Expert Systems with Applications 28 (2005) 655–665

www.elsevier.com/locate/eswa

Hybrid mining approach in the design of credit scoring models

Nan-Chen Hsieh*
Department of Information Management, National Taipei College of Nursing, No. 365, Min-Ten Road 11257, Taipei, Taiwan, ROC

Abstract

Unrepresentative data samples are likely to reduce the utility of data classifiers in practical application. This study presents a hybrid mining
approach in the design of an effective credit scoring model, based on clustering and neural network techniques. We used clustering
techniques to preprocess the input samples with the objective of indicating unrepresentative samples into isolated and inconsistent
clusters, and used neural networks to construct the credit scoring model. The clustering stage involved a class-wise classification process. A
self-organizing map clustering algorithm was used to automatically determine the number of clusters and the starting points of each cluster.
Then, the K-means clustering algorithm was used to generate clusters of samples belonging to new classes and eliminate the unrepresentative
samples from each class. In the neural network stage, samples with new class labels were used in the design of the credit scoring model.
The proposed method demonstrates by two real world credit data sets that the hybrid mining approach can be used to build effective credit
scoring models.
q 2005 Elsevier Ltd. All rights reserved.

Keywords: Data mining; Credit scoring model; Clustering; Class-wise classification; Neural network

1. Introduction of the deficiencies of the linear discriminant analysis model.

In response to the recent growth of the credit industry and
the management of large loan portfolios, the industry is
actively developing credit and behavioral scoring models.
Credit scoring models help to decide whether to grant credit
to new applicants using customer’s characteristics such as
age, income and martial status (Chen & Huang, 2003).
Behavioral scoring models help to analyze purchasing
behavior of existing customers (Setiono, Thong, & Yap,
1998). This study utilizes a hybrid mining approach in the
design of credit scoring models to support credit approval
decisions.
A simple parametric statistical model, linear discrimi-
nant analysis, was the first model employed for credit
scoring. However, analysts have questioned the appropri-
ateness of linear discriminant analysis for credit scoring
because of the categorical nature of the credit data and the
fact that the covariance matrices of the good and bad credit
classes are not likely to be equal. Researchers are now
investigating more sophisticated models to overcome some
* Tel./fax: C886 2 2822 7101 2200.
E-mail address: nchsieh@mail1.ntcn.edu.tw
0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2004.12.022
One of the efforts is leading to the investigation of
non-parametric statistical methods, that is, neural networks
for scoring applications.
Swales and Yoon (1992) used neural networks to
differentiate stocks. They found that the neural networks
performed significantly better than the linear multiple
discriminant models. Tam and Kiang (1992) compared the
neural network approach with a linear classifier model,
logistic regression model, neural network model, and
decision tree model to predicate bank failures. They
demonstrated that neural networks are more accurate,
adaptive, and robust than to other methods. Zhang et al.
(1999) similarly showed that neural networks are signifi-
cantly better than logistics regression models in bankruptcy
predication. Desai, Crook, and Overstreet. (1996) per-
formed an experiment that the performance of discriminant
analysis is comparable to the performance of back-
propagation neural networks in classifying loan applicants
into good and bad credit. They pointed out that more
customized architectures might be necessary for building
effective generic models to classify consumer loan appli-
cations in the credit union environment. West (2000)
investigated the credit scoring accuracy of five neural
656 N.-C. Hsieh / Expert Systems with Applications 28 (2005) 655–665
network models, and reported that a hybrid architecture of
neural network models should be considered for credit
scoring applicants. That is, non-parametric and hybrid
design architecture will be very useful in developing
effective credit scoring models. The meaning of an effective
credit scoring model is two-fold, relating to accuracy and
easy interpretation of classified results.
Generally, hybrid design architectures for creating
more accurate classifiers that uses neural networks guided
by clustering algorithms (Gopalakrishnan, Sridhar, &
Krishnamurthy, 1995; Sung, 1998) or genetic algorithms
(Kim & Street, 2004) were proposed. The other studies focus
on integrating the multivariate analysis and neural network to
increase clustering accuracy. Punj and Steward (1983)
proposed a two-stage method that combines Ward’s
minimum variance method and the K-means method.
Balakrishnan et al. (1996) integrated unsupervised FSCL
neural networks with the K-means method. Kuo, Ho, and Hu
(2002) proposed a two-stage method, which uses the self-
organizing map to determine the number of clusters and then
employs the K-means algorithm to classify samples into
clusters. This study addresses the benefit by investigating a
simple but effective hybrid utility of clustering and neural
network techniques in the design of a credit scoring model.
As shown in Fig. 1, the proposed hybrid scoring model has
two processing stages. In the clustering stage, samples are
grouped into homogeneous clusters. In particular, unrepre-
sentative samples are indicated as ‘isolated’ and ‘inconsist-
inconsistent’ clusters. Herein, isolated clusters are thinly
populated clusters; inconsistent clusters are those clusters
with inconsistent class values. To improve the performance
Fig. 1. A hybrid mining credit scoring system.
of the credit scoring model, isolated clusters should be
eliminated. Since inconsistent clusters may be important and
eliminating them may cause the loss of valuable information,
investigating inconsistent clusters in detail will be very
