Knowledge-Based Systems 24 (2011) 1380–1388

Contents lists available at ScienceDirect

Knowledge-Based Systems
journal homepage: www.elsevier.com/locate/knosys

The random subspace binary logit (RSBL) model for bankruptcy prediction
Hui Li a,d,⇑, Young-Chan Lee b, Yan-Chun Zhou c, Jie Sun a
a
School of Economics and Management, Zhejiang Normal University, P.O. Box 62, 688 YingBinDaDao, Jinhua, Zhejiang 321004, China
b
Division of Economics & Commerce, Dongguk University, GyeongJu Campus, Gyeongju, Gyeongbuk 780-714, Republic of Korea
c
School of Business, Ningbo University, 818 FengHuaLu Road, Ningbo, Zhejiang 315211, China
d
College of Engineering, The Ohio State University, 470 Hitchcock Hall, 2070 Neil Avenue, Columbus, OH 43210, USA

a r t i c l e i n f o a b s t r a c t

Article history: This paper proposes the random subspace binary logit (RSBL) model (or random subspace binary logistic
Received 6 January 2011 regression analysis) by taking the random subspace approach and using the classical logit model to gen-
Received in revised form 21 June 2011 erate a group of diverse logit decision agents from various perspectives for predictive problem. These
Accepted 21 June 2011
diverse logit models are then combined for a more accurate analysis. The proposed RSBL model takes
Available online 29 June 2011
advantage of both logit (or logistic regression) and random subspace approaches. The random subspace
approach generates diverse sets of variables to represent the current problem as different masks. Differ-
Keywords:
ent logit decision agents from these masks, instead of a single logit model, are constructed. To verify its
Bankruptcy prediction
Random subspace binary logit
performance, we used the proposed RSBL model to forecast corporate failure in China. The results indi-
Group decision of predictive models cate that this model significantly improves the predictive ability of classical statistical models such as
Corporate failure prediction multivariate discriminant analysis, logit model, and probit model. Thus, the proposed model should make
Probit logit model more suitable for predictive problems in academic and industrial uses.
Multivariate discriminant analysis Ó 2011 Elsevier B.V. All rights reserved.

1. Introduction multivariate normality. To address this problem, Martin [25] and

Investors, creditors, bankers, stockholders, and managers need
effective tools for managing various risks associated with their
decisions. Researches in this topic includes: Cho [9], Pai et al.
[33], among others. One such tool is bankruptcy prediction, which
refers to prediction of business failure through financial variables
[11,14,15,26–28,34,36,38]. Early studies of bankruptcy prediction
typically employed discriminant analysis. Fitzpatrick [10] provided
an in-depth interpretation of bankruptcy variables and trends (i.e.,
he employed a type of multiple variable analysis); Beaver [3] pro-
posed a framework for employing univariate analysis for bank-
ruptcy prediction; and Altman [1] employed multivariate
discriminant analysis (MDA) to predict bankruptcy. Because
discriminant analysis is easy to understand, interpret, and explain,
it has been suitable for industrial use. MDA assumes dichotomous
data, a multivariate normal distribution, equal variance-covariance
matrices across two groups, a specified prior probability of two
groups, and the absence of multicollinearity [2]. However, the
assumption of multivariate normality is often violated for bank-
ruptcy data without effective solutions (e.g., only several financial
variables distribute normally for bankruptcy data from China).
Further, univariate normality is not a sufficient condition for
⇑ Corresponding author at: School of Economics and Management, Zhejiang
Normal University, P.O. Box 62, 688 YingBinDaDao, Jinhua, Zhejiang 321004, China.
Tel.: +86 579 8229 8602.
E-mail address: lihuihit@gmail.com (H. Li).
Ohlson [29] used a conditional probability model to forecast bank-
ruptcy. This type of model uses the nonlinear maximum likelihood
method for estimating probability of bankruptcy. According to the
assumption about probability distribution, logit and probit models
assume a logistic distribution and a cumulative normal distribu-
tion, respectively. The relationship between financial variables
and the probability of bankruptcy is typically assumed to be linear.
Thus, linear logit model has typically been used. These two classi-
cal statistical models have been widely used for forecasting bank-
ruptcy for many years [17,18,39,16]. Even now, firms frequently
employ logit model to calculate probability of bankruptcy for their
customers. Previous studies have attempted to improve the effec-
tiveness of these two classical models by optimizing single model.
For example, Refs. [32,31] proposed the Tabu method, a type of fea-
ture selection method, for discriminant analysis and logit model.
In the past two decades, a number of studies have investigated
performance of intelligent models on bankruptcy prediction (e.g.,
case-based reasoning with the k-nearest neighbor as the heart,
classification and regression tree, support vector machine, among
others). However, the two statistical models are still very popular
(particularly for industrial use) because they are well-known mod-
els for bankruptcy prediction and are easy to model, interpret, and
explain. Logit model is used more frequently because it is less
demanding than MDA. However, logit model has a drawback. Most
of the recent studies (e.g., [22,23,40,19]) have demonstrated that
predictive abilities of classical statistical models (e.g., logit model)
are relative weak.

