35 Years of Studies On Business Failure - An Overview of The Classic Statistical Methodologies and Their Related Problems

The British Accounting Review 38 (2006) 63–93
www.elsevier.com/locate/bar
35 years of studies on business failure: an overview

of the classic statistical methodologies and their
related problems
Sofie Balcaena,*, Hubert Oogheb
a
Ghent University, Kuiperskaai 55 E, B-9000 Gent, Belgium
b
Vlerick Leuven Gent Management School, Reep 1, B-9000 Gent, Belgium
Received 19 November 2004; revised 6 September 2005; accepted 7 September 2005
Abstract
Over the last 35 years, business failure prediction has become a major research domain
within corporate finance. Numerous corporate failure prediction models have been developed,
based on various modelling techniques. The most popular are the classic cross-sectional
statistical methods, which have resulted in various ‘single-period’ or static models, especially
multivariate discriminant models and logit models. To date, there has been no clear overview
and discussion of the application of classic statistical methods to business failure prediction.
Therefore, this paper extensively elaborates on the application of (1) univariate analysis, (2)
risk index models, (3) multivariate discriminant analysis, and (4) conditional probability
models in corporate failure prediction. In addition, because there is no clear and
comprehensive analysis in the existing literature of the diverse problems related to the
application of these methods to the topic of corporate failure prediction, this paper brings
together all problem issues and enlarges upon each of them. It discusses all problems related
to: (1) the classical paradigm (i.e. the arbitrary definition of failure, non-stationarity and data
instability, sampling selectivity, and the choice of the optimisation criteria); (2) the neglect
of the time dimension of failure; and (3) the application focus in failure prediction
modelling. Further, the paper elaborates on a number of other problems related to the use of
a linear classification rule, the use of annual account information, and neglect of the
multidimensional nature of failure. This paper contributes towards a thorough understanding
* Corresponding author. Tel.: C32 9 264 35 07.

E-mail address: sofie.balcaen@ugent.be (S. Balcaen).
0890-8389/$ - see front matter q 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.bar.2005.09.001
64 S. Balcaen, H. Ooghe / The British Accounting Review 38 (2006) 63–93
of the features of the classic statistical business failure prediction models and their related
problems.
q 2005 Elsevier Ltd. All rights reserved.
Keywords: Business failure; Failure prediction models; Overview; Problem topics
1. Introduction
Over the last 35 years, the topic of business failure prediction has developed into a
major research domain within corporate finance. Many academic studies have been
dedicated to finding the best corporate failure prediction model. With a view to predicting
the risk of business failure and to accurately classifying firms according to their financial
health, academic researchers from throughout the world have used various modelling
techniques, each having distinct assumptions and specific computational complexities.
The most popular methods are the classic cross-sectional statistical methods, which have
resulted in numerous static failure prediction models.
In 1966, Beaver (1967a) pioneered a corporate failure prediction model with
financial ratios. He developed a univariate model—a univariate discriminant analysis
model—using a number of financial ratios, selected by a dichotomous classification
test. In response to Beaver, Tamari (1966); Moses and Liao (1987) used risk index
models to predict failure, which are simple and intuitive point systems, based on
different ratios. Altman (1968) introduced a statistical multivariate analysis technique
called multiple discriminant analysis (MDA) to the problem of company failure
prediction and estimated a Z-score model. MDA is ‘a statistical technique used to
classify an observation into one of several a priori groups dependent upon the
observation’s individual characteristics. [it] attempts to derive a linear [or quadratic]
combination of these characteristics which ‘best’ discriminates between the groups’
(Altman, 1968, p. 592). Over the years, there have been an enormous number of
studies based on Altman’s Z-score model.Altman et al. (1977) adjusted the original
Z-score model into a new, better performing Zeta analysis model. Until the 1980s, the
MDA technique dominated the literature on business failure prediction. The majority
of MDA studies used a linear MDA model, but quadratic MDA has also been applied.
Since the 1980s, the use of MDA has decreased (Dimitras et al., 1996), but it remains
a generally accepted standard method and it is frequently used as a baseline method
for comparative studies (Altman and Narayanan, 1997). MDA has been replaced by
less demanding statistical techniques such as logit analysis (LA), probit analysis (PA)
and linear probability modelling (LPM). These methods resulted in conditional
probability models (Zavgren, 1983; Zavgren, 1985; Doumpos and Zopoudinis, 1999),
consisting of a combination of variables that best distinguish between failing and non-
failing firms. Ohlson (1980) pioneered using LA in company failure prediction,
whereas Zmijewski (1984) was the pioneer in applying PA. Until now, LA has been a
very popular method in business failure prediction. The number of studies using PA is
much smaller, probably because this technique requires more computations (Gloubos
and Grammatikos, 1988; Dimitras et al., 1996).
S. Balcaen, H. Ooghe / The British Accounting Review 38 (2006) 63–93 65
To date, the topic of business failure prediction has lacked a clear overview and
discussion of the classic cross-sectional statistical methods. Therefore, a first aim of this
paper is to elaborate on the application of these methods in corporate failure prediction.
We discuss the main features and assumptions of the various classic statistical methods
and provide an overview of a large number of academically developed corporate failure
prediction models based on these methods.
Despite the popularity of the classic statistical methods, significant problems relating to
the application of these methods to corporate failure prediction remain. In the existing
literature, there is no clear and comprehensive analysis of these diverse problems, so this
paper aims to bring together all the problems and issues and to enlarge upon each of them.
First, it discusses all problems related to the ‘classical paradigm’, in particular, the
arbitrary definition of failure, non-stationarity and data instability, sampling selectivity,
and the choice of the optimization criteria. Furthermore, it enlarges upon the problems
related to the neglect of the time dimension of failure and the application focus in failure
prediction modelling. Finally, it elaborates on a number of problems related to the use of a
linear classification rule, the use of annual account information, and neglect of the
multidimensional nature of failure.
The remainder of this paper is structured as follows. Section 2 gives an overview and
discussion of the application of the classic cross-sectional statistical techniques in
corporate failure prediction. Section 3 discusses the various problems related to the
application of these modelling methods to business failure prediction.
2. Classic statistical failure prediction models
The following classic cross-sectional statistical methods have been widely used in the
development of corporate failure prediction models: univariate models, risk index models,
MDA models, and conditional probability models, such as logit, probit and linear
probability models (Zavgren, 1983;Van Wymeersch and Wolfs, 1996; Atiya, 2001). MDA
is by far the most dominant classic statistical method, followed by logit analysis (Altman
and Saunders, 1998). The classic statistical failure prediction models involve a
classification procedure to classify firms as failing or non-failing. Two types of
misclassifications can be made: a type I error is made when a failing firm is misclassified
as a non-failing firm, whereas a type II error is made when a non-failing firm is wrongly
assigned to the failing group.
2.1. Univariate analysis
In a univariate failure prediction model, an optimal cut-off point is estimated for each
measure or ratio in the model and a classification procedure is carried out separately for
each measure, based on a firm’s value for the measure and the corresponding optimal cut-
off point. The univariate modelling technique is extremely simple and the application does
not require any statistical knowledge. On the other hand, univariate analysis is based on the
stringent assumption of a linear relationship between all measures and the failure status.
2.2. Risk index models
The risk index model of Tamari (1966) is a simple and intuitive point system, which
includes various ratios. A firm is attributed a certain number of points between 0 and 100
according to the values of the ratios involved in the model, so that higher total points
indicate a better financial situation. Points are allocated so that the most important ratios
have higher weights (i.e. a higher maximum of points). However, the allocation of weights
is subjective.
Moses and Liao (1987) presented another interesting risk index model. First, optimal
cut-off points are determined for each of the composing ratios based on a univariate
analysis. Next, a dichotomous variable is created for each of the ratios and assigned a score
of one in the case where a firm’s ratio value exceeds the optimal cut-off point. Then, the
risk index simply adds the values of the dichotomous variables, so that a high score is
associated with a financially healthy situation.
2.3. Multiple discriminant analysis
An MDA model consists of a linear combination of variables, which provides the best
distinction between the failing and non-failing firms. As linear MDA is by far the most
popular MDA method, we will not further elaborate on the quadratic MDA method. The
discriminant function for a linear MDA model is as follows (Lachenbruch, 1975): DiZ
d0Cd1Xi1Cd2Xi2C.CdnXin, where Di is the discriminant score for firm i; Xij is the value
of attribute Xj (with jZ1,.,n) for firm i; d0 is the intercept; and dj is the linear
discriminant coefficient for attribute j. Several firm characteristics or attributes are
combined into one single multivariate discriminant score,Di. Di has a value between KN
and CN and gives an indication of a firm’s financial health. In most studies, a low
discriminant score indicates poor financial health1. Although MDA is called a ‘continuous
scoring’ system, we have to bear in mind that a discriminant score is simply an ordinal
measure that allows the ranking of firms. In addition, it should be noted that it is possible
that variables that seem insignificant on a univariate basis actually supply significant
information in the multivariate context of the MDA model (Altman, 1968) or that some
coefficients have unexpected, counter-intuitive signs (Ooghe and Verbaere, 1985).
Furthermore, it should be stressed that the coefficients of the MDA model do not indicate
the relative importance of the composing variables because they cannot be interpreted like
the b-coefficients of a regression2 (Altman, 1968; Blum, 1974; Joy and Tollefson, 1975;
Eisenbeis, 1977; Taffler, 1983). In a classification context, firms are classified as failing or
non-failing based on their discriminant score and the optimal cut-off point of the MDA
model: if their discriminant score, Di, is less than the cut-off point, they are classified as
1
In some studies—for example, in Ooghe et al. (1994a)—the MDA model is defined in the opposite direction.
A high discriminant score, Di, then indicates poor financial health and, hence, the score is seen as a risk measure.
2
In contrast, Scott (1978); Blum (1974); Eisenbeis (1977); Joy and Tollefson (1975) argued that the
standardized coefficients can be used to evaluate the importance of the individual variables. Nevertheless, we
need to recognize that the attempt to assess the role of the individual coefficients is inappropriate in view of the
purpose of MDA (Zavgren, 1985).
failing, whereas if their score exceeds or equals the cut-off point, they are classified as non-
failing. In this way, firms are assigned to the group they most closely resemble. In the strict
sense, an MDA-based classification statement cannot be considered a prediction.
However, in practice, when a firm is classified as failing because it most resembles
a group of firms failing in the next year, this classification is treated as a prediction that ‘the
firm will fail in year tC1’ (Blum, 1974).
The MDA technique starts with several assumptions (Edmister, 1972; Eisenbeis, 1977;
Zavgren, 1983; Karels and Prakash, 1987). First, MDA assumes that that the dataset is
dichotomous; that is, groups are discrete, non-overlapping, and identifiable. Second, the
use of MDA is based on three restrictive assumptions: (1) multivariate normally
distributed independent variables; (2) equal variance-covariance matrices across the
failing and non-failing group; and (3) specified prior probability of failure and
misclassification costs. Although certain studies have stressed the importance of the
first two restrictive assumptions, most MDA failure prediction studies do not check
whether the data satisfy the assumptions. As a result, the MDA modelling technique is
often applied in an inappropriate way and the resulting MDA models are not suited for
generalization (Joy and Tollefson, 1975; Eisenbeis, 1977; Richardson and Davidson,
1984; Zavgren, 1985).
First, the assumption of multivariate normality is often violated (Deakin, 1976; Taffler
and Tisshaw, 1977; Barnes, 1987), resulting in biased significance tests and error rates
(Eisenbeis, 1977; Richardson and Davidson, 1984; McLeay and Omar, 2000). Some
researchers, having tested for univariate normality and finding non-normality, then try to
approximate univariate normality by transforming the variables or by trimming the
outliers3. However, despite their efforts to approximate univariate normality, they ignore
the following facts: (1) univariate normality is not a sufficient condition for multivariate
normality; (2) transformation may change the interrelations among the variables
(Eisenbeis, 1977; Ezzamel and Mar-Molinero, 1990), thus distorting the MDA model;
and (3) outlier trimming may cause a significant loss of information (Ezzamel and
Mar-Molinero, 1990).
Second, the data rarely satisfy the assumption of equal dispersion matrices, which
results in biased significance tests. In case of unequal dispersion matrices, a quadratic
MDA model needs to be used (Joy and Tollefson, 1975; Eisenbeis, 1977; Zavgren, 1983).
However, in practice, researchers avoid working with quadratic MDA models because
they are very complex and seem to outperform linear MDA models only in cases where
there are large samples, a small number of independent variables, and very substantial
3
There is ample evidence that financial ratio variables, which are prevalent in MDA models, generally exhibit
non-normal distributions (Barnes, 1982; Ooghe and Verbaere, 1985; Ezzamel and Mar-Molinero, 1990; McLeay
and Omar, 2000). Univariate normality is often approximated by a variable transformation: a reciprocal or
logarithmic transformation (Taffler, 1983), a log transformation (Altman et al., 1977), or a square root or log-
normal transformation (Deakin, 1976). Another way to deal with non-normality is trimming the outliers.
Trimming is done by outlier deletion or by ‘windsorizing’, which involves changing an outlier’s value into the
value of the closest non-outlier (Taffler, 1983; Barnes, 1987; Ezzamel and Mar-Molinero, 1990; Ooghe et al.,
1995; McLeay and Omar, 2000).
differences in the dispersion matrices. Instead, researchers transform the data so as to

