A loan default refers to the failure of a borrower to fulfill the agreed repayment obligations within the stipulated

schedule, either partially or in full.

Defaults may occur due to various factors, including financial instability,

unemployment, excessive debt burden, unexpected economic conditions, or poor credit behavior. From the lender’s

perspective, loan defaults represent a significant source of financial loss and increased operational risk, as they directly

impact asset quality, liquidity, and profitability.

Consequently, accurately identifying high-risk borrowers prior to

(ML), a subfield of artificial intelligence, has the capability to efficiently handle large volumes of data and extract

meaningful patterns from complex datasets. Algorithms belonging to the same or similar domains present two major

assessment challenges. The first challenge involves comparative evaluation among algorithms applied to the same

dataset within a single study, while the second challenge arises when comparing outcomes reported by different authors

across related studies. The present study addresses both evaluation perspectives, as discussed in prior works.

This study adopts a classification-based approach using decision tree–based ensemble models to support accurate and

timely prediction of borrowers’ likelihood of loan default. Loan default prediction is widely recognized as a critical

and risk-sensitive task for financial institutions, as it directly influences lending decisions and risk mitigation strategies.

scoring models, and rule-based decision systems. While these methods are effective for modeling linear relationships,

they exhibit limited capability in capturing complex non-linear patterns present in large-scale financial datasets.

Predicting loan defaults in such datasets is a challenging task due to the presence of multiple interacting non-linear

features, a concern relevant across various industries, including FMCG and automation sectors.

A loan default typically occurs when a borrower fails to meet scheduled repayment obligations, resulting in financial

distress for the lender. The ability to identify potential defaults at an early stage enables financial institutions to

proactively manage and mitigate future risks. Consequently, evaluating and assessing the performance of machine

learning algorithms in this domain has become a significant area of both academic research and practical application.

Performance assessment focuses on determining how effectively models can distinguish between potential defaulters

Gradient Boosting Machines, XGBoost, and LightGBM.

These ensemble models exhibit superior performance

due to their ability to capture complex, multi-level interactions among financial attributes such as credit history, income

stability, debt-to-income ratio, and loan characteristics.

Rendering to the research since 2019 to 2025, ensemble or collaborative machine learning approaches mainly Gradient

Boosting, Random Forest and XGBoost, LightGBM, consistently outstrip the classical statistical technique methods

for loan default prediction. Reviews in similar vicinities which supports the study are as follows:

anticipated prejudiced logistic regression with L2 penalty dataset and TF-IDF features on Chinese credit

analysed with the Lending Club dataset by utilizing Random Forest, SVM, and

Logistic Regression algorithm. The Random Forest algorithms performed best in association of SMOTE improved

philological model on P2P loans (2007–2020), attaining high accuracy and extra interpretability for officials and initial

default exposure. Luca Barbaglia et al.

outstripped logistic regression. The variables Interest rate, LTV, and local economy were considered as the key

The study presented cost-sensitive threshold fine-tuning for directive. Kang et al.

(2025) considered the Kaggle loan

data (148k records) and process the same with Random Forest, XGBoost and LightGBM models, utilizing SMOTE

for balance the dataset. It is observed that LightGBM achieved best (Acc 0.9764, Prec 0.9747, Rec 0.9503). The

significant features remained interest rate and credit type in association of target variable

boosting, XGBoost and LightGBM) for forecasting loan defaults.

performance as:

H1: The machine learning algorithms are significantly performing to deal the prediction of the loan defaults.

Fig. 1 Illustrate the research methodology. The experimental methodology of the research includes five main segments.

The first step is concern from the assortment of the proper dataset. It is necessity of dataset that appropriate credit and

demographic data should be available. The second stage is concern from the Data Pre-processing; the key deliberation

is treatment of missing values, data encoding, normalization and dealing the class imbalances. Feature engineering and

selection with REF, correlation analysis and valuations of tree analysis. The machine training process will be complete

with 80:20 ratio. At this time 80% dataset is used to train the machine and 20% of the dataset endure for the testing.

Subsequently train the machine several machine learning algorithm / models are used for the experimental. Here we

are considering Random Forest, Gradient Boosting Machine, XGBoost and LightGBM model od machine learning for

the prediction of loan defaults. The performance of the model is assessed by metrics of Confusion matrix, accuracy,

precision, recall, F1-score, ROC AUC and feature performance.

www.kaggle.com/datasets/taweilo/loan approval-classification-data, retrieved on 10 October 2025. It comprises about

45,000 records and 14 variables, which includes both numerical and categorical features. It comprises client and loan-

related features which are related to forecasting loan defaults.

Table 1: Key variables used for modelling.

the entire features are on a similar scale and enhance convergence for algorithms alike Random Forest, Logistic

Regression, Gradient Boosting and neural networks.

procedure established by Microsoft. It is an enactment of the Gradient Boosting Decision Tree (GBDT) outline,

intended to deliver rapid training speediness, lesser memory ingesting, and well correctness, chiefly for large-scale

and high-dimensional datasets. In R, LightGBM is obtainable by the package LightGBM. which permits handlers to

execute classification, regression, and ranking jobs competently.

LightGBM performs by construct an ensemble of puny learners, characteristically decision trees, in a consecutive way.