0950-7051/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.knosys.2011.06.015
 H. Li et al. / Knowledge-Based Systems 24 (2011) 1380–1388
Predictive accuracy is one of the most important indicators of
model effectiveness. If logit model is relative weak in predictive
accuracy, the loss saved by using logit in bankruptcy prediction
in industry will be smaller. However, it is difficult to introduce
new models for industrial users for the following reasons. (1) Intel-
ligent models are complex approaches for the problem, compared
with statistical models. (2) Industrial practitioners are already
familiar with the classical approaches. One is not likely to replace
a familiar tool as long as it continues to perform reasonably well.
Thus, the shortcoming of logit model—relative poor performance
in terms of bankruptcy prediction—should be addressed to make
it more suitable for industrial use. One may argue that the loss
resulting from the use of logit model instead of intelligent models
for bankruptcy prediction is not substantial because the difference
in reductions in predictive error between statistical models and
intelligent models is commonly 3%. However, a decision tool that
can reduce predictive error by 3% could potentially save the indus-
try approximately $1.2 billion annually [41]. Because logit model is
widely used for bankruptcy prediction, we need to enhance its
predictive ability for forecasting bankruptcy while preserving its
advantages.
From the perspective of management science and ensemble
learning, using various decision agents with diverse opinions is
effective in improving performance of predictive model [30,20].
The predictive ability of a committee of decision agents may ex-
ceed that of a single decision agent. Logit model can be regarded
as a decision agent for bankruptcy prediction. The current problem
can be described and represented by some variables as a mask
from which a decision agent will be constructed. The random sub-
space method is an approach for producing various representations
(i.e., masks) that can be used to generate different decision agents.
This approach injects randomness into problem representation by
randomly selecting variables with replacement. This means that
different variable sets are used to construct logit model.
Thus, to improve the analysis performance of logit model, the
present study combines random subspace approach with binary
logit model to generate the simple random subspace binary logit
(RSBL) model that takes into account different decision agents’
opinions. The results of practical application verifying its ability
to forecast corporate failure in China indicate that the proposed
model can forecast corporate failure significantly better than clas-
sical statistical models (i.e., MDA, logit model, and probit model).
The paper is organized as follows. Section 2 introduces binary logit
model and random subspace approach. Section 3 proposes the
RSBL model. Section 4 uses the proposed model to forecast corpo-
rate failure in China, and Section 5 concludes.
2. Binary logit model and random subspace approach
2.1. Binary logit model for bankruptcy prediction
Logit model is used to forecast probability of an event by fitting
data (represented by some variables) to the logistic curve. The term
‘‘logit,’’ introduced by Berkson [4], is borrowed from probit model,
a similar model introduced by Bliss [5]. In the field of bankruptcy
prediction, the occurrence of an event refers to corporate failure.
Further, the variables include financial variables calculated from
firm’s public financial statements. For example, the probability of
a firm declaring bankruptcy in the future can be determined by
analyzing various variables for the firm’s profitability (e.g., various
financial ratios). Assume that the result of bankruptcy prediction is
either corporate failure or non-failure. The number of Xi is known,
and the probability of proi is unknown. Here Xi refers to the ith
object (observation), and a total of m observations exist. For each
observation, a set of variables can be used to inform the probability
1381
of an event (e.g., a bankruptcy). Assume that there are k variables,
namely x1, x2, . . ., xk. The logit value of the unknown binomial prob-
ability is modeled as a linear function:
 
probabilityi
logitðprobabilityi Þ ¼ ln
1  probabilityi
¼ b0 þ b1 xi1 þ b2 xi2 þ    þ bk xik : ð1Þ
The unknown parameter bj can be estimated through maximum
likelihood estimation of generalized linear models. The greater the
value of bj is, the more the jth variable contributes to prediction.
Further, b0 refers to an intercept, and bj (j = 1, 1, . . ., k), the regres-
sion coefficient of the jth variable. Finally,
1
probabilityi ¼ ðb0 þb1 xi1 þb2 xi2 þþbk xi Þ
: ð2Þ
1þe k
Assume that
z ¼ b0 þ b1 x1 þ b2 x2 þ    þ bk xk : ð3Þ
The following function, which is dependent on z, is referred to as lo-
gistic regression:
1
f ðzÞ ¼ ; ð4Þ
1 þ ez
where f(z) 2 (0, 1) represents the probability of an event. Typically,
the cutoff value is 0.5. Logistic regression is useful for describing the
relationship between financial variables and the probability of cor-
porate failure.
2.2. Random subspace approach
Random subspace approach refers to the construction of deci-
sion models through random selection of a number of variables
from a given set of variables [13]. This approach reflects a type of
ensemble method and is known to be able to improve predictive
accuracy [6,7,8]. Each time a random subspace is generated, a deci-
sion agent will be produced by constructing a model on top of the
representation of the current problem. Finally, all decision agents
are integrated by a simple vote by the committee. This approach
is described as follows. Consider the condition with m observations
in a k-dimensional space:
fðx1;j ; x2;j ; . . . ; xi;j ; . . . ; xk;j Þjxi;j g; where i 2 f1; mg; j 2 f1; kg: ð5Þ
i is the number of observations; j is the number of variables; and xi,j
refers to the value of the ith observation for the jth variable. The
random subspace represented by the randomly selected variables
is expressed as follows:
fðx1;j ; x2;j ; . . . ; xm;j Þjxi;j g;
where
ð6Þ
xi;j ¼ xi;j for i 2 I; and
xi;j ¼ Null for i R I:
I is the k0 -dimensional subset of {1, 2, . . ., k}; and k0 6 k. Here we as-
sume that k0 = k. Thus, the total number of variables in subspaces is
the same as that of those variables in the initial space. Because some
of the same variables are selected, the actual number of variables is
less than k. The creation of random space is repeated P times. Each
time, a random generator from 1 to k0 is used to select a variable to
be used in the subspace, and the process is repeated k0 times. The
source of randomness is based on re-sampling P times with replace-
ment the blocks of m observations for each k variable. As a result, P
random subspaces are created as different masks for the current
problem. The same algorithm is implemented on top of these masks
to generate diverse decision agents. According to Ho [13], this
approach reflects a type of stochastic discrimination to increase
1382 H. Li et al. / Knowledge-Based Systems 24 (2011) 1380–1388

