Journal of Small Business and Enterprise Development

Emerald Article: The bias of unhealthy SMEs in bankruptcy prediction models

J. Samuel Baixauli, Antonina Mdica-Milo
Article information:
To cite this document:
J. Samuel Baixauli, Antonina Mdica-Milo, (2010),"The bias of unhealthy SMEs in bankruptcy prediction models", Journal of Small
Business and Enterprise Development, Vol. 17 Iss: 1 pp. 60 - 77
Permanent link to this document:
http://dx.doi.org/10.1108/14626001011019134
Downloaded on: 26-06-2012
References: This document contains references to 20 other documents
To copy this document: permissions@emeraldinsight.com
This document has been downloaded 1237 times since 2010. *
Users who downloaded this Article also downloaded: *
James DeLisle, Terry Grissom, (2011),"Valuation procedure and cycles: an emphasis on down markets", Journal of Property
Investment & Finance, Vol. 29 Iss: 4 pp. 384 - 427
http://dx.doi.org/10.1108/14635781111150312
Franois Des Rosiers, Jean Dub, Marius Thriault, (2011),"Do peer effects shape property values?", Journal of Property
Investment & Finance, Vol. 29 Iss: 4 pp. 510 - 528
http://dx.doi.org/10.1108/14635781111150376
David Wyman, Maury Seldin, Elaine Worzala, (2011),"A new paradigm for real estate valuation?", Journal of Property Investment &
Finance, Vol. 29 Iss: 4 pp. 341 - 358
http://dx.doi.org/10.1108/14635781111150286
Access to this document was granted through an Emerald subscription provided by BIBLIOTECA CENTRALA UNIVERSITARA EUGEN TODORAN TIM
For Authors:
If you would like to write for this, or any other Emerald publication, then please use our Emerald for Authors service.
Information about how to choose which publication to write for and submission guidelines are available for all. Please visit
www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
With over forty years' experience, Emerald Group Publishing is a leading independent publisher of global research with impact in
business, society, public policy and education. In total, Emerald publishes over 275 journals and more than 130 book series, as
well as an extensive range of online products and services. Emerald is both COUNTER 3 and TRANSFER compliant. The organization is
a partner of the Committee on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative for digital archive
preservation.
*Related content and download information correct at time of download.
The bias of unhealthy SMEs in
bankruptcy prediction models
J. Samuel Baixauli and Antonina Modica-Milo
Department of Management and Finance, University of Murcia,
Espinardo, Spain
Abstract
Purpose This paper aims to construct a nancial health indicator to dene the degree of nancial
health in order to decontaminate the estimation sample and to make predictions that are not biased by
unhealthy rms.
Design/methodology/approach The binomial logit model is used to examine the likelihood that
a rm will go bankrupt. In order to evaluate the accuracy of the estimated models, measures proposed
by the Basel Committee on Banking Supervision are applied: cumulative accuracy prole (CAP) and
the receiver operating characteristics (ROC).
Findings The proposed nancial health indicator permits the heterogeneity of the rms to be
reduced as well as identifying a strong rm sample to estimate the bankruptcy probability accurately.
Originality/value A drawback of all bankruptcy prediction models comes from the fact that
bankruptcy is an example of a homogeneous observable qualitative response while non-bankruptcy
would be expected to be represented by a healthy rm. However, the non-bankruptcy rms are
heterogeneous and their actual probabilities of bankruptcy are non-observable. The article adds to the
previous literature on SMEs bankruptcy prediction by using a nancial health indicator to construct
the estimation sample and to make accurate bankruptcy predictions.
Keywords Bankruptcy, Business failures, Mathematical modelling
Paper type Research paper
Introduction
The small and medium-sized enterprise (SME) sector is often viewed as the incubator
of employment, innovation and growth (Craig et al., 2003). The reports by the
observatory of European SMEs use statistics on the number of enterprises, total
employment and production by rm size to provide an overview of the current
situation in the SME sector in Europe. Regardless of how they are measured, not only
are most enterprises in Europe small, but they also account for a signicant amount of
European work experience and economic activity. For instance, in 2003 there were
more than 19 million enterprises in Europe, providing jobs for almost 140 million
people. In contrast, there were only about 40,000 large enterprises in existence, which
accounted for only 0.2 per cent of all enterprises. So, the vast majority of enterprises in
Europe are SMEs. Zingales (2000) points out that, empirically, the attention paid to
large companies has lead researchers to ignore the rest of the young and small rms,
which do not have access to public markets.
SMEs are nancially more constrained than large rms and are less likely to have
access to the capital market. Nowadays, due to the lack of capital market data, the bank
sector uses techniques based on models where nancial ratios are combined and
The current issue and full text archive of this journal is available at
www.emeraldinsight.com/1462-6004.htm
The authors thank Fundacion Cajamurcia and the Spanish Government (Project ECO
2008-02846) for nancial support.
JSBED
17,1
60
Journal of Small Business and
Enterprise Development
Vol. 17 No. 1, 2010
pp. 60-77
qEmerald Group Publishing Limited
1462-6004
DOI 10.1108/14626001011019134
weighted to produce a probability of bankruptcy in order to estimate the bankruptcy
probability. Bankruptcy is disruptive and costly to owners, investors and
communities. Despite the existence of many contributions, the bankruptcy
prediction continues to be an important issue. As Altman and Saunders (1998) point
out, an increase in the number of bankruptcies, more competitive margins on loans,
and a trend towards disintermediation by the highest quality and largest borrowers
have led to the development of new and more sophisticated credit-scoring and
early-warning systems. Altman and Saunders (1998) review new models of credit risk
measurement, for example the KMV model, models based on the term structure of yield
spreads, mortality-default rate models, and neural network models. These models are
based on capital market data and quoted rms. However, the difculty of SMEs in
accessing the capital market restricts the use of such models. In Europe only a very
small percentage of SMEs are quoted rms.
Many bankruptcy prediction models thatcombine nancial ratios have been
proposed in the literature. Of these, the models of Beaver (1968), Altman (1968) and
Altman et al. (1977) are worthy of note. A drawback of all bankruptcy prediction
models comes from the fact that bankruptcy is an example of a homogenous
observable qualitative response, while non-bankruptcy would be expected to be
represented by a healthy rm. However, non-bankruptcy rms are heterogeneous and
their degree of nancial health is non-observable.
Our article adds to the previous literature on SMEs bankruptcy prediction by
identifying the characteristics of a healthy rm and by evaluating the effects of using
or not using these characteristics in the sample design on the accuracy of the
bankruptcy prediction models. In the evaluation process, statistical methodologies
recommended by the Basel Committee on Banking Supervision are used. Our proposal
is compared with the traditional methodology based on unbalanced samples and on
matching by size and industry.
The structure of the paper is as follows: the next section discusses bankruptcy
prediction models and the denition of healthy rms in order to exclude rms with
high probability of bankruptcy. The subsequent section outlines methods used to
analyse the data. The next section covers the data collection and discusses the ndings
of the analysis. The nal section concludes with a discussion of the ndings.
Bankruptcy prediction and the bias of non-bankruptcy rms
According to Berryman (1983) there are different denitions of failure:
.
earning a rate of return signicantly and continually below prevailing rates on
similar investments;
.
ceased operations; or
.
termination for any reason.
When the use of bankruptcy prediction models is the main issue, failure is dened as
only those rms that are declared bankrupt, and non-bankrupt rms are any other
rms. Such denitions mean that rms with a much more general denition of
nancial distress or failure are also included among non-bankrupt rms. They have
high bankruptcy probability even though they cannot be classied as bankrupt rms.
For example, a failure denition could include discontinued ownership (ownership
The bias of
unhealthy SMEs
61
changes), business discontinuance (the business ceases to exist) or failing to make a
go of it.
Often, rms that satisfy the wide denition of failure are not bad rms to invest in.
In this sense, a change in ownership may not imply failure because the business may
continue under a different owner. Watson and Everett (1999) believe that
discontinuance may signal actual failure because business resources may have been
reallocated to more protable areas. Headd (2003) makes a compelling argument that
discontinuance may not be associated with failure. Many nancially strong rms may
cease to continue because of an existing strategy. The meaning of discontinuance is too
wide and includes rms that sell because they are successful and are offered a good
price and rms that are taken over. In this sense, Cochran (1986) points out that
businesses may be discontinued because of extraneous factors such as retirement or
illness, because alternative opportunities present themselves, or under some denitions
of discontinuance, even because the business is sold at a prot. As Watson and Everett
(1996) point out, the broader the denition of failure, the higher the failure rate for
small rms. In order to show that bankruptcy probabilities are unobservable variables,
Carter and Van Auken (2006) dene failure as only those rms that declare
bankruptcy. All this evidence gives risk managers and researchers a range of rms
with different degrees of failure and different bankruptcy probability. In fact, Watson
and Everett (1999) test the relation of the probability of failure with the years of life of a
business and with the barriers to entry in an industry sector by considering three
different denitions of failure (bankruptcy, discontinuance of ownership and failing to
make a go of it).
The sample design employed by many research studies has been to match a set of
bankrupt rms with the same number of non-bankrupt rms, often controlling for size
and industry (see Altman, 1968; Theodossiou et al., 1996; Carter and Van Auken, 2006,
among others). Matching unhealthy non-bankrupt rms with bankrupt rms in order
to estimate the parameters of a bankruptcy prediction model produces a bias in the
estimated parameters and a reduction in the ability to predict bankruptcy accurately.
As a consequence, it is necessary to make a previous selection in order to include
non-bankrupt rms with good nancial health in a model.
As the degree of nancial health is unobservable, in this article an indicator of
nancial strength is constructed. If this indicator is positively correlated with the
nancial health, then to match non-bankruptcy rms characterised by having high
values of the nancial health indicator with bankrupt rms should allow us to estimate
bankruptcy prediction models with less prediction error than the models that match
non-bankrupt rms characterised by low values of the nancial health indicator with
bankrupt rms.
Statistical records are satisfactory for listing bankruptcies or discontinuances, but
they cannot cope with identifying the severity of failures. The denition of
non-bankruptcy should be devoid of unsuccessful rms with a low degree of nancial
health. The economic denitions of failure take protability as a common denominator.
In this line, Fredland and Morris (1976) state that any rm earning a rate of return on
investment that is less than the rms opportunity cost is a failure. Following the
denition of Altman (1968), a non-bankrupt rm is a failure if it is earning a rate of
return on invested capital which is signicantly and continually below prevailing rates
on similar investments. Causes of failure can be classied as endogenous (internal to
JSBED
17,1
62
the rm) and exogenous (external to the rm). Exogenous causes, such as exchange
rates or high interest rates, ascribe their failure to problems in operational management
and endogenous causes are widely related to poor operational management, high costs
of production, poor marketing policy and poor personal management (Hall and Young,
1991).
We select four measures of protability to construct our nancial health indicator:
.
the return on assets;
.
capital turnover;
.
return on equity; and
.
net worth variation.
As the prot measures may be managed gures subject to deliberate manipulations,
we add the auditing opinion:
.
Return on assets (FR
1
), measured by earnings before interest and taxes over total
assets. This variable measures the economic efciency of a rm, i.e. its capacity
to generate prots independently of its nancial structure, debt cost, taxes or the
protability required by the owners.
.
Capital-turnover ratio (FR
2
), measured by the sales over total assets. This ratio is
a standard nancial ratio illustrating the ability of the rms assets to generate
sales. It is a measure of managements capability in dealing with competitive
conditions.
.
Return on equity (FR
3
), measured by the net income to net worth. This variable
measures the nancial protability, which depends on the economic protability
as well as on the nancial structure, debt cost and the protability required by
the owners. A positive return on equity reects the opportunity for retaining
earnings.
.
Change of net worth (FR
4
), measured by the net worth annual percentage change.
This variable represents the solvency increase or decrease.
.
Auditing opinion (d
5
): this measures the quality of the nancial statements.
Edmister (1972) tested the same technique as Altman (1968) for small business
with some success but warned that three consecutive nancial statements of
good quality must be available for analysis of a small rm.
The expression of the nancial health indicator is given by:
FS
X
4
j1
I FR
j
$ P
5
d
5
; 1
where FR is the nancial ratio, P
5
is the fth percentile, d
5
is a dummy variable which
takes the value 1 if the auditing opinion is favourable and 0 otherwise, and I(.) is the
indicator function, which takes the value 1 if FR is higher than or equal to P
5
and 0
otherwise.
If this is a good indicator of the bankruptcy probability of healthy rms, it will
allow us to form decontaminated samples and the prediction models will give much
more accurate bankruptcy predictions. If a rm shows a value of the nancial health
The bias of
unhealthy SMEs
63
indicator below the fth percentile, it reveals that it has an abnormal reduced indicator
value, which means that it is nancially a weak rm. It is expected that, if FS is a good
proxy of nancial health, using rms with an FS value equal to 5 (healthy rms)
should allow us to predict bankruptcy more accurately than using rms with FS equal
to 0 (unhealthy rms).
The indicator FS is evaluated on several bankruptcy prediction models to avoid
results in which it could be inuenced by the selected model. The models we use are
those of Beaver (1966), Altman (1968), and Altman et al. (1977). These models are the
most widely cited bankruptcy prediction models in the credit risk literature. Also,
Garc a et al. (1997) presented a bankruptcy prediction model that represents the
reference in the Spanish market. All the models above examine multivariate ratios in
large and small rms. They are displayed in equations (2)-(5):
P
ij
f CASH
i
; NIN
i
; LEV
i
; WC
i
; LIQ
i
; NCRE
i
Beaver; 1966; 2
P
ij
f WC
i
; CP
i
; ROA
i
; CAP
i
; ACT
i
Altman; 1968; 3
P
ij
f ROA
i
; S
ROA
i
; DS
i
; CP
i
; LIQ
i
; CAP
i
; SIZE
i
Altman et al:; 1977; 4
P
ij
f QLIQ
i
; ATD
i
; INT
i
; AMO
i
; EAR
i
Garcia et al:; 1997: 5
The ratios employed are:
.
Beaver (1966) CASH, cash ow to total debt; NIN, net income to total assets;
LEV, total debt to total assets; WC, working capital over total assets; LIQ,
current assets to current debt; NCRE, no-credit interval measured as quick assets
minus current liabilities to operating expenses minus depreciation, depletion,
and amortisation.
.
Altman (1968) WC, working capital over total assets; CP, retained earnings
over total assets; ROA, earnings before interest and taxes over total assets; CAP,
for application purposes a ve-year average of the total market value equity over
total capital has been substituted by book value equity over total capital; ACT,
sales over total assets.
.
Altman et al. (1977) ROA, earnings before interest and taxes over total assets;
S
ROA
, standard error of estimating of ROA; DS, earnings before interest and
taxes over total interest payments; CP, retained earnings over total assets; LIQ,
current assets to current debt; CAP, for application purposes a ve-year average
of the total market value equity over total capital has been substituted by book
value equity over total capital; SIZE, logarithmic transformation of total asset.
.
Garc a et al. (1997) QLIQ, quick assets over current libialities; ATD, total asset
to total debt; INT, total interest payments to sales; AMO, annual amortization to
amortizable assets; EAR, earnings before taxes over total debt.
Methodology
We use the binomial logit model to examine the likelihood that a rm will go bankrupt.
Logit analysis uses a set of nancial variables to predict bankruptcy probability,
assuming that bankruptcy probability is logistically distributed, i.e. the cumulative
JSBED
17,1
64
bankruptcy probability takes a logistic functional form and is, by denition,
constrained to fall between 0 and 1. In the prediction evaluation process we distinguish
between the in-sample validation and the out-of-sample validation, in the estimation
year, when applying the model to the non-used sample in the estimation process, as
well as in the next two years to the whole population of rms.
Under non-normality, we choose the logistic regression model with maximum
likelihood estimators. The regression model is specied as follows:
P
ij

