EUROPEAN

JOURNAL
OF OPERATIONAL
RESEARCH
ELSEVIER European Journal of Operational Research 90 (1996) 487-513

Theory and Methodology

A survey of business failures with an emphasis on prediction

methods and industrial applications
A.I. Dimitras a, S . H . Z a n a k i s u,. , C. Z o p o u n i d i s a
a Decision Support Systems Laboratory, Technical University of Crete, Chania 73132, Greece
b Department of Decision Sciences and Information Systems, College of Business Administration, Florida International University,
Miami, FL 33 199, USA
Received January 1995

Abstract

The considerable interest in the prediction of business failures is reflected in the large number of studies
presented in the literature. Various methods have been used to construct prediction models. This paper provides a
review of the literature and a framework for the presentation of this information. Articles can be classified according
to the country, industrial sector and period of data, as well as the financial ratios and models or methods employed.
Relationships and research trends in the prediction of business failure are discussed.

Keywords: Business failures; Statistical analysis

1. Introduction A unifying theory of business failure has not

been developed, in spite of a few notable efforts.
In recent years a large n u m b e r of researchers These include Wilcox's (1971) ruin model equat-
and practitioners (credit managers, financial ana- ing failure to zero net worth and Scapens et al.
lysts, etc.) have worked on the prediction of busi- (1981) catastrophic theory approaches. In several
ness failures. Failure can be defined in many cases, alternative options to declared failure are
ways, depending on the specific interest or condi- considered. E x a m p l e s are liquidation or reorgani-
tion of the firms under examination. This ac- zation of bankrupt firms (Casey et al., 1986) and
counts for the many definitions of failure in the acquisition by or merger with a healthy company
literature. According to a general definition, fail- (Bulow and Shoven, 1978). These alternatives will
ure is the situation that a firm cannot pay lenders, not be considered in our survey.
preferred stock shareholders, suppliers, etc., or a Clearly, failure affects a firm's entire existence
bill is overdrawn, or the firm is bankrupt accord- and it has a high cost to the firm, the collabora-
ing to law. All these situations result in a discon- tors (firms and organizations), the society and the
tinuity of the firm's operations. country's economy (Warner, 1977). Beaver (1968)
showed the decline of stock market price as a
firm was coming closer to failure. Altman (1969)
* Corresponding author. provided similar conclusions. Consequently, the

0377-2217/96/$15.00 © 1996 Elsevier Science B.V. All rights reserved

SSDI 0377-2217(95)00070-4
488 A.L Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513

prediction of failure is important for all those several countries for the prediction of business
involved: owners or shareholders, managers, failure. Jones (1987) examined the techniques
workers, lenders, suppliers, clients, the commu- used for bankruptcy prediction in USA. Keasey
nity and the government. The development and and Watson (1991) explored the limitations and
use of models able to predict failure can be very usefulness of methods used for the prediction of
important for them in two different ways. First, as firm financial distress.
'early warning systems', such models are very These reviews do not include recent methods
useful to those (managers, authorities, etc.) that or applications in the field of business failure.
can take action to prevent failure. These actions There is clearly a need for a comprehensive sur-
include the decision about merger of the dis- vey of literature on business failures including
tressed firm, liquidation or reorganization type world-wide applications, models/methods and
and associated costs (Casey et al., 1986). characteristics (e.g. financial ratios). The aim of
Second, such models can also be useful in this paper is to reduce this need and provide a
aiding decision makers of financial institutions in discussion of findings based on a new framework,
evaluating and selecting firms to collaborate with which includes classification of the studies by
or to invest in. Such decisions have to take into country, method and industrial sector.
account the opportunity cost and the risk of fail-
ure. In that regard, these models are close to
those developed for the credit granting problem. 2. Financial ratios and other variables
Credit granting problem is different from the
failure problem, but lenders of a firm can use the Financial ratios were introduced early as char-
failure models to assess the probability of failure acteristics able to predict the failure of a finn.
for that firm. This probability can be an impor- The early studies were using only the ratios from
tant factor or even the main criterion in credit specific year(s) to make predictions. However,
granting decisions for financially distressed firms. failure is a continuous process. This means that
Furthermore, the idea of classifying a firm as although the appraisal of failure happens at a
healthy or not is common to both the failure and certain time, it is the result of a specific policy of
credit granting problems. Credit applicants are the firm for a number of years. Therefore, the
classified into groups according to the acceptance values of the ratios should be inspected over time
or not of their credit application. to provide full information about the progress of
Since business failure prediction became a field a firm. To get this information over time re-
of study, researchers introduced a plethora of searchers used the time trend, the coefficient of
methods for the classification and selection of variation and shift away from the trend in the
firms. Different views, requirements and reliabil- period(s) prior to failure (e.g. Meyer and Pifer,
ity needs have led researchers in using more 1970; Dambolena and Khoury, 1980; Falbo, 1991).
sophisticated methods, already applied to other In a comparison work Collins (1983) concluded
scientific fields. This diversity and large interest that Meyer and Pifer's variables did not improve
on this subject have been addressed partially in a the predictive ability compared to a simpler
few review articles. Vernimmen (1978) examined model.
various failure models and criticized their contri- A large number of ratios has been proposed in
bution and limits. Scott (1981) investigated the the literature. Courtis (1978) made an attempt to
empirical models developed as well as the identify the variables useful in predictive studies.
bankruptcy theories presented to identify the In his survey 79 financial ratios were identified
overlap between them, focusing mainly on US from various studies and were grouped in three
studies. Zavgren (1983) surveyed different meth- main categories: (a) profitability ratios; (b) man-
ods and empirical models developed for the pre- agerial performance ratios and (c) solvency ratios.
diction of corporate failure in the USA. Altman Later studies proposed additional financial ratios
(1984) presented a review of models developed in for inclusion in failure prediction analysis.
 A.I. Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513 489

A firm is not standing alone. Its performance ber of director resignations and appointments
and survival are influenced by several factors; e.g. and the changes in directors' shareholdings. The
the environment and its changes, as well as na- social importance of the firm and the strength of
tional and international economic conditions. In its bank relationship (Suzuki and Wright, 1985)
turn these can affect any model built at that time, could be also critical. The narrative portion of
possibly producing occasionally good predictions. financial statements and the information it con-
Therefore a good model, constructed under nor- tains were proposed by Tennyson et al. (1990) to
mal circumstances, may be unable to predict fail- explain bankruptcy. The analysts' forecasts on
ure successfully during periods of difficulties. earnings per share were proposed by Moses (1990)
Macro-economic variables for failure prediction and decomposition measures analysis were sug-
were proposed by Foster (1986) and Rose et al. gested by Booth and Hutchinson (1989).
(1982). Mensah (1984) noted that different eco- The lack of a normal distribution fit to most
nomic environments as well as different sectors financial ratios has been noticed for different
lead to different models for the prediction of types of firms by several authors (Eisenbeis, 1977;
failure. Such a conclusion is also derived from Karels and Prakash, 1987; Ezzamel and Mo-
Tables A.4a and A.4b in Appendix A, where not linero, 1990; Zanakis and Walter, 1994). This may
even two models are identical for those variables create poor model fits a n d / o r predictions for
found significant. Among different countries, sec- statistical methods, such as discriminant analysis
tors and periods of time the financial ratios able (more so in quadratic than linear).
to predict failure differ from study to study.
Furthermore, there are studies proposing
models for specific sectors. A cluster analysis by 3. Framework
Gupta and Huefner (1972) showed that firms in
different sectors have different characteristics. After an extensive but not necessarily exhaus-
Additionally, in different sectors, there are vari- tive review of journal articles written in English
ous adapted characteristics able to provide pre- or French, we identified a list of 158 articles on
dictions about firms. The failure of firms in each the subject. These papers were published in the
sector can be better predicted by consideration of period 1932-1994 in various journals specialized
these characteristics. So, there are special studies in accounting, finance, operations research and
about financial institution failure prediction (i.e. decision science. Working papers and disserta-
Booth et al., 1989; Lane et al., 1986; Martin, tions on the subject were not included; books and
1977; Santomero and Vinso, 1977; Zanakis and monographs were not considered, unless they
Walter; 1994) or hospital closure (Wertheim and were dedicated to the subject of this survey. All
Lynn, 1993). publications are listed in the References for in-
On the other hand, a company's performance formation purposes. However, in order to keep
and future may be influenced by characteristics this review manageable our analysis was re-
other than financial, some of them qualitative, stricted only to: (a)journal articles presenting
such as management, personnel, products, equip- models and (b) pertaining to industrial and retail
ment, etc. Zopounidis (1987) employs a set of application. The former condition (a) excluded
'strategic criteria' to assess the risk of failure of from the analysis all methodological and exposi-
French firms. These were: quality of manage- tory articles without any models. The latter con-
ment, research and development level, diversifi- dition (b) excluded essentially only a few banking
cation stage, market trend, market niche/posi- applications; this was deemed necessary to avoid
tion, cash out method and world market share. adding a large number of financial ratios specific
Similar propositions were taken by Shaw and only to bank operations. Inclusion of such diverse
Gentry (1988), while Peel et al. (1986) proposed entities would have affected negatively the depth
other qualitative variables, such as the changes in and readability of this survey. This selectivity
the lag in reporting accounts of a firm, the num- produced 47 journal articles, identified in Ap-
490 A.L Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513

pendix A, written mainly by scholars and only a Suggestions have also been made for defining
few by practitioners. more than two groups according to the associated
The tables in Appendix A depict the informa- level of risk (Zopounidis, 1987). Because of the
tion contained in the above 47 articles, based on general acceptance of the two group classifica-
a proposed framework; namely according to tion, the interest has been mainly focused on
country, year of publication, industry type, sample dichotomous classification methods, being re-
periods and method (Table A.3). A detailed list- ferred to as discriminating approaches. Methods
ing of the financial ratios included in the models in this category include discriminant analysis and
presented in these articles is summarized in Ta- its alternatives, logit/probit analysis and linear
bles A.4a and A.4b. The models are identified by probability models. Recursive partitioning algo-
the author(s) name(s) and year of publication rithms employ a Bayesian scheme for group dis-
cited in the reference list. If a paper contains crimination. Prediction of time to failure, instead
more than one model, these are characterized by of group membership, was the aim of survival
the method used, thus producing a total of 59 analysis adopted for this problem.
models presented in these 47 articles. Tables A.1 Statistical methods were the earliest employed
and A.2 in Appendix A define the abbreviations by researchers for the prediction of business fail-
used for methods and ratios respectively. Summa- ures. The increasing progress on such methods
rizing and coding this information in Appendix A has led researchers to apply and modify these
was a difficult task due to lack of uniformity on methods to financial problems with numerous
common language and ratio definitions between financial characteristics. An interesting such al-
authors. ternative is the use of mathematical programming
An examination of the articles in our survey for solving the discriminant analysis problem. The
revealed the most frequent journal outlets for ambiguity of measuring some qualitative charac-
papers on business failures. Three accounting/fi- teristics has led to different approaches such as
nance journals (Journal of Banking and Finance, individual or group judgments (Libby, 1975b;
Journal of Business Finance and Accounting and Houghton, 1984; Houghton and Senagupta, 1984),
Journal of Accounting Research) are the most multicriteria decision aid and expert systems.
frequent sources of such articles with 15, 14 and These methods are reviewed in Appendix B and
13 articles respectively. These are followed by the next sections, and illustrated with typical lit-
OMEGA (8 publications), Decision Sciences and erature examples.
The Journal of Finance (7 each) and European
Journal of Operational Research (6 articles).
5. Summary of methods, sectors and country ap-
plications
4. Survey of methods and models
A variety of methodologies have appeared in The techniques for business failure prediction,
the literature for modeling business failures. Each as presented earlier in this article, consist mainly
method has its own assumptions and different of three parts:
contributions in the field of business failure. The (i) sample selection and collection of data (vari-
basic assumption on which most of the failure ables and sample sizes);
prediction methods are based, is that firms can (ii) selection of method and specific variables
generally be split into groups, usually two; the (ratios) to develop a predictive model;
group of healthy and the group of failing firms. (iii) model validation, i.e. statistical significance
Accordingly, firms are characterized by a variable and accuracy of results.
y such that For each one of these parts, considerable discus-
sion exists in the literature. The selection of the
(01 ifthei-thfirmisbankrupt, method can be the most important part. This
Yi = if the i-th firm is not bankrupt. (1) selection depends on the data to be analyzed and
 A.I. Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513 491

