Journal of Business and Management Studies

ISSN: 2709-0876
DOI: 10.32996/jbms
JBMS
AL-KINDI CENTER FOR RESEARCH
Journal Homepage: www.al-kindipublisher.com/index.php/jbms AND DEVELOPMENT

| RESEARCH ARTICLE

Machine Learning-Based Risk Prediction Model for Loan Applications: Enhancing

Decision-Making and Default Prevention

Anchala Chouksey✉1 , Md Shihab Sadik Shovon2 , Nikhil Rao Tannier3 , Proshanta Kumar Bhowmik4 ,
Miraz Hossain5 , Md Shafiqur Rahman6 , Md Khalilor Rahman7 , and Md Sazzad Hossain8

1Masters of Science in Financial Mathematics, University of North Texas, Denton, Texas

2
Master of Science in Engineering Management, College of Graduate and Professional Studies, Trine University.
3Master’s in Artificial Intelligence, University of North Texas

4Department of Business Analytics, Trine University, Angola, IN, USA

5Master of Business Administration, Westcliff University

6MBA in Management Information System, International American University

78MBA, Business Analytics, Gannon University, Erie, PA, USA

Corresponding Author: Anchala Chouksey, E-mail: [email protected]

| ABSTRACT
The primary objective of this research was to develop a machine learning model for loan application risk prediction that achieves
maximum reliability in decision-making while minimizing risks of default. This study focused on credit application risk
assessment in the context of the USA finance industry because challenges and opportunities in this industry are unique in their
manner. The dataset for this analysis comprises in-depth records of applicants for loans that exhibit a vast range of characteristics
of borrowers, credit history, and repayment behaviors. Comprehensive in scope, the rich dataset has variables that span age,
earnings, employment status, and locality alongside other crucial finance variables such as credit scores, debt-to-income ratio,
and repayment performance. For model selection, we utilized a variety of machine learning algorithms, including Logistic
Regression, Random Forest Classifier, and XG-Boost. The Random Forest and XG-Boost models closely aligned with actual data,
showing high accuracy. The integration of predictive modeling of advanced levels within loan decision processes has far-
reaching consequences on building lender confidence within risk assessments. By using evidence-driven facts through machine
learning models, lenders can make better-informed decisions that better reflect greater insight into borrower behavior and
attributes of risk. Looking ahead, numerous directions of future research can advance AI capability and AI-based loan risk
assessment software. A critical direction is investigating how to use deep learning techniques, which have shown much promise
in numerous fields of endeavor through their ability to learn complex nonlinear relationships within large datasets.

| KEYWORDS
Loan Risk Prediction, Machine Learning, Credit Scoring, Default Prevention, Financial Decision-Making, USA

| ARTICLE INFORMATION
ACCEPTED: 15 November 2023 PUBLISHED: 12 December 2023 DOI: 10.32996/jbms.2023.5.6.13

I. Introduction
Background and Setting

Accurate risk measurement is of greatest significance in credit approval to determine how well a finance organization manages its
portfolio. Not only does the ability to predict how likely it is that a debtor would not repay a loan affect individual credit decisions
but it also affects economic balance on a larger scope by preventing inefficient allocation of funds (James, 2021). Over the last few

Copyright: © 2023 the Author(s). This article is an open access article distributed under the terms and conditions of the Creative Commons
Attribution (CC-BY) 4.0 license (https://creativecommons.org/licenses/by/4.0/). Published by Al-Kindi Centre for Research and Development,
London, United Kingdom.
Page | 160
 JBMS 5(6): 160-176

decades, the United States has witnessed alarming defaults on loans that have imposed severe economic implications on lenders
in the way of greater write-offs and lower profitability. All these have served to make it clear that more sophisticated methods of
debtor risks need to be employed in their analysis, considering that the economic environment has become more sophisticated
and fluid in dynamics (Hsu et al., 2019).

Historically, financial institutions have employed traditional credit score models that typically take only several variables into
account to establish the creditworthiness of a borrower based on credit record, level of earnings, and debt-to-income ratio. As
good starting points, though these have remained, these have not considered multi-dimensional aspects of borrower actions or
all of the variables that would contribute to loan default (Babaei et al., 2020). They may not consider crucial variables such as
economic conditions, locality of living, or even social determinants of health that have far-reaching effects on the repayment
capacity of a borrower. There is, thus, a need to move from these conventional methods to more advanced, data-facilitated
methods that take advantage of machine learning and predictive analytics (Gupta, 2021).

According to Jui et al. (2023), the evolution from conventional credit scoring to AI-based predictive analytics is a paradigmatic
breakthrough in how finance companies establish risk determination. Machine learning algorithms have the potential to analyze
vast amounts of disparate data from varied origins to offer more refined insight into borrower activity and risk profiles than ever
before. As per Nasiruddin et al. (2023), through pattern recognition that may not necessarily reveal itself through conventional
analysis, machine learning allows for more informed lender choices that ultimately lead to more robust loan performance and
reduced default risks. As digital disruption sweeps through finance companies, integration of machine learning in processes for
judging risks on loans is not only apropos but necessary for competitiveness and fiscal soundness.

Problem Statement

Rana et al. (2023), argued that despite the technological breakthroughs, traditional risk measurement models have substantial
limitations in correctly categorizing high-risk borrowers. They typically make judgments based on records that might not accurately
reflect economic conditions or new developments at present or in the future, thus causing potential errors in the classification of
borrowers. They might even create biases by disproportionately affecting specific population groups based on past credit
dispensations. This poses a double challenge to lenders to balance their need to sustain their acceptance levels to sustain their
business while at the same time reducing their exposure to risks to escape expensive defaults.

Moreover, the economic environment is constantly fluctuating with changing economic conditions, regulations, and consumer
demand. As a result of these changes, lenders must work with outdated models that cannot account for these fluid conditions. As
a result of these imbalances, severe financial losses accrue that have the potential to erode consumer confidence in finance
companies. As such, cutting-edge solutions that have the potential to address these challenges by delivering a more inclusive and
balanced approach to credit risk measurement are needed (Sizan et al., 2023).

Research Objectives

The primary objective of this research is to develop a machine learning model for loan application risk prediction that achieves
maximum reliability in decision-making while minimizing risks of default. Through the utilization of the potential of machine
learning algorithms, this model aims to make more precise judgments of the risks of borrowers so that lenders would be in a better
position to identify high-risk applicants more accurately and make more sound choices when accepting or rejecting loans. Not
only does this approach attempt to make lending organizations more financially secure, but it also hopes to make lending more
inclusive by eliminating biases in traditional credit scores.

To achieve that objective, research would involve aggregating and comparing disparate databases of characteristics of borrowers,
credit records, economic factors, and other factors that have a potential impact on the repayment of loans. Through machine
learning approaches, research would explore some of these algorithms and their performance in their ability to predict defaults
on loans to identify the most suitable model for deployment in real-life lending scenarios. Research results would be of invaluable
assistance to lenders to make their processes of determination of risks more effective and to regulators to drive good practice in
lending in the finance sector.