1
1 e
2b
0
X
; 6
where P
ij
is the probability that rm i is bankrupt, X is a vector of measured
characteristics for rm i, and b is the unknown parameter vector.
The estimates of the parameters of this model yield the bankruptcy probabilities for
a given rm. The vector of measured characteristics used contains the list of empirical
variables described in equations (2)-(5). The nancial ratios are calculated at the end of
the year prior to the bankrupt year, except the standard error of estimating of ROA of
Altman et al. (1977), which is computed over seven years. If the model, as represented
by the likelihood ratio statistic, indicates that the model ts the data signicantly, we
then move on to interpret parameter estimates. We also compute the McFadden R
2
.
In order to classify the rms belonging to the sample under the estimated
bankruptcy probabilities, a cut-off must be xed which allows the sample to be
dichotomise and the prediction errors to be determined. The type I error is dened as
those non-bankrupt rms which have already been classied as bankrupt and the type
II error is dened as those bankrupt rms which have been classied as non-bankrupt.
The total error is obtained by weighting type I and type II errors.
Since type I and type II errors are conditioned by the selected cut-off, we apply two
evaluation measures proposed by the Basel Committee on Banking Supervision
(BCBS), which are constructed for all possible cuts-off:
(1) the cumulative accuracy prole (CAP); and
(2) the receiver operating characteristics (ROC).
Given a variety of rating methodologies, the question is which of these methodologies
deliver acceptable discriminatory power between bankrupt and non-bankrupt rms.
The Basel Committee on Banking Supervision (2005) has published a working paper
summarising statistical methodologies for assessing discriminatory power.
Accuracy ratio of cumulative accuracy prole
To obtain the cumulative accuracy prole (CAP) curve, all rms are rst ordered by
their respective scores, from riskiest to safest, i.e. from the rm with the highest score
to the rm with the lowest score. For a given fraction x of the total number of rms, the
CAP curve is constructed by calculating the percentage d(x) of the bankruptcy rms
whose rating scores are equal to, or higher than, the maximum score of fraction x. This
is done for x ranging from 0 per cent to 100 per cent. Figure 1 illustrates the CAP
curves.
A real credit model lies somewhere in between the two extremes of a perfect rating
credit model and a random model. In a perfect rating model, the CAP increases linearly
to 1 then remains constant, since a perfect rating model will assign the highest scores
The bias of
unhealthy SMEs
65
to bankrupt rms. In a random model, the fraction x of all rms with the highest rating
scores will contain x per cent of all bankrupt rms. The accuracy ratio (AR) is dened
as the ratio of the area between the real credit model and the random model a
r
, and the
area between the perfect rating model and the random model, a
p
. The closer the AR is
to 1, the better the rating model:
AR
CAP