Table 1 try. Some studies employed more than one

Number of studies by country method to develop failure prediction models; so
Country No. of studies the total frequency of methods is larger than the
Australia 2 number of 47 studies considered in this analysis.
Canada 1 Some countries (i.e. Finland, Greece) present
Finland 4 only a few case study applications, because the
France 5
Greece 7 subject of business failure has been analyzed only
Israel 1 recently and by a handful of authors. From the
Italy 2 countries for which more than two methods are
Japan 1 used, the primarily employed methods are dis-
Sweden 1 criminant and logit analysis. One would suspect
Netherlands 1
UK 9 that the special interest in specific industries in
USA 13 different countries would favor different models.
This does not seem to be the case here. Most
studies used three types of firms in their samples:
the objectives of the study. The data selection is manufacturing firms, combination of manufactur-
influenced by their availability or reliability, the ing and retailing firms or firms from several in-
definition of failure or underlying failure theory dustrial sectors. Only two studies focused on a
and the study objectives. specific industrial sector, the textile industry (Alt-
The prediction of business failures, being an man et al., 1974; Michalopoulos et al., 1993).
area of high interest, was faced in many countries These two studies (one in France and one in
by various authors. This implies that failure pre- Greece) probably indicate that the textile indus-
diction is a field of world-wide interest. Table 1 try is of high interest in the specific countries. In
summarizes by country the number of studies general we found no relation between the indus-
contained in Table A.3, Appendix A, of 47 indus- trial sectors and the methods used for the predic-
trial and retail applications. This table reflects tion of failure.
the world-wide interest and the relative fre-
quency by country.
6. Summary of ratios by country application
One may naturally wonder whether different
methods are more frequently used in some coun- Business failure prediction models are based
tries. Table 2 summarizes this frequency by coun- on the financial characteristics of firms in the

Table 2
Use of methods a by country
COUNTRY DA LPM Logit Probit RPA Electre ES SA UA Other
Australia 1 1
Canada 1
Finland 3 1 1
France 4 1
Greece 3 5 3 1 1
Israel 1
Italy 2
Japan 1
Sweden 1
Netherlands 1
UK 6 4
USA 4 8 1 1

Total 26 6 15 2 2 1 1 1 1 1
a See Table A.1, Appendix A, for method definitions.
492 A.I. Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513

form of financial ratios. Most of these models The financial ratios employed in the 47 indus-
included entirely or mainly financial characteris- trial applications we reviewed are presented in
tics (financial ratios) of the firms available. Ta- Tables A.4a, b. Table 3 summarizes by country
bles A.4a and A.4b present a large number of the most commonly used ratios (those with total
ratios used in predictive studies of industrial fail- frequency less than 4 were not included). It should
ures. be noted that one ratio, (S-TP)/TC, was used in
As can be easily seen in these tables, there are 5 UK models, four of which are from one study
no financial characteristics common to all predic- referring to different years before failure.
tive studies. The ratios found useful in earlier The most frequently used financial ratios are
studies (e.g. Altman, 1968) were the first under W C / T A (16 times in 5 countries), T D / T A (15
consideration by many researchers and subse- models in 4 countries), and CA / CL , E B I T / T A
quently used in later studies. and N I / T A with 12, 12 and 11 models respec-
The availability of data was a major limitation tively. Ratios preferred or found more significant
in the selection of the variables. As the number in different country studies can easily be identi-
of variables available was rather high in many fied in Table 3. For example, T D / T A was en-
cases, some further selection was feasible. Too countered more frequently in US studies (but was
many ratios in a model may cause multicollinear- never used in the UK), while CF/S prevailed in
ity, which of course can be reduced or eliminated Finnish studies. Financial ratios often encoun-
by factor analysis. This approach has been over- tered in the studies of one country sometimes
looked by many authors. Another favored ap- were ignored in the other countries. For example
proach involves stepwise procedures, which re- the ratios T D / T A and N I / T A found large appli-
duce the number of variables in the model but cation in US studies but they were not used at all
pay no attention to multicollinearity. Additionally in UK studies.
a large number of characteristics in a model may
cause difficulties in data collection for real cases.
7. Concluding remarks
Furthermore, as Hamer (1983) mentioned, the
variable set should be constructed on the basis of This paper presents a comprehensive survey of
(a) minimizing the cost of data collection and (b) literature on business failures, summarized ac-
maximizing the model applicability. cording to a new framework. We first identified a

Table 3
Use of financial ratios a by country

WC TD CA EBIT NI CF QA CF RE S GP NI Cash PBT S-TP Inv. QA TA

TA TA CL TA TA TD CL S TA TA TA SE TA S TC S TA GNP

Australia - - 2 1 1 - - - 1 . . . . . . . .
Canada 1 1 1
Finland - 1 1 3 1 2 1 6 - 2 . . . . . . . .
France 1 . . . . 1 - 1 - 1 . . . . . . . .
Greece 5 5 2 1 2 - - - 1 - 6 1 . . . . .
Israel - - 1 . . . . . . . . . . . . . . .
Italy . . . . . . . . . . . 1 . . . . . .
Japan 1 . . . . . . . . . . . 1 . . . . .
Sweden - - - 1 . . . . . . . . . . . 1 - -
Nether . . . . . . . . . 1 1 - 1 . . . . .
lands
UK 5 - 1 - - 5 - 1 - - 3 2 5 5 2 1 2
USA 4 8 5 5 7 6 3 1 3 2 - - 2 - - 1 3 2

Total 16 15 12 12 11 9 9 8 7 7 6 6 5 5 5 4 4 4

a See Table A.2, Appendix A, for ratio definitions.

 A.L Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513 493

list of 158 articles on this topic and then proceed Table A.2
to analyze those articles that contained predictive Code names for financial quantities and ratios
models for mainly industrial failures. AE Administrative Expenses
AP Accounts Payable
The major conclusions of this analysis can be
APP Average Payment Period (for accounts payable)
summarized as follows: AV Added Value
- There is a world-wide interest for business CA Current Assets
failure studies. Such studies were made in indus- Ca. Cash
trial countries (USA, UK, France) as well as in CF Cash Flow
CL Current Liabilities
countries under development (Greece).
D Depreciation
- The discriminant analysis method was the most EBIT Earnings Before Interest & Taxes
frequently used in business failure studies. Logit FA Fixed Assets
analysis ranks second among the methods used. FAP Free Assets Percentage ( = non-collateralized
Recently, some multicriteria methods have been tangible assets divided by the total
tangible assets)
employed in business failure studies, primarily to
GFA Gross Fixed Assets
take into account qualitative variables. The ques- GNP Gross National Product
tion remaining is whether these methods can GP Gross Profit
respond efficiently to the problem. IE Interest Expenses
- Many methods appeared mainly after the 80's Inv. Inventory
for the prediction of business failure in order to LA Liquid Assets
LTD Long Term Debt
overcome the limitations of DA. MVE Market Value of Equity
- The most important financial ratios came from CE Capital Employed
the solvency category (e.g. W C / T A , TD/TA). NCI No Credit Interval
The profitability ratios were also important, indi- NI Net Income
NP Notes Payable
cating that the viability of a firm depends to a
NW Net Worth
large extent on profit making. OE Operating Expenses
Although the framework of this study involved PBD Profit Before Depreciation
only financial ratios as predictors of failure, the PBT Profit Before Taxes
research trend is on using non-financial a n d / o r Prod Production ( = Sales + Inventory)
QA Quick Assets
qualitative variables (which of course can be ac-
R Receivables
commodated by our framework). RE Retained Earnings
The survey of methods, bibliography and anal- ROI Return on Investment
ysis of articles on business failures presented in S Sales
this paper are hoped to be useful to teachers, SC Shareholders' Capital
researchers and practitioners in this important SE Shareholders' Equity
SOP Stock Option Percentage ( = number of shares of
area. common stock potentially issued from
Outstanding stock options as a percentage of the
number of outstanding shares of common stock)
A p p e n d i x A
SP Stock Price
STD Short Term Debt
T Taxes
Table A. 1 TCr Total Creditors
Code names for methods TP Trading Profit
DA Discriminant Analysis TA Total Assets
LPM Linear Probability Model TC Total Capital
Probit Probit Analysis TD Total Debt
Logit Logit Analysis TE Total Expenses
RPA Recursive Partitioning Algorithm TL Total Liabilities
SA Survival Analysis WC Working Capital
UA Univariate Analysis
ES Expert Systems
Table A.3
Characteristics of industrial failures studies
No. Country Author (year publ.) Industry type Sample period Method
1 Australia Izan (1984) Several 1963-74 DA
2 Australia Messier & Hansen (1988) Land development firms - ES
3 Canada Altman & Levailee (1981) Manufacturing/retailing 1970-79 DA
4 Finland Laitinen (1991) Several - DA
5 Finland Laitinen (1992) Manufacturing 1980-85 DA
6 Finland Laitinen (1993) Several 1986-88 LPM
7 Finland Luoma & Laitinen (1991) Manufacturing/retailing - SA
8 Finland Luoma & Laitinen (1991) Manufacturing/retailing - DA
9 Finland Luoma & Laitinen (1991) Manufacturing/retailing - Logit
10 France Altman et al. (1974) Textile industry 1968-69 DA
11 France Conan & Holder (1979) Manufacturing 1970-75 DA
12 France Collongues (1977) Manufacturing - DA(1)
13 France Collongues (1977) Manufacturing - DA (2)
14 France Micha (1984) Manufacturing 1975-80 DA
15 France Zollinger (1982) - 1975-76 Electre
16 Greece Gloubos & Grammatikos (1988) Manufacturing 1977-81 DA
17 Greece Gloubos & Grammatikos (1988) Manufacturing 1977-81 Logit, Probit, LPM
18 Greece Grammatikos & Gloubos (1984) Manufacturing 1977-81 DA
19 Greece Grammatikos & Gloubos (1984) Manufacturing 1977-81 LPM
20 Greece Michalopoulos et al. (1993) Textile industry - RPA
21 Greece Slowinski & Zopounidis (1994) Several 1988 RS
22 Greece Theodossiou (1991) Manufacturing - DA, Logit, LPM
23 Greece Vranas (1991) Manufacturing 1979-84 LPM, Logit
24 Greece Vranas (1992) Manufacturing 1980-84 LPM
25 Israel Tamari (1964) - 1956-60
26 Italy Appetiti (1984) Manufacturing 1979-81 DA (1)
27 Italy Appetiti (1984) Manufacturing 1979-81 DA (2)
28 Italy Falbo (1991) Manufacturing - DA
29 Japan Takahashi et al. (1984) - 1977 DA
30 Sweden Skogsvik (1990) Mining/manufacturing 1966-80 Probit
31 Netherlands Bilderbeek (1977) a _ 1950-79 DA
32 UK Lis (1972) b _ 1964-72 DA
33 UK Peel & Peel (1987) Manufacturing 1982-85 DA, Logit
34 UK Peel et al. (1986) Manufacturing - Logit
35 UK Keasey & McGuinness (1990) Several 1976-84 Logit (1 year prior)
36 UK Keasey & McGuinness (1990) Several 1976-84 Logit (2 years prior)
37 UK Keasey & McGuinness (1990) Several 1976-84 Logit (3 years prior)
38 UK Keasey & McGuinness (1990) Several 1976-84 Logit (4 years prior)
39 UK Keasey & McGuinness (1990) Several 1976-84 Logit (5 years prior)
40 UK Keasey et al. (1990) Several 1976-84 Multilogit
41 UK Taffler (1976) b Manufacturing 1968-73 DA
42 UK Taffler (1977) b Manufacturing 1969-76 DA
43 UK Taffler (1983) - 1972-77 DA
44 UK Tisshaw (1976) b Manufacturing 1975-76 DA
45 USA Altman (1968) Manufacturing 1946-65 DA
46 USA Altman et al. (1977) Manufacturing/retailing 1969-75 DA
47 USA Beaver (1966) - - UA
48 USA Dambolena and Khoury (1980) Manufacturing/retailing 1969-75 DA
49 USA Deakin (1972) - 1965-75 DA
50 USA Frydman et al. (1985) Manufacturing/retailing 1971-81 DA
51 USA Frydman et al. (1985) Manufacturing/retailing 1971-81 RPA(1)
52 USA Frydman et al. (1985) Manufacturing/retailing 1971-81 RPA(2)
53 USA Gilbert et al. (1990) Several 1974-83 Logit
54 USA Lau (1987) - 1971-80 Logit
55 USA Ohlson (1981) Manufacturing 1970-76 Logit
56 USA Platt & Platt (1990) Several 1972-76 Logit
57 USA Tennyson et al. (1990) - 1980 Logit
58 USA Zavgren (1985) Several 1972-78 Logit
59 USA Zmijhewski (1985) - 1972-78 Probit
a Cited by Altman (1984).
b Cited by Taffler (1984).
 A.I. Dimitras et al. /European Journal of Operational Research 90 (1996) 487-513 495

