DEVELOPING LOAN DEFAULT PREDICTION MODEL USING MACHINE LEARNING TECHNIQUES

Abstract

Loan defaults pose a significant risk to financial institution, leading to substantial financial losses and impacting their stability and profitability. Existing predictive models often overlook key borrower characteristics, resulting in less accurate predictions. This study aims to improve loan default prediction by integrating borrower-specific features and loan characteristics using a blending ensemble model. Specifically, we focus on borrower characteristics such as business location, loan product type, yearly business income, location of collateral, total years of experience, and educational status, which are used by some Ethiopian banks for risk assessment but have been underexplored in previous studies. We employ three base models: logistic regression, multilayer perceptron, and random forest. These models are combined using a weighted average blending ensemble approach to enhance predictive performance. The dataset, consisting of 18,184 records from a single bank, was split using an 70/30 ratio for training and testing. Our findings demonstrate that the blending ensemble model outperform individual base models in predicting loan defaults, achieving higher accuracy (98.62%), precision, recall, and F1-score. The most significance predictors identified includes sex, collected total, educational status, employment status, and age, while gender and marital status shower lesser impact. This study contributes to the field by providing a more robust predictive model that incorporates underexplored borrower characteristics, offering financial institutions a more accurate tool for risk assessment and decision-making