=Paper=
{{Paper
|id=Vol-2546/paper07
|storemode=property
|title=Credit scoring model for microfinance organizations
|pdfUrl=https://ceur-ws.org/Vol-2546/paper07.pdf
|volume=Vol-2546
|authors=Svitlana O. Yaroshchuk,Nonna N. Shapovalova,Andrii M. Striuk,Olena H. Rybalchenko,Iryna O. Dotsenko,Svitlana V. Bilashenko
}}
==Credit scoring model for microfinance organizations==
https://ceur-ws.org/Vol-2546/paper07.pdf
115
Credit scoring model for microfinance organizations
Svitlana O. Yaroshchuk, Nonna N. Shapovalova[0000-0001-9146-1205],
Andrii M. Striuk[0000-0001-9240-1976], Olena H. Rybalchenko[0000-0001-8691-5401],
Iryna O. Dotsenko[0000-0001-7912-2497] and Svitlana V. Bilashenko[0000-0002-4331-7425]
Kryvyi Rih National University, 11, Vitalii Matusevych Str., Kryvyi Rih, 50027, Ukraine
yaroschucksvetlana@gmail.com, shapovalovann09@gmail.com,
andrey.n.stryuk@gmail.com, ellinaryb@gmail.com,
irado441@gmail.com, SvitlanaViktorivnaBilashenko@gmail.com
Abstract. The purpose of the work is the development and application of models
for scoring assessment of microfinance institution borrowers. This model allows
to increase the efficiency of work in the field of credit. The object of research is
lending. The subject of the study is a direct scoring model for improving the
quality of lending using machine learning methods. The objective of the study:
to determine the criteria for choosing a solvent borrower, to develop a model for
an early assessment, to create software based on neural networks to determine
the probability of a loan default risk. Used research methods such as analysis of
the literature on banking scoring; artificial intelligence methods for scoring;
modeling of scoring estimation algorithm using neural networks, empirical
method for determining the optimal parameters of the training model; method of
object-oriented design and programming. The result of the work is a neural
network scoring model with high accuracy of calculations, an implemented
system of automatic customer lending.
Keywords: neural network, machine learning, lending, scoring.
1 Introduction
Increasing the profitability of credit operations is directly related to the quality of credit
risk assessment [1]. In recent years, there has been a rapid increase in retail lending.
Competition is increasing, the range of services is expanding, the process of obtaining
a loan is being simplified, and the decision-making time is significantly reduced. The
quality and speed with which a credit request is generated, as well as the reliability and
simplicity of this process are crucial factors in a complex competitive process.
An important component of the bank’s stable development under conditions of
volatile financial position is the compliance of the risk management system with
modern standards of quality of management, as well as the degree of protection against
unpredictable external influences. Thus, banking organizations need to introduce
scoring systems that allow to resolve qualitatively emerging issues.
Scoring is a way of quickly evaluating a potential customer of a bank or microfinance
organization. The assessment is performed by analyzing the borrower’s questionnaire
___________________
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0).
116
and calculating each customer’s score according to the rules set out in the specific
financial structure. The scoring system is a special program with a built-in algorithm
for deciding on certain parameters. The reliability and quality of the response depends
on the quality of the algorithm. Therefore, it is important not only to create a scoring
model, but also to minimize the error of results.
2 Research apparatus
The aim of the study is the theoretical justification and development of a scoring model
for microfinance borrowers.
Objectives of the study:
1. To analyze the needs of the lending industry in applying scoring assessment, types
of scoring and optimal solutions for this sector.
2. To consider methods of constructing a model for scoring assessment, choose the
most optimal one.
3. To choose a model architecture, create software for practical demonstration of
scoring.
4. To process the input data, evaluate the initial result and achieve maximum accuracy
of the system.
The object of research is creating software for scoring of borrowers.
The subject of research is the development of a scoring assessment model for
microfinance organizations.
Research methods: analysis of the literature on banking scoring; artificial
intelligence methods for scoring; modeling of scoring estimation algorithm using neural
networks, empirical method for determining the optimal parameters of the training
model; method of object-oriented design and programming.
The practical significance of the obtained results is a software for microfinance
organizations that helps to assess the risk of issuing a loan to a specific borrower,
thereby improving the efficiency of these institutions.