Appendix B that were applied to discriminate between failed

and non-failed firms.
B.1. Univariate statistical methods Earlier methods tried to predict failure by
using data from one to five years prior failure;
they were not able to estimate really the time of
Univariate statistical methods were the first
failure. Lane et al. (1986) employed a propor-
ones used in an attempt to discriminate between
tional hazard model, proposed by Cox, for the
healthy and failing firms. Some of the early stud-
prediction of bank failure, called survival analysis
ies were those of Fitzpatrick (1932), Smith and
(SA). A full description of the method is provided
Winakor (1935), and Merwin (1942). Later, Beaver
by Cox and Oakes (1984).
(1966) introduced a univariate technique for the
SA assumes that both failed and non-failed
classification of firms in two groups by using
firms in the sample are from the same popula-
some financial ratios. The ratios were used indi-
tion. Non-failed firms are considered as some
vidually and a cut-off score was calculated for
kind of censored observations. The risk of failure
each ratio on the basis of minimizing misclassifi-
is measured by the survival time that is calculated
cations. In Beaver's study, the ratios Cash
for each firm. Assuming that T is the time on
Flow/Total Debts, Net Income/Total Assets and
which a firm will fail, the survivor function S(t) is
Total Debts/Total Assets were the ones provid-
the probability that the time T is greater than t.
ing the highest discrimination capability. A de-
The probability F(t) for a firm to fail before t is
tailed description of the univariate method is
given by Beaver (1968) and Elam (1975). The F(t) = 1 - S(t). (B.1)
univariate method was later criticized, in spite of The hazard function h(t) is given by
its good predictions, because of the correlation
among ratios and the associated difficulty in get- h(t) =f(t)/S(t) = -S'(t)/S(t), (B.2)
ting clear signals when different variables provide and
conflicting predictions for a firm. Also, there is a h( t [x ) = h o exp(x't~), (B.3)
large number of factors that describe efficiently
the financial status of a firm, so that a single ratio where x is the variable vector (vector of charac-
cannot include the full information (see Edmis- teristics) and t~ is the coefficient vector.
ter, 1972). Even though univariate methods re- Vector a, which is the parametric part of the
ceived considerable criticism, they have paved the function, is estimated by a technique similar to
road for subsequent research on business failures. that of maximum likelihood; ho(t), the non-para-
Researchers recognized that business failure can metric part, is calculated by setting x = 0.
be caused and affected by many different factors The survivor function SOl x) is then
at the same time. The idea of adopting a multi- S ( t l x ) = So(t ) exp(~'x), (B.4)
variate index of risk was rather old. Tamari (1964)
presented such a model for Israeli firms, using six where So(t) is given by
variables weighed according to the opinion of
financial analysts, economists and credit analysts. So(t ) = exp( - f/ho( u) du). (B.5)
Different multivariate statistical methods were
As ho(t) is distribution free, SA is a 'semipara-
used in an attempt to relate the predictions to
metric' methodology, and it partially skips the
more than one variable, in order to try to improve
criticisms against parametric techniques. Further-
classification results.
more, the method appears to be more natural in
dealing with the failure problem.
B.2. Survival analysis SA was employed by Luoma and Laitinen
(1991) for the prediction of business failures.
There were some methods, earlier used in Their sample consisted of 36 industrial and retail-
medicine, biochemistry or other scientific fields, ing failed Finnish companies, matched by size
496 A.L Dimitras et aL / European Journal of Operational Research 90 (1996) 487-513

Table A.4a
Financial ratios included in industrial failure models
WC TD CA EBIT NI CF QA CF RE S GP NI Cash PBT S-TP Inv. QA TA
No.
TA TA CL TA TA TD CL S TA TA TA SE TA S TC S TA GNP
1 +a +
2 + + +
3 + + +
4 + + + + +
5 + + +
6 + + + +
7 + +
8 +
9 + +
10

11
12 +
13
14 +
15 +
16 + + + +
17 + + +
18 + +
19 + +
20 + + + +

21
22 + + +
23
24 + +
25 +
26
27
28
29 +, (+) (+)b
30 +

31 + +
32 + +
33 + + (+) (+)
34 + (+)
35 +
36 + + +*e
37 + + +
38 +
39 +
40 + + + + +*

41
42
43
44 +
45 + + + +
46 + +,(+) +
47 + •+- +
 A.L Dimitras et al. /European Journal of Operational Research 90 (1996) 487-513 497

CA EBIT GP IE MVE MVE S Salaries SC AV CA Cash Cash CL CL FA

APP No.
TA IE S GP TC TD nFA AV TC TA TD CL S TA TD TA
+ + 1
2
3
4
5
6
+ + 7
+ + 8
+ 9
+ + 10

+ 11
+ 12
13
+ + 14
15
16
17
+ 18
19
20

21
22
23
24
25
26
27
+ +* 28
-t- 29
30

-I- 31
32
+* 33
34
35
36
37
+ 38
39
+ 40

41
+ + 42
+ + 43
44
45
+ + 46
47
498 A.I. Dimitras et al. ~European Journal of Operational Research 90 (1996) 487-513

Table A.4a (continued)

Financial ratios included in industrial failure models
WC TD CA EBIT NI CF QA CF RE S GP NI Cash PBT S-TP Inv. QA TA
No.
TA TA CL TA TA TD CL S TA TA TA SE TA S TC S TA GNP
48 + +
49 + + + + + + + +
50 + + + + + +

51
52
53 + (+)
54 +
55 + +, (+) +* (+)
56 +,(+) +,(+)
57 + + + +
58 + + + + +
59 + + +

a + indicates that the ratio is used in the corresponding model.

b ( + ) indicates that the ratio is used in the corresponding model somehow transformed.
c + . indicates that the ratio is used in the corresponding model inversed.

and industry type to 36 non-failed firms. The goal ent means but equal dispersion matrices. The
of the method was to calculate the survival time objective of this method is to obtain the linear
starting from the end of accounting period, as- combination of the independent variables that
suming that at this time the failure process starts. maximizes the variance between the populations
This time is not necessarily a natural start of the relative to within group variance. The method
failure process. This model consisted of financial estimates a discriminant function that is a coeffi-
ratios (see Appendix A), as well as a measure of cient vector A, (a 1, a 2 , . . . , a , ) , and a constant
the size; it performed satisfactory compared to term a 0. The linear combination of the variables
D A and logit analysis. provides for each firm a Z-score, according to
The interpretation of results, according to the
Z i=a O+alXil +a2xi2+a3xi3+ ... +anXin ,
expected failure time, provides decision makers
(B.6)
with important information about a firm. The SA
method, although a viable alternative to statistical where Z i is the Z-score for firm i and
methods, has not been often applied for the xil, xi2 . . . . . X/n are the n independent variables
prediction of business failure. for firm i.
A cut-off score is calculated according to the
B.3. Discriminant analysis a-priori probabilities of group membership and
the costs of misclassification. Based on its Z-score
Discriminant analysis (hereafter DA) is a mul- and the cut-off score, a firm is classified to the
tivariate analytical method. D A was used in a failure or the non-failure group. If the assump-
large number of studies for the development of tion of equality of dispersion matrices is not
models able to predict business failure. First, we satisfied, then quadratic D A instead of linear D A
consider that any firm i is characterized by a may be advantageous. Altman et al. (1981) pro-
vector X the elements of which are measure- vided a detailed description of D A and its finan-
ments of n independent variables x (predictors). cial applications.
For the two populations (failing and non-failing Altman first proposed D A for the failure pre-
firms) it is assumed that the independent vari- diction in 1968. This study was of much interest
ables are distributed within each group according as it was the first to relate the classification of
to a multivariate normal distribution with differ- firms to more than one variable, using DA. He
 A.L Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513 499

