ArticlePDF Available

Abstract

The paper reviews the best-developed and most frequently applied methods of credit scoring employed by commercial banks when evaluating loan applications. The authors concentrate on retail loans - applied research in this segment is limited, though there has been a sharp increase in the volume of loans to retail clients in recent years. Logit analysis is identified as the most frequent credit-scoring method used by banks. However, other nonparametric methods are widespread in terms of pattern recognition. The methods reviewed have potential for application in post-transition countries.
152
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
SHORT PAPERS
UDC: 336.77
JEL Classification: B41, C14, D81, G21, P43
Keywords: credit scoring; discrimination analysis; banking sector; pattern recognition; retail loans
Credit Scoring Methods
Martin VOJTEK – EvÏen KOâENDA*
1. Introduction
Despite the proliferation of banking services, lending to industry and
the public still constitutes the core of the income of commercial banks and
other lending institutions in developed as well as post-transition countries.
From the technical perspective, the lending process in general is a relatively
straightforward series of actions involving two principal parties. These ac-
tivities range from the initial loan application to the successful or unsuc-
cessful repayment of the loan. Although retail lending belongs among
the most profitable investments in lenders’ asset portfolios (at least in de-
veloped countries), increases in the amounts of loans also bring increases
in the number of defaulted loans, i.e. loans that either are not repaid at all
or cases in which the borrower has problems with paying debts. Thus,
the primary problem of any lender is to differentiate between “good” and
“bad” debtors prior to granting credit. Such differentiation is possible by us-
ing a credit-scoring method. The goal of this paper is to review credit-scor-
ing methods and elaborate on their efficiency based on the examples from
the applied research. Emphasis is placed on credit scoring related to retail
loans.
We survey the methods which are suitable for credit scoring in the retail
segment. We focus on retail loans as sharp increase in the amounts of loans
for this clientele has been recorded in the last few years and another in-
crease can be expected. This dynamic is highly relevant for post-transition
countries. In the last few years, banks in the Czech and Slovak Republics
have allocated a significant part of their lending to retail clientele. In 2004
alone, Czech and Slovak banks recorded 33.8% and 36.7% increases in re-
tail loans, respectively. Hilbers et al. (2005) review trends in bank lending
to the private sector, with a particular focus on Central and Eastern Euro-
pean countries, and find that rapid growth of private sector credit conti-
nues to be a key challenge for most of these countries. In the Czech and Slo-
vak Republics the financial liabilities of households formed 11 % and 9 %
* CERGE-EI, joint-place of Charles University and Academy of Sciences, Prague
(martin.vojtek@cerge-ei.cz), (evzen.kocenda@cerge-ei.cz)
The authors thank Petr Franûk, Vilém Semerák, an anonymous referee and participants
of the Algoritmy 2005 (Podbanské) and MME 2005 (Hradec Králové) conferences for their
helpful comments. The support of the Grant Agency of the Czech Republic under Grant
No. 402/05/0931 is gratefully acknowledged. The usual disclaimer applies.
of the GDP, respectively, in 2004.
1
Since the average ratio of financial lia-
bilities to GDP in the older 15 members of the European Union is about five
times higher, and taking into account the recent trend of decreasing inte-
rest rates, it is expected that the amount of loans to retail clientele will con-
tinue to increase.
Empirical studies on credit scoring with respect to retail loans are infre-
quent in the relevant literature on developed markets and, to our best
knowledge, no such empirical study exists with respect to retail loans in
post-transition countries. We conjecture that this is due to the sensitivity
of information associated with privacy laws that results in an aversion of
banks to provide this data. Thus, most of the credit-scoring literature deals
with industry loans, i.e. loans received by firms. Industry-loans credit scor-
ing (or, in general, rating assignment) is different from that of commercial
loans in several instances. Primarily, the amounts lent are much lower in
the case of retail lending. Most importantly, there are different decision
variables used in the decision process regarding industry loans: various ra-
tios of financial indicators, for example. To assess the efficiency of various
scoring methods we primarily draw on the empirical studies that deal with
retail loans. When appropriate, we also use those that deal with industry
loans since they often provide a focused illustration of a particular scoring
method’s abilities.
2
In order to properly distinguish the different aspects of
retail credit scoring, we also provide a comprehensive discussion on the in-
dicators that are typically important in the credit-scoring models and are
employed in the literature.
The rest of the paper is organized as follows: in Section 2 we introduce
the concept of incurred costs that is essential to credit scoring; then, in
the following sub-sections, we review the most widely used credit-scoring
methods and provide representative results from the empirical literature.
Section 3 concludes with a brief assessment of the reviewed methods. The ap-
pendix contains indicators typical to retail credit-scoring models as known
from the literature and practice.
2. Techniques Used in Credit Scoring
The process of credit scoring is very important for banks as they need to
segregate “good borrowers” from “bad borrowers” in terms of their credit-
worthiness. This is a classic example of asymmetric information, where
abank has to reveal hidden data about its client. Credit scoring in its au-
tomated form is often the only way to assess creditworthiness, as banks do
not have enough resources to treat each small exposure individually.
The methods generally used for credit scoring are based on statistical pat-
tern-recognition techniques.
3
These sophisticated methods defy the per-
ception that often regards credit scoring as being a simple data-mining pro-
cess. Statistical pattern recognition is a growing research area. Renault and
153
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
1
These numbers cover only the banking sector and not various financing companies, etc. Hence
there is a large space for expansion in the financial liabilities in both countries.
2
Altman and Narayanan (1997) provide a broad review of the corporate failure models and their
classification.
de Servigny (2004) claim that “[...] this means that most current statistical
models applied to credit risk are lagging far behind the state-of-the-art
methodologies. As a consequence, we anticipate that in coming years banks
will be catching up with the integration of nonparametric techniques and
machine learning models.”
In this paper we review the best developed and most frequently applied
methods of credit scoring employed by banks when evaluating applications
for loans based on the (private) information provided by the borrower.
4
We
will discuss problems connected with implementing these approaches as
well.
An essential concept relevant to credit scoring is that of incurred costs
associated with the probabilities of repayment of loans. For simplicity let
us assume that the population of loans consists of two groups or
classes Gand Bthat denote loans that (after being granted) will turn good
and bad in the future, respectively. Good loans are repaid in full and on
time. Bad loans are subject to various degrees of default.
5
Usually the class sizes are very different, so that for the probability that
arandomly chosen candidate (customer) belongs to group G, denoted as p
G
,
we have p
G
>p
B
. Let xbe a vector of independent variables
6
(called mea-
surement vector) used in the process of deciding whether an applicant be-
longs to group Gor B. Let the probability that an applicant with measure-
ment vector xbelongs to group Gis p(G|x), and that of Bis p(B|x). Let
the probability p(x|G) indicate that a good applicant has measurement vec-
tor x. Similarly, for a bad applicant the probability is p(x|B). The task is to
estimate probabilities p(.|x) from the set of given data about applicants
which turn out to be good or bad and to find a rule for how to partition
the space Xof all measurement vectors into the two groups A
G
and A
B
based
on these probabilities so that in A
G
would be the measurement vectors of
applicants who turn out to be good and vice versa. However, it is usually
not possible to find perfect classification as it may happen that the same
vector is given by two applicants where one is good and the other is bad.
Therefore it is necessary to find a rule that will minimize the costs of
the agency providing credit connected with the misclassification of appli-
cants. Let us denote c
G
as the costs connected with misclassifying a good
applicant as bad and c
B
as the costs connected with classifying a bad ap-
plicant as good. Usually c
B
>c
G
, because costs incurred due to misclassify-
154
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
3
The first major use of credit scoring dates back to the 1960s, when credit card business grew
up and the automatized decision process was a must. Credit scoring was fully recognized in
the USA by the 1975 Equal Opportunity Act, which stated that any discrimination can be based
only on statistical assessments.
4
Discussing the issue that applicants are subject to various external influences such as exter-
nal social or financial changes, which can change their probability of default, is beyond the scope
of this paper, however. This can be incorporated into the behavioral scoring model. There is
apractice to offer loans to clients based mainly on behavioral credit scoring (average balance
on their checking accounts, etc.). This practice is common with banks trying to aggressively in-
crease their credit retail portfolio.