Scope of the Study

This study will focus on credit application risk assessment in the context of the USA finance industry because challenges and
opportunities in this industry are unique in their manner. The application of machine learning to increase predictability and finance
protection is of particular importance in light of the economic conditions at the moment that are subject to changes in interest
rates, inflation on the rise, and greater uncertainty. Directing research to the United States lending environment shall provide
results that shall yield implementable counsel that shall be easy for finance organizations to apply in their environment.

Page | 161
Machine Learning-Based Risk Prediction Model for Loan Applications: Enhancing Decision-Making and Default Prevention

Furthermore, the relevance of this research moves beyond individual lenders to policymakers and regulators as well. As lending
evolves further, guidelines that foster good lending practices, mitigate risks, and protect consumers need to be established more
than ever before. By illustrating that machine learning is capable of accurately anticipating risks in loans, this research hopes to
make its contribution to the greater discourse on technology and regulation in finance in building a more secure financial system.

II. Literature Review

Loan Risk Prediction and Default Trends in the USA

Chiamka (2023), reported that the examination of the United States loan default and forecast risks displays a complicated
interrelation of several economic variables, consumer behaviors, and credit policies that have immense implications on repayment
levels of credit. Over the past decades, levels of defaults have varied depending on general economic conditions such as
employment levels, levels of interest, and general economic cycles. For instance, in economic boom conditions, levels of defaults
fall as people have higher incomes and more secure employment that makes it easy for them to make their repayment schedules.
Kumari et al. (2022), found that during economic recessions like that of 2007-2008, however, levels of defaults increase as people
face layoffs, reduced earnings, and more economic strain. This pattern displays why it is crucial to take economic conditions into
account in anticipating repayment risks of credit.

Moreover, specific economic conditions have been found to forecast default on loans. For example, higher levels of unemployment
have greater associated default levels, as loss of employment has a clear impact on someone’s ability to make payments on their
loans. Similarly, higher levels of interest have the potential to escalate repayment difficulties for individuals with variable-rate loans,
consequently elevating their potential for default (Jillson, 2021). Housing conditions in markets further play a vital part in
influencing the performance of loans, more so for mortgage loans. Declining property values have potential negative equity for
property owners that translates to defaults as individuals realize that their property is more than that which they owe on it (Piskunov
et al., 2023).

Spagher et al. (2018), asserted that the impact of credit policies on credit risks cannot be overlooked. Regulations such as the
Dodd-Frank Wall Street Reform and Consumer Protection Act have been instituted to provide greater consumer protection while
reducing systemic risks in the wake of the credit crunch. As much as these policies helped stabilize credit conditions, these policies
have levied additional costs of compliance on lenders that may limit credit to certain categories of credit-seekers. Urban et al.
(2022), ascertained that as lenders adjust their credit policies following these rules, they must balance their need to manage risks
with that of credit provision to credit-seekers in a sensitive manner. This interrelation gives way to heterogeneity in probabilities
of default across demographical categories that require sophisticated measures of risks that account for economic as well as
regulatory conditions at large.

Traditional vs. Machine Learning-Based Loan Risk Assessment

According to Sizan et al. (2023), the traditional approaches to credit measurement of loan credit have largely relied on conventional
credit scoring approaches such as the FICO score that consolidates a borrower's credit history, repayment profile, and outstanding
balances into a numerical credit score that captures creditworthiness in all its facets. Other measures of debt-to-income (DTI) ratio
have similarly acted as pillar measures in the creditworthiness of a borrower by providing insight into their ability to make debt
payments every month in proportion to their revenues. As much as these conventional approaches have provided credit
measurement of credit risk with a basic platform to work from, these have several disadvantages that make their performance in
loan defaults' predictability questionable at times.

One significant disadvantage of traditional credit scoring methods is that these methods make assumptions based on records that
may not reflect economic conditions or specific situations of borrowers at present. For instance, creditworthy borrowers may face
economic challenges from sudden events like illness or sudden loss of jobs that traditional methods may not consider (Babaei et
al., 2020). Besides that, traditional scoring methods have challenges in coping with the detailed nuances of the actions of borrowers
and all of the variables that may impact repayment capacity. Over-simplification of these factors results in inappropriate
categorizing of risks that may deny credit to creditworthy individuals or grant credit to high-risk individuals unfairly, ultimately
contributing to greater potential for defaults (Caffo et al., 2022).

In contrast, credit risk assessment through machine learning has several advantages that offset these disadvantages. Employing
algorithms that possess the capability to sift through large amounts of data from disparate sources, machine learning models
possess the potential to discern sophisticated interlinkages and patterns that traditional methods may not discern. They possess
the potential to consider a vast array of variables, from borrower details to economic signals to even behavioral signals, to provide
more inclusive credit risk measurement (Gupta, 2021). Machine learning models possess the potential to improve over time by
constantly learning from fresh details to make their forecasts more accurate and their estimations more precise. Possessing the

Page | 162
 JBMS 5(6): 160-176

potential to discern sophisticated risk patterns makes machine learning a forceful tool for credit risk improvement that has the
potential to make more accurate estimations for lenders to make more prudent judgments and ultimately reduce default rates
(Chiamaka, 2023).

Machine Learning in Financial Decision-Making

Hsu et al. (2019), articulated that the application of credit risk analysis through machine learning in finance depicts how these
technologies have the potential to make lending operations more efficient and more dependable. Supervised algorithms have
received special attention for their credit risk forecast performance. They have been trained on experiences in the past so that they
have learned from past outcomes to apply to analyze fresh credit requests. Tools like logistic regression, decision trees, and neural
networks have found extensive application in developing forecast models that establish the probability of default based on several
characteristics of the borrowers.

As per James (2021), one of the greatest benefits of supervised learning models is that they can take in a broad array of input
features to provide a more in-depth view of borrower risk. For instance, whereas traditional credit scores might focus on credit
history alone, machine learning models can include other variables like social media activity, online activity, or even other alternate
data points like utility payments that offer insight into how financially sound a borrower might be. Through their multi-faceted
approach, lenders have a greater capacity to determine high-risk applicants more accurately and make their lending strategy more
specific to their target clientele.

Success stories of AI-enabled loan approval processes further establish the potential of machine learning in credit decision-making.
Financial organizations have implemented these technologies to automate their credit processes, enhance their capabilities in
credit risk evaluation, and provide more sophisticated experiences to their customers. For instance, organizations like Zest-Finance
and Upstart have employed machine learning algorithms to take unconventional points of view to make credit decisions for people
who would have gone unnoticed by conventional scoring methods (Jillson, 2021). Not only do these developments extend credit
to underserved people, but they also reduce default rates and improve the performance of credits. Through utilizing capabilities
of machine learning, these organizations take center stage in redefining credit risk analysis and management, ultimately resulting
in a more inclusive and solid financial ecosystem (Kumari et al., 2022)

Research Gaps

Urban et al. (2022), argued that regardless of the possibilities of credit risk evaluation through machine-learning methods, several
research challenges must still be investigated and further developed. Foremost of these challenges is how to develop explainable
yet adaptable AI-based credit risk models that would have the capability to respond to changes in economic conditions as well as
in the behaviors of borrowers. As economic conditions fluctuate, it is needed that risk evaluation models remain adaptable by
continuously refining their algorithms to include new information and wisdom. Concurrently, however, their responsiveness must
accompany explainability so that lenders, as well as borrowers, would be aware of why credit judgments have been made through
these AI-based credit risk models. The opaqueness of many of these machine-learning algorithms has the potential to cause
mistrust and suspicion on the part of consumers if these judgments have not been properly justified or explained to them (Piskunov
et al., 2023)

Additionally, there is not enough technology in existence for real-time measurement of loan risks that would provide instant
insight into the profile of risks of borrowers. Most credit scoring in traditional methods captures static snapshots of a borrower’s
credit history that may not accurately reflect their current economic situation enough. Having the potential to ascertain loan risks
in real-time would enable lenders to counter changes in the circumstances of their borrowers more flexibly, hence enabling more
effective mitigation of defaults and more effective portfolio management (Gupta, 2021). There is a need for measurement in real-
time in an age of rapid technological advancement and changing consumer behaviors in specific. Closing these research gaps
would enable the finance industry to move towards more progressive, inclusive, and effective credit measurement practices,
ultimately leading to more solid lending conditions (Caffo et al., 2022).