CA EBIT GP IE MVE MVE S Salaries SC AV CA Cash Cash CL CL FA

APP No.
TA IE S GP TC TD nFA AV TC TA TD CL S TA TD TA
48
+ + + 49
+ (+) 50

+ + 51
52
53
54
+ + 55
56
+ 57
58
59

selected 33 manufacturing firms failed in the pe- mine the sensitivity of the model to distinguish
riod 1946-1965 matched by industry and asset between firms facing permanently and temporal
size to 33 non-failed firms. Applying linear DA, financial problems.
he constructed a 5-variable model to classify Moyer (1977) pointed out that Altman's (1968)
b a n k r u p t / n o n - b a n k r u p t firms into the two model had poor predictive ability and he used a
groups. The ratios used in the model were chosen stepwise D A method to construct a better model.
because of their popularity in the relative litera- Joy and Tollefson (1975) and the relative replies
ture and their potential relevancy to the study. by Altman and Eisenbeis (1978) and Scott (1978)
These financial ratios were calculated for up to discussed the applicability of D A to financial
five years prior to the failure. problems. Based on Altman's (1968) model, Joy
The discriminant function proposed was and Tollefson (1975) focused their criticism on
Z -- 0.021X~ + 0.014X 2 + 0.033X 3 + 0.006X 4 the predictive ability, the relative discriminatory
power of variables and classification efficiency of
+ 0 . 9 9 9 X 5,
DA.
where: Altman et al. (1977), using quadratic DA, con-
X 1= WC/TA. structed a 7-variable model called the 'Zeta ®
X 2 RE/TA.
= model'. This study was based on a sample of 53
X 3 = EBIT/TA. manufacturing and retailing firms that went
X 4 = MVE/TD. bankrupt in the period of 1969-1975. The
X 5 = S/TA. bankrupt firms were matched by industry group
(See Table A.2 for interpretation of the used and year of data to 58 non-bankrupt firms. They
code names.) used log-transformations of variables to improve
The cut-off score minimizing the total number their normality. To overcome the assumption of
of misclassifications was 2.675. Firms with a Z- equal dispersion matrices, required by linear DA,
score less than this cut-off score were assigned as they used quadratic DA. The variables entering
bankrupt. A test of the model was provided on a the model were:
holdout sample consisting of 99 distressed firms X 1= EBIT/TA.
with 'similar asset size to the original group. As X 2 = TA.
the holdout sample's firms were really assigned to X 3 = EBIT/IE.
neither of the two initial subsamples of firms X4 = CA/CL.
(bankrupt/non-bankrupt), the goal was to deter- X 5 = RE/TA.
500 A.L Dimitras et al. / European Journal o f Operational Research 90 (1996) 487-513

Table A.4b
Financial ratios and other ratios included in industrial failure models
F1.F IE IE Inv. NI NI NI NI NW PBD PBT R SC SE WC FA
No. NCI
TD S TE-De nWC CL NCe S TD TD TD CL S TA TA LTD NW
1
2
3 +a
4
5 +
6 +
7
8
9
10

11 +
12 +
13
14
15
16
17
18 +
19 +
20

21
22
23 + +
24 + +
25 +(1/) +
26 +
27
28 + + +
29
30 +

31
32
33
34 +
35
36 +
37
38
39
40 +

41 +
42 (+)
43 + +
44
45
46
47
 A.L Dimitras et al. ~European Journal of Operational Research 90 (1996) 487-513 501

No. Other ratios

1 Funded Debt/Shareholders Funds
2
3 Rate of growth of Equity-Rate of Growth of TA
4
5
6
7 ROI *APP Official Expenses/corrected expenses
8
9
10 Av/Prod. EBIT/SE LTD/(LTD + SE) (LTD + SE)/STD SE/S SE/TD

11 (CA-Inv.)/Inv. EBIT/TD (LTD + SE)/(TD + SE)

12
13 EBIT/S WC/Inv.
14 FA/AV (LTD + SE)/CE Trade Debt/Purchases (WIP - customers' prepayment + trade acc. rcvble)/output
15 SE/LTD NW/(SE + LTD) PBT/Prod. W C / N e e d s in financing
16 GP/CL
17
18 NP/TA
19
20 (CA-Inv)/CL TD/SE

21 (General & Administrative Expences)/S

22 LTD/TA
23 (R + Ca)/CA
24 QA/(TE-D) (R + LA)/CL
25 Prod./Inv. Prod./WC
26 DE/GFA
27 I E/ AV
28 D E / AV
29 NW/TA
30 T/PBT IE/TD

31 AP/S
32
33
34 S/TD
35
36
37 Capital Gearing
38
39
40 Capital Gearing

41 TD/NCe WC/NW
42
43
44 QA/NCe
45
46
47
502 A.L Dimitras et at,./European Journal of Operational Research 90 (1996) 487-513

Table A.4b (continued)

Financial ratios and other ratios included in industrial failure models
FI" F IE IE Inv. NI NI NI NI NW PBD PBT R SC SE WC FA
No. NCI
TD S TE-De nWC CL NCe S TD TD TD CL S TA TA LTD NW
48 (+) + +,(+)
49
50

51
52
53
54 (+)
55
56
57
58
59
a -F indicates that the ratio is Used in the corresponding model.

b ( + ) indicates that the ratio is used in the corresponding model somehow transformed.
c + • indicates that the ratio is used in the corresponding model inversed.

X 6 = MVE/TC. modified DA to get the probability of a firm to be

X 7 = Stability of variable X 1 over a ten-year tree. a member of the two classes (bankrupt or non-
The development of the Zeta ® model created bankrupt). Tamer (1983) made some modifica-
an increased interest in the application of DA. tions of DA method deriving for each firm a
The method was also employed for the classifica- PAS-score (Performance Analysis). The PAS-
tion of firms in different areas. The extensive score allows the relative performance of a firm to
application and study of DA pointed out rather be tracked throughout the whole performance
early the problems of the method. Eisenbeis range and within the at risk region. Taffler's
(1977) summarized 7 main problems in the appli- (1983) PAS-score method has been improved by
cation of it: Giordano (1986).
1) violation of the assumption of multivariate Studies employing DA for the business failure
normal distribution of the variables; prediction are presented in Appendix A. DA was
2) use of linear instead of quadratic discrimi- also applied in other classification problems in
nant functions when the group dispersions are the field of finance. Lane (1972) applied DA in
unequal; the credit-scoring problem (sometimes similar to
3) unsuitable interpretation of the role of the the bankruptcy problem, as mentioned before).
independent variables; Recently, Laitinen (1992), studying venture capi-
4) reduction in dimensionality; tal decisions, used DA for the prediction of fail-
5) group definition; ure of newly founded firms.
6) inappropriate choice of a priori probabili- DA has been highly criticized, thus forcing
ties a n d / o r costs of misclassifications; researchers to suggest new methods and compare
7) problems in estimating classification error them to DA. This explains why DA is used in so
rates to assess the performance of the models. many studies in the business failure literature
Analogous problems were also reported by (see Table 2 and Appendix A), although after the
Malecot in his research work on the applicability 80's its use has decreased.
of DA (Malecot, 1981, 1986, 1991). In summary, DA provides the decision maker
Other researchers suggested improvements to with a dichotomous classification of the firms.
DA to get better prediction results. Deakin (1972) This classification, although important, does not
 A.I. Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513 503

No. Other ratios

48 Funded d e b t / n e t WC
49 CA/S QA/S WC/S
50

51
52
53
54 WC/TD
55
56
57
58 R/Inv.
59

provide any estimate of the associated risk of Serious statistical problems were mentioned in
failure. On the basis of this idea the next step in the application of LPM. The error terms are
failure prediction was to use methods and models heteroskedastic and their distribution is not nor-
able to provide a probability of failure. Re- mal. Also there is a problem of interpretation, as
searchers proposed linear conditional probability the predicted value of probability can lie outside
models and logit/probit analysis. the (0-1) interval. Altman et al. (1981) provided
some comments and solutions for these. On the
B.4. Linear probability model other hand, while underlying assumptions of D A
and LPM are not similar, the results of the meth-
As the probability of failure was an attractive ods are identical. These might explain in part why
way for dealing with failure, researchers devel- the method did not find large application for the
oped models estimating the probability of failure. prediction of business failure (see Table 2); how-
Linear probability models (LPM) were proposed ever, there are studies where LPM was preferred
as an alternative to DA. over D A for convenience (e.g. Theodossiou, 1991).
T h e linear probability model is a special case of Meyer and Pifer (1970) first employed LPM
ordinary least squares (OLS) regression with a for bank bankruptcy prediction. Other studies
dichotomous (0-1) dependent variable. A de- employing LPM for business failure prediction
tailed description of the method is given by Gu- are presented in Appendix A.
jarati (1988). The method assumes that the
dummy variable y in Eq. (1), that represents the B.5. Logit and probit analyses
membership of a firm i in one of the specified
Multivariate conditional probability models
groups, is a linear combination of n characteris-
were later introduced into the failure prediction
tics of the firm. By transformation (see Collins
literature. These methods, based on a cumulative
and Green, 1982) the probability Pi for a firm to
probability function, provide the probability of a
fail is given by
firm belonging to one of the prescribed classes,
ei = ao + a l X i l -I- a 2 x i 2 + a3xi3 -I- • • • + a n X i n given the financial characteristics of the firm.
(B.7) In the logit method the probability of a firm i
to go bankrupt, given the vector of variables X i,
where a0, a 1. . . . . a , are OLS estimates and is
xi~, xi2 . . . . . x~, are the n independent variables
for firm i. P( X i, a) = F( 6 + aXi), (B.8)
504 A.L Dimitras et al. / European Journal of OperationalResearch 90 (1996)487-513