3 Theoretical foundations of banking scoring
Credit is an important category of a market economy that reflects the real ties and
relationships of economic life in society. The loan originated from the practical needs
of production development, its adaptation to the conditions of permanent capital
shortage – monetary and material resources.
Credit relations operate in the system of economic relations. They are based on the
movement of a special kind of capital – loan capital.
In today’s context, the approach to credit organization has changed fundamentally:
there has been a shift from object to direct lending to entities. This means that the
emphasis in the lending mechanism has shifted from the selection of the entity to the
entity’s valuation. Commercial and partnership relations between the parties to the
117
agreement exclude the creditor’s dictate in determining the object of credit. The risky
operations that give the highest income to the bank need to study not only the
effectiveness of the activities (projects) under which the funds are allocated, but also
the creditworthiness of the client.
A borrower’s creditworthiness is his ability to fully and timely settle financial
obligations.
The borrower’s creditworthiness, unlike its solvency, does not record any insolvency
for the current period or for any date, but predicts its solvency in the near term.
One way to organize credit relations is to qualitatively assess the creditworthiness of
the borrower. Commercial banks are in dire need of information about the
creditworthiness of farms. Their profitability and liquidity depend largely on the
financial position of the customers, since the reduction of the risk when performing
loan operations can be achieved only based on studying the creditworthiness of clients.
One of the most effective tools for such assessment is the scoring system. Credit
scoring enables to make a quick and qualitative decision on a loan application. In
addition, its reliability and simplicity are crucial factors in the complex competition
[10].
In general, credit scoring can be defined as an assessment of the level of credit risk
that results from the processing of various credit history data, which directly or
indirectly affects the level of payment discipline [1].
Applying credit scoring, that is, a systematic approach to dealing with credit
applications as a whole, allows the bank to:
Increase the loan portfolio by reducing the number of unjustified refusals of loan
applications;
Improve the accuracy of the borrower’s valuation;
Reduce the level of defaults;
Speed up the borrower’s valuation process;
Create centralized accumulation of borrower data;
Reduce provisions for possible losses on credit liabilities;
Quickly and qualitatively evaluate the dynamics of changes in the credit account of
the individual borrower and the credit portfolio as a whole.
All of these have many advantages over a conventional customer rating system and
establish the bank’s performance.
4 Definition of creditworthiness of the client
Determining the borrower’s creditworthiness is an important step in approving a loan
application. The main task of the lender is to assess all the risks associated with the
possibility of non-repayment of the funds provided.
Credit companies consider the following nuances when evaluating creditworthiness
[3; 10; 15]:
The financial position of the potential borrower;
118
Debt load – the ratio of existing liabilities and the requested loan to the applicant’s
principal income;
Credit reputation of the client;
The value of the property owned by the borrower;
The applicant’s social status, personality, career advancement and other factors.
Banking organizations analyze the ability of an individual to pay on a loan. In this case,
not only the borrower’s monthly income and expenses, but also other factors are taken
into account. For example, the risk of job loss and other insured events [15].
Microfinance companies use cheap and fast valuation methods. Therefore, it will
take a minute to conduct a scoring test. This is very handy for small loans. However,
when it comes to large bank loans, credit professionals can apply all of these methods
in combination. This approach will make a specific prediction.
5 Data and methods of scoring in microfinance
5.1 Structure of the main modules
Designing a system that solves the problem of credit scoring can be divided into two
main modules: data processing, which includes bringing data to a format that is
favorable for computer computing, and directly the module of calculations of the loan
decision, containing the interpretation of the algorithm selected for solving the problem
teaching. The output of the first module is the input for the second. Therefore, the
quality of the result depends on the degree of processing of the primary data.
Data processing includes:
Data preparation: delete duplicate records, non-informative data columns, records
with many null values. Assess the significance of each trait included in the training
sample by conducting correlation and regression analyzes.
Data conversion: categorical features are reduced to a vector form and numeric
values should be reduced to a single standard, such as the interval [0, 1] or [-1, 1].
The calculation module consists of:
Choosing the architecture, parameters of the algorithm.
Model training on the selected algorithm.
