Enhancing Credit Risk Assessment in Loan Approval: Performance Evaluation of Machine Learning Models

Abstract

Effective credit risk assessment is paramount for the financial stability of lending institutions. Traditional methods often suffer from inefficiency and reliance on manual judgment. This research addresses these limitations by developing and evaluating three distinct machine learning models—Logistic Regression, Random Forest, and XGBoost—for loan approval prediction using a synthesized credit dataset. We detail the entire methodology, including comprehensive data preprocessing steps (encoding, normalization, and outlier capping) and hyperparameter tuning. Performance is rigorously evaluated based on Accuracy, F1-score, and Area Under the Curve (AUC). Our results demonstrate that the XGBoost classifier achieves superior predictive performance, with an Accuracy of 0.9187 and an F1-score of 0.8636, significantly outperforming the other models. The findings validate the use of gradient boosting techniques for automating and minimizing credit risk in the loan approval pipeline.