XGBOOST MODEL FOR DEFAULT PREDICTION IN CREDIT SCORING OF CONVENTIONAL BANK

Abstract

We develop an XGBoost-based classification model for predicting loan default in the context of credit scoring for a conventional commercial bank in Indonesia. The model aims to improve predictive performance in identifying high-risk borrowers using historical loan data. For this purpose, we use two years of consumer loan records, consisting of over forty thousand observations, including borrower demographic, credit score, and loan characteristics. To address the severe class imbalance within the dataset, we employ the Random OverSampling Examples (ROSE) technique on the training subset. The model is trained and evaluated using standard classification performance metrics, including precision, recall, F1 score, and area under both ROC and Precision-Recall curves. Our results show that the XGBoost model performs well in detecting non-defaults with high sensitivity and precision, particularly in the training set. However, performance on the test set indicates a significant drop in recall for the default class, suggesting model overfitting under imbalanced conditions. These findings highlight the potential and limitations of using ensemble learning methods such as XGBoost in real-world credit risk evaluation, especially when data imbalance remains a major concern.