International Research Journal of Finance and Economics

ISSN 1450-2887 Issue 12 (2007)

© EuroJournals Publishing, Inc. 2007
http://www.eurojournals.com/finance.htm

Bank Failure Prediction Using Modified Minimum

Deviation Model

Ali Argun Karacabey

Ankara University, Faculty of Political Sciences, 06590 Cebeci Ankara, Turkey
E-mail: karacabe@politics.ankara.edu.tr

Abstract

In discriminant analysis which is one of the widely used failure prediction

technique, a group of observations whose memberships are already identified, are used for
the measurement of weight estimates of a function by minimizing their group
misclassifications. Since 1980’s a group of researchers studied on developing non
parametric discriminant methods. Non parametric models give to the analyst the
opportunity to add new conditions (constraints) to the model. In this paper a new non
parametric discriminant model is proposed. This new model which is a modified version of
well known minimum deviation model, helps the analyst by choosing the optimal variables
to predict the discriminant function. Using the data set of Turkish commercial banks for the
period of 1994-2001, the modified model is tested, the validity and the prediction
performance are compared with a two stage prediction process that employs both factor
and discriminant anlysis.

I. Introduction
Bank failure prediction has always been an interesting topic for several reasons. First of all,
performance assessment, in other words distinguishing between good and bad, is itself an important
area of study. Besides, since banks are dominant actors in financial systems of all countries, bank
failures have more significant effects than failures of non-financial enterprises do. Consequently, the
demand to reduce the costs of bank failures by taking measures in advance makes this issue interesting.
Being able to predict failures also connotes displaying the differences between the successful and
unsuccessful. Determining the features distinguishing the successful and unsuccessful also takes place
within this prediction activity.
Studies on failure prediction continue since 1960s. “Discriminant analysis” is the leading
technique among the interesting techniques used in these studies in which usually parametrical
methods are employed. Applications of discriminant analysis, which tries to predict the weights of
observations, group memberships of which have already been determined, in the discriminant function
in such a way that minimizes their misclassifications, (Sueyoshi, 2001) are not only seen in prediction
of financial failures of enterprises but also in engineering sciences, in the field of medical science and
in the fields of marketing, management and finance (Asparoukhov and Stam, 1997). The most
commonly known work on use of discriminant analysis in financial failure prediction is the study by
Altman and Nelson (1968).
Since 1980s studies have been conducted on mathematical programming based discriminant
analysis. Theoretical studies on mathematical programming based discriminant analysis which were
first conducted by Freed and Glover(1981a, 1981b) in the early 80s were carried on by Glover (1990),
Ragsdale and Stam (1991) and Gupta and others (1990). These studies mostly focused on applicability
International Research Journal of Finance and Economics - Issue 12 (2007) 148

