UNIVERSITY OF NAIROBI

COLLEGE OF BIOLOGICAL AND PHYSICAL SCIENCES

SCHOOL OF COMPUTING AND INFORMATICS

MASTER OF SCIENCE IN COMPUTATIONAL INTELLIGENCE

LOAN DEFAULT PREDICTION USING MACHINE LEARNING :

A CASE OF MOBILE BASED LENDING

AUTHOR: GOLDA TRACEY KISUTSA

REG NO: P52/35489/2019

SUPERVISOR: DR. WANJIKU NG’ANGA

A project report submitted in partial fulfillment of the Master of Science degree in

Computational Intelligence.

2021
DECLARATION

I hereby declare that this research project is my own work, and has to the best of my
knowledge, not been submitted to any other institution of higher learning.

Name: Golda Kisutsa Registration: P52/35489/2019

Signature: Date:

This research project has been submitted as partial fulfilment of the requirements for the award
of the Degree of Master of Science in Computational Intelligence at the University of Nairobi
with my approval as the faculty supervisor.

Supervisor: Dr. Wanjiku Ng’anga

Signature: Date:

i
ABSTRACT

Over the last decade, digital credit has been the fastest growing financial innovation in Kenya.
This has largely been attributed to by technological innovations and mobile phone penetration
enabling expanded access to financial services to individuals who were previously unbanked.
Overall access to formal financial services now stands at 83%, up from 67% in 2016, and 88% of
the adult population has access to a mobile money account (KFSD , 2019).

Precise credit risk assessment also known as loan default prediction is crucial to the functioning
of lending institutions. Traditional credit score models are constructed with demographic
characteristics, historical payment data, credit bureau data and application data. In online mobile
based lending, borrower’s fraudulent risk is higher. Hence, credit risk models based on machine
learning algorithms provide a higher level of accuracy in predicting default.

The main objective of this project is to predict loan default by applying machine learning
algorithms. The proposed methodology involves data collection , data pre-processing , data
analysis , model selection and performance evaluation . This project takes data of previous
customers to whom on a set of parameters loan were approved. The machine learning model is
then trained on that record to get accurate results. The main machine learning algorithms applied
are logistic regressions, naïve bayes and decision trees. The performance of the machine learning
models are then compared using performance metrics and the best machine learning algorithm is
selected to predict the loan default.

ii
TABLE OF CONTENTS

Declaration........................................................................................................................................i

Abstract............................................................................................................................................ii

Table of Contents............................................................................................................................iii

List of Figures.................................................................................................................................vi

List of Tables.................................................................................................................................vii

Chapter 1: Introduction....................................................................................................................1

1.1 Background............................................................................................................................1

1.2 Problem Statement.................................................................................................................2

1.3 Significance of Study.............................................................................................................2

1.4 Research Objectives...............................................................................................................2

1.5 Justification............................................................................................................................2

Chapter 2: Literature Review...........................................................................................................3

2.1 Introduction............................................................................................................................3

2.2 Traditional Credit Risk Assessment.......................................................................................3

2.2.1 Linear Regression...........................................................................................................3

2.2.2 Discriminant Analysis.....................................................................................................3

2.2.3 Probit Analysis and Logistic Regression........................................................................4

2.2.4 Judgment-Based Models.................................................................................................4

2.3 Machine Learning approaches in Credit Scoring...................................................................4

2.3.1 Decision Trees................................................................................................................5

2.3.2 Random Forest................................................................................................................5

2.3.3 Logistic Regression.........................................................................................................5

2.3.4 Neural Network...............................................................................................................6

iii
 2.3.5 Naïve Bayes....................................................................................................................6

2.4 Related Research Work..........................................................................................................7

2.4.1. Prediction For Loan Approval Using Machine Learning Algorithm.............................7

2.4.2 Loan Prediction Using Decision Tree and Random Forest............................................7

Chapter 3: Methodology..................................................................................................................8

3.1 Introduction............................................................................................................................8

3.2 Research Design.....................................................................................................................8

3.3 Data Collection......................................................................................................................9

3.4 Conceptual Design.................................................................................................................9

3.5 Proposed Model...................................................................................................................10

3.5.1 Design Requirements....................................................................................................11

3.6 Data Preprocessing...............................................................................................................12

3.6.1 Data Cleaning................................................................................................................13

3.6.2 Data Reduction..............................................................................................................14

3.6.3 Feature Engineering......................................................................................................15

3.6.4 Exploratory Data Analysis............................................................................................16

3.6.5 Converting Categorical Variables.................................................................................20

3.6.6 Standard Scaler.............................................................................................................20

3.6.7 Handling Outliers..........................................................................................................21

3.6.8 Modelling......................................................................................................................24

3.6.9 Model Testing...............................................................................................................25

3.7 Performance Metrics............................................................................................................27

3.7.1 Confusion Matrix..........................................................................................................27

3.7.2 Accuracy.......................................................................................................................27

3.7.3 Precision........................................................................................................................28
iv
 3.7.4 Recall............................................................................................................................28

3.7.5 Specificity.....................................................................................................................28

3.7.6 F1 Score........................................................................................................................28

Chapter 4: Results and Discussions...............................................................................................29

4.1 Introduction..........................................................................................................................29

4.2 Results..................................................................................................................................29

4.2.1 Decision Tree................................................................................................................30

4.2.2 Logistic Regression.......................................................................................................32

4.2.3 Naïve Bayes..................................................................................................................34

4.3 Discussion............................................................................................................................36

Chapter 5: Conclusions and recommendations..............................................................................38

5.1 Introduction..........................................................................................................................38

5.2 Conclusions..........................................................................................................................38

5.3 Limitations of the Research.................................................................................................39

5.4 Recommendations and Future Work....................................................................................39

References......................................................................................................................................40

Appendices.................................................................................................................................43

Appendix A: Gantt Chart...........................................................................................................43

v
LIST OF FIGURES

Figure 2.1 Neural Networks.............................................................................................................6