a
r
a
p
: 7
Accuracy ratio of receiver operating characteristic
The decision as to which rms will not go bankrupt during the next period and which
rms will is made by introducing a cut-off value C, as in Figure 2. Thus, a rm with a
bankruptcy probability lower than C is classied as a non-bankrupt rm and a rm
with a bankruptcy probability higher than C is classied as a bankrupt rm.
Consequently, a hit rate HR(C) is dened as the fraction of bankrupt rms that are
classied correctly for a given cut-off value C:
Figure 1.
Cumulative accuracy
prole curves
Figure 2.
Distribution of
bankruptcy probability for
bankrupt and
non-bankrupt rms
JSBED
17,1
66
HRC
HC
N
B
: 8
In equation (8), H(C) is the number of bankruptcies predicted correctly with the cut-off
value C, and N
B
is the total number of bankruptcies in the sample. The false alarm rate
FAR(C) is dened as:
FARC
FC
N
NB
; 9
where F(C) is the number of non-bankrupt rms that were classied incorrectly as
bankrupt by using the cut-off C and N
NB
is the total number of non-bankruptcy rms.
In Figure 2, HR(C) is the area to the right of the cut-off value C under the probability
distribution of the bankrupt rms, while FAR(C) is the area to the right of C under the
probability distribution of the non-bankrupt rms.
The receiver operating characteristic (ROC) curve is a plot of HR(C) versus FAR(C),
illustrated in Figure 3. To construct the ROC curve, HR(C) and FAR(C) are computed
for different possible cut-off values in the range 0 to 1. Like the CAP curve, the larger
the area under the ROC curve, the better the model. The area is called AUROC.
In a random model with no discriminative power, AUROC is equal to 0.5. In a
perfect model AUROC is equal to 1. Engelmann et al. (2003) showed a measure between
0 to 1 as an accuracy ratio for ROC, which is computed using the following expression:
AR
ROC
2 AUROC 21 2
Z
1
0
HRFARdFAR 21: 10
Data and empirical results
The period is divided into two subperiods to estimate the bankruptcy prediction
models:
(1) from January 1994 to December 2000 we compute the nancial variables; and
(2) from January 2001 to December 2004 we estimate the models and evaluate their
predictive capacity.
Figure 3.
Receiver operating
characteristic curves
The bias of
unhealthy SMEs
67
Firms belonging to a nancial sector were excluded from the sample. Our sample
consists of rms which satisfy the European Commission denition of SMEs and
which have accounting data from January of 1994 to December 2004. Following the
European Commission denition, we consider SMEs as:
.
companies with fewer than 250 employees;
.
sales below e40 million; and
.
total assets under e27 million.
We use a database that includes nancial information of more than 190,000 Spanish
rms. This database is called SABI, and is managed by Informa SA.
In the database, 76 bankrupt rms were found during the period January 2001 to
December 2004. From the 76 bankrupt rms, 27 rms were deleted due to lack of
continuous information in at least one of the ve years previous to the bankruptcy
event, and one rm was deleted because it was not considered a SME. Therefore, the
nal sample is formed by 48 bankrupt rms. This gives a total sample of 2,211 rms in
2000, where 2,194 are non-bankrupt rms and 17 are bankrupt rms. In 2001, the
sample consists of 2,130 rms, where 2,113 are non-bankrupt rms and 17 are
bankrupt rms. In 2002 there are 2,299 rms, consisting of 2,292 non-bankrupt rms
and seven bankrupt rms. In 2003 there are a total of 2,310 rms, of which 2,303 are
non-bankrupt rms and seven are bankrupt rms.
We equate the number of non-bankrupt rms selected from each year to the number
of bankrupt rms for such year. We select three samples of non-bankrupt rms:
(1) the high nancial health sample (strong rms);
(2) the low nancial health sample (weak rms); and
(3) the classical sample (control rms).
We also use the whole sample to estimate the model (total rms). In the high nancial
health sample, the non-bankrupt rms are selected from the rms that obtain an
indicator value equal to 5. In the low nancial health sample, the non-bankrupt rms
are selected from those rms that obtained an indicator value equal to 0. In the classical
sample, the non-bankrupt rms are selected from the rms that belong to the same
sector as the bankrupt rms.
Using each of these four samples, in 2001 we estimate the bankruptcy prediction
models of Beaver (1966), Altman (1968), Altman et al. (1977) and Garc a et al. (1997).
Then, given the estimated bankruptcy probabilities, we measure the in-sample
prediction capacity and the out-of-sample predictive ability of such models. The
in-sample prediction capacity is measured using the rms that were not employed in
the estimation process in 2001, and the out-sample prediction capacity is computed
using the whole rm population in the two next years. We repeat the analyses,
re-estimating the models in 2002. We do not repeat the estimation in 2003 and 2004
because of the small size of the bankrupt sample.
In the estimation process, we estimate several logit models. The estimation results of
the Beaver (1966, 1968) model are reported in Table I. These results show that the
samples formed by weak rms and total rms have the lowest McFaddens R
2
in 2001,
while the control rm sample has a quite good value in 2001, taking into account this
type of data, but a very small one in 2002. The signicance of 22 log L the statistics
JSBED
17,1
68
2
0
0
1
Y
e
a
r
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
C
o
n
s
t
a
n
t
2
1
5
.
3
2
2
1
1
.
3
4
*
2
1
7
.
5
0
2
1
0
.
9
8
*
2
6
.
8
7
2
2
.
4
2
2
2
2
.
1
3
*
2
1
0
.
8
1
*
(
1
1
.
4
0
)
(
5
.
5
6
)
(
1
7
.
1
0
)
(
2
.
4
8
)
(
9
.
4
9
)
(
3
.
3
0
)
(
1
2
.
9
4
)
(
3
.
2
5
)
C
A
S
H
1
7
.
4
7
1
9
.
3
4
1
5
.
6
5
2
7
.
4
7
2
0
.
4
3
2
5
.
5
8
2
2
2
.
3
9
2
3
.
6
3
(
2
5
.
0
6
)
(
1
2
.
1
6
)
(
4
1
.
0
6
)
(
7
.
4
9
)
(
2
0
.
0
6
)
(
9
.
8
0
)
(
1
8
.
1
9
)
(
7
.
6
4
)
N
I
N
2
7
1
.
3
4
6
.
3
8
2
1
4
2
.
0
7
*
1
.
6
3
2
6
.
9
6
9
.
0
1
2
0
.
9
8
2
1
5
.
6
4
(
5
1
.
7
5
)
(
9
.
7
0
)
(
8
0
.
0
5
)
(
9
.
2
3
)
(
2
7
.
9
1
)
(
1
3
.
0
8
)
(
2
2
.
6
8
)
(
1
0
.
7
4
)
L
E
V
1
5
.
9
2
1
1
.
7
8
*
2
0
.
1
9
6
.
3
1
*
1
2
.
1
8
5
.
9
9
2
1
.
5
5
8
.
4
9
*
(
1
2
.
2
6
)
(
5
.
8
5
)
(
1
9
.
0
0
)
(
2
.
6
8
)
(
7
.
8
7
)
(
3
.
7
2
)
(
5
.
6
1
)
(
2
.
7
1
)
W
C
2
1
.
3
2
2
4
.
4
9
2
0
.
6
1
2
3
.
6
2
8
.
2
9
5
.
6
1
2
3
0
.
7
2
1
.
7
4
(
1
0
.
7
8
)
(
4
.
6
1
)
(
1
2
.
6
6
)
(
2
.
7
2
)
(
1
4
.
7
6
)
(
4
.
6
7
)
(
2
0
.
4
5
)
(
4
.
1
8
)
L
I
Q
3
.
1
4
1
.
6
5
4
.
1
1
1
.
6
8
*
2
2
.
4
4
2
2
.
2
5
2
3
.
5
6
2
0
.
3
0
(
4
.
6
5
)
(
1
.
6
4
)
(
4
.
3
9
)
(
0
.
8
7
)
(
7
.
7
5
)
(
1
.
9
0
)
(
1
3
.
8
8
)
(
2
.
2
1
)
N
C
R
E
2
.
6
6
0
.
8
2
0
.
8
2
2
0
.
3
8
2
2
.
0
4
2
1
.
5
3
2
3
.
2
4
2
0
.
2
8
(
4
.
1
3
)
(
1
.
2
2
)
(
5
.
2
3
)
(
0
.
9
0
)
(
5
.
2
6
)
(
2
.
2
7
)
(
3
.
7
6
)
(
1
.
0
7
)
2
2
l
o
g
L
2
2
.
6
1
1
6
.
4
2
3
1
.
8
2
5
8
.
6
9
2
4
.
0
5
9
.
3
2
1
4
.
3
8
6
5
.
7
5
x
2
p
-
v
a
l
u
e
0
.
0
0
0
0
.
0
1
1
7
0
.
0
0
0
0
0
.
0
0
0
0
0
.
0
0
0
5
0
.
1
5
6
5
0
.
0
2
5
7
0
.
0
0
0
0
M
c
F
a
d
d
e
n
R
2
0
.
4
7
9
7
0
.
3
4
8
4
0
.
6
7
5
1
0
.
2
9
4
3
0
.
5
1
0
3
0
.
1
9
7
7
0
.
3
0
5
1
0
.
3
3
1
9
A
c
c
u
r
a
c
y
i
n
-
s
a
m
p
l
e
P
e
r
c
e
n
t
a
g
e
c
l
a
s
s
i