5
For a precise definition of default according to the new Basel II framework, see (BIS, 2004).
6
In practice the variables are often correlated. However, some methods can deal with such a data
problem as multicollinearity. We comment on this issue later on when describing various tech-
niques. Assumption of independence is maintained here for ease of exposition.
ing a bad customer are financially more damaging than costs associated
with the former kind of error. If applicants with xare assigned to
class Gthe expected costs are c
B
p(B|x) and the expected loss for the whole
sample is c
B
xA
G
p(B|x)p(x) + c
G
xA
B
p(G|x)p(x), where p(x) is a proba-
bility that the measurement vector is equal to x. This is minimized when
into group Gsuch applicants are assigned who have their group of mea-
surement vectors
A
G
= x|c
B
p(B|x) c
G
p(G|x)
which is equivalent to
c
B
A
G
=
x
p(G|x) –––––––
(1)
c
B
+ c
G
Without loss of generality we can normalize misclassification costs to
c
B
+c
G
=1. In that case the rule for classification is to assign an applicant
with xto class Gif p(G|x)>c
B
and otherwise to class B.
Of course, an important task is to specify the cost of lending errors for re-
tail loans and, in doing so, more accurately specify the optimal cutoff-score
approach to credit scoring. Namely the bank has to choose the optimal trade-
-off between profitability and risk. Loan policies that are too restrictive may
ensure minimal costs in terms of defaulted loans, but the opportunity costs
of rejected loans may exceed potential bad debt costs and thus profit is not
maximized. Conversely, policies that are too liberal may result in high losses
from bad debts.
2.1 Linear Discriminant Analysis
The aim of Linear Discriminant Analysis (hereinafter “LDA”) is to clas-
sify a heterogeneous population into homogeneous subsets and further
the decision process on these subsets. We can assume that for each ap-
plicant there is a specific number of explanatory variables available.
The idea is to look for such a linear combination of explanatory variables,
which separates most subsets from each other. In a simple case of two
subsets, the goal is to find the linear combination of explanatory vari-
ables, which leaves the maximum distance between means of the two sub-
sets.
In a general case we consider the distributions p(x|G) and p(x|B) which
are multivariate normal distributions with the common variance. Then
equation (1) reduces to
A
G
=
x
w
i
x
i
> c
(1’)
as follows from the econometrics theory. Here x
i
are explanatory variables,
w
i
are associated coefficients (weights) in the linear combination of ex-
planatory variables. If one takes s(x) = w
i
x
i
then it is possible to dis-
criminate according to this score and thus to reduce the problem to only one
dimension.
155
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
The discriminant analysis was introduced by Fisher (1936) who searched
for the best way to separate two groups using linear combination of vari-
ables.
7
Eisenbeis (1977) criticized this method by stating that the rule is
optimal only for a small class of distributions. However, Hand and Henley
(1997) claim that “if the variables follow a multivariate ellipsoidal distri-
bution (of which the normal distribution is a special case), then the linear
discriminant rule is optimal”.
8
Other critiques state that there is a selection bias due to the fact that
alearning sample for a credit-scoring system is made of applicants to whom
credit has been granted. That means that results are biased when applied
to the whole population. Eisenbeis (1977) also saw problems in the defini-
tion of bad and good groups in the case when no clear boundary is between
them and under the assumption that the covariance matrices of the two dis-
tributions are equal. In this case the use of quadratic discriminant analysis
instead of the linear case is appropriate. Problems also arise when one wants
to test for the significance of individual variables as one does not have the as-
sumption of normality and therefore cannot perform statistical inference.
9
Altman (1968), who was the first to apply discriminant analysis, con-
structed the so-called z-score, which is a linear combination of several ex-
planatory variables for the case of the corporate credit granting problem.
10
He found the model to be extremely accurate in correctly predicting
bankruptcy.
As we have mentioned, the advantages of the LDA method are that it is
simple, it can be very easily estimated and it actually works very well; it is
often used by banks for credit-scoring purposes. The disadvantage is that
LDA requires normally distributed data but the credit data are often non-
-normal (and categorized).
2.2 Logit Analysis
As stated earlier, distribution of the credit information data is usually
non-normal, and this fact may theoretically pose a problem when conduct-
ing an LDA.
11
One way to overcome the problems with non-normality of
data is to use an extension of the LDA model that allows for some para-
metric distribution. In this case a suitable extension is a generalized linear
model known as logit model. Given a vector of application characteristics x,
the probability of default pis related to vector xby the relationship:
p
log
––––
= w
0
+
w
i
logx
i
(2)
1–p
156
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
7
Fisher (1936) suggested (under assumption of common sample variance) looking for the li-
near combination of explanatory variables which leaves the maximum distance between means
of the two classes.
8
The proof of this claim can be found in (Webb, 2002).
9
Many of these issues are addressed in the review by Rosenberg and Gleit (1994).
10
The variables used are Sales/Total assets (TA), Working capital/TA, Retained Earnings/TA,
Earnings before Interest and Taxation/TA, Market Value of Equity/Book Value of Total Debt.
11
On the other hand, Reichert, Cho and Wagner (1983) argue that the non-normality of credit
information is not a limitation for the use of LDA from the empirical point of view.
One of the advantages over linear discriminant analysis is the use of
the maximum likelihood method for the estimation of parameters w
i
. An-
other advantage is that one can provide the probabilities of being in some
concrete class. The logit method can also deal with categorized data; the so-
lution is to take dummy variables for each category of data. Several stu-
dies found that that logit model outperforms discriminant analysis.
12
More recently, logit analysis became one of the main approaches of clas-
sification in credit scoring in the practices of banks. The coefficients ob-
tained have the same values as in the studies that employed the LDA
decision rule. Nevertheless, they are obtained under much weaker as-
sumptions. Actual classification results are similar for both types of re-
gression and because both are sensitive to high correlation among ex-
planatory variables, one should ensure that there are no such variables
left in the training set.
13
Another disadvantage of this method is the sen-
sitivity to missing values (all observations with missing values have to
be deleted). Similarly, as with linear discriminant analysis, the logit ap-
proach is limited by a parametric form of model.
Lawrence and Arshadi (1995) used the logit model for the analysis of
the management of problem loans and of determinants of resolution choices
using a series of borrower and bank variables. In the area of mortgage lend-
ing, Campbell and Dietrich (1983) utilized a logit model to show that the age
of a mortgage, the loan-to-value ratio, interest rates, and unemployment
rates are significant in explaining mortgage prepayments, delinquencies
and defaults. Gardner and Mills (1989), recognizing that delinquent bor-
rowers do not necessarily end up in default, employ a logit regression model
to estimate the probability of default for currently delinquent loans. They
recommend that bankers use this method to identify the severity of the prob-
lem and thereby formulate an appropriate response to the delinquency. Re-
cently, Charitou, Neophytou and Charalambous (2004) found that the logit
method is superior to other methods in predicting defaults.
2.3 k-nearest Neighbor Classifier
The k-nearest neighbor classifier serves as an example of the non-para-
metric statistical approach. This technique assesses the similarities be-
tween the pattern identified in the training set and the input pattern. One
chooses a metric on the space of applicants and takes k-nearest neighbor
(hereinafter “k-NN”) of the input pattern that is nearest in some metric
sense. A new applicant will be classified in the class to which the majority
of the neighbors belong (in the case when the costs of misclassification are
equal) or according to the rule expressed by equation (1). This means that
this method estimates the p(G|x) (or p(B|x)) probability by the proportion
of G(or B) class points among the k-nearest neighbors to the point xto be
classified.
157
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
12
See for example (Wiginton, 1980).
13
The training set is the data sample with known properties that serves to calibrate and ve-
rify the model performance before the model is applied.
The first to use this method were Chatterjee and Barcun (1970). Identi-
fying the advantages of this method, Henley and Hand (1996) stated that
the non-parametric nature of the method enables modeling of irregularities
in the risk function over the feature space. The k-NN method has been found
to perform better than other non-parametric methods such as kernel me-
thods when the data are multidimensional. It is a fairly intuitive procedure
and as such it could be easily explained to business managers who would
need to approve its implementation. It can also be used dynamically by
adding applicants when their class becomes known and deleting old appli-
cants to overcome problems with changes in population over time.