Testing the model on a deferred sample, determining the calculation error with
further correction [9].
5.2 Methods of scoring
Credit scoring is a typical machine learning task. It refers to the type of supervised
learning (training with the teacher) [4], namely to the problems of classification,
because the solution of the task is reduced to the identification of risks of granting credit
for two types (classes): “good” and “bad”. The good ones will be those customers who
are likely to repay the loan, the bad ones – those who will have a delay of more than 3
119
months. Sometimes, the “bad” risks include those customers who repay loans early or
within a specified period, and the bank does not have time to profit from such clients.
The credit scoring problem can be solved by different machine learning
classification methods [7]. These include [2]:
Statistical methods based on discriminant analysis (linear regression, logistic
regression);
Different linear programming options;
Classification tree or recursion-partition algorithm;
Neural networks;
Genetic algorithm;
Method of nearest neighbors.
Traditional and most common are regression methods, primarily linear multivariate
regression [2]. The disadvantage of the model is that on the left side of the equation is
a probability that takes values from 0 to 1, and variables on the right can take any values
from –∞ to +∞. In addition, this model is unstable to emissions; any sudden value that
gets out of the picture can lead to the wrong answer.
Linear programming also leads to a linear scoring model. It is impossible to carry
out a completely accurate classification of “bad” and “good” clients, but it is desirable
to minimize the error. The task is to find the weights for which the error will be minimal
[17].
Classification trees are a method that allows the observation or object to be assigned
to a particular class of categorical dependent variable according to the values of one or
more predictor variables. Classification trees are tailored to the graphical
representation, so they have a more convenient look for human understanding. The
disadvantages are instability, small changes in the data can significantly change the
built decision tree, the problem of finding the optimal depth of the tree, the complexity
of data gaps.
The genetic algorithm is based on an analogy with the biological process of natural
selection. In the field of lending, it looks like this: there is a set of classification models
that can be “mutated”, “crossed”, and as a result, the “strongest” model is selected,
which gives the most accurate classification.
When using the nearest-neighbor method, a unit of measure is selected to determine
the distance between clients. All clients in the sample are given a specific spatial
position. Each new client is classified based on which clients – good or bad – are more
around him [2].
Neural Networks – a common solution to classification problems. Artificial neural
network – a mathematical model, which is built on the principle of organization and
operation of biological neural networks, is a system of connected and interacting simple
processes [8].
In scoring, the use of neural networks has the least use compared to other methods.
Nevertheless, the neural network has significant advantages. These advantages include
the possibility of automatic learning of the model, the versatility of working with
different scales of measurement of dependent and independent variables, the ability to
approximate any continuous function of dependence. The mathematical model of
120
neural network scoring makes it possible, based on a set of known characteristics of a
research object, to predict a specific characteristic that is unknown to the researcher [6;
18].
The neural network model meets all the above-mentioned needs of the domain, such
as speed and precision of calculations, scalability of data, possibility of introduction of
new characteristics without significant deterioration of quality of estimation, easy
modernization of the system on demand. Therefore, based on the advantages of using
neural networks, this method was chosen to solve the credit-scoring problem.
Building a neural network has its own characteristics and steps that you must go
through to get a truly high-quality model.
Building a neural network starts with data preparation. At this point, it is necessary
to delete duplicate records, non-informative data columns, records that have many null
values. At the same time, there should be a sufficient number of examples for neural
network training. There is an imperial rule that establishes the recommended ratio
between the number of training examples that have input and the correct answers and
the number of connections in the neural network: X<10.
For the facts included in the training sample, it is advisable to estimate its
significance in advance by conducting correlation and regression analyzes and to
consider the ranges of their possible changes. All this is an important component in the
methodology of building a scoring model [3].
The second stage is the conversion of the initial data, taking into account the nature
and type of problem solved by the neural network, the means of presentation of
information are selected. For example, categorical features should be reduced to a
vector form and numerical values should be reduced to a single standard, such as the
interval [0, 1] or [–1, 1].
The third stage is the choice of network architecture. It is necessary to define such
parameters as the number of network layers, the number of neurons of each layer, and
the activation functions to be used [18].
The number and type of independent variables in the model determine the number
of neurons in the input layer. For categorical variables, it is advisable to use one input
neuron for each category, with only one neuron in the group being activated for each
observation.
