Optimizing Profit Scoring in P2P Lending Using Prepayment and IRR Analysis

Peer-to-peer (P2P) lending platforms face significant challenges in accurately assessing borrower risk and predicting default probability due to the complex nature of lending data characterized by high dimensionality, severe class imbalance, and intricate relational dependencies. This work develops a comprehensive machine learning pipeline for credit and profit scoring in P2P lending environments. A critical innovation of this work is the identification and incorporation of prepayments as a distinct class, addressing the financial impact of early loan settlements on expected interest returns. We conduct a systematic comparison of default prediction performance between binary classification (default vs. non-default) and multi-class classification (default vs. non-default vs. prepaid) frameworks using three machine learning algorithms: Logistic Regression, Random Forest, and XGBoost. To evaluate the practical financial implications of misclassification errors, we implement Internal Rate of Return (IRR) analysis through simulated cash flow modeling. This approach quantifies the actual monetary impact of prediction errors on lending profitability. The findings demonstrate that misclassification costs are substantially higher in binary classification models compared to multi-class approaches, suggesting that incorporating prepayment prediction significantly improves both predictive accuracy and financial outcomes.