developing 34.014 18.322 30.752 17.923 33.979 18.785 3.098
OWNERS
Total 5.711 7.255 5.171 6.074 5.537 5.690 9.226
**
Developed 5.768 7.368 5.736 7.230 5.205 6.127 2.742
Developing 4.821 5.193 4.515 4.310 5.611 5.604 2.021
SUBS
Total 91.983 195.844 32.808 85.374 17.551 68.655 114.935
***
Developed 96.572 200.980 49.574 112.187 64.359 152.735 6.737
**
Developing 20.214 33.317 13.333 22.096 7.120 8.649 14.017
***
(B)
LIST 1 2 3
Total 466 214 214
Developed 438 115 39
Developing 28 99 175
Notes:
***
Significant at the 1% level;
**
significant at the 5% level;
*
significant at the 10% level; variables are defined in Table III.
C. GAGANIS ET AL. 108
Copyright # 2007 John Wiley & Sons, Ltd. J. Multi-Crit. Decis. Anal. 14: 103–111 (2006)
DOI: 10.1002/mcda
Wallis test to assess the means’ differences across
the groups.
SIZE, PROVIS and SUBS are the only variables
that are statistically significant in all cases.
However, while higher size results in higher
soundness in the cases of the total sample and
the developing markets sub-sample, the results are
mixed in the case of the developed markets sub-
sample, with an average SIZE equal to 4.398
(Group 1), 4.010 (Group 2) and 4.462 (Group 3).
Lower PROVIS results in higher soundness which
can be attributed to the perception that lower
provisions correspond to lower probability of
non-performing loans and higher asset quality.
As in the case of SIZE, the impact of SUBS on
soundness depends on the status of the market.
CAP and LIQ are significant in the case of the
total sample and in developed countries, although
not in developing countries. ROAA and EX-
PENSES on the other hand are significant in the
case of the two sub-samples, although not in
the case of the total sample. Higher ROAA results
in higher soundness, however, the results are
mixed with respect to EXPENSES. Finally, OWN-
ERS is significant only in the case of the total
sample.
Table V presents the average weights of the
criteria over the 10 replications. The two most
important criteria are PROVIS and CAP, with
weights equal to 20 and 19.76%, respectively.
Hence, as in previous studies (Poon et al., 1999;
Poon and Firth, 2005; Pasiouras et al., 2006) and
consistent with the univariate results, lower asset
quality (in terms of loan portfolio) results in lower
soundness. As for CAP, our finding is consistent
with the view that capital is important for banks
for several reasons. For example, capital serves as
the last line of defence against the risk of bank’s
insolvency, as any losses a bank suffers could be
potentially written off against capital. Even in the
case that insolvency becomes unavoidable, capital
protects to some degree depositors, creditors and
investors (Le Bras and Andrews, 2004). Further-
more, as mentioned by Theodore (1999) capital
allows the leveraging of a bank’s growth and
diversification, and a tight solvency position would
be an obstacle to do so.
The average weight of MARKET equals
15.66% showing that the country where the
banks operate has an important impact on their
classification. In other words, we find that operat-
ing in a developed market results in higher
soundness. EXPENSES and ROAA also carry
an average weight above 10%, indicating that
higher efficiency in expenses management and
higher profitability result in higher bank sound-
ness. The least important variables are SIZE and
LIST. Especially, the latter carries an average
weight equal to 0%, indicating that it makes no
difference whether the bank is listed in a stock
exchange or not.
Table VI presents the average classification
accuracies over the 10 replications. Panel A
corresponds to the training sample and Panel B
to the validation sample. For benchmarking
purposes, two additional models are developed
through DA and ordinary logistic regression
(OLR) using the same input variables and 10-fold
cross-validation process for estimating and testing
the models.
UTADIS obtains the highest overall classifica-
tion accuracy in the training sample that equals
Table V. Weights of criteria (variables) in the UTADIS
model (averages of 10 replications)
Criteria (variables) Weights (%)
PROVIS 20.00
CAP 19.76
MARKET 15.66
EXPENSES 10.31
ROAA 10.21
LIQ 9.08
OWNERS 8.79
SUBS 4.58
SIZE 1.60
LIST 0.00
Note: Variables are defined in Table III.
Table VI. Classification results (averages over 10
replications): (A) Training and (B) Validation
Group
1 (%) 2 (%) 3 (%) Overall (%)
(A)
UTADIS 84.10 53.89 74.54 70.84
OLR 92.63 24.02 73.37 63.34
DA 90.17 27.1 79.78 65.68
(B)
UTADIS 83.47 50.49 72.75 68.91
OLR 92.63 23.33 72.68 62.88
DA 90.48 26.22 78.47 65.06
Notes: UTADIS, UTilite´ s Additives DIScriminantes; OLR,
ordinary logistic regression; DA, discriminant analysis.
