A Deep Learning-based Model for P2P Microloan Default Risk Prediction

Abstract

This study presents a new deep-learning model for predicting default risk in peer-to-peer (P2P) microlending platforms. The model integrates convolutional neural networks (CNNs) and short-term (LSTM) networks to capture both spatial and temporal patterns in lending data. An extensive database including 150,000 loan records from a major P2P platform was used, including 78 characteristics related to borrowers, loan characteristics, and platform-specific metrics. The model uses a hybrid selection method that combines filtering and wrapping methods to identify the most relevant parameters. An ensemble learning strategy is implemented, combining deep learning models with gradient boosting and random forest classifiers. The experimental results show the best model performance, achieving an accuracy of 92.34% and an AUC-ROC of 0.9687, outperforming the scoring model and the machine learning model. Factor analysis shows that annual income, debt-to-income ratio, and credit score are the most important factors in predicting bad credit. This study provides insight into the interpretation of the SHAP and LIME criteria, improving transparency in credit risk assessment. The findings have important implications for P2P lending platforms and investors, providing better risk management strategies and more informed decision-making capabilities in microloan evaluation.