approximate equal dispersion matrices and then apply linear MDA (Taffler, 1982).
Third, ignoring prior probabilities of failure and the costs of types I and II errors in the
definition of the optimal cut-off score of the MDA model may result in a misleading
estimate of model accuracy (Edmister, 1972; Eisenbeis, 1977; Deakin, 1977; Zavgren,
1983; Hsieh, 1993; Steele, 1995). Most researchers neglect specifying prior probabilities
and error costs and simply minimize the total error rate4 because, in practice, the
specification of error costs and population proportions seems to be difficult and very
subjective5 (Steele, 1995). Therefore, researchers implicitly assume equal misclassifi-
cation costs and sample proportions equal to the population proportions. However,
neglecting misclassification costs generally leads to relatively high type I errors and
relatively low type II errors, whereas neglecting population frequencies generally implies
an excessively strong focus on reducing type I errors and results in relatively high type II
errors. Nevertheless, it is possible to overcome the specification of prior probabilities and
error costs by reporting the model’s classification results for different cut-off values
(Ooghe and Verbaere, 1985; Pompe and Bilderbeek, 2000) or by reporting a lower and a
higher cut-off score, which is called the ‘black-grey-white method’ (Edmister, 1972).
A final assumption of MDA is the absence of multicollinearity. Although Eisenbeis
(1977); Altman and Eisenbeis (1978) pointed out that multicollinearity is irrelevant in
MDA models, most authors agree that severe correlation among the independent variables
may cause unstable and difficult-to-explain parameter estimates and a misleading model
accuracy (Edmister, 1972; Joy and Tollefson, 1975; Joy and Tollefson, 1978; Ooghe et al.,
1994a; Back et al., 1996; Doumpos and Zopoudinis, 1999).
2.4. Conditional probability models
Conditional probability models allow the use of the non-linear maximum likelihood
method to estimate the probability of failure conditional on a range of firm characteristics.
These models are based on a certain assumption concerning the probability distribution.
Whereas logit models assume a logistic distribution (Maddala, 1977; Hosmer and
Lemeshow, 1989) and probit models assume a cumulative normal distribution (Theil,
1971), in linear probability models, the relationship between the variables and the failure
probability is assumed to be linear (Altman et al., 1981). As LA is the most popular
conditional probability method in business failure prediction, we will not elaborate further
on probit and linear probability models.
In LA, a non-linear maximum likelihood estimation procedure is used to obtain the
parameter estimates of the following logit model (based on Hosmer and Lemeshow, 1989,
p. 25 and Gujarati, 2003, p. 595–615): P1(Xi)Z1/[1CexpK(boCb1Xi1Cb2Xi2C.C
bnXin)]Z1/[1CexpK(Di)], where P1(Xi) is the probability of failure given the vector of
4
Altman et al. (1977); Taffler (1982) are two of a small number of studies that do consider the ratio of both
error costs.
5
For example, when specifying population proportions, fluctuating failure rates require the determination of a
certain reference period or the calculation of an average population proportion (Eisenbeis, 1977). This problem
becomes even greater in the case of pooled samples of observations (from different time periods).
attributes Xi; Xij is the value of attribute j (with jZ1,.,n) for firm i; bj is the coefficient for
attribute j; b0 is the intercept; and Di is the ‘logit’ of firm i. The LA model combines
several firm characteristics or attributes into a multivariate probability score, which
indicates the firm’s failure probability or vulnerability to failure. It allows for categorical
qualitative variables (Ohlson, 1980; Keasey and Watson, 1987). The logistic function
implies that the logit score P1 has a value in the [0,1] interval and is increasing in Di. When
the failed status is coded as one (zero), a high (low) logit score indicates a high failure
probability and, hence, poor financial health. The failure probability P1 follows the logistic
distribution. The underlying logistic function of the LA model implies that an extremely
healthy (weak) company, as compared to a firm that has an average financial health, must
experience a proportionally larger deterioration (amelioration) in its variables in order to
deteriorate (ameliorate) its financial health score (Laitinen and Kankaanpää, 1999). The
estimated coefficients bj can be interpreted separately as representing the importance or
significance of each of the independent variables in the explanation of the estimated failure
probability (Ohlson, 1980; Mensah, 1984; Zavgren, 1985), provided that there is no
multicollinearity among the variables. In a classification context, the essence of the LA
model is that it assigns firms to the failing or the non-failing group based on their logit
score and a certain cut-off score for the model. In the case where a high logit score
indicates a high failure probability, a firm is classified into the failing group if its logit
score exceeds the cut-off point and into the non-failing group if its score is lower than or
equal to the cut-off point. Similarly to MDA, the LA model is based on the resemblance
principle: firms are assigned to the group they most closely resemble.
When applying LA, no assumptions are made regarding prior probabilities of
failure or the distribution of the independent variables—LA does not require
multivariate normal distributed variables or equal dispersion matrices (Ohlson, 1980;
Zavgren, 1983). Therefore, the LA method is commonly considered as less demanding
than MDA.
Nevertheless, LA has two basic assumptions. First, the LA method assumes the
dependent variable to be dichotomous, with the groups being discrete, non-
overlapping, and identifiable. Second, the cost of type I and type II error rates
should be considered when defining the optimal cut-off score of the logit model.
However, due to the subjectivity of these misclassification costs (Steele, 1995), in
practice, most researchers minimize the total error rate and, hence, implicitly assume
equal misclassification costs (Zavgren, 1985; Koh, 1992; Hsieh, 1993). Ohlson (1980);
Ooghe et al. (1993); Ooghe et al. (1994a) are a few of the rare studies that explicitly
acknowledge the impact of the choice of the error costs on the corresponding error
rates. They reported the performance results of their LA model for various cut-off
points associated with various error costs. In contrast, Koh (1992) argued that
ignorance of error costs is not a serious problem and that the choice of the optimal
cut-off point is robust to different misclassification costs.
Furthermore, it should be stressed that LA models are extremely sensitive to
multicollinearity (Ooghe et al., 1993; Ooghe et al., 1994a; Doumpos and Zopoudinis,
1999), as well as to outliers and missing values (Joos et al., 1998b). The multicollinearity
problem in LA models is often severe (Tucker, 1996) because most LA models are based
on financial ratios, which in se are highly correlated because they frequently share the
same numerator or denominator.
Finally, although logit models do not require the variables to be normally distributed,
there is evidence that they do remain sensitive to extreme non-normality (McLeay and
Omar, 2000).
Table 1
Overview of classic statistical business failure prediction models
Method Failure prediction models

Univariate analysis Beaver (1967a)
Risk index models Tamari (1966)
Moses and Liao (1987)
MDA models Altman (1968)
Deakin (1972)
Edmister (1972)
Blum (1974)
Altman et al. (1977)
Deakin (1977)
Taffler and Tisshaw (1977)
van Frederikslust (1978)
Bilderbeek (1979)
Dambolena and Khoury (1980)
Taffler (1982), model from 1974
Ooghe and Verbaere (1985)
Taffler (1983)
Micha (1984)
Betts and Belhoul (1987)
Gombola et al. (1987)
Gloubos and Grammatikos (1988)
Declerc et al. (1991)
Laitinen (1992)
Lussier and Corman (1994)
Altman et al. (1995)
Conditional probability models Ohlson (1980)
Swanson and Tybout (1988)
Zavgren (1983)
Zmijewski (1984)
Gentry et al. (1985a)
Zavgren (1985)
Keasey and Watson (1987)
Peel and Peel (1987)
Aziz et al. (1988)
Gloubos and Grammatikos (1988)
Keasey and McGuinness (1990)
Platt and Platt (1990)
Ooghe et al. (1993)
Sheppard (1994)
Lussier (1995)
Mossman et al. (1998)
Charitou and Trigeorgis (2000)
Becchetti and Sierra (2002)
Charitou et al. (2004)
2.5. Overview of classic statistical business failure prediction models
Table 1 provides an overview of a large number of academically developed classic

statistical business failure prediction models. This table does not present an exhaustive
overview of all classic statistical models that have been developed to date. It includes only
models developed by academic researchers that are published in the academic literature,
frequently cited and considered to have added significantly to the literature on business
failure. For a supplementary list of failure prediction models, please refer to Zavgren
(1983); Altman (1984); Taffler (1984); Jones (1987); Keasey and Watson (1991); Ooghe
et al. (1995); Dimitras et al. (1996); Altman and Narayanan (1997); Altman and Saunders
(1998).
3. Problems related to the classic statistical methods
The classic statistical failure prediction models are subject to various problems.
First, several problems stem from the ‘classical paradigm’, which fails to take into
account some important aspects of business failure prediction. Secondly, there are
problems related to the neglect of the time dimension of failure. Third, a range of
problems results from the application focus in failure prediction modelling. Finally,
various problems arise because of the use of a linear classification rule, the use of
annual account information, and neglect of the multidimensional nature of failure.
3.1. Problems related to the classical paradigm
Failure prediction modelling is about supervised classification, which starts from the
classical paradigm that, ‘given a set of firms with known descriptor variables and
known outcome class membership, a rule is constructed which allows other companies
to be assigned to an outcome class on the basis of their descriptor variables’ (Hand,
2004). However, this paradigm clearly fails to take account of some important aspects
of the real problem of corporate failure prediction. Moreover, it ignores some important
sources of uncertainty in the classification problem as follows: (1) the arbitrary
definition of failure; (2) non-stationarity and data instability; (3) sampling selectivity;
and (4) the arbitrary choice of the optimisation criteria (Hand, 2004). In this context,
corporate failure prediction studies are subject to over-modelling, which means
optimising the fit of the model to the presented problem. As a result, the predictive
results of classic statistical failure prediction models—in fact, they are classification
results—are misleading because the models are unstable and sample specific. In this
respect, Joy and Tollefson (1975) argued that the predictive abilities of many models
have been exaggerated because researchers have confused the ex-post classification
results with ex-ante predictive abilities. Similarly, Moses and Liao (1987) concluded
that, ‘while statistical analysis is a great tool in building a good model, it also lends the
model an aura of reliability that is misleading’ (p. 27). Therefore, the enthusiasm
caused by the overall good predictive results of the classic statistical corporate failure
prediction models should be mitigated. In this context, Joy and Tollefson (1975);
why defining bankruptcy is shit
Moyer (1977) suggested that the proof is in the eating, which means that before one
can really have confidence in the predictive abilities of a failure prediction model, it
needs to be tested on data subsequent to its construction. In addition, Taffler (1983,
1984) stressed the importance of testing the efficiency of failure models on new
samples.
3.1.1. Arbitrary definition of failure