predictive accuracy by combining weak models that do not have full
discriminate power for the same problem. Combining invariant
models has a chance to increase their discriminative power.
The randomly selected variables represent different masks for
the current problem, which means that different masks have dif-
ferent component variables reflecting different perspectives of
the classification problem. Thus, the corresponding models based
on different masks focus on different perspectives of the current
problem. These models have different discriminant abilities, which
are on the foundation of their perspectives. This random subspace
approach attempts to derive only the different perspectives from
the initial representation of the current problem. These perspec-
tives represent the interests of different committees. This is consis-
tent with the operation of committees in the real world, that is,
committees include representatives from different groups. Finally,
the voting mechanism is used to produce a consensus inside the
committee of voters or decision agents.
3. The random subspaces binary logit model
The RSBL model is constructed by combining logit model with
random subspace approach. To generate a committee of logit
decision agents, we take the random subspace approach to produce
diverse masks of the current problem for logit model. Based on
Eq. (6), we assume that the rth mask is expressed as follows:
fðxr1;j ; xr2;j ; . . . ; xrm;j Þjxri;j g;
0
where i 2 f1; mg; j 2 f1; k g: ð7Þ
According to this expression, the current problem is represented by
a mask using k0 variables from the initial set of k variables. Some
variables in the initial set of variables may emerge more than once
in the rth mask, whereas others may not emerge at all. The k0 vari-
ables are selected as follows: First, a random number generator is
used to produce a random number between (0, 1). Then the random
number is multiplied by the total number of variables, namely k, to
be a feature indicator. The upper integer of the feature indicator is
used to identify the selected feature. Third, this process is repeated
k times, resulting in k features. These k features are used to repre-
sent sample as a mask. The whole procedure is repeated P times. Fi-
nally, P masks are obtained for the sample representation. After the
sets of subspace variables are selected, the logit decision agents are
constructed on the m observations with P masks.
By constructing Eq. (4) on the basis of Eq. (7), we transfer logit
model as follows:
f ðzr Þ ¼ 1þe1zr ;
where ð8Þ
zr ¼ br0 þ br1 xr1 þ br2 xr2 þ    þ brk0 xrk0 :
Eq. (7) are the same, then all the logit decision agents are the same.
As a result, combining the results for different logit agents’ deci-
sions produces no difference. Thus, using randomness to increase
the diversity of committee members is critical. When logit agents
have their decisions, the committee will reach a consensus through
the following mechanism:
X
P
Z¼ wr  f ðzr Þ
r¼1
X
P
wr
¼
r¼1
1 þ expðzr Þ
ð10Þ
X
P
wr
¼ ;
r¼1
1 þ expððbr0 þ br1 xr1 þ br2 xr2 þ    þ brk0 xrk0 ÞÞ
where
0
i 2 f1; mg; j 2 f1; k g; xri 2 ð1; þ1Þ; wr 2 ½0; 1:
wr is the importance of the opinion of the rth logit decision agent;
P
and wr = 1. Further, Z 2 (0, 1) indicates the score whether an event
is likely to occur. Z is an integration of P logit decision agents. Be-
cause randomness is injected into the proposed RSBL model, the
weight of each logit decision agent can be the same, that is, 1/P.
By this use of the logit decision committee to produce the prob-
ability of an event (e.g., bankruptcy prediction), the analysis per-
formance of logit model can be enhanced. The cutoff value is still
0.5. If Z P 0.5, then the consensus of the logit decision committee
(i.e., the RSBL model) is that the event will happen with the prob-
ability of Z. Otherwise, the consensus is that the event will not
occur.
According to Hastie et al. [12] and Lim [24], combining several
weak decision agents can produce a powerful committee. We use a
total of P logit decision agents. These agents are independent of
one another in that randomness is injected into the construction
of the proposed model. For a simple vote by the committee, we
can assume that P is odd. Let Xr express a random variable reflect-
ing the prediction of the rth logit model. Suppose that the perfor-
mances of P logit decision agents are expressed as the variable
accuracy, which is derived based on Eq. (10). Thus,
X r  BernoulliðaccuracyÞ: ð11Þ
The variable accuracy is determined by the proportion of correct
decisions by the logit agent in all decisions. The number of accurate
predictions by the RSBL model (i.e., the committee of logit decision
agents) is defined as follows:
X
P
Y¼ X r  BinomialðP; accuracyÞ: ð12Þ
r¼1