III. Data Collection and Exploration

Dataset Overview

The dataset for this analysis comprises in-depth records of applicants for loans that exhibit a vast range of characteristics of
borrowers, credit history, and repayment behaviors. Comprehensive in scope, the rich dataset has variables that span age, earnings,
employment status, and locality alongside other crucial finance variables such as credit scores, debt-to-income ratio, and
repayment performance. Data for the set is taken from several banks, credit agencies, and online lenders to offer good coverage
of disparate lender-client records and credit policies. By aggregating these varied databases, the set offers rich insight for forecast
analysis of loan risks and improvement of default estimations.
Page | 163
Machine Learning-Based Risk Prediction Model for Loan Applications: Enhancing Decision-Making and Default Prevention

Feature Selection

S/No Feature/Attribute Description

01. Credit Utilization Ratio Proportion of how much credit outstanding on credit cards by
a borrower to their credit limits, or how much credit is used
about their credit limits.
02. Employment Stability Coded variable for how many years of employment at their
current job someone has accumulated, which may indicate their
earnings stability level.
03. Loan-to-Value Ratio (LTV) Proportion of the loan to the appraised property value for
secured financing to ascertain the probability of default
concerning collateral value.
04. Payment History Yes or no indicator that states if the debtor has ever previously
been in arrears on payments, which is a good predictor of
future repayment performance.
05. Account Age: The credit account age of the debtor might indicate their credit
experience in operating their accounts.
06. Borrower debt payments in proportion to their gross earnings
Debt-to-Income Ratio (DTI) every month, presenting their debt burden in sharp focus.
07. Recent Credit Inquiries: Recent credit inquiries on the credit record of the debtor
demonstrate credit shopping activity or distress on their part.
08. Geographic Risk Factor A feature that assigns levels of risk based on where the
borrower is situated, considering regional economic conditions
as well as historical default experience.
09. Social Media Sentiment Aggregated sentiment from social media activity on or related
to the borrower offers additional behavioral signals that may be
linked to creditworthiness.
Data Preprocessing

In the pre-processing of data, we used a systematic approach to make the dataset consistent and analyzable. To start with, we
preprocessed by cleansing the data by eliminating errors like duplicates and erroneous entries to make the general quality of the
dataset better. Thereafter, we preprocessed by scaling numerical variables to have similar ranges to make all variables have similar
effects on model performance. To fill in missing values, imputation processes like replacing with mean or median for numerical
variables and replacing with mode for categorical variables to have a low loss of information but not impact the integrity of the
dataset were used. Last but not least, feature extraction was used to identify relevant variables for model training by applying
processes like correlation analysis and recursive feature elimination to finally make the dataset more refined to improve model
performance by selecting features that significantly contributed to the predictability of loan defaults.

Exploratory Data Analysis (EDA)

According to Nasiruddin et al. (2023), exploratory Data Analysis (EDA) is part of the research methodology that comprises
inspection and graphical presentation of data to unveil inherent structures, patterns, and outliers before launching onto proper
modeling. There are numerous crucial objectives that EDA accomplishes: it allows researchers to acquaint themselves with
variables' distributions and associations, to determine potential outliers, and to ascertain the goodness of the data through
descriptive statistics and graphical displays such as histograms, boxplots, and scatterplots. Jui et al. (2023), asserted that by
providing insight into the structure of the data as well as to potential points of further inspection or pre-processing that may need
to occur, EDA guides the next research steps, impacting feature choice and hypothesis formation. Overall, not only does EDA make
the analysis more robust, but it provides more insight into the data so that model-building work is based on full awareness of its
characteristics.

Monthly Loan Application Trend

The code snippet plots a line graph of the application trend for loans monthly. It begins by resetting the plot style to "white" from
within the seaborn library. The plot size of 12x5 inches is set for the plot. At its center is the Data Frame df, which is grouped by
column 'Application Date' for the month of every column ('Month') and counted for unique user IDs ('UID') in every column
('Month'). This count is shown on a line graph with markers ('o') to identify every month's count uniquely. The plot has the title
"Monthly Loan Application Trend," the title for the x-axis as "Month," and the title for the y-axis as "Number of Applications."

Page | 164
 JBMS 5(6): 160-176

Finally, x-axis ticks are set to display months of the year for easy comparison of months of maximum applications and general
observation of the trend.

Output:

Figure 1: Monthly Loan Application Trend

The chart of a pattern of monthly application of loans illustrates the variability in the level of application received
throughout the year, with a clear indication of the pattern of seasonality. Notably, figures exhibit highs in application in specific
months, which would have shown months of higher borrowing activity, having arisen from economic conditions or seasonal factors.
For instance, months 3 and 6 have maximum figures of application, which would have shown people borrowing more in the spring
months and initial months of summer due to higher consumer spending or fiscal cycle budgeting. Conversely, months 7 and 10
have low figures for an application that would have shown low consumer spending or consumer confidence in these months.
Overall, through these pattern observations, crucial points for lenders to make changes in their strategy and resource allocation
have been highlighted, underscoring how awareness of seasonality in borrowing application patterns would make forecastability
and lender decision-making more effective.

Loan Amount Distribution

The implemented code snippet makes a histogram of how loan amounts are distributed. It makes a plot of size 10x5
inches by calling plt.figure(). At its center is sns.histplot(), which makes a histogram of column 'Amount' of Data Frame df. Using
parameter bins=30 makes the histogram have 30 bins so that it offers detailed insight into how the distribution looks. kde=True
parameter overlays on the histogram of column 'Amount' of Data Frame df a Kernel Density Estimate curve so that it offers a
smooth view of how the data is distributed. The plot is further enhanced by giving it the title "Loan Amount Distribution" and label
on the x-axis as "Loan Amount." Last but not least, plt.show() makes the histogram that has been produced viewable so that its
frequency and variability of loan amounts can be analyzed.
Output:

Figure 2: Loan Amount Distribution

Page | 165
Machine Learning-Based Risk Prediction Model for Loan Applications: Enhancing Decision-Making and Default Prevention

The histogram of financing demand by financing amount has striking revelations of lender demand for borrowing and
lender preferences. There is a dramatic spike at $10,000 that identifies that financing amount as the most requested by applicants,
with over 1,200 cases reported that would suggest that it may identify a popular financing demand for applicants. There is evident
bimodality in financing demand by counts of requests by financing amount with secondary highs at lower ranges around $2,500
and $5,000 that identify substantial counts of requests for low-end financing that may identify financing for applicants that need
financing for low-scale or brief financing. There is significantly lower demand for higher financing requests for higher ranges of
$15,000 or higher with decreasing counts of requests by higher financing ranges. There is more indication of these dynamics from
an overlay of the smoothed curve that illustrates decreasing counts by financing ranges as financing ranges increase higher than
$10,000. Overall, this analysis identifies clustering of financing requests around specific financing ranges that provide good insight
for lenders to customize their offers to fulfill financing demand of applicants well.