helpful in understanding customers. In the neural network
stage, the neural network used samples with new class labels
to create models for predicate consumer loans.
For a better understanding of our study, Section 2 begins
with an overview of credit scoring models in general.
Section 3 offers a hybrid credit scoring model and considers
why it should perform better than other credit scoring
models. Following this discussion, Section 4 empirically
tests the hybrid credit scoring model using two real world
credit data sets. Finally, Section 5 discusses the findings of
the experiment and offers observations about practical
applications and directions for future research.
2. Description the analysis methodology
2.1. Credit and behavioral scoring models
Credit and behavioral scoring models (Thomas, 2000) are
one of the most successful applications of statistical and
operational research modelling in finance and banking, and
the number of scoring analysts in the industry is constantly
increasing. The main objective of both credit and behavioral
scoring models is to classify samples into homogeneous
groups (Lancher, Coates, Shanker, & Fant, 1995). Hence,
scoring problems are related to classification by statistical
analysis (Hand, 1981; Hsieh, 2004; Johnson & Wichern,
 N.-C. Hsieh / Expert Systems with Applications 28 (2005) 655–665
1998; Morrison, 1990), especially classification by neural
networks in the field of data mining (Lancher et al., 1995).
Applying to credit approval databases, classification analysis
for credit scoring is used to categorize a new applicant as
either accepted or rejected with respect to characteristics
such as age, income and martial status (Chen & Huang,
2003). On the other hand, classification analysis for
behavioral scoring is used to describe the behavior of
existing customers by using behavioral scoring variables and
also to predict future purchasing behavior or credit status of
existing customers (Setiono et al., 1998).
Until recently, the building of both scoring models has
been always based on a pragmatic approach; therefore, a best
and most standard scoring model for every unique circum-
stance undoubtedly does not exist. Most previous studies
have focused on building more accurate credit or behavioral
scoring models and increasing the accuracy of the classifi-
cation model with various statistical techniques. However,
creating a good performance model for customer credit
management is difficult because unrepresentative samples
exist. Therefore, even when building highly accurate scoring
models, misclassification patterns appear frequently (Kim &
Sohn, 2004).
This study intends to draw significantly from data mining
perspectives, providing general model which integrates
clustering and neural network techniques for applicant
approval analysis, including necessary preprocessing of the
credit data sets. Finding isolated and inconsistent clusters to
support a standard model building process will be very
useful. The proposed hybrid credit scoring model can serve
as a tool to validate the effect of clustering techniques in
practical credit scoring analysis applications.
2.2. Unrepresentative samples prevent the utility
of credit scoring models
To date, many studies have proposed various credit
scoring models to find accurate boundaries to classify
the samples. Since there are practical reasons to prevent
perfect classification, an absolutely accurate credit scoring
model does not exist. The learning of credit scoring models
is based on historical samples with known classes, which are
usually two classes: ‘good credit’ and ‘bad credit’. In
theory, every new sample has a fixed class membership to
the known classes, but this theory is unachievable in real
world application, since samples with known classes are
quite limited, and the collected samples are often unrepre-
sentative of the population to be analyzed.
Fig. 2 shows how unrepresentative samples prevent
accurate predictions when the classes are determined by a
two-dimensional feature space. The ideal classification
boundary is a simple line. The sample points with circles are
limited and unrepresentative, and cause credit scoring models
to find an inaccurate classification boundary. However, if
samples are further classified into more classes, the accuracy
of credit scoring models increases. Anyway, the basic
657
Fig. 2. Unrepresentative samples prevent accurate predictions.
requirement of correct classification requires that feature
values of samples conform to specify class membership. In
many situations, the correct class membership is not
absolutely known. The uncertainty of class membership of
the samples will increase with the learning complexity of the
credit scoring models. Clustering is a technique that uses
sample’s feature values to classify samples into clusters with
shared characteristics. Preprocessing samples by clustering
technique is useful in defining proper class membership.
2.3. Assessing hybrid architecture to the credit
scoring analysis
For credit scoring analysis, many studies have shown that
neural networks perform significantly better than statistical
techniques such as linear discriminate analysis,
multiple discriminate analysis and logistic regression
analysis (Desai et al., 1996; Lancher et al., 1995; Malhotra
& Malhotra, 2003). The application of neural networks to
credit scoring analysis is a promising research area and is a
challenge for a variety of marketers (Kim & Sohn, 2004;
Malhotra & Malhotra, 2003; Sharda & Wilson, 1996;
Vellido, Lisboa, & Vaughan, 1999; Zhang, Hu, Patuwo, &
Indro ,1999). Lee, Chiu, Lu, and Chen. (2002) explored the
performance of credit scoring by integrating the back-
propagation neural networks with traditional discriminate
analysis. Hush and Horne (1993) presented the application
of clustering techniques to build more accurate classifiers in
the context of RBF-based neural networks. Hruschka and
Natter (1999) suggested that the feedforward neural
networks obtained better solutions comparable to those of
partitional clustering techniques, such as K-means. Gopa-
lakrishnan et al. (1995) presented that by employing
clustering techniques, the training effort of neural networks