14
When performing the k-NN methodology, a very important step is
the choice of the metric used. Henley and Hand (1996) describe the choice
of the metric and the choice of the number of nearest neighbors to consider.
Acommonly used metric is the standard Euclidean norm given by
1
(x, y) = [(xy)
T
(xy)]
1/2
(3)
where xand yare measurement vectors.
However, when the variables are in different units or categorized
15
, it is
necessary to use some appropriate standardization of variables as well as
to select some data-dependent version of the Euclidean metric such as:
2
(x, y) = [(xy)
T
A(xy)]
1/2
(4)
where Ais a n nmatrix with nnumber of variables. As matrix Acan de-
pend on xwe can define two types of metrics according to how Ais selected:
local metrics are those where Adepends on x;global metrics are those where
Ais independent of x.
16
The choice of the number of nearest neighbors chosen (k) determines
the bias/variance trade-off in the estimator. The khas to be much smaller
than the smallest class. A simulation study by Enas and Choi (1986) sug-
gested that k
n
2/8
or n
3/8
is reasonable.
17
Recently, Hand and Vinciotti (2003) observed that in problems where there
are two unbalanced classes, the fact that kis finite (and thus asymptotic
158
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
14
The method is often superior over linear discriminant analysis; this is shown for example by
Ripley (1994), who compares data linear discriminant, neural networks, k-NN and various other
classification methodologies.
15
As mentioned earlier, many variables in measurement vector xare categorized and measured
on different scales. One way to overcome this problem is to introduce convenient dummy vari-
ables. Another strategy is to use the so-called weights of evidence (w
ij
) where the j-th attribute
of the i-th characteristic is given by w
ij
= (p
ij
/q
ij
), where p
ij
is the number of those classified in
Gclass in attribute jof characteristic idivided by the total number of good risks and similarly
q
ij
is a proportion of bad risks in attribute jof characteristic i.
16
Henley and Hand (1996) proposed a global metric given by the parametrization A=I+Dww
T
,
where wis the direction orthogonal to equiprobability contours for p(G|x), and Dis a distance
parameter. The direction wcan be calculated using linear regression weights.
17
On the other hand a choice of kvia cross-validation on the misclassification rate is often
adopted in empirical literature. The cross-validation consists of dividing the training sample on
msubsets and then using m– 1 subsets to train; the last set is used as the test set, and this is
repeated for each subset.
properties do not hold) results in a non-monotonic relationship between
the kand the proportion of each class correctly classified. That means, in
general, that a larger kmay not yield better performance than a smaller k.
For example if the number of points from the smaller class is less than
(1 – c
B
)
–1
, then the best classification rule for predicting class Gmember-
ship is to use k=1.
Holmes and Adams (2002) realized that there is a lack of a formal frame-
work for choosing the kand that the method can only make discrete pre-
dictions by reporting the relative frequencies which have no probabilistic
interpretation. They tried to overcome these difficulties by proposing
the Bayesian approach, which integrates over the choice of k. Such approach
leads to the conclusion that marginal predictions are given as proper pro-
babilities.
2.4 Classification and Regression Trees (CART)
Classification and Regression Trees (CART) is a nonparametric method
that is due to Breiman, Friedman, Olshen and Stone (1984). It is a flexible
and potent technique; however, it is used in banking practice chiefly only
as a supporting tool to accompany the parametric estimation methods de-
scribed earlier. It serves, for example, in the process to select regressors or
characteristics with the highest explanatory power. The CART method em-
ploys binary trees and classifies a dataset into a finite number of classes.
It was originally developed as an instrument for dealing with binary re-
sponses and as such it is suitable for use in credit scoring where the default
and non-default responses are contained in data. The CART method was
later refined in subsequent editions of Breiman, Friedman, Olshen and
Stone (1984).
Similarly as with the methodologies reviewed earlier, we make the as-
sumption of having a training set of measurement vectors X
T
={x
j
} along
with information whether an individual jdefaulted, or not (hence, y
j
is coded
1 or 0 respectively). The CART tree consists of several layers of nodes:
the first layer consists of a root node; the last layer consists of leaf nodes.
Because it is a binary tree, each node (except the leaves) is connected to
two nodes in the next layer. The root node contains the entire training set;
the other nodes contain subsets of this set. At each node, this subset is di-
vided into 2 disjoint groups based on one specific characteristic x
i
from
the measurement vector. If x
i
is ordinal, the split results from the fact, re-
lated to a particular individual, as to whether x
i
>c, for some constant c. If
the previous statement is true, an individual jis classified into the right
node; if not, an individual is classified into the left node. A similar rule ap-
plies, if x
i
is a categorized variable.
The characteristic x
i
is chosen among all possible characteristics and
the constant cis chosen so that the resulting subsamples are as homoge-
neous in yas possible. In other words: x
i
and care chosen to minimize the di-
versity of resulting subsamples (diversity in this context will be defined
presently). The classification process is a recursive procedure that starts at
the root node and at each further node (with exception of the leaves) one
single characteristic and a splitting rule (or constant c) are selected. First,
159
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
the best split is found for each characteristic. Then, among these charac-
teristics the one with the best split is chosen.
The result of this procedure is that subsamples are more homogeneous
than the parental sample. The procedure ends when the node contains only
individuals with the same y
j
or it is not possible to decrease diversity fur-
ther. For illustration purpose, let p(.|t) be the proportion of Gand Bgroups
present at node t. As an example of the diversity functions can be taken
aGini index function, which is defined as d(t)=p(G|t)p(B|t). The value
of constant cthat is instrumental to splitting between nodes is defined as
to minimize the weighted diversify in the daughter nodes (i.e. nodes to
which node tis parental). Formally, the aim is to choose cthat minimizes
p
L
d (t
L
)+p
R
d(t
R
),where p
L
and p
R
are the proportions of individuals going
into nodes t
L
and t
R
respectively. The completed tree is usually very large
but algorithms exist for pruning it into a simpler final tree.
18
The advan-
tages of the CART method in credit scoring are that it is very intuitive,
easy to explain to management, and it is able to deal with missing obser-
vations. The major disadvantage is the computational burden in case of
large datasets since at each node every characteristic has to be examined.
Very often the resulting tree is quite large so that the process of model-
-learning becomes too time-consuming. Some empirical studies also note
that often the trees are not stable since small change in a training set may
considerably alter the structure of the whole tree.
19
Asignificant problem
is also the fact that CART optimizes only locally on a single variable at
atime and thus it may not minimize the overall costs of misclassification.
The first to use the CART method in the credit scoring area were Fryd-
man, Altman and Kao (1985) who found it to outperform LDA.
20
Relevant
to retail lending is the study by Devaney (1994), who used logit and CART
methods to choose which financial ratios are the best predictors of house-
holds default and found that both methods differ substantially in selecting
the ratios: the ones chosen by the CART method were not so important ac-
cording to logit regression. A very recent addition to the empirical litera-
ture dealing with the CART method is (Feldman – Gross, 2005). The au-
thors use this method for mortgage default data and discuss the pros and
cons of CART in relation to traditional methods.
Extensions of the CART method cover the Bayesian approach to tree con-
struction that uses Bayesian techniques for node splitting, pruning and ave-
raging of multiple trees (Denison – Mallick – Smith, 1988). A comparative
study of pruning methods for CART is provided in (Esposito – Malerba –
Semeraro, 1997). An acceleration procedure in splitting the tree is discussed
in (Mola – Siciliano, 1997).
160
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
18
The choice of the pruning algorithms can be found, for example, in (Breiman – Friedman –
Olshen – Stone, 1984). The most common and efficient ones are based on the fact that if one
tries to select a subtree of the maximal tree that minimizes the estimated misclassification costs,
alarge number of trees yield approximately the same estimated misclassification costs. There-
fore it is reasonable to stop the search for the best pruned tree once a subtree with similar mis-
classification costs to the maximal tree is found.