Figure 3.1 Research Design.............................................................................................................8
Figure 3.2 CRISP-DM Methodology...............................................................................................9
Figure 3.3 Proposed Model............................................................................................................10
Figure 3.4 Importing Python Libraries..........................................................................................12
Figure 3.5 Train Data.....................................................................................................................12
Figure 3.6 Removing Missing Values...........................................................................................13
Figure 3.7 Preprocessed Data........................................................................................................14
Figure 3.8 Creating Target Variable..............................................................................................15
Figure 3.9 Univariate analysis.......................................................................................................17
Figure 3.10 Gender vs Default.......................................................................................................18
Figure 3.11 Education vs Default..................................................................................................18
Figure 3.12 New Credit Customer vs Default...............................................................................19
Figure 3.13 Employment Status vs Default...................................................................................19
Figure 3.14 Marital Status vs Default............................................................................................19
Figure 3.15 Converting Categorical Variables..............................................................................20
Figure 3.16 Scaling data................................................................................................................20
Figure 3.17 Normalized Income Total...........................................................................................22
Figure 3.18 Normalized Amount...................................................................................................23
Figure 3.19 Variable Declaration...................................................................................................24
Figure 3.20 Splitting Data..............................................................................................................24
Figure 3.21 Test Data Preprocessing.............................................................................................25
Figure 3.22 Handling Outliers in Test Data...................................................................................26
Figure 4.1 Decision Tree Results...................................................................................................31
Figure 4.2 Logistic Regression Results.........................................................................................33
Figure 4.3 Naïve Bayes Results.....................................................................................................35

vi
LIST OF TABLES

Table 3.1 Confusion Matrix...........................................................................................................27

Table 4.1 Dataset............................................................................................................................29
Table 4.2 Decision Tree Confusion matrix...................................................................................30
Table 4.3 Decision Tree Performance............................................................................................31
Table 4.4 Logistic Regression Confusion matrix..........................................................................32
Table 4.5 Logistic Regression Performance..................................................................................33
Table 4.7 Naive Bayes Confusion matrix.....................................................................................34
Table 4.8 Naïve Bayes Performance..............................................................................................35
Table 4.10 Confusion Matrix Comparison.....................................................................................36
Table 4.11 Performance Comparison.............................................................................................37

vii
 CHAPTER 1: INTRODUCTION

1.1 Background

Kenya has made tremendous progress in electricity connectivity , internet penetration and mobile
network coverage. Mobile cellular subscriptions per 100 people has significantly grown from 0.4
in 2000 to 80.4 in 2016. (World Development Indicators , 2016). This rapid telecommunications
and infrastructure development coupled with the global decline in cellphone prices has been
harnessed by companies to provide value-added services such as digital credit.

Credit providers have traditionally required interaction between agents and clients, risk
assessment based on previous financial history, and loans delivered into a bank account. This
excluded those without a bank account or access to a bank branch and those with undocumented
financial histories, This hurdle was readily overcome by digital credit, which refers to loans that
are supplied and repaid digitally, generally using a cell phone. Digital credit is instant , loan
decisions are automated based on a set of rules applied on available data and it is managed
remotely. (CGAP, 2016).

Digital credit has evolved to incorporate a number of different business models. The first model
is a bank and mobile network operator partnership such as M-Shwari by NCBA Bank and
Safaricom. The second model is a non-bank lender and mobile network operator such as Kopa
Cash by Jumo and Airtel Kenya. The third model is a bank utilizing mobile network operator
channels such as MCo-op Cash by Co-operative bank that uses USSD. The fourth model is non-
bank mobile internet applications which involves non-bank lenders disbursing loans through
smartphone mobile application such as Branch and Tala.

Mobile lending platforms use predictive analytics like transaction history , call logs, text
messages, contact lists , age , education, and income to arrive at a credit worthiness score
and limit. When analyzing first-time borrowers, alternative digital data is especially significant,
whereas repayment-based credit history becomes more important for subsequent loan
applications. This research project aims to evaluate the application of machine learning technique
1
to improve the predicted loan default rates.

2
1.2 Problem Statement

Financial mobile lending institutions use credit scoring models to evaluate potential loan default
risks. These models generate a score that translates the likelihood of defaulting on a loan, making
lending decisions easier. Developing credit scoring model is time consuming. These models are
also fixed and do not easily evolve with changing customer behavior to predict default more
accurately. Machine learning approaches can help enhance the accuracy of loan default
prediction.

1.3 Significance of Study

Credit risk assessment is crucial to the success of lending institutions since customer credit risk
affects profitability directly. Traditional procedures are inefficient and time-consuming. The goal
of this study is to investigate the use of machine learning approaches in loan prediction that is
more dynamic and adaptable to changing client data. These techniques will also provide higher
accuracy in predicting loan default.

1.4 Research Objectives

i. To review the existing literature on techniques applied in loan default prediction.

ii. To design a machine learning model that can be forecast loan default.
iii. To train and test the machine learning model to predict loan default.
iv. To evaluate the performance of the model in predicting loan default.

1.5 Justification

The importance of credit risk evaluation has significantly increased with digital credit.
Financial institutions have developed advanced systems to access the credit worthiness of their
customers. The objective of this research is to explore the use of machine learning algorithms to
predict loan default and improve the accuracy of default prediction.

3
 CHAPTER 2: LITERATURE REVIEW

2.1 Introduction

This chapter reviews the existing literature on the traditional credit risk assessment methods, the
machine learning algorithms that are increasingly being used to evaluate credit risk and the
relevance of the literature to the problem statement.

2.2 Traditional Credit Risk Assessment

The type of data used in traditional credit scoring is historical data which includes
bank transactional data such as past credit, records of late payment, credit bureau checks and
commercial data such as financial statements and length of credit history (World Bank, 2019).

2.2.1 Linear Regression

Linear regression is used to determine the linear relationship between explanatory and target
variables. The assumption in a general linear regression issue is that there is a dependent or
response variable yi that is impacted by independent variables. xi1, xi2. ,xiq. A regression
model
can express this relation: yi= 1 xi1 + 2 xi2 +....+ q xiq + where 1, 2,..., q are fixed regression
parameters and is a random error or noise parameter. Before attempting to construct a linear
model, it is critical to assess whether there is a link between the variables of interest.