may be drastically reduced, and clustering techniques can be
used to build more robust classifiers. Natter (1999) found
that a feedforward neural network incorporated both
clustering and discriminate analysis, producing more
homogeneous solutions.
Importantly, these authors noted that weaknesses in the
built models are due to the model’s architecture, rather than
658 N.-C. Hsieh / Expert Systems with Applications 28 (2005) 655–665
a failure of neural network itself in general. They thus infer
that neural networks can effectively be used in credit
scoring, but the architecture design of the credit scoring
model determines the degree of success observed. Along
these lines, this study investigates a hybrid credit scoring
model based on clustering and neural networks for building
more effective classifiers. Various issues, such as number
and nature of training samples, and the network architecture,
can be resolved with the help of clustering techniques.
3. Offering a hybrid credit scoring model
Typically, the network architecture affects the ability of a
neural network. The design parameters of a neural network
include network topology, number of input nodes, number
of nodes in each hidden layers, number of output nodes and
activation function selected. However, the quality of
training samples leads to misclassification, and the mis-
classification problem is due to unrepresentative samples
that drastically reduce the accuracy of the neural network
classifier. A way of identifying the unrepresentative samples
is to look for isolated and inconsistent clusters.
Isolated clusters are thinly populated clusters; inconsist-
ent clusters are clusters with inconsistent class values. A
fundamental assumption made here is that the isolated
clusters are far away from the core clusters. To improve the
quality of classification, isolated clusters can be eliminated
from training samples. However, inconsistent clusters may
be important and eliminating them may cause the loss of
valuable information. Handling inconsistent clusters in
detail and investigating their properties will be very helpful
for understanding customers. A way of achieving this goal is
to employ clustering techniques to preprocess the training
samples. Finally, we stand to gain if we can keep the feature
values of training samples more homogeneous, would allow
an efficient classifier to be build. The training samples
filtered by clustering are used to successively train the credit
scoring model.
The proposed credit scoring model is a hybrid approach
using clustering techniques to preprocess samples into
homogeneous clusters, and neural network techniques to
build classifiers. The clustering process depends on
sample’s feature values and optionally makes use of the
original class label to generate clusters. The cluster labels
are thus added as new class labels for each sample. This is a
so-called class-wise classification process, that is, the
clustering process generated clusters of samples belonging
to class and eliminates the unrepresentative samples from
each class.
Since the quality of clustering is not easy to control, Punj
and Steward (1983) suggest that a two-stage clustering
method consisting of a hierarchical model, such as Ward’s
minimum variance method, followed by a non-hierarchical
method, such as K-means method, can provide a better
solution. The key is that hierarchical methods can determine
the candidate number of clusters and starting point of each
cluster, and non-hierarchical methods can provide better
clustering result with the specified information. This study
has chosen to replace the hierarchical method with a
computational intelligent technique, self-organizing map.
For the sake of in the hierarchical method, once an
observation has been assigned to a cluster, it should not
be moved. However, a self-organizing map can continually
update the observation to the closest cluster even initial
conditions have been assigned. Therefore, the candidate
number of clusters and the starting point of each cluster can
easily be determined. The second stage follows the
non-hierarchical, K-means method, to determine the final
solution due to its efficiency. Next, an intelligent search
method is used to find the best neural network architecture,
then this neural network applies the clustering results to
predicate consumer loans. In Section 4, two experiments
were conducted to show how this hybrid credit scoring
model proceeded, and verify that the proposed model
performs better than the other approaches.
4. Empirical analysis
4.1. Real world credit data sets
Two real world credit data sets from the UCI Repository
of Machine Learning Databases (http://www.niaad.liacc.up.
pt/statlog/datasets.html) were used to test the predicate
accuracy of the credit scoring models. The German credit
data set consisted of a set of loans given to a total of 1000
applicants, 700 samples of creditworthy applicants and 300
samples where credit should not be extended. For each
applicant, 20 variables described credit history, account
balances, loan purpose, loan amount, employment status,
and personal information. The Australian credit data set is a
similar data set with 690 samples, in which 468 samples
were accepted and maintain good credit and 222 samples
were accepted, but became delinquent. To protect the
confidentiality of these data sets, sensitive variables of these
two credit sets were transformed to symbolic data. This
study employed all the variables as inputs, whereas some
previous studies used summary variables (Kuo et al., 2002),
or ranking importance variables (Kim & Sohn, 2004; Lee
et al., 2002; Sung, 1998) to build the credit scoring models.
4.2. A two-stage clustering technique to preprocess
the credit date sets
The goal of developing a neural network is to find an
appropriate topology for the network and then train the
network so that the network gradually learns the desired
input–output functionality. A network can be trained in two
ways, by supervised and unsupervised learning. In super-
vised learning, the network is presented with training
samples of known input–output relationships. The trained
 N.-C. Hsieh / Expert Systems with Applications 28 (2005) 655–665 659