19
See for example (Hastie – Tibshirani – Friedman, 2001).
20
Chandy and Duett (1990) compared trees with logit and LDA and found that these methods
are comparable in results to a sample of commercial papers from Moody’s and S&P’s. Zhang
(1998) generalized CART to multiple binary responses and used it for medical data.
2.5 Neural Networks
The last reviewed method is so-called neural networks. A neural network
(NNW) is a mathematical representation inspired by the human brain and
its ability to adapt on the basis of the inflow of new information. Mathe-
matically, NNW is a non-linear optimization tool. Many various types of
NNW have been specified in the literature.
21
The NNW design called multilayer perceptron (MLP) is especially suit-
able for classification and is widely used in practice. The network consists
of one input layer, one or more hidden layers and one output layer, each
consisting of several neurons. Each neuron processes its inputs and gene-
rates one output value that is transmitted to the neurons in the subsequent
layer. Each neuron in the input layer (indexed i=1,...,n) delivers the value
of one predictor (or the characteristics) from vector x. When considering de-
fault/non-default discrimination, one output neuron is satisfactory. In each
layer, the signal propagation is accomplished as follows. First, a weighted
sum of inputs is calculated at each neuron: the output value of each neu-
ron in the proceeding network layer times the respective weight of the con-
nection with that neuron. A transfer function g(x) is then applied to this
weighted sum to determine the neuron’s output value. So, each neuron in
the hidden layer (indexed j=1,..., q) produces the so-called activation:
a
j
= g
i
w
ij
x
i
(5)
The neurons in the output layer (indexed k=1,..., m) behave in a man-
ner similar to the neurons of the hidden layer to produce the output of
the network:
y
k
= f
j
w’
ik
a
j
= f
j
w’
jk
g
i
w
ij
x
i

(6)
where w
ij
and w
jk
are weights.
22
There are two stages of optimization. First, weights have to be initialized,
and second, a nonlinear optimization scheme is implemented. In the first
stage, the weights are usually initialized with some small random number.
The second stage is called learning or training of NNW. The most popular
algorithm for training multilayer perceptrons is the back-propagation al-
gorithm. As the name suggests, the error computed from the output layer
is back-propagated through the network, and the weights are modified ac-
cording to their contribution to the error function. Essentially, back-propa-
gation performs a local gradient search, and hence its implementation; al-
though not computationally demanding, it does not guarantee reaching
161
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
21
For a thorough exposition see (Bishop, 1995).
22
The Sigmoid (or logistic) function f(x) = 1/(1 + exp(x)) or hyperbolic tangent function f(x) =
= (exp (x) – exp(–x))/(exp(x) + exp(–x)) is usually employed in the above network output for func-
tions fand g. The logistic function is appropriate in the output layer if we have a binary clas-
sification problem, as in credit scoring, so that the output can be considered as default proba-
bility. According to the theory (Bishop, 1995), the NNW structure with a single hidden layer is
able to approximate any continuous bounded integrable function arbitrarily accurately.
aglobal minimum. For each individual, weights are modified in such a way
that the error computed from the output layer is minimized.
NNWs were described in the 1960s but their first use in the credit-scor-
ing-related literature appears only at the beginning of 1990s. Altman (1994)
employed both LDA and a neural network to diagnose corporate financial
distress for 1,000 Italian firms and concluded that neural networks are not
aclearly dominant mathematical technique compared to traditional statis-
tical techniques such as discriminant analysis, and that LDA compares
rather well to the neural network model in decision accuracy. On the other
hand, with different results, Tam and Kiang (1992) studied the application
of the neural network model to Texas bank-failure prediction for the period
1985–1987; they compared the NNW prediction accuracy with that of LDA,
logistic regression, k-NN, and a decision tree model. Their results suggest
that the NNW is the most accurate, followed by linear discriminant analy-
sis, logistic regression, decision trees, and k-NN.
In direct relation to retail credit scoring, Desay, Crook and Overstreet
(1996) investigate a multilayer perceptron neural network, a mixture of
an expert’s neural network, linear discriminant analysis, and logistic re-
gression for scoring credit applicants in the credit union industry. Their re-
sults indicate that customized neural networks offer a very promising ave-
nue if the measure of performance is the percentage of bad loans correctly
classified. However, if the measure of performance is the percentage of good
and bad loans correctly classified, logistic regression models are compar-
able to the neural networks approach.
West (2000) investigated the credit-scoring accuracy of five various neu-
ral network models. The neural network credit-scoring models were tested
using 10-fold cross-validation with two real-world data sets (with both Ger-
man and Australian credit data). Results were benchmarked against more
traditional methods reviewed in this article. Their research suggested that
logistic regression is a good alternative to the neural models. Logistic re-
gression was slightly more accurate than the neural network models for
the average case, which includes some inferior neural network training ite-
rations.
The major drawback of NNWs is their lack of explanation capability. While
they can achieve a high prediction accuracy rate, the reasoning behind why
and how the decision was reached is not available. For example, in a case
of a denied loan it is not possible to determine which characteristic(s) was
exactly the key one(s) to prompt rejection of the application. Consequently,
it is very difficult to explain the decision results to managers.
23
3. Concluding Remarks
In this paper we have identified the most common methods used in the pro-
cess of credit scoring of applicants for retail loans. Our review concentrates
on the most relevant methods, which correspond to their use in the prac-
162
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
23
Baesens, Setiono, Mues and Vanthienen (2003) presented the results from analyzing three
real-life datasets using NNW rule extraction techniques, i.e. how to clarify NNW decisions by
explanatory rules that capture the learned knowledge embedded in networks.
tice of banks. According to the personal experience of the authors, most lo-
cal banks in the Czech and Slovak Republics use models based on the logit
method, which is an extension of the linear discriminant analysis and is
very tractable and convenient. Other methods such as CART or neural net-
works are used mainly as support tools, either in the process of selecting
variables or in the process of the model-quality evaluation. The lesser known
k-NN method is not used at all or is used very rarely. These facts are very
surprising, as the alternative (nonparametric) methods have excellent po-
tential in pattern recognition and they are very competitive with logit re-
gression. It seems that this potential is unrecognized by the local banks and
this reality is not far from the claim by Renault and de Servigny (2004)
mentioned in Section 2.
Answering the question of which method to choose is not straightforward
and depends mainly on the bank’s preferences, data availability, its cha-
racteristics, etc. As follows from our short survey, the various methods are
often very comparable in results. This fact can be partly explained by
the mathematical relationships between these models: for example,
the NNW can be seen as a generalization of the logit method. Often, there
is no superior method for diverse data sets. However, the logit method is
the most favored method in practice, mainly due to (almost) no assumptions
imposed on variables, with the exception of missing values and multi-
collinearity among variables. Contrary to this, non-parametric methods can
deal with missing values and multicollinearity (or correlations) among vari-
ables, but often are computationally demanding. The rules that are con-
structed on the basis of some of these methods can be hard to explain to
amanager as well as to a client, however. For example, despite the fact that
neural networks have the potential to produce excellent results, the inability
to explain why can present a serious drawback to applying this method in
practice.
In any event, as the amounts of retail loans increase, the quality and
methodology improvements in the credit-scoring processes become impera-
tives for commercial banks. This is especially important during periods of
sharp increase in optimism about future earnings, which often prompts
households to borrow and spend. If real performance falls below these ex-
pectations, the severity of incurred losses may be very high for commercial
banks. The post-transition developments on the retail-loans market in
the Czech and Slovak Republics can serve as an example of the necessity
to advance credit methods in order to protect banks as well as their cus-
tomers.
163
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
APPENDIX
Indicators that Are Typically Important in Retail Credit Scoring Models
Variables or indicators that are typical to the retail segment of credit scoring are
given in this appendix. To answer the question of what the main determinants of de-
fault are, it is necessary to design a model specification containing the right vari-
ables. The following brief outline concentrates on those variables that most frequently
come from the relevant literature; the variables are presented in Table 1. They have
two common features: first is their soundness in helping to estimate the probability
of default of an applicant; second is their explanatory power when a credit-scoring
method is employed to analyze a loan application. The variables can be divided into
four main categories as Table 1 indicates. In the following text we will briefly dis-
cuss the importance of each of these categories.