2.2.2 Discriminant Analysis

Discriminant analysis is a credit scoring approach developed by Sir Ronald Fisher in 1936 to
distinguish between two groups (Fisher, 1936). The most basic form is a two-category label,
such as default versus nondefault. Linear discriminant analysis was the first statistical tool used
to systematically explain which firms went bankrupt using accounting ratios and other financial
indicators in default prediction.
4
2.2.3 Probit Analysis and Logistic Regression

The inverse standard normal distribution of the probability is modelled as a linear combination of
characteristics in the probit (probability unit) model. The log of odds is used by the logit (logistic
unit) function. The log of the label's odds ratio is described as a linear combination of attributes
in the logit model. The following formula is used to predict log odds ratios:
logit(p) = β0 + β1 × x1 + β2 × x2 + . . . + βn × xn (5)

2.2.4 Judgment-Based Models

Multiple strategies are used to create judgment-based models. The analytic hierarchy process
(AHP), which is a structured approach for organizing and analyzing complex decisions, is one
such method. The decision makers break down their choice problem into a hierarchy of easier-to-
understand subproblems, each of which may be studied individually. Human judgments, not just
the underlying data, are used to complete the evaluations in the AHP, according to Bana e Costa,
Barroso, and Soares (2002)

2.3 Machine Learning Approaches in Credit Scoring

Machine learning is based on the development of algorithms that can take in data and apply
statistical analysis to anticipate an output, as well as update outcomes when new data becomes
available. There are three different forms of machine learning. The purpose of supervised
learning is to present a computer with labeled data and estimate the mapping function to the
point that you can predict the output variables (y) for that data when you have new input data (x).
Clustering and association problems are examples of unsupervised learning, in which a computer
is supplied with unlabeled, uncategorized data and the system's algorithms act on the data
without prior training. Reinforcement learning is a type of learning algorithm that learns by
interacting with its surroundings. When the agent performs successfully, he is rewarded, and when
he performs wrong, he is penalized. By maximizing its reward and reducing its penalty, the agent
learns without the need for human intervention.

5
 2.3.1 Decision Trees

A supervised learning approach used to tackle classification and regression problems is decision
trees. To tackle the prediction problem, decision trees use tree representation, in which the
external node and leaf node of a tree represent attribute and class labels, respectively.

The categorical variable decision tree contains categorical target variables that are separated into
categories, such as yes or no. A continuous variable decision tree is one that has a continuous
target variable, such as an individual's income, which can be forecasted using information such
as their occupation, age, and other continuous factors.

2.3.2 Random Forest

Random forest is a machine learning algorithm that is supervised. The bagging approach is used
to train a forest, which is a collection of decision trees. Random forest generates a large number
of decision trees and then combines them to get a consistent and accurate classification. The
Random Forest algorithm has the advantage of being able to be used for both classification and
regression analysis.

2.3.3 Logistic Regression

A classification procedure called logistic regression is used to describe data and explain the
relationship between one dependent binary variable and one or more independent variables.
Binary logistic regression has three major assumptions:
i. The dependent variable must be either binary or dichotomous (e.g. yes or no).

ii. The data should be free of outliers, which can be determined by converting continuous
predictors to standardized scores.

iii. The predictors should not have a lot of strong correlations (multicollinearity). The
independent variables must be unrelated to one another. A correlation matrix among the
predictors can be used to examine this.

6
2.3.4 Neural Network

Artificial Intelligence includes neural networks, which are a type of learning model affected by
the activity of organic neurons. The neural network is made up of nodes that process the data
provided to them and send the results to other nodes. Each node's output is known as the
activation or node value. Weights are assigned to the nodes, which can be changed to assist the
network learn. The magnitude of an input's influence on an output is represented by these
weights. A linear, ramp, move sigmoid, hyperbolic, or Gaussian activation function is used to
perform the net linear calculation. Because it can recognize non-linear regions, the Multilayer
Perceptron Model is utilized to detect fraud.

Figure 2.1 Neural Networks

2.3.5 Naïve Bayes

The Bayes Theorem is used to build a collection of classification algorithms known as Naive
Bayes classifiers. They're employed in a variety of fields for things like prediction and anomaly
detection. Each node represents a variable, and the arcs reflect the relationship between them,
and they are represented by a graph with nodes and directed linkages between them. Although no
information on the Nave Bayes is accessible in fraud detection, the set of variables that cause the
frauds can be computed using the same theorem.

7
2.4 Related Research Work

2.4.1. Prediction for Loan Approval using Machine Learning Algorithm

This study examines banking systems and how banks might reduce non-performing loans or
defaulting loans by forecasting loan defaulters (Kadam, Nikam, Aher, Shelke & Chandgude,
2021). The project's major goal is to use support vector machines and the Nave Bayes method to
forecast loan safety. The subject of forecasting loan defaulters is studied using a critical
predictive analytics technique. Data collection, data pre-processing, model selection, model
evaluation, classification, and result are all part of the suggested model. This project gathers
information from prior clients of various banks who were accepted for loans based on a set of
criteria. To generate reliable results, the machine learning model is then trained on that record..
When comparing the Nave Bayes algorithm to the support vector machine, it was found that the
Nave Bayes algorithm produced the most accurate predictions.

2.4.2 Loan Prediction using Decision Tree and Random Forest

Every year, the number of people or organizations seeking for a loan in India grows.
(Madaan, Kumar, Keshri, Jain, & Nagrath, 2021). Banks must put in a lot of effort to determine
whether a customer can pay back the loan amount on time or not (defaulter or non-defaulter).
The purpose of this paper is to determine the nature, background, and credibility of the client
who is seeking a loan. To cope with the challenge of granting or rejecting a loan request, or in
short loan prediction, we apply exploratory data analysis. The purpose of this paper is to see if a
loan made to a specific individual or organization will be approved. Feature analysis, data
cleaning, exploratory data analysis modeling, and testing are all part of the research's
recommended methodology. Random forest and decision trees were the machine learning
methods used. The decision tree approach had a 73 percent accuracy whereas the random forest
classifier had an accuracy of 80 percent.

