Predicting Default Risk in Buy Now, Pay Later Services: A Comparative Study of Random Forest and XGBoost

Buy Now, Pay Later (BNPL) has emerged as one of the fastest-growing segments of the global financial technology (FinTech) industry, offering consumers flexible payment options while creating new challenges in credit risk management. The rapid expansion of BNPL services has increased the likelihood of borrower defaults, making accurate and timely default prediction essential for maintaining portfolio quality and supporting sustainable lending practices. This study investigates the application of supervised machine learning techniques for binary classification to predict whether BNPL users are likely to default or successfully repay their obligations. Two widely used ensemble learning algorithms, Random Forest and XGBoost, were developed and compared using a dataset comprising 10,345 BNPL transactions.
Model performance was evaluated using four key metrics: Accuracy, Area Under the Receiver Operating Characteristic Curve (AUC-ROC), Recall, and the F1-score. Experimental results indicate that the Random Forest model consistently outperformed the XGBoost model, achieving an overall accuracy of 97% while demonstrating superior discriminative capability in distinguishing between defaulting and non-defaulting customers. The findings highlight the potential of machine learning methods to strengthen credit risk assessment by providing more accurate and reliable default predictions. Consequently, these predictive models can support financial technology institutions in enhancing lending decisions, minimizing credit losses, optimizing risk management strategies, and promoting responsible growth within the rapidly expanding BNPL ecosystem.