Loan Term Distribution

The executed code snippet made a count plot of term distribution. It makes a plot of size 8x4 inches by using plt.figure().
At its center is sns.countplot() that counts how many of each unique term in column 'Term' of Data Frame df occurs and displays
these in bars. This gives easy insight into how many of each term of loan occurs. To further enhance that plot, it is assigned a title
of "Loan Term Distribution," an x-axis label of "Loan Term (Months)," and a y-axis label of "Count." Finally, plt.show() displays the
count plot that has been created so that analysis of how many of each term of loan occurs can be performed.
Output:

Figure 3: Loan Term Distribution

The histogram of the breakdown of the terms of loans offers an interesting insight into the demand for how long to repay their
loans by borrowers. There is a clear concentration in the term of 60 months that has over 2,500 of these requests, which implies
that borrowers like lengthy repayment schedules that have lower payments that fit their budgets more. There is consistently
decreasing demand for short repayment schedules that have much lower counts for repayment schedules of 12 months and 18
months, which implies that borrowers may not like to have quick repayment schedules so much. Terms of repayment of 36 months
and 48 months have mid-interest but not nearly that of the term of repayment of 60 months. Overall, that breakdown shows that
borrowers tend to like lengthy repayment schedules that make their repayment more affordable to their budgets by having more
months to repay their loans in full.

Approval Rate by Employment

The code snippet generates a bar plot visualizing the relationship between employment type and loan approval success rate. It
begins by setting the figure size to 8x4 inches using plt.figure(). The core of the visualization is sns.barplot(), which created a bar
plot with 'Employment Type' on the x-axis and the 'Success' rate on the y-axis. This implicitly calculates and displays the proportion
of successful loan applications for each employment type. The plot is then labeled with the title "Approval Rate by Employment
Type," an x-axis label "Employment Type," and a y-axis label "Approval Rate". Finally, plt.show() displays the generated bar plot,
allowing for a comparison of loan approval rates across different employment types.

Page | 166
 JBMS 5(6): 160-176

Output:

Figure 4: Approval Rate by Employment

Feature Correlation Matrix

The histogram of loan approval rates by category of employment reveals distinct differences in loan approval levels by category
of employment. The most prominent loan-approval rate is shown by Employment Type 1, approximately 0.15, indicating that
individuals within this category have better loan-approval probability compared to others. The reverse is seen where there is a
lower loan-approval rate of approximately 0.05, as shown by Employment Type 2, indicating that individuals within this category
may face tighter lending conditions or lender perceptions of higher risk. A moderate loan-approval rate is shown by Employment
Type 0, while that of Employment Type 3 is worst, indicating that individuals within this category have better difficulty accessing
loan approval. The variance bars portray variance within each category of employment, providing insight into lending non-
homogeneity. The overall observation is that there is evident evidence of how the category of employment influences loan approval
probability, indicating that loan assessors may need to adjust lending terms based on distinct risk factors within each category of
employment.

Feature Correlation Matrix

The code generates a heatmap of Data Frame df's correlation matrix. It begins by making plt.figure()'s size is eight by 6 inches. The
heart of visualization is sns.heatmap(), where it computes and displays df's correlation matrix. Df.corr() function computes columns'
pairwise correlations. annot=True displays correlations' values within each cell of the heatmap. cmap="cool warm" is what color
scheme is to use cool color gradation where negative correlations are cool while positive correlations are warm. fmt=".2f" is what
is shown by annot=True but only up to two decimals only. The plot is labeled "Feature Correlation Matrix." Finally, plt.show()
displays a created heatmap where relationships can easily be identified along with how strong or weak each is within a set of
features of the dataset.

Output:

Figure 5: Feature Correlation Matrix

Page | 167
Machine Learning-Based Risk Prediction Model for Loan Applications: Enhancing Decision-Making and Default Prevention

The feature correlation matrix helps provide an overall insight into how features of the dataset correlate with each other
on the -1 to 1 range of correlation coefficients. The most striking relationship is that of loan amount and loan term, where there is
a 0.57 coefficient suggesting that there is a moderate positive relationship where loan amount increases. Loan terms also tend to
increase, possibly suggesting that customers prefer manageable monthly installments. The date of application is weakly correlated
to most features, possibly suggesting that it might have no significant effect on loan outcomes. The job type of applicants and
loan use have very weak correlations to most features, possibly suggesting that these attributes might have no significant relation
to loan success or the success of loan requests. The success feature is weakly negative to amount (-0.11), possibly suggesting that
loan amounts and terms might have no relation to loan success, or weakly negative to loan terms (-0.10), possibly suggesting that
loan terms might have no relation to loan success. The feature correlation matrix helps provide insight into key relationships
between features, informing follow-on analysis or possible modeling activities.
Loan Approval Rate by Loan Amount Range

The code script in Python compares loan amount vs. loan success by binning loan amounts into bins and graphing each loan
amount range vs its respective success percentage. It generates a new Data Frame column 'Loan Amount Bin' by binning the Data
Frame column 'Amount' into bins of 0-5k, 5k-10k, 10k-20k, 20k-50k, 50k-100k using pd.cut(). It creates a bar plot of loan amount
bins vs success percentage using sns.barplot() by putting 'Loan Amount Bin' on the x-axis and success percentage on the y-axis.
The argument ci=None suppresses the displaying of confidence intervals, and palette="cool warm" is set to specify the color
scheme. The title of the plot is set to "Loan Approval Rate by Loan Amount Range," along with axis titles. Finally, plt.show() is
invoked to plot the created bar plot, allowing observation of how success percentages of loans vary by loan amount range.

Output:

Figure 6: Loan Approval Rate by Loan Amount Range

The histogram of loan approval percentages by loan size range reveals significant differences in the probability of approval by
loan size range. The most approving range is that of $0-5k, approximately 0.12, suggesting that lower loan amounts have a better
probability of approval, possibly due to lower lender perceptions of risk. The second-highest approving range is that of $5-10k,
also approximately 0.12, suggesting that applicants requesting slightly larger loans continue to have a good probability of
approval. The approving percentage falls significantly, though, for the range of $10-20k, to approximately 0.04, and continues to
fall off for loans of greater than $50k, suggesting that larger loans have tighter lending restrictions and possibly higher levels of
corresponding risk factors. The trend reveals that applicants have increasing difficulty as loan size increases, suggesting that larger
loans must have lending criteria carefully evaluated by lenders on corresponding levels of risk while possibly reconsidering lending
thresholds on larger loans.