The architecture of the source layer of the neural network is also dictated by the
structure of the problem. One output neuron is created for each dependent quantitative
variable.
No techniques have been developed to determine the number of hidden layers and
the number of neurons in them. In practice, these parameters are experimentally
determined by analyzing the quality of approximation provided by networks of
different sizes. Thus, a network with an input layer of 58 inputs, one hidden layer with
30 neurons and 2 outputs was selected.
At the input of a neuron, we have a vector of parameters. These are the results of the
collection of billing information about a potential customer, presented in numerical
form = , ,…, .
In this case, each client is responsible for the class . In total, there are two classes
of the set Y with the following values: 1 – to give credit, 0 – not to give. The neural
121
network, in fact, must find the optimal separating hypersurface in the vector space, the
dimension of which will correspond to the number of features. Learning the neural
network in this case is to find such values (coefficients) of the weight matrix, in which
the neuron responsible for the class will give values close to one in cases where credit
is approved, and values close to zero, if not.
As we can see from formula 1, the result of the neuron ℎ ( ) is a function of
activation from the sum of the product of the input parameters to the coefficients
required ∙ in the learning process:
ℎ ( )= ∑| | ∙ (1)
It is desirable to interpret the value derived from a neuron in the range [0, 1] as the
probability of belonging to a class. Therefore, such a monotonous smooth function is
required, which will display elements of the set of real numbers in the range from zero
to one. The activation function of sigmoid (2) is the best way to do this. The function
graph is shown in Figure 1.
( ∙ )= (2)
Fig. 1. Plot of sigmoid function
The fourth stage is learning the network. In the selected data set, each sample object is
assigned a class to which it belongs. Therefore, we are tasked with the type of
supervised learning [16]. Therefore, in the learning process, the network must review
the sample many times, each complete passage of the sample is called the learning age.
To train the model, you need to split the data into two parts – the actual training and the
test.
We apply the standard separation in the ratio of 80% and 20% respectively [11].
Neural network training should be understood as the weighting for each of the traits
based on the results obtained from past data views. The backpropagation method is
selected for weight correction. See [5] for more on this method.
It has been determined experimentally that 64 epochs are required to select the
optimal weights for a given neural network. 221,712 training sample records are
processed for each epoch. The results are Table 2.
122
Table 1. The results of calculating epochs
Epoch number Calculation time, s Error Accuracy
1 22 0.4871 0.6860
2 21 0.3066 0.7271
3 22 0.2472 0.7748
4 21 0.2128 0.7951
5 23 0.1970 0.8155
… … … …
61 22 0.0304 0.9518
62 22 0.0275 0.9529
63 22 0.0255 0.9537
64 22 0.0251 0.9540
The fifth step is to test the received neural network model on a delayed sample. The
test sample includes only those records that did not participate in the training network.
We have 55,428 records. The accuracy of the calculations on the test sample is 95.4%.
In this case, the error in the decision to issue a loan – 0.00469, the error in the loan
prohibition decision is 0.00406.
6 Software architecture and operation
6.1 Architecture
The client server architecture of the program was chosen to develop the software for
rapid assessment of customer solvency.
Client-server is a software architecture model of two parts, client systems and server
systems, both communicating over a computer network or on the same computer. A
client-server application is a distributed system made up of both client and server
software. The client-server application provides a better way to share the workload. The
client process always initiates a connection to the server, while the server process
always waits for requests from any client.
The client-server relationship describes the relationship between the client and how
it makes a service request to the server, and how the server can accept these requests,
process them, and return the requested information to the client [12].
6.2 Description of software operation
To build a functional diagram, the IDEF0 methodology was chosen, which is
considered the classic method of the process approach to design. In IDEF0, the system
that is being modeled is represented as a set of interrelated works (functions, activities).
To develop a functional diagram, 4 main functions of the program were presented,
which are presented on blocks A1–A4 (Fig. 2). Each of these functions has input and
output data, control information and mechanisms through which the function can be
executed.
123
Fig. 2. Functional program scheme
Block A1. Block maintaining the customer profile. At the input, it has data coming from
a bank client. This data is written to the database in a format structured in accordance
with the bank policy and the accepted questionnaire template. The bank operator
maintains the questionnaire. It can both enter data, view it, perform a search, delete it.