If a variable emerges more than once in the rth mask, then coeffi- Because we assume that P is odd, we express P as P = 2K + 1, where
cient of the variable indicates the number of times it emerges. By K is a nonnegative integer. Let Zr = Probability (Y P K + 1). Then the
this means, the effectiveness of randomness is used to influence predictive accuracy of the RSBL model by voting is calculated as
the results. Further, f(zr) is in the range of (0, 1) because x ranges follows:
from (1, +1). X
P
By using P masks for the current problem, we can obtain P dif- ZP ¼ PC r accuracyr ð1  accuracyÞPr ; ð13Þ
ferent logit decision agents based on Eq. (8). The set of logit deci- r¼Kþ1
sion agents is expressed as follows:
where Cr is a turning parameter, namely P!/r!/(P  r)!. Based on
fðf ðz1 Þ; f ðz2 Þ; . . . ; f ðzP ÞÞjf ðzr Þg; where r 2 f1; Pg: ð9Þ Lam and Suen [21], we will obtain the following results:
8
Here all f(zr) are in the range of (0, 1). The P logit decision agents < < 0:5 when accuracy < 0:5;
>
constitute the model committee for the current problem (e.g., bank- Z 2Kþ1 ¼ 0:5 when accuracy ¼ 0:5; ð14Þ
ruptcy prediction). Because Eq. (7) is used to inject the randomly >
:
> 0:5 when accuracy > 0:5:
generated masks for the initial problem into the modeling process
of logit model, committee members of decision agents in Eq. (9) The RSBL model follows the following procedure:
model different perspectives of the current problem. Randomness
makes different logit decision agents to produce diverse decisions. Step 1: Randomly generate P masks from the original variables
As a result, their diversity is assured. Note that if all masks from with replacement. These masks are used to represent the
 H. Li et al. / Knowledge-Based Systems 24 (2011) 1380–1388
Fig. 1. Mechanism of the RSBL model.
available observations from different perspectives. By con-
structing diverse masks for the current problem, we will under-
stand and model the problem more effectively.
Step 2: Fit P logit decision agents by using all available masks for
observations. These P logit decision agents represent P different
perspectives on the initial data. This means that diverse masks
holding knowledge of classification are transferred to a commit-
tee of logit decision agents.
Step 3: Make P predictions about the same target by using P dif-
ferent logit decision agents. New problems will be analyzed by
using the committee of logit decision agents, each of which ful-
fils the analysis independently.
Step 4: Produce a consensus for the RSBL model (i.e., the com-
mittee of logit decision agents). A simple vote by the committee
of logit decision agents is used to integrate the opinions of each
logit model. This produces the results for the RSBL model.
Fig. 1 shows the mechanism of the RSBL model.
4. Empirical design and results
4.1. Bankruptcy prediction for Chinese firms
To verify the feasibility and effectiveness of the RSBL model, we
examined its ability to predict corporate failure in China. We ob-
tained the data from the Shanghai Stock Exchange and the Shenz-
hen Stock Exchange, which included a total of 135 pairs of firms
representing those that declared bankruptcy and those that did
not. There is no rigid mathematical definition of an outlier. Typi-
cally, approximately 1 in 370 deviates by three times the standard
deviation for normally distributed data [35]. This study defines
those firms with variable values deviating from the mean by more
than three times the standard deviation as outliers. In addition,
there is no rigid mathematical definition of corporate failure, and
models of bankruptcy prediction identify only a certain part of cor-
porate failure. Further, an effective model for identifying outlier
failures may not be the same as that for indentifying common fail-
ures. Thus, removing outliers from data is useful for making a mod-
el for identifying common failures more effective, that is, if outliers
are included, then the accuracy of the model should decrease not
only for common failures but also for outlier failures. In this regard,
we excluded outliers and observations with missing values and ob-
tained 153 observations for short-term bankruptcy prediction and
216 observations for medium-term bankruptcy prediction [37]. For
short-term bankruptcy prediction, we used data one year prior to
corporate failure, and for medium-term bankruptcy prediction,
we used data two years prior to corporate failure.
We used a total of 30 financial variables as the initial ratios to
represent the observations; none of these variables had missing
values. To determine whether the RSBL model can significantly
improve predictive accuracy, we employed two feature selection
approaches—stepwise MDA and a two-tailed t-test at the 95%
1383
confidence level—to produce four more data sets. Thus, we ob-
tained a total of six data sets. Finally, as shown in Appendix, we
used the variables to initially represent the observations for the
two types of data sets [22]. Table 1 summarizes the six data sets.
Using the initial data sets, we compared predictive performance
of the RSBL model with that of the classical statistical models (i.e.,
MDA, logit model, and probit model). We set the total number of
logit decision agents as 11 because Breiman [6] noted that most
of the improvements from bagging will be achieved within 10 rep-
lications. However, there will be no consensus on the binary clas-
sification without additional processes if each group has five votes.
To address this problem, we used an odd number (11). We then
tested the following null hypotheses:
 H1: Predictive performance of the RSBL model is significant bet-
ter than that of MDA.
 H2: Predictive performance of the RSBL model is significant bet-
ter than that of logit model.
 H3: Predictive performance of the RSBL model is significant bet-
ter than that of probit model.
To test these hypotheses, we divided all the observations into
two groups. We used one group (approximately 70% of observa-
tions) as the training and calibration data set for the modeling
and the other group (approximately 30% of observations) as unseen
data set for testing predictive performance of the models. We re-
peated this split 200 times to make the results sensible in statistic.
Thus, we employed the multiple holdout method, which reduces
the bias in performance estimation of models much better than
the one-time holdout method. One-tailed paired-samples t test
was employed to verify hypotheses.
4.2. Results
The performance measure, i.e., accuracy, is the proportion of
right decisions of a logit agent on testing samples in total decisions.
The values for statistic analysis are accuracies of models on the 30%
testing dataset. Table 2 shows the results for predictive perfor-
mance of the RSBL model and the three classical statistical models
on the six data sets (the six statistical indices refer to the mean, the
median, the minimum, the maximum, the standard deviation, and
the range). The best-time index refers to the number of times the
Table 1
The six datasets for bankruptcy prediction.
Dataset Description
1 Data set for short-term prediction with all variables
2 Data set for short-term prediction with t-test variables
3 Data set for short-term prediction with MDA variables
4 Data set for medium-term prediction with all variables
5 Data set for medium-term prediction with t-test variables
6 Data set for medium-term prediction with MDA variables
1384 H. Li et al. / Knowledge-Based Systems 24 (2011) 1380–1388

