Profit-oriented loan default prediction for the financial industry: a fusion framework with interpretability

Abstract Loan default risk prediction has become crucial for financial institutions because it can help decision-makers successfully distinguish good from bad loan applicants and earn potential economic benefits. The core of default prediction is the construction of a well-performing default prediction model with good interpretability. Inspired by this need, a novel fusion prediction framework with interpretability is proposed, integrating extreme gradient boosting (XGBT) into the construction of a random forest (RF) to reduce prediction variance and bias. This model extends the original bootstrap-aggregation framework of RF to a novel bootstrap-boosting (XGBT)-aggregation framework, effectively combining the strengths of both models and addressing the bias‒variance tradeoff in machine learning. Several real-world credit datasets from the financial industry are leveraged to assess the prediction performance of the proposed fusion framework. The experimental results reveal that the proposed framework significantly outperforms alternative models, showing notable improvements in both default prediction accuracy and profitability. The interpretability analysis highlights the key features that significantly influence the prediction results of the proposed fusion framework, which helps practitioners better identify defaulters and make more profitable decisions.