of mathematical programming techniques on discriminant analysis and their formulation. In later years,
Wallin and Sundgren (1995), Sueyoshi (2001) conducted studies on evaluation of results of application
of these methods. Besides, studies aiming at developing new models which are compatible with new
mathematical programming based discriminant analysis were also conducted. While Sueyoshi (1997,
2003) tried to combine discriminant analysis and data envelopment analysis using goal programming
and mixed integer programming, Loucopoulos and Pavur(1997) suggested a model which enables
dividing into three-group classification. Lam, Choo and Moy (1996) developed a two-stage model
which makes classification after minimizing deviations from group averages. Although there are
several models in different fields in the literature, which were developed with different models, there is
a consensus on “minimum sum of deviations model” (MSD) as the model which gives the most proper
results in a significant portion of studies (Lam, Choo and Moy, 1996; Karacabey, 2003).
Determination of characteristics of banks, which are used in both parametric and non-
parametric methods for failure prediction, is a separate problem. The questions asking “which financial
ratios should be used” or “how these financial ratios should be selected”, which are not as interesting
as failure prediction but are very important since they have direct effects on these studies, are generally
ignored in studies. While Barnes (1987) asserts that ratios used in analysis are generally selected
according to their popularity; in some other studies, it is seen that the financial ratios used in similar
studies are taken as samples. Laitinen (1991) emphasizes that “a priori” selection of financial ratios
used in analysis makes it difficult to support the obtained model theoretically, whereas Zavgren (1983)
asserts that use of multiple ratios which are close to each other will result in generating results which
are peculiar to the sample. Horrigan (1965) stated that financial ratios are interdependent and therefore
the information which is expected to be obtained from financial ratios can be obtained from a more
limited number of ratios.
Factor analysis is used in most of the studies which are conducted to prevent use of another
financial ratio, which repeats the information provided by a financial ratio, in the model. There are
several studies which generate factors that cover the information in the financial ratios, using factor
analysis (Pinches, Mingo and Carruthers, 1973; Johnson, 1978; Laurent, 1979; Mear and Firth, 1986;
Aktaş and others, 2001; Canbaş and others, 2004) which can be described as a technique which
simplifies and reduces complex and numerous data available (Kline, 1994, p.12). Although factor
analysis can cope with this problem, this technique causes loss of information.
Therefore, financial failure prediction consists of at least two stages. The first stage of the
process involves determination of ratios or factors which will be used in the analysis and the second
stage involves financial failure prediction. In this study, two new models which will serve to reduce
this process which has two or more stages to single stage are suggested. The suggested models are
developed based on classical minimum sum of deviations model. Therefore, it is possible to name
these models as “modified minimum sum of deviations” model. While these models are developed, the
aim is to enable selection of desired number of ratios among numerous ratios in such a way that will
make the best discrimination.
Modified minimum sum of deviation models will be discussed in the second section of the
article. In the following sections, these new models will be applied using the data from private Turkish
Commercial Banks, the results will be compared with the results of classical minimum deviation model
in which factors, which are formed by factor analysis, are used and validity of the models will be
discussed. The study will conclude with a general evaluation section.

2. Weighted Minimum Sum of Deviation Model

Let us assume that N firms (i=1,2…,n) are evaluated using m (j=1,2…m) independent variables (Xj)
and a binary classification is made. It is known that among these n firms, n1 firms belong to the first
group (i G1) and n2 firms belong to the second group (i G2). To generate a discriminant function with
constant term K0 the MSD model is
Min ∑di (1.1)
149 International Research Journal of Finance and Economics - Issue 12 (2007)

st.
x1ai1 + x2ai2 +…+ xjaij + di ≥ K0 (i G1) (1.2)
x1ai1 + x2ai2 +…+ xjaij – di ≤ K0 - ε (i G2) (1.3)
∑xj =1 (1.4)
xj, di ≥ 0 (1.5)
*
Solution of this model will give the optimum break point ( K 0 ), the value of deviation variable
for each unit ( d i* ) and the optimum values of weights of independent variables ( x *j ).
Some problems may occur in the application of this model which is perfect in terms of
mathematical modeling and which can generate proper results. While this model restricts the maximum
values of the weights by equating the sum of weights to one, this constraint will be satisfied also when
the weight of only one variable is 1. Thus, in a case when you believe that ten characteristics are
important in terms of classification, the model may take only one characteristic into consideration. It is
possible to get rid of this problem by voluntarily defining a lower limit which weights can take, like in
Gem-flex which is a mathematical programming based artificial criterion used for macroeconomic
performance measurement (Cherchye,2001). In this case, if θ is defined as lower limit; it will be
enough to add
xj ≥ θ (2)
constraint to the model.
In case numerous criterions are used, this constraint can be a compulsion which may cause
deadlock. In such cases, in order to make the model more flexible, obliging the use of some of the
determined characteristics may be a more meaningful solution. For example, in a case in which n
characteristics are determined, it can be asked to take at least s of these n characteristics into
consideration with a weight greater than the determined lower limit (θ). In this case, in order to extend
the scope of the model, instead of adding the constraint (2) it will be necessary to add the following
constraints:
xj + Myj ≥ θ (3.1)
∑yj < n-s (3.2)
yj [0,1] (3.3)
Although these constraints guarantee that the selected characteristics will take a specific weight,
they do not help which characteristics to select. The method used in bank failure prediction is
important as well as the variables used. It is possible to follow two different methods in selecting the
ratios used in the study. One of these methods is to follow the similar studies conducted formerly and
conduct the study on the basis of the ratios used in these studies. This method may cause problems
especially when applied in countries like Turkey which have a very dynamic and changing structure.
The second method is to determine the factors which will take place in the analysis by using factor
analysis before failure prediction. Although it is a commonly used method and can overcome this
problem, factor analysis means that some of the information on the data will be lost.
The method which is suggested in this study is determination of ratios to be selected within the
mathematical model defined above. Some arrangements are necessary for the model to realize this
function. Firstly, the ratios to be used should be grouped. For example, in bank failure prediction, the
proper classification of the ratios should be as follows: capital adequacy ratios, profit capital ratios,
liquidity ratios, ratios related to income and expenditure structure and ratios related to the quality of
assets. After ratios are divided into “m” groups, the constraints which will enable the model to select
the most proper ratio from each ratio group should be added. Constraints to be added will both enable
each group to be taken into consideration by determining the lower limits of weights which are
appointed to each group and determine the ratios which will make the most proper classification.
In order to add these characteristics to the model, constraints similar to the above constraint 3
should be redefined for each ratio group. Accordingly, the constraints which will be added to the
model, provided that minimum weight of each group in the model is θ for m ratio groups, will be as
follows:
International Research Journal of Finance and Economics - Issue 12 (2007) 150