network can be used to mimic the presented input–output

behavior, and then the network is tested to verify whether it
is able to produce correct output. Several supervised neural
network models exist. West (2000) has investigated the
credit scoring accuracy of five neural network models, for
both the German and Australian credit data sets. However,
due to the quality of samples, the improvement of correct
classification rate is limited.
In unsupervised learning, the output data is not available
and usually not even known beforehand. Therefore, the
network tries to find similarities between input samples.
Similar samples form clusters that automatically constitute
the output of the neural network. As reported in Sung
(1998), clustering samples before training contributes to the
network’s achieving a higher accuracy. In here, two credit
data sets have predefined labels describing their loan status.
However, credit data sets might contain unrepresentative
Fig. 3. Applicant distribution map of German credit set.
samples, which will affect the accuracy and representation
of credit scoring models. For this reason, this study utilizes a good credit applicants, and the bad credit applicants would
two-stage learning method (Kuo et al., 2002), which have more of an opportunity to repay the delayed payments
employs a self-organizing map to determine the number than the other bad credit applicants. However, the applicants
of clusters and the starting points of each cluster and then with the same credit status but distributed over different
employs the K-means algorithm to find the final solution. clusters may have different degrees of uncertainty to their
Since no rule exists for determining the best training credit status.
parameters, the training parameters of both clustering Table 1 shows that neither credit data sets has significant
methods are obtained by trial and error. After many times isolated clusters. For easy understanding of the variations of
of testing, the best segmentation number of the two credit credit status, Figs. 3 and 4 depict the applicant distribution
data sets was found to be four. maps of two credit data sets. In the German credit data set,
Suppose that the input samples are representative, and cluster-1 has 235 samples of good credit and 3 samples of
then most of the applicants should be validated clearly in bad credit. Since the proportion of samples having bad
terms of credit status, whether good or bad. However, some credit is small in cluster-1, these samples can be ignored at
applicants may be misclassified by those credit scoring this time. Cluster-2 is an inconsistent cluster, which contains
models, mainly due to unrepresentative samples, or possibly samples of both good and bad credit. The behavioral trend
uncertain definition of class membership of the samples. of these applicants might be changed in the near future.
Worthy of notice, the future behavioral trend of these To indicate this behavioral trend in the built credit scoring
unrepresentative applicants might be changed in the near model, cluster-2 was further divided into cluster-21 and
future (Kim & Sohn, 2004). For the applicants in the
inconsistent cluster, the good credit applicants would have
more of an opportunity to delay future payments than other
Table 1
The clustering results of German and Australian credit sets