164
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
Demographic Financial Employment Behavioral
Indicators Indicators Indicators Indicators
1. Age of borrower 1.Total assets 1. Type of employment 1. Checking account
of borrower (CA)
2. Sex of borrower 2. Gross income 2. Length of current 2. Average balance
of borrower employment on CA
3. Marital status 3. Gross income 3. Number of employ- 3. Loans outstanding
of borrower of household ments over the last
xyears
4. Number 4. Monthly costs 4. Loans defaulted
of dependants of household or delinquent
5. Home status 5. Number of payments
per year
6. District of address 6. Collateral/guarantee
TABLE 1 Indicators that Are Typically Important in Retail Credit Scoring Models
The first category contains demographic indicators. These variables typically do
not have the highest importance, but they are useful in capturing various regional,
gender, and other relevant differences. For example, it is often found that older women
are less risky than young men. In general, the risk of default decreases with age and
is also lower for married applicants with dependants. Home owners also represent
aless risky category due to a house as collateral (more on collateral in the fourth ca-
tegory). Relations like this can help to better discriminate between good/bad appli-
cants.
The second category contains data on a financial situation. When considering a loan
application, a bank needs to know what other available resources a household has,
what its incomes and costs are, and consequently from these items of information,
what the realistic or potential maximum possible monthly payment is. The impor-
tance of these variables is evident.
The source of income and employment status constitute the third set of variables.
Typically, in developed countries, a large proportion of people are self-employed and
this category frequently receives a lower score in the assessment of loan applications
than employed people. This is due to the fact that stability of employment may pro-
vide a sign of stability of payments. The character and length of employment are also
decisive factors: frequent change of low-skilled jobs invites a low score.
The behavioral characteristics of the fourth category are the first-rate information
that can be used for credit scoring. This type of information significantly lowers
the problem of asymmetric information between a bank and a client. If a client has
some history with a bank, then that bank can easily verify, for example, the history
of average balances in a checking account(s), the inflow and outflow of money from
that checking account(s), etc. The bank knows if the client has already had a loan,
whether this loan was successfully repaid, or whether it involved some problems.
Banks often share this type of information, since previous default/delinquency is a se-
rious determinant of future problems with repaying debts. The existence, type, and
value of collateral are also a part of the category. Collateral is often a key, and in
the case of certain loans the dominant, factor in determining a bank’s lending deci-
sion. Collateral is also a forceful factor in a client’s decision to repay the debt. From
this point of view, real estate serve as the best collateral. The threat of losing
one’s house in the event of default is a critical factor for a client.
REFERENCES
ALTMAN, E. (1968): Financial Ratios, Discriminant Analysis and the Prediction of Corporate
Bankruptcy. Journal of Finance, vol. 23, 1968, pp. 589–609.
ALTMAN, E. (1994): Corporate Distress Diagnosis: Comparisons Using Linear Discriminant
Analysis and Neural Networks (the Italian Experience). Journal of Banking and Finance, vol. 18,
1994, pp. 505–529.
ALTMAN, E. – NARAYANAN, P. (1997): An International Survey of Business Failure Classifi-
cation Models. Financial Markets, Institutions and Instruments, vol. 6, 1997, no. 2, pp. 1–57.
BAESENS, B. – SETIONO, R. – MUES, CH. – VANTHIENEN, J. (2003): Using Neural Net-
work Rule Extraction and Decision Tables for Credit-Risk Evaluation. Management Science,
vol. 49, 2003, pp. 312–329.
BIS (2004): International Convergence of Capital Measurement and Capital Standards: A Re-
vised Framework. Basel Committee on Banking Supervision, Bank for International Settle-
ments, June 2004.
BISHOP, C. M. (1995): Neural Networks for Pattern Recognition. Oxford University Press, 1995.
BREIMAN, L. – FRIEDMAN, J. H. – OLSHEN, R. A. – STONE, C. J. (1984): Classification and
Regression Trees. Pacific Grove, CA, Wadsworth, 1984.
CAMPBELL, T. S. – DIETRICH, J. K. (1983): The Determinants of Default on Insured Con-
ventional Residential Mortgage Loans. Journal of Finance, vol. 38, 1983, pp. 1569–1581.
CHANDY, P. R. – DUETT, E. H. (1990): Commercial Paper Rating Models. Quarterly Journal
of Business and Economics, vol. 29, 1990, pp. 79–101.
CHARITOU, A. – NEOPHYTOU, E. – CHARALAMBOUS, C. (2004): Predicting Corporate Fai-
lure: Empirical Evidence for the UK. European Accounting Review, vol. 13, 2004, pp. 465–497.
CHATTERJEE, S. – BARCUN, S. (1970): A Nonparametric Approach to Credit Screening. Jour-
nal of American Statistical Association, vol. 65, 1970, pp. 150–154.
DENISON, D. G. T. – MALLICK, B. K. – SMITH, A. F. M. (1988): A Bayesian CART Algorithm.
Biometrica, vol. 85, 1988, pp. 363–377.
DESAY, V. – 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, vol. 95, 1996, pp. 24–37.
DEVANEY, S. (1994): The Usefulness of Financial Ratios as Predictors of Household Insolvency:
Two Perspectives. Financial Counseling and Planing, vol. 5, 1994, pp. 15–24.
EISENBEIS, R. A. (1977): Pitfalls in the Application of Discriminant Analysis in Business, Fi-
nance and Economics. Journal of Finance, vol. 32, 1977, pp. 875–900.
ENAS, G. G. – CHOI, S. C. (1986): Choice of the Smoothing Parameter and Efficiency of k-near-
est Neighbor Classification. Computers and Mathematics with Applications, vol. 12, 1986,
pp. 235–244.
ESPOSITO, F. – MALERBA, D. – SEMERARO, G. (1997): A Comparative Analysis of Methods
for Pruning Decision Trees. IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 19, 1997, pp. 476–491.
165
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
FELDMAN, D. – GROSS, S. (2005): Mortgage Default: Classification Tree Analysis. Journal of
Real Estate Finance and Economics, vol. 30, 2005, pp. 369–396.
FISHER, R. A. (1936): The Use of Multiple Measurements in Taxonomic Problems. Annals of
Eugenic, vol. 7, 1936, pp. 179–188.
FRYDMAN, H. – ALTMAN, E. – KAO, D.-L. (1985): Introducing Recursive Partitioning for Fi-
nancial Classification: The Case of Financial Distress. The Journal of Finance, vol. 40, 1985,
pp. 269–291.
GARDNER, M. J. – MILLS, D. L. (1989): Evaluating the Likelihood of Default on Delinquency
Loans. Financial Management, vol. 18, 1989, pp. 55–63.
HAND, D. J. – HENLEY, W. E. (1997): Statistical Classification Methods in Consumer Credit
Scoring. Journal of the Royal Statistical Society, Series A (Statistics in Society), vol. 160, 1997,
pp. 523–541.
HAND, D. J. – VINCIOTTI, V. (2003): Choosing kfor Two-Class Nearest Neighbour Classifiers
with Unbalanced Classes.Pattern Recognition Letters, vol. 24, 2003, pp. 1555–1562.
HASTIE, T. – TIBSHIRANI, R. – FRIEDMAN, J. H. (2001): The Elements of Statistical Learn-
ing: Data Mining, Inference and Prediction. (Springer Series in Statistics) New York, Springer
Verlag, 2001.
HENLEY, W. E. – HAND, D. J. (1996): A k-nearest Neighbour Classifier for Assessing Consumer
Credit Risk. Statistician, vol. 45, 1996, pp. 77–95.
HILBERS, P. – OTKER-ROBE, I. – PAZARBASIOGLU, C. – JOHNSEN, G. (2005): Assessing
and Managing Rapid Credit Growth and the Role of Supervisory and Prudential Policies. In-
ternational Monetary Fund Working Paper, no. 05/151.
HOLMES, C. C. – ADAMS, N. M. (2002): A Probabilistic Nearest Neighbour Method for Statis-
tical Pattern Recognition. Journal of Royal Statistical Society. Series B, vol. 64, 2002,
pp. 295–306.