Block A2. After the data are entered into the database and the electronic application
is generated, the customer data is searched in external sources. As a rule, these are
Internet portals or credit bureaus. This approach helps to obtain more detailed
information about the client and his solvency.
Block A3. Data processing. It provides functionality for preparing data for further
calculations, all non-informative data is deleted, data is brought to a single range. Data
processing is based on the analysis of data requirements.
Block A4. The processed network dataset enters the neural network model. The
structure and quality of functioning of the neural network are determined by its
architecture. The model gives the result of a loan to the borrower.
All blocks as the executing mechanism have software and hardware.
The system is a module for the banking system, so this imposes certain requirements
on the program interface.
Firstly, an official and minimalistic style of design should be maintained. Light
background colors of the components and dark text color on them.
Secondly, all elements should be located at an optimal distance from each other, to
prevent errors that in the banking sector can cost both time and money.
Thirdly, the size of the inscriptions should be sufficient for the readability of the text.
The text on the components must match the actual functionality that the component
executes.
An important part of the interface is the presence of messages about the actions
performed in the program, as well as the status of request processing.
124
The developed software consists of several tabs: “Clients”, “Data Analysis” and
“Settings”.
The Customers tab has buttons for managing customer data. It is possible to add new
clients, remove them from the database, search the client for the database, and directly
score on the selected client.
The developed software consists of several tabs: “Clients”, “Data Analysis” and
“Settings”.
The Customers tab (Fig. 3) has buttons for managing customer data. It is possible to
add new clients, remove them from the database, search the client for the database, and
directly score on the selected client.
Fig. 3. Customers tab
To view the analytics data to monitor the status of credit disbursements and other
parameters, go to the Data Analysis tab (Fig. 4). The graph will be automatically
generated according to the selected criterion from the drop-down menu.
Fig. 4. Data Analysis Tab
125
The user can also change the neural network settings by going to the Settings tab. The
banking system is dynamic and sometimes it is necessary to re-evaluate the weight of
the features. In the list of features, you should select the ones that will be included in
the new model using the multiple choice in the box. You can select the number of
hidden layers and the number of neurons in them.
7 Results
As a result, a system of automatic crediting of clients in the sphere of microfinance was
created. For this system, the best artificial neural network architecture with all the
relevant settings was selected.
Using this decision support system on the artificial neural network platform to make
serious decisions such as credit decisions significantly simplify lending operations,
reduce the risk of default.
Testing the system gave fairly accurate results, which indicates a high degree of trust
in the software.
The result of the program (Fig. 5):
Fig. 5. The result of the program
8 Conclusion
The study showed how artificial neural networks could be used to build an automatic
customer lending system.
In accordance with this goal, the following results were obtained: the current state
of the problem and task of crediting clients are analyzed; the modern decision support
systems are analyzed and the choice of artificial neural network systems as the basic
technological platform is grounded; the data model necessary for the proper functioning
126
of the system is built; selected the best neural network architecture for a specific task;
automatic system of crediting of clients is implemented.
Building a neural-boundary model for the credit-scoring problem has shown that this
method is one of the best for finding the most accurate solution for issuing a loan, as
evidenced by a very low calculation error.
The introduction of such a system in a microfinance institution helps to improve the
performance of the institution, automate the processes associated with credit loans,
reduce the risk of errors and fraud.
References
1. Aleshin, V.A, Rudayeva, O.O.: Kreditnyj skoring kak instrument povysheniya kachestva
bankovskogo risk-menedzhmenta v sovremennyh usloviyah (Credit scoring as an
instrument for improving the quality of banking risk management in current conditions).