8
 CHAPTER 3: METHODOLOGY

3.1 Introduction

This chapter reviews the research design and its justification , the sources of data obtained and its
relevance to the problem statement , the methods for data collection , the data analysis methods
applied and their justifications.

3.2 Research Design

This research project considers the problem of selecting the relevant predictor variables in a
classification. Based on the factors collected during the loan application process, the classification
goal is to forecast if a specific borrower is likely to fail on their loan. The design involves data
collection , data preprocessing , data analysis , model building by applying decision trees, logistic
regression and naive bayes algorithms. Evaluating the performance of the models.

Data Collection
Test data set
Feature Engineering Data missing
Handling Preprocessing
values Outlier Treatment

Data Analysis
Univariate Analysis Bivariate Analysis

Model Building
Logistic Regression Decision Trees Naïve Bayes

Evaluate Performance

Figure 3.1 Research Design

9
3.3 Data Collection

The data collection method is secondary data from Bondora which is a financial lending
institution based in Estonia. The data was obtained from the company website on their public
reports page. The dataset included loan data from the year 2009 to 2021 which is 176,680 rows
and 62 columns. The amount of training data required for a machine learning algorithm varies
according to the complexity of the model, the pattern in the data and the correlation
between attributes. The rule of 10 states the amount of training data needed for a well
performing model should be 10x the number of parameters in the model. Hence , the dataset for
the year 2020 is extracted for the purpose of this study which contains 17,933 rows and 62
columns. The data preprocessing includes extracting the relevant features , handling missing
values, and handling outliers. Finally, the data is split into a train and test set with 70% training
data and 30% test data.

3.4 Conceptual Design

The conceptual design used to develop the model is the CRISP-DM methodology that
incorporates six design phases.

Figure 3.2 CRISP-DM Methodology

10
 i. Business understanding. The first step is to comprehend the research's background, the
problem description, and how the proposed project will achieve the goals.

ii. Data understanding. The second stage requires collection of data listed in the project
resources. This involves describing the data requirements and exploring key data
attributes.

iii. Data preparation. The third stage involves cleaning the data to handle any missing values.

iv. Modelling. This involves determining the modelling technique and testing the design.

v. Evaluation. The is an evaluation of the results achieved to determine the performance of

the model with the best accuracy.

vi. Deployment. The last stage is the implementation of the model.

3.5 Proposed Model

The model will be able to predict whether a loan applicant will default on a given loan.
The system architecture is as below.

Decision Tree

Predict
Train model Naive Bayes Default 1
Loan best model

Logistic
Regression 0

Figure 3.3 Proposed Model

Step 1 : The Loan application goes through the trained model where the three classification
algorithms are applied.
Step 2: The machine learning with the best performance in accuracy is selected.
Step 3 : The machine learning algorithm is applied to the loan application.
Step 4: The machine learning algorithm determines the probability of default. 1, being true and 0
11
being false.

12
3.5.1 Design Requirements

Hardware and Software Requirements

The hardware requirements for the project include a laptop with at least 4GB ram running
windows or Linux operating system. The software requirements include a code editor. This
project uses Microsoft Visual Studio which is a code editor redefined and optimized for building
and debugging modern web and cloud applications.

Python Libraries

The machine learning models are implemented using python version 3.7 on a Jupyter notebook
with the listed libraries: numpy, pandas , matplotlib, seaborn , and sklearn.

i. Jupyter notebooks are a web-based interface in which you can write, visualize, and
execute python code in cells. It is good for exploratory analysis and enable to run
individual code cells.

ii. Numpy is a Python library that may be used to work with multi-dimensional arrays,
linear algebra, the Fourier transform, and matrices.

iii. Pandas is a data manipulation and analysis package written in Python.

iv. Matplotlib is a Python package that allows you to create static, animated, and interactive
visualizations.

v. Seaborn is a matplotlib-based python data visualization package. It has a high-level

interface for creating visually appealing and instructive statistics visuals.

vi. Sklearn is a Python toolkit that allows you to create machine learning and statistical
models including clustering, classification, and regression.

13
3.6 Data Preprocessing

Data preprocessing entails converting raw data into a comprehensible format that a machine
learning model can understand. The data is loaded on a Jupyter notebook in Microsoft Visual
Studio and the python libraries numpy, pandas, matplotlib , seaborn and sklearn are imported.

The dataset contains 17,933 rows and 62 columns before preprocessing. The data preprocessing
involves data cleaning which involves handling missing values, data transformation which
involves normalizing the data and data reduction which involves using only relevant features and
discarding duplicate values of less relevant attributes.

Figure 3.4 Importing Python Libraries

Figure 3.5 Train Data

14
3.6.1 Data Cleaning

The first step of preprocessing is data cleaning by checking and eliminating any missing values
because they affect the accuracy of the model. This is achieved by either filling the missing
values with a mean or mode function or by dropping all missing values. In this case the missing
values are dropped.

Figure 3.6 Removing Missing Values

15
3.6.2 Data Reduction

The next step of the data preprocessing is data reduction. This is used to remove duplicate
features e.g., ‘LoanId’ when there’s ‘LoanNumber’, ‘DateofBirth’ when the feature ‘Age’ is
present. Features relating to dates excluding ‘DefaultDate’ are deleted. The multiple values of
income are also deleted since they are already aggregated in ‘IncomeTotal’. The data is reduced
to 17,933 rows and 20 columns.

Figure 3.7 Preprocessed Data

16
3.6.3 Feature Engineering

When utilizing machine learning to create a predictive model, feature engineering is the act of
choosing and modifying variables in a dataset. The ‘Status’ and ‘DefaultDate’ variables will be
used to create the target variable , ‘Default’. The ‘Status’ variable cannot be used since it has
three unique values current, late and repaid. Late can also not be treated as default since in some
records the loan status is late however the default date is null which implies the loan was not
defaulted but was only late. The ‘DefaultDate’ informs when a borrower defaulted. Combining
both the ‘Status’ feature and ‘DefaultDate’ feature will enable to create the target variable
‘Default’. This is achieved by filtering the loan status to current and checking the default dates to
create a new target variable called ‘Default’ that will have the values 0 if default and 1 if loan is
not default. The ‘Status’ and ‘DefaultDate’ features are removed once the target variable is
created.