LAWRENCE, E. – ARSHADI, N. (1995): A Multinomial Logit Analysis of Problem Loan Reso-
lution Choices in Banking. Journal of Money, Credit and Banking, vol. 27, 1995, pp. 202–216.
MOLA, F. – SICILIANO, R. (1997): A Fast Splitting Procedure for Classification Trees. Statis-
tics and Computing, vol. 7, 1997, pp. 209–216.
REICHERT, A. K. – CHO, C.-C. – WAGNER, G. M. (1983): An Examination of Conceptual Is-
sues Involved in Developing Credit-scoring Models. Journal of Business and Economic Statis-
tics, vol. 1, 1983, pp. 101–114.
RENAULT, O. – SERVIGNY, A. de (2004): The Standard & Poor’s Guide to Measuring and Ma-
naging Credit Risk. 1st ed. McGraw-Hill, 2004.
RIPLEY, B. D. (1994): Neural Networks and Related Method for Classification. Journal of
the Royal Statistical Society. Series B, vol. 56, 1994, pp. 409–456.
ROSENBERG, E. – GLEIT, A. (1994): Quantitative Methods in Credit Management: A Survey.
Operations Research, vol. 42, 1994, pp. 589–613.
TAM, K. Y. – KIANG, M. Y. (1992): Managerial Applications of Neural Networks: The Case of
Bank Failure Predictions. Management Science, vol. 38, 1992, pp. 926–947.
WEBB, A. R. (2002): Statistical Pattern Recognition. 2nd edition. John Willey & Sons, 2002.
WEST, D. (2000): Neural Network Credit Scoring Models. Computers & Operations Research,
vol. 27, 2000, pp. 1131–1152.
WIGINTON, J. C. (1980): A Note on the Comparison of Logit and Discriminant Models of Con-
sumer Credit Behavior. Journal of Financial and Quantitative Analysis, vol. 15, 1980,
pp. 757–770.
ZHANG, H. (1998): Classification Trees for Multiple Binary Responses. Journal of the Ameri-
can Statistical Association, vol. 93, 1998, pp. 180–193.
166
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
SUMMARY
JEL Classification: B41, C14, D81, G21, P43
Keywords: banking sector; credit scoring; discrimination analysis; pattern recognition; retail loans
Credit-Scoring Methods
Martin VOJTEK – Evžen KOČENDA: CERGE-EI, Prague (martin.vojtek@cerge-ei.cz)
(evzen.kocenda@cerge-ei.cz)
The paper reviews the best-developed and most frequently applied methods of credit
scoring employed by commercial banks when evaluating loan applications. The au-
thors concentrate on retail loans – applied research in this segment is limited, though
there has been a sharp increase in the volume of loans to retail clients in recent years.
Logit analysis is identified as the most frequent credit-scoring method used by banks.
However, other nonparametric methods are widespread in terms of pattern recogni-
tion. The methods reviewed have potential for application in post-transition coun-
tries.
167
Finance a úvûr – Czech Journal of Economics and Finance, 56, 2006, ã. 3-4
... In recent years, many studies have shown that Artificial Intelligence techniques such as Artificial Neural Networks (West 2000), Decision Trees (Jiang et al. 2009) Case-Based Reasoning (Shin and Han 2001) and Vector Machines (Schebesch and Stecking 2005) can be used as alternative methods for credit rating and show higher prediction accuracy. According to Vojtek and Koèenda (2006) discriminant analysis was first used by Fisher (1936), who explored the best way to separate two groups using a linear combination of variables. Eisenbeis (1977), Hand and Henley (1997) where also contributed to the research of Linear Discriminant Analysis in credit rating. ...
... The most recent and increasingly used method in assessing creditworthiness according to Vojtek and Koèenda (2006) is so-called neural networks. A neural network (NNW) is a mathematical representation inspired by the human brain and its ability to adapt based on the influx of new information. ...
Chapter
The climate change is a global externality that requires international efforts. The European Union has subscribed international agreements and internally has launched the European Union 2020 Strategy, with specific environmental targets. In this context, it is important to measure the environmental damage and economic growth relationship in the European Union. The paper aims to evaluate the decoupling of the gross domestic product and greenhouse gas emissions and to analyse the effects of population growth and gross domestic product on increase of greenhouse gas emissions and energy inefficiency in the members of the European Union. To do that, we use Data Envelopment Analysis models, which were applied to analyse the inefficiency levels and the population growth and gross domestic product decoupling of the 28 member countries of the European Union using data from the years 1995, 2000, 2005, 2010 and 2014. It was found that Luxembourg, Austria, Denmark, Malta, UK and Italy are the economies that turn out to be the most decoupled, and there is a reduction of inefficiency rates from 2010 to 2014 that could be affected by the great recession of 2009.
... However, because of the growth of the credit card business at that time period, banks began to increasingly rely on automatic decision processes, giving rise to the notion and practice of credit scoring. As discussed in [1], this notion was fully recognized in the USA by the 1975 Equal Opportunity Act, which stated that any discrimination can be based only on statistical assessments. ...
... An extension of the WOE-based methodology usually applied for credit scoring purposes has been presented in this work. The main aims of this extension are as follows: (1) addressing the lack of adaptation of the standard WOE-based approach to deal with interacting variables; and, related to the previous point, (2) enhancing the predictive capability of the WOE-based methodology while simultaneously safeguarding its interpretability and, particularly, its main interpretability tool, that is, balance scorecards. ...
Article
Full-text available
The term credit scoring refers to the application of formal statistical tools to support or automate loan-issuing decision-making processes. One of the most extended methodologies for credit scoring include fitting logistic regression models by using WOE explanatory variables, which are obtained through the discretization of the original inputs by means of classification trees. However, this Weight of Evidence (WOE)-based methodology encounters some difficulties in order to model interactions between explanatory variables. In this paper, an extension of the WOE-based methodology for credit scoring is proposed that allows constructing a new kind of WOE variable devised to capture interaction effects. Particularly, these new WOE variables are obtained through the simultaneous discretization of pairs of explanatory variables in a single classification tree. Moreover, the proposed extension of the WOE-based methodology can be complemented as usual by balance scorecards, which enable explaining why individual loans are granted or not granted from the fitted logistic models. Such explainability of loan decisions is essential for credit scoring and even more so by taking into account the recent law developments, e.g., the European Union’s GDPR. An extensive computational study shows the feasibility of the proposed approach that also enables the improvement of the predicitve capability of the standard WOE-based methodology.
... In recent years, many studies have shown that Artificial Intelligence techniques such as Artificial Neural Networks (West 2000), Decision Trees (Jiang et al. 2009) Case-Based Reasoning (Shin and Han 2001) and Vector Machines (Schebesch and Stecking 2005) can be used as alternative methods for credit rating and show higher prediction accuracy. According to Vojtek and Koèenda (2006) discriminant analysis was first used by Fisher (1936), who explored the best way to separate two groups using a linear combination of variables. Eisenbeis (1977), Hand and Henley (1997) where also contributed to the research of Linear Discriminant Analysis in credit rating. ...
... The most recent and increasingly used method in assessing creditworthiness according to Vojtek and Koèenda (2006) is so-called neural networks. A neural network (NNW) is a mathematical representation inspired by the human brain and its ability to adapt based on the influx of new information. ...
Chapter
Full-text available
China’s programme for inter-regional economic connectivity, the Belt and Road Initiative (BRI), will develop new trade routes and production chains linking predominantly developing world economies with the Chinese market, making a unique contribution to global economic development but disrupting power relativities. At the same time, problems with debt servicing, corruption and a range of other political risks such as the role of Chinese state-owned enterprises have emerged in some BRI partner countries and with some projects. These have encouraged an increasingly negative commentary, particularly following the shift of the US rhetoric since 2017, characterising China as a strategic competitor. Conventional geopolitical, realist and liberal internationalist analysis appears to have reached foregone conclusions that the BRI will generate zero sum benefits to China and costs to others, yet there is a significant research gap in empirical case studies of what is actually taking place on the BRI. To date, firms engaged in BRI projects have been overwhelmingly Chinese, but it is likely that there will be future internationalisation of BRI projects and firms (and governments) will need to assess the full range of political risks of participation. Political risks are non-market risks that may affect business outcomes. While most political risk analysis has usually taken place at the national (country) level or at the firm (project) level, this paper utilises a political risk framework to identify political risks generated by the BRI at the geopolitical and trans-national level, as well as at the country and project levels. Under each of these categories, this paper discusses how risks are constructed, assesses available evidence, agency of stakeholders and identifies where further empirical investigation and theoretical development are required.