Table 2
Results of bankruptcy prediction (six data sets).

model provided the optimal value. The values in rectangular indi-
cate the best performance result. Table 3 shows the results of t-test
for the three hypotheses.
We compared performance of the RSBL model with that of the
other models. As shown in Table 4, the RSBL model predicted cor-
porate failure much better than the classical models. The second
column indicates the error rate, and the third column refers to
the error rate for the RSBL model. The fourth, fifth, and sixth col-
umns refer to the differences in the error rate between MDA and
the RSBL model; logit model and the RSBL model; and probit model
and the RSBL model, respectively. A positive value means that the
classical model made more errors in bankruptcy prediction than
the RSBL model, whereas a positive value means that the classical
model made fewer errors. The seventh, eighth, and ninth columns
indicate the superiority of the RSBL model over MDA, Logit, and
Probit, respectively, in percentage terms. The underlined values
indicate the cases in which MDA, logit model, and probit models
performed better than the RSBL model.

Table 3
Significance test (one-tailed test).

⁄⁄⁄
Significant at the 1% level.
 H. Li et al. / Knowledge-Based Systems 24 (2011) 1380–1388 1385

Table 4
Comparative results using random subspaces binary logit model as a benchmark.

Data set RSBL MDA-RSBL Logit-RSBL Probit-RSBL MDA-RSBL

RSBL (%) Logit-RSBL
(%) Probit-RSBL
RSBL (%)
RSBL

1 Mean error 15.81 1.37 5.49 5.52 8.67 34.72 34.91

Median error 15.56 0.00 5.55 6.66 0.00 35.67 42.80
SD 4.89 0.46 1.22 1.28 9.41 24.95 26.18
Max. error 28.89 4.44 8.89 8.89 15.37 30.77 30.77
Min. error 4.44 0.00 2.23 4.45 0.00 50.23 100.23
Range 24.44 4.45 6.67 4.45 18.21 27.29 18.21
2 Mean error 14.83 0.89 5.23 5.17 6.00 35.27 34.86
Median error 15.56 0.00 4.44 4.44 0.00 28.53 28.53
SD 4.31 0.22 1.75 1.82 5.10 40.60 42.23
Max. error 24.44 2.23 15.56 11.12 9.12 63.67 45.50
Min. error 2.22 0.00 2.22 2.22 0.00 100.00 100.00
Range 22.22 2.22 13.34 8.89 9.99 60.04 40.01
3 Mean error 11.54 0.20 0.96 1.23 1.73 8.32 10.66
Median error 11.11 0 1.11 2.22 0 9.99 19.98
SD 4 0.07 0.3 0.29 1.75 7.50 7.25
Max. error 22.22 2.22 2.22 2.22 9.99 9.99 9.99
Min. error 4.44 4.44 4.44 4.44 100.00 100.00 100.00
Range 17.78 6.66 6.66 6.66 37.46 37.46 37.46
4 Mean error 19.37 1.57 4.58 4.78 8.11 23.64 24.68
Median error 18.75 1.56 4.69 6.25 8.32 25.01 33.33
SD 4.27 0.20 0.92 0.71 4.68 21.55 16.63
Max. error 31.25 1.56 7.81 3.12 4.99 24.99 9.98
Min. error 6.25 3.12 1.56 3.12 49.92 24.96 49.92
Range 25 1.56 6.25 0 6.24 25.00 0.00

5 Mean error 18.14 2.09 3.41 3.64 11.52 18.80 20.07

Median error 18.75 1.56 3.12 3.12 8.32 16.64 16.64
SD 3.98 0.43 0.96 1.07 10.80 24.12 26.88
Max. error 28.12 4.69 7.82 7.82 16.68 27.81 27.81
Min. error 9.37 1.56 0.00 1.56 16.65 0.00 16.65
Range 18.75 6.25 7.81 9.38 33.33 41.65 50.03
6 Mean error 16.50 0.20 0.62 1.01 1.21 3.76 6.12
Median error 16.41 0.79 0.78 0.78 4.81 4.75 4.75
SD 4.35 0.29 0.15 0.1 6.67 3.45 2.30
Max. error 29.69 0 0 0 0 0.00 0.00
Min. error 6.25 0 0 1.56 0 0.00 24.96
Range 23.44 0 0 1.56 0 0.00 6.66
Average – – – – 4.18 21.98 22.79