Average Loan Amount by Loan Term

The implemented code script generates a loan amount by loan term heatmap by first producing a pivot table called pivot_table
using df.pivot_table(). The Data Frame df is averaged by 'Term,' and the meaning of 'Amount' is taken by each term using
df.pivot_table(). It generates a plot of size eight by 6 inches using plt.figure(). The backbone of visualization is sns.heatmap(),
displaying the pivot table as a heatmap. cmap='Blues' is utilized to set the color scheme to blues. annot=True is utilized to plot
values of the average loan amount within each cell of the heatmap, and fmt=".0f" is utilized to format values of the average loan
amount to integers. The plot is finally decorated by adding the title "Average Loan Amount by Loan Term" using plt.title(), adding
a label on the x-axis using plt. Label (), adding label on the y-axis using plt. Label (), and finally, displaying the generated heatmap
using plt.show().
Page | 168
 JBMS 5(6): 160-176

Output:

Figure 7: Average Loan Amount by Loan Term

The histogram of loan amount by loan terms reveals a clear trend of increasing loan amounts corresponding to increasing loan
terms, i.e., loan terms of larger lengths are generally associated with larger loan amounts on average. For instance, the loan amount
of a 12-month loan is approximately $1,997, increasing substantially to approximately $10,460 for a loan of 48-month terms,
suggesting that there is clear correspondence of loan size with loan terms of larger lengths. The maximum loan amount is reached
by 60-month terms, up to $11,255, suggesting that customers of extended terms of loans make bigger loan requests, presumably
to make lower monthly installments within loan terms of larger lengths. The trend can also reflect that customers like extended
terms of loans for big funding requests, whereas customers like to make lower installments every month within loan terms of larger
lengths. Overall, this evidence verifies that loan terms should also be considered along with loan amounts, as it is helpful to
determine lending strategies and borrower demand.

Loan Applications by Day of the Week

The provided script was supposed to plot loan requests by the day of the week on a bar plot. It starts by adding a new column,
"Day of Week," to Data Frame df by taking the names of the days out of "Application Date." The size of the figure is set to 10 by 5
to make it readable. The function sns. Counterplot of library Seaborn is invoked to count and plot the number of requests by each
day, ordered by Monday through Sunday, to make it evident what each is supposed to represent. The plot is titled "Loan
Applications by Day of the Week," and the axis is labeled accordingly, where "Day of the Week" is on the axis at the bottom and
"Number of Applications" is on the axis at the top. Finally, plt.show() is invoked to make the plot visible on screen. The visualization
can establish trends of loan requests by day, a valuable insight that can guide lenders in busy times of requests.

Output:

Figure 8: Loan Applications by Day of the Week

Page | 169
Machine Learning-Based Risk Prediction Model for Loan Applications: Enhancing Decision-Making and Default Prevention

The histogram of loan requests by weekday reveals distinct trends of borrower activity throughout the week. Wednesday stands
out by far, with its peak of well over 600 loan requests, suggesting that midweek is maybe a convenient or desirable point to make
loan requests. Following closely behind are Monday and Tuesday, each of which has significant levels of requests, while on
Thursday, there is a decline, suggesting that there is a shift in borrower activity throughout the week. Requests continue downward
on Friday, Saturday, and Sunday, with Sunday showing a minimum of around 300 loan requests. The trend here is of loan requests
peaking in the middle of the week, suggesting that loan processors can learn how to most efficiently process loan requests on or
around that point while also suggesting that weekend loan requests are maybe lower. Overall, trends like this can inform marketing
campaigns and operational planning within lending institutions.

IV. Methodology

Feature Engineering

As per Jui et al. (2023), in the feature engineering process, one should prioritize picking predictive features that have maximum
contribution to loan approval outcomes. Critical features were credit history, income, debt-to-income ratio, and loan purpose since
features like these are well recognized to have impacts on the creditworthiness of borrowers. We also created new features to
enhance the interpretability of the model along with predictive power. For instance, we created a feature of "credit utilization" by
taking the outstanding debt-to-credit availability ratio, which gives insight into how well or poorly credit is being utilized by a
borrower. We also converted continuous features like income into discrete bins to view trends better. Such prudent selection and
feature engineering created a good groundwork for the modeling process that followed, making sure that the models were able
to make use of the most meaningful data.

Model Building and Training

For model selection, we utilized a variety of machine learning algorithms, including Logistic Regression, Random Forest Classifier,
and XG-Boost. Logistic Regression is easy to interpret and is therefore appropriate to use as a baseline model. According to Sizan
et al. (2023), the Random Forest Classifier is also helpful due to its robustness to overfitting along with its ability to learn complex
feature interactions. XG-Boost, also known to have good performance on structured data, is incorporated to leverage its gradient-
boosting properties. The reason behind using these is that features of the dataset have a moderate feature count along with a
binary target variable (approval of loan or not). It is possible to test simple or complex models using this set of features on
predictive accuracy (Rana et al., 2023).

Model Optimization and Performance

To optimize our model performance, we employed hyperparameter tuning through cross-validation and grid search strategies.
We compared systematically differing combinations of parameters within our models to identify the most effective parameters
that give maximum predictive power. We also employed feature importance to establish what features were most important to
our model outcomes, enhancing interpretability and informing potential business decision-making. We also compared our model
precision and recall through confusion matrices, offering insight into how well our models were approving or denying loans. As
per Jui et al. (2023), with this multi-step process of checking our performance, our models were effective and interpretable,
supporting our loan risk assessment goals.

Evaluation Metrics

We employed varied measures of evaluation to establish how well each of the models had functioned, including F1-score, precision,
recall, accuracy, and ROC-AUC. Overall, the correctness of the model predictions was evaluated using accuracy, while precision
and recall helped establish how well the model had functioned concerning positive class predictions. The F1-score, balancing out
precision and recall, was particularly useful where there had been an unbalanced class distribution. The use of the ROC-AUC metric
permitted us to establish how well the model had discriminated between positive and negative classes using varied thresholds of
classification. Sizan et al. (2023), by making comparative performance judgments using these measures, we were better able to
establish the most effective model to use for loan risk prediction, ensuring that our final selection had been effective and
appropriate to the aim of our project.

V. Results and Analysis

Model Performance Evaluation

a) Logistic Regression Modelling

This code script demonstrates constructing and analyzing Logistic Regression model to predict loan success or failure. It begins
the pre-processing of data: features (X) are separated by dropping columns ‘Success’ and ‘UID,’ categorical features are converted
Page | 170
 JBMS 5(6): 160-176

into numerical by one-hot encoding through pd.get_dummies(), missing values are imputed, and float data type is assigned to it.
The target feature (y) is assigned to the ‘Success’ column. The data is split into a train set (80%) and a test set (20%) through
train_test_split() by preserving stratification on the target feature distribution. A Logistic Regression is initialized with maximum
iterations set to 1000 and is trained on the train set. It is passed through a test set to predict, and its performance is analyzed
through accuracy, classification report (with precision, recall, F1-score), and confusion matrix. All of these metrics provide a clear
view of how accurately the model can predict loan success or failure.