... In recent years, many studies have shown that Artificial Intelligence techniques such as Artificial Neural Networks (West 2000), Decision Trees (Jiang et al. 2009) Case-Based Reasoning (Shin and Han 2001) and Vector Machines (Schebesch and Stecking 2005) can be used as alternative methods for credit rating and show higher prediction accuracy. According to Vojtek and Koèenda (2006) discriminant analysis was first used by Fisher (1936), who explored the best way to separate two groups using a linear combination of variables. Eisenbeis (1977), Hand and Henley (1997) where also contributed to the research of Linear Discriminant Analysis in credit rating. ...
... The most recent and increasingly used method in assessing creditworthiness according to Vojtek and Koèenda (2006) is so-called neural networks. A neural network (NNW) is a mathematical representation inspired by the human brain and its ability to adapt based on the influx of new information. ...
Chapter
The purpose of this paper is to test the effect of non-parametric methodology of ANNs on enhancing decision-making procedures compared to classic multivariate Regression models. We implement the two methods on decision-making for loan allowances and on a valuation of collaterals using data from a Small Medium Greek Bank. Using sensitivity analysis, we also find the key points where decision-making results change. We provide all the estimations and a comparable matrix is produced in order to point out the similarities and differences between the two methodologies. Also, the key points in decision-making approach are presented. The practical implication is that crucial for decision-making approach in the banking industry and affects bank’s risk management. From the presentation of comparison between the two methodologies, we also provide the relationship between independent variables along with decision-making options. The debate in academic society for theoretical and non-theoretical approaches in decision-making procedure is crucial the last decades. Our paper contributes also to this academic debate, by presenting advantages and disadvantages per methodology. This study demonstrates a practical tool for decision roles and units who are trying to define the best practice for decision-making procedures.
... But because of the seriousness of credit evaluation work, strongly interpretability models are used in most works. Vojtek et al. [2] use linear discriminant analysis (LDA) and logistic regression (LR) models for credit evaluation. The author explains that these two models are widely used in bank credit evaluation work due to their simplicity and strong interpretability. ...
Article
Full-text available
In view of the real-world problem of falsity and errors credit data, and the performance degradation of the credit evaluation model caused by these problems, we proposed an outlier detection algorithm, which considered two characteristics of class-imbalance and cost-sensitive in credit data. We use an anomaly detection model called EIF to optimize the credit evaluation models. EIF uses the EasyEnsemble algorithm to construct balanced data sets, and train an Isolation Forest model for anomaly detection by the balanced datasets with different disturbances. On the one hand, the balanced dataset ensures that the class-imbalance problem is solved by undersampling, on the other hand, each sub-model learns from the overall minority class samples in order to solve the cost-sensitive problem. Experiments were performed on UCI German dataset, and the test set with fake data was constructed by correlation. Compared with other anomaly detection algorithms in common credit evaluation models, the EIF-optimized model has a higher F1 score and a lower cost-sensitive error rate. In conclusion, the EIF model is effective in enhancing the performance of the credit evaluation model for forged credit datasets.
... Physical world dataset of credit score has gathered enormous amounts of data about customers and their payment and financial transactions. Feature selection methods are utilized based on very high magnitude data to minimize the magnitude by eliminating unnecessary and excess features to boost the predictive accuracy of data mining algorithms [18,19]. ...
Article
In banking sector credit score plays a very important factor. It is important to find which customer is valid and which is not valid for loan. Now to classify customer’s credit score is used. Based on this credit score of customers the bank will decide whether to approve loan or not. In banks there are major failures due to credit risks. We can automate this by using various Machine learning algorithms to identify loan defaulters. To classify and predict the customers here various Machine learning techniques like gradient boosting, random forest and Feature Selection technique along with Decision Tree are used. Using these algorithms we accurately classify valid and invalid customers for loan. Designed model can classify their customers into good and bad applicants and train the model for getting the better accuracy of the customer data.
... 22.º do RGPD, relativo ao reduto daquilo que é ou dever ser entendida, em contornos gerais e abstratos, como decisão exclusivamente automatizada à escala dos Sistemas de Interface Homem-Máquina(SIHM), i.e., de dispositivos que permitem ao ser humano interagir com o software (Pais-Vieira, M., & Pais-Vieira, C., 2019; Matos, 2009;Johannsen,1982), como sucede, in casu, com a camada de consulta materializada no Agenteinterface (Silva & Analide, 2011b). É, consecutivamente, verosímil o não controlo antropomorfizado em momento algum sobre os resultados (outcomes) perquiridos através do modelo de auto-aprendizagem, a partir do qual se infere (ou antes, define) o perfil solvente de um pretenso mutuário-beneficiário (output) das saídas algorítmico-modelares relativas às correlações intermediadas no ato de manipulação rotulada (labeled) e autónoma das vinte categorias (ou atributos) iniciais, também eles de natureza pessoal (Morgado Rebelo & Analide, 2021), embora estas assinalem matrizes de conhecimento demográficas, financeiras, relativas à situação de emprego e comportamentais (Silva & Analide, 2011a;Vojtek & Koèenda, 2006). Logo, será exclusivamente automatizada toda a antecipação prescritiva que, em função do nível de precisão quase perfeito, exteriorizará a tendência que qualquer supervisor humano tenha para a complacência, dependência e confiança exacerbada num sistema de IA que, aprioristicamente, concretiza níveis de performance melhores do que o de qualquer operador humano (Zerilli [et al.], 2019). ...
Preprint
Full-text available
Nos dias de hoje, há cada vez mais atos normativos da União Europeia, como o RGPD em particular, que impedem os Engenheiros do Direito de inovar, de pensarem a la design thinking, impelindo-os a fazerem do mais impossível que cogitável para olvidar coimas draconianas. O presente estudo de caso concernente à implementação de um SMA para a classificação no crédito confrontara as díspares aproximações indagadas em sede de criptografia homomórfica. Destarte, concluíra-se que o Agente-modelo decisão, caso empregue técnicas parciais ou totalmente homomorfizadas, será inevitavelmente sujeito a coimas correspondentes a 20 000 000 EUR ou, no caso de uma empresa, até 4 % do seu volume de negócios anual a nível mundial. Preterirá sempre a exatidão principiológica prevista na al. c) do n.º 1 do art. 5.º, ex vi da al. a) do n.º 5 do art. 83.º, ambos do RGPD. Afigura-se, então, que o tecido empresarial estará sujeito ao dilema intransponível ad impossibilita nemo tenetur, a um beco judicativo-subsuntivo sancionatório sem saída à vista. No atual estado de arte, é completamente inexequível solucionar o trade-off dos enunciados autenticidade e confidencialidade na RNA multicamada, treinada, validada e implementada via codificação soft que preconiza o processo LWE segundo o código-fonte de um algoritmo-aprendiz de retropropação. Tal sucede, na medida em que a utilização de esquemas criptográficos homomórficos, tanto parciais como totais, sobre a configuração de Criptonets ou RNA Discretas, respetivamente, não só preterem o princípio da exatidão, como, mesmo caso executem melhores níveis de performance, não salvaguardam uma tutela efetiva da privacidade, entendida que seja no bastião da reserva sobre a intimidade da vida privada. Também aqui, quem só (de) Direito sabe, nem o próprio ramo domina. Palavras-Chave: Criptografia Homomórfica, Dilema ad impossibilita nemo tenetur, Redes Neuronais Artificiais, Sistema Multi-Agente de Classificação no Crédito. Abstract: Nowadays, more and more normative acts of the European Union, such as the GDPR, prevent legal engineers from innovating, from thinking of Law as engineering, pushing them to do the most impossible and not valuable legal artefacts in order to avoid draconian administrative fines. The case study concerning implementing an MAS DAI-based sysren for credit classification challenges the diverse approaches queried in homomorphic cryptography. Thus, this essay concluded that the Decision Agent-based Model, employing the partially or totally
... With the advents of new technologies and the development of expertise across the industry, the traditional credit scoring methods have now transformed to become automated as banks started looking for a way to utilise their limited resources to no longer have to "treat each small exposure individually" [21]. However, along this much-needed digital transformation are concerns about different obstacles that can challenge the implication of these new approaches. ...
