Domain-Informed Default Risk Prediction: Interpretable Features and Model Performance Analysis

The rapid development of peer-to-peer lending platforms has provided a more efficient financing channel for the consumer credit market. However, limited borrower information and the regulatory environment increase the difficulty of predicting default risks. In the problem of default prediction, there are usually challenges such as sample imbalance and how to select appropriate models. In the field of credit default prediction, existing research mostly focuses on model selection and model optimization, while less attention is paid to improving model performance by integrating financial domain knowledge to construct interpretable features. Therefore, this paper proposes three new features - DTI_Interest (debt-to-income ratio for interest), stability score, and loan density, which respectively describe repayment pressure, financial stability, and borrowing intensity. Experiments based on five different models show that although the new features slightly improve the prediction performance, this improvement is not statistically significant. This study provides new ideas for P2P risk feature engineering and highlights the potential value of domain knowledge in credit assessment.