A MULTICRITERIA DECISION FRAMEWORK 109
Copyright # 2007 John Wiley & Sons, Ltd. J. Multi-Crit. Decis. Anal. 14: 103–111 (2006)
DOI: 10.1002/mcda
70.84%, while the corresponding figures for DA
and OLR are 65.58 and 63.34%, respectively. All
the three models classify correctly a high propor-
tion of banks from Group 1 that ranges between
84.19% (UTADIS) and 92.63% (OLR), followed
by banks in Group 3, but only a relatively smaller
proportion of banks in Group 2 that is between
24.02% (OLR) and 53.89% (UTADIS). The poor
performance in terms of classifying banks into
intermediate groups has been observed in past
studies as well (e.g. Pasiouras et al., 2007) and can
be attributed to the fact that banks falling in this
category might be closely related either to banks in
the lower band of Group 1 or the upper band of
Group 3, hence making their correct classification
a difficult task. The results of a t-test, used to
assess the differences in the classification accura-
cies among the methods, indicate that the accura-
cies achieved by UTADIS are significantly
different from the ones obtained by DA and
OLR at the 1% level.
Turning to the accuracies in the validation data
set, the highest overall accuracy (68.91%) is again
achieved by UTADIS, with group accuracies equal
to 83.47% (Group 1), 50.49% (Group 2) and
72.75% (Group 3). OLR and DA classify correct
62.88 and 65.06% of the banks in sample (i.e.
overall). As in the case of the training sample, all
three models classify correctly a higher proportion
of banks from Groups 1 and 3, and a relatively
lower proportion of banks from Group 2. The
difference between UTADIS and DA is now
statistically significant at the 5% level, whereas
the one between UTADIS and OLR remains
significant at the 1% level.
5. CONCLUSIONS
In this paper, we developed a multicriteria model,
through UTADIS, to classify banks into three
groups on the basis of their soundness. The sample
consisted of 894 banks from 79 countries. The
model was developed through a 10-fold cross-
validation procedure using six financial and four
non-financial variables, while the ratings of Fitch
formed the basis for assessing the soundness of
banks.
The results indicate that asset quality (as
measured by loan loss provisions), capitalization
and the market where banks operate are the most
important criteria (in terms of weights) in classify-
ing the banks. Profitability and efficiency in
expenses management are also important attri-
butes, whereas size and listing in a stock exchange
are the least important ones.
UTADIS classified correctly 70.84 and 68.91%
of the banks in training and validation samples
accordingly, and was more efficient than models
developed through DA and OLR for benchmark-
ing purposes. Furthermore, the differences in the
classification accuracies achieved by UTADIS and
the ones obtained by DA and OLS were statisti-
cally significant in both the training and the
validation samples. Finally, all the techniques
experienced difficulties in classifying banks in
intermediate situation.
The present study could be extended in various
ways, such as the classification of banks into more
groups, and the inclusion of additional bank-
specific variables into the model. It would also be
worthwhile to incorporate further country-specific
variables into the analysis, reflecting the regulatory
and economic environment in the markets where
banks operate. Finally, one could benchmark the
developed model against the alternative classifica-
tion techniques or develop integrated models.
ACKNOWLEDGEMENTS
We would like to thank two reviewers, and
participants at the 18th International Conference
on Multiple Criteria Decision Making (2006) for
helpful comments and suggestions that helped us
improve earlier versions of the paper circulated
under the title ‘An assessment of banks’ credit-
worthiness: a multicriteria approach’.
REFERENCES
Aldrich JH, Nelson FD. 1984. Linear Probability, Logit,
and Probit Models. Sage Publication: CA.
Barniv R, McDonald JB. 1999. Review of categorical
models for classification issues in accounting and
finance. Review of Quantitative Finance and Account-
ing 13: 39–62.
Belkaoui A. 1983. Industrial Bonds and the Rating
Process. Greenwood Press: Westport.
Canbas S, Cabuk A, Kilic SB. 2005. Prediction of
commercial bank failure via multivariate statistical
analysis of financial structures: the Turkish case.
European Journal of Operational Research 166:
528–546.
Claessens S, Demirgu¨ c¸ -Kunt A, Huizinga H. 2001. How
does foreign entry affect domestic banking markets?
Journal of Banking and Finance 25(5): 891–911.
C. GAGANIS ET AL. 110
Copyright # 2007 John Wiley & Sons, Ltd. J. Multi-Crit. Decis. Anal. 14: 103–111 (2006)
DOI: 10.1002/mcda
Cole RA, Gunther JW. 1998. Predicting bank failures:
a comparison of on- and off-site monitoring
systems. Journal of Financial Services Research 13(2):
103–117.
Demirguc-Kunt A, Detragiache E, Tressel T. 2006.
Banking on the principles: compliance with Basel core
principles and bank soundness. World Bank Policy
Research Working Paper 3954, June.
Doumpos M, Zopounidis C. 2002. Multicriteria Deci-
sion Aid Classification Methods. Kluwer Academic
Publishers: Dordrecht.
Ederington L, Yawitz JB, Roberts BE. 1987. The
information content of bond rating. The Journal of
Financial Research 10(3): 211–226.
Frecka TJ, Hopwood WS. 1983. The effects of outliers
on the cross-sectional distributional properties of
financial ratios. The Accounting Review 58: 115–128.