e
d
8
5
.
2
9
7
9
.
4
1
9
1
.
1
8
9
9
.
4
1
8
8
.
2
4
6
4
.
7
1
7
6
.
4
7
9
9
.
3
0
T
y
p
e
I
e
r
r
o
r
(
p
e
r
c
e
n
t
)
1
7
.
6
5
2
3
.
5
3
1
1
.
7
6
0
.
0
0
1
1
.
7
6
4
1
.
1
8
2
3
.
5
3
0
.
0
5
T
y
p
e
I
I
e
r
r
o
r
(
p
e
r
c
e
n
t
)
1
1
.
7
6
1
7
.
6
5
5
.
8
8
7
6
.
4
7
1
1
.
7
6
2
9
.
4
1
2
3
.
5
3
8
2
.
3
5
A
R
C
A
P
0
.
9
2
9
8
0
.
8
7
2
8
0
.
9
1
4
0
0
.
7
8
7
9
0
.
9
5
2
4
0
.
7
9
8
1
0
.
8
5
0
6
0
.
7
8
2
9
A
R
R
O
C
0
.
7
6
2
5
0
.
7
5
8
6
0
.
7
1
6
5
0
.
3
0
7
9
0
.
7
9
0
1
0
.
5
7
3
3
0
.
8
4
6
1
0
.
3
1
9
9
N
o
t
e
s
:
T
h
e
s
a
m
p
l
e
i
s
c
o
m
p
o
s
e
d
o
f
1
7
b
a
n
k
r
u
p
t

r
m
s
.
T
h
e
d
e
p
e
n
d
e
n
t
v
a
r
i
a
b
l
e
t
a
k
e
s
t
h
e
v
a
l
u
e
o
n
e
f
o
r
t
h
e
1
7
b
a
n
k
r
u
p
t

r
m
s
a
n
d
z
e
r
o
f
o
r
t
h
e
n
o
n
-
b
a
n
k
r
u
p
t

r
m
s
.
S
t
a
n
d
a
r
d
e
r
r
o
r
s
a
r
e
g
i
v
e
n
i
n
p
a
r
e
n
t
h
e
s
e
s
.
*
S
i
g
n
i

c
a
n
t
a
t
t
h
e
5
p
e
r
c
e
n
t
l
e
v
e
l
Table I.
Estimates of the logit
target prediction model
based on the Beaver
(1966) variables
The bias of
unhealthy SMEs
69
behaves likewise. The percentage of correctly classied companies during the
estimation period is relatively high using the strong rm and the control rm samples.
The results for the total rmsample are over-valued due to bankrupt rms having a very
small weight in the total sample. On xing a cut-off equal to 0.5, the weak rm sample
has high type I and II errors. In the total sample, the type I error is insignicant while the
type II error is very high. The CAPand ROCmeasures indicate that a strong rmsample
allows the model to be estimated with the best in-sample predictive capacity.
The results concerning the Altman (1968), Altman et al. (1977) and Garc a et al.
(1997) models are likewise summarised in Tables II, III, and IV, respectively.
In general, Tables II IV illustrate that the strong rm and control rm samples
have the highest McFaddens R
2
in 2001 while the strong rm sample does so in 2002.
The percentage of correctly classied companies during the estimation period is
relatively high using the strong rms, from 82.35 per cent to 88.24 per cent. The CAP
and ROC measures point out that the strong rm sample allows the model to be
estimated with the best in-sample predictive capacity with values up to 0.9751. The
total rm sample has the highest type II error, up to 99.41 per cent, which means that to
predict that no rm is going to be bankrupt minimises the total error, since it
represents a reduced percentage of the total number of rms used in the estimation
process. On the other hand, the control rm sample shows different results depending
on both the year and the model. In this sense, the percentage correctly classied goes
from 64.71 per cent to 91.18 per cent, the type II error goes from 5.88 per cent to 35.29
per cent and the type I error goes from 5.88 per cent to 41.18 per cent. The CAP and
ROC values of the control rm sample are lower than the strong rm sample and
higher than the weak rm sample.
The out-of-sample predictive capacity of the models is reported in Tables V and VI.
When the total sample is considered to estimate the model, out-of-sample results
cannot be computed in the estimation year.
Concerning the strong rm sample, Table V reports that, in 2001, the type I error
ranges from 24.38 per cent to 29.72 per cent, the type II error is equal to 11.76 per cent in
all models and the CAP values go from 0.871 to 0.893 while the ROC values go from
0.7423 to 0.8520. These values do not worsen when the model is applied one or two
years after the estimation. In the worst case, the type I error rises to 33.35 per cent, the
type II error rises to 42.86 per cent, the CAP decreases to 0.801 and the ROC rises to
0.662. In general, the weak rm sample gives higher type I and II errors and lower ROC
and CAP values than the strong rm and the control rm samples. The control rm
sample has higher values of CAP and ROC as well as lower type I and II errors than the
weak rm sample. In contrast, the total rm sample has the highest type II error.
Table VI shows that, as in 2001, in 2002 the strong rm sample shows the most
accuracy. Type I error in 2002 goes from 28.16 per cent to 30.76 per cent, type II error
ranges from 5.88 per cent to 11.76 per cent in all the models and the CAP values go from
0.872 to 0.88 while ROC values go from 0.79 to 0.814. The results remain similar in 2003
and 2004. In the worst case, the type I error is 28.85 per cent, the type II error rises to 28.57
per cent, the CAP decreases to 0.785 and the ROC rises to 0.712. In general, the weak rm
sample gives higher type I and II errors and lower ROC and CAP values than the strong
rmandthe control rmsamples. It must be highlighted that the control rmsample has
higher values of CAPand ROCandlower type I and II errors than the weak rmsamples.
In contrast, the total rm sample has the highest type II error.
JSBED
17,1
70
2
0
0
1
2
0
0
2
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
C
o
n
s
t
a
n
t
2
.
2
4
*
0
.
7
5
5
.
2
2
*
2
3
.
0
6
*
1
.
6
0
2
1
.
5
1
0
.
1
9
2
2
.
7
6
*
(
1
.
2
8
)
(
0
.
9
2
)
(
2
.
2
1
)
(
2
.
2
1
)
(
1
.
5
8
)
(
1
.
1
0
)
(
1
.
1
6
)
(
0
.
6
4
)
W
C
1
.
8
2
2
1
.
6
2
1
.
6
2
2
2
.
1
6
4
.
2
0
5
.
3
2
3
.
2
1
1
.
0
0
(
3
.
0
7
)
(
2
.
1
4
)
(
7
.
0
2
)
(
1
.
4
4
)
(
4
.
4
8
)
(
3
.
4
7
)
(
2
.
6
4
)
(
1
.
8
1
)
C
P
2
0
.
3
5
1
.
7
7
2
2
.
9
7
2
.
3
4
2
9
.
8
7
9
.
8
2
*
0
.
6
3
1
.
1
5
(
7
.
1
5
)
(
3
.
6
7
)
(
8
.
1
5
)
(
2
.
7
2
)
(
9
.
1
3
)
(
5
.
9
0
)
(
6
.
5
3
)
(
3
.
5
6
)
R
O
A
2
1
1
.
3
0
4
.
4
3
2
3
4
.
8
3
*
2
8
.
1
4
*
2
1
0
.
5
8
2
0
.
1
2
2
7
.
6
9
2
1
2
.
2
3
*
(
9
.
9
6
)
(
3
.
0
8
)
(
1
7
.
6
0
)
(
3
.
7
0
)
(
7
.
8
8
)
(
4
.
9
4
)
(
5
.
9
7
)
(
4
.
1
6
)
C
A
P
2
6
.
0
2
*
2
3
.
3
5
2
6
.
0
4
2
3
.
9
0
*
2
5
.
0
1
2
1
0
.
1
8
*
2
2
.
5
9
2
5
.
9
7
*
(
3
.
3
0
)
(
2
.
8
0
)
(
5
.
0
6
)
(
1
.
2
4
)
(
3
.
0
8
)
(
4
.
2
9
)
(
2
.
3
9
)
(
2
.
2
5
)
A
C
T
0
.
0
4
2
0
.
0
3
2
0
.
6
2
2
0
.
1
7
1
.
0
9
2
.
0
6
*
0
.
3
1
0
.
1
1
(
0
.
4
5
)
(
0
.
3
6
)
(
0
.
4
7
)
(
0
.
3
0
)
(
0
.
8
2
)
(
0
.
8
9
)
(
0
.
5
1
)
(
0
.
2
6
)
2
2
l
o
g
L
2
0
.
2
9
7
.
8
0
2
7
.
4
2
5
4
.
8
2
2
5
.
3
7
1
4
.
1
2
7
.
2
2
5
2
.
5
1
x
2
p
-
v
a
l
u
e
0
.
0
0
1
1
0
.
1
6
7
4
0
.
0
0
0
0
0
.
0
0
0
0
0
.
0
0
0
1
0
.
0
1
4
8
0
.
2
0
4
6
0
.
0
0
0
0
M
c
F
a
d
d
e
n
R
2
0
.
4
4
3
0
5
0
.
1
6
5
5
0
.
5
8
1
7
0
.
2
7
4
9
0
.
5
3
8
3
0
.
2
9
9
7
0
.
1
5
3
2
0
.
2
6
5
0
A
c
c
u
r
a
c
y
i
n
-
s
a
m
p
l
e
P
e
r
c
e
n
t
a
g
e
c
l
a
s
s
i