German credit data Australian credit data

Good credit, 700 (1) Bad Good credit, 468 (1) Bad
credit, credit,
300 (2) 222 (0)
Cluster Number of Credit Cluster Number of Credit
ID samples status ID samples status
1 238 235 (1) 1 278 0
3 (2) 278 (1)
2 211 105 (1) 2 190 0
106 (2) 190 (1)
3 191 0 3 106 106 (0)
191 (2) 0
4 360 360 (1) 4 116 116 (0)
0 0
Fig. 4. Applicant distribution map of Australian credit set.
660 N.-C. Hsieh / Expert Systems with Applications 28 (2005) 655–665
cluster-22 according to their original credit status. More-
over, cluster-1 and cluster-4 are clusters with different
degrees of uncertainty to good credit status. Finally, the
German credit set has five clusters. The original class labels
were ignored and each sample was set up with new class
labels by the cluster identifiers. On the other hand, the
sample quality of the Australian credit set was relatively
good to the German credit set. Each sample was given a new
class label using the cluster identifiers.
4.3. Building the credit scoring model
To provide a reliable estimate and minimize the impact
of data dependency in developing credit scoring models,
k-fold cross-validation was used to generate random
partitions of the credit data sets (West, 2000). In this
procedure, the credit data set was divided into k
independent groups. A neural network was trained using
the first kK1 groups of samples and the trained neural
network was tested using the kth group. This procedure
was repeated until each of the groups has been used as a
test set once. The overall scoring accuracy was reported
as an average across all k groups. A merit of cross-
validation is that the credit scoring model is developed
with a large proportion of the available data and that all
the data is used to test the resulting models. In this
experiment, the value of k was set to 10 and thus forms a
10-fold cross-validation. An estimate from 10-fold cross
validation is likely to be more reliable than an estimate
from a common practice of using a singleton holdout set.
The selection of the neural network is another important
factor in building a credit scoring model. This study used
supervised feedforward MLP (multi-layer perceptrons)
neural networks trained by back-propagation and gradient
descent, in which the output of the network acts as
feedback and re-enters the network as an input, afford more
sensitive modeling of real data by iteratively adjusting the
weights until an optimal solution is found. Network
architecture decisions included the number of hidden
layers and number of neurons in each layer. For the neural
input layer, each input variable with a small number of
discrete integer values was recoded as a nominal variable,
and the categorical variables were encoded using the ‘One-
to-N’ encoding method. The numeric variables were scaled
by ‘Minimax’ encoding method in the range (Chen &
Huang, 2003) before being fed to the network input,
according to the logistic activation function property.
Finally, the number of neurons in the input layer of the
neural networks was the number of variables in the input
samples. One-to-N encoding was used for the neural output
layer with an output neuron dedicated to each of the credit
decision outcomes.
The number of hidden layers and the number of neurons
in each hidden layer are more difficult to define. As
recommended by Hornik, Stinchcombe, and White (1989),
one hidden layer network is sufficient to model a complex
system with any desired accuracy, therefore the employed
neural network has just one hidden layer. For the size of
the hidden layer, a relatively large hidden layer creates a
more flexible scoring model and errors tend to have a low
bias component with a large variance caused by the
tendency for the model to over-fit the training samples,
while a relatively small hidden layer results in a model
with a higher error bias and a lower variance. Therefore,
the design of the hidden layer involves a tradeoff between
error components.
For each of the two credit data sets, an intelligent
search process was used to determine the network
architecture in the feedforward MLP neural network.
The search process started with one hidden layer, and set
the number of hidden neurons equal to the number of
inputs divided by 2. Then neurons were gradually added
to the hidden layer one at a time, retraining each network
at least three times with different initial weight randomiz-
ations and saving the best for comparison with other
architectures. The search process stopped until there is no
further improvement in network performance. In this
experiment, the maximum number of hidden neurons was
rarely required to exceed the number of inputs by more
than two times.
Tables 2 and 3 summarize the results of the neural
network models. To compare results, 10-fold cross-
validation technique was employed, using the confusion
matrix to observe the error percentage and then evaluate the
result. For example, the good credit error is the proportion of
creditworthy applicants who are classified as a bad credit.
Conversely, the bad credit error is the proportion of
applicants who are not creditworthy but are incorrectly
identified as creditworthy. The performance measure
indicates the proportion of samples which are correctly
classified. Each network was trained at least 10 repetitions.
The measures in each repetition were the average in
proportion to the samples of training and testing sets,
while the final results in each group were an average of the 3
best from 10 repetitions. After being preprocessed by
clustering, the accuracy of neural network models increased
significantly.
German credit set was then used to interpret the credit
scoring results. Appendix A lists the distribution of the
relative importance for each input variable using the neural
network. The error sensitivity indicates the performance of
the network if the corresponding input variable is unavail-
able. In general, important input variables have a high error,
indicating that the network performance deteriorates badly
if they are not present. The ratio sensitivity presents the ratio
between the error and the baseline error. If the ratio
sensitivity is closer to one, then indicating the input variable
contributes most to the performance of the network. Finally,
the rank summarizes the input variables in order of
importance.
Herein, the sensitivity analysis of the neural network and
the order of most significant input variables indicate
 N.-C. Hsieh / Expert Systems with Applications 28 (2005) 655–665 661