Figure 3.8 Creating Target Variable

17
3.6.4 Exploratory Data Analysis

There are two types of independent variables in the data set. Categorical features which include
‘Gender’ , ‘Education’ , ‘MaritalStatus’ and numerical features which include ‘IncomeTotal’ and
‘Amount’.

Univariate Analysis

This term refers to data that consists solely of observations on a single attribute. The basic goal
of univariate analysis is to characterize the data and discover patterns within it. The data is
visually shown using graphing. The primary goal of graphs is to convey data, summarize data,
enhance verbal descriptions, describe and explore data, facilitate comparisons, avoid distortion,
and stimulate thought about the data. The bar graph is the specific graph that has been used. On
the y (vertical) and x (horizontal) axes, the graph is labeled (horizontal axis).The categorical and
ordinal features explored include ‘Gender’, ‘Education’, ‘Marital Status’, ‘EmploymentStatus’,
‘EmploymentDurationCurrentEmployer’ , ‘NewCreditCustomer’.

Observations
− 55% (8,961) of the loans are default.

− 70% of the loan applicants are male.

− Nearly 40% have a secondary education while approximately 30% have a higher education.

− Nearly 25% are fully employed.

− 50% are new credit customers.

18
Figure 3.9 Univariate analysis

19
Bivariate Analysis

The analysis of two variables with the goal of identifying the empirical link is known as bivariate
analysis. The following categorical variables: ‘Gender’, ‘Education’ , ‘EmploymentStatus’ ,
‘MaritalStatus’ , ‘New credit customer’ will be compared to the dependent variable ‘Default’.

Observations
i. Male loan applicants default more than female.
ii. Those with Secondary education default most than the other education status.
iii. New Credit Customers default more than existing credit customers.
iv. Those who have more than 5 years employment duration default more.

Figure 3.10 Gender vs Default

Figure 3.11 Education vs Default

20
Figure 3.12 New Credit Customer vs Default

Figure 3.13 Employment Status vs Default

Figure 3.14 Marital Status vs Default

21
3.6.5 Converting Categorical Variables

Sklearn requires all inputs to be numeric. The categorical variables are converted to numerical
variables using label encoder. The values ‘NewCreditCustomer’ , ‘Restructured’ ,
‘EmploymentDurationCurrentEmployer’ will be converted to numerical values.

Figure 3.15 Converting Categorical Variables

3.6.6 Standard Scaler

To turn data into a distribution with a mean of 0 and a standard deviation of 1, use a standard
scaler.

Figure 3.16 Scaling data

22
3.6.7 Handling Outliers

Outliers are data points that are far apart from other similar points, which could be due to
measurement variability or experimental errors. The range and distribution of attribute values are
particularly important to machine learning algorithms. Outliers in the data might cause the
training process to be misled, resulting in longer training times, fewer accurate models, and
inferior outcomes. To analyze the data and detect any outliers, data visualization is employed.

There are four methods for dealing with outliers in a dataset. Remove the outlier records entirely
to get rid of them. By setting a value range, you can limit the data of outliers. If the data is out of
scope for the intended variable, assign a new value. Using techniques such as log transformation,
data can be transformed.

The ‘IncomeTotal' and ‘Amount' variables include some outliers and are skewed, as can be seen
in the dataset. The log transformation is used to normalize the data. The log transformation is
used to skewed data in order to approximate normality. Because the dataset has a log-normal
distribution, the log-transformed data will have a normal or near-normal distribution as well,
reducing skewness.

23
Normalizing Income Total Variable.

Figure 3.17 Normalized Income Total

24
 i. Normalizing Amount Variable

Figure 3.18 Normalized Amount

25
3.6.8 Modelling

i. Declaration of Variables

The independent variables are declared in x. These include the features ‘NewCreditCustomer’ ,
‘Gender’, ‘Education’ , ‘EmploymentStatus’ and ‘Restructured’. These features are selected
since they are categorical variables. The dependent value is declared in y which is ‘Default’.

Figure 3.19 Variable Declaration

ii. Splitting Data into Train and Test Set

Figure 3.20 Splitting Data

26
3.6.9 Model Testing

i. Preprocessing. This entails loading the test data, dealing with missing values, and
deleting features that won't help predict loan default. Feature alignment as well to ensure
that the features in the test data align with the features in the model.

Figure 3.21 Test Data Preprocessing

27
ii. Handling Outliers

The data in the columns ‘IncomeTotal' and ‘Amount' are skewed to the right, indicating that
majority of the data is skewed to the right due to outliers. Outliers affect the mean and standard
deviation. This can be removed using log transformation to reduce the larger values and

normalize.

Figure 3.22 Handling Outliers in Test Data

ii. Loan Default Prediction

Selecting the independent categorical variables same as the features.

28
Figure 3.23 Loan Default Prediction

29
3.7 Performance Metrics

3.7.1 Confusion Matrix

This produces a matrix that describes the model's overall performance.

Predicted
Positive Negative
Positive True Positive False Positive
Actual
Negative False Negative True
Negative

Table 3.1 Confusion Matrix

True Positives are instances where the forecast is YES and the actual output is YES.
True Negatives: When the prediction is NO and the actual output is also NO.
False Positives are instances where the prediction is YES but the actual outcome is NO.
False Negatives: When the predicted outcome is NO but the actual outcome is YES.

3.7.2 Accuracy

This is the ratio of number of right predictions to the total number of input samples.

Accuracy for the matrix is calculated as:

True Positive + True Negative

Total Sample

30
3.7.3 Precision

When compared to the total number of positively predicted values, this metric shows the number
of True Positives that are truly positive.

Precision = True Positive .

True Positive + False Positive

3.7.4 Recall

The Recall Metric shows how many True Positives the model has classified out of the total
number of samples that should have been positive.

