International Journal of Advanced Engineering Research and

Science (IJAERS)
Peer-Reviewed Journal
ISSN: 2349-6495(P) | 2456-1908(O)
Vol-8, Issue-9; Sep, 2021
Journal Home Page Available: https://ijaers.com/
Article DOI: https://dx.doi.org/10.22161/ijaers.89.20

Credit Risk Analysis Applying Logistic Regression, Neural

Networks and Genetic Algorithms Models
Eric Bacconi Gonçalves1, Maria Aparecida Gouvêa2

1Department of Marketing, São Paulo State University (USP), Brazil

2Department of Business Administration, São Paulo State University (USP), Brazil

Received: 14 Aug 2021, Abstract—Most large Brazilian institutions working with credit
Received in revised form: 15 Sep 2021, concession use credit models to evaluate the risk of consumer loans. Any
improvement in the techniques that may bring about greater precision of a
Accepted: 22 Sep 2021,
prediction model will provide financial returns to the institution. The first
Available online: 30 Sep 2021 phase of this study introduces concepts of credit and risk. Subsequently,
©2021 The Author(s). Published by AI with a sample set of applicants from a large Brazilian financial institution,
Publication. This is an open access article three credit scoring models are built applying these distinct techniques:
under the CC BY license Logistic Regression, Neural Networks and Genetic Algorithms. Finally,
(https://creativecommons.org/licenses/by/4.0/). the quality and performance of these models are evaluated and compared
to identify the best. Results obtained by the logistic regression and neural
Keywords—credit risk, credit scoring models,
network models are good and very similar, although the first is slightly
genetic algorithms, logistic regression, neural
better. Results obtained with the genetic algorithm model are also good,
networks.
but somewhat inferior. This study shows the procedures to be adopted by a
financial institution to identify the best credit model to evaluate the risk of
consumer loans. Use of the best fitted model will favor the definition of an
adequate business strategy thereby increasing profits.

I. INTRODUCTION in financial institutions aiming to speed up evaluation of

With the currency stability achieved by the Economical proposals.
Plano Real in 1994, financial loans became a good Models of analysis for extension of credit known as
business for the banks that no longer made such large models of credit scoring are based on historical
profits from currency devaluation(Bresser-Pereira & information from the databank on existing clients, in order
Nakano, 2002). To replace this profitability, the need to to assess whether the prospective client will have a greater
increase investment alternatives was felt at the end of the chance of being a good or bad payer. The models of credit
inflation period. Thereafter institutions have endeavored to scoring are added to the institution’s systems permitting
expand their credit portfolios. However, loans could not be on-line credit evaluation.
offered at random to all the applicant clients, therefore 1.1 Objectives of the Study
ways to evaluate the candidates were required.
Based on the data of a sample, the intention is to:
Some years ago, when applying for a loan, the client
• Develop three credit scoring models by using
filled in a proposal for evaluation by one or more
three statistical/computational techniques: Logistic
analysts(Abdou & Pointon, 2011). They then issued an
Regression, Neural Networks, Genetic Algorithms
opinion regarding the request. Although effective, the
process was slow because it did not accommodate the • Compare the models developed in terms of the
analysis of many requests. As such, the model for the quality of fitness and prediction indicators;
analysis of the concession of credit was initially introduced • Propose a model for the classification of clients

www.ijaers.com Page | 198

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

II. THEORETICAL BASIS In the universe of consumer credit, pledge of future

In this section, the theoretical concepts that will payment involves the idea of risk. As the future cannot be
support the theme of this work will be presented. fully predicted, all consumer credit involves risk, because
assurance of payment does not exist (Lewis, 1992).
2.1 Consumer Credit
Analysis of credit is charged with the task of estimating
The expression consumer credit may be understood as the risk involved in the concession or not of credit.
a form of trade where a person obtains money, goods or
The maximum risk that the institution may accept
services and vouches to pay for this in the future, adding a
relies on the policy adopted by the company. Risk
premium (interest) to the original value (Crook et al.,
presented by the applicant is of major significance for the
2007).
process of credit concession, and various queries must be
Currently, consumer credit is a large industry operating considered in its evaluation.
worldwide. Major retailers spur their sales by supplying
2.3 Evaluation of the Credit Risk
credit. Automobile companies, banks and other segments
utilize consumer credit lines as an additional alternative to Evaluation of risk is the main issue for concession of
make profit. On the other hand, consumer credit injects credit. If the risk is poorly evaluated the company will
resources into the economy, permitting production and certainly lose money, be it because of acceptance of clients
economic expansion of a country, thereby bringing who will generate losses to the business or because of the
development to the nation (Lewis, 1992). refusal of good clients who would generate profits for the
business. Companies who have a better evaluation than
However to make credit widely available does not
their competitors in the concession of credit have an
mean to distribute credit at random to all those requesting
advantage over the others as they are less vulnerable to the
it; there is a factor associated to consumer credit which is
consequence of the wrong decisions when providing
crucial in the decision of making credit available or not:
credit.
the risk.
Evaluation of risk of a potential client can be carried
2.2 Credit Risk
out in two ways:
On the financial market, credit risk is the oldest form of
1. By judgment, a more subjective way involving a
risk (Caouette et al., 2008). It is the upshot of a financial
more qualitative analysis;
transaction, contracted between the supplier of funds
(giver of credit) and the user (taker of credit). Prior to any 2. By classifying the taker by means of evaluation
sophistication resulting from financial engineering, the models, involving a more quantitative analysis.
mere act of lending a sum to someone entails the Currently, almost all large sized companies working
probability of it not being repaid, the uncertainty regarding with concession of credit use a combination of both.
return. This is, in essence, the credit risk which may be The models called credit scoring are used for the
defined as the risk of a counterpart, in an agreement of evaluation of risk of credit by classification of the
credit concession, not to meet his/her obligation. applicant. They permit measurement of the credit
According to Caouette et al. (2008 p.1), “if credit may applicant’s risk, to support the decision taking (concession
be defined as the expectation of receiving a sum of money or not of credit).
in a given period, credit risk is a chance that this 2.4 Credit Scoring Models
expectation is not fulfilled”.
The pioneer of credit models was Henry Wells,
The activity of credit concession is a basic function of executive of the Spiegel Inc. who developed a credit
banks, therefore credit risk takes on a relevant role in the scoring model during the Second World War (Lewis,
composition of an institution’s risks and may be found in 1992).
the operations where there is a transfer of money to the
Wells needed tools that would allow inexperienced
clients as well as in those where there is only a possibility
analysts to perform credit evaluation, because many of its
of usage, the pre-conceded limits. Primary types of a bank
qualified employees had been recruited for the War.
credit operation are: loans, financing, discount of payables,
advancement to depositors, advancement of exchange, During the fifties the scoring models were
leasing operations, surety bonds and warranties etc. disseminated in the American banking industry. The first
models were based upon pre-established weights for
In these operations risk may take on different forms; to
certain given characteristics, summing the points to reach a
be conceptually familiar with them helps to orient
classification score.
management and mitigation.