where F(6 + ~X i) is the cumulative logistic func- obtain better classification accuracy. Zavgren
tion: (1985) developed a measure of the information
1 contained in a logistic function using measures of
F( 6+aXi ) -<a +,~x,) " (B.9) entropy to assess the uncertainty of unexpected
l+e failure. Keasey and McGuinness (1990) criticized
The logit method provides the probability the results of such an approach as not applicable
P ( X i, 4) for a firm to go bankrupt. Based on that in their application of UK firms. Keasey et al.
probability a firm is classified as bankrupt or (1990) developed a multilogit model to classify
non-bankrupt, using a 'cut-off probability, at- firms according to the time they are expected to
tempting to minimize the type I and type II fail.
errors. The model coefficients are obtained by While logit analysis seems preferable to DA,
maximizing the log-likelihood function. because of the limitations of DA, comparative
Logit analysis was first proposed for bank fail- studies between the two methods have not proved
ure prediction by Martin (1977) and for the pre- a higher classification accuracy for all cases and
diction of business failure by Ohlson (1980). types of samples (e.g., Press and Wilson, 1978;
Ohlson selected 105 industrial firms failed in the Collins and Green, 1982). Hamer (1983) com-
period 1970-1976. All firms had to have been pared DA to the logit method for different data
traded on the stock exchange during the three sets, concluding that the models derived are of
years before failure. The non-failed firms were comparable ability to assess the probability of
selected at random. The goal was to construct failure.
three models able to predict firm failure up to Logit analysis has found considerable applica-
three years prior to actual failure. tions in failure predictions. Many models in sev-
The variables inserted to the three models eral countries were constructed as stated in ap-
estimated by Ohlson were: pendix A and Table 2. Studies presented after
1981 used mostly logit analysis, mainly because of
S 1 = Iog(TA/GNP).
the limitations of DA.
X2= TL/TA. Probit models are similar to the logit ones.
X3= WC/TA. The main difference between them is that for the
calculation of probability
X4 = CL/CA.
X5=( ~ ifTL>TA,
P( X i, a) = F( 6 + ~Xi),
ifTL <TA. F(6 + aX i) is the cumulative standard normal dis-
X 6 = NI/TA. tribution function:
X 7= FPO/TL,
1
P ( a ' l - a X i ) = fa_:~xi(27,r)l/2 e - z 2 / 2 dz. (B.10)
where FPO is Funds Provided by Operation.
1 if NI < 0 for the last two years, Maximum likelihood estimation is employed as
X8-- 0, otherwise. in logit analysis. Grablowsky and Talley (1981)
NI t - Nit_ 1 compared probit analysis and DA for classifying
credit applicants. They tested the uniqueness of
g9-- INItl + I N I , - I I '
coefficients, the multicollinearity effect and clas-
where NI t is net income for the most recent sification accuracy, concluding that probit analy-
period. sis can be an alternative to DA.
Appendix A presents some more studies for As shown in Table 2, studies using probit
the prediction of business failure using logit anal- analysis are much fewer than logit. This might be
ysis. The basic techniques of logit analysis were because probit analysis requires more computa-
explored and extended by other researchers to tional effort compared to logit, as it involves
 A.L Dimitras et aL / European Journal of Operational Research 90 (1996) 487-513 505

non-linear estimation (Gloubos and Gram- CashFlow/ TotalDebt

matikos, 1988).

B.6. Recursive partitioning algorithm

The recursive partitioning algorithm (RPA) is a C a s h / ~
non-parametric classification technique. The
<0 ~ 7 8 6
method starts with the sample of firms, their
financial characteristics, the actual group classifi-
cation, the prior probabilities and the misclassifi-
MarketValueof EquityY..LL~
TotalCapitalization
cation costs. A binary classification tree is built,
where a rule is associated to any node. These are,
usually, univariate rules; that is a certain financial
characteristic and a cut-off point that minimize
the cost of misclassification for the rest of the < ~ 0 9 5
firms. The risk of misclassification in any node t,
R(t), is
1 n,(t)
n2(t )
Fig. B.la. The RPAt tree in Frydman, Altman and Kao
R(t) = (C21 + C12)PIP2p(t) (1985).
N2 NI '

(B.11)
where:
N1, N2 : The total number of firms in each
CashFlow/TotalDebt
group (failed and non-failed firms).
nl(t), n2(t) : The number of firms in each group
on node t.
C21 : Cost of misclassifying a firm in
group 1 while it is in group 2.
C12 : Cost of misclassifying a firm in Fig. B.lb. Tile RPA2 tree in Frydman, Altman and Kao
group 2 while it is in group 1.
(1985).
PI' P2 : Prior probability of a firm to be a
member of group 1 or group 2. compare resulting classification trees to models
p(t) : Probability of classifying a firm on derived by DA. A sample of 58 bankrupt indus-
node t. trial companies and of 142 non-bankrupt manu-
After the classification tree is constructed, the facturing and retailing ones was selected at ran-
risk of the final nodes and the risk for the entire dom from the period 1971-1981. Two R P A clas-
tree is calculated. For the classification of any sification trees and two D A models were con-
new object (firm), the object descends the tree structed and compared in the study.
and falls into a final node that identifies the The R P A trees, R P A 1 and R P A 2 (see Fig.
group membership for the specific firm and the B.la and Fig. B.lb respectively) were constructed
associated probability. Breiman et al. (1984), pro- for different costs of misclassification. For R P A 1,
vided an extended description of the method, c12 = 20 and c21 = 1 and for R P A 2, c12 = 50 and
including theory of binary trees, splitting rules, c21 -- 1. For both trees, the prior probabilities for
etc. bankrupt and non-bankrupt groups (see Eq.
Frydman et al. (1985) first employed R P A as (B.11)) were set at (P1, P 2 ) = (0.02, 0.98). From
an alternative method to study the failure prob- the various trees constructed according to the
lem. The purpose of this study was to introduce above parameters, the trees with the smallest
R P A for the prediction of business failure and to cross validation costs were selected.
506 A.L Dimitras et al. / European Journal o f Operational Research 90 (1996) 487-513

The D A models were constructed using a for- One of the main advantages of R P A is that the
ward stepwise procedure. The first model, D A 1 binary tree easily explains failure for a specific
included 10 variables, while model D A 2 included firm. This simplicity is eliminated if, instead of
only the four most significant variables as pro- single variable rules, rules of linear combination
vided by the stepwise method. These discrimina- of characteristics are used. In such cases the
tion functions were resulting model can not easily explain failure.
The estimation of the tree is also difficult. From
Z 1 = - 1.761 + 5.452X 1 + 1.758X 2 + 0.505X 3
a decision m a k e r point of view, R P A just classi-
+ 1.850X 4 - 6.292X 5 - 1.021X 6 + 8.970X 7 fies firms into categories of risk. It does not
permit comparisons between firms in the same
- 1.995X 8 + 3.482X 9 - 1.033X10,
category, making it difficult to assess the relative
Z 2 = - 4 . 0 4 1 + 5.322X 1 + 0.622X 2 + 0.712X 3 performance of firms.
Appendix A presents studies using R P A for
+ 1.149X 4,
failure prediction. In spite of its attractiveness,
where: R P A did not find a large application for the
X 1= NI/TA. prediction of business failure, as shown in Table
X 2 = CA/EL 2. Nevertheless, the method was proposed for
X 3 = log(TA). other classification problems in finance. Marais
S 4 = MVE/TC. et al. (1984) and Srinivasan and Kim (1988) used
X5 = CA/TA. R P A for commercial bank loans classifications.
X 6 = CF/TD.
X 7 = QA/TA. B. 7. Mathematical programming
X 8 = QA/CL.
X 9 = EBIT/TA. Mathematical programming methods were em-
X10 = log(IE + 15). ployed early on to solve the problem of grouping
The results and classification efficiency of R P A items into one or more groups. Different cluster-
trees in this study were good; however, direct ing and discrimination methodologies based on
comparison with D A results was difficult, as D A mathematical programming were proposed in
provides a continuous scoring system while R P A many studies. The basic goal of these approaches
provides categories of risk for failure. was to escape the assumptions and restrictions of
R P A has the characteristics of both the uni- D A and improve the classification accuracy. Sev-
variate and multivariate methods. Classification eral mathematical programming formulations and
on a node is m a d e upon the rule of a single discrimination criteria were proposed. Such for-
variable (although rules in the form of linear mulations were those of minimizing the sum of
combination of financial characteristics can also absolute deviations (Freed and Glover, 1981a),
be assigned to a node). On the other hand the minimizing the maximum deviation (Freed and
method uses a sequence of nodes, i.e. a sequence Glover, 1981b) and minimizing the n u m b e r of
of characteristics to classify a firm. misclassifications (Bajgier and Hill, 1982). Some
R P A is a forward selection method. It does studies, like the one by Glover (1990), proposed
not review previous classifications while it intro- some 'hybrid' mathematical programming formu-
duces new classification rules and this can result lations. Koehler and Erenguc (1990) provided a
in a reappearing of the same variable to a later review of mathematical programming discrimina-
stage, with a different cut-off score. T h e r e is also tion methods.
a problem of overfitting, as continuation of parti- M a h m o o d and Lawrence (1987) used mathe-
tioning process can result in a tree were each matical p r o g r a m m i n g discrimination ( M P D )
firm is classified by one terminal node. To avoid methods in the bankruptcy problem. More re-
such problems different trees of various degrees cently, G u p t a et al. (1990) applied linear goal
of complexity are derived and tested. programming, as proposed by Freed and Glover
 A.I. Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513 507

(1981a,b), for the failure problem. Formulating contrast, Bajgier and Hill (1982) and Markowski
the intra-group and the inter-group differences and Markowski (1987), using MPD methods in-
between firms, this model sets a score for each cluding qualitative variables, and Rubin (1990),
firm and boundaries for group discrimination. demonstrated that for some classes of problems,
The method generates a hyperplane of the form linear programming discrimination methods per-
Ax = b. (B.12) form usually better than DA. While MPD meth-
ods are distribution free and flexible (Freed and
Given the n observations (firms) A i and the Glover, 1981a,b) this fact makes them interesting
two groups of firms G 1 and G2, this method alternatives to statistical discrimination proce-
looks for a linear transformation of x and a dures. The infrequent application of these recent
'cut-off (boundary) point b. The goal is to maxi- and promising methods in failure prediction does
mize the weighted sum of squares and minimize not allow us to fully evaluate their usefulness.
the weighted sum of boundary violations. These MPD methods were not used only for business
are expressed by the following formulation: failure prediction (see Appendix A for more stud-
ies). Srinivasan and Kim (1988) used a formula-
Min ~ Piai- ~ qidi (B.13) tion proposed by Freed and Glover (1981a) for
i-1 i=1 credit granting classifications.
subject to
Aix+di<b+a i, Ai~G 1,
B.8. Multicriteria decision aid/support methods

