The Japanese Society for Artificial Intelligence

Predicting loan defaults is an important task in peer-to-peer (P2P) lending platforms to mitigate financial risks and ensure platform sustainability. Traditional methods rely heavily on structured data such as credit scores, income, and loan amounts. While recent studies also have considered integration of textual data from applications, they overlook the latent link between the textual narratives and structured data. This study focuses on the discrepancy—the misalignment between a borrower's actual financial state (captured by structured data) and their self-reported or perceived circumstances (reflected in textual narratives). Borrowers exhibiting poor self-awareness, overconfidence, or dishonesty are more likely to create such discrepancies, making them prone to default. This study proposes a novel framework to extract and quantify these discrepancies by analyzing borrower-submitted narratives alongside structured financial data. Our framework demonstrates that incorporating text-data discrepancies enhances AUC performance, providing a new perspective for risk assessment. Our model achieves performance comparable to traditional soft-information extraction methods while delivering superior interpretability through its explainable, discrepancy-driven framework.