www.ijaers.com Page | 199

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

More extensive use of the models in the sixties (1997)further stress Discriminant Analysis, Linear
transformed business in the American market (Thomas, Regression and Decision Trees as methods that can be
2000). Not only companies in the financial area, but also used in practice. There is no method that is clearly better
the large retailers began to use credit scoring models to than the others, everything depends upon how the elected
carry out credit sales to their consumers. Retailers such as technique fits the data.
Wards, Bloomingdale’s and J.C. Penney were some of the 6. Definition of the comparison criteria of the
pioneers in this segment. models
In Brazil the background is shorter. Financial Measurement for the comparison of the models will be
institutions started to make an intensive use of credit defined here, normally by the rate of hits and the
scoring models only in the mid-nineties. Kolmogorov-Smirnov (KS) statistics.
There are some steps to be followed to construct a 7. Selection and implementation of the best model
credit scoring model; such as:
The best model is chosen using the previously defined
1. Survey of a historical background of the clients criteria. As such, the implementation of the model must be
The basic supposition to construct a model of credit programmed. The institution must adjust its systems to
evaluation is that the clients have the same behavior receive the final algorithm and program its utilization in
pattern over time; therefore models are constructed based coordination with the other areas involved.
upon past information. The availability and quality of the
data bank are fundamental for the success of the model
III. METHODOLOGICAL PROCEDURES
(Jain et al., 2020)
3.1 Description of the Study
2. Classification of clients according to their
behavior pattern and definition of the dependent variable A financial institution wishes to grant loans to its
clients and therefore it requires a tool to assess the level of
In addition to good and bad clients there are also the
risk associated to each loan to support the decision making
excluded clients, those who have peculiar characteristics
process. To set up this project, information on the history
and should not be considered (for instance, workers in the
of the clients that contracted personal credit was made
institution) and the indeterminate clients, those on the
available.
threshold of being good or bad, still without a clear
position about them. In practice, institutions consider only The product under study is personal credit. Individual
the good and bad clients to build the model because it is credit is a rapid and practical consumer credit operation.
much easier to work with binary response models. This The purpose of the loan does not need to be stated, and the
tendency to work only with good and bad clients is also loan will be extended according to the applicant’s credit
noticed in academic works (Amaral & Iquiapaza, 2020; scoring.
Gonçalves et al., 2013; Locatelli et al., 2015; Ríha, 2016). Another characteristic of the product in question is the
3. Selection of a random sample representative of lack of requirement of goods as a guarantee of
the historical background payment.The modality with pre-fixed interest rates with
the loan terms ranging from 1 to 12 months was focused
It is important that the samples of good and bad clients
for this study.
have the same size so as to avoid any possible bias due to
size difference. There is no fixed number for the sample; 3.2 The Data
however Lewis (1992)suggests a sample of 1,500 good To carry out this study a random selection was made in
clients and 1,500 bad clients to achieve robust results. a universe of clients of the bank, 10,000 credit contracts,
Habitually three samples are used, one for building of the considered as good and 10,000 considered as bad. All
model, another for the validation of the model and a third these contracts had already matured, that is to say the
to test the model. sample was collected after the due date of the last
4. Descriptive analysis and preparation of data installment of all contracts. This is an historical database
with monthly information on the utilization of the product.
This consists of analyzing, according to statistic
Based upon this structure, the progress of the contract
criteria, each variable that will be utilized in the model.
could be accompanied and particularized when the client
5. Choice and application of techniques to be used did not pay one or more installments.
in the construction of the model
In the work, the sample is divided into three sub-
Logistic Regression, Neural Networks and Genetic samples coming from the same universe of interest: one
Algorithms will be used in this work. Hand & Henley