Every novel tree is trained to accurate the errors produced by the preceding ensemble of trees by diminishing a stated

loss function. Dissimilar to old-style gradient boosting, LightGBM usages a leaf-wise growing tactic in its place of

level-wise development. In such method, the algorithm cultivates the tree by excruciating the leaf through the

uppermost loss lessening, follow-on in profounder trees and enhanced accurateness. Though, to avoid overfitting, the

limitation max_depth might be established to bound tree deepness.

The IDE of R is known as the R studio, is an environment which implement the capabilities of R language at the single

interface. It gives a user-friendly environment and interface with facilities of coding, reporting and visualization. Also,

the library and package are comprising according to machine learning algorithm implemented in R. In R Random

Forest uses “randomForest” and “ranger”, for Gradient Boosting “gbm” and “caret”, for XGBoost library is “xgboost”

and for LightGBM library is “lightgbm”.

classifications with actual and predicted form. It also provides a thorough breakdown of model performance. Fig. 4 is

depicting binary classification (2 × 2) matrix arrangement.

The key advantage of the confusion matrix is it gives inside of values with types of error like False positive (FP) and

recall.

Accuracy: It is basically the proportion of appropriately predicted cases to the total cases. It measures in what way

frequently the model is accurate all-inclusive.

F1 Score: It is the harmonic mean of Precision and Recall, which equilibriums the trade-off among them.

Recall: It is also known as sensitivity or true positive rate. It assesses the ability of model to accurately recognize all

positive cases.

(4)

ROC Curve: It is denoted as “receiver operating characteristics” curve, designs the true Positive Rate “Recall” in

contradiction of the False Positive Rate “FPR” at diverse classification beginnings.

(5)

The confusion metrics is utilised to create ground for the measurement of performance of the machine learning models.

The models are assessed on the testing dataset and are presented below to deliver a comprehensive detail of their

classification performance. The matrices are vital for analysis of classification errors, exactly particularization of the

number of true positives, true negatives, false positives, and false negatives. This is vital to know how well every

According to the values of is TN = 6820 which accurately predicted non-defaulters, FP = 179 non-defaulters forecast

as defaulters, FN = 455 defaulters forecast as non-defaulters and TP = 1546 accurately forecast defaulters (Fig. 5).

The accuracy is approximately 93% which shows that model is accurate for maximum predictions, but it might be

misleading when dataset is imbalanced. The precision approximately 89.6% which shows that maximum people

forecast as defaulters are really defaulters. Thus now 10% of forecasted defaulters are false positives. The recall

The values received from confusion matrix are as the values of is TN = 6800 which accurately predicted non-defaulters,

FP = 200 non-defaulters forecast as defaulters, FN = 463 defaulters forecast as non-defaulters and TP = 1537 accurately

forecast defaulters (Fig. 6).

The Gradient Boosting Machine model achieves an accuracy of approximately 92.6%, indicating strong overall

predictive performance. However, due to class imbalance, accuracy alone may be misleading. The precision of 88.5%

suggests that most customers predicted as defaulters are indeed defaulters, with around 11% false positives, which

may result in lost business opportunities or rejected loan applications. The recall of 76.8% indicates that the model

correctly identifies nearly 77% of actual defaulters, while 23% (FN = 463) are misclassified as non-defaulters, posing

a potential financial risk to the bank. Additionally, 200 false positives represent non-defaulters incorrectly classified

as defaulters.

Fig. 6: Confusion matrix for Gradient Boosting machine algorithm.

The model achieves an accuracy of approximately 93.2%, indicating strong overall predictive performance. The

precision of about approximately suggests that most customers predicted as defaulters are indeed defaulters, with

nearly 11% false positives, which may result in lost business opportunities or rejected loan applications. The recall of

approximately 79% model accurately recognizes approximately 79% of real defaulters so in this case 21% of defaulters

are unexploited (FN = 421). False Negatives (421) means these are actual defaulters which predicted as non-defaulters

which considered as “risk to the bank”. False Positives (195) means these are non-defaulters forecasted as defaulters,

possible lost business or disallowed loan applications.

approximately 88.8% shows that maximum people forecast as defaulters are really defaulters. Thus now 11% of

forecasted defaulters are false positives. The recall approximately 79.8%, shows model accurately recognizes

approximately 80% of real defaulters thus 20% of defaulters are misclassified as non-defaulters (FN = 404). False

Negatives (404) means these are actual defaulters which predicted as non-defaulters as “risk to the bank”. False

Positives (201) means these are non-defaulters forecasted as defaulters, possible lost business or disallowed loan

applications.

Accuracy: Considered four models achieved admirable accuracy, which displays vigorous whole predictive

competence. The model LightGBM has been noted with highest accuracy value as 0.933, trailed meticulously by model

XGBoost (0.932), model Random Forest (0.930), and model Gradient Boosting (0.926). These protests the ensemble

tree-based representations are enormously capable for loan default forecasting.

F1-Score: The F1-Score equipoises precision and recall, mostly important for excessive datasets. At this time the model

Precision and Recall: the precision assess correctly forecast defaults out of whole forecast defaults; however, recall

assesses correctly forecast defaults out of actual defaults. The model LightGBM achieved the highest recall (0.798),

as it identifies as extensively held of actual defaulters, while model Random Forest had the highest precision (0.896),

observing few false positives. The model XGBoost provides a steady performance by precision value (0.890) and

recall value (0.789).

From the performance metrics, it is observed that that entire models’ performances are outstandingly on the utilised

dataset. Even the consequence of the individual models is very good, thus hypothesis for the study is accepted. It is

noted that LightGBM model somewhat outperformed other models in almost each metrics. The model LightGBM