e
d
8
2
.
3
5
7
0
.
5
9
8
8
.
2
4
9
9
.
3
2
8
8
.
2
4
7
6
.
4
7
6
4
.
7
1
9
9
.
3
0
T
y
p
e
I
e
r
r
o
r
(
p
e
r
c
e
n
t
)
2
3
.
5
3
3
5
.
2
9
1
7
.
6
5
0
.
6
4
1
7
.
6
5
2
3
.
5
3
4
1
.
1
8
0
.
0
0
T
y
p
e
I
I
e
r
r
o
r
(
p
e
r
c
e
n
t
)
1
1
.
7
6
2
3
.
5
3
5
.
8
8
8
3
.
0
3
5
.
8
8
2
3
.
5
3
2
9
.
4
1
8
8
.
2
4
A
R
C
A
P
0
.
9
2
3
1
0
.
8
4
6
1
0
.
8
6
8
3
0
.
7
8
1
0
0
.
9
5
0
2
0
.
7
3
0
3
0
.
8
9
8
0
0
.
7
6
8
7
A
R
R
O
C
0
.
7
4
2
4
0
.
6
0
0
7
0
.
7
1
7
1
0
.
2
7
0
2
0
.
8
0
2
2
0
.
4
9
5
9
0
.
6
7
2
2
0
.
2
4
6
9
N
o
t
e
s
:
T
h
e
s
a
m
p
l
e
w
a
s
c
o
m
p
o
s
e
d
o
f
1
7
b
a
n
k
r
u
p
t

r
m
s
.
T
h
e
d
e
p
e
n
d
e
n
t
v
a
r
i
a
b
l
e
t
a
k
e
s
t
h
e
v
a
l
u
e
1
f
o
r
t
h
e
1
7
b
a
n
k
r
u
p
t

r
m
s
a
n
d
0
f
o
r
t
h
e
n
o
n
-
b
a
n
k
r
u
p
t

r
m
s
.
S
t
a
n
d
a
r
d
e
r
r
o
r
s
a
r
e
s
h
o
w
n
i
n
p
a
r
e
n
t
h
e
s
e
s
.
*
S
i
g
n
i

c
a
n
t
a
t
t
h
e
5
p
e
r
c
e
n
t
l
e
v
e
l
Table II.
Estimates of the logit
target prediction model
based on the Altman
(1968) variables
The bias of
unhealthy SMEs
71
2
0
0
1
2
0
0
2
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
C
o
n
s
t
a
n
t
1
6
.
7
7
2
0
.
6
5
7
.
2
9
2
1
.
5
4
4
.
3
6
1
3
.
7
1
1
2
.
9
6
*
0
.
2
6
(
1
1
.
6
9
)
(
6
.
8
3
)
(
9
.
4
6
)
(
3
.
4
6
)
(
8
.
1
5
)
(
8
.
3
5
)
(
7
.
0
8
)
(
3
.
3
5
)
R
O
A
2
4
.
9
5
2
2
.
8
3
*
2
3
9
.
0
9
*
2
5
.
4
5
2
2
.
8
2
1
7
.
1
2
2
6
.
8
7
2
9
.
0
3
*
(
2
0
.
0
3
)
(
1
0
.
3
6
)
(
1
9
.
4
2
)
(
3
.
9
8
)
(
8
.
9
0
)
(
1
0
.
6
2
)
(
7
.
0
8
)
(
4
.
5
1
)
S
R
O
A
1
7
4
.
3
1
3
0
.
0
8
2
2
3
.
5
3
7
.
0
9
3
6
.
2
8
5
3
.
5
1
*
1
2
.
0
4
1
8
.
0
4
*
(
1
1
3
.
6
1
)
(
2
1
.
0
6
)
(
2
4
.
4
6
)
(
4
.
8
8
)
(
2
9
.
7
0
)
(
2
2
.
7
8
)
(
1
4
.
4
5
)
(
7
.
4
2
)
D
S
1
.
2
3
0
.
2
2
*
0
.
1
3
0
.
0
3
0
.
3
3
0
.
0
1
0
.
4
0
0
.
0
2
(
1
.
0
4
)
(
0
.
1
2
)
(
0
.
3
2
)
(
0
.
0
8
)
(
0
.
4
1
)
(
0
.
1
3
)
(
0
.
3
1
)
(
0
.
0
9
)
C
P
9
.
7
5
5
.
4
0
2
7
.
3
7
5
.
4
2
2
1
0
.
2
7
2
.
7
8
2
1
.
6
4
0
.
3
0
(
1
1
.
2
3
)
(
5
.
7
7
)
(
8
.
9
0
)
(
3
.
3
1
)
(
9
.
8
7
)
(
7
.
1
8
)
(
7
.
7
0
)
(
3
.
6
0
)
L
I
Q
2
4
.
8
1
2
0
.
6
3
0
.
0
4
0
.
1
4
6
.
1
5
2
.
0
3
6
.
3
2
1
.
7
9
(
1
6
.
4
5
)
(
2
.
6
8
)
(
9
.
5
1
)
(
1
.
8
0
)
(
5
.
6
2
)
(
4
.
3
4
)
(
3
.
7
4
)
(
1
.
8
6
)
C
A
P
2
5
6
.
3
1
2
1
3
.
1
4
*
2
8
.
3
1
2
1
2
.
4
2
*
2
1
2
.
0
4
*
2
1
2
.
1
4
*
2
7
.
0
9
2
9
.
8
8
*
(
3
5
.
1
1
)
(
6
.
3
4
)
(
1
1
.
6
1
)
(
3
.
4
7
)
(
6
.
8
7
)
(
7
.
3
3
)
(
5
.
1
6
)
(
3
.
3
4
)
S
I
Z
E
2
1
.
4
7
0
.
1
2
2
0
.
1
7
2
0
.
1
5
2
0
.
1
9
2
1
.
4
5
*
2
1
.
3
2
*
2
0
.
3
6
(
1
.
1
7
)
(
0
.
7
1
)
(
1
.
0
1
)
(
0
.
3
6
)
(
0
.
9
1
)
(
0
.
8
6
)
(
0
.
7
5
)
(
0
.
3
7
)
2
2
l
o
g
L
3
2
.
2
1
1
7
.
9
3
2
7
.
5
7
6
2
.
2
2
2
7
.
1
2
2
2
.
0
3
1
3
.
6
6
6
1
.
8
6
x
2
p
-
v
a
l
u
e
0
.
0
0
0
0
0
.
0
1
2
0
.
0
0
0
3
0
.
0
0
0
0
0
.
0
0
0
3
0
.
0
0
2
5
0
.
0
5
7
5
0
.
0
0
0
0
M
c
F
a
d
d
e
n
R
2
0
.
6
8
3
4
0
.
3
8
0
5
0
.
5
8
5
0
0
.
3
1
2
0
0
.
5
7
5
4
0
.
4
6
7
3
0
.
2
8
9
9
0
.
3
1
2
2
A
c
c
u
r
a
c
y
i
n
-
s
a
m
p
l
e
P
e
r
c
e
n
t
a
g
e
c
l
a
s
s
i