Terra economicus. 10(2), 27–30 (2012)
2. Allison, P.D. (ed.): Logistic regression using the SAS system: theory and application. SAS
Institute, Stanford (2012)
3. Anderson, R.: The credit scoring toolkit: theory and practice for retail credit risk
management and decision automation. Oxford University Press, New York (2007)
4. Coelho, L.P., Richert, W.: Building Machine Learning Systems with Python. Packt
Publishing, Birmingham (2013)
5. Flach, P.: Machine Learning: The Art and Science of Algorithms that Make Sense of Data.
Cambridge University Press, Cambridge (2012)
6. Haykin, S.: Neural Networks and Learning Machines, 3rd edn. Pearson, New Jersey (2008)
7. Kiv, A., Semerikov, S., Soloviev, V., Kibalnyk, L., Danylchuk, H., Matviychuk, A.:
Experimental Economics and Machine Learning for Prediction of Emergent Economy
Dynamics. In: Kiv, A., Semerikov, S., Soloviev, V., Kibalnyk, L., Danylchuk, H.,
Matviychuk, A. (eds.) Experimental Economics and Machine Learning for Prediction of
Emergent Economy Dynamics, Proceedings of the Selected Papers of the 8th International
Conference on Monitoring, Modeling & Management of Emergent Economy (M3E2 2019),
Odessa, Ukraine, May 22-24, 2019. CEUR Workshop Proceedings 2422, 1–4. http://ceur-
ws.org/Vol-2422/paper00.pdf (2019). Accessed 17 Aug 2019
8. Lewis, E.M. An introduction to credit scoring. Athena Press, London (1992)
9. Luo, F.L., Unbehauen, R.: Applied Neural Networks for Signal Processing. Cambridge
University Press, Cambridge (1997)
10. Mays, E. (ed.): Handbook of credit scoring. Global Professional Publishing, Chicago (2001)
11. Rojas, R.: Neural Networks: A Systematic Introduction. Spring-Verlag, Berlin (1996)
12. Saternos, C.: Client-Server Web Apps with JavaScript and Java. O’Reilly Media,
Sebastopol (2014)
13. Semerikov, S.O., Teplytskyi, I.O., Yechkalo, Yu.V., Markova, O.M., Soloviev, V.N., Kiv,
A.E.: Computer Simulation of Neural Networks Using Spreadsheets: Dr. Anderson,
Welcome Back. In: Ermolayev, V., Mallet, F., Yakovyna, V., Kharchenko, V., Kobets, V.,
Korniłowicz, A., Kravtsov, H., Nikitchenko, M., Semerikov, S., Spivakovsky, A. (eds.)
Proceedings of the 15th International Conference on ICT in Education, Research and
Industrial Applications. Integration, Harmonization and Knowledge Transfer (ICTERI,
2019), Kherson, Ukraine, June 12-15 2019, vol. II: Workshops. CEUR Workshop
127
Proceedings 2393, 833–848. http://ceur-ws.org/Vol-2393/paper_348.pdf (2019). Accessed
30 Jun 2019
14. Semerikov, S.O., Teplytskyi, I.O., Yechkalo, Yu.V., Kiv, A.E.: Computer Simulation of
Neural Networks Using Spreadsheets: The Dawn of the Age of Camelot. In: Kiv, A.E.,
Soloviev, V.N. (eds.) Proceedings of the 1st International Workshop on Augmented Reality
in Education (AREdu 2018), Kryvyi Rih, Ukraine, October 2, 2018. CEUR Workshop
Proceedings 2257, 122–147. http://ceur-ws.org/Vol-2257/paper14.pdf (2018). Accessed 30
Nov 2018
15. Siddiqi, N.: Credit risk scorecard: developing and implementing credit scoring. John Wiley
and Sons, New Jersey (2006)
16. Sorokin, A.S.: K voprosu validacii modeli logisticheskoj regressii v kreditnom skoringe
(On the validation of the logistic regression model in credit scoring). Naukovedenie 2.
http://naukovedenie.ru/PDF/173EVN214.pdf (2014). Accessed 10 Nov 2019
17. Sorokin, A.S.: Postroenie skoringovyh kart s ispolzovaniem modeli logisticheskoj regressii
(Construction of scoring maps using a logistic regression model). Naukovedenie 2.
http://naukovedenie.ru/PDF/180EVN214.pdf (2014). Accessed 10 Nov 2019
18. Sorokin, S.V., Sorokin, A.S.: Ispolzovanie nejrosetevyh modelej v povedencheskom
skoringe (Use of neural network models in behavioral scoring). Prikladnaja informatika
10(2(56)), 92–109 (2015)