www.ijaers.com Page | 200

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

for construction of the model, 8,000 data (4,000 good and

e  'X eZ
4,000 bad), the second for validation of the constructed p( X ) = E ( Y ) = =
model, 6,000 data (3,000 good and 3,000 bad) and the 1 + e  'X 1+ e Z
third also with 6,000 (with the same equal division) to test Initially, in this work all variables will be included for
the model obtained. the construction of the model; however in the final logistic
3.3 The Variables model, only some of the variables will be selected. The
choice of the variables will be done by means of the
The available explanatory variables have
method forward stepwise, which is the most widely used in
characteristics that can be divided into two groups:
models of logistic regression.
Reference File Variables, and Variables of Utilization and
Restriction. Reference File Variables are related to the Fensterstock (2005)points out the following advantages
client and the Utilization and Restriction Variables regard in using logistic regression for the construction of models:
the restriction of credit and notes about the client’s other • The generated model takes into account the
credit operations existing in the market. correlation between variables, identifying relationships
The Reference File Variables as well as those of that would not be visible and eliminating redundant
Utilization and Restriction are collected when the client variables;
contracts the product. • It takes into account the variables individually
3.4 Definition of the Dependent Variable and simultaneously;
This definition of the Dependent Variable, also called • The user may check the sources of error and
Performance Definition, is directly related to the optimize the model.
institution’s credit policy. For the product under study, In the same text, the author further identifies some
clients delinquent for 60 or more days were considered disadvantages of this technique:
Bad (default) and clients with a maximum delinquency of
• In many cases preparation of the variables takes a
20 days were considered Good.
long time;
Clients designated as undetermined represent a group
• In the case of many variables the analyst must
whose credit behavior is not sufficiently clear to assign
perform a pre-selection of the more important, based upon
them as good or bad customers. In practice, clients who are
separate analyses:
not clearly defined as good or bad are analyzed separately
by the credit analyst, based upon qualitative analysis. • Some of the resulting models are difficult to
implement.
3.5Logistic Regression
3.6 Artificial Neural Networks
In the models of logistic regression, the dependent
variable is, in general a binary variable (nominal or Artificial Neural Networks are computational
ordinal) and the independent variables may be categorical techniques that present a mathematical model based upon
(as long as dichotomized after transformation) or the neural structure of intelligent organisms and who
continuous(Almeida et al., 2020). acquire knowledge through experience.

The model of Logistic Regression is a particular case It was only in the eighties that, because of the greater
of the Generalized Linear Models(Lopes et al., 2017). The computational power, neural networks were widely studied
function which characterizes the model is given by(Ye & and applied. Rojas (1996)underlines the development of
Bellotti, 2019): the backpropagation algorithm as the turning point for the
popularity of neural networks.
 p( X ) 
ln   = ' X = Z An artificial neural network model processes certain
 1 − p( X )  characteristics and produces replies like those of the
human brain. Artificial neural networks are developed
' = ( 0 , 1 ,  2 ,...,  n ) : vector of the parameters using mathematical models in which the following
associated to the variables suppositions are made (Rojas, 1996):
p(X)=E(Y=1|X): probability of the individual has been 1. Processing of information takes place within the
classified as good, given the vector X. so-called neurons;
This probability is expressed by (Gonçalves et al., 2. Stimuli are transmitted by the neurons through
2013): connections;

www.ijaers.com Page | 201

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

3. Each connection is associated to a weight which, 1. Feedforward networks with a single layer are the
in a standard neural network, multiplies itself upon simpler network, in which there is only one input layer and
receiving a stimulus; one output layer. Some networks utilizing this architecture
4. Each neuron contributes for the activation are: the Hebb Network, perceptron, ADALINE, among
function (in general not linear) to determine the output others.
stimulus (response of the network). 2. Multilayered feedforward networks are those
The pioneer model by McCulloch and Pitts having one or more intermediate layers. The multilayer
(McCulloch & Pitts, 1943)for one processing unit (neuron) perceptron networks (MLP), MADALINE and of a radial
can be summarized in: base function are some of the networks utilizing this
architecture.
• Signals are presented upon input;
3. Recurrent networks: in this type of network, the
• Each signal is multiplied by a weight that
output layer has at least one connection that feeds back the
indicates its influence on the output of the unit;
network. The networks called BAM (Biderectional
• The weighted sum of the signals which produces Associative Memory) and ART1 and ART2 (Adaptative
a level of activity is made; Resonance Theory) are recurring networks.
• If this level exceeds a limit, the unit produces an The most important quality of neural networks is the
output. capacity to “learn” according to the environment and
thereby improve their performance (Deiu-merci & Mayou,
There are input signals X 1 , X 2 ,..., X p and
2018).
corresponding weights W1, W2 ,..., Wp and the limit There are essentially three types of learning:
being k. 1. Supervised Learning: in this type of learning the
In this model the level of activity is given by: expected reply is indicated to the network. This is the case
p
of this work, where a priori it is already known whether
a =  Wi X i the client is good or bad.
i =1 2. Non-supervised Learning: in this type of learning
the network must only rely on the received stimuli; the
And the output is given by:
network must learn to cluster the stimuli;
y = 1, if a  k
3. Reinforcement Learning: in this type of learning,
y = 0, if a < k behavior of the network is assessed by an external
Three characteristics must be taken into account in the reviewer.
definition of a model of neural networks: the form of the Berry & Linoff (2004) point out the following positive
network called architecture, the method for determination points in the utilization of neural networks:
of the weights, called learning algorithm; and the
• They are versatile: neural networks may be used
activation function.
for the solution of different types of problems such as:
Architecture relates to the format of the network. Every prediction, clustering or identification of patterns;
network is divided in layers, usually classified into three
• They are able to identify non-linear relationships
groups(Akkoç, 2012):
between variables;
• Input Layer where the patterns are presented to
• They are widely utilized, can be found in various
the network;
software.
• Intermediate or Hidden layers in which the major
As for the disadvantages the authors state:
part of processing takes place, by means of the weighted
connections, they may be viewed as extractors of • Results cannot be explained: no explicit rules are
characteristics; produced, analysis is performed inside the network and
only the result is supplied by the “black box”;
• Output Layer, in which the end result is
concluded and presented. • The network can converge towards a lesser
solution: there are no warranties that the network will find
There are basically three main types of architecture:
the best possible solution; it may converge to a local
feedforward networks with a single layer; feedforward
maximum.
networks with multiple layers and recurring networks.