Output:

Classification Report:
precision recall f1-score support

0 0.91 1.00 0.95 668

1 0.00 0.00 0.00 68

accuracy 0.91 736

macro avg 0.45 0.50 0.48 736
weighted avg 0.82 0.91 0.86 736

Table 1: Logistic Regression Classification Report

The table above shows the test set result of a Logistic Regression model to predict loan outcomes, where overall accuracy is
90.76%. The classification report also reveals marked differences between the two classes: on class "0" (denied loans), the model
is good on precision (0.95) and recall (1.00), i.e., it is good at picking nearly all loans that were denied without too many false
alarms. For class "1" (approval of loans), its precision and recall are very poor (both 0.00), i.e., the model is poor at making good
predictions of loans that were approved, i.e., no true positives are picked up. The values of support also indicate that there were
668 of class "0" but only 68 of class "1" within the test set. The confusion matrix also reveals this skewness, where there were 668
true negatives but no true positives, i.e., there is likely to have been a skewness of the class that will need to be addressed within
runs of the model in the future. Overall, though good at picking loan denials, its poor performance on loan approvals is where
there is most scope for improvement.

b) Random Forest Classifier Modelling

The code snippet demonstrated how to train and test a Random Forest Classifier model. It begins by importing the Random Forest
Classifier within sci-kit-learn's ensemble package. A Random Forest model is initialized with 100 trees (n-estimators=100) and a
deterministic random state for reproducibility purposes (random-state=42). The model is trained on the train set (X-train, y-train).
Predictions on a test set (X-test) are created, followed by checking how well or badly the loan requests are classified by the model
using metrics like accuracy, classification report that has precision, recall, F1-score, etc., and a confusion matrix. All of these metrics
provide a good overall insight into how well or badly loan requests are classified by the model.

Output:

Classification Report:
precision recall f1-score support

0 0.91 1.00 0.95 668

1 0.00 0.00 0.00 68

accuracy 0.90 736

macro avg 0.45 0.50 0.47 736
weighted avg 0.82 0.90 0.86 736

Table 2: Random Forest Classification Report

The table depicts test outcomes of using a Random Forest model to predict loan outcomes, yielding 90.33% accuracy. The
classification report verifies that on class "0" (denied loans), the model is very accurate (0.95) but perfect on recall (1.00), i.e., it
identifies nearly all loans that have been denied without resulting in much of a false positive result. But on class "1" (approved
Page | 171
Machine Learning-Based Risk Prediction Model for Loan Applications: Enhancing Decision-Making and Default Prevention

loans), the model is poor on precision and recall, each of which is 0.00, i.e., it fails to predict loan approvals, resulting in no true
positives. The values of support also indicate that there are 668 of class "0" but only 68 of class "1." The confusion matrix also
reveals this same weakness, indicating 665 true negatives and three false positives but no true positives. It means that while the
model is good on loan denial detection, it is very poor on loan approvals, indicating that there is a persistent weakness of class
unbalance that will have to be adjusted by modeling in the future.

c) XG-Boost Modelling

The code snippet demonstrated how to train and test an XG-Boost Classifier model. It starts by importing XGB-Classifier from the
library xg-boost. The XG-Boost model is created using use-label-encoder=False to make sure that there is no label encoding
problem and eval_metric='logloss' to specify what metric is used to train the model. The model is trained on a train set (X-train, y-
train). Predict on the test set (X-test) is created, and its performance is evaluated using accuracy, a classification report that has
precision, recall, F1-score, and a confusion matrix. All of these metrics provide an overall insight into how well XG-Boost is
performing on loan classification tasks.

Output:

Classification Report:
precision recall f1-score support

0 0.91 0.98 0.95 668

1 0.28 0.07 0.12 68

accuracy 0.90 736

macro avg 0.60 0.53 0.53 736
weighted avg 0.85 0.90 0.87 736

Table 3: XG-Boost Classification Report

The table above presents the test result of an XG-Boost-based loan outcome predictor that had a global accuracy of 89.67%. The
classification report also shows that on class "0" (denied loans), the model is good, where its precision is 0.95, and its recall is 0.98,
meaning that it is good at selecting most loan denials accurately. For class "1" (approved loans), however, the performance of the
model is poor, where its precision is only 0.12, but its recall is 0.74, meaning that it identifies only a proportion of approved loans
accurately but makes many false negatives. The values of the support also reflect that there are 668 instances of class "0" and 68
instances of class "1." The confusion matrix also reports 63 true negatives, five false positives, 13 false negatives, and 51 true
positives, meaning that although the model is good at selecting loan denials, it is far behind on selecting approvals, judging by its
relatively poor F1-score of 0.87 on class "1." It is evidence of ongoing issues of class imbalance and that there is a need to improve
on selecting loan approvals accurately.

The code in the Python program compared how well each of the three machine learning models - Logistic Regression, Random
Forest, and XG-Boost - is doing by running each on a test set. It makes a function called evaluate-model that takes a trained model
and test set, makes predictions on it, and returns a dictionary of accuracy, precision, recall, and F1-score. It iterates through a
dictionary of models, runs each through the function defined, and puts each result into another dictionary. The resulting dictionary
is converted to a Pandas Data Frame to make it easy to use and view. The Data Frame is finally ordered by accuracy in reverse
order and printed out, making it easy to compare how well each of the four metrics is represented by each of the four models,
making it easy to select a model based on your preference of what is most important.

Output:

Model Performance Comparison:

Accuracy Precision Recall F1 Score
Logistic Regression 0.907609 0.000000 0.000000 0.000000
Random Forest 0.903533 0.000000 0.000000 0.000000
XGBoost 0.896739 0.277778 0.073529 0.116279

Page | 172
 JBMS 5(6): 160-176

Figure 9: Model Performance Comparison

The table and chart above also display a comparative snapshot of performance measures of three models—Logistic Regression,
Random Forest, and XG-Boost—used in loan outcomes prediction. It is important to observe that each of the models has very
poor values of precision and recall on class "1" (approval of loans), where Logistic Regression and Random Forest have values of
0.00, indicating that no approved loans were picked up by either of them. XG-Boost is better on its part but by very thin margins,
where its precision is 0.28, and its recall is 0.07, but here also, its values indicate poor performance on loan approvals. For overall
accuracy, values of the models range from 89.67 by XG-Boost to 90.76 by Logistic Regression, indicating that though overall there
is good performance by each of the models, this is due to good performance on class "0" or loan denial by each of the models.
The Random Forest and XG-Boost models closely aligned with actual data, showing high accuracy. The fact that there are no
positive values on class "1" by each of the models is an indicator that there is a persistent class imbalance that needs redressing
to predict loan approvals better.

Feature Importance Analysis

Feature importance analysis is also critical to establish the predictive power of each of the attributes that contribute to loan default
predictability. By employing feature importance measures such as permutation importance and tree-based feature importance
measures, we were able to establish that there is a set of key variables that have significant impacts on loan default probability.
Among them, credit history is by far the most important, owing to its strong association with the repayment behavior of customers.
Notably, this is expected, given that credit history is generally representative of the reliability of borrowers along with historical
financial conduct, meaning that customers with better credit ratings have lower default probabilities. The debt-to-income ratio is
also demonstrated to be important, where values of increasing ratio reflect increasing financial burden on customers. The feature
importance analysis demonstrated that customers with DTI ratio values of above 40% have higher default probabilities, taking into
account that customers may have challenges honoring financial obligations. The income stability, represented by the history of
employment along with the income category of salaried versus freelance, is also demonstrated to be important. The loan purpose
is also demonstrated to have significant impacts on repayment probability, where loans taken on investments or business activities
may have different risk profiles compared to consumption loans taken by customers. Overall, findings on attributes of borrowers
contribute to better interpretability of the model but also provide lending institutions valuable insight to screen out applicants of
high risk along with better optimization of processes of determination of risk.