Although MDA models and conditional probability models assume a dichotomous
dependent variable (populations of failing and non-failing firms that are well defined
and clearly separated from each other), the classic statistical corporate failure
prediction models are based on an arbitrary separation of firms into failing and non-
failing populations. First, the definition of failure itself is arbitrary. In the majority of
studies, a juridical definition of failure is used, mostly bankruptcy6 (Dirickx and Van
Landeghem, 1994; Ward and Foster, 1997; Van Caillie, 1999; Daubie and Meskens,
2002; Charitou et al., 2004). On the other hand, some models are based on a criterion
of ‘financial distress’7 (Keasey and Watson, 1991; Hill et al., 1996; Kahya and
Theodossiou, 1996; Doumpos and Zopoudinis, 1999; Platt and Platt, 2002) or on
failure-related events such as cash insolvency (Laitinen, 1994), loan default (Ward
and Foster, 1997), capital reconstructions, major closures, forced disposals of large
parts of the firm, informal government support, and loan covenant renegotiations with
bankers (Taffler and Agarwal, 2003). The criterion of failure is chosen arbitrarily,
whether a juridical definition of failure or a financial distress definition is used. A
second arbitrary factor is the way in which the definition of failure is applied in order
to separate failing and non-failing firms. As the failure definition is always applied to
a certain arbitrarily chosen year or time period, the separation of firms into failing and
non-failing populations is artificial: the two populations are only mutually exclusive
within the chosen time period (Ooghe and Verbaere, 1985; Ooghe and Joos, 1990;
Ooghe et al., 1995; Altman and Eisenbeis, 1978).
In reality, corporate failure is not a well-defined dichotomy. Therefore,
dichotomizing failure seems inappropriate and the classic statistical modelling
techniques, which are based on a dichotomy assumption, are applied inappropriately
to the topic of corporate failure prediction. In fact, the arbitrary definition of failure
may have serious consequences for the resulting failure prediction model. First, the
developed model may distinguish accurately between the arbitrarily chosen classes in
the failure definition, but fail to do so between the classes of real interest. If a model
6
The juridical definition of failure is popular because it provides an objective criterion that allows firms to be to
be separated easily into two populations (Ooghe and Joos, 1990; Ooghe et al., 1993; Dirickx and Van Landeghem,
1994; Ooghe et al., 1995; Charitou et al., 2004). A second reason is that the moment of failure can be objectively
dated.
7
The following criteria have been used in financial distress definitions: several years of negative net operating
income, suspension of dividend payments, major restructuring or layoffs (Platt and Platt, 2002), low interest
coverage ratio, negative EBIT, negative net income before special items (Platt and Platt, 2004), losses, selling
shares to private investors, entering into a capital restructuring or a reorganization, a few years of negative
shareholders’ funds or accumulated losses (McLeay and Omar, 2000). Keasey and Watson (1991) explicitly
mentioned that the criterion of financial distress is ‘arbitrary in nature’.
is developed on a group of risky, failing firms and a group of non-risky, surviving

firms (Cybinski, 2000, Cybinski, 2001), the strong ex-post classification power of this
model—which is not surprising (Wood and Piesse, 1987)—may be very misleading.
The ex-ante predictive accuracy for a sample consisting of different kinds of firms,
including ‘grey zone’ firms, may be much smaller8. The way in which the two
populations are separated is of major importance, not only for the further development
of the failure prediction model9, but also for the predictive usefulness of the model in
practice. Second, the application of the failure definition to an arbitrarily chosen year
or time period may result in a selection bias (Shumway, 1999) or contaminated
populations (Taffler, 1982; 1983)10 and, hence, may yield performance results that
cannot be generalized.
In addition, although bankruptcy is widely used in failure definitions, there are some
specific problems related to this particular juridical definition of failure. First, as the
declaration of bankruptcy is based mainly on liquidity and/or solvency figures, the
estimation samples for bankruptcy prediction models may be contaminated by firms that
are declared bankrupt despite the fact that they do not show any other real signs of failure.
For example, some companies may have filed for bankruptcy in order to get rid of their
debts and restart their activities with a clean sheet. In these firms, bankruptcy reflects a
strategic decision. In addition, some bankruptcies may result from an unexpected event,
such as a natural disaster. In this context, Hill et al. (1996) distinguished ‘sudden
bankruptcies’, which did not show any previous indications of financial distress, and Davis
and Huang (2004) introduced the concept of ‘accidental bankruptcy’. Contaminated
bankruptcy samples may result in bankruptcy prediction models with poor forecasting
abilities. Second, bankruptcy prediction models ignore the fact that bankruptcy is only one
of a large range of possible endings of the failure process. For instance, failure may result
in other juridical exits, such as merger, absorption, dissolution, or liquidation. Finally,
bankruptcy models do not account for the fact that there may be a long time lag between
the moment when a firm experiences serious problems that make it impossible to operate
in a normal way, or the moment when it ceases to record annual accounts, and the final
juridical exit in the form of a bankruptcy (Theodossiou, 1993; Ooghe et al., 1995; Pompe
and Bilderbeek, 2000). In other words, the timing of bankruptcy may be much later than
the real moment of failure if, for example, the bankruptcy declaration is postponed because
of rescue attempts (in the case of large and economically important firms).
8
Jones (1987), for example, suggested that, ‘the accuracy in predicting bankruptcy among marginal companies,
rather than quite healthy and quite distressed companies, may be the real test of a model’s usefulness’ (p. 147). To
this end, Gilbert et al. (1990) studied the performance of a ratio-based failure model for a sample of problem
companies, in which obviously strong firms were excluded. They found that such a failure model performed
poorly in identifying likely bankruptcies.
9
In the case of an empirical selection of discriminating variables, it is highly likely that the definition of failure
will influence the variable selection (Ooghe et al., 1995; Blazy, 2000; Van Caillie and Dighaye, 2002). However,
in contrast, Hayden (2003) found a similar selection of variables for three different definitions of failure.
10
A possible solution is using an extended time frame for the construction of the populations, as in Ooghe et al.
(1993). However, Back et al. (1997) argued that it is better not to use an extended time frame and to include all
types of non-failing companies, even those with many failing characteristics.
3.1.2. Non-stationarity and data instability

Using a classic statistical failure prediction model based on supervised classification in
a predictive context requires that the relationships among the variables are stable over time
and remain the same in future samples. The classical paradigm assumes that the
distributions of the variables involved in the model do not change over time. This implies
‘stationarity’, which means a stable relationship between the independent variables and
the dependent variable (Edmister, 1972; Zavgren, 1983; Mensah, 1984; Jones, 1987), and
stable inter-correlations between the independent variables (Edmister, 1972; Zavgren,
1983). This is closely related to the notion of ‘data stability’, also called ‘population drift’
(Mensah, 1984; Wood and Piesse, 1987). However, there is ample evidence of data
instability and non-stationarity (Barnes, 1982; Richardson and Davidson, 1984;
Zmijewski, 1984). Data instability may be the result of changes in inflation or interest
rates, and/or the phases of the business cycle (Mensah, 1984), or it may be due to changes
in the competitive nature of the market, corporate strategy, and/or technology (Wood and
Piesse, 1987). Dambolena and Khoury (1980) suggested that the data instability problem
of financial ratios is generally the greatest for firms that are about to fail. Consequently, we
can argue that the classic statistical failure prediction models generally suffer from
stationarity problems (Moyer, 1977; Mensah, 1984; Charitou et al., 2004).
In addition to the predictive use of the classic statistical models, the popular practice of
pooling data across different years11 in the estimation process of classic statistical failure
prediction models requires that the relationships among the variables are stable over time
(Altman and Eisenbeis, 1978; Zmijewski, 1984). In other words, it requires stationarity
and data stability.
Non-stationarity and data instability may have severe consequences for the classic
statistical failure prediction models. First, they may result in models having poor
predictive abilities in future-dated samples, even if a model has good ex-post classification
results (Mensah, 1984)12. Second, data instability causes the models to be fundamentally
unstable or not robust over time. Re-estimation of failure models on new data generally
results in other coefficients. Consequently, classic statistical failure prediction models may
need redevelopment or updating from time to time (Joy and Tollefson, 1975; Taffler, 1982;
Mensah, 1984; Keasey and Watson, 1991; Dirickx and Van Landeghem, 1994; Ooghe et
al., 1994a). Finally, due to non-stationarity and data instability, classic failure prediction
models based on pooled estimation samples may be based on temporarily distorted data
and this may result in inconsistent coefficient estimates (Platt et al., 1994) and low
accuracy levels (Back et al., 1997).
In attempts to overcome or attenuate the problems related to data instability, some
11
Pooled estimation samples of failing firms, which are used in the great majority of corporate failure prediction
studies, consist of firms that are failing in different years. Pooling allows a considerable number of firms to be
included in the failing sample and increases the representativeness of the resulting model (Ooghe and Joos, 1990).
12
For example, Moyer (1977) analysed the classification performance of Altman’s model on a future-dated
sample and found a much lower accuracy than the level reported in the original Altman study. Pompe and
Bilderbeek (2000) pointed out that lower predictive performances for alternative samples show up particularly in
periods of a downward economic evolution. This conclusion is in line with the relationship suggested above
between data instability and changes in the business cycle.
researchers use (in)stability measures (Dambolena and Khoury, 1980; Betts and Belhoul,
1987), industry-relative ratios (Platt and Platt, 1990), or deflated financial ratios (Mensah,
1983). However, Platt and Platt (1991) demonstrated that industry-relative variables do
not significantly increase the stability of the estimated coefficients.
3.1.3. Sampling selectivity

The classical paradigm is based on the assumption that a random sampling design is
used. For this reason, if a classic statistical failure prediction model is eventually to be
used in a predictive context, the estimation samples of failing and non-failing firms should
be representative of the whole population of firms (Ooghe and Joos, 1990). Nevertheless,
in the great majority of classic failure prediction models, non-random samples of firms
with available annual accounts are used. (For example: Altman, 1968; Deakin, 1972;
Blum, 1974; Altman et al., 1977; Taffler and Tisshaw, 1977; van Frederikslust, 1978;
Dambolena and Khoury, 1980; Ohlson, 1980; Zavgren, 1983; Chalos, 1985; Gentry et al.,
1985a; Keasey and Watson, 1987; Aziz et al., 1988; Gloubos and Grammatikos, 1988;
Keasey and McGuinness, 1990; Altman et al., 1995; Mossman et al., 1998).
First, non-random samples may result from over-sampling the failing companies.
Because of the low frequency rate of failing firms in the economy, most researchers draw a
state-based sample, thereby over-sampling the failing firms. This may cause a choice-
based sample bias (Zmijewski, 1984; Platt and Platt, 2002).
Second, non-random samples may be the result of applying a ‘complete data’ sample
selection criterion (Taffler, 1982; Ooghe and Verbaere, 1985; Declerc et al., 1991), which
may lead to a sample selection bias (Zmijewski, 1984). More particularly, failing firms are
more likely to have incomplete data because they tend to be younger and smaller than non-
failing firms.
Finally, non-random samples may result from using matched pairs of failing and non-
failing firms13 (Ohlson, 1980; Platt and Platt, 2002). In this respect, Scott (1981) stated
that, ‘most bankruptcy models are derived using a paired-sample technique. Part of the
sample contains data from firms that eventually failed; the other part contains
contemporaneous data from firms that did not fail’ (p. 320). Examples of matched
sampling models can be found in Altman (1968); Blum (1974); Taffler and Tisshaw
(1977); van Frederikslust (1978); Bilderbeek (1979); Zavgren (1983); Zavgren (1985);
Keasey and McGuinness (1990),Platt and Platt (1990); Mossman et al. (1998); Charitou
and Trigeorgis (2000); Charitou et al. (2004). Usually, pairing is done on criteria of size,
industry and/or age.
Non-random estimation samples may have serious consequences for the resulting
classic statistical failure prediction model. Clearly, it is pointless to build a highly accurate
model for the available data when the model is likely to be inaccurate for failure prediction
in new cases. If the estimation samples are non-random, it might be expected that the
parameter estimates are biased (Zmijewski, 1984) and that the overall ex-post
13
Matching is a common practice, because it enables the researcher to control for some variables that are
believed to have some predictive power but are not included in the set of prediction variables (Zavgren, 1983;
Ooghe and Verbaere, 1985; Jones, 1987; Keasey and Watson, 1991).
classification accuracy is different from the ex-ante predictive performance. In other