4.3. Discussion
4.3.1. Absolute results
This study proposes the RSBL model to improve predictive
ability of logit model for corporate failure because logit model is
preferred by industrial users. Table 2 shows the resampling results.
The RSBL model performed much better than all the classical sta-
tistical models (e.g., MDA, logit model, and probit model). The RSBL
model produced the best mean accuracy for the first five data sets
(84.19%, 85.17%, 88.46%, 80.63%, and 81.86%). For the sixth data set,
the RSBL model was outperformed by MDA, but there was little dif-
ference in their performance (83.50% vs. 83.70%, respectively).
Thus, the results for mean accuracy suggest that the RSBL model
is more suitable for bankruptcy prediction than MDA, logit model,
and probit model. In terms of median accuracy for short-term
bankruptcy prediction, both the RSBL model and MDA produced
the best ratio for the first three data sets (84.44%, 84.44%, and
88.89%, respectively). For medium-term bankruptcy prediction,
the RSBL model showed the best ratio for the fourth and fifth data
sets (81.25%). These results indicate that the RSBL model per-
formed at least as well as the other models. In terms of minimum
accuracy (i.e., the maximum error rate), the RSBL model produced
the best ratio for the six data sets (71.11%, 75.56%, 77.78%, 68.75%,
71.88%, and 70.31%, respectively). Thus, the RSBL model outper-
formed MDA, logit model, and probit model in terms of maximum
accuracy (i.e., the minimum error rate). The accuracy results for the
mean, the median, the minimum, and the maximum suggest that
the RSBL model is more accurate than MDA, logit model, and probit
model for bankruptcy prediction.
We considered the standard deviation and the range to exam-
ine the stability of the models. In terms of the standard deviation,
the RSBL model generated the best ratio on the datasets repre-
sented by all variables and t-test variables: 4.89 and 4.31 for
the first and second data sets, respectively, and 4.37 and 3.98
for the fourth and fifth data sets, respectively. However, MDA
produced the best standard deviation for the third (3.91) and
sixth (4.06) data sets. These results suggest that the RSBL model
performs better than the other models. In terms of the range,
the RSBL model showed the best ratio for the first (24.44), second
(22.22), third (17.78), fifth (18.75), and sixth (23.44) data sets.
Thus, these results indicate that the RSBL model is superior to
and thus more suitable than MDA, logit model, and probit model
for bankruptcy prediction.
The best-time index refers to the number of times the model
provided the best ratio. As shown in Table 2, the RSBL model
provided the best ratio for the six data sets 6, 6, 4, 5, 5, and 3
times, respectively. This indicates that the RSBL model is more
accurate than MDA, logit model, and probit model on five occa-
sions and less accurate than to MDA on one occasion. Thus, these
absolute results indicate that the RSBL model can dramatically
improves the predictive ability of statistical models, including
logit model.
1386 H. Li et al. / Knowledge-Based Systems 24 (2011) 1380–1388
As shown in Table 2, the RSBL model performed slightly better
than the other models for the initial data sets and performed
slightly better or even worse when the feature selection method
was used. The feature selection method improved the predictive
performance of classical models, and thus, all the models performed
better with this method. More specifically, when this method was
used, the RSBL model increased the accuracy of the classical models
just slightly or even had a negative effect. This indicates that group
decisions by logit agents are more useful when a single model does
not provide excellent performance. If a single model produces
excellent performance, then the use of group decisions by logit
agents may reduce the accuracy of the model.
4.3.2. Test of significance
We tested the three hypotheses, which addressed whether
there will be any difference in predictive performance between
the RSBL model and MDA, logit model, and probit model. As shown
in Table 3, H1 was accepted at the 1% level of significance for the
first, second, fourth, and fifth data sets. H2 and H3 were also ac-
cepted at the 1% level of significance for the first, second, fourth,
and fifth data sets. Thus, H1, H2, and H3 were accepted four times.
For the third set, the test did not accept H1 but accepted H2 and H3
at the 1% level of significance. For the sixth data set, the test did not
accept H1 but accepted H2 and H3 at the 1% level of significance.
Thus, H2 and H3 were accepted two more times. In sum, H1 was ac-
cepted four times, whereas H2 and H3, six times. These results indi-
cate that the RSBL model has significant positive effects on
predictive performance of the three classical statistical models
for bankruptcy prediction, verifying the effectiveness of the pro-
posed RSBL model.
4.3.3. Results of performance improvement
Though we have provided some evidence that RSBL provides
significant improvement on predictive ability of MDA, logit, and
probit, yet one may wonder how worse the three classical statis-
tical models are compared with RSBL. As shown in Table 4, al-
most all of percentage point’s indicators of error improvement
are positive. This means that MDA, logit and probit were inferior
to RSBL because they produced more error rates. MDA, logit and
