Model Construction of Consumer Credit Ratings Based on Nonparametric Statistical Testing and Machine Learning Fusion

Abstract

With the rapid growth of consumer credit, there is a need to assess it. Given the limitations of traditional credit scoring methods, which rely on linear assumptions or lack interpretability, and artificial intelligence (AI) methods, more robust models are needed. Against this backdrop, this study proposes a novel consumer credit rating model that integrates nonparametric statistical testing with machine learning to increase accuracy and interpretability. The research focuses on screening statistically significant features via Kolmogorov‒Smirnov tests (e.g., D=0.28, p=0.002 for annual income) and Mann‒Whitney U tests (U=210, p=0.008) while addressing multicollinearity via Spearman rank correlation. A fusion framework combining the Cox proportional hazards model and conformal prediction is employed to generate confidence intervals for default probability. The results demonstrate the model’s superior performance, achieving an area under the curve (AUC) of 0.775, higher recall (0.