Recall = True Positive .

True Positive + False Negative

3.7.5 Specificity

The number of True Negatives classified by the model out of the total number of samples that
should have been classified as negative is called specificity.

Specificity = True Negative .

True Negative + False Positive

3.7.6 F1 Score

The Harmonic Mean of precision and recall is the F1 Score. F1 Score has a range of [0, 1]. It
demonstrates how exact your classifier is, i.e. how many instances it correctly classifies, as well
as how robust it is, i.e. how many examples it does not miss. The higher the F1 Score, the better
our model's performance.

F1 Score = 2* Precision * Recall

Precision + Recall

31
 CHAPTER 4: RESULTS AND DISCUSSIONS

4.1 Introduction

This Chapter discusses the results obtained from the developed model of loan default prediction.
The results are analyzed with respect to the research objectives and how they relate to the
problem statement and outlined methodology.

4.2 Results

The research addresses the problem of improving the accuracy of predicting loan default in
mobile lending. The developed model is used to classify default and non-default loans by
applying machine learning algorithms. The model is developed from historical loan data of
borrowers using the categorical variable of their loan default status. The model was able to
successfully recognize the behavior patterns of the borrowers and predict the probability of
default of new loan applications. The performance metrics applied to evaluate the performance
of the model were instrumental in determining the best algorithm to determine the probability
of default. These metrics include confusion matrix, precision, recall and F1 score. The decision
tree algorithm had the highest accuracy and precision of 0.64, however it is also equally
important to avoid the misclassification of default loans as non-default loans as this results in
loss. The comparison of the classification results obtained from the three machine learning
algorithms which are decision trees logistic regression and Naive bayes shows the efficiency of
using machine learning algorithms in predicting loan default. The results presented also
accomplish the objectives of the research which is to design a machine learning loan prediction
model , train and test the model and evaluate the performance of the model. The dataset includes
16,046 samples. The dataset distribution is 44% is non default and 56% default.

Non-default Default
(Positive) (negative) Samples
1 = 7,085(44%) 0 = 8,961 (56%) 16,046
Training data (70%) 4,959 6,273 11,232
Test data (30%) 2,126 2,688 4,814

Table 4.1 Dataset

32
4.2.1 Decision Tree

The confusion matrix stating the proportion of correctly classified as well as those misclassified
for each category gives a fulfilling picture of the test result. The result is rounded off to nearest
integer.
- The distribution shows 63% of the data was correctly classified while 27% was
misclassified.

- Misclassifications can be split into type 1 (false positive) and type 2 (false positive
errors). The model contains 11% type 1 errors which means 513 loans were classified as
non- default, but they actually default. This is a problem because the mobile lending
institution will incur a loss when they customers are issued a loan.

- Type 2 error contains 1,233 cases (26%) where the model classifies a borrower will
default but they actually did not default. This doesn’t result in a loss however limits the
mobile lender from issuing loans to customers who would have fulfilled their obligations
and repaid.

Actual
Confusion Matrix Non-default Default
N=4,814 (Positive) (Negative)
Non- default (Positive) Predicted correct Type 1 error
Predicted (True positive) (False positive)

2186 (46%) 513 (11%)

Default (Negative) Type 2 error Predicted correct
(False negative) (True negative)

1233 (26%) 882 (17%)

Table 4.2 Decision Tree Confusion matrix

33
 Accuracy Precision Recall F1score
0.64 0.64 0.64 0.62

Table 4.3 Decision Tree Performance

The model is also evaluated across accuracy , precision , recall and F1 score. The model obtained
a 64% result in accuracy and recall , 63% in precision and 62% in F1 Score.

- Accuracy is the total number of correct predictions made to determine if a loan would be
default or non-default from the entire sample.

- Precision shows the actual number of non-default predictions out of all the all the values
predicted as non-default. This is achieved by:

- Recall shows the percentage of non-default predictions are correct over the total number
of samples that should have been non-default.

- F1 Score is the harmonic mean between precision and recall.

In the case of loan default prediction, the recall metric is most important to determine the
percentage of non-default predictions from all values that should have been predicted as default.

Figure 4.1 Decision Tree Results

34
4.2.2 Logistic Regression

The confusion matrix stating the proportion of correctly classified as well as those misclassified
for each category gives a fulfilling picture of the test result. The result is rounded off to nearest
integer.
- The distribution shows 63% of the data was correctly classified while 27% was
misclassified.

- Misclassifications can be split into type 1 (false positive) and type 2 (false positive
errors). The model contains 12% type 1 errors which means 555 loans were classified as
non- default, but they actually default. This is a problem because the mobile lending
institution will incur a loss when they customers are issued a loan.

- Type 2 error contains 1,221 cases (25%) where the model classifies a borrower will
default but they actually did not default. This doesn’t result in a loss however limits the
mobile lender from issuing loans to customers who would have fulfilled their obligations
and repaid.

Actual
Confusion Matrix N=4,814
Non-default Default
(Positive) (Negative)
Predicted Non- default (Positive) Predicted correct Type 1 error
(True positive) (False positive)

2144 (45%) 555 (12%)

Default (Negative) Type 2 error Predicted correct
(False negative) (True negative)

1221 (25%) 894 (18%)

Table 4.4 Logistic Regression Confusion matrix

35
The model is also evaluated across accuracy , precision , recall and F1 score. The model obtained
a 63% result in accuracy , recall and precision and 62% in F1 Score.

- Accuracy is the total number of correct predictions made to determine if a loan would be
default or non-default from the entire sample.

- Precision shows the actual number of non-default predictions out of all the all the values
predicted as non-default.

- Recall shows the percentage of non-default predictions are correct over the total number
of samples that should have been non-default.

- F1 Score is the harmonic mean between precision and recall.

Accuracy Precision Recall F1score

0.63 0.63 0.63 0.62

Table 4.5 Logistic Regression Performance

Figure 4.2 Logistic Regression Results

36
4.2.3 Naïve Bayes

The confusion matrix stating the proportion of correctly classified as well as those misclassified
for each category gives a fulfilling picture of the test result. The result is rounded off to nearest
integer.
- The distribution shows 61% of the data was correctly classified while 29% was
misclassified.