xj + Myjk ≥θ k= 1....m (4.1)

∑yjk < k-1 (4.2)
yj [0,1] (4.3)
Thus, the model which is suggested to be used for bank failure prediction and determines the
ratios which will be used to classify banks are as follows:
Min ∑di (5.1)
st.
x1ai1 + x2ai2 +…+ xjaij + di ≥ K0 (i G1) (5.2)
x1ai1 + x2ai2 +…+ xjaij – di ≤ K0 - ε (i G2) (5.3)
xj + Myjk ≥θ (k= 1....m) (5.4)
∑yjk < k-1 (5.5)
∑xj =1 (5.6)
yj [0,1] (5.7)
xj, di ≥ 0 (5.8)
Another alternative of mixed integer programming model is to make the total weight of each
group greater than a determined lower limit. In this case, instead of constraints 5.4 and 5.5, the
following constraint will be added to the model above for each group.
∑xjk ≥ θ (k= 1....m) (6)
Both mixed integer programming model and linear programming model which determines
group weight will guarantee usage of at least one characteristic from each group. Certainly, adding new
compulsive constraints to the model, compared to minimum deviation model, may have negative
effects on classification power of the model. However, financial interpretation results is also important
as well as the numerical results of financial failure prediction.

3. Application on Private Turkish Commercial Banks

The period between 1999 and 2001 has been an important transition process for Turkish banking. In
this period, 18 banks were transferred to Saving Deposit Insurance Fund for various reasons. Most of
these banks failed to discharge their liabilities with their assets and some of them were mismanaged.
Regardless of their reasons, the extent to which these banks which were transferred to the fund were
harmful for the system and the economy in general sense is very important.
In this study, analysis of the classes which consist of banks transferred to the fund and banks
continuing their operations will be realized applying discriminant analysis with the mathematical
programming techniques, which were explained in the previous section in order to determine the
variables which makes the banks in these classes different from each other, determine the weights of
these variables and determine whether financial failures can be predicted in advance. Besides the
capability of amended and weighted minimum deviation model to distinguish between successful and
unsuccessful, its capability to select the variables which will be used in this analysis is another aim of
this study. The method used in the study for this purpose can be briefly summarized as follows: Firstly,
capabilities of applicable ratios to distinguish between groups have been investigated statistically and
the ratios which do not have this capability have been eliminated. Next, minimum deviation model
weighted with remaining ratios and mixed integer programming models were applied.
As explained above, these two new models are improved models in order to perform the duty of
factor analysis. Therefore, validity of the new models has been investigated by comparing the factors
obtained by applying factor analysis to the variables with the results of application of classical MSD
method and three methods in order to analyze whether the models perform this duty.