Table 2
Classification of applicants (German credit data)

German credit data (raw data) German credit data (preprocessed by clustering)
Good credit error Bad credit error Performance Good credit error Bad credit error Performance
Group 1 0.1396 0.1444 0.7740 0.0087 0.0173 0.9906
Group 2 0.1285 0.1335 0.7850 0.0092 0.0487 0.9809
Group 3 0.1650 0.1948 0.8236 0.0109 0.0319 0.9869
Group 4 0.1261 0.1711 0.7988 0.0078 0.0350 0.9839
Group 5 0.1542 0.1392 0.7645 0.0361 0.0874 0.9598
Group 6 0.2475 0.2530 0.7585 0.0069 0.0445 0.9829
Group 7 0.1939 0.2609 0.7844 0.0304 0.0586 0.9933
Group 8 0.1849 0.2011 0.8012 0.0041 0.0298 0.9889
Group 9 0.1472 0.1369 0.8233 0.0379 0.0923 0.9866
Group10 0.1960 0.1704 0.7855 0.0353 0.0696 0.9922
Network architecture 20-15-1 20-15-5
(input–hidden–output)

variables that are worth looking at in detail. Since the
variable ‘Telephone’ is unconcerned with differentiating
applicants, then ‘Status of existing checking account’,
‘Present employment since’ and ‘Personal status and sex’
were the three most differentiating variables. On the other
hand, ‘Duration in month’, ‘Foreign worker’ and ‘Age in
years’ were the three least differentiating variables.
Appendix B depicts the statistical summarized data for
each cluster to verify the proposed method. Only the 12
most significant variables to each cluster were presented in
the order of relative importance.
Statistical summarized data was observed for whole
input variables to determine the differences between the
input variables of each group. The first group (cluster-1,
cluster-4, cluster-21) consisted of clusters tending toward
good credit status. The second group (cluster-22,
cluster-3) consisted of clusters tending toward bad credit
status. The third group (cluster-21, cluster-22) consists of
clusters tending toward changed credit status in the near
future. The first group had nine variables with differences
between cluster-1 with cluster-21, and four variables had
differences between cluster-4 with cluster-21 at 20%
level. Therefore, cluster-21 was more similar to cluster-4
than to cluster-1. Moreover, we compared the input
variables of cluster-1 and cluster-4, the characteristics of
‘Status of existing checking account’, ‘Present employ-
ment since’, ‘Savings account/bonds’, ‘Personal status
and sex’, and ‘Present residence since’ of cluster-1 were
better than those in cluster-4. Moreover, cluster-21 and
cluster-22 were very similar, and the three most
differentiating variables, ‘job’, ‘credit amount’ and ‘age
in years’, were relatively unimportant. Therefore, we can
conclude that cluster-1’s overall applicant credit status
can be judged better than that of cluster-4, and the
behavioral trend of applicants in cluster-21, cluster-22
might be changed in the near future. Also, comparing
cluster-22 and cluster-3, the applicants in cluster-3 were
younger than those in cluster-22, and the average
applicant’s ‘installment rate in percentage of disposable’
in cluster-3 is higher than in cluster-22. So the credit
status of applicants in cluster-3 might be worse than that
of those applicants in cluster-22.

Table 3
Classification of applicants (Australian credit data)

Australian credit data (raw data) Australian credit data (preprocessed by clustering)
Good credit error Bad credit error Performance Good credit error Bad credit error Performance
Group 1 0.0945 0.0961 0.9043 0.0545 0.0044 0.9797
Group 2 0.1377 0.1312 0.8980 0.0143 0.0787 0.9916
Group 3 0.1520 0.1502 0.8783 0.0417 0.1123 0.9733
Group 4 0.1654 0.1608 0.8666 0.0323 0.0794 0.9850
Group 5 0.1177 0.1223 0.9133 0.3208 0.0923 0.9866
Group 6 0.1217 0.1142 0.9000 0.0369 0.1092 0.9716
Group 7 0.1416 0.13128 0.8880 0.0370 0.1344 0.9750
Group 8 0.1295 0.1229 0.9050 0.0336 0.0996 0.9800
Group 9 0.1435 0.1386 0.8850 0.0338 0.0731 0.9860
Group10 0.1431 0.1289 0.8950 0.0360 0.1512 0.9700
Network architecture 13-15-1 13-20-4
(input–hidden–output)
662 N.-C. Hsieh / Expert Systems with Applications 28 (2005) 655–665