probit were inferior to RSBL respectively by 4.18%, 21.98%, and
22.79% from the perspective of average performance of the six
statistical indices. Note that MDA, logit and probit outperformed
RSBL by 100% in terms of minimum error since these three mod-
els produced 100% accuracies in maximum accuracy while RSBL
produced 95.56% in this index. Without this outlier, MDA, logit
and probit were worse than RSBL, respectively, by 7.16%,
25.47%, and 26.30% in terms of predictive performance. RSBL
was significantly better than logit and probit in terms of predic-
tive ability. RSBL was also significantly better than MDA in
predictive ability since either 4.18% or 7.16% is a significant
improvement on predictive performance. These results provide
clear evidence that the proposed model can dramatically im-
proves the predictive ability of logit model.
5. Conclusion and limitations
We combined logit model and random subspace approach to
propose the RSBL model and used it to forecast corporate failure
in China. The results indicate that the proposed model performs
significantly better than the classical models (i.e., MDA, logit mod-
el, and probit model) in predicting corporate failure. Predictive
abilities of MDA, logit and probit are worse than that of RSBL,
respectively, by 4.18%, 21.98%, and 22.79% (or 7.16%, 25.47%, and
26.30% after deleting outliers). Thus, these results suggest that
the RSBL model helps industrial users to minimize loss from
corporate bankruptcy in uncertain environments and demonstrate
the feasibility and effectiveness of the RSBL for bankruptcy predic-
tion. The RSBL model is not complex: It injects randomness and
committee decisions into logit model. Industrial users typically
construct logit models for bankruptcy prediction by using all avail-
able data. In this regard, the RSBL model extends logit model
slightly by randomly subspacing the available data and then induc-
ing a vote inside the committee of logit models. The results indi-
cate that this treatment allows logit model to retain its advantages.
This study has some limitations. We attempted to integrate
bagging with logit model for bankruptcy prediction, but this type
of integration did not increase predictive performance for corpo-
rate failure in China. Thus, future research should focus on mak-
ing this combination work. The random subspace approach
cannot be directly integrated with MDA because the replaced
variables disobey the assumptions of MDA. Thus, future research
should examine how MDA could be integrated with random sub-
space approach for bankruptcy prediction. Further, data from
other countries (e.g., the US and UK) should be used for a better
understanding of RSBL. Although this study focuses on binary lo-
git model, there are several types of logit models, including mul-
tinomial logit, mixed logit, nested logit, exploded logit, and
ordered logit models. Thus, future research should extend the
RSBL model to other types of logit models. The RSBL model is
limited by its computational burden, which exceeds that of a
single logit model. However, this computation can be imple-
mented with little difficulty if some computer-aided tools are
available. Logit model can be used for real-world risk manage-
ment through manual computation or computer-aided computa-
tion. When a single logit model is used with manual
computation, the user simply develops a logit model from avail-
able data and uses the model to predict the risk associated with
the current problem. However, users typically employ some
computer-aided tools (e.g., SPSS, SAS, and Matlab) to implement
the single logit model. In this regard, the RSBL model should not
be used manually because it is not easy to generate random
numbers manually. With computer-aided tools, the RSBL model
does not require 11 times the work that a single logit model re-
quires just because there are 11 logit decision agents. First, the
whole data set should be re-featured with replacement to be
11 data sets. Further, those computer-aided tools that can gener-
ate a random number between (0, 1) should be used. By inte-
grating the random number with the total number of features,
a computer-aided tool can retrieve a randomly selected feature.
By repeating the selection process the times of the total number
of features, we can obtain a randomly generated mask, i.e., a set
of features for the data representation, and achieve the data set.
By repeating this process 11 times with a computer-aided tool,
we can obtain 11 data sets and produce 11 logit decision agents
for the data sets, which are used to generate 11 decision out-
puts. We can induce a consensus for the logit decision group
through a vote, and for this, computer-aided tools are ideal.
Although the computational burden associated with the RSBL
model is likely to be similar to that associated with a single logit
model, the proposed model does require more computing re-
sources than statistical models. In this regard, future research
should focus on developing a software package.
Acknowledgements
This study was supported in part by the National Natural Sci-
ence Foundation of China (No. 70801055) and the Zhejiang Provin-
cial Natural Science Foundation of China (No. Y7100008). The
authors are grateful to Prof. Hamido Fujita (the Editor in Chief)
and the three anonymous referees for their helpful comments
and recommendations.
 H. Li et al. / Knowledge-Based Systems 24 (2011) 1380–1388 1387