3.1. Data
In this study, data related to 39 commercial banks have been used. While 21 out of these 39 banks
which were included in the analysis as of the end of 2001 continue their operation, the remaining 18
151 International Research Journal of Finance and Economics - Issue 12 (2007)

banks are transferred to the fund. In the analysis, it has been investigated whether the mentioned
mathematical programming based models can predict whether the banks transferred to the fund will
experience financial failure using the data of 3 years before they were transferred to the fund.
Accordingly, the end of 2001 was taken as the base year for the banks continuing their operations and
the data of 2000, 1999 and 1998 were used for these banks. Base years for failed banks were accepted
as the year when they were transferred to the fund. For example, if a bank was transferred to the fund
in 1999, the data group for this bank included the data of 1998, 1997 and 1996. In appendix A, the
years in which the banks in the unsuccessful class were transferred to the fund and the data used for
them in the analysis are shown.
In classification of the banks transferred to the fund and the banks continuing their operation,
the financial ratios published by the Banks Association of Turkey (BAT) for 1994-2000 were used.
These ratios are classified in five groups, which are capital ratios, ratios related to the quality of assets,
liquidity ratios, profit capital ratios and ratios related to income and expenditure structure, in
accordance with the classification of BAT. Ratios used in the study are given in appendix B.
As it was mentioned before, firstly, it was investigated with t-test whether 27 ratios given in the
table are statistically different from each other among the groups. Each ratio was divided into two
groups for successful and unsuccessful banks and it was investigated whether the group averages were
statistically different from each other. Averages, standard deviations and t values for ratio groups are
given in table 1.
International Research Journal of Finance and Economics - Issue 12 (2007) 152
Table 1: Classification Power of the Ratios

Ratios Group Mean St. Deviation t-value Prob.

Succesful 17,205 11,276
S1 2.298 .035
Failed -1,557 31,558
Succesful 25,233 22,842
S2 2.763 .013
Failed 3,496 18,788
Succesful 9,916 9,755
S3 2.264 .037
Failed -12,050 37,857
Succesful 7,353 7,180
S4 2.242 .039
Failed 0,310 10,529
Succesful 215,817 187,900
S5 -0.366 .719
Failed 244,040 246,709
Succesful 30,6717 15,022
A1 -1.025 .320
Failed 35,235 10,056
Succesful 3,869 5,003
A2 -1.620 .124
Failed 44,5828 105,436
Succesful 17,183 14,617
A3 0.384 .706
Failed 16,238 14,386
Succesful 66,199 23,612
A4 0.234 .818
Failed 64,836 21,889
Succesful 46,023 22,175
L1 2.828 .012
Failed 30,609 13,844
Succesful 59,985 31,445
L2 3.087 .007
Failed 35,252 17,473
Succesful 40,458 25,247
L3 1.075 .297
Failed 33,315 16,643
Succesful 2,410 2,808
K1 1.819 .087
Failed -10,915 30,445
Succesful 32,534 35,007
K2 0.927 .367
Failed -12,049 191,302
Succesful 49,134 47,739
K3 1.540 .142
Failed -212,227 718,420
Succesful 3,314 3,569
K4 1.837 .084
Failed -10,339 30,789
Succesful 1,691 2,656
K5 -1.417 .175
Failed 21,872 60,038
Succesful 0,447 0,699
K6 -1.394 .181
Failed 5,633 15,694
Succesful 12,286 6,545
G1 1.385 .184
Failed 4,546 21,540
Succesful 213,830 70,334
G2 3.521 .003
Failed 144,041 43,355
Succesful 1,385 66,344
G3 1.649 .117
Failed -40,442 89,634
Succesful 118,860 17,133
G4 2.750 .014
Failed 94,255 31,552
Succesful 36,596 9,815
G5 -3.052 .007
Failed 62,188 37,840
Succesful 16,761 6,027
G6 -3.329 .004
Failed 30,420 14,292
Succesful 18,227 5,443
G7 -3.556 .002
Failed 46,272 31,639
Succesful 101,806 21,463
G8 0.936 .362
Failed 24,476 346,149
Succesful 59,245 12,286
G9 -2.424 .027
Failed 70,535 10,678
153 International Research Journal of Finance and Economics - Issue 12 (2007)

As it is seen, 12 out of 27 selected financial ratios are statistically different for successful and
unsuccessful banks. In other words, it can be said that 15 ratios do not have the distinguishing feature
between these two groups. Therefore, these 15 ratios were eliminated. While ratios related to the
quality of assets and profit capital ratios do not take place among 12 ratios which have distinguishing
feature between successful and unsuccessful banks, four of capital ratios, two of liquidity ratios and six
of income expenditure ratios took place. The correlations between these selected 12 ratios are given in
table 3.