words, the reported accuracy cannot be generalized (Piesse and Wood, 1992) and may be
very misleading14.
Over-sampling of failing firms is very likely to result in an overstatement of the ex-post
accuracy of the model because of an understatement of the misclassification error rate for
the over-sampled failing group of firms. Such a model will pay more attention to
accurately classifying the failing companies at the expense of more misclassifications of
non-failing firms15 (Zavgren, 1983; Zmijewski, 1984; Piesse and Wood, 1992; Platt and
Platt, 2002). In contrast, the use of a complete data sample selection criterion is likely to
result in an understatement of the misclassification error rate for the non-failing group of
firms. Furthermore, if some characteristics are over- or under-represented in the matched
samples, due to the use of matched pairs, the likely result is a sample-specific model
(Zavgren, 1983), biased model coefficients (Zmijewski, 1984), and misleading indications
of the model’s predictive accuracy (Keasey and Watson, 1991). Moreover, if the matching
criteria show any link with the failure probability, this may result in a selection bias and in
a restricted failure prediction model (Taffler, 1982). For this reason, Eisenbeis (1977)
stated that, ‘every effort should be made to avoid arbitrary grouping’ (p. 889).
Nevertheless, in many studies the matching criteria are chosen arbitrarily (Ohlson,
1980; Ooghe and Verbaere, 1985; Peel and Peel, 1987; Ooghe et al., 1993). Finally, it is
obvious that if a failure prediction model is estimated on samples that are non-random with
respect to certain general characteristics—such as industry, size and age—the application
of the model to the under-represented types of firms may give inefficient predictions.
3.1.4. Choice of the optimization criteria

When fitting a classic statistical failure prediction model, one tries to optimize a certain
measure of goodness-of-fit. Various goodness-of-fit measures are possible, such as the
percentage of correct classifications (i.e. the unweighted error rate), the type I or type II
classification error, the Gini-coefficient (Ooghe and Balcaen, 2002; Ooghe et al., 2003),
R2-type measures, measures based on entropy (Zavgren, 1985), and likelihood. Different
criteria generally lead to different models, and, therefore, it is clear that a failure prediction
model largely depends on the choice of the optimization measure. In failure prediction
studies, a certain performance measure is selected arbitrarily and, in this context, the
classical paradigm fails to take account of the real nature of corporate failure prediction.
Nevertheless, it should be emphasized that there is little point in finding a model that
minimizes a performance measure that is irrelevant to one’s aims. Marginal improvements
of the irrelevant criterion could have adverse effects for the relevant criterion. In
particular, likelihood is seldom the appropriate performance measure16 (Hand, 2004).
14
In contrast, Zmijewski (1984) found that non-random samples do not significantly affect the overall accuracy
rates or the statistical inferences on the impact of the independent variables. Only type I and type II errors are
influenced.
15
This may be beneficial as the cost of a type I error is usually considered to be much larger than the cost of a
type II error.
16
Likelihood is generally viewed as a good goodness-of-fit measure for use in cases where one does not know
the real criterion (Hand, 2004).
3.2. Problems related to the neglect of the time dimension of failure
The static classic statistical failure prediction models ignore the fact that companies
change over time, which causes various problems and limitations.
First, almost all classic statistical failure prediction models are subject to a number of
problems related to using only one single observation (i.e. one annual account) for each
firm in the estimation samples. The use of one single observation is based on the implicit
assumption that consecutive annual accounts are independent, repeated measurements.
Obviously, this assumption is not met: the observations are not entirely independent
(Dirickx and Van Landeghem, 1994).
An initial problem in this context relates to the snapshot character of the annual
account. The choice of when to observe a firm may introduce a selection bias in the
resulting model (Mensah, 1984; Shumway, 1999). For example, it is possible that a
healthy firm suffering from a temporarily adverse situation is classified as failing by
the model. A second problem is that the static failure prediction models do not
account for time-series behaviour. Although many authors have argued that the
prediction of failure should depend on more than one annual account or a change in
financial health17 (Bilderbeek, 1979; Shumway, 1999), past information regarding
corporate performance has been ignored (Theodossiou, 1993; Dirickx and Van
Landeghem, 1994; Kahya and Theodossiou, 1996). A third problem is that in an ex-
ante predictive context, the repeated application of a failure prediction model to
consecutive annual accounts of one particular firm may result in a whole list of
potentially conflicting predictions (Keasey et al., 1990; Dirickx and Van Landeghem,
1994). This is called the signal inconsistency problem (Keasey et al., 1990; Luoma
and Laitinen, 1991).
Second, as the fixed score output of the classic statistical failure prediction models
clearly contradicts general intuition, it might be argued that static models, such as
MDA and logit models, are not suited to corporate failure prediction. For example,
Altman and Eisenbeis (1978) pointed out that ‘.the concept of prediction [.] is not
strictly applicable to the standard discriminant analysis, which does not explicitly
incorporate a time dimension’ (p. 186). In fact, the principal aim of a classic
statistical model is to summarize information to determine whether a firm’s profile
more closely resembles the profile of a failing firm or that of a non-failing firm.
Altman et al. (1981) referred to this as the concept of ‘resemblance’. In this context,
Taffler (1982; 1983) and Taffler and Agarwal (2003) stressed that an MDA model
analyses the following question: ‘Does the company have a profile more similar to the
failing group of companies from which the model was developed or the non-failing
group?’. A model score below a certain threshold highlights financial difficulties (i.e.
the firm ‘might’ fail), but it does not indicate that the company ‘will’ fail. Classic
17
Tamari (1966) indicated the importance of trend analysis. For the same reason, Edmister (1972) included the
trend of financial ratios as failure indicators in his failure prediction model. Similarly, Dambolena and Khoury
(1980); Betts and Belhoul (1987) found that the inclusion of (in)stability measures improves the classification
results of failure prediction models, and Chalos (1985) pointed out that trend data could reveal interesting
information.
failure prediction models have a retrospective character. What is demonstrated is that

failing and non-failing firms have dissimilar characteristics, not that the variables have
predictive power18 (Ooghe and Joos, 1990). Their retrospective character and the
notion of resemblance cause the classic statistical failure prediction models to have a
descriptive or pattern recognition nature (Taffler, 1982; Taffler, 1983; Keasey and
Watson, 1991; Taffler and Agarwal, 2003). Therefore, in the strict sense, the
developed models must not be seen as prediction devices, but rather as
communication devices. For example, instead of using z-scores as indications of
failure probabilities, it seems more appropriate to analyse a firm’s financial health by
examining the evolution of its z-scores over time (Taffler, 1983; Ooghe and Joos,
1990).
Nevertheless, in practice, the classic statistical failure prediction models are used as a
prediction device. In this context, the time frame-based failure prediction models, which
are estimated on observations one, two, or three years prior to the event of failure, are
considered to provide a certain classification statement on the failure/survival status in
years tC1, tC2, or tC3 (Deakin, 1972). It should be emphasized that the use of a short-
term model estimated on observations one year prior to failure should be restricted to
classification statements concerning year tC1, because such a model is likely to become
unreliable for long-term failure predictions (Lane et al., 1986). In fact, in this case, the
model has little or no meaning. Moreover, re-estimation or re-development on
observations from other years prior to failure would very likely result in other coefficients
and/or variable selections.
Finally, the fact that the classic statistical failure prediction models do not treat
company failure as a process results in some serious drawbacks. Besides approaching
failure as a discrete event (Altman, 1984), as reflected in the dichotomous dependent
variable, the static failure prediction models do not consider information on the progress
and the dynamics of the failure process to predict failure (Van Wymeersch and Wolfs,
1996). Instead, they assume that failure is a steady state (Luoma and Laitinen, 1991;
Laitinen, 1993; Laitinen and Kankaanpää, 1999). The underlying failure process is
assumed to be stable over time and no phases are distinguished. However, this strongly
contrasts with reality. It is obvious that firm failure is not a sudden and unexpected event
(Luoma and Laitinen, 1991). Instead, failure is the result of a failure process or a failure
path, which may consist of several phases, each characterized by specific behaviour of
certain variables or specific symptoms of failure19. Moreover, the classic statistical failure
prediction models do not consider possible differences in failure paths: they assume that all
companies follow a uniform failure process. However, in practice, there are obviously a
wide variety of failure paths (Laitinen, 1991).
The steady-state assumption and the uniform failure-process assumption of the static
failure prediction models may have serious consequences. For example, in an empirically
founded model, the relative importance of the variables and the ex-ante predictive
18
When predicting company failure, the crucial problem is making an inference in the opposite, prospective
direction, from variables to failure (Johnson, 1970).
19
As indicated by Daubie and Meskens (2002), the relative importance of the variables for the detection of
failure is not constant over time.
accuracy of the model will be implicitly determined by the frequency of occurrence of the
different phases of the failure process and the different kinds of failure paths in the
estimation sample of failing firms20.
3.3. Problems related to the application focus
Most classic statistical failure prediction models are the result of ‘putting the cart
before the horse’ (Cybinski, 2001). They have been developed without a complete
understanding of the nature of company failure, sometimes due to commercial pressure.
The large majority of the models reported in the academic literature, are strongly
application driven: researchers apply an outsider’s or external viewpoint and a financial
approach to company failure. They construct failure models on the basis of several
statistical techniques and the financial data of a sample of failing and non-failing
companies in order to predict failure on a rather short-term basis, one to three years prior to
failure (Cybinski, 2001; Van Caillie and Arnould, 2001). Scott (1981) explained the
application focus as follows: ‘Most bankruptcy models are derived using a paired-sample
technique [.]. A number of plausible and traditional financial ratios are calculated from
financial statements that were published before failure. Next, the researcher searches for a
formula based either on a single ratio or a combination of ratios that best discriminates
between firms that eventually failed and firms that remained solvent’ (p. 320). The models
are simply the outcome of a statistical search through a number of plausible financial
indicators in order to empirically find some characteristics that distinguish between failing
and non-failing firms. There seems to be no consensus on the superior predictor variables,
or on the superior modelling method. In this respect, the two following sections discuss the
various problems related to the application focus and, more particularly, the problems
concerning the variable selection and the selection of the modelling method.
3.3.1. Variable selection