- Misclassifications can be split into type 1 (false positive) and type 2 (false positive
errors). The model contains 23% type 1 errors which means 1109 loans were classified as
non- default, but they actually default. This is a problem because the mobile lending
institution will incur a loss when they customers are issued a loan.

- Type 2 error contains 780 cases (16%) where the model classifies a borrower will default
but they actually did not default. This doesn’t result in a loss however limits the mobile
lender from issuing loans to customers who would have fulfilled their obligations and
repaid.

Actual
Confusion Matrix N=4,814
Non-default Default
(Positive) (Negative)
Predicted Non- default (Positive) Predicted correct Type 1 error
(True positive) (False positive)

1590 (33%) 1109 (23%)

Default (Negative) Type 2 error Predicted correct
(False negative) (True negative)

780 (16%) 1335 (28%)

Table 4.7 Naive Bayes Confusion matrix

37
The model is also evaluated across accuracy , precision , recall and F1 score. The model obtained
a 63% result in accuracy , recall and precision and 62% in F1 Score.

- Accuracy is the total number of correct predictions made to determine if a loan would be
default or non-default from the entire sample.

- Precision shows the actual number of non-default predictions out of all the all the values
predicted as non-default.

- Recall shows the percentage of non-default predictions are correct over the total number
of samples that should have been non-default.

- F1 Score is the harmonic mean between precision and recall.

Accuracy Precision Recall F1score

0.61 0.62 0.61 0.61

Table 4.8 Naïve Bayes Performance

Figure 4.3 Naïve Bayes Results

38
4.3 Discussion

The distribution shows decision trees and logistic regression classified 63% of the data was
correctly classified while 27% was misclassified. Naives Bayes classified 61% of the data was
correctly classified while 29% was misclassified. Misclassifications can be split into type 1 (false
positive) and type 2 (false positive errors).

Type 1 error refers to the loans that were classified as non-default, but they default. This is a
problem because the mobile lending institution will incur a loss when they customers are issued
a loan. Decision trees had the lowest type 1 error at 11% followed by Logistic regression at 12%
and lastly Naïve Bayes at 33%. This is measured by precision.

Type 2 error refers to the loans that were classified as default, but they did not default. This
doesn’t result in a loss however limits the mobile lender from issuing loans to customers who
would have fulfilled their obligations and repaid. Decision trees had the highest type 2 error at
26% followed by Logistic regression at 25% and lastly Naïve Bayes at 16%. This is measured by
recall.

Actual
Non-default (Positive) Default (Negative)
Confusion Matrix
N=4,814
Non- default (Positive) Predicted correct Type 1 error
(True positive) (False positive)
Decision Tree 2186 (46%) 513 (11%)
Logistic Regression 2144 (45%) 555 (12%)
Naïve Bayes 1590 (33%) 1109 (23%)
Default (Negative) Type 2 error Predicted correct
Predicted (False negative) (True negative)
Decision Tree 1233 (26%) 882 (17%)
Logistic Regression 1221 (25%) 894 (18%)
Naïve Bayes 780 (16%) 1335 (28%)

39
Table 4.10 Confusion Matrix Comparison

40
The machine learning algorithms have various assumptions which affect the performance results.

Decision Trees Logistic Regression Naïve Bayes

The variables are categorical The variables are categorical The variables are categorical
At the beginning, the whole The predictor variables are The predictor variables are
training set is considered as independent independent
the root.
Records are distributed There is no multicollinearity
recursively on the basis of (this occurs when two or
attribute values more exploratory variables
are highly correlated.
There are no extreme outliers

Decision tree has the highest accuracy , recall leading with a minimal margin in comparison with
Logistic regression. Naives Bayes has the least accuracy , precision , recall and F1 score.
The performance metrics are important in giving a wholistic view of the algorithms. In the case
of loan default prediction that has a direct impact on revenue, recall has a significance impact.
Recall shows how many True Positives the model has classified from the total number of all
samples that should have been identified as positive. The slightest margin makes a significance
as they determine the overall accuracy and the best lending decision to reduce the risk of default.

Accuracy Precision Recall F1 Score

Decision Trees 0.64 0.64 0.64 0.62
Logistic Regression 0.63 0.63 0.63 0.62
Naïve Bayes 0.60 0.62 0.61 0.61

Table 4.11 Performance Comparison

41
 CHAPTER 5: CONCLUSIONS AND RECOMMENDATIONS

5.1 Introduction

This Chapter discusses the value of the research , the limitations of the research , the conclusion
in accordance with the evidence presented and recommendations for further study.

5.2 Conclusions

The application of machine learning techniques in the financial sector with the goal of profit
maximization has seen a rising interest over the last few years. There has been increasing number
of research conducted in the areas of credit scoring, risk management and bankruptcy prediction
using machine learning approaches. Rapid telecommunications and infrastructure development
in Kenya coupled with the global decline in cellphone prices has led to an increase in mobile
lending. Customers who were previously unbanked can now access digital credit from their
mobile phones. This new dynamic creates an opportunity and a challenge to mobile lending
institutions on how best to make lending decisions to determine whether a customer will default
on a loan.

This research proposes machine learning as a method to improve the accuracy of loan default
predictions. This better understanding of customer behaviors to improve the prediction of loan
default will contribute to tremendous financial benefit in the mobile lending sector. This research
successfully explores the features of loan data that contribute to the risk of loan defaults.
Exploratory data analysis shows the correlation of various features with loan default to select the
most appropriate features to train the machine learning model. The train and test data set are then
applied to three machine learning algorithms to determine the one with the most accurate results.
Key performance metrics which include confusion matrix, accuracy , precision and recall and
applied to evaluate the best machine learning technique in loan default prediction.