Table 2: Correlations Between the Selected Ratios

s1 s2 s3 s4 l1 l2 g2 g4 g5 g6 g7 G9
s1 1,000
s2 0,866 1,000
s3 0,948 0,809 1,000
s4 0,915 0,903 0,903 1,000
l1 0,243 0,173 0,355 0,307 1,000
l2 0,374 0,354 0,448 0,461 0,962 1,000
g2 0,403 0,446 0,411 0,249 0,004 0,032 1,000
g4 0,693 0,514 0,694 0,552 0,298 0,344 0,515 1,000
g5 -0,252 -0,301 -0,372 -0,417 -0,441 -0,435 -0,116 -0,173 1,000
g6 -0,202 -0,280 -0,237 -0,227 -0,246 -0,239 -0,506 -0,270 0,814 1,000
g7 -0,472 -0,490 -0,577 -0,530 -0,425 -0,432 -0,479 -0,432 0,900 0,893 1,000
g9 -0,022 -0,234 0,045 -0,032 0,069 0,018 -0,541 0,116 0,386 0,666 0,455 1,000

Before they were subjected to factor analysis, these 12 ratios which were determined to have
distinguishing feature were normalized in such a way that the most desirable value was “1” and the
least desirable value was “0”. In this normalization stage, the first 9 ratios were converted into the
following equation:
X ij − min j
nij =
max j − min j
and the last three ratios were converted into the following equation:
X ij − max j
nij =
min j − max j
The factors to be used in this stage of analysis in the classical MSD model were generated
using principle component analysis and varimax rotation. After the factor analysis which was made
using principle components method, 3 factors were found out whose eigenvalue is greater than one.
These three factors explain approximately 83% of the total change in 12 financial ratios (table 3).

Table 3: Eigenvalues of the Factors

Factors Eigenvalues Component Cumulative

1 5,838 48,650 48,650
2 2,367 19,723 68,373
3 1,831 15,258 83,631
4 ,872 7,263 90,894
5 ,665 5,539 96,433
6 ,190 1,583 98,016
7 ,117 ,975 98,991
8 6,261E-02 ,522 99,512
9 3,004E-02 ,250 99,763
10 1,255E-02 ,105 99,867
11 1,089E-02 9,071E-02 99,958
12 5,024E-03 4,187E-02 100,000
International Research Journal of Finance and Economics - Issue 12 (2007) 154

In order to make a more distinct classification, factor rotation was made in accordance with the
VARIMAX method; therefore it became possible to determine the loading ratios which will be used to
determine which factor represents which financial rate and to define factors. This method reduces the
numbers of variables with high loading ratios. The results are given in table 4. In the table, variables
which have high loading ratios for each factor are grouped and variables which have low loading ratios
(<0.5) are ignored.

Table 4: Factor Loadings

Factor1 Factor2 Factor3

S1 .965
S3 .930
S4 .891
S2 .879
G4 .753
G6 .937
G9 .851
G7 .783
G2 .597
G5 -.704
L1 .932
L2 .890

As it is seen clearly in the table, although the first factor basically represents capital ratios, ratio
of total incomes to total expenditures are also included in this factor. The other two factors have
relatively more homogenous structures. While factor 2 consists of only income expenditure ratios,
factor 3 includes liquidity ratios. The analysis which are indicated for one single year (year -1) in this
section formed the principle data for models to be applied and the applications which are conducted for
the years -2 and -3 in order to evaluate the prediction successes of models were also based on these
ratios and factors.
The rest of the analysis is allocated for the operation of the models explained above using the
data prepared up to this stage.

3.2. Application of Models and Results

As it was mentioned in the previous section, three different models were applied to the database and
results were achieved. The analysis process which was followed is summarized in figure 1.
155 International Research Journal of Finance and Economics - Issue 12 (2007)
Figure 1:

Formation of the Database

(Compilation of 27 Financial Ratios)

Determination of Ratios with Distinguishing Feature

(Determination of 12 Ratios)

Factor Analysis Application of Application of Mixed

(Formation of 3 Weighted Minimum Integer Minimum Sum of
Factors) Sum of Deviations Deviations Model
Model

Application of
Classical Minimum
Sum of Deviations
Model

In case the results achieved from the models suggested in this study are as successful as or
more successful than the results achieved from the minimum deviation sums model which was applied
using the factors of prediction success of these models, validity of these models will be evident.
Prediction successes of the models according to the results related to the past 1, 2 and 3 years are
summarized in table 5.