www.ijaers.com Page | 202

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

3.7 Genetic Algorithms Some of the disadvantages pointed out in literature are:
The idea of genetic algorithms resembles the evolution • They continue to be seldom used for problems of
of the species proposed by Darwin: the algorithms will assessment of risk credit (Fensterstock, 2005)
evolve with the passing of generations and the candidates • Require a major computational effort (Berry &
for the solution of the problem one wants to solve “stay Linoff, 2004)
alive” and reproduce(Silva et al., 2019).
• Are available in only a few softwares(Berry &
The algorithm is comprised of a population which is Linoff, 2004)
represented by chromosomes that are merely the various
Criteria for Performance Evaluation
possible solutions for the proposed problem. Solutions that
are selected to shape new solutions (starting from a cross- To evaluate performance of the model two samples
over) are selected according to the fitness of the parent were selected, one for validation and the other for test.
chromosomes. Thus, the more fit the chromosome is, the Both were of the same size (3,000 clients considered good
higher the possibility of reproducing itself. This process is and 3,000 considered bad, for each one). In addition to the
repeated until the rule of halt is satisfied, that is to say to samples, other criteria are used, which are presented in this
find a solution very near to that hoped for. section.
Every genetic algorithm goes through the following 3.8 Score of Hits
stages: The score of hits is measured by dividing the total of
Start: initially a population is generated formed by a clients correctly classified, by the number of clients
random set of individuals (chromosomes) that may be included in the model.
viewed as possible solutions for the problem. Similarly, the score of hits of the good and bad clients
Fitness: a function of fitness is defined to evaluate the can be quantified.
“quality” of each one of the chromosomes. In some situations it is much more important to identify
Selection: according to the results of the fitness a good client than a bad client (or vice versa); in such
function, a percentage of the best fit is maintained while cases, often a more fitting weight is given to the score of
the others are rejected (Darwinism). hits and a weighted mean of the score of hits is calculated.
Cross-over: two parents are chosen and based upon In this work, as there is not a priori information on
them an offspring is generated, based on a specific cross- what would be more attractive for the financial institution
over criterion. The same criterion is used with another (identification of the good or bad clients), the product
chromosome and the material of both chromosomes is between the score of hits of good and bad clients (Ih) will
exchanged. If there is no cross-over, the offspring is an be used as an indicator of hits to evaluate the quality of the
exact copy of the parents. model. This indicator will privilege the models with high
scores of hits for both types of clients. The greater the
Mutation is an alteration in one of the genes of the
indicator is the better will be the model.
chromosome. The purpose of mutation is to avoid that the
population converges to a local maximum. Thus, should 3.9 The Kolmogorov-Smirnov Test
this convergence take place, mutation ensures that the The Kolmogorov-Smirnov (KS) is the other criterion
population will jump over the minimum local point, often used in practice and used in this work(Fonseca et al.,
endeavoring to reach other maximum points. 2019; Lin, 2013; Machado, 2015).
Verification of the halt criterion: once a new generation The KS test is a non-parametric technique to determine
is created, the criterion of halt is verified and should this whether two samples were collected from the same
criterion not have been met, one returns to the stage of the population (or from populations with similar
fitness function. distributions)(Jaklič et al., 2018). This test is based on the
The following positive points in the utilization of accumulated distribution of the scores of clients
genetic algorithms must be highlighted: considered good and bad.
• Contrariwise to neural networks they produce To check whether the samples have the same
explicable results (Berry & Linoff, 2004) distribution, there are tables to be consulted according to
the significance level and size of the sample (Siegel &
• Their use is easy (Berry & Linoff, 2004)
Castellan Jr, 2006). In this work, as the samples are large,
• They may work with a large set of data and tendency is that all models reject the hypothesis of equal
variables (Fensterstock, 2005) distributions. The best model will be that with the highest

www.ijaers.com Page | 203

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

value in the test, because this result indicates a larger • The variables commercial telephone and home
spread between the good and bad. telephone were recoded in the binary form as ownership or
not;