e
d
8
5
.
2
9
7
6
.
4
7
8
5
.
2
9
9
9
.
4
1
8
5
.
2
9
8
2
.
3
5
7
3
.
5
3
9
9
.
3
4
T
y
p
e
I
e
r
r
o
r
(
p
e
r
c
e
n
t
)
1
7
.
6
5
2
9
.
4
1
1
7
.
6
5
0
.
0
0
1
7
.
6
5
1
7
.
6
5
2
9
.
4
1
0
.
0
0
T
y
p
e
I
I
e
r
r
o
r
(
p
e
r
c
e
n
t
)
1
1
.
7
6
1
7
.
6
5
1
1
.
7
6
7
6
.
4
7
1
1
.
7
6
1
7
.
6
5
2
3
.
5
3
8
2
.
3
5
A
R
C
A
P
0
.
9
7
5
1
0
.
8
6
8
2
0
.
9
1
6
2
0
.
7
9
6
7
0
.
9
6
6
1
0
.
7
9
3
6
0
.
9
3
8
9
0
.
7
8
5
3
A
R
R
O
C
0
.
7
4
2
4
0
.
7
1
5
4
0
.
7
1
6
6
0
.
3
1
3
0
0
.
8
1
5
4
0
.
5
7
0
6
0
.
7
5
5
5
0
.
2
9
7
6
N
o
t
e
s
:
T
h
e
s
a
m
p
l
e
w
a
s
c
o
m
p
o
s
e
d
o
f
1
7
b
a
n
k
r
u
p
t

r
m
s
.
T
h
e
d
e
p
e
n
d
e
n
t
v
a
r
i
a
b
l
e
t
a
k
e
s
t
h
e
v
a
l
u
e
1
f
o
r
t
h
e
1
7
b
a
n
k
r
u
p
t

r
m
s
a
n
d
0
f
o
r
t
h
e
n
o
n
-
b
a
n
k
r
u
p
t

r
m
s
.
S
t
a
n
d
a
r
d
e
r
r
o
r
s
a
r
e
s
h
o
w
n
i
n
p
a
r
e
n
t
h
e
s
e
s
.
*
S
i
g
n
i

c
a
n
t
a
t
t
h
e
5
p
e
r
c
e
n
t
l
e
v
e
l
Table III.
Estimates of the logit
target prediction model
based on the Altman et al.
(1977) variables
JSBED
17,1
72
2
0
0
1
2
0
0
2
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
C
o
n
s
t
a
n
t
6
.
9
7
*
5
.
9
7
*
1
0
.
0
5
3
.
3
8
*
6
.
4
6
*
4
.
7
2
*
2
.
4
8
1
.
5
6
(
4
.
2
2
)
(
2
.
7
6
)
(
6
.
7
8
)
(
1
.
9
4
)
(
3
.
1
8
)
(
2
.
3
7
)
(
2
.
3
1
)
(
1
.
5
5
)
Q
L
I
Q
1
.
8
5
2
0
.
1
3
4
.
9
8
2
0
.
6
3
1
.
7
7
0
.
4
5
2
2
.
1
6
2
1
.
1
4
(
2
.
2
7
)
(
1
.
4
3
)
(
3
.
9
2
)
(
0
.
9
9
)
(
1
.
9
8
)
(
1
.
5
3
)
(
1
.
5
1
)
(
0
.
9
8
)
A
T
P
2
8
.
7
8
*
2
3
.
9
4
2
1
0
.
8
5
2
5
.
9
6
2
6
.
7
7
*
2
2
.
1
5
2
2
.
7
6
2
3
.
9
3
*
(
4
.
3
8
)
(
2
.
4
2
)
(
6
.
8
0
)
(
1
.
7
1
)
(
2
.
7
1
)
(
1
.
7
6
)
(
2
.
0
0
)
(
1
.
1
1
)
I
N
T
4
0
.
1
4
2
4
.
8
9
1
1
3
.
7
2
*
1
6
.
1
2
*
9
3
.
7
5
2
2
1
.
3
4
6
.
9
1
1
8
.
0
5
*
(
3
5
.
3
8
)
(
8
.
2
4
)
(
6
3
.
2
5
)
(
6
.
2
6
)
(
5
8
.
4
0
)
(
2
0
.
0
9
)
(
2
0
.
7
1
)
(
1
0
.
1
5
)
A
M
O
2
6
.
5
5
2
4
.
5
0
2
.
1
8
2
0
.
0
6
2
0
.
4
1
2
.
8
8
4
.
2
9
2
0
.
8
8
(
1
7
.
9
7
)
(
5
.
2
1
)
(
1
2
.
5
6
)
(
3
.
1
8
)
(
6
.
5
7
)
(
4
.
2
7
)
(
5
.
2
9
)
(
2
.
9
6
)
E
A
R
2
1
6
.
1
5
8
.
1
1
*
2
2
8
.
6
3
2
6
.
8
2
*
2
1
.
1
9
2
.
8
5
2
1
.
6
6
2
9
.
7
4
*
(
9
.
8
6
)
(
4
.
1
3
)
(
1
6
.
4
4
)
(
3
.
0
7
)
(
5
.
6
1
)
(
3
.
3
0
)
(
4
.
4
9
)
(
3
.
1
3
)
2
2
l
o
g
L
2
6
.
4
9
1
0
.
5
0
3
1
.
8
5
6
1
.
0
4
2
5
.
9
4
9
.
5
5
5
.
6
9
5
2
.
7
7
x
2
p
-
v
a
l
u
e
0
.
0
0
0
0
0
.
0
6
2
0
.
0
0
0
0
0
.
0
0
0
0
0
.
0
0
0
0
0
0
.
0
8
9
1
0
.
3
3
7
2
0
.
0
0
0
0
M
c
F
a
d
e
n
R
2
0
.
5
6
2
0
0
.
2
2
2
9
0
.
6
7
5
7
0
.
3
0
6
1
0
.
5
5
0
4
0
.
2
0
2
5
0
.
1
2
0
8
0
.
2
6
6
4
A
c
c
u
r
a
c
y
i
n
-
s
a
m
p
l
e
P
e
r
c
e
n
t
a
g
e
c
l
a
s
s
i

e
d
8
8
.
2
4
6
1
.
7
6
9
1
.
1
8
9
9
.
2
8
8
8
.
2
4
6
7
.
6
5
6
4
.
7
1
9
9
.
3
0
T
y
p
e
I
e
r
r
o
r
(
p
e
r
c
e
n
t
)
1
1
.
7
6
3
5
.
2
9
5
.
8
8
0
.
0
5
1
7
.
6
5
3
5
.
2
9
3
5
.
2
9
0
.
0
0
T
y
p
e
I
I
e
r
r
o
r
(
p
e
r
c
e
n
t
)
1
1
.
7
6
4
1
.
1
8
1
1
.
7
6
8
8
.
2
4
5
.
8
8
2
9
.
4
1
3
5
.
2
9
8
8
.
2
4
A
R
C
A
P
0
.
9
5
9
2
0
.
8
5
5
2
0
.
8
7
5
1
0
.
8
1
4
2
0
.
9
6
8
3
0
.
7
1
9
0
0
.
8
4
6
1
0
.
7
7
4
4
A
R
R
O
C
0
.
7
9
8
7
0
.
6
2
9
5
0
.
7
5
9
9
0
.
2
6
0
1
0
.
8
1
4
7
0
.
4
7
6
4
0
.
6
1
8
3
0
.
2
3
5
3
N
o
t
e
s
:
T
h
e
s
a
m
p
l
e
w
a
s
c
o
m
p
o
s
e
d
o
f
1
7
b
a
n
k
r
u
p
t

r
m
s
.
T
h
e
d
e
p
e
n
d
e
n
t
v
a
r
i
a
b
l
e
t
a
k
e
s
t
h
e
v
a
l
u
e
1
f
o
r
t
h
e
1
7
b
a
n
k
r
u
p
t

r
m
s
a
n
d
0
f
o
r
t
h
e
n
o
n
-
b
a
n
k
r
u
p
t

r
m
s
.
S
t
a
n
d
a
r
d
e
r
r
o
r
s
a
r
e
s
h
o
w
n
i
n
p
a
r
e
n
t
h
e
s
e
s
.
*
S
i
g
n
i

c
a
n
t
a
t
t
h
e
5
p
e
r
c
e
n
t
l
e
v
e
l
Table IV.
Estimates of the logit
target prediction model
based on the Garc a et al.
(1997) variables
The bias of
unhealthy SMEs
73
2
0
0
1
2
0
0
2
2
0
0
3
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
B
e
a
v
e
r
(
1
9
6
8
)
T
y
p
e
I
e
r
r
o
r
2
9
.
7
2
7
0
.
6
5
3
2
.
6
3