5. Conclusions
Input variables Rank Error Ratio Comments

This study has presented a hybrid mining approach in the Credit history 6 0.1862 3.9473 A30: no credits taken/all
credits paid back duly;
design of the credit scoring models. The experience of A31: all credits at this
constructing neural networks for the credit scoring problem bank paid back duly;
indicated that clustering is valuable in building networks of A32: existing credits
paid back duly till now;
high effectiveness. Not only designing an accurate classifier, A33: delay in paying off
clustering techniques were successfully employed to pre- in the past; A34: critical
process the input samples for the purpose of indicating account/other credits
existing (not at this
unrepresentative samples. The clustering results consisted of bank)
normal, isolated and inconsistent three kinds of clusters. Property 7 0.1861 3.9443 A121: real estate; A122:
Dividing samples using this method will be an effective if not A121, building
society savings agree-
means of understanding behavioral trends of each group of ment/life insurance;
applicants, and in building effective credit scoring models. A123: if not A121/
The experimental results demonstrate that such a hybrid A122, car or other;
A124: unknown/no
approach is simple but efficient in the application of credit property
market. How to utilize experimental tools, such as sensitivity Purpose 8 0.1792 3.7984 A40: car (new); A41: car
analysis or genetic algorithm, to identify important input (used); A42: furniture/
equipment; A43: radio/
variables, and the soft-computing method, such as fuzzy sets, television; A44: dom-
to encode inputs into more informative data values, in estic appliances; A45:
conjunction with the domain expert’s knowledge to interpret repairs; A46: education;
A47: (vacation-miss-
the results, are issues worthy of further investigation. ing); A48: retraining;
A49: business; A410:
others
Installment 9 0.1719 3.6440
rate in percen-
tage of dispo-
sable income
Appendix A. Sensitivity analysis of the relative Present resi- 10 0.1596 3.3837
importance of input variables (German credit data set) dence since
Job 11 0.1539 3.2631 A171: unemployed/
unskilled non-resident;
Input variables Rank Error Ratio Comments A172: unskilled-resi-
dent; A173: skilled
Telephone 1 0.4150 8.7968 A191: none; A192: yes, employee/official;
registered under the A174:management/self-
customers name
employed/highly quali-
Status of exist- 2 0.2362 5.0070 A11: x!0 DM; A12: fied employee/officer
ing checking 0%x!200 DM; A13: Other install- 12 0.1385 2.9364 A141: bank; A142:
account xR200 DM/salary ment plans stores; A143: none
assignments for at least 1
Housing 13 0.1371 2.9060 A151: rent; A152: own;
year; A14: no checking A153: for free
account Credit amount 14 0.0979 2.0740 Amount of credit
Present 3 0.2111 4.4737 A71: unemployed; A72:
Number of 15 0.0890 1.8859
employment x!1 year; A73: 1%x! existing credits
since 4 years; A74: 4%x!7 at this bank
years; A75: xR7 years Other debtors/ 16 0.0869 1.8419 A101: none;
Personal status 4 0.1988 4.2136 A91: male, divorced/
guarantors A102: co-applicant;
and sex separated; A92: female, A103: guarantor
divorced/separated/mar- Number of 17 0.0817 1.7306
ried; A93: male, single; people being
A94: male, married/
liable to pro-
widowed; A95: female: vide mainten-
single ance for
Savings 5 0.1950 4.1327 A61: x!100 DM;
Duration in 18 0.0807 1.7099
account/bonds A62: 100%x!500 DM;
month
A63: 500%x!1000 Foreign 19 0.0771 1.6346 A201: yes; A202: no
DM; A64: xR1000 DM; worker
A65: unknown/without
Age in years 20 0.0554 1.1747 Applicant’s age
savings account
 N.-C. Hsieh / Expert Systems with Applications 28 (2005) 655–665 663