A i x + d i >__b + ai, A i ~ G2, (B.14) Many real-world failure problems include

non-quantitative variables, usually other than fi-
b+ Exi=M,
nancial, e.g. managerial or marketing. Such vari-
where: ables are difficult but not impossible to insert in
the last constraint is considered as a 'normalizing 'classical' discrimination methods; even though
constraint'. there are some studies on the use of qualitative
a i is the degree to which an A i violates the variables (e.g., Gilbert, 1968; Press and Wilson,
'cut-off' point, 1988; Libby, 1975a). To overcome these difficul-
d i is the measure of the distance between A i and ties, multicriteria decision aid methods were pro-
the 'cut-off' point; di, a i >__O, posed in order to rank companies according to
Pi and qi are weights for a i and d i respectively, the level of risk (CNME, 1973; Diakoulaki et al.,
and 1992; Moscarola and Roy, 1977; Zopounidis, 1987;
x and b are unrestricted in sign. Zopounidis, 1990). These methods included both
Gupta et al. (1990) used the same variables as quantitative (financial ratios) and qualitative vari-
Altman (1968) in a sample consisting of firms in ables for the assessment of a company perfor-
several industries, covering the period 1971-1986 mance.
and using data for up to five years prior to Later, the above tools were inserted in a deci-
failure. The MPD model provided better results sion support framework and the created system
than an alternative DA model. The Mahmood was called a multicriteria decision support system
and Lawrence (1987) study, however, concluded (MCDSS). The characteristic of an MCDSS is
the opposite, indicating sensitivity to the nature that it allows the decision maker to interact with
of the sample. the system at any time, setting the parameters of
The results of mathematical programming dis- the problem and data to be used. A typical
crimination methods were usually compared to MCDSS is that developed by Siskos et al. (1994).
those provided by DA using either real or simu- Basic parts of the system are the data base and
lated data. Some MPD methods were criticized the models base. Data base includes basic finan-
for providing arbitrary and not very helpful classi- cial statements (balance sheet and income state-
fication scores (Glorfeld and Gaither, 1982). In ment) and also qualitative information (informa-
508 A.I. Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513

tion security, market trend, quality of manage- ing to the ELECTRE family were popular in
ment, etc.) for firms. Model base includes finan- several European studies (cf. Roy and Bouyssou,
cial analysis methods providing common size 1993; Siskos et al., 1984) but no so much in the
statements, financial ratios and graphical repre- USA.
sentation of them, as well as methods for the Rough sets approach deals with vagueness,
assessment of firm performance (DA, Principal caused by granularity in the representation of a
Components Analysis, multicriteria decision decision situation, in sorting problems. Thus, it
methods, etc.). The user of the system can select enables one to: (a) discover minimal subsets of
the information useful to the analysis and meth- criteria ensuring an acceptable quality of classifi-
ods to be employed - often interactively. The cation of the firms, and (b) derive decision rules
multicriteria idea for predicting failure or assess- to be used to support decisions. Slowinski and
ing the performance of a firm led to the develop- Zopounidis (1994) used the rough sets approach
ment of some MCDSSs. Mareschal and Brans for the classification of 39 Greek firms in three
(1991), Massaglia and Ostanello (1991) and Zo- categories of risk (failed/uncertain/non-failed).
pounidis et al. (1992) propose such systems that The criteria found most important and employed
are able to help in financial decision-making, in the decision rules were: general and adminis-
involving financial and qualitative variables to trative expenses/sales; manager's work experi-
assess a company's performance. ence; market niche-position; special competitive
As an alternative to statistical approach, multi- advantages of firms.
criteria methods were suggested for the predic-
tion of business failure. The ELECTRE method
(Roy, 1991), and Rough Sets method (Slowinski
B.9. Expert systems
and Zopounidis, 1994) have been proposed to
overcome the limitations of DA and other statis-
tical methods. The development and application of artificial
The ELECTRE family of multicriteria meth- intelligence led some researchers to employ ex-
ods (cf. Roy, 1991) belongs to the French school, pert systems and neural networks in the
which uses the majority rule within an outranking bankruptcy problem. (For instance, neural net-
relationship, instead of the American school, works are only used for bank bankruptcy predic-
which uses the unanimity rule within the concept tion, see Dutta and Shekhar (1992), Tam (1991),
of dominance (Pareto optimality). The particular- and Tam and Kiang (1992).) Messier and Hansen
ity of the ELECTRE family (and of the French (1988) made an attempt to use expert systems
school) is to refuse the possibility of total com- (ES) for the prediction of business failure. They
pensation between the alternative's performance used a 'data-driven' method developed by Quin-
on the criteria, and then to accept incomparabil- lan (1983), which is based on the Concept Learn-
ity and intransitivity. The ELECTRE method was ing System proposed by Hunt et al. (1966). The
employed by Zollinger (1982) to assess business objective of the method was to take firms of
failure risk, using the following criteria: known classes (bankrupt/non-bankrupt) de-
X 1 = EBIT/TA. scribed by a fixed set of attributes (financial,ra-
X 2 = CF/S. tios), and then to generate a production system
X 3 = W C / N F , where NF = Needs in Financing. using the attributes which correctly classify all the
X 4 N W / ( T L + SE)
= firms of the sample. The rules at each stage (i.e.,
X 5 = SE/LTD. the variable and the cut-off score) were defined
The ELECTRE method proceeds to the sort- by using measures of entropy and selecting the
ing of firms in categories that adapts well to the minimum entropy rule. A decision tree was de-
purpose of risk failure, The ELECTRE family rived from the production system rules. The deci-
accepts incomparability and intransitivity using sion tree provided by the method is rather similar
straightforward techniques. The methods belong- to the RPA's decision tree. So, ES's decision
 A.L Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513 509

IF Current Assets~Current Liabilities < 1.533 and

IF Earnings~Total Tangible Assets < O.092 and
IF Retained Earnings/Total Tangible Assets < 0.064 and
IF Retained Earnings/Total Tangible Assets < -0.942
THEN the business will not fail
ELSE the firm will fail
ELSE the firm will not fail
ELSE the firm will not fail
ELSE the firm will not fail

Fig. B.2. The production system proposed by Handen and Messier (1988).

trees are also able to explain the prediction of References

failure or non-failure.
Messier and Hansen's study was based on a Altman, E.I. (1968), "Financial ratios, discriminant analysis
sample of 23 Australian land development firms. and the prediction of corporate bankruptcy", The Journal
The sample consisted of 8 bankrupt and 15 non- of Finance, 23, 589-609.
Altman, E.I. (1969), "Corporate bankruptcy, potential stock-
bankrupt firms. The methodology develops a pro- holder returns and share valuation", The Journal of Fi-
duction system as presented in Fig. B.2. The nance, December, 887-900.
classification accuracy of the production system Altman, E.I. (1983), Corporate Financial Distress, Wiley, New
was encouraging in this small case but, as the York.
authors mentioned, there are some limitations in Altman, E.I. (1984), "The success of business failure predic-
tion models. An international survey", Journal of Banking
the use of such inductive methods; e.g., computa- and Finance 8/2, 171-198.
tional burden when the sample is large. A conflict Altman, E.I., and Eisenbeis, R.A. (1978), "Financial applica-
arises when two firms have similar attributes but tions of discriminant analysis: A clarification", Journal of
belong to different classes. Although the authors Financial and Quantitative Analysis 13, 185-195.
proposed some possible solutions, they pointed Altman, E.I., and Lavallee, M.Y. (1981), "Business failure
classification in Canada", Journal of Business Administra-
that more research is needed in this area. tion, Summer, 147-164.
In addition to inductive learning methods, the Altman, E.I., and Spivack, J. (1983), "Predicting bankruptcy:
common knowledge engineering process of inter- The value line relative financial strength systems vs. the
viewing experienced decision makers and incor- zeta bankruptcy classification approach", Financial Ana-
poration of their knowledge in production rules lysts Journal, November-December, 60-67.
Altman, E.I., Hadelman, R.G., and Narayanan, P. (1977),
was also employed. The model quality is still "Zeta analysis, a new model to identify bankruptcy risk of
affected by the amount and reliability of informa- corporations", Journal of Banking and Finance 1, 29-51.
tion. Altman, E.I., Avery, R., Eisenbeis, R., and Stinkey, J. (1981),
Expert systems were introduced for both fail- Application of Classification Techniques m Business, Bank-
ure and credit granting problems. Elmer and ing and Finance, Contemporary Studies in Economic and
Financial Analysis, Vol. 3, JAI Press, Greenwich.
Borowski (1988) proposed an ES approach for Altman, E.I., Margaine, M., Schlosser, M., and Vernimmen,
bankruptcy at savings & loans institutions. Credit P. (1974), "Statistical credit analysis in the textile industry:
granting studies using ES were those of Bouw- A French experience", Journal of Financial and Quantita-
man (1983), Cronan et al. (1991), Duchessi and tive Analysis 9, 195-211.
Belardo (1987), Duchessi et al. (1988), Levine and Appetiti, S. (1984), "Identifying unsound finns in Italy. An
attempt to use trend variables", Journal of Banking and
Pomerol (1989), Pinson (1989,1991), Ruparel and Finance 8/2, 269-279.
Srinivasan (1992), Shaw and Gentry (1988) and Bajgier, S.M., and Hill, A.V. (1982), "An experimental com-
Srinivasan and Ruparel (1990). parison of statistical and linear programming approaches
510 A.L Dimitras et al. /European Journal of Operational Research 90 (1996) 487-513