The great majority of the classic statistical failure prediction models result from an
empirically based variable selection. They start from an extensive initial battery of
variables, often arbitrarily chosen on the basis of their popularity in the literature and their
predictive success in previous research21. The theoretical basis for the selection of
variables has always been limited, which has prevented a better selection of variables
(Karels and Prakash, 1987; Dirickx and Van Landeghem, 1994). From this initial battery
of variables, a final set of variables is selected, largely based on statistical or empirical
20
For example, a model estimated on data one year prior to failure, mainly considering the final phase of the
failure process, will perform poorly when it is used for failure prediction many years prior to failure. In the same
way, the sample construction with respect to the different phases of the failure process may explain the unstable
coefficients when a model estimated on data one year prior to failure is re-estimated for a sample of annual
accounts two and three years prior to failure (Laitinen, 1993). Furthermore, if an estimation sample mainly
contains ‘acute failure’ firms, this may explain why the resulting model performs poorly when applied to a sample
with a high frequency of ‘chronic failure’ firms.
21
The selection of variables based on popularity may be problematic because popular ratios are more likely to
be subject to window dressing and, hence, are more likely to be unreliable (Beaver, 1967b).
considerations because there is no theory indicating which (financial) variables are the best
predictors22 (Scott, 1981). When summarizing the variables used in classic failure
prediction models, there seems to be little agreement concerning which variables are the
best in distinguishing between failing and non-failing companies23 (Edmister, 1972; Back
et al., 1996; Altman and Narayanan, 1997; Altman and Saunders, 1998; Mossman et al.,
1998; Becchetti and Sierra, 2003). In this respect, Zavgren (1983) stated that the lack of
knowledge concerning the relevant predictors has seriously thwarted the development of a
scientific approach to corporate failure prediction.
Although it is the most popular method of variable selection, the empirically based
variable selection has some serious drawbacks (Keasey and Watson, 1991). First, there are
strong indications that the choice of variables is sample specific and, therefore, the
resulting failure prediction model is very likely to be sample specific (Edmister, 1972;
Zavgren, 1983; Zavgren, 1985). The empirical findings may not be suitable for
generalization (Edmister, 1972; Gentry et al., 1987). Moreover, because of the empirical
selection of variables, the resulting model may be not diversified, and may even show
counter-intuitive signs for some coefficients. The models of Bilderbeek (1979); Zavgren
(1985); Gloubos and Grammatikos (1988); Keasey and McGuinness (1990) are only a few
of the large number of models with unexpected coefficients24, which have been caused by
a high correlation among the individual ratios (Moses and Liao, 1987; Ooghe and Balcaen,
2002; Hayden, 2003). This contradicts the general viewpoint that a good failure prediction
model should include some carefully chosen variables from the whole spectrum of
financial analysis—liquidity, indebtedness, profitability, and activity (Dambolena and
Khoury, 1980)—and that it should use these variables in the intuitively right sense25. In
this context, Karels and Prakash (1987) warned that a careful selection of predictor
variables is needed to improve the predictive performances of failure models.
Although the variable selection in the large majority of studies is application driven and
done empirically, some researchers have a (limited) theory or theoretical frameworks
underlying their variable selection. Most of these studies are based on a certain cash-flow
theory. Examples include: Beaver (1967a); Blum (1974); Gentry et al. (1985b); Aziz et al.
(1988); Aziz and Lawson (1989); Ooghe et al. (1993), and Charitou et al. (2004). Other
theoretical models are the gambler’s ruin model (Wilcox, 1971), the framework of option
pricing (Charitou and Trigeorgis, 2000), and the integrated ratio model (Ooghe
22
Empirical considerations include the statistical significance of the estimated parameters, the individual
discriminating ability of each of the variables, the classification results of different combinations of variables or
stepwise methods, the signs of the variables’ coefficients, principal components analysis, and factor analysis. In
this context, we can refer to the term ‘brute empiricism’.
23
Dimitras et al. (1996); Daubie and Meskens (2002) provided an extensive overview of financial ratios
included in corporate failure prediction models. According to Daubie and Meskens (2002), the most frequently
used financial ratios are: current assets/current liabilities, working capital/total assets, EBIT/total assets, quick
assets/current liabilities, and net income/total assets. These ratios also appear in the study by Dimitras et al.
(1996).
24
See Ooghe and Balcaen (2002) for a discussion of the counter-intuitive signs of some coefficients in these
models.
25
Ooghe and Balcaen (2002) showed that well performing models are the result of a diversified combination of
variables, used in the intuitively right sense.
and Verbaere, 1985). It is obvious that a theory-based variable selection reduces the scope
for statistical over-fitting (Scott, 1981).
In this context, it should be mentioned that there is ample evidence demonstrating
that the largest gains in classification accuracy are made by relatively simple models
with a small number of predictors. First, in line with the 80/20 law (Pareto’s Principle)
and the law of diminishing returns, the marginal improvement in the accuracy of a
model is expected to decrease as the model complexity increases. In other words, the
initial crude versions of a failure prediction model are likely to provide the greatest
reduction in uncertainty relating to the failure prediction (Hand, 2004). When there is a
reasonably strong mutual correlation between the predictor variables, the simplest
models contribute substantially more to reducing unexplained variance than do the
refined and more complex models. Much of the predictive power of additional predictors
has already been accounted for by the existing predictors through their correlations
(Hand, 2004).
3.3.2. Selection of a modelling method

Supervised classification concerning business failure with classic statistical modelling
methods involves the choice of a method: MDA, LA, PA, or LPM. Despite the extensive
literature, there seems to be no superior modelling method. It is impossible to ascertain a
superior method from an examination of a large range of comparative studies that compare
the ex-post classification results and/or the ex-ante prediction abilities of these different
kinds of failure prediction models. Most studies reach heterogeneous conclusions and
point in different directions26. As a result of this lack of consensus, the selection of
modelling method is left entirely to the researcher.
On the other hand, there is ample evidence supporting the assertion that large gains in
classification accuracy are made by the relatively simple modelling methods, whereas the
more sophisticated methods yield rather small marginal improvements. The simple MDA
models can produce over 90% of the predictive power that can be achieved by the more
complex models and they are less likely to over-fit (Hand, 2004).
3.4. Other problems
3.4.1. The linearity assumption

As mentioned in Section 1, both the univariate analysis method and the MDA technique
are based on the assumption of linearity, which is a very strong assumption.
Univariate corporate failure prediction models assume that if a firm’s value for a certain
predictor is higher (lower) than a certain cut-off point, this signals strong (poor) financial
health. An increase (decrease) in a variable then indicates an increase (decrease) in
financial health. However, in practice, this assumption is often violated (Keasey and
Watson, 1991): some variables indicate financial problems when they have a very low or a
very high value. For this reason, the univariate modelling technique is often applied
incorrectly, which results in questionable classifications.
26
See Ooghe and Balcaen (2004) for a comprehensive overview of comparative studies.
Similarly, MDA involves a linear classification rule. A discriminant score above

(below) a certain cut-off point automatically signals a healthy (poor) financial situation.
However, this classification rule intuitively conflicts with the fact that some of the
composing variables might not show a linear relationship to financial health.
3.4.2. The use of annual account information

The majority of the classic cross-sectional models use only annual account information
in the form of financial ratios in order to predict failure (Dimitras et al., 1996). Financial
ratios are used because they are hard, objective measures and because they are based on
publicly available information (Micha, 1984; Laitinen, 1992; Dirickx and Van
Landeghem, 1994). On the other hand, financial ratios have been subject to many
criticisms, and annual account-based models have proven to suffer from some serious
drawbacks. We would like to emphasize here that, despite the criticisms, the role of
financial ratios in failure prediction is very important.
An initial problem related to the use of annual account information (financial ratios) is
that the obligation to prepare and/or publish annual accounts is restricted and mostly
depends on criteria concerning firm type and/or firm size. In most countries (including, for
example, the USA, the UK, and Germany) only large firms—those which meet certain
criteria concerning asset size, sales level, and/or number of employees—are obliged to
publish their annual accounts. Therefore, many failure prediction models have been
restricted to large businesses.
Second, when predicting failure on the basis of financial ratios, researchers implicitly
assume that the annual accounts give a fair and true view of the financial situation.
However, it seems reasonable to assume the opposite. First, there is much anecdotal and
academic evidence that firms in general, and unhealthy or failing firms in particular, have
incentives to manipulate or manage their annual account figures (Ooghe and Joos, 1990;
DeFond and Jiambalvo, 1994; Sweeney, 1994; Ooghe et al., 1995; Burgstahler and Dichev,
1997; Degeorge et al., 1999; Rosner, 2003). By means of creative accounting practices,
failing firms adjust their earnings upwards and give a more positive presentation of their
financial situation, especially when the moment of failure is very near (Argenti, 1976;
Watts and Zimmerman, 1986; Ooghe and Joos, 1990; Ooghe et al., 1995; Dechow et al.,
1996; Blazy, 2000; Charitou and Lambertides, 2003; Rosner, 2003). Second, annual
accounts may be unreliable, especially in smaller firms, because of the lack of an internal
control system (Keasey and Watson, 1986; 1987) or because of annual account
adjustments made by the auditor in the light of a bankruptcy filing (referred to as
‘accommodated’ annual accounts) (Charitou and Lambertides, 2003). Because of the
unreliability of annual account information, failure prediction models based on financial
ratios may be distorted and their practical usefulness may be limited.
Third, corporate failure prediction models based on annual account data (financial
ratios) suffer from the occurrence of extreme ratio values, errors, and missing values. As a
result of extreme ratio values, models may be strongly biased or contaminated (Moses and
Liao, 1987) and may show biased coefficients for the ratios involved. Furthermore, as a
result of annual account errors, models based on erroneous annual account information
may become worthless. The problem of annual account errors is often underestimated. For
example, several studies on the formal quality of Belgian annual accounts have shown that
the quality of many annual accounts is poor, especially in small firms (Dirickx and Van
Landeghem, 1994, p. 453; see also Ooghe and Joos, 1990). Finally, a large number of
annual accounts have missing values. In many studies, annual accounts with missing
values are simply deleted from the analysis. Possible solutions to these annual account
problems are to trim the ratios with extreme values at certain percentiles and to replace the
missing values by mean or random values (Tucker, 1996).
A fourth problem is that many short-term models (for failure prediction one or two
years prior to failure) suffer from the absence of annual accounts at the end of the failure
process. More particularly, many firms stop preparing annual accounts one or two years
prior to bankruptcy. In these cases, researchers generally observe the last published annual
account and, consequently, they implicitly study the real moment of failure instead of
bankruptcy. A related problem appears when the short-term model is to be applied in
practice to predict the failure of firms in year tC1. In most cases, failing firms delay the
publication of their annual account when they approach failure (bankruptcy). In this
context, Deakin (1977) remarked that, ‘in many cases that [annual] report is delayed for
failing companies and may not be available until the failure event’ (p. 75). In these cases, it
is impossible to apply the model in the appropriate way and, as a result, the model becomes
useless.
Fifth, failure prediction models based on financial ratios implicitly assume that all
relevant failure or success indicators—both internal and external—are reflected in the
annual accounts. However, it is clear that not all relevant information is reflected in the
annual accounts. In this context, Argenti (1976) stated that, ‘while these [financial] ratios
may show that there is something wrong . I doubt whether one would dare to predict
collapse or failure on the evidence of these ratios alone’ (p. 138). In addition, Zavgren
(1985) pointed out that, ‘any econometric model containing only financial statement
information will not predict with certainty the failure or nonfailure of a firm’ (p. 22-23).
Furthermore, Maltz et al. (2003) mentioned that the use of financial measures as sole
indicators of organizational performance is limited. For this reason, some authors have
advised including non-accounting or qualitative failure indicators in failure prediction
models (Ohlson, 1980; Zavgren, 1983; Keasey and Watson, 1987; Lussier and Corman,
1994; Sheppard, 1994; Lussier, 1995; Slowinski and Zopoudinis, 1995; Doumpos and
Zopoudinis, 1999; Becchetti and Sierra, 2003; Daubie and Meskens, 2002; Lehmann,
2003). This might be particularly appropriate when studying small firms, which often lack
reliable annual account information. Examples of non-accounting and qualitative
indicators are staffing, management experience, education, age, motivation, social skills,
and the leadership quality of the owner/manager, the number of partners, the existence of a
plausible long-term business strategy, productive efficiency, customer concentration,
dependence on one or a few large suppliers, subcontracting status, export status, the
presence of large competitors in the same region, relationship with banks, level of
diversification, profitability of the industry, industry growth rate, market share, number of
joint ventures, characteristics of the board of directors, group relations, stock value, and
events reflecting management (corrective) actions, such as reductions in dividends,
qualified audit opinions as to whether a firm is a going concern, troubled debt
restructurings, and violations of debt agreements (Ooghe and Joos, 1990; Flagg et al.,
1991; Hall, 1994; Lussier and Corman, 1994; Sheppard, 1994; Becchetti and Sierra, 2003;
Lehmann, 2003). General firm characteristics concerning industry type and firm size have
also proven to be very important variables in failure prediction (Bilderbeek, 1978; Taffler,
1983; Mensah, 1984; Taffler, 1984; Laitinen, 1992; Hill et al., 1996; Blazy, 2000; Daubie
and Meskens, 2002; Ooghe et al., 2003). Firm age may also be an important indicator. In
addition, because a firm never stands alone, a failure prediction model should consider
information about the external environment, including both the macroeconomical
situation—for example, the interest rates, the business cycle, and the availability of
credit (Zavgren, 1983)—and industry-specific conditions, such as the prospects of the
industry. In this respect, Johnson (1970) pointed out that financial ratios ‘do not contain
information about [.] the intervening economic conditions’, adding that, ‘the riskiness of
a given value for [a] ratio changes with the business cycle’ (p. 1166–1167). Similarly,
Richardson et al. (1998) asserted that accounting-based failure models generally do not
control for any changes that may occur in the information content of accounting data
because of recession. Macroeconomic and industry-specific variables that may have an
impact on business failure are interest rates and the occurrence of credit shocks (Swanson
and Tybout, 1988), macroeconomic instability (Bhattacharjee et al., 2002), the potential of
the market, industry profitability, and competition (Lehmann, 2003). Finally, socio-
scientific factors can be taken into account when predicting failure (Bijnen and Wijn,
1994).
An additional problem concerning the use of financial ratios in failure prediction
models relates to the fact that financial ratios are constructed from different components,
each of them reflecting other information on a firm’s financial health. It is possible that a
financial ratio does not discriminate between failing and non-failing firms but that its
components do differ between these types of firms (Beaver, 1967b). For this reason, it
might be interesting to analyse the components of financial ratios, instead of the ratios
themselves. A similar remark applies to the overall ratios, which are composed of different
detailed ratios. It is possible that an overall ratio (for example, the rotation of total assets)
does not signal poor financial health because, although one detailed ratio (for example, the
profitability of total assets) has a negative impact on financial health, it is offset by another
detailed ratio (for example, the profit margin of sales) that has a positive impact on
financial health. For the timely detection of problems, analysis of the detailed ratios may
be necessary (Bilderbeek, 1978, 1979).
Finally, although many studies have compared the predictive abilities of accrual-based
financial ratios and cash flow-based ratios, there seems to be no consensus as to which
types of financial ratios are the best failure indicators. Some studies have suggested using
cash flow-based funds flow components instead of accrual-based financial ratios in failure
prediction modelling (Gentry et al., 1985a; Gentry et al., 1987; Aziz and Lawson, 1989;
Declerc et al., 1990) or, at least, improving model accuracy by adding cash flow ratios to
models based on accrual-based financial ratios (Gombola and Ketz, 1983; Gentry et al.,
1985b; Sharma and Iselin, 2003). Other studies have opposed using of cash flow-based
models (Casey and Bartczak, 1984; Gentry et al., 1985a), or have suggested that cash flow
ratios do not provide any additional information in failure prediction models (Casey and
Bartczak, 1984; Gombola et al., 1987). In this respect, it should be mentioned that value-
added ratios have been neglected in failure prediction research. Despite the fact that a
study by Declerc et al. (1991) showed that value-added ratios do increase a model’s
classification results when incorporated into a failure prediction model along with other
financial ratios, only a small number of studies have used value-added ratios.
3.4.3. The neglect of the multidimensional nature of failure