Appendix B. Statistical summarized data

664 N.-C. Hsieh / Expert Systems with Applications 28 (2005) 655–665

References
Balakrishnan, P. V. S., Cooper, M. C., Jacob, V. S., & Lewis, P. A. (1996).
Comparative performance of the FSCL neural net and K-means
algorithm for market segmentation. European Journal of Operational
Research, 93, 346–357.
Chen, M. C., & Huang, S. H. (2003). Credit scoring and rejected instances
reassigning through evolutionary computation techniques. Expert
Systems with Applications, 24, 433–441.
Desai, V. S., Crook, J. N., & Overstreet, G. A. (1996). A comparison of neural
networks and linear scoring models in the credit union environment.
European Journal of Operational Research, 95(1), 24–37.
 N.-C. Hsieh / Expert Systems with Applications 28 (2005) 655–665
Gopalakrishnan, M., Sridhar, V., & Krishnamurthy, H. (1995). Some
applications of clustering in the design of neural netwroks. Pattern
Recognition Letters, 16, 59–65.
Hand, D. J. (1981). Discrimination and classification. New York:
Wiley.
Hornik, K., Stinchcombe, M., & White, H. (1989). Multilayer feedforward
networks are universal approximations. Neural Networks, 2, 336–359.
Hruschka, H., & Natter, M. (1999). Comparing performance of feedforward
neural nets and K-means for cluster-based market segmentation.
European Journal of Operational Research, 114, 346–353.
Hsieh, N. C. (2004). An integrated data mining and behavioral scoring
model for analyzing bank customers. Expert Systems With Applications,
27(4), 623–633.
Hush, D. R., & Horne, B. G. (1993). Progress in supervised neural
networks. IEEE Signal Processing Magazine, 10(1), 8–39.
Johnson, R. A., & Wichern, D. W. (1998). Applied multivariate statistical
analysis (4th ed.). Upper Saddle River, NJ: Prentice-Hall.
Kim, Y. S., & Sohn, S. Y. (2004). Managing loan customers using
misclassification patterns of credit scoring model. Expert Systems with
Applications, 26, 567–573.
Kim, Y. S., & Street, W. N. (2004). An intelligent system for customer
targeting: A data mining approach. Decision Support Systems, 37(2),
215–228.
Kuo, R. J., Ho, L. M., & Hu, C. M. (2002). Integration of self-organizing
feature map and K-means algorithm for market segmentation.
Computers and Operations Research, 29, 1475–1493.
Lancher, R. C., Coats, P. K., Shanker, C. S., & Fant, L. F. (1995). A neural
network for classifying the financial health of a firm. European Journal
of Operational Research, 85(1), 53–65.
Lee, T. S., Chiu, C. C., Lu, C. J., & Chen, I. F. (2002). Credit scoring using
the hybrid neural discriminate technique. Expert Systems with
Applications, 23, 245–254.
Malhotra, R., & Malhotra, D. K. (2003). Evaluating consumer loans using
neural networks. Omega, 31(2), 83–96.
665
Morrison, D. F. (1990). Multivariate statistical methods. New York, NY:
McGraw-Hill.
Natter, M. (1999). Conditional market segmentation by neural networks: A
Monte-Carlo study. Journal of Retailing and Consumer Services, 6,
237–248.
Punj, G., & Steward, D. W. (1983). Cluster analysis in marketing research:
Review and suggestions for applications. Journal of Marketing
Research, 20, 134–148.
Setiono, R., Thong, J. Y. L., & Yap, C. S. (1998). Symbolic rule extraction
from neural networks—An application to identifying organizations
adopting IT. Information and Management, 34(2), 91–101.
Sharda, R., & Wilson, R. (1996). Neural network experiments in business
failures predication: A review of predictive performance issues.
International Journal of Computational Intelligence and Organiz-
ations, 1(2), 107–117.
Sung, A. H. (1998). Ranking importance of input parameters of neural
networks. Expert Systems with Applications, 15, 405–411.
Swales, G., & Yoon, Y. (1992). Applying artificial neural networks to
investment analysis. Financial Analysis Journal, 48, 78–80.
Tam, K. Y., & Kiang, M. (1992). Managerial applications of neural
networks: The case of bank failure predications. Management Science,
38(7), 926–947.
Thomas, L. C. (2000). A survey of credit and behavioural scoring:
Forecasting financial risk of lending to consumers. International
Journal of Forecasting, 16, 149–172.
Vellido, A., Lisboa, P. J. G., & Vaughan, J. (1999). Neural networks in
business: A survey of applications (1992w1998). Expert Systems with
Applications, 17, 51–70.
West, D. (2000). Neural network credit scoring models. Computers and
Operations Research, 27, 1131–1152.
Zhang, G., Hu, M. Y., Patuwo, B. E., & Indro, D. C. (1999). Artificial neural
networks in bankruptcy prediction: General framework and cross-
validation analysis. European Journal of Operational Research, 116,
16–32.