to the discriminant problem", Decision Sciences 13, 604- Deakin, E.B. (1972), "A discriminant analysis of predictors of
617. business failure", Journal of Accounting Research, Spring,
Beaver, W.H. (1966), "Financial ratios as predictors of 167-179.
failure", in: Empirical Research in Accounting: Selected Diakoulaki, D., Mavrotas, G., and Papayannakis, L. (1992),
Studies, Supplement to Journal of Accounting Research 5, "A multicriteria approach for evaluating the performance
179-199. of industrial firms", OMEGA 20/4, 467-474.
Beaver, W.H. (1968), "Market prices, financial ratios and the Duchessi, P., and Belardo, S. (1987), "Lending analysis sup-
prediction of failure", Journal of Accounting Research, port system (LASS): An application of a knowledge-based
Autumn, 179-192. system to support commercial loan analysis", IEEE Trans-
Booth, D.E., Adam, P., Ahkam, S.N., and Osyk, B. (1989), "A actions on Systems, Man, and Cybernetics 17/4, 608-616.
robust multivariate procedure for the identification of Duchessi, P., Shawky, H., and Seagle, J.P. (1988), "A knowl-
problem Saving and Loan institutions", Decision Sciences edge-engineered system for commercial loan decisions",
20, 320-333. Financial Management 17/3, 57-65.
Booth, P., and Hutchinson, P. (1989), "Distinguishing be- Dutta, S., and Shekhar, S. (1992), "Generalization with neural
tween failing and growing firms: A note on the use of networks: An application in the financial domain", Jour-
decomposition measure analysis", Journal of Banking and nal of lnformation Science and Technology 1/4, 309-330.
Finance 16/2, 267-271. Edmister, R.A. (1972), "An empirical test of financial ratio
Bouwman, M. (1983), "Human diagnostic reasoning by com- analysis for small business failure prediction", Journal of
puter: An illustration from financial analysis", Manage- Financial and Quantitative Analysis 7, 1477-1493.
ment Science 29/6, 653-672. Eisenbeis, R.A. (1977), "Pitfalls in the application discrimi-
Breiman, L., Friedman, J.H., Olsen, R.A., and Stone, C.J. nant analysis in business and economics", The Journal of
(1984), Classification and Regression Trees, The Wadsworth Finance 32, 875-900.
Statistics/Probability Series, Wadsworth & Brooks/Cole Eisenbeis, R.A. (1978), "Problems in applying discriminant
Advanced Books & Software, Pacific Grove, CA. analysis in credit scoring models", Journal of Banking and
Bulow, J.I., and Shoven, J.B. (1978), "The bankruptcy deci- Finance 2, 205-219.
sion", The Bell Journal of Economics, Autumn, 437-456. Elam, E. (1975), "The effect of lease data on the predictive
Caevel, J. de (1983), "Un syst6me de clignotants-credits", ability of financial ratios", The Accounting Review, 15-43.
Revue Barulue, No. Special, 43-47. Elmer, P.J., and Borowski, D.M. (1988), "An expert system
Casey, M., McGee, V., and Stinkey, C. (1986), "Discriminat- approach to financial analysis: The case of S & L
ing between reorganized and liquidated firms in bankruptcy", Financial Management, Autumn, 66-76.
bankruptcy", The Accounting Review, April, 249-262. Ezzamel, M., and Molinero, C.M. (1990), "The distributional
Chalos, P. (1985), "Financial distress: A comparative study of properties of financial ratios in UK manufacturing compa-
individual, model and committee assessments", Journal of nies", Journal of Business Finance and Accounting 17/1,
Accounting Research, Autumn, 527-543. 1-29.
CNME (1973), "M6thode de d6cision multicrit~re appliq6e Falbo, P. (1991), "Credit scoring by enlarged discriminant
l'6valuation de l'entr6prise", No. 58, 1-27. analysis", OMEGA 19/4, 275-289.
Collins, R.A. (1983), "An empirical comparison of bankruptcy Fitzpatrick, P.I. (1932), "A comparison of the ratios of the
prediction models", Financial Management, Summer, 52- successful industrial enterprises with those of failed com-
57. panies", The Accountants Publishing Company.
Collins, R.A., and Green, R.D. (1982), "Statistical methods Foster, G. (1986), Financial Statements Analysis, Prentice-Hall,
for bankruptcy forecasting", Journal of Economics and London.
Business 34, 349-354. Freed, N., and Glover, F. (1981a), "Simple but powerful goal
Collongues, Y. (1977), "Ratios financiers et prevision des programming models for discriminant problems", Euro-
faillites des petites et moyennes entreprises, Revue Banque pean Journal of Operational Research 7, 44-60.
365, 963-970. Freed, N., and Glover, F. (1981b), "A linear programming
Courtis, J.K. (1978), "Modelling a financial ratios categoric approach to the discriminant problem", Decision Sciences
framework", Journal of Business Finance and Accounting 12, 68-74.
5/4, 371-386. Frydman, H., Adtman, E.I., and Kao, D.L. (1985), "Introduc-
Cox, D.R., and Oakes, D. (1984), Analysis of Survival Data, ing recursive partitioning for financial classification: The
Chapman & Hall, London. case of financial distress", The Journal of Finance XL/1,
Cronan, T.P., Glorferd, L.W., and Perry, L.G. (1991), "Pro- 269-291.
duction system development for expert systems using a Gilbert, E.S. (1968), "On discrimination using qualitative vari-
recursive partitioning induction approach: An application ables", American Statistical Association Journal, Decem-
to mortgage, commercial, and consumer lending", Deci- ber, 1399-1412.
sion Sciences 22, 812-845. Gilbert, L.R., Menon, K., and Schwartz, K.B. (1990), "Predic-
Dambolena, I.G., and Khoury, S.J. (1980), "Ratio stability and ting bankruptcy for firms in financial distress", Journal of
corporate failure", The Journal of Finance 35, 1017-1026. Business Finance and Accounting 17/1, 161-171.
 A.L Dimitras et al. /European Journal of Operational Research 90 (1996) 487-513 511

Giordano, Y. (1986), "Indicateur synth6tique de position- Keasey, K., and Watson, R. (1991), "Financial distress predic-
nement: Un nouvel outil de gestion pour la firme", Revue tion models: A review of their usefulness", British Journal
Banque 456, 879-884. of Management 2, 89-102.
Glorfeld, L.W., and Gaither, N. (1982), "On using linear Keasey, K., McGuinness, P., and Short, H. (1990), "Multilogit
programming in discriminant problems", Decision Sciences approach to predicting corporate failure - Further analy-
13, 166-171. sis and the issue of signal consistency", OMEGA 18/1,
Glover, F. (1990), "Improved linear programming models for 85 -94.
discriminant analysis", Decision Sciences 21, 771-785. Koehler, G.J., and Erenguc, S.S. (1990), "Minimizing misclas-
Gujarati, D.N. (1988), Basic Econometrics, 2nd ed., McGraw- sifications in linear discriminant analysis", Decision Sci-
Hill, Singapore. ences 21, 63-85.
Gupta, Y.P., Rao, R.P., and Bagghi, P.K. (1990), "Linear goal Laitinen, E.K. (1991), "Financial ratios and different failure
programming as an alternative to multivariate discriminant processes", Journal of Business Finance and Accounting
analysis: A Note", Journal of Business Finance and Ac- 18/5, 649-673.
counting, 17/4, 593-598. Laitinen, E.K. (1992), "Prediction of failure of a newly
Gloubos, G., and Grammatikos, T. (1988), "The success of founded firm", Journal of Business Venturing 7, 323-340.
bankruptcy prediction models in Greece", Studies in Bank- Laitinen, E.K. (1993), "Financial predictors for different
ing and Finance 7, 37-46. phases of the failure process", OMEGA 21/2, 215-228.
Grablowsky, B.J., and Talley, W.K. (1981), "Probit and dis- Lane, S. (1972), "Submarginal credit risk classification", Jour-
criminant factors for classifying credit applicants: A com- nal of Financial and Quantitative Analysis 7, 1379-1386.
parison", Journal of Economics and Business 33, 254-261. Lane, W.R., Looney, S.W., and Wansley, W. (1986), "An
Grammatikos, T., and Gloubos, G. (1984), "Predicting application of the Cox proportional hazards model to bank
bankruptcy of industrial firms in Greece", Spoudai, The failure", Journal of Banking and Finance 10, 511-531.
University of Piraeus Journal of Economics, Business, Statis- Lau, A.H.L. (1987), "A five-state financial distress prediction
tics and Operations Research 3-4, 421-443. model", Journal of Accounting Research 25/1, 127-138.
Gupta, M.C., and Huefner, R.J. (1972), "A cluster analysis Levine, P., and Pomerol, J.-Ch. (1989), "Knowledge represen-
study of financial ratios and industry characteristics", tation by schemata in financial expert systems", Theory
Journal of Accounting Research, Spring, 77-95. and Decision 27, 147-161.
Hamer, M.M. (1983), "Failure prediction: Sensitivity of classi- Libby, R. (1975a), "Accounting ratios and the prediction of
fication accuracy to alternative statistical methods and failure: Some behavioral evidence", Journal of Accounting
variable sets", Journal of Accounting and Public Policy 2, Research 13/1, 150-161.
289-307. Libby, R. (1975b), "The use of simulated decision makers in
Houghton, K.A. (1984), "Accounting data and the prediction information evaluation", The Accounting Review, July,
of business failure: The setting of priors and the age of 475-489.
data", Journal of Accounting Research 22/1, 361-368. Luoma, M., and Laitinen, E.K. (1991), "Survival analysis as a
Houghton, K.A., and Senagupta, R. (1984), "The effect of tool for company failure prediction", OMEGA 19/6, 673-
prior probability disclosure and information set construc- 678.
tion on bankers' ability to predict failure", Journal of Mahmood, M.A., and Lawrence, E.C. (1987), "A performance
Accounting Research 22/2, 768-775. analysis of parametric and nonparametric discriminant
Hunt, E.B., Marin, J., and Stone, P.T. (1966), Experiments in approaches to business decision making", Decision Sci-
Induction, Academic Press, New York. ences 18, 308-326.
Izan, H.Y. (1984), "Corporate distress in Australia", Journal Malecot, J.F. (1981), "Limites des modules de pr6vision de
of Banking and Finance 8, 303-320. d6faillances", Finance 2/4, 291-315.
Jensen, R.E. (1971), "A cluster analysis study of financial Malecot, J.F. (1986), "Sait-on vraiment pr6voir les d6fail-
performance of selected business firms", The Accounting lances d'entreprises?", ISMEA, Revue Sciences de Gestion,
Review XLVI, 36-56. Decembre, 55-82.
Jones, F.L. (1987), "Current techniques in bankruptcy predic- Malecot, J.F. (1991), "Analyses th6oriques des d6faillances
tion", Journal of Accounting Literature 6, 131-164. d'entreprises: Une revue de la litt6rature", Cahier de
Joy, O.M., and Tollefson, J.O. (1975), "On the financial recherche No. 9108, CEREG, Universite de Paris-
applications of discriminant analysis", Journal of Financial Dauphine.
and Quantitative Analysis 101, 723-739. Marais, M.L., Patell, J.M., and Wolfson, M.A. (1984), "The
Karels, G.V., and Prakash, A. (1987), "Multivariate normality experimental design of classification models: An applica-
and forecasting of business bankruptcy", Journal of Busi- tion of recursive partitioning and bootstrapping to com-
ness Finance and Accounting 14, 573-593. mercial bank loan classifications", Journal of Accounting
Keasey, K., and McGuinness, P. (1990), "The failure of UK Research 22, 87-113.
industrial firms for the period 1976-1984: Logistic analysis Mareschal, B., and Brans, J.P. (1991), "BANKADVISER: An
and entropy measures", Journal of Business Finance and industrial evaluation system", European Journal of Opera-
Accounting 17/1, 119-135. tional Research 54, 318-324.
512 A.I. Dimitras et al. /European Journal of Operational Research 90 (1996) 487-513