IV. RESULTS • The variables commercial ZIP Code and home

ZIP Code were initially clustered according to the first
This section will cover the methods to treat variables,
three digits, next the relative risk of each layer was
the application of the three techniques under study and the
calculated and later a reclustering was made according to
results obtained by each one of them, comparing their
the similar relative risk, the same procedure adopted
performance. For descriptive analysis, categorization of
byHand & Henley (1997);
data and application of logistic regression the SPSS for
Windows v.21.0 software was used, the software SAS • The variable salary of the spouse was discarded
Enterprise Miner. 14.1 was used for the selection of the from the analysis because much data was missing;
samples and application to the neural network; for the • Two new variables were created, percentage of
genetic algorithm a program developed in Visual Basic by the amount loaned on the salary and percentage of the
the authors was utilized. amount of the installment on the salary. Both are
4.1 Treatment of the Variables quantitative variables, which where categorized in the
same way as the remainder.
Initially, the quantitative variables were categorized.
4.2 Logistic Regression
The deciles (values below which 10%, 20% etc. of the
cases fall) of these variables were initially identified for For the estimation of the model of logistic regression, a
categorization of the continuous variables. Starting from sample of 8,000 cases equally divided in the categories of
the deciles, the next step was to analyze them according to good or bad was utilized.
the dependent variable. The distribution of good and bad Initially, it is interesting to evaluate the logistic
clients was calculated by deciles and then the ratio relationship between each independent variable and the
between good and bad was calculated, the so called dependent variable TYPE.
relative risk (RR). Since one of the objectives of this analysis was to
Groups presenting a similar relative risk (RR) were re- identify which variables are more efficient for the
grouped to reduce the number of categories by variable. characterization of the two types of bank clients, a
The relative risks were also calculated for the stepwise procedure was utilized. The elected method of
qualitative variables to reduce the number of categories, selection was forward stepwise.
whenever possible. According to (Gouvêa et al., With categorical variables, evaluation of the effect of
2012)there are two reasons to make a new categorization one particular category must be done in comparison with a
of the qualitative variables. The first is to avoid categories reference category. The coefficient for the reference
with a very small number of observations, which may lead category is 0.
to less robust estimates of the parameters associated to Variables with a logistic coefficient estimated negative
them. The second is the elimination of the model indicate that the focused category, with regard to the
parameters, if two categories present a close risk, it is reference, is associated to a decrease of the odds and
reasonable to group them in one single class. therefore a decrease in the probability of having a good
Besides clustering of categories, RR helps to client.
understand whether this category is more connected to There are two statistical tests to evaluate the
good or to bad clients. This method of clustering significance of the final model: the chi-square test of the
categories is explained by Hand & Henley (1997) change in the value of – 2LL (-2 times the log of the
When working with the variables made available, heed likelihood) and the Hosmer and Lemeshow test.
was given to the following: Table 1 presents the initial value of – 2LL, considering
• The variables gender, first acquisition and type of only the model’s constant, its end value, the improvement
credit were not re-coded as they are already binary and the descriptive level to measure its significance.
variables;
• The variable profession was clustered according
to the similarity of the nature of jobs;

www.ijaers.com Page | 204

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

Table1: Chi-Square test p

-2LL Chi-Square Degrees of Significance g =  Wi X i
(improvement) freedom i =1

11090.355 is the weighted sum of the neuron inputs.

9264.686 1825.669 28 0.000 Training of the networks consists in finding the set of
Wi weights that minimizes one function of error. In this
work for the training will be used the Back propagation
The model of 28 variables disclosed that the reduction algorithm. In this algorithm the network operates in a two
of the -2LL measure was statistically significant. step sequence. First a pattern is presented to the input layer
The Hosmer and Lemeshow test considers the of the network. The resulting activity flows through the
statistical hypothesis that the predicted classifications in network, layer by layer until the reply is produced by the
groups are equal to those observed. Therefore, this is a test output layer. In the second step the output achieved is
of the fitness of the model to the data. compared to the desired output for this particular pattern.
If not correct, the error is estimated. The error is
The chi-square statistic presented the outcome 3.4307,
propagated starting from the output layer to the input layer,
with eight degrees of freedom and descriptive level equal
and the weights of the connections of the units of the inner
to 0.9045. This outcome leads to the non rejection of the
layers are being modified, while the error is
null hypothesis of the test, endorsing the model’s
backpropagated. This procedure is repeated in the
adherence to the data.
successive iterations until the halt criterion is reached.
4.3 Neural Network
In this model the halt criterion adopted was the mean
In this work, a supervised learning network will be error of the set of validation data. This error is calculated
used, as it is known a priori whether the clients in question by means of the module of the difference between the
are good or bad. According to Potts (1998: 44), the most value the network has located and the expected one. Its
used structure of neural network for this type of problem is mean for the 8,000 cases (training sample) or the 6,000
the multilayer perceptron (MLP) which is a network with a cases (validation sample) is estimated. Processing detected
feedforward architecture with multiple layers. Consulted that the stability of the model took place after the 94th
literature (Akkoç, 2012; Deiu-merci & Mayou, 2018; iteration. In the validation sample the error was somewhat
Olson et al., 2012; Ríha, 2016) supports this statement. larger (0.62 x 0.58), which is common considering that the
The network MLP will also be adopted in this work. model is fitted based upon the first sample.
The MLP networks can be trained using the following Initially, the bad classification is of 50%, because the
algorithms: Conjugate Descending Gradient, Levenberg- allocation of an individual as a good or bad client is
Marquardt, Back propagation, Quick propagation or Delta- random; with the increase of the iterations, the better result
bar-Delta. The most common (Rojas, 1996)is the Back of 30.6% of error is reached for the training sample and of
propagation algorithm which will be detailed later on. 32.3% for the validation sample.
The implemented model has an input layer of neurons, Some of the statistics of the adopted network are in
a single neuron output layer, which corresponds to the table 2.
outcome whether a client is good or bad in the
Table2:Neural network statistics
classification of the network. It also has an intermediate
layer with three neurons, since it was the network which Obtained statistics Test Validation
presented the best outcomes, in the query of the higher Misclassification of cases 0.306 0.323
percentage of hits as well as in the query of reduction of
Mean error 0.576 0.619
the mean error. Networks which had one, two or four
neurons were also tested in this work. Mean square error 0.197 0.211

Each neuron of the hidden layer is a processing Degrees of freedom of the 220
element that receives n inputs weighted by weights Wi. model
The weighted sum of inputs is transformed by means of a Degrees of freedom of the 7780
nonlinear activation function f(.). error
The activationfunctionused in Total degrees of freedom 8000
1
thisstudywillbethelogisticfunction , where
1 + e ( −g )

www.ijaers.com Page | 205

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

Besides the misclassification and the mean error, the Fitness Function: each client was associated to the
square error and the degrees of freedom are also presented. estimate of a score and classified as good or bad. By
The average square error is calculated by the average of comparing with the information already known a priori on
the squares of the differences between that observed and the nature of the client, the precision of each chromosome
that obtained from the network. can be calculated. The indicator of hits (Ih), will be the
The number of degrees of freedom of the model is fitness function, that is to say, the greater the indicator the
related to the number of estimated weights, to the better will be the chromosome.
connection of each of the attributes to the neurons of the Selection: In this work an elitism of 10% was used for
intermediate layer and to the binding of the intermediate each new generation, the twenty best chromosomes are
layer with the output. maintained while the other hundred and eighty are formed
4.4 Genetic Algorithms by cross over and mutation.