In univariate models, the neglect of the multidimensional nature of corporate failure is
the cause of serious problems. First, the univariate model conflicts with reality in that
Table 2
Overview of the problem topics and related issues
Problem category Problem topics Related problem issues

Classical paradigm A. Arbitrary definition of failure Arbitrary definition of failure
Arbitrary application of failure
definition
B. Non-stationarity and data Predictive context
instability
Pooled estimation samples
C. Sampling selectivity Over-sampling of failing firms
Complete data sample selection
criterion
Matched pairs
D. Choice of the optimization
criteria
Neglect of the time dimension of A. Use of one single observation Snapshot character
failure (annual account)
Neglect of time-series behaviour
Predictive context
B. Fixed score output/concept of
resemblance/descriptive nature
C. Failure not seen as a process Steady-state assumption
Assumption of uniform failure
process
Application focus A. Variable selection Empirically based variable selec-
tion
B. Selection of modelling method Ad hoc selection
Other problems A. Use of a linear classification
rule
B. Use of annual account infor- Restricted obligation to prepare
mation (financial ratios) annual accounts
Assumption of fair and true view
Extreme ratio values, errors, and
missing values
Absence of annual accounts
Annual accounts do not reflect all
relevant failure indicators
Information content of com-
ponents and detailed ratios
No consensus on type of financial
ratios
C. Neglect of multidimensional
nature of failure
the financial status of a company is a complex, multidimensional concept, which cannot be

analysed by one single ratio. Moreover, as firm classification with a univariate model can
only occur for one ratio at a time, this may result in inconsistent and confusing
classifications for different ratios on the same firm (Altman, 1968; Zavgren, 1983). These
inconsistency problems may be considerable and, consequently, it may be impossible to
get a clear picture of a firm’s financial health based on a univariate model.
3.5. Overview
Table 2 gives an overview of all problems related to the classic statistical failure
prediction models discussed above. This table summarizes the four main problem
categories, the corresponding problem topics, and the related issues.
4. Conclusion
The topic of corporate failure prediction has become a major research domain within
corporate finance. Over the last 35 years, many academic studies have been dedicated to
searching for the best failure prediction model that can classify companies according to
their financial health or failure risk. The classic cross-sectional statistical methods have
been the most popular methods, with MDA and LA being the dominant techniques. A huge
number of single-period or static classification models have been developed.
Four types of classic statistical methods have been applied in corporate failure
prediction studies: (1) univariate analysis, (2) risk index models, (3) MDA, and (4)
conditional probability models (logit, probit, and linear probability models). Each of these
methods has its specific features and assumptions. For example, apart from assuming a
dichotomous dataset, the technique of MDA assumes multivariate normally distributed
independent variables, equal variance-covariance matrices across the failing and non-
failing group, and specified misclassification costs and prior probability of failure. On the
other hand, the less demanding LA method has no assumptions regarding the distribution
of the independent variables or the prior probability of failure. LA assumes only a
dichotomous dependent variable, consideration of the costs of type I and type II error rates,
and the absence of multicollinearity.
Although the classic statistical methods of MDA and conditional probability models
have proven to be very popular in corporate failure prediction, there appear to be several
problems related to the application of these methods to corporate failure prediction
modelling.
First, a group of problems is related to the fact that the classical paradigm, on which the
classic statistical methods of MDA and LA are based, fails to take account of some
important aspects of the real problem of business failure prediction. As a result, corporate
failure prediction studies are subject to over-modelling. The models concerned may report
strongly misleading predictive results and may be very unstable and sample specific. A
first issue in this context is the arbitrary definition of failure. Dichotomizing failure is
conflicts with reality and leads to an inappropriate application of the classic statistical
modelling techniques. An arbitrary definition of failure may result in models with
a misleading classification power and a weak predictive usefulness in practice. A second