Practical Implications for Lenders

The insights of feature importance analysis have significant practical implications on how loan approval processes can be
streamlined by lenders without compromising the effective management of risk. From analyzing the outcomes of the model,
lenders can give priority to key factors such as credit history and DTI ratios throughout the underwriting process. It makes possible
a more refined loan approval process that enables lenders to give priority to applicants who have good repayment ability. For
instance, lenders can have tighter terms on applicants who have poor credit history or high DTI ratios while providing softer terms
on applicants who have good income sources and good credit history. The findings also point out that lenders can institute
targeted financial education programs that enhance the creditworthiness of prospective applicants, especially applicants that have
thin credit history or have high DTI ratios. By enhancing financial literacy, lenders can ensure that people make good financial
decisions that enhance credit history, hence default reduction levels.

Page | 173
Machine Learning-Based Risk Prediction Model for Loan Applications: Enhancing Decision-Making and Default Prevention

Moreover, striking a balance between financial inclusivity and managing default risks is important. The lender can employ these
findings to tailor loan products that are suitable to meet varied borrower needs without enhancing default risks excessively. For
example, offering lower loan amounts or extended repayment periods to higher-risk borrowers can ensure financial inclusivity
without enhancing default levels excessively. Besides, employing non-traditional sources of data, such as rent or utility payment
history, can provide better insight into a borrower’s creditworthiness, especially of poor or non-credit history borrowers. By doing
this, default levels can be reduced, but financial inclusivity can also be assured by extending funding to good borrowers while
keeping default risks within control. Overall, implementing model findings within loan decision processes can yield better-informed
decision-making that fosters better lending conditions that are balanced by social responsibility along with profitability.

VI. Practical Applications

Impact on loan decision making

The integration of predictive modeling of advanced levels within loan decision processes has far-reaching consequences on
building lender confidence within risk assessments. By using evidence-driven facts through machine learning models, lenders can
make better-informed decisions that better reflect greater insight into borrower behavior and attributes of risk. Not only is this
analytical process better suited to improve reliability within risk assessments, but also lender confidence in loan forecasting of
performance is elevated. The ability of the models to identify customers of high risk through effective feature importance analysis
makes it possible for lenders to better personalize underwriting processes. The result is that lenders can enact differentiated
strategies on borrower profiles, ensuring credit is extended to individuals who have demonstrated good repayment capability. The
use of real-time loan forecasting of risk can also make the processes of approval much simpler. By utilizing real-time data analytics,
lenders can approve loan requests on the fly, reducing loan times of approval and overall operational effectiveness. Not only is
this decision process better suited to improve customer satisfaction through quick responses, but it is also better suited to let
lenders control their risk exposure by responding to changing conditions of economies and borrower situations on a prompter
scale.

Regulatory compliance and fair lending

As lenders increasingly turn to AI-based risk assessment software, ensuring that such programs remain compliant with US financial
laws is of top importance. Regulatory bodies aim to ensure fair lending by compelling lending institutions to make lending
decisions without discriminating on racial or gender or covered characteristic bases. It is, hence, critical that lenders ensure that
their machine-learning programs remain compliant with the Equal Credit Opportunity Act (ECOA) and the Fair Housing Act (FHA).
For this, thorough tests of programs need to detect and neutralize bias that can result unintentionally through the use of programs
or through datasets that train programs. It is not merely compliance but also doing what is right by ensuring that there is no fair
lending law breach on its part. The lender must ensure that its AI system is transparent, explaining how decisions are reached and
that what is employed to make lending decisions is transparently justified. By actively eliminating bias and ensuring fair lending
law compliance, lenders can improve credibility and foster better relationships with diverse populations. By doing this proactively,
lenders can lead to fairer lending decisions that make the financial environment fairer overall.

Scalability and Future Applications

The scalability of loan risk assessment through the use of machine learning-based models opens up exciting possibilities for
extended use, particularly in mortgage lending and lending to small business ventures. With distinct features and corresponding
risk profiles of these loan products, there is ample scope for adapting existing models to suit the needs of these markets. For
instance, mortgage lending is beset by added complications of property values and repayment history that can be integrated
within predictive models to make better-informed decision-making possible. Similarly, lending to small business ventures presents
distinct challenges like varying cash flows and business sustainability tests that can be addressed by upgrading the models to
employ industry-specific parameters of risk. Besides, the integration of predictive models within AI-based fraud-detection software
can significantly enhance safety within lending processes. By employing advanced analytical parameters to detect abnormal trends
that can point to fraudulent transactions, lending institutions can secure business processes and ensure that customers, too, remain
safe from scams. Not only is this dual process enhancing lending process integrity, but it also makes every player safe within its
realm. With increasing technological upgrades, integration of risk assessment and fraud detection will increasingly assume greater
importance, opening up possibilities of a well-rounded lending framework that is balanced on both parameters of efficiency and
safety.

Page | 174
 JBMS 5(6): 160-176

VII. Discussion and Future Directions

Challenges of AI-Based Loan Risk Measurement

The implementation of AI-based loan risk assessment systems is beset by challenges that must be navigated to ensure that they
remain effective and morally applicable within financial institutions. Foremost of these is data privacy, given that financial
transactions involve private financial information by its very nature. With greater use of big datasets to train predictive models
using machine learning, lenders must adhere to rigid data protection laws like the General Data Protection Regulation (GDPR) and
the California Consumer Privacy Act (CCPA), which impose restrictions on how consumer data can be accessed, held, and utilized.
The laws mandate that lenders maintain effective data governance processes that protect private individuals' information. Failing
to do so might have far-reaching legal implications, together with loss of consumer trust. The second most important challenge is
balancing predictive performance with the interpretability of predictive models. While complex predictive models can have good
predictive performance, predictive models are often "black boxes" that make it unreasonably complex or difficult to establish how
choices are reached. Such unexplained decision-making makes it problematic for lenders to defend choices to customers or
regulators. While building predictive models that improve predictive performance, lenders must ensure that predictive models
remain interpretable and accountable to ensure that customers have trust in predictive models together with compliance with
regulation requirements.

Limitations of the Study

This study is far from ideal, and its limitations can have consequences on its findings' generalizability and the applicability of its
developed models. A fundamental limitation is that of availability and representativeness of the datasets on which its models were
developed and validated. The use of datasets that are biased to specific demographics or locations can result in datasets that
poorly reflect the overall population of loan applicants. Such limitations can result in well-validated models within specific contexts
that poorly generalize to diverse populations or lending contexts. A second fundamental limitation is that of bias within its training
datasets that can have tremendous adverse consequences on the fairness of developed models. When historical lending datasets
reflect systemic bias—e.g., discriminatory lending to or within specific groups—these can perpetuated by and indeed amplified
through the use of machine learning-based modeling. The resulting models can, therefore, disadvantage specific loan applicants
or fail to accurately reflect diverse population-based risk profiles. Such limitations make it critical that well-rounded and
representative datasets are employed, along with actively seeking out and eliminating bias throughout the process of developing
models. Awareness of these limitations is critical to inform future research and ensure that AI-based loan risk assessments are
effective and fair.