The genetic algorithm was used to find a discriminate Cross-over: to chose the parents for cross-over the
equation permitting to score clients, and later, separate the method known as roulette wheel was used for selection
good from the bad according to the score achieved. The among these twenty chromosomes that were
equation scores the clients and those with a higher score maintained(Oreski et al., 2012). In this method, each
are considered good, while the bad are those with a lower individual is given one probability of being drawn
score. This route was adopted by Metawa et al., (2017) and according to its value of the fitness function.
Picinini et al. (2003). For the process of exchange of genetic material a
The implemented algorithm was similar to that method known as uniform cross-over was used(Galvan,
presented in Picinini et al. (2003). Each one of the 71 2016). In this type of cross-over each gene of the offspring
categories of variables was given an initial random weight. chromosome is randomly chosen among the genes of one
To these seventy one coefficients, one more was of the parents, while the second offspring receives the
introduced, an additive constant incorporated to the linear complementary genes of the second father.
equation. The value of the client score is given by: Mutation: in the mutation process, each gene of the
chromosome is independently evaluated. Each gene of
( )
72
S j =  w i p ij , where each chromosome has a 0.5% probability of undergoing
i =1 mutation. Whenever a gene is chosen for mutation, the
genetic alteration is performed, adding a small scalar value
S j = Score obtained by client j k in this gene. In the described experiment a value ranging
from -0.05 and + 0.05 was randomly drawn.
w i = Weight relating to the category i
Verification of the halt criterion: a maximum number
p ij = binary indicator equal to 1, if the client j has of generations equal to 600 was defined as the halt
criterion. After six hundred iterations, the fit chromosome
the category i and 0, conversely.
will be the solution.
The following rule was used to define if the client is
Results of the algorithm that had the highest Indicator
good or bad:
of hits are presented here.
If S j  0 , the client is considered good After execution of the algorithm, variables with a very
small weight were discarded. In the work by Picinini et al.
If S j  0 , the client is considered bad
(2003) the authors consider that the variables with a
As such, the problem the algorithm has to solve is to weight lower than 0.15 or higher than -0.15 would be
find the vector W= [ w1, w 2 ,..., w 72 ] resulting in a discarded because they did not have a significant weight
for the model. In this work, after performing a sensitivity
classification criterion with a good rate of hits in analysis, it was decided that the variables with a weight
predicting the performance of payment of credit. higher than 0.10 or lower than – 0.10 would be considered
Following the stages of a genetic algorithm, one has: significant for the model. This rule was not applied for the
Start: a population of 200 individuals was generated constant, which was proven important for the model even
with each chromosome holding 72 genes. The initial with a value below cutoff.
weight w i of each gene was randomly generated in the 4.5 Evaluation of the Models’ Performance
interval [-1, 1] (Picinini et al., 2003) After obtaining the models the three samples were
scored and the Ih and KS were calculated for each of the

www.ijaers.com Page | 206

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

models. Table 3 shows the results of classification reached

by the three models.
Table 3: Classification results
Training Validation Test
Logistic Bad Good % Correct Bad Good % Correct Bad Good % Correct
Regression
Bad 2833 1167 70.8 2111 889 70.4 2159 841 72.0
Good 1294 2706 67.7 1078 1922 64.1 1059 1941 64.7
Total 4127 3873 69.2 3189 2811 67.2 3218 2782 68.3
Neural Networks
Bad 2979 1021 74.5 2236 764 74.5 2255 745 75.2
Good 1430 2570 64.3 1177 1823 60.8 1193 1807 60.2
Total 4409 3591 69.4 3413 2587 67.7 3448 2552 67.7
Genetic
Algorithms
Bad 2692 1308 67.3 1946 1054 64.9 2063 937 68.8
Good 1284 2716 67.9 1043 1957 65.2 1073 1927 64.2
Total 3976 4024 67.6 2989 3011 65.1 3136 2864 66.5

All presented good classification results, because, KS values in all models can be considered good.
according toPicinini et al. (2003): “credit scoring models Again, Picinini et al. (2003) explain: “The Kolmogorov-
with hit rates above 65% are considered good by Smirov test (KS) is used in the financial market as one of
specialists”. the efficiency indicators of the credit scoring models. A
The hit percentages were very similar in the models of model which presents a KS value equal or higher than 30
logistic regression and neural network and were somewhat is considered good by the market”. Here again, the logistic
lesser for the model of genetic algorithms. Another regression and neural network models exhibit very close
interesting result is that, except for genetic algorithms, the results, superior to those achieved by the genetic
models presented the greatest rate of hits for bad clients, algorithm.
with a higher than 70% rate for bad clients in the three In choosing the model that best fits these data and
samples of the logistic and neural network models. analyzing according to the Ih and KS indicators, the model
Table 4 presents results of the criteria Ih and KS which built by logistic regression was elected. Although results
were chosen to compare the models. were very similar to those achieved by neural networks
this model presented the best results in the test sample,
Table 4: Comparison indexes
suggesting that it is best fit for application in other
Ih Training Validation Test databases. Nevertheless, it must be highlighted that the
Logistic regression 47.9 45.1 46.6 adoption of any one of the models would bring about good
results for the financial institution.
Neural network 47.9 45.3 45.3
Genetic algorithm 45.7 42.3 44.2
V. CONCLUSION
KS Training Validation Test
The objective of this study was to develop credit
Logistic regression 38 35 37 scoring predictive models based upon data of a large
Neural network 39 35 35 financial institution by using Logistic Regression,
Artificial Neural Networks and Genetic Algorithms.
Genetic algorithm 34 30 32
When developing the credit scoring models some care
must be taken to guarantee the quality of the model and its

