Incorporating Macroeconomic Indicators to Improve P2P Lending Credit Risk Modeling: A Machine Learning-Based Heterogeneity Analysis

Abstract

This study investigates the improvement effect of macroeconomic indicators on credit risk modeling in the P2P lending market, focusing on near-prime borrower groups and short-term loan segments. Based on 171,644 loans from Lending Club platform during 2012-2013, this paper compares the performance of four machine learning algorithms (logistic regression, random forest, XGBoost, and LightGBM) under baseline feature sets (borrower/loan characteristics only) and enhanced feature sets (including regional macroeconomic indicators). Empirical results demonstrate: (1) Logistic regression achieves optimal performance-stability balance after integrating macroeconomic indicators, with AUC improving 1.08% to 0.6854, KS statistic increasing 3.82%, while maintaining 1.0% overfitting degree; (2) Model performance improvement exhibits significant heterogeneity, with 36-month short-term loans achieving 1.41% AUC improvement, and near-prime borrowers (FICO 620-680) showing 2-3 times higher sensitivity to unemployment rate fluctuations compared to other groups; (3) SHAP analysis reveals macroeconomic indicators contribute 3%-5% to predictive power, with interaction terms between macro and baseline features significantly outperforming single macro indicators. The study provides empirical support for P2P platforms to construct group-specific risk pricing models.