Abstract Digital lending and alternative finance systems, particularly Buy Now, Pay Later (BNPL) services, have expanded access to credit but introduced new challenges for accurate and transparent credit risk assessment. Borrowers in these ecosystems often lack extensive financial histories, making it difficult for traditional scoring models to capture their financial, behavioral, and socioeconomic characteristics. In addition, many existing machine learning approaches operate as black-box models, limiting interpretability and raising concerns regarding regulatory compliance and trust. This study proposes an optimization-driven hybrid machine learning framework that integrates gradient boosting models with nature-inspired metaheuristic optimization to enhance both predictive performance and interpretability in credit risk assessment. The proposed approach incorporates systematic data preprocessing, handling of class imbalance, and feature engineering to extract meaningful patterns from a publicly available dataset of 1,000 loan applications with 16 predictive attributes. Hyperparameters of the predictive models are optimized through iterative refinement, enabling efficient exploration of the search space and improved generalization. To ensure transparency, the framework provides feature-level explanations that identify the most influential variables contributing to default prediction. The model is evaluated using multiple performance metrics, demonstrating improved stability and predictive capability across cross-validation folds. Unlike conventional black-box approaches, the proposed framework balances accuracy with interpretability, making it suitable for deployment in regulated financial environments. The findings demonstrate that combining gradient boosting with optimization techniques yields a robust and explainable solution for credit risk prediction. This study contributes to the advancement of interpretable artificial intelligence in digital lending by offering a practical and transparent modeling framework that supports reliable decision-making.
Mohammad et al. (Wed,) studied this question.