Predicting Loan Default Using Deep Learning model: A Case Study of Hibret Bank

Abstract
In recent years, financial institutions have faced increasing challenges in credit risk management, where accurate loan default prediction is crucial for stability and profitability. Because bank revenues depend on loan repayments, reducing defaults is essential to sustain performance. This study explores the use of advanced deep learning and hybrid models for predicting loan defaults in Hibret Bank, aiming to enhance credit risk assessment and decision-making. We developed five algorithms: three deep learning models (TabNet, 1DCNN, and LSTM) and two hybrid models (XGBoost+LSTM ensemble stacking and 1D-CNN+LSTM). The study utilized 43,467 loan records, with preprocessing including min–max and robust scaling for numerical features, label and target encoding for categorical attributes, and feature selection with embedding extraction using TabNet attention masks. To address class imbalance, resampling techniques such as SMOTE, ADASYN, SVMSMOTE, and SMOTEENN were applied, with SMOTEENN providing the most balanced. Unlike previous Ethiopian studies that relied primarily on traditional machine learning or single deep learning models, our approach demonstrated significant improvements. Among all models, the XGBoost+LSTM ensemble stacking achieved the best performance with an accuracy of 98.2% and recall of 91.2% for the default class, demonstrating strong detection of high risk borrowers. These results highlight that hybrid ensemble stacking algorithms offer a practical and effective solution for predicting loan defaults, contributing to improved risk management practices in the Ethiopian banking sector.