Evaluating the Predictive Effectiveness of Digital Credit Scoring Compared with Traditional Credit Assessment

Abstract

This paper is an in-depth empirical analysis of the digital credit scoring processes versus traditional credit assessment processes in the fast-growing fintech landscape in China. Based on the 6 credit scoring models, which are based on traditional statistical algorithms (logistic regression, linear discriminant analysis, PBOC-based scoring), and sophisticated machine learning algorithms (Random Forest, XGBoost, Deep Neural Networks), we systematically analyses 6 Chinese credit datasets with 95,000 observations of peer-to-peer lending platforms, commercial banks, and digital financial services providers. The findings indicate that digital models based on alternative data of Alipay, online shopping websites, and social networks are significantly better at all metrics of evaluation than the traditional methods XGBoost has an AUC-ROC value of 0.932 over the logistic regression AUC-ROC value of 0.748, which is a 24.6 percent increase in predictive power. The results indicate that credit scoring using machine learning can boost default prediction rates by 42 percent of thin-file borrowers, as well as increase financial inclusion among the heterogeneous urban and rural population of China. This study will offer evidence-based information to Chinese financial institutions, and fintech platforms that are undergoing the transformation of the traditional systems based on the PBOC to holistic digital credit evaluation systems in accordance with the regulatory environment in China.