www.ijaers.com Page | 207

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

later applicability. Precautions in the sampling, clear rry+linoff&ots=KYIpqpPXxG&sig=GfmnZURJGHyQzm

definition of criteria for the classification of good and bad CH0YUPuiAoXO0&redir_esc=y#v=onepage&q=berry
clients and treatment of variables in the database prior to linoff&f=false
[6] Bresser-Pereira, L. C., & Nakano, Y. (2002). Uma
application of the techniques were the measures taken in
Estratégia de Desenvolvimento com Estabilidade. Brazilian
this study, aiming to optimize results and minimize errors.
Journal of Political Economy, 22(3), 533–563.
The three models presented suitable results for the https://doi.org/10.1590/0101-31572002-1246
database in question, which was supplied by a large retail [7] Caouette, J. B., Narayanan, P., Nimmo, R., & Altman, E. I.
bank operating in Brazil. The logistic regression model (2008). Managing Credit Risk: The Great Challenge for
presented slightly better results to the model built by Global Financial Markets (2nd ed.). John Wiley & Sons.
[8] Crook, J. N., Edelman, D. B., & Thomas, L. C. (2007).
neural networks and both were better than the model based
Recent developments in consumer credit risk assessment.
on genetic algorithms.
European Journal of Operational Research, 183(3), 1447–
This study did not aim at a more detailed approach of 1465. https://doi.org/10.1016/j.ejor.2006.09.100
the techniques focused. Neural networks and genetic [9] Deiu-merci, K. K., & Mayou, M. (2018). Network Data
algorithms presented an extensive range of structures and Security for the Detection System in the Internet of Things
variations that may (and must) be better explored. Genetic with Deep Learning Approach. International Journal of
Advanced Engineering Research and Science, 5(6), 208–
algorithms, as they are a rather flexible method, not yet
213. https://doi.org/10.22161/ijaers.5.6.34
widely researched in problems of credit concession, may
[10] Fensterstock, A. (2005). Credit scoring and the next step.
be used in diverse forms to optimize results. Business Credit, 46–50.
In this type of problem, new techniques such as [11] Fonseca, S. E., Santos, A. de O., Pereira, M. V. L., &
survival analysis should not be overlooked and merit Camargos, M. A. de. (2019). Análise do Impacto de
attention in future studies. Variáveis Macroeconômicas no Desempenho Financeiro e
Endividamento de Empresas Listadas na B3. Revista
Universo Contábilbil, 14(4), 93–114.
REFERENCES https://doi.org/10.4270/ruc.2018429
[12] Galvan, P. (2016). Educational Evaluation and Prediction
[1] Abdou, H. A., & Pointon, J. (2011). CREDIT SCORING,
of School Performance through Data Miningand Genetic
STATISTICAL TECHNIQUES AND EVALUATION
Algorithms. International Journal of Advanced
CRITERIA: A REVIEW OF THE LITERATURE.
Engineering Research and Science, 3(10), 215–220.
Intelligent Systems in Accounting, Finance and
https://doi.org/10.22161/ijaers/3.10.34
Management, 18(2–3), 59–88.
[13] Gonçalves, E. B., Gouvêa, M. A., & Mantovani, D. M. N.
https://doi.org/10.1002/isaf.325
(2013). Análise de risco de crédito com o uso de regressão
[2] Akkoç, S. (2012). An empirical comparison of
logística. Revista Contemporânea de Contabilidade,
conventional techniques, neural networks and the three
10(20), 139–160. https://doi.org/10.5007/2175-
stage hybrid Adaptive Neuro Fuzzy Inference System
8069.2013v10n20p139
(ANFIS) model for credit scoring analysis: The case of
[14] Gouvêa, M. A., Gonçalves, E. B., & Mantovani, D. M. N.
Turkish credit card data. European Journal of Operational
(2012). Aplicação De Regressão Logística E Algoritmos
Research, 222(1), 168–178.
Genéticos Na Análise De Risco De Crédito. Revista
https://doi.org/10.1016/j.ejor.2012.04.009
Universo Contábil, 84–102.
[3] Almeida, F. P., Gouveia, R. G. L. de, Lima, M. K. G. de,
https://doi.org/10.4270/ruc.2012214
Ribeiro, F. A. B. S., Mendonça, J. P., & Oliveira, J. do N.
[15] Hand, D. J., & Henley, W. E. (1997). Statistical
(2020). Co-occurrence of Economic, Political and
Classification Methods in Consumer Credit Scoring: a
Environmental Factors in the Perception of Social Groups
Review. Journal of the Royal Statistical Society: Series A
in the Municipality of Uberlândia (Minas Gerais, Brazil)
(Statistics in Society), 160(3), 523–541.
About Notified Cases of Dengue. International Journal of
https://doi.org/10.1111/J.1467-985X.1997.00078.X
Advanced Engineering Research and Science, 7(4), 145–
[16] Jain, A., Patel, H., Nagalapatti, L., Gupta, N., Mehta, S.,
156. https://doi.org/10.22161/ijaers.74.17
Guttula, S., Mujumdar, S., Afzal, S., Sharma Mittal, R., &
[4] Amaral, G. H. de O., & Iquiapaza, R. A. (2020).
Munigala, V. (2020). Overview and Importance of Data
Determinantes de Inadimplência e de Recuperação de
Quality for Machine Learning Tasks. Proceedings of the
Crédito em um Banco de Desenvolvimento. BASE –
ACM SIGKDD International Conference on Knowledge
Revista de Administração e Contabilidade Da Unisinos,
Discovery and Data Mining, 3561–3562.
17(3), 483–519. https://doi.org/10.4013/base.173.05
https://doi.org/10.1145/3394486.3406477
[5] Berry, M. J. a., & Linoff, G. S. (2004). Data mining
[17] Jaklič, J., Grublješič, T., & Popovič, A. (2018). The role of
techniques: for marketing, sales, and customer relationship
compatibility in predicting business intelligence and
management. In Portal.Acm.Org.
analytics use intentions. International Journal of
https://books.google.com.br/books?hl=pt-
Information Management, 43(August), 305–318.
BR&lr=&id=AyQfVTDJypUC&oi=fnd&pg=PR37&dq=be