Markowski, C.A., and Markowski, E.P. (1987), "An experi- evidence on predicting private company failure", Account-
mental comparison of several approaches to the discrimi- ing and Business Research 18/69, 57-66.
nation problem with both qualitative and quantitative vari- Peel, M.J., Peel, D.A., and Pope, P.F. (1986), "Predicting
ables", European Journal of Operational Research 28, 74- corporate failure - Some results for the UK corporate
78. sector", OMEGA 14/1, 5-12,
Mar Molinero, C., and Ezzamel, M. (1991), "Multidimen- Pinson, S. (1989), "Credit risk assessment and meta-judge-
sional scaling applied to corporate failure", OMEGA 19/4, ment", Theory and Decision 27, 117-133.
259-274. Pinson, S. (1991), "A multi-expert architecture for credit risk
Martikainen, T., and Ankelo, T. (1991), "On the instability of assessment: The CREDEX system", in: D.E. O'Leary and
financial patterns of failed firms and the predictability of P.R. Watkins (eds.), Expert Systems in Finance, Elsevier
corporate failure", Economic Letters 35, 209-214. Science Publishers, Amsterdam, 37-64.
Martin, D. (1977), "Early warning of bank failure. A logit Platt, H.P., and Platt, M.B. (1990), "Development of a class of
regression approach", Journal of Banking and Finance 1, stable predictive variables: The case of bankruptcy predic-
249-276. tion", Journal of Business Finance and Accounting 17/1,
Massaglia, M., and Ostanello, A. (1991), "N-TOMIC: A deci- 31-51.
sion support for multicriteria segmentation problems", in: Press, S.J., and Wilson, S. (1978), "Choosing between logistic
P. Korhonen (ed.), International Workshop on Multicriteria regression and discriminant analysis", Journal of the Amer-
Decision Support, Helsinki 7-11, 1989, Lecture Notes in ican Statistical Association 73/364, 699-705.
Economics and Mathematical Systems 356, Springer- Quinlan, J.R. (1983), "Learning efficient classification proce-
Verlag, Berlin, 167-174. dures and their application to chess and games", in: R.S.
Mensah, Y. (1984), "An examination of the stationarity of Michalski, J.G. Carbonell and T.M. Mitchell (eds.), Ma-
multivariate bankruptcy prediction models: A methodolog- chine Learning: An Artificial Intelligence Approach, Tioga,
ical study", Journal of Accounting Research 22/1, 380-395. Palo Alto, CA.
Merwin, C.L. (1942), "Financing small corporations: In five Rose, P.S., Andrews, W.T. and Giroux, G.A. (1982), "Predic-
manufacturing industries, 1926-1936", National Bureau of ting business failure: A macroeconomic perspective",
Economic Research. Journal of Accounting and Finance 6/1, 20-31.
Messier, W.F., and Hansen, J.V. (1988), "Including rules for Roy, B. (1991), "The outranking approach and the founda-
expert system development: An example using default and tions of ELECTRE methods", Theory and Decision 31,
bankruptcy data", Management Science 34/12, 1403-1415. 49-73.
Meyer, P.A., and Pifer, H.W. (1970), "Prediction of bank Roy, B., and Bouyssou, D. (1993), Aide Multicrit~re h la
failures", Journal of Finance 25, 853-868. DEcision: M~thodes et Cas, Economica, Paris.
Micha, B. (1984), "Analysis of business failures in France", Rubin, P.A. (1990), "A comparison of linear and parametric
Journal of Banking and Finance 8, 281-291. approaches to the two-group discriminant problem", Deci-
Michalopoulos, M., Zopounidis, C., and Kouris, I. (1993), "An sion Sciences 21, 373-386.
integrated DSS for the assessment of the risk of failure", Ruparel, B., and Srinivasan, V. (1992), "A dedicated shell for
in: Applied Stochastic Models and Data Analysis, World designing expert credit support systems", Decision Support
Scientific Publishing, New Jersey, 606-625. Systems 8, 343-359.
Moscarola, J., and Roy, B. (1977), "Procddure automatique Salchenberger, L.M., Cinar, E.M., and Lash, N.A. (1992),
d'examen de dossiers fond6e sur une segmentation tri- "Neural networks: A new tool for predicting thrift fail-
chotomique en prdsence de crit~res multiples", RAIRO ures", Decision Sciences 23, 899-916.
Recherche OpHationnelle 11/2, 145-173. Santomero, A.M., and Vinso, J.D. (1977), "Estimating the
Moses, O.D. (1990), "On analysts' earnings forecasts for fail- probability of failure for commercial banks and banking
ing firms", Journal of Business Finance and Accounting systems", Journal of Banking and Finance 1, 185-205.
17/1, 101-118. Scapens R.W., Ryan, R.J., and Flecher, L. (1981), "Explaining
Moyer, R.C. (1977), "Forecasting financial failure: A re-ex- corporate failure: A catastrophe theory approach", Jour-
amination", Financial Management, Spring, 11-17. nal of Business Finance and Accounting 8/1, 1-26.
Ohlson, J.A. (1980), "Financial ratios and the probabilistic Scott, E. (1978), "On the financial applications of discrimi-
prediction of bankruptcy", Journal of Accounting Research, nant analysis: Comment", Journal of Financial and Quanti-
Spring, 109-131. tative Analysis 13, 201-205.
O'Leary, D.E. (1993), "On bankruptcy information systems", Scott, J. (1981), "The probability of bankruptcy: A compari-
European Journal of Operational Research 56, 67-79. son of empirical predictions and theoretical models",
O'Neill, H.M., and Duker, J. (1986), "Survival and failure in Journal of Banking and Finance 5, 317-344,
small business", Journal of Small Business Management, Sharma, S., and Mahajan, V. (1980), "Early warning indica-
30-37. tors of business failure", Journal of Marketing 44, 80-89.
Peel, M.J. (1987), "Timeliness of private company reports Shaw, M.J., and Gentry, J.A. (1988), "Using an expert system
predicting corporate failure", Investment Analysis 83, 23- with inductive learning to evaluate business loans", Finan-
27. cial Management 17/3, 45-56.
Peel, M.J., and Peel, D.A. (1987), "Some further empirical Siskos, Y, Washer, G., and Winkels, H.M. (1984), "Outrank-
 A.L Dimitras et al. / European Journal of Operational Research 90 (1996) 487-513 513

ing approaches versus MAUT in MCDM", European Jour- model to analyse corporate behavior in Israel", Journal of
nal of Operational Research 16, 270-271. Banking and Finance 8, 293-302.
Siskos, Y., Zopounidis, C., and Pouliezos, A. (1994), "An Tennyson, B.N., Ingram, R.W., and Dugan, M.T. (1990),
integrated DSS for financing firms by an industrial devel- "Assessing the information content of narrative disclo-
opment bank in Greece", Decision Support Systems 12, sures in explaining bankruptcy", Journal of Business Fi-
151-168. nance and Accounting 17/3, 391-410.
Skogsvik, K. (1990), "Current cost accounting ratios as predic- Theodossiou, P. (1991), "Alternative models for assessing the
tors of business failure: The Swedish case", Journal of financial condition of business in Greece", Journal of
Business Finance and Accounting 17/1, 137-160. Business Finance and Accounting 18/5, 697-720.
Slowinski, R., and Zopounidis, C. (1994), "Rough-set sorting Vernimmen, P. (1978), "Panorama des recherches portant sur
of firms according to bankruptcy risk", in: M. Paruccini le risque du creancier", Analyse Financidre 1, 54-61.
(ed), Applying Multiple Criteria Aid for Decision to Envi- Vranas, A.S. (1991), "Probability models for the forecasting of
ronmental Management, Kluwer Academic Publishers, Greek industrial firms' failure" (in Greek), Spoudai, The
Dordrecht, 339-357. University of Piraeus Journal of Economics, Business, Statis-
Smith, R.F., and Winakor, A.H. (1935), "Changes in the tics and Operations Research 4/4, 431-448.
financial structure of unsuccessful corporations", Bureau Vranas, A.S. (1992), "The significance of financial character-
of Business Research, University of Illinois, Urbana, IL. istics in predicting business failure: An analysis in the
Srinivasan, V., and Kim, Y.H. (1987), "Credit granting: A Greek context", Foundations of Computing and Decision
comparative analysis of classification procedures", The Sciences 17/4, 257-275.
Journal of Finance XLII/3, 665-683. Warner, J.B. (1977), "Bankruptcy costs: Some evidence", The
Srinivasan, V., and Kim, Y.H. (1988), "Designing expert fi- Journal of Finance 32/2, 337-347.
nancial systems: A case study of corporate credit manage- Wertheim, P., and Lynn, M.L. (1993), "Development of a
ment", Financial Management, Autumn, 32-44. prediction model for hospital closure using financial ac-
Srinivasan, V., and Ruparel, B. (1990), "CGX: An expert counting data", Decision Sciences 24/3, 529-546.
support system for credit granting", European Journal of Wilcox, J.W. (1971), "A simple theory of financial ratios as
Operational Research 45, 293-308. predictors of failure", Journal of Accounting Research,
Stein, J.H. von, and Ziegler, W. (1984), "The prognosis and Autumn, 389-395.
surveillance of risks from commercial credit borrowers", Zanakis, S.H., and Walter, G.A. (1994), "Discriminant char-
The Journal of Finance 8, 249-268. acteristics of US banks acquired with or without federal
Suzuki, S., and Wright, R.W. (1985), "Financial structure and assistance", European Journal of Operational Research 77,
bankruptcy risk in Japanese companies", Journal of Inter- 440-465.
national Business Studies, Spring, 97-110. Zavgren, C.V. (1983), "The prediction of corporate failure:
Taffler, R.J. (1982), "Forecasting company failure in the UK The state of the art", Journal of Financial Literature 2,
using discriminant analysis and financial ratio data", Jour- 1-37.
nal of the Royal Statistical Society. Series. A. General 145/3, Zavgren, C.V. (1985), "Assessing the vulnerability to failure of
342-358. American industrial firms. A logistic analysis", Journal of
Tarrier, R.J. (1983), "The z-score approach to measuring Business Finance and Accounting 12/1, 19-45.
company solvency", The Accountant's Magazine 87/921, Zmijewski, M.E. (1984), "Methodological issues related to the
91-96. estimation of financial distress prediction models", Studies
Taffler, R.J. (1984), "Empirical models for the monitoring of on Current Econometric Issues in Accounting Research,
UK corporations", Journal of Banking and Finance 2, 59-82.
199-227. Zollinger, M. (1982), "L'analyse multicrit~re et le risque de
Takahashi, K., Kurokawa, Y., and Watase, K. (1984), "Corpo- cr6dit aux entreprises", Revue Fran~aise de Gestion, 56-66.
rate bankruptcy prediction in Japan", Journal of Banking Zopounidis, C. (1987), "A multicriteria decision making
and Finance 8, 229-247. methodology for the evaluation of the risk of failure and
Tam, K.Y. (1991), "Neural network models and the prediction an application", Foundations of Control Engineering 12/1,
of bank bankruptcy", OMEGA 19/5, 429-445. 45-67.
Tam, K.Y., and Kiang, M.Y. (1992), "Managerial applications Zopounidis, C. (1990), La Gestion du Capital Risque, Econom-
of neural networks: The case of bank failure predictions", ica, Paris.
Management Science 38/7, 926-947. Zopounidis, C., Pouliezos, A., and Yannacopoulos, D. (1992),
Tamari, M. (1964), "Financial ratios as a means of forecasting "Designing a DSS for the assessment of company perfor-
bankruptcy", Economic Review, Bank of Israel, Jerusalem. mance and viability", Computer Science in Economics and
Tamari, M. (1984), "The use of a bankruptcy forecasting Management 5, 41-56.