2
9
.
1
1
6
7
.
1
6
3
3
.
4
6
0
.
0
0
3
3
.
3
5
6
5
.
3
9
3
7
.
6
3
0
.
4
8
T
y
p
e
I
I
e
r
r
o
r
1
1
.
7
6
1
7
.
6
5
5
.
8
8

2
3
.
5
3
3
5
.
2
9
2
3
.
5
3
9
4
.
1
2
1
4
.
2
9
1
4
.
2
9
1
4
.
2
9
8
5
.
7
1
A
R
C
A
P
0
.
8
3
9
8
0
.
5
3
1
6
0
.
7
4
8
4

0
.
8
0
1
0
.
4
7
3
0
.
7
1
3
0
.
7
8
6
0
.
8
3
4
0
.
5
1
2
0
.
7
7
1
0
.
7
8
6
A
R
R
O
C
0
.
7
6
2
5
0
.
7
0
0
6
0
.
7
5
8
6

0
.
7
3
3
0
.
5
9
8
0
.
6
7
9
0
.
2
1
2
0
.
8
1
6
0
.
6
5
2
0
.
7
7
6
0
.
3
2
1
A
l
t
m
a
n
(
1
9
6
8
)
T
y
p
e
I
e
r
r
o
r
2
4
.
3
8
2
6
.
3
9
2
8
.
0
8

2
2
.
9
1
2
4
.
1
4
2
7
.
4
5
0
.
0
0
2
3
.
4
8
2
3
.
4
8
2
8
.
9
4
0
.
4
8
T
y
p
e
I
I
e
r
r
o
r
1
1
.
7
6
2
3
.
5
3
5
.
8
8

2
3
.
5
3
2
9
.
4
1
1
7
.
6
5
9
4
.
1
2
1
4
.
2
9
2
8
.
5
7
1
4
.
2
9
8
5
.
7
1
A
R
C
A
P
0
.
8
9
3
7
0
.
7
8
4
3
0
.
8
3
0
5

0
.
8
7
3
0
.
7
3
8
0
.
8
1
3
0
.
7
4
7
0
.
8
7
3
0
.
7
7
5
0
.
8
3
8
0
.
7
5
9
A
R
R
O
C
0
.
7
4
2
3
0
.
6
0
0
7
0
.
7
1
7
1

0
.
7
2
1
0
.
5
7
1
0
.
6
4
6
0
.
2
2
1
0
.
7
3
7
0
.
5
7
8
0
.
7
1
0
0
.
3
0
0
A
l
t
m
a
n
e
t
a
l
.
(
1
9
7
7
)
T
y
p
e
I
e
r
r
o
r
2
6
.
1
6
4
4
.
3
5
2
7
.
5
3

2
6
.
3
1
4
1
.
6
5
2
7
.
5
4
0
.
0
0
2
2
.
7
4
3
6
.
5
3
2
8
.
3
3
0
.
8
3
T
y
p
e
I
I
e
r
r
o
r
1
1
.
7
6
1
7
.
6
5
1
1
.
7
6

3
5
.
2
9
4
1
.
1
8
1
7
.
6
5
8
8
.
2
4
2
8
.
5
7
2
8
.
5
7
1
4
.
2
9
7
1
.
4
3
A
R
C
A
P
0
.
8
8
1
8
0
.
7
5
0
4
0
.
7
7
4
4

0
.
8
0
1
0
.
6
7
4
0
.
7
4
0
0
.
7
4
9
0
.
8
4
4
0
.
7
7
5
0
.
7
8
6
0
.
7
8
1
A
R
R
O
C
0
.
8
5
2
0
0
.
7
1
6
6
0
.
7
1
5
4

0
.
6
6
2
0
.
6
2
9
0
.
6
5
4
0
.
2
6
3
0
.
7
2
5
0
.
7
1
4
0
.
7
3
6
0
.
3
6
1
G
a
r
c

a
e
t
a
l
.
(
1
9
9
7
)
T
y
p
e
I
e
r
r
o
r
2
4
.
4
8
4
7
.
9
0
2
6
.
7
1

2
4
.
9
4
4
6
.
1
0
2
5
.
9
3
0
.
0
0
5
2
5
.
6
7
4
1
.
3
4
2
5
.
0
1
0
.
4
8
T
y
p
e
I
I
e
r
r
o
r
1
1
.
7
6
4
7
.
0
6
1
1
.
7
6

3
5
.
2
9
2
9
.
4
1
1
7
.
6
5
9
4
.
1
2
1
4
.
2
9
4
2
.
8
6
1
4
.
2
9
8
5
.
7
1
A
R
C
A
P
0
.
8
7
1
1
0
.
7
0
0
7
0
.
8
0
5
8

0
.
8
0
9
0
.
7
4
2
0
.
7
4
3
0
.
7
8
5
0
.
8
4
6
0
.
6
9
8
0
.
7
6
5
0
.
7
8
3
A
R
R
O
C
0
.
7
9
8
7
0
.
6
2
9
5
0
.
7
5
9
9

0
.
6
8
1
0
.
6
4
4
0
.
6
8
0
0
.
2
2
4
0
.
7
9
9
0
.
5
7
7
0
.
7
0
6
0
.
3
1
3
N
o
t
e
:
T
h
e
o
u
t
-
o
f
-
s
a
m
p
l
e
p
r
e
d
i
c
t
i
v
e
a
b
i
l
i
t
y
o
f
t
h
e
m
o
d
e
l
s
i
s
m
e
a
s
u
r
e
d
o
n
t
h
e
w
h
o
l
e
p
o
p
u
l
a
t
i
o
n
o
f

r
m
s
i
n
2
0
0
1
,
2
0
0
2
a
n
d
2
0
0
3
Table V.
Out-of-sample accuracy
of the models estimated
in 2001
JSBED
17,1
74
2
0
0
2
2
0
0
3
2
0
0
4
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
S
t
r
o
n
g

r
m
s
W
e
a
k

r
m
s
C
o
n
t
r
o
l

r
m
s
T
o
t
a
l

r
m
s
B
e
a
v
e
r
(
1
9
6
8
)
T
y
p
e
I
e
r
r
o
r
2
8
.
9
6
2
4
.
8
9
6
1
.
2
4

2
8
.
8
5
2
3
.
8
3
6
3
.
3
8
1
.
1
4
2
8
.
1
8
2
3
.
2
7
6
3
.
7
4
1
.
3
1
T
y
p
e
I
I
e
r
r
o
r
1
1
.
7
6
2
9
.
4
1
2
3
.
5
3

1
1
.
7
6
2
3
.
5
3
1
1
.
7
6
2
9
.
4
1
1
4
.
2
9
5
7
.
1
4
2
8
.
5
7
4
2
.
8
6
A
R
C
A
P
0
.
8
8
0
0
.
4
0
6
0
.
8
3
5

0
.
8
7
9
0
.
4
2
9
0
.
7
3
3
0
.
7
9
6
0
.
8
5
6
0
.
4
7
3
0
.
6
2
2
0
.
7
5
4
A
R
R
O
C
0
.
7
9
0
0
.
5
7
3
0
.
6
1
6

0
.
7
9
0
0
.
5
1
5
0
.
6
3
0
0
.
4
0
8
0
.
8
3
1
0
.
4
4
7
0
.
6
8
1
0
.
5
7
7
A
l
t
m
a
n
(
1
9
6
8
)
T
y
p
e
I
e
r
r
o
r
2
8
.
1
6
3
1
.
5
2
2
0
.
7
8

2
8
.
1
1
3
0
.
9
9
2
3
.
0
9
0
.
7
1
2
6
.
5
3
3
0
.
2
6
2
2
.
9
3
0
.
6
5
T
y
p
e
I
I
e
r
r
o
r
5
.
8
8
1
7
.
6
5
2
9
.
4
1

5
.
8
8
5
.
8
8
1
1
.
7
6
3
5
.
2
9
2
8
.
5
7
2
8
.
5
7
1
4
.
2
9
4
2
.
8
6
A
R
C
A
P
0
.
8
7
3
0
.
7
4
0
0
.
7
7
6

0
.
8
4
9
0
.
7
4
4
0
.
8
1
3
0
.
7
7
4
0
.
7
8
5
0
.
7
1
4
0
.
7
5
9
0
.
7
6
3
A
R
R
O
C
0
.
8
0
2
0
.
4
9
5
0
.
6
7
2

0
.
7
7
5
0
.
5
4
0
0
.
7
4
5
0
.
3
3
3
0
.
7
1
2
0
.
6
6
9
0
.
6
8
0
0
.
5
2
8
A
l
t
m
a
n
e
t
a
l
.
(
1
9
7
7
)
T
y
p
e
I
e
r
r
o
r
2
8
.
5
8
4
0
.
5
1
2
6
.
9
3