Preprint
Full-text available
With the rapid adoption of Internet technologies, digital footprints have become ubiquitous and versatile to revolutionise the financial industry in digital transformation. This paper takes initiatives to investigate a new paradigm of the unified credit assessment with the use of federated artificial intelligence. We conceptualised digital human representation which consists of social, contextual, financial and technological dimensions to assess the commercial creditworthiness and social reputation of both banked and unbanked individuals. A federated artificial intelligence platform is proposed with a comprehensive set of system design for efficient and effective credit scoring. The study considerably contributes to the cumulative development of financial intelligence and social computing. It also provides a number of implications for academic bodies, practitioners, and developers of financial technologies.
... Martin Vojtek and Evzen Kocenda concluded that most local banks in the Czech Republic and Slovakia are using logit method-based models in paper [15]. Other methods like CART or neural networks are primarily used as support tools in the variable selection process or the process of model quality evaluation. ...
Article
Full-text available
With the improving banking sector in recent times and the increasing trend of taking loans, a large population applies for bank loans. But one of the major problem banking sectors face in this ever-changing economy is the increasing rate of loan defaults, and the banking authorities are finding it more difficult to correctly assess loan requests and tackle the risks of people defaulting on loans. The two most critical questions in the banking industry are (i) How risky is the borrower? and (ii) Given the borrower’s risk, should we lend him/her? In light of the given problems, this paper proposes two machine learning models to predict whether an individual should be given a loan by assessing certain attributes and therefore help the banking authorities by easing their process of selecting suitable people from a given list of candidates who applied for a loan. This paper does a comprehensive and comparative analysis between two algorithms (i) Random Forest, and (ii) Decision Trees. Both the algorithms have been used on the same dataset and the conclusions have been made with results showing that the Random Forest algorithm outperformed the Decision Tree algorithm with much higher accuracy.
... With the advents of new technologies and the development of expertise across the industry, the traditional credit scoring methods have now transformed to become automated as banks started looking for a way to utilise their limited resources to no longer have to "treat each small exposure individually" [21]. However, along this much-needed digital transformation are concerns about different obstacles that can challenge the implication of these new approaches. ...
Chapter
Full-text available
With the rapid adoption of Internet technologies, digital footprints have become ubiquitous and versatile to revolutionise the financial industry in digital transformation. This paper takes initiatives to investigate a new paradigm of the unified credit assessment with the use of federated artificial intelligence. We conceptualised digital human representation which consists of social, contextual, financial and technological dimensions to assess the commercial creditworthiness and social reputation of both banked and unbanked individuals. A federated artificial intelligence platform is proposed with a comprehensive set of system design for efficient and effective credit scoring. The study considerably contributes to the cumulative development of financial intelligence and social computing. It also provides a number of implications for academic bodies, practitioners, and developers of financial technologies.
Article
Full-text available
Many static and dynamic models have been used to assist decision making in the area of consumer and commercial credit. The decisions of interest include whether to extend credit, how much credit to extend, when collections on delinquent accounts should be initiated, and what action should be taken. We survey the use of discriminant analysis, decision trees, and expert systems for static decisions, and dynamic programming, linear programming, and Markov chains for dynamic decision models. Since these models do not operate in a vacuum, we discuss some important aspects of credit management in practice, e.g., legal considerations, sources of data, and statistical validation of the methodology. We provide our perspective on the state-of-the-art in theory and in practice.
Article
Feed‐forward neural networks are now widely used in classification problems, whereas non‐linear methods of discrimination developed in the statistical field are much less widely known. A general framework for classification is set up within which methods from statistics, neural networks, pattern recognition and machine learning can be compared. Neural networks emerge as one of a class of flexible non‐linear regression methods which can be used to classify via regression. Many interesting issues remain, including parameter estimation, the assessment of the classifiers and in algorithm development.
Book
The methodology used to construct tree structured rules is the focus of this monograph. Unlike many other statistical procedures, which moved from pencil and paper to calculators, this text's use of trees was unthinkable before computers. Both the practical and theoretical sides have been developed in the authors' study of tree methods. Classification and Regression Trees reflects these two sides, covering the use of trees as a data analysis method, and in a more mathematical framework, proving some of their fundamental properties.
Article
Most commonly used credit screening methods are based on Fisher's linear discriminant function for quantitative variates. The variables actually used in credit scoring are usually, however, qualitative occurring as high/low, or present/ absent. We present a nonparametric approach to the credit screening problem. In our method we do not assume multivariate normality nor do we apply any arbitrary scaling procedures to the qualitative variables. We classify an observation to that group with which it has most in common, this being done so as to minimize expected loss from misclassification. It is a variant of the “closest neighbor” rule. The misclassification probabilities of our screening rule are obtained by a jack-knife method. An empirical method for the selection of variates for screening is also given. The possibilities for adopting the method to an on-line computer system is discussed with an illustrative example.
Article
The last 30 years have seen the development of credit scoring techniques for assessing the creditworthiness of consumer loan applicants. Traditional credit scoring methodology has involved the use of techniques such as discriminant analysis, linear or logistic regression, linear programming and decision trees. In this paper we look at the application of the k-nearest-neighbour (k-NN) method, a standard technique in pattern recognition and nonparametric statistics, to the credit scoring problem. We propose an adjusted version of the Euclidean distance metric which attempts to incorporate knowledge of class separation contained in the data. Our k-NN methodology is applied to a real data set and we discuss the selection of optimal values of the parameters k and D included in the method. To assess the potential of the method we make comparisons with linear and logistic regression and decision trees and graphs. We end by discussing a practical implementation of the proposed k-NN classifier.
Article
The purpose of this study was to examine the usefulness of financial ratios as predictors of household insolvency. Financial ratios were developed for 1,934 households using data from the Survey of Consumer Finances. Two statistical methods---logistic regression and a classification tree procedure (CART)---were used for analysis. The 1983 Liquidity ratio was the most important predictor of 1986 insolvency according to the logistic regression while the 1983 Assets/Liabilities ratio was the most important variable in the classification tree. The Gross Annual Debt Payments to Disposable Income ratio was second in importance for each of the two methods. Implications for financial educators, counselors, and planners are offered.
Article
Credit-risk evaluation is a very challenging and important management science problem in the domain of financial analysis. Many classification methods have been suggested in the literature to tackle this problem. Neural networks, especially, have received a lot of attention because of their universal approximation property. However, a major drawback associated with the use of neural networks for decision making is their lack of explanation capability. While they can achieve a high predictive accuracy rate, the reasoning behind how they reach their decisions is not readily available. In this paper, we present the results from analysing three real-life credit-risk data sets using neural network rule extraction techniques. Clarifying the neural network decisions by explanatory rules that capture the learned knowledge embedded in the networks can help the credit-risk manager in explaining why a particular applicant is classified as either bad or good. Furthermore, we also discuss how these rules can be visualized as a decision table in a compact and intuitive graphical format that facilitates easy consultation. It is concluded that neural network rule extraction and decision tables are powerful management tools that allow us to build advanced and userfriendly decision-support systems for credit-risk evaluation.
Article
This paper introduces a neural-net approach to perform discriminant analysis in business research. A neural net represents a nonlinear discriminant function as a pattern of connections between its processing units. Using bank default data, the neural-net approach is compared with linear classifier, logistic regression, kNN, and ID3. Empirical results show that neural nets is a promising method of evaluating bank conditions in terms of predictive accuracy, adaptability, and robustness. Limitations of using neural nets as a general modeling tool are also discussed.