Appendix A. Variables used for bankruptcy prediction of China companies

No. Names of 30 variables Short-term prediction Medium-term prediction

MDA variables t-test variables MDA variables t-test variables
1 Gross income/sales O O
2 Net income/sales O O
3 Ebit/total asset O O
4 Net profit/total assets O O O
5 Net profit/current assets O O O
6 Net profit/fixed assets O
7 Profit margin O O
8 Net profit/equity O O O
9 Account receivable turnover
10 Inventory turnover
11 Account payable turnover O
12 Total assets turnover O O O
13 Current assets turnover O O O
14 Fixed assets turnover O O
15 Current ratio O O
16 Cash/current liability O O
17 Asset-liability ratios O O O O
18 Equity/debt ratio O O
19 Liability/tangible net asset O O
20 Liability/equity market value O O
21 Interest coverage ratio O
22 Growth rate of primary business O
23 Growth rate of total assets O O O
24 Current assets/total assets O
25 Fixed assets/total assets O
26 Equity/fixed assets
27 Current liability/total liability O O
28 Earning per share O O O
29 Net assets per share O O
30 Cash flow per share O O

References [17] K. Keasey, R. Watson, Non-financial symptoms and the prediction of small
company failure: Atest of Argenti’s hypotheses, Journal of Business Finance
and Accounting 14 (3) (1987) 335–354.
[1] E.I. Altman, Financial ratios discriminant analysis and the prediction of
[18] K. Keasey, P. McGuiness, H. Short, Multilogit approach to predicting corporate
corporate bankruptcy, The Journal of Finance 23 (4) (1968) 589–609.
failure – further analysis and the issue of signal consistency, Omega 18 (1)
[2] S. Balcaen, H. Ooghe, 35 years of studies on business failure: an overview of the
(1990) 85–94.
classical statistical methodologies and their related problems, The British
[19] P.R. Kumar, V. Ravi, Bankruptcy prediction in banks and firms via statistical
Accounting Review 38 (2006) 63–93.
and intelligent techniques – a review, European Journal of Operational
[3] W. Beaver, Financial ratios predictors of failure, Journal of Accounting Research
Research 180 (2007) 1–28.
4 (1966) 71–111.
[20] L.I. Kuncheva, Combining Pattern Classifiers: Methods and Algorithms, Wiley,
[4] J. Berkson, Application of the logistic function to bio-assay, Journal of the
NJ, 2004.
American Statistical Association 39 (1944) 357–365.
[21] L. Lam, C.Y. Suen, Application of majority voting to pattern recognition: an
[5] C.I. Bliss, The method of probits, Science 79 (1934) 409–410.
analysis of its behavior and performance, IEEE Transaction on Systems, Man,
[6] L. Breiman, Bagging predictors, Machine Learning 24 (1996) 123–140.
and Cybernetics 27 (1997) 553–568.
[7] L. Breiman, Arcing classifiers, Annals of Statistics 26 (1998) 801–849.
[22] H. Li, J. Sun, Ranking-order case-based reasoning for financial distress
[8] L. Breiman, Random forecast, Machine Learning 45 (2001) 5–32.
prediction, Knowledge-Based Systems 21 (8) (2008) 868–878.
[9] V. Cho, MISMIS – a comprehensive decision support system for stock market
[23] H. Li, J. Sun, Forecasting business failure in China using case-based reasoning
investment, Knowledge-Based Systems 23 (6) (2010) 626–633.
with hybrid case representation, Journal of Forecasting 29 (5) (2010) 486–501.
[10] P.J. Fitzpatrick, A comparison of ratios of successful industrial enterprises with
[24] N. Lim, Classification by ensembles from random partitions using logistic
those of failed firms, Certified Public Accountant, New York, 1932.
regression models, PhD Thesis, Department of Applied Mathematics and
[11] W. Hardle, Y. Lee, D. Schafer, et al., Variable selection and oversampling in the
Statistics, Stony Brook University, 2007.
use of smooth support vector machines for predicting the default risk of
[25] D. Martin, Early warning of bank failure: a logit regression approach, Journal of
companies, Journal of Forecasting 28 (6) (2009) 512–534.
Banking and Finance 1 (1977) 249–276.
[12] T. Hastie, R. Tibshirani, J. Friedman, The Elements of Statistical Learning: Data
[26] T.E. McKee, Rough sets bankruptcy prediction models versus auditor signaling
Mining Inference and Prediction, Springer, New York, NY, 2001.
rates, Journal of Forecasting 22 (8) (2003) 569–586.
[13] T.K. Ho, The random space method for constructing decision forests, IEEE
[27] T.E. McKee, M. Greenstein, Predicting bankruptcy using recursive partitioning
Transactions on Pattern Analysis and Machine Intelligence 20 (8) (1998) 832–
and a realistically proportioned data set, Journal of Forecasting 19 (3) (2000)
844.
219–230.
[14] Y. Hu, J. Ansell, Retail default prediction by using sequential minimal
[28] C. Nam, T. Kim, N. Park, et al., Bankruptcy prediction using a discrete-time
optimization technique, Journal of Forecasting 28 (8) (2009) 651–666.
duration model incorporating temporal and macroeconomic dependencies,
[15] R. Hwang, K. Cheng, J. Lee, A semiparametric method for predicting
Journal of Forecasting 27 (6) (2008) 493–506.
bankruptcy, Journal of Forecasting 26 (5) (2010) 317–342.
[29] J. Ohlson, Financial ratios and the probabilistic prediction of bankruptcy,
[16] S. Jones, D.A. Hensher, Predicting firm financial distress: a mixed logit model,
Journal of Accounting Research 18 (1) (1980) 109–131.
Accounting Review 79 (4) (2004) 1011–1038.
1388 H. Li et al. / Knowledge-Based Systems 24 (2011) 1380–1388
[30] D.E. O’Leary, Knowledge acquisition from multiple experts: an empirical study,
Management Science 44 (1998) 1049–1058.
[31] J. Pacheco, S. Casado, L. Nunez, A variable selection method based on tabu
search for logistic regression models, European Journal of Operational
Research 199 (2) (2009) 506–511.
[32] J. Pacheco, S. Casado, L. Núñez, O. Gómez, Analysis of new variable selection
methods for discriminant analysis, Computational Statistics and Data Analysis
51 (3) (2006) 1463–1478.
[33] P.F. Pai, M.F. Hsu, M.C. Wang, A support vector machine-based model for
detecting top management fraud, Knowledge-Based Systems 24 (2) (2011)
314–321.
[34] P. Ravisankar, V. Ravi, Financial distress prediction in banks using group
method of data handling neural network, counter propagation neural network
and fuzzy ARTMAP, Knowledge-Based Systems 23 (8) (2010) 823–831.
[35] D. Ruan, G. Chen, E.E. Kerre, G. Wets (Eds.), Intelligent Data Mining:
Techniques and Applications. Studies in Computational Intelligence, vol. 5,
Springer, 2005.
[36] J. Sun, H. Li, Data mining method for listed companies’ financial distress
prediction, Knowledge-Based Systems 21 (1) (2008) 1–5.
[37] J. Sun, H. Li, Financial distress prediction based on serial combination of
multiple classifiers, Expert Systems with Applications 36 (4) (2009) 8659–
8666.
[38] C.F. Tsai, Feature selection in bankruptcy prediction, Knowledge-Based
Systems 22 (2) (2009) 120–127.
[39] F.M. Tseng, L. Lin, A quadratic interval logit model for forecasting bankruptcy,
Omega 33 (2005) 85–91.
[40] F.M. Tseng, Y.C. Hu, Comparing four bankruptcy prediction models: logit,
quadratic interval logit, neural and fuzzy neural networks, Expert Systems
with Applications 37 (3) (2010) 1846–1853.
[41] D. West, S. Dellana, J. Qian, Neural network ensemble strategies for financial
decision applications, Computers and Operations Research 32 (2005)
2543–2559.