2
3
.
8
3
3
9
.
0
2
2
1
.
3
0
0
.
9
5
2
3
.
3
2
3
4
.
0
4
2
0
.
6
7
0
.
8
7
T
y
p
e
I
I
e
r
r
o
r
1
1
.
7
6
1
7
.
6
5
2
3
.
5
3

5
.
8
8
1
1
.
7
6
1
1
.
7
6
2
9
.
4
1
2
8
.
5
7
0
.
0
0
1
4
.
2
9
2
8
.
5
7
A
R
C
A
P
0
.
8
7
7
0
.
7
1
6
0
.
7
7
9

0
.
8
5
6
0
.
7
5
8
0
.
7
8
7
0
.
7
7
3
0
.
8
1
5
0
.
8
1
4
0
.
8
1
2
0
.
7
6
3
A
R
R
O
C
0
.
8
1
5
0
.
5
7
0
0
.
7
7
5

0
.
7
7
4
0
.
6
4
5
0
.
7
2
1
0
.
3
6
7
0
.
7
3
6
0
.
7
2
6
0
.
7
5
6
0
.
6
0
6
G
a
r
c

a
e
t
a
l
.
(
1
9
9
7
)
T
y
p
e
I
e
r
r
o
r
3
0
.
7
6
3
8
.
8
1
1
6
.
3
3

2
7
.
7
2
3
7
.
8
4
1
7
.
5
0
0
.
5
2
2
5
.
7
1
3
6
.
0
4
1
7
.
7
6
0
.
4
8
T
y
p
e
I
I
e
r
r
o
r
5
.
8
8
2
9
.
4
1
3
5
.
2
9

1
1
.
7
6
5
.
8
8
1
1
.
7
6
3
5
.
2
9
1
4
.
2
9
4
2
.
8
6
1
4
.
2
9
5
7
.
1
4
A
R
C
A
P
0
.
8
7
2
0
.
7
4
3
0
.
7
4
6

0
.
8
5
3
0
.
7
5
4
0
.
8
0
6
0
.
7
8
8
0
.
8
5
0
0
.
6
0
5
0
.
7
5
8
0
.
7
4
5
A
R
R
O
C
0
.
8
1
4
0
.
4
7
6
0
.
6
1
8

0
.
7
5
4
0
.
5
2
4
0
.
6
7
8
0
.
3
1
4
0
.
8
2
7
0
.
4
8
7
0
.
6
1
5
0
.
3
9
8
N
o
t
e
:
T
h
e
o
u
t
-
o
f
-
s
a
m
p
l
e
p
r
e
d
i
c
t
i
v
e
a
b
i
l
i
t
y
o
f
t
h
e
m
o
d
e
l
s
i
s
m
e
a
s
u
r
e
d
o
n
t
h
e
w
h
o
l
e
p
o
p
u
l
a
t
i
o
n
o
f

r
m
s
i
n
2
0
0
1
,
2
0
0
2
a
n
d
2
0
0
3
Table VI.
Out-of-sample accuracy
of the models estimated
in 2002
The bias of
unhealthy SMEs
75
Conclusions
Many efforts have been devoted to developing models to predict bankruptcy
probability and banks, investors and rms are interested in using these models.
However, since the models have been developed in the aim to estimate the bankruptcy
probability of quoted rms, less attention has been paid to non-quoted SMEs. Usually,
the bankruptcy probability of non-quoted SMEs is studied by analysing nancial
ratios using discriminating techniques. These models are based on constructing an
estimation sample using bankrupt and non-bankrupt rms, controlling by size and
industry or unbalancing the sample using all rms.
In our article, we propose an alternative to control the heterogeneity of healthy
rms. In doing this, we construct a nancial health indicator to dene the degree of
nancial health. Our results are obtained under four different models of credit scoring.
We show in-sample and out-of-sample bankruptcy predictions. We apply measures
proposed by the Basel Committee on Banking Supervision to evaluate the credit
scoring of the models depending on the criterion considered to select the rms.
Focusing on the non-bankrupt rms, the nancial health indicator allows rms to
be classied before the estimation process. This procedure permits the heterogeneity of
the rms to be reduced as well as identifying a strong rm sample to estimate the
bankruptcy probability accurately. The in-sample and out-of-sample evaluation based
on the CAP and ROC indicators, proposed by the BCBS, leads to the conclusion that the
models estimated under the strong rm sample are much more accurate.
References
Altman, E.I. (1968), Financial ratios, discriminant analysis and the prediction of corporate
bankruptcy, Journal of Finance, Vol. 23 No. 4, pp. 589-609.
Altman, E.I. and Saunders, A. (1998), Credit risk measurement: developments over the last
20 years, Journal of Banking and Finance, Vol. 21, pp. 1721-42.
Altman, E.I., Haldeman, E. and Narayanan, P. (1977), Z analysis: a new model to identify
bankruptcy risk of corporations, Journal of Banking and Finance, Vol. 10, pp. 29-54.
Basel Committee on Banking Supervision (2005), Studies on the validation of internal rating
systems, Working Paper No. 14, Basel Committee on Banking Supervision, Basel.
Beaver, W.H. (1966), Financial ratios as predictor of failure, Empirical Research in Accounting:
Selected Studies, Supplement to Journal of Accounting Research, Vol. 5, pp. 71-111.
Beaver, W.H. (1968), Alternative accounting measures as predictor of failure, The Accounting
Review, Autumn, pp. 112-22.
Berryman, J. (1983), Small business failure and bankruptcy: a survey of the literature,
European Small Business Journal, Vol. 1 No. 4, pp. 47-59.
Carter, R. and Van Auken, H. (2006), Small rm bankruptcy, Journal of Small Business
Management, Vol. 44 No. 4, pp. 493-512.
Cochran, B. (1986), Small business mortality rates: a review of the literature, Journal of Small
Business Management, Vol. 19 No. 4, pp. 50-9.
Craig, B., Jackson, W. and Thomson, J. (2003), On SBA-guaranteed lending and economic
growth, working paper, Federal Reserve Bank of Cleveland, Cleveland, OH, April.
Edmister, R. (1972), An empirical test of nancial ratio analysis for small business failure
prediction, Journal of Financial and Quantitative Analysis, Vol. 7 No. 2, pp. 1477-93.
JSBED
17,1
76
Engelmann, B., Hayden, E. and Tasche, D. (2003), Testing rating accuracy, Risk, January,
pp. 82-6.
Fredland, J. and Morris, E. (1976), A cross section analysis of small business failure, American
Journal of Small Business, Vol. 1, July, pp. 7-18.
Garc a, D., Arques, A. and Calvo-Flores, A. (1997), Un modelo discriminante para evaluar el
riesgo bancario en los creditos a empresas, Revista Espan ola de Financiacion y
Contabilidad, Vol. 24 No. 82, pp. 175-200.
Hall, G. and Young, B. (1991), Factors associated with insolvency amongst small rms,
International Small Business Journal, Vol. 9 No. 2, pp. 54-63.
Headd, B. (2003), Redening business success: distinguishing between closure and failure,
Small Business Economics, Vol. 21, pp. 51-61.
Thedodossious, P., Kahya, E., Saidi, R. and Philippatos, G. (1996), Financial distress and
corporate acquisitions: further empirical evidence, Journal of Business Finance and
Accounting, Vol. 23, pp. 699-719.
Watson, J. and Everett, J. (1996), Do small business have high failure rates?, Journal of Small
Business Management, Vol. 34 No. 4, pp. 45-52.
Watson, J. and Everett, J. (1999), Small business failure rates: choice of denition and industry
effects, International Small Business Journal, Vol. 17 No. 2, pp. 31-47.
Zingales, L. (2000), In search of new foundations, Journal of Finance, Vol. 55 No. 4, pp. 1623-53.
About the authors
J. Samuel Baixauli is an Associate Professor at the Department of Management and Finance at
the University of Murcia, Spain. He received his PhD in Business Administration for work on
nancial economics. He has been research visitor at the Department of Economics at the
University of York, UK. His main research areas include nancial modeling and estimation of
market and credit risk. He has published on these topics in international journals such as
European Journal of Operational Research, Journal of Financial Research, Review of Quantitative
Finance and Accounting and European Journal of Finance. J. Samuel Baixauli is the
corresponding author and can be contacted at: sbaixaul@um.es
Antonina Modica-Milo is an Associate Professor at the Department of Management and
Finance at the University of Oriente, Venezuela. She received her PhD in Business Economics
from the Department of Management and Finance at the University of Murcia, Spain. Her main
research areas include nancial modeling and estimation of credit risk. She has experience
working as a nancial advisor for relevant private companies.
The bias of
unhealthy SMEs
77
To purchase reprints of this article please e-mail: reprints@emeraldinsight.com
Or visit our web site for further details: www.emeraldinsight.com/reprints