Table 5: Classification Power of the Models

Real Group
-1.Yıl Factor Analysis MSD Weighted MSD Mixed Integer MSD
Un-failed Failed Un-failed Failed Un-failed Failed
Model Un-failed 15 0 20 3 20 2
Classification Failed 6 18 1 15 1 16
Classification Power %71,4 %100 %95,2 %83,3 %95,2 %88,9
Total Classification Power %84,6 %89,7 %92,3
Real Group
-2. Yıl Factor Analysis MSD Weighted MSD Mixed Integer MSD
Un-failed Failed Un-failed Failed Un-failed Failed
Model Un-failed 18 3 19 2 19 3
Classification Failed 3 15 2 16 2 15
Classification Power %85,7 %83,3 %90,5 %88,9 %90,5 %83,3
Total Classification Power %84,6 %89,7 %87,2
Real Group
-3. Yıl Factor Analysis MSD Weighted MSD Mixed Integer MSD
Un-failed Failed Un-failed Failed Un-failed Failed
Model Başarılı 17 4 16 4 16 4
Sınıflandırması Başarısız 4 14 5 14 5 14
Classification Power %81,0 %77,8 %76,2 %77,8 %76,2 %77,8
Total Classification Power %79,5 %76,9 %76,9

As it is seen clearly in table 6, prediction successes of the models suggested in this study are
higher than the classical model supported with factor analysis. In all three models, it is seen that
prediction success falls when data gets older. It is possible to say that the most successful model is the
International Research Journal of Finance and Economics - Issue 12 (2007) 156

mixed integer MSD model. Mixed integer minimum deviation model is the most successful model for
that year with a prediction success of 92, 3%. It is seen that totally three banks were zeroed incorrectly
in the first year. While the weighted model zeroed four banks incorrectly in the same year, the model
supported with factor analysis zeroed six banks incorrectly.
It can be said that prediction success of the model suggested in this study is quite high in cases
when data delayed for two or three years are used. Among the models which tried to predict failure two
years in advance, the weighted model, mixed integer model and the model with factor analysis zeroed
four, five and six banks incorrectly respectively. Percentages of success of the models are 89.7%,
87.2% and 84.6% respectively. In failure prediction three years in advance, the model with factor
analysis zeroed 8 banks incorrectly while the other two models zeroed 9 banks incorrectly. Prediction
success of the model with factor analysis was 79.5% whereas prediction success of these two models
was 76.9%.
When a general evaluation of the results generated by the models is made, it can be said that the
models suggested in this study and which determine the variables within themselves in the framework
of desired conditions achieve better results compared to the three-stage failure prediction process.

4. General Evaluation and Conclusion

Use of mathematical programming based models in financial failure prediction is increasing by the
day. It can be said that studies on this issue are gathered under two main groups. While one group tries
to develop new mathematical models aiming at increasing prediction success, the other group is
interested in the results of application of these models. However, the variables to be taken into
consideration in usage of models have not been mentioned very much.
Flexibility is one of the most important characteristics of mathematical programming methods,
which is different from other methods with the same aim. We can enable models to take different cases
into consideration by adding new constraints. In this study, mathematical programming based classical
discriminant models have been enabled to determine the classical variable which is used themselves
with a couple of small corrections. The suggested model has been tested by making a comparison with
a three-stage financial failure process which includes statistical methods, factor analysis and
discriminant analysis. The results show that the results achieved from the suggested models are better
at financial failure prediction compared to the designed three-stage process.
157 International Research Journal of Finance and Economics - Issue 12 (2007)

