Comparative Study of Personal Credit Default Risk Prediction Based on Different Machine Learning Models

School of Statistics and Data Science, Capital University of Economics and Business, 100070 Beijing, China

^* Corresponding author: lihaonan@arizona.edu

Abstract

The accurate prediction of individual credit default risk is crucial for maintaining the stability of the financial system. This study aims to establish a systematic comparative framework to evaluate the performance of five representative machine learning models in the task of predicting individual credit default risk. The research first preprocesses a public dataset containing 150,000 customer records and applies SMOTE technology to address the inherent class imbalance problem. Subsequently, the five models are trained under the same benchmark and evaluated based on multiple dimensions including accuracy, precision, recall, F1-Score, and AUC-ROC. The experimental results indicate that ensemble learning models (especially random forests and LightGBM) significantly outperform traditional models. Among them, random forests show the best performance in recall and AUC- ROC, effectively identifying potential defaulting customers; while LightGBM achieves the optimal results in F1-Score and precision. Additionally, feature importance analysis reveals that the utilization rate of unsecured credit limits and the number of family members are key factors affecting credit default risk. The conclusions of this study provide valuable guidance for financial institutions in selecting the most appropriate credit risk prediction model based on different business needs, and offer practical references for future related research.