www.ijaers.com Page | 208

Eric Bacconi Gonçalves et al. International Journal of Advanced Engineering Research and Science, 8(9)-2021

https://doi.org/10.1016/j.ijinfomgt.2018.08.017 Engineering Research and Science, 6(7), 218–222.

[18] Lewis, E. M. (1992). An Introduction to Credit Scoring. https://doi.org/10.22161/ijaers.6726
[19] Lin, R. (2013). The application and assessment of [32] Thomas, L. C. (2000). A survey of credit and behavioural
consumer credit scoring models in measuring consumer scoring: forecasting financial risk of lending to consumers.
loan issuing risk of commercial banks in China. May. International Journal of Forecasting, 16(2), 149–172.
[20] Locatelli, R. L., Afonso, T., Ramalho, W., & Silvério, R. https://doi.org/10.1016/S0169-2070(00)00034-0
A. de O. (2015). Determinantes da inadimplência no [33] Ye, H., & Bellotti, A. (2019). Modelling recovery rates for
crédito habitacional direcionado a classe média emergente non-performing loans. Risks, 7(1), 1–17.
brasileira. Revista de Finanças Aplicadas, 1(1), 1–30. https://doi.org/10.3390/risks7010019
http://repositorio.uninove.br/xmlui/handle/123456789/1112
[21] Lopes, M. G., Ciribeli, J. P., Massardi, W. D. O., &
Mendes, W. D. A. (2017). Análise dos indicadores de
inadimplência nas linhas de crédito para pessoa física: um
estudo utilizando modelo de regressão logística. Estudos
Do CEPE, 46, 75.
https://doi.org/10.17058/cepe.v0i46.11099
[22] Machado, A. R. (2015). Collection Scoring via Regressão
Logística e Modelo de Riscos Proporcionais de Cox.
Universidade de Brasília.
[23] McCulloch, W. S., & Pitts, W. (1943). A logical calculus of
the ideas immanent in nervous activity. The Bulletin of
Mathematical Biophysics 1943 5:4, 5(4), 115–133.
https://doi.org/10.1007/BF02478259
[24] Metawa, N., Hassan, M. K., & Elhoseny, M. (2017).
Genetic algorithm based model for optimizing bank lending
decisions. Expert Systems with Applications, 80, 75–82.
https://doi.org/10.1016/J.ESWA.2017.03.021
[25] Olson, D. L., Delen, D., & Meng, Y. (2012). Comparative
analysis of data mining methods for bankruptcy prediction.
Decision Support Systems, 52(2), 464–473.
https://doi.org/10.1016/j.dss.2011.10.007
[26] Oreski, S., Oreski, D., & Oreski, G. (2012). Hybrid system
with genetic algorithm and artificial neural networks and its
application to retail credit risk assessment. Expert Systems
with Applications, 39(16), 12605–12617.
https://doi.org/10.1016/j.eswa.2012.05.023
[27] Picinini, R., Oliveira, G. M. B., & Monteiro, L. H. A.
(2003). Detecção de problemas de crédito em empresas de
pequeno porte usando redes neurais e algoritmo genético.
SIMPÓSIO BRASILEIRO DE AUTOMAÇÃO
INTELIGENTE.
[28] Ríha, J. (2016). Artificial Intelligence Approach to Credit
Risk [Charles University].
file:///E:/Downloads/DPTX_2013_2_11230_0_415651_0_
151649.pdf
[29] Rojas, R. (1996). Neural networks: a systematic
introduction. In Springer Science & Business Media.
https://books.google.com.br/books?hl=pt-
BR&lr=&id=4rESBwAAQBAJ&oi=fnd&pg=PA3&ots=V
Bf8cRZWqP&sig=wKOJYQs4mZa3iR1F56RB-
rzB6zM&redir_esc=y#v=onepage&q&f=false
[30] Siegel, S., & Castellan Jr, N. J. (2006). Estatística não-
Paramétrica Para Ciências do Comportamento (2nd ed.).
Bookman.
[31] Silva, M. F. da, Silva, W. G. da, Carvalho, R. L. de, Silva,
E. M. da, & Almeida, T. da S. (2019). Analysis of Genetic
Algorithm for synthesis digital systems modeled in finite
state machine. International Journal of Advanced

www.ijaers.com Page | 209