42
5.3 Limitations of the research

Machine learning algorithms are limited to the dataset used to train and test the model.
Data is governed by data protection laws making it challenging to access primary data for the
purpose of research. This research was conducted using secondary open data from bondora.com
that is available to the public. This limits the generalization of the model as it is specific towards
the dataset used to train and test the model. It would be beneficial to look comprehensively at the
main features that are relevant to the characteristic that drive default and can be applied.
Further limitation is in reference to the variables provided in the dataset, although the dataset is
open due to data protection laws some factors may not be available to the public and these may
have had an impact on the predictions of the probability of default. Lastly, the research focused
on the probability of default in a default state however default loans may still be recovered
during the recollection process,

5.4 Recommendations and Future Work

This research explores using machine learning algorithms to improve the accuracy of predicting
loan default. This model will be instrumental to mobile lending institutions in evaluating their
customer credit risk. The best performing model in the research which is decision trees achieves
an accuracy of about 64%. This is a fair performance and can further be improved through
different methods of parameter tuning and feature selection which may possibly yield
improvements in the model performance. It may also be beneficial to do a cross validation with
other sources of open dataset as they become more accessible to compare the performance of the
model. Since the research is also limited to the probability of default in a default state , further
exploration may be made in determining the expected return of the loan based on borrower’s
characteristics , loan characteristics the recollection of loans processes.

43
REFERENCES

[1].Oates, B. (2006). Researching information systems and computing. London: SAGE.

[2].Gwer, F., Odero, J., & Totolo, E. (2020). Digital credit audit report: Evaluating the
conduct and practice of digital lending in Kenya – Financial Sector Deepening Kenya.
Retrieved
18 October 2020, from https://fsdkenya.org/publication/digital-credit-audit-report-
evaluating-the-conduct-and-practice-of-digital-lending-in-kenya/

[3].Credit scoring approaches guidelines - World Bank. (n.d.). Retrieved October 19, 2020,
from http://pubdocs.worldbank.org/en/935891585869698451/CREDIT-SCORING-
APPROACHES-GUIDELINES-FINAL-WEB.pdf

[4].Central Bank of Kenya, Kenya National Bureau of Statistics and FSD Kenya. 2016. The
2016 FinAccess household survey.

[5].Sousa, M., Gama, J. and Brandão, E., 2016. A new dynamic modeling framework for
credit risk assessment. Expert Systems with Applications, 45, pp.341-351.

[6].Arun, K., Ishan, G., & Sanmeet, K. (2016). Loan Approval Prediction based on Machine
Learning Approach. IOSR Journal Of Computer Engineering, 18(3), 79-81.

[7].Kadam, A., Nikam, S., Aher, A., Shelke, G., & Chandgude, A. (2021). Prediction for
Loan Approval using Machine Learning Algorithm. International Research Journal of
Engineering and Technology, 08(04), 4089-4092.

[8].Blessie, E., & Rekha, R. (2019). Exploring the Machine Learning Algorithm for
Prediction the Loan Sanctioning Process. International Journal of Innovative Technology
and Exploring Engineering, 9(1), 2714-2719.

[9].Aithal, V., & Jathanna, R. (2019). Credit Risk Assessment using Machine Learning
Techniques. International Journal of Innovative Technology and Exploring Engineering,
9(1), 3482-3486. doi: 10.35940/ijitee.a4936.119119

44
[10]. Mohammadi, N., & Zangeneh, M. (2016). Customer Credit Risk Assessment using
Artificial Neural Networks. International Journal of Information Technology and
Computer Science, 8(3), 58-66. doi: 10.5815/ijitcs.2016.03.07

[11]. Blessie, C. (2019). Exploring the Machine Learning Algorithm for Prediction the
Loan Sanctioning Process. International Journal of Innovative Technology and Exploring
Engineering, 9(1), 2714-2719. doi: 10.35940/ijitee.a4881.119119

[12]. Uddin, S., Khan, A., Hossain, M., & Moni, M. (2019). Comparing different
supervised machine learning algorithms for disease prediction. BMC Medical Informatics
and Decision Making, 19(1). doi: 10.1186/s12911-019-1004

[13]. Madaan, M., Kumar, A., Keshri, C., Jain, R., & Nagrath, P. (2021). Loan default
prediction using decision trees and random forest: A comparative study. IOP Conference
Series: Materials Science and Engineering, 1022, 012042. doi: 10.1088/1757-
899x/1022/1/012042

[14]. Cook, T., & McKay, C. (2015). How M-Shwari works: The story so far. Forum 10,
Washington, DC: CGAP and FSD Kenya. Retrieved from
http://www.cgap.org/sites/default/files/Forum-How-M-Shwari-Works-Apr-2015.pdf

[15]. Bhandari, M. (2020, October 19). Predict Loan Eligibility using Machine Learning
Models. Medium. https://towardsdatascience.com/predict-loan-eligibility-using-machine-
learning-models-7a14ef904057

[16]. Xu, Z. (. (2021, March 10). Loan default prediction with Berka dataset. Medium.
https://towardsdatascience.com/loan-default-prediction-an-end-to-end-ml-project-with-
real-bank-data-part-1-1405f7aecb9e

[17]. Massaoudi, T. (2020, October 23). ML basics : Loan prediction. Medium.

https://towardsdatascience.com/ml-basics-loan-prediction-d695ba7f31f6

45
[18]. Aleksandrova, Y. (2021). Comparing Performance of Machine Learning
Algorithms for Default Risk Prediction in Peer-to-Peer Lending. TEM Journal, 133-143.
https://doi.org/10.18421/tem101-16

[19]. Madaan, M., Kumar, A., Keshri, C., Jain, R., & Nagrath, P. (2021). Loan default
prediction using decision trees and random forest: A comparative study. IOP Conference
Series: Materials Science And Engineering, 1022, 012042. doi: 10.1088/1757-
899x/1022/1/012042

[20]. Bondora.com. Bondora.com. (2021). Retrieved 4 July 2021, from

https://www.bondora.com/en/public-reports.

[21]. Jafar Hamid, A., & Ahmed, T. (2016). Developing Prediction Model of Loan Risk
in Banks Using Data Mining. Machine Learning and Applications: An International
Journal, 3(1), 1-9.

46
APPENDICES

Appendix A: Gantt Chart

47