Hammer PL, Kogan A, Lejeune MA. 2006. Modeling
country risk ratings using partial orders. European
Journal of Operational Research 175: 836–859.
Hopwood WS, McKeown JC, Mutchler JF. 1988. The
sensitivity of financial distress prediction models to
departures from normality. Contemporary Accounting
Research 5: 284–298.
Horrigan JO. 1966. The determination of long-term
credit standing with financial ratios. Journal of
Accounting Research 4: 44–62.
Huang Z, Chen H, Hsu Ch-J, Chen W-H, Wu S. 2004.
Credit rating analysis with support vector machines
and neural networks: a market comparative study.
Decision Support Systems 37: 543–558.
Kaplan RS, Urwitz G. 1979. Statistical models of bond
ratings: a methodological inquiry. Journal of Business
52: 231–261.
Kim JW. 1993. Expert systems for bond rating: a
comparative analysis of statistical, rule-based
and neural networks systems. Expert Systems 10:
167–171.
Kolari J, Glennon D, Shin H, Caputo M. 2002.
Predicting large US commercial bank failures. Journal
of Economics and Business 54: 361–387.
Kosmidou K, Zopounidis C. 2007. Predicting US
commercial bank failures via a multicriteria approach.
International Journal of Risk Assessment and Manage-
ment, in press.
Lanine G, Vander Vennet R. 2006. Failure prediction in
the Russian bank sector with logit and trait recogni-
tion models. Expert Systems with Applications 30:
463–478.
Le Bras A, Andrews D. 2004. Bank Rating Methodology:
Criteria Report, 25 May, FitchRatings. Available at
www.fitchratings.com
Mar-Molinero C, Serrano-Cinca C. 2001. Bank failure:
a multidimensional scaling approach. The European
Journal of Finance 7: 165–183.
Noreen E. 1988. An empirical comparison of probit and
OLS regression hypothesis tests. Journal of Account-
ing Research 26: 119–133.
Pantalone C, Platt MB. 1987. Predicting commercial
bank failure since deregulation. New England Eco-
nomic Review 13: 37–47.
Pasiouras F, Gaganis Ch, Doumpos M. 2007. A multi-
criteria discrimination approach for the credit rating of
Asian banks. Annals of Finance 3: 351–367.
Pasiouras F, Gaganis Ch, Zopounidis C. 2006. The
impact of bank regulations, supervision, market
structure and bank characteristics on individual bank
ratings: a cross-country analysis. Review of Quantita-
tive Finance and Accounting 27(4): 403–438.
Pinches GE, Mingo KA. 1973. A multivariate analysis of
industrial bond ratings. Journal of Finance 28: 1–18.
Poon WPH, Firth M. 2005. Are unsolicited credit
ratings lower? International evidence from bank
ratings. Journal of Business Finance and Accounting
32(9 & 10): 1741–1771.
Poon WPH, Firth M, Fung H-G. 1999. A multivariate
analysis of the determinants of Moody’s bank
financial strength ratings. Journal of International
Financial Markets, Institutions and Money 9: 267–283.
Ravi Kumar P, Ravi V. 2007. Bankruptcy prediction in
banks and firms via statistical and intelligent techni-
ques}a review. European Journal of Operational
Research 180: 1–28.
Richardson FM, Davidson LF. 1984. On linear dis-
crimination with accounting ratios. Journal of Busi-
ness Finance and Accounting 11: 511–525.
Rose PS, Kolari JW. 1985. Early warning systems as a
monitoring device for bank condition. Quarterly
Journal of Business and Economics 24(Winter): 43–60.
Sinkey J. 1975. A multivariate statistical analysis of the
characteristics of problem banks. Journal of Finance
30: 21–36.
Stone M, Rasp J. 1991. Tradeoff in the choice between
logit and OLS for accounting choice studies. The
Accounting Review 66: 170–187.
Tam KY, Kiang MY. 1992. Managerial applications of
neural networks: the case of bank failure predictions.
Management Science 38(7): 926–947.
Theodore SS. 1999. Bank Credit Risk Rating Methodol-
ogy (An Analytical Framework for Banks in Developed
Markets), Moody’s Investors Service, Global Credit
Research, April.
Thomson JB. 1991. Predicting bank failure in the 1980s.
Economic Review, Federal Reserve Bank of Cleveland,
Volume Q1: 9–20.
Tung WL, Quek C, Cheng P. 2004. GenSo-EWS: a novel
neural-fuzzy based early warning system for predicting
bank failures. Neural Networks 17: 567–587.
Yim J, Mitchell H. 2005. Comparison of country risk
models: hybrid neural networks, logit models, dis-
criminant analysis and cluster techniques. Expert
Systems with Applications 28(1): 137–148.
Zopounidis C, Doumpos M. 1999. A multicriteria
decision aid methodology for sorting decision pro-
blems: the case of financial distress. Computational
Economics 14: 197–218.
A MULTICRITERIA DECISION FRAMEWORK 111
Copyright # 2007 John Wiley & Sons, Ltd. J. Multi-Crit. Decis. Anal. 14: 103–111 (2006)
DOI: 10.1002/mcda