References
[1] AKTAŞ, R./S.KARACAER/A.A.KARACABEY (2001), “Mali Oranlar Arasındaki Ilişkilerin
Faktör Analizi Ile Incelenmesi”, Muhasebe Bilim Dünyası Dergisi, Cilt 3, Sayı 1, 9-28
[2] ALTMAN, E.I./NELSON, A.D. (1968), “Financial Ratios, Discriminant Analysis and the
Prediction of Corporate Bankruptcy”, Journal of Finance, 23: 589-609
[3] ASPAROUKHOV, O.K/STAM, A. (1997), “Mathematical programming formulations for two-
group classification with binary variables”, Annals of Operations Research, 74: 89-112
[4] BARNES, P. (1987), “The Analysis and Use of Financial Ratios: A Review Article”, Journal of
Business Finance&Accounting, 449-461
[5] CANBAS, S./CABUK, A.,/KILIC, S.B., (2004), “Prediction of commercial bank failure via
multivariate statistical analysis of financial structures: The Turkish case”, European Journal of
Operational Resarch, (Article In Press)
[6] FREED, N./GLOVER, F. (1981a), “A linear programming approach to the discriminant
problem”, Decision Sciences, 12: 68-74
[7] FREED, N./GLOVER, F. (1981b), “Simple but powerful goal programming formulations for
the discriminant problem”, European Journal of Operational Resarch, 7: 44-60
[8] GLOVER, F. (1990), “Improved linear programming models for discriminant analysis”,
Decision Sciences, 21: 771-785
[9] GUPTA, Y./R. RAO, BAGCHI, P. (1990), “Linear goal programming as an alternative to
multivariate discriminant analysis”, Journal of Business Finance and Accounting, 17: 593-598
[10] HORRIGAN, J.O. (1965), “Some Emprical Bases of financial Ratio Analysis”, Accounting
Review, 558-568
[11] JOHNSON, W.B. (1978), “The Cross Sectional of Financial Patterns”, Journal of Business
Finance&Accounting, 207-214
[12] KARACABEY, A.A. (2003) “Matematiksel Programlama Ile Sınıflandırma”, Savunma
Bilimleri Dergisi, Vol.2 N0:1
[13] KLINE, P. (1994) An Easy Guide to Factor Analysis, Routledge
[14] KONNO, H./KOBAYASHI, H. (2000), “Failure discrimination and rating of enterprises by
semi-definite programming”, Asia-Pacific Financial Markets, 7: 261-273
[15] LOUCOPOULOS, C./PAVUR, R., (1997), “Experimental evaluation of the classificatory
performance of mathematical programming approaches to the three-group discriminant
problem: The case of small samples”, Annals of Operations Research, 191-209
[16] LAURENT, C.R. (1979), “Improving the Efficiency and Effectiveness of Financial Ratio
Analysis”, Journal of Business Finance&Accounting, 401-413
[17] MEAR, R./M.FIRTH (1986), “ A Note on the Financial Variable and Ratio Structure of New
Zealand Listed Companies”, Accounting and Finance, 47-56
[18] PINCHES, G.E./K.A.MINGO (1973), “A Multivariate Analysis of Industrial Bond Ratings”,
Journal of Finance, 1-18
[19] RAGSDALE, C./STAM,A (1991), “Mathematical programming formulations for the
discriminant problem: an old dog does new tricks”, Decision Sciences, 22: 296-306
[20] SUEYOSHI, T. (1999), DEA–Discriminant Analysis In the View of Goal Programming.
European Journal of Operational Research 115, 564–582.
[21] SUEYOSHI, T. (2001), “Extended DEA- Discriminant Analysis, European Journal of
Operational Research, 131, 324-352
[22] SUEYOSHI, T. (2004), “Mixed Integer Programming Approach of Extended DEA-
Discriminant Analysis, European Journal of Operational Research, 152, 45-55
[23] WALLIN, J./SUNDGREN, S. (1995), “Using Linear Programming to Predict Business Failure:
An Empirical Study”, NAN Research Papers, (http://accounting.rutgers.edu/NAN/Papers/ )
[24] ZAVGREN, C.V. (1983), “The Prediction of Corporate Failure: The State of the Art”, Journal
of Accounting Literature, 1-18
International Research Journal of Finance and Economics - Issue 12 (2007) 158