Future Research Horizons

Looking ahead, numerous directions of future research can advance AI capability and AI-based loan risk assessment software. A
critical direction is investigating how to use deep learning techniques, which have shown much promise in numerous fields of
endeavor through their ability to learn complex nonlinear relationships within large datasets. Applying deep learning-based
programs to loan risk forecasting can potentially unlock richer insight into nuanced borrower behavior and determinants of risk
that might lie hidden in classical machine learning programs. A second compelling direction of research is building hybrid models
that connect financial and behavioral analytics. Such models can employ quantitative financial measures—e.g., credit ratings and
income levels—along with qualitative behavioral measures—e.g., spending habits and repayment history. By taking advantage of
each category of information, these hybrid models can provide richer insight into borrower risk, providing better predictive power
and better-informed lending decision-making. Lastly, there is room to research broader social consequences of these advanced
models, including ensuring fairness of operation along with maximizing predictive power. With AI continuing to evolve, bringing
innovative techniques and cross-disciplinary views will be important to building greater insight and use of AI-based loan risk
assessment, leading to better lending practices that are also fairer.

VIII. Conclusion

The primary objective of this research was to develop a machine learning model for loan application risk prediction that achieves
maximum reliability in decision-making while minimizing risks of default. This study focused on credit application risk assessment
in the context of the USA finance industry because challenges and opportunities in this industry are unique in their manner. The
dataset for this analysis comprises in-depth records of applicants for loans that exhibit a vast range of characteristics of borrowers,
credit history, and repayment behaviors. Comprehensive in scope, the rich dataset has variables that span age, earnings,
employment status, and locality alongside other crucial finance variables such as credit scores, debt-to-income ratio, and
repayment performance. For model selection, we utilized a variety of machine learning algorithms, including Logistic Regression,
Random Forest Classifier, and XG-Boost. The Random Forest and XG-Boost models closely aligned with actual data, showing high
accuracy. The integration of predictive modeling of advanced levels within loan decision processes has far-reaching consequences

Page | 175
Machine Learning-Based Risk Prediction Model for Loan Applications: Enhancing Decision-Making and Default Prevention

on building lender confidence within risk assessments. By using evidence-driven facts through machine learning models, lenders
can make better-informed decisions that better reflect greater insight into borrower behavior and attributes of risk. Looking ahead,
numerous directions of future research can advance AI capability and AI-based loan risk assessment software. A critical direction
is investigating how to use deep learning techniques, which have shown much promise in numerous fields of endeavor through
their ability to learn complex nonlinear relationships within large datasets.

Funding: This research received no external funding.

Conflicts of Interest: The authors declare no conflict of interest.
Publisher’s Note: All claims expressed in this article are solely those of the authors and do not necessarily represent those of
their affiliated organizations, or those of the publisher, the editors and the reviewers.

References

[1] Babaei, G., Pagnottoni, P., & Do, T. T. (2020). EXPLAINABLE MACHINE LEARNING FOR LENDING DEFAULT CLASSIFICATION. Chicago: Sage
Press
[2] Caffo, B. S., D'Asaro, F. A., Garcez, A., & Raffinetti, E. (2022). Explainable artificial intelligence models and methods in finance and
healthcare. Frontiers in Artificial Intelligence, 5, 970246.
[3] Chiamaka, A. (2023). Determinants of Credit Default in Banking in Nigeria. International Journal of Modern Risk Management, 1(2), 33-44.
[4] de Laat, P. B. (2017). Big data and algorithmic decision-making: can transparency restore accountability. Acm Sigcas Computers and
Society, 47(3), 39-53.
[5] Gupta, S. (2021). Impact of artificial intelligence on financial decision making: A qualitative study. Journal of Cardiovascular Disease
Research,, 12(6), 2130-2137.
[6] Heires, K. (2017). Flaws in the data: in an era of data-driven decision-making, businesses must avoid the faulty models and algorithms that
can poison their results. Risk Management, 64(3), 38-43.
[7] Hsu, T. C., Liou, S. T., Wang, Y. P., & Huang, Y. S. (2019, May). Enhanced recurrent neural network for combining static and dynamic features
for credit card default prediction. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
(pp. 1572-1576). IEEE.
[8] James, C. (2021). AI-Powered Tools for Automated Financial Decision-Making. New York. Harvard Press
[9] Jillson, E. (2021). Aiming for truth, fairness, and equity in your company’s use of AI. Federal Trade Commission, 19.
[10] Jui, A. H., Alam, S., Nasiruddin, M., Ahmed, A., Mohaimin, M. R., Rahman, M. K., ... & Akter, R. (2023). Understanding Negative Equity Trends
in US Housing Markets: A Machine Learning Approach to Predictive Analysis. Journal of Economics, Finance and Accounting Studies, 5(6),
99-120.
[11] Kumari, M. R., HMRITM, D., Gurpreet, I. M., Khanna, I. A., HMRITM, H., Miglani, I. R., & Singhal, I. R. MACHINE LEARNING MODEL FOR
HEART DISEASE PREDICITION USING TKINTER FOR GRAPHIC USER INTERFACE.
[12] Nasiruddin, M., Al Mukaddim, A., & Hider, M. A. (2023). Optimizing Renewable Energy Systems Using Artificial Intelligence: Enhancing
Efficiency and Sustainability. International Journal of Machine Learning Research in Cybersecurity and Artificial Intelligence, 14(1), 846-881.
[13] Piskunov, R., Moskalenko, O., & Krill, V. (2023). ANALYTICAL SUPPPORT FOR BANK LOAN TRANSACTIONS. Financial and credit systems:
prospects for development, 2(9), 18-24.
[14] Rana, M. S., Chouksey, A., Das, B. C., Reza, S. A., Chowdhury, M. S. R., Sizan, M. M. H., & Shawon, R. E. R. (2023). Evaluating the Effectiveness
of Different Machine Learning Models in Predicting Customer Churn in the USA. Journal of Business and Management Studies, 5(5), 267-
281.
[15] Spangher, A., Ustun, B., & Liu, Y. (2018, February). Actionable recourse in linear classification. In Proceedings of the 5th workshop on
fairness, accountability and transparency in machine learning.
[16] Sizan, M. M. H., Das, B. C., Shawon, R. E. R., Rana, M. S., Al Montaser, M. A., Chouksey, A., & Pant, L. (2023). AI-Enhanced Stock Market
Prediction: Evaluating Machine Learning Models for Financial Forecasting in the USA. Journal of Business and Management Studies, 5(4),
152-166.
[17] Urban, S., Moosbrugger, N., Hellwig, M., & Dobler, M. (2022). Sustainable smart services for financial institutions: A framework to support
banks in ESG integration in the risk analysis process of loans. In Smart Services Summit (pp. 143-151). Cham: Springer Nature Switzerland.

Page | 176