issue in the context of the classical paradigm concerns non-stationarity and data
instability. As both the predictive use of a failure prediction model and the use of pooled
estimation samples require stationarity and data stability, non-stationarity and data
instability may result in fundamentally unstable models, which have poor predictive
abilities in future-dated samples. Third, as the classical paradigm is based on the
assumption of random sampling, sampling selectivity—due to the over-sampling of failing
firms, the application of a complete data sample selection criterion, or the use of matched
pairs—may have serious consequences for the resulting models in terms of biased
parameter estimates, misleading accuracy levels, and a restricted predictive usefulness. A
final issue in the context of the classical paradigm concerns the arbitrary choice of the
optimization criterion.
A second group of problems relate to the neglect of the time dimension of failure. More
particularly, when one single observation only (i.e. one annual account) is used for each
firm in the estimation samples, the snapshot character of the annual account may introduce
a selection bias and inconsistency problems. Furthermore, the classic models, which have
a fixed score output, have a retrospective and descriptive nature and are based on the
concept of resemblance. Therefore, they are not suitable as prediction devices. Moreover,
due to the steady-state assumption and the uniform failure path assumption of the classic
failure prediction models, failure is not treated as a process and models may be biased
towards prediction of certain specific kinds of failure phases or failure processes.
Third, as a result of the application focus in failure prediction modelling, there is no
consensus on the superior predictor variables, or on a superior modelling method. The
empirically based variable selection may result in sample-specific models, which are not
diversified and even show counter-intuitive signs for some coefficients.
Finally, there are a number of other problems related to the use of a linear classification
rule, the use of annual account information, and neglect of the multidimensional nature of
failure.
In this paper, several issues viewed in isolation by earlier studies are considered
together, thus contributing towards a better understanding of the features of the classic
statistical failure prediction models and their related problems.
The alternative methods for modelling business failure—such as multi-logit analysis,
survival analysis, machine-learning decision trees, expert systems and neural networks—
are beyond the scope of this study. However, as the literature does not provide a clear
overview of the application of alternative methods to the topic of business failure
prediction, further research concerning these methods is necessary. In this context, it
would be interesting to systematically compare the performances of the various classic
statistical methods and the alternative modelling methods.
References
Altman, E.I., 1968. Financial ratios, discriminant analysis and the prediction of corporate bankruptcy.
The Journal of Finance 23 (4), 589–609.
Altman, E.I., 1984. The success of business failure prediction models—An international survey. Journal of
Banking and Finance 8, 171–198.
Altman, E.I., Eisenbeis, R.A., 1978. Financial applications of discriminant analysis: a clarification. Journal of
Financial and Quantitative Analysis, 185–195.
Altman, E.I., Narayanan, P., 1997. An international survey of business failure classification models. Financial
Markets, Institutions and Instruments 6 (2), 1–57.
Altman, E.I., Saunders, A., 1998. Credit risk measurement: developments over the last 20 years. Journal of
Banking and Finance 21 (11–12), 1721–1742.
Altman, E.I., Haldeman, R.G., Narayanan, P., 1977. ZETA analysis: a new model to identify bankruptcy risk of
corporations. Journal of Banking and Finance 1 (1), 29–51.
Altman, E.I., Avery, R.B., Eisenbeis, R.A., Sinkey, J.F., 1981. Application of classification techniques in
business, banking and finance. JAI Press, Greenwich (Connecticut).
Altman, E.I., Eom, Y.H., Kim, D.W., 1995. Failure prediction: evidence from Korea. Journal of International
Financial Management and Accounting 6 (3), 230–249.
Argenti, J., 1976. Corporate Collapse: the Causes and Symptoms. Mc Graw-Hill, Londen.
Atiya, A.F., 2001. Bankruptcy prediction for credit risk using neural networks: a survey and new results. IEEE
Transactions on Neural Networks 12 (4), 929–935.
Aziz, A., Lawson, G., 1989. Cash flow reporting and financial distress models: testing of hypotheses. Financial
Management 18 (1), 55–63.
Aziz, A., Emanuel, D.C., Lawson, G.H., 1988. Bankruptcy prediction—An investigation of cash flow based
models. Journal of Management Studies 25 (5), 419–437.
Back, B., Laitinen, T., Sere, K., Van Wezel, M., 1996. Choosing bankruptcy predictors using discriminant
analysis, logit analysis and genetic algorithms, Technical Report from Turku Centre for Computer Science,
Report number 40, September, pp. 1–18.
Back,B., Laitinen, T., Hekanaho, J., Sere, K., 1997. The effect of sample size on different failure prediction
methods, Technical Report from Turku Centre for Computer Science, Report number, 155, December, pp.
1–23.
Beaver, W., 1967. Financial ratios predictors of failure. Empirical research in accounting: selected studies 1966.
Journal of Accounting Research 4 (Suppl.), 71–111.
Beaver, W., 1967b. Alternative accounting measures as predictors of failure. American Accounting Association
Manuscript Contest for 1967, 113–121.
Barnes, P., 1982. Methodological implications of non-normality distributed financial ratios. Journal of Business
Finance and Accounting 9 (1), 51–62.
Barnes, P., 1987. The analysis and use of financial ratios: a review article. Journal of Business Finance and
Accounting 14 (4), 449–461.
Becchetti, L., Sierra, J., 2003. Bankruptcy risk and productive efficiency in manufacturing firms. Journal of
Betts, J., Belhoul, D., 1987. The effectiveness of incorporating stability measures in company failure models.
Journal of Business Finance and Accounting 14 (3), 323–334.
Bhattacharjee, A., Higson, C., Holly, S., Kattuman, P., 2002. Macro economic instability and business exit:
determinants of failures and acquisitions of large UK firms, Cambridge Working Papers in Economics,
Department of Applied Economics, University of Cambridge, Working paper number 0206, March.
Bijnen, E.J., Wijn, M.F., 1994. Corporate prediction models: ratios or regression analysis?, Research memoranda,
Faculty of Economics and Business Administration, Tilburg University, Paper number, 658, May, pp. 1–23.
Bilderbeek, J., 1978. Het voorspellen van falingen. Financiële kengetallen als thermometer voor de
ondernemingsdoorlichting. Economisch en Sociaal Tijdschrift 32 (1), 5–25.
Bilderbeek, J., 1979. An empirical study of the predictive ability of financial ratios in the Netherlands. Zeitschrift
Für Betriebswirtschaft, 388–407.
Blazy, R., 2000. La faillite, éléments d’analyse économique, Préface de Pierre Morin, Paris (France): Economica.
Blum, M., 1974. Failing company discriminant analysis. Journal of Accounting Research 12 (1), 1–25.
Burgstahler, D., Dichev, I., 1997. Earnings management to avoid earnings decreases and losses. Journal of
Accounting and Economics 24 (1), 99–126.
Casey, C., Bartczak, N., 1984. Cash flow: it’s not the bottom line. Harvard Business Review 4, 60–66.
Chalos, P., 1985. Financial distress: a comparative study of individual, model and committee assessments.
Journal of Accounting Research 23 (2), 527–543.
Charitou, A., Trigeorgis, L., 2000. Option-based bankruptcy prediction, Paper presented at 6th Annual Real
Options Confernce, Paphos, Cyprus, 4–6 July, pp. 1–25.
Charitou, A., Lambertides, N., 2003. Earnings management prior to bankruptcy, Working Paper, University of
Cyprus, pp. 1–46.
Charitou, A., Neophytou, E., Charalambous, C., 2004. Predicting corporate failure: empirical evidence for the
UK. European Accounting Review 13 (3), 465–497.
Cybinski, P., 2000. The path to failure: where are bankruptcy studies at now? Journal of Business and
Cybinski, P., 2001. Description. Explanation, Prediction: the Evolution of Bankruptcy Studies? Managerial
Finance 27 (4), 29–44.
Dambolena, I., Khoury, S., 1980. Ratio stability and corporate failure. Journal of Finance 33 (4), 1017–1026.
Daubie, M., Meskens, N., 2002. Business failure prediction: a review and analysis of the literature, Working
Paper, Department of Productions and Operations Management, Catholic University of Mons (Belgium), pp.
1–15.
Davis, A.H.R., Huang, X., 2004. The stock performance of firms emerging from Chapter 11 and accidental
bankruptcy, Paper presented at the FMA Meeting, New Orleans, USA, pp. 6–9 October 2004.
Deakin, E., 1972. A discriminant analysis of predictors of business failure. Journal of Accounting Research 10
(1), 167–179.
Deakin, E., 1976. On the nature of the distribution of financial accounting ratios: some empirical evidence. The
Accounting Review 51 (1), 90–97.
Deakin, E., 1977. Business failure prediction: an empirical analysis. In: Altman, B., Sametz, A. (Eds.), Financial
crisis: institutions and markets in a fragile environment. Wiley, New York, pp. 72–98.
Declerc, M., Heins, B., Van Wymeersch, Ch. 1990. Flux financiers et prévision de faillite: une analyse
comportementale de l’entreprise, Paper presented at Colloque International sur «l’analyse financière par les
flux: mode ou modèle?», Institut d’ Administration des Entreprises, Nice (France), December, pp. 1–26.
Declerc, M., Heins, B., Van Wymeersch, Ch., 1991. The use of value added ratios in statistical failure prediction
models: some evidence on Belgian annual accounts, Paper presented at the 1991 Annual Congress of the
European Accounting Association, April 10-12 1991, Maastricht (The Netherlands), pp. 1–24.
DeFond, M.L., Jiambalvo, J., 1994. Debt covenant violation and manipulation of accruals. Journal of Accounting
and Economics 17, 145–176.
Degeorge, F., Patel, J., Zeckhauser, R., 1999. Earnings management to exceed thresholds. Journal of Business 72
(1), 1–33.
Dimitras, A., Zanakis, S., Zopoudinis, C., 1996. A survey of business failures with an emphasis on failure
prediction methods and industrial applications. European Journal of Operational Research 90 (3), 487–513.
Dirickx, Y., Van Landeghem, G., 1994. Statistical failure prevision problems. Tijdschrift voor Economie en
Management 39 (4), 429–462.
Doumpos, M., Zopoudinis, C., 1999. A multicriteria discrimination method for the prediction of financial distress:
the case of Greece. Multinational Finance Journal 3 (2), 71–101.
Edmister, R., 1972. An empirical test of financial ratio analysis for small business failure prediction. Journal of
Financial and Quantitative Analysis, 1477–1493.
Eisenbeis, R.A., 1977. Pitfalls in the application of discriminant analysis in business. Journal of Finance 32 (3),
875–900.
Ezzamel, M., Mar-Molinero, C., 1990. The distributional properties of financial ratios in UK manufacturing
companies. Journal of Business Finance and Accounting 17 (1), 1–29.
Flagg, J.C., Giroux, G.A., Wiggins, C.E., 1991. Predicting corporate bankruptcy using failing firms. Review of
Financial Economics 1 (1), 67–79 (Fall).
Gentry, J.A., Newbold, P., Whitford, D.T., 1987. Funds flow components, financial ratios and bankruptcy. Journal
of Business Finance & Accounting 14 (4), 595–606.
Gentry, J.A., Newbold, P., Whitford, D.T., 1985a. Classifying bankrupt firms with funds flow components.
Journal of Accounting Research 2 (1), 146–160.
Gentry, J.A., Newbold, P., Whitford, D.T., 1985b. Predicting bankruptcy: if cash flow’s not the bottom line, What
is? Financial Analysts Journal 41 (5), 47–56.
Gilbert, L.R., Menon, K., Schwartz, K.B., 1990. Predicting bankruptcy for firms in financial distress. Journal of
Business Finance and Accounting 17 (1), 161–171.
Gloubos, G., Grammatikos, T., 1988. The success of bankruptcy prediction models in Greece. Studies in Banking
and Finance 7, 37–46.
Gombola, M.J., Ketz, J., 1983. Note on cash flow and classification patterns of financial ratios. Accounting
Review 58 (1), 105–114.
Gombola, M., Haskins, M., Ketz, J., Williams, D., 1987. Cash flow in bankruptcy prediction. Financial
Management, 55–65.
Gujarati, 2003. Basic Econometrics, Fourth Ed. McGraw-Hill, London.
Hall, G., 1994. Factors distinguishing survivors from failures amongst small firms in the UK construction sector.
Journal of Management Studies 31 (5), 737–760.
Hand D.J., 2004. Marginal classifier improvement and reality, Paper presented at the Symposium on Data Mining,
Ghent University (Belgium), May 2004.
Hayden, E., 2003. Are credit scoring models sensitive with respect to default definitions? Evidence from the
Austrian Market, Dissertation Paper, Department of Business Administration, University of Vienna, Austria,
pp. 1–43.
Hill, N.T., Perry, S.E., Andes, S., 1996. Evaluating firms in financial distress: an event history analysis. Journal of
Applied Business Research 12 (3), 60–71.
Hosmer, D.W., Lemeshow, S., 1989. Applied logistic regression. Wiley, New York.
Hsieh, S., 1993. A note on the optimal cutoff point in bankruptcy prediction models. Journal of Business Finance
& Accounting 20 (3), 457–464.
Johnson, C.G., 1970. Ratio analysis and the prediction of firm failure. Journal of Finance, 1166–1168.
Jones, F.L., 1987. Current techniques in bankruptcy prediction. Journal of Accounting Literature 6, 131–164.
Joos, P., Vanhoof, K., Ooghe, H., Sierens, N. 1998b. Credit classification: a comparison of logit models and
decision trees, Proceedings Notes of the Workshop on Application of Machine Learning and Data Mining in
Finance, 10th European Conference on Machine Learning, April, 24, Chemnitz (Germany), pp. 59–72.
Joy, O.M., Tollefson, J.O., 1975. On the financial applications of discriminant analysis. Journal of Financial and
Quantitative Analysis 10 (5), 723–739.
Joy, O.M., Tollefson, J.O., 1978. Some clarifying comments on discriminant analysis. Journal of Financial and
Quantitative Analysis, 197–200.
Kahya, E., Theodossiou, P., 1996. Predicting corporate financial distress: a time-series CUSUM methodology,
Paper presented at the Third Annual Conference of the Multinational Finance Association, June, 1996, pp.
1–38.
Karels, G.V., Prakash, A.J., 1987. Multivariate normality and forecasting of business bankruptcy. Journal of
Business Finance & Accounting 14 (4), 573–593.
Keasey, K., Watson, R., 1986. The prediction of small company failure: some behavioural evidence for the UK.
Accounting and Business Research 17, 49–57.
Keasey, K., Watson, R., 1987. Non-financial symptoms and the prediction of small company failure: a test of
Argenti’s hypotheses. Journal of Business Finance & Accounting 14 (3), 335–354.
Keasey, K., McGuinness, P., 1990. The failure of UK industrial firms for the period 1976–1984, logistic analysis
and entropy measures. Journal of Business Finance & Accounting 17 (1), 119–135.
Keasey, K., Watson, R., 1991. Financial distress models: a review of their usefulness. British journal of
Keasey, K., McGuinness, P., Short, H., 1990. Multilogit approach to predicting corporate failure: further analysis
and the issue of signal consistency. Omega International Journal of Management Science 18 (1), 85–94.
Koh, H.C., 1992. The sensitivity of optimal cutoff points to misclassification costs of Type I and Type II errors in
the going-concern prediction context. Journal of Business Finance & Accounting 19 (2), 187–197.
Lachenbruch, P.A., 1975. Discriminant Analysis. Hafner Press, New York.
Laitinen, E.K., 1991. Financial ratios and different failure processes. Journal of Business Finance and Accounting
18 (5), 649–673.
Laitinen, E.K., 1992. Prediction of failure of a newly founded firm. Journal of Business Venturing 7, 323–340.
Laitinen, E.K., 1993. Financial predictors for different phases of the failure process. Omega International Journal
of Management Science 21 (2), 215–228.
Laitinen, E.K., 1994. Traditional versus operating cash flow in failure prediction. Journal of Business Finance and
Accounting 21 (2), 195–217.
Laitinen, T., Kankaanpää, M., 1999. Comparative analysis of failure prediction methods: the Finnish case.
European Accounting Review 8 (1), 67–92.
Lane, W.R., Looney, S.W., Wansley, J.W., 1986. An application of the Cox proportional hazards model to bank
failure. Journal of Banking and Finance 10, 511–531.
Lehmann, B., 2003. Is it worth the while? The relevance of qualitative information in credit rating, Working
Paper presented at the EFMA 2003 Meetings, Helsinki (Finland), 25–29 June, 2003 pp. 1–25.
Luoma, M., Laitinen, E.K., 1991. Survival analysis as a tool for company failure prediction. Omega International
Journal of Management Science 19 (6), 673–678.
Lussier, R.N., 1995. A nonfinancial business success versus failure prediction model for young firms. Journal of
Small Business Management 33 (1), 8–20.
Lussier, R.N., Corman, J., 1994. A success vs. failure prediction model of the manufacturing industry, Paper
presented at the Conference of the Small Business Institute Director’s Association, San Antonio, Texas,
February 1994, Paper number, 48, pp. 1–5.
Maddala, G.S., 1977. Econometrics. McGraw-Hill, New York.
Maltz, A.C., Shenhar, A.J., Reilly, R.R., 2003. Beyond the balanced scorecard: refining the search for
organizational success measures. Long Range Planning 36, 187–204.
McLeay, S., Omar, A., 2000. The sensitivity of prediction models tot the non-normality of bounded an unbounded
financial ratios. British Accounting Review 32, 213–230.
Mensah, Y.M., 1983. The differential bankruptcy predictive ability of specific price-level adjustments: some
empirical evidence. Accounting Review 58 (2), 228–246.
Mensah, Y.M., 1984. An examination of the stationarity of multivariate bankruptcy prediction models: a
methodological study. Journal of Accounting Research 22 (1), 380–395.
Micha, B., 1984. Analysis of business failures in France. Journal of Banking and Finance 8, 281–291.
Moyer, R.C., 1977. Forecasting financial failure: A re-examination. Financial Management 6 (1), 11–17.
Mossman, Ch.E., Bell, G.G., Swartz, L.M., Turtle, H., 1998. An empirical comparison of bankruptcy models.
Financial Review 33 (2), 35–54.
Moses, D., Liao, S.S., 1987. On developing models for failure prediction. Journal of Commercial Bank Lending
69, 27–38.
Ohlson, J., 1980. Financial ratios and the probabilistic prediction of bankruptcy. Journal of Accounting Research
18 (1), 109–131.
Ooghe, H., Verbaere, E., 1985. Predicting business failure on the basis of accounting data: The Belgian
experience. The International Journal of Accounting 9 (2), 19–44.
Ooghe, H., Joos, P., 1990. Failure prediction, explanation of misclassifications and incorporation of other relevant
variables: result of empirical research in Belgium, Working paper, Department of Corporate Finance, Ghent
University (Belgium).
Ooghe, H., Balcaen, S., 2002. Are failure prediction models transferable from one country to another? An
empirical study using Belgian financial statements, Proceedings of the 9th Annual Conference of the
Multinational Finance Society, 30/06/02-03/07/02, Cyprus.
Ooghe, H., Balcaen, S., 2004. Alternative methodologies in studies on business failure: do they produce better
results than the classical statistical methods? Working Paper Series, Faculty of Economics and Business
Administration, Ghent University (Belgium), Paper number, 04/249 June 2004.
Ooghe, H., Joos, P., De Vos, D., 1993. Risico-indicator voor een onderneming aan de hand van falingspredictie-
modellen. Accountancy en Bedrijfskunde Kwartaalschrift 18 (3), 3–26.
Ooghe, H., Joos, P., De Vos, D., De Bourdeaudhuij, C., 1994a. Towards an improved method of evaluation of
financial distress models and presentation of their results, Working Paper, Department of Corporate Finance,
Ghent University (Belgium), January, pp. 1–22.
Ooghe, H., Joos, P., De Bourdeaudhuij, C., 1995. Financial distress models in Belgium: the results of a decade of
empirical research. International Journal of Accounting 30, 245–274.
Ooghe, H., Camerlynck, J., Balcaen, S., 2003. The Ooghe-Joos-De Vos failure prediction models: a cross-
industry validation. Brussels Economic Review 46 (1), 39–70.
Peel, M.J., Peel, D.A., 1987. Some further empirical evidence on predicting private company failure. Accounting
and Business Research 18, 57–66.
Piesse, J., Wood, D., 1992. Issues in assessing MDA models of corporate failure: a research note. British
Accounting Review 24, 33–42.
Platt, H.D., Platt, M.B., 1990. Development of a class of stable predictive variables: the case of bankruptcy
prediction. Journal of Business Finance & Accounting 17 (1), 31–51.
Platt, H.D., Platt, M.B., 1991. A note on the use of industry-relative ratios in bankruptcy prediction. Journal of
Platt, H.D., Platt, M.B., 2002. Predicting corporate financial distress: reflections on choice-based sample bias.
Journal of Economics and Finance 26 (2), 184–199.
Platt, H.D., Platt, M.B., 2004. Industry-relative ratios revisited: the case of financial distress, Paper presented at
the FMA 2004 Meeting, New Orleans (USA), pp. 6–9, 2004.
Platt, H.D., Platt, M.B., Pedersen, J.G., 1994. Bankruptcy discrimination with real variables. Journal of Business
Finance & Accounting 21 (4), 491–510.
Pompe, P., Bilderbeek, J., 2000. Faillissementspredictie: een vergelijking tussen lineaire discriminant analyse en
neurale netwerken. Economisch Sociaal Tijdschrift 2, 215–242.
Richardson, F.M., Davidson, L.F., 1984. On linear discrimination with accounting ratios. Journal of Business
Richardson, F.M., Kane, G.D., Lobingier, P., 1998. The impact of recession on the prediction of corporate failure.
Journal of Business Finance and Accounting 2 (1-2), 167–186.
Rosner, R.L., 2003. Earnings manipulation in failing firms. Contemporary Accounting Research 20 (2), 361–408.
Scott, E., 1978. On the financial applications of discriminant analysis: comment. Journal of Financial and
Quantitative Analysis, 201–205.
Scott, J., 1981. The probability of bankruptcy: a comparison of empirical predictions and theoretical models.
Journal of Banking and Finance 5 (3), 317–344.
Sharma, D.S., Iselin, E.R., 2003. The relative relevance of cash flow and accrual information for solvency
assessments: a multi-method approach. Journal of Business Finance and Accounting 30 (7–8), September,
1115–1140.
Sheppard, J.P., 1994. Strategy and bankruptcy: an exploration into organizational death. Journal of Management
20 (4), 795–833.
Shumway, T., 1999. Forecasting bankruptcy more accurately: a simple hazard model, Working Paper, University
of Michigan Business School, USA, July, pp. 1–31.
Slowinski, R., Zopudinis, C., 1995. Application of the rough set approach to evaluation of bankruptcy risk.
Intelligent Systems in Accounting, Finance and Management 4, 27–41.
Steele, A., 1995. Going concern qualifications and bankruptcy prediction, Paper presented at the 18th Annual
Congress of the European Accounting Association, Birmingham (UK), 10–12 May, pp. 1–28.
Swanson, E., Tybout, J., 1988. Industrial bankruptcy determinants in Argentina. Studies in Banking and Finance
7, 1–25.
Sweeney, A.P., 1994. Debt-covenant violations and managers’ accounting responses. Journal of Accounting and
Economics 17 (3), 281–308.
Taffler, R.J., 1982. Forecasting company failure in the UK using discriminant analysis and financial ratio data.
Journal of the Royal Statistical Society 145, 342–358 (Part 3).
Taffler, R.J., 1983. The assessment of company solvency and performance using a statistical model. Accounting
and Business Research 15 (52), 295–307.
Taffler, R.J., 1984. Empirical models for the monitoring of UK corporations. Journal of Banking and Finance 8,
199–227.
Taffler, R.J., Tisshaw, H., 1977. Going, Going, Gone—Four Factors Which Predict. Accountancy 88, 50–54.
Taffler, R.J., Agarwal, V., 2003. Do statistical failure prediction models work ex ante or only ex post?, Paper read
in the Deloitte & Touche Lecture Series on credit risk, University of Antwerp (Belgium), February.
Tamari, M., 1966. Financial ratios as a means of forecasting bankruptcy. Management International Review 4,
15–21.
Theil, H., 1971. Principles of econometrics. J. Wiley and Sons, New York.
Theodossiou, P., 1993. Predicting shifts in the mean of a multivariate time series process: an application in
predicting business failures. Journal of the American Statistical Association 88 (422), 441–449.
Tucker, J., 1996. Neural networks versus logistic regression in financial modelling: a methodological comparison,
Proceedings of the 1996 World First Online Workshop on Soft Computing (WSC1), Nagoya University
(Japan), August, pp. 19–30.
Van Caillie, D., 1999. Business failure prediction models: what is the theory looking for? Paper presented at the
Second International Conference on Risk and Crisis Management, Liège (Belgium), May, pp. 1–14.
Van Caillie, D., Arnould, S., 2001. The follow-up of financial value creation indicators to prevent bankruptcy in
Belgian SMEs: an empirical multivariate approach, Paper presented at the 2001 Babson College Research
Conference on Entrepreneurship, Jonköping (Sweden), June.
Van Caillie, D., Dighaye, A., 2002. The concept of «economic added result», a new tool to prevent bankruptcy?
Paper presented at the European Accounting Association Congress, Copenhagen (Denmark), April, pp. 1–30.
van Frederikslust, R.A.I., 1978. Predictability of corporate failure: models for prediction of corporate failure and
for evaluation of corporate debt capacity. PhD thesis in Economic Sciences, Martinus Nijhoff Social Science
Division, Leiden/Boston, Erasmus University, Rotterdam (The Netherlands).
Van Wymeersch, Ch., Wolfs, A. 1996. La ‘trajectoire de faillite’ des entreprises: une analyse chronologique sur
base des comptes annuels, Working Paper, Faculté des Sciences Economiques et Sociales, Université Notre-
Dame de la Paix (Belgium), Paper number 172, pp. 1–32.
Ward, T.J., Foster, B.P., 1997. A note on selecting a response measure for financial distress. Journal of Business
Watts, R.L., Zimmerman, J.L., 1986. Positive Accounting Theory. Prentice Hall Publishing Company.
Wilcox, J., 1971. A simple theory of financial ratios as predictors of failure. Journal of Accounting Research 9 (2),
389–395.
Wood, D., Piesse, J., 1987. The information value of MDA based financial indicators. Journal of Business Finance
and Accounting 14 (1), 27–38.
Zavgren, C., 1983. The prediction of corporate failure: the state of the art. Journal of Accounting Literature 2,
1–000mm37.
Zavgren, C.V., 1985. Assessing the vulnerability to failure of American industrial firms: a logistic analysis.
Journal of Business Finance and Accounting 12 (1), 19–45.
Zmijewski, M.E., 1984. Methodological issues related to the estimation of financial distress prediction models.
Journal of Accounting Research 22 (Suppl.), 59–86.