Appendix
Appendix A: Sample of Banks

BANK Code Date of Failure Year -1 Year -2 Year -3

Adabank A.Ş. D1 - 2000 1999 1998
Akbank T.A.Ş. D2 - 2000 1999 1998
Alternatif Bank A.Ş. D3 - 2000 1999 1998
Anadolubank A.Ş. D4 - 2000 1999 1998
Birleşik Türk Körfez Bankası A.Ş. D5 - 2000 1999 1998
Denizbank A.Ş. D6 - 2000 1999 1998
Fiba Bank A.Ş. D7 - 2000 1999 1998
Finans Bank A.Ş. D8 - 2000 1999 1998
Koçbank A.Ş. D9 - 2000 1999 1998
MNG Bank A.Ş. D10 - 2000 1999 1998
Oyak Bank A.Ş. D11 - 2000 1999 1998
Pamukbank T.A.Ş. D12 - 2000 1999 1998
Şekerbank T.A.Ş. D13 - 2000 1999 1998
Tekstil Bankası A.Ş. D14 - 2000 1999 1998
Turkish Bank A.Ş. D15 - 2000 1999 1998
Türk Dış Ticaret Bankası A.Ş. D16 - 2000 1999 1998
Türk Ekonomi Bankası A.Ş. D17 - 2000 1999 1998
Türkiye Garanti Bankası A.Ş. D18 - 2000 1999 1998
Türkiye İmar Bankası T.A.Ş. D19 - 2000 1999 1998
Türkiye İş Bankası A.Ş. D20 - 2000 1999 1998
Yapı ve Kredi Bankası A.Ş. D21 - 2000 1999 1998
Bank Ekspres A.Ş. D22 1998 1997 1996 1995
Bank Kapital Türk A.Ş. D23 2000 1999 1998 1997
Demirbank T.A.Ş. D24 2000 1999 1998 1997
Egebank A.Ş. D25 1999 1998 1997 1996
Eskişehir Bankası T.A.Ş. D26 1999 1998 1997 1996
Etibank A.Ş. D27 2000 1999 1998 1997
Interbank D28 1999 1998 1997 1996
Sümerbank A.Ş. D29 1999 1998 1997 1996
Türk Ticaret Bankası A.Ş. D30 1997 1996 1995 1994
Türkiye Tütüncüler Bankası Yaşarbank A.Ş. D31 1999 1998 1997 1996
Yurt Ticaret ve Kredi Bankası A.Ş. D32 1999 1998 1997 1996
Bayındırbank A.Ş. D33 2001 2000 1999 1998
Ege Giyim Sanayicileri Bankası A.Ş. D34 2001 2000 1999 1998
İktisat Bankası T.A.Ş. D35 2001 2000 1999 1998
Kentbank A.Ş. D36 2001 2000 1999 1998
Milli Aydın Bankası T.A.Ş. D37 2001 2000 1999 1998
Sitebank A.Ş. D38 2001 2000 1999 1998
Toprakbank A.Ş. D39 2001 2000 1999 1998
159 International Research Journal of Finance and Economics - Issue 12 (2007)
Appendix B:

(Shareholders’ Equity + T.Income)/ Total Asssets S1

(Shareholders’ Equity + T.Income)/(Deposits and Non Deposit Funds) S2
Capital
ratios
Net Working Capital/Total Assets S3
(Shareholders’ Equity + T.Income)/(Total Assets + Contingencies and Commitments S4
Fx Position / Shareholders’ Equity S5
Loans/Total Assets A1
Quality
Asset

Non Performing Loans / Loans A2

Permanent Assets / Total Assets A3
Fx Assets / Fx Liabilities A4
Liquid Assets / Total Assets L1
Liqu
idity

Liquid Assets / (Deposits and Non Deposit Funds) L2

Fx Liquid Assets / Fx Liabilities L3
Net Income (Loss) / Average total Assets K1
Profitability

Net Income (Loss) / Average Shareholders’ Equity K2

Ratios

Net Income (Loss) / Average Share K3

Income Before Tax / Average Total Assets K4
Provision for Loan Loses / Loans K5
Provision for Loan Loses / Total Assets K6
Net Income After Provisions / Average Total Assets G1
Income-Expenditure

Interest Income / Interest Expenses G2

Structure Ratios

Non-Interest Income / Non-Interest Expenses G3

Total Income /Total Expenses G4
Interest Income / Average Profitable Assets G5
Interest Expenses / Average Non-Profitable Assets G6
Interest Expenses / Average Profitable Assets G7
Interest Income / Total Income G8
Non-Interest Income/Total Income G9