35 Years of Studies On Business Failure - An Overview of The Classic Statistical Methodologies and Their Related Problems

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

35 Years of Studies On Business Failure - An Overview of The Classic Statistical Methodologies and Their Related Problems

Uploaded by

Copyright:

Available Formats

The British Accounting Review 38 (2006) 63–93

35 years of studies on business failure: an overview

Received 19 November 2004; revised 6 September 2005; accepted 7 September 2005

* Corresponding author. Tel.: C32 9 264 35 07.

Keywords: Business failure; Failure prediction models; Overview; Problem topics

2. Classic statistical failure prediction models

2.1. Univariate analysis

2.2. Risk index models

2.3. Multiple discriminant analysis

differences in the dispersion matrices. Instead, researchers transform the data so as to

2.4. Conditional probability models

Method Failure prediction models

2.5. Overview of classic statistical business failure prediction models

Table 1 provides an overview of a large number of academically developed classic

3. Problems related to the classic statistical methods

3.1. Problems related to the classical paradigm

72 S. Balcaen, H. Ooghe / The British Accounting Review 38 (2006) 63–93

3.1.1. Arbitrary definition of failure

is developed on a group of risky, failing firms and a group of non-risky, surviving

3.1.2. Non-stationarity and data instability

3.1.3. Sampling selectivity

classification accuracy is different from the ex-ante predictive performance. In other

3.1.4. Choice of the optimization criteria

3.2. Problems related to the neglect of the time dimension of failure

failure prediction models have a retrospective character. What is demonstrated is that

3.3. Problems related to the application focus

3.3.1. Variable selection

3.3.2. Selection of a modelling method

3.4. Other problems

3.4.1. The linearity assumption

Similarly, MDA involves a linear classification rule. A discriminant score above

3.4.2. The use of annual account information

3.4.3. The neglect of the multidimensional nature of failure

Problem category Problem topics Related problem issues

the financial status of a company is a complex, multidimensional concept, which cannot be

a misleading classification power and a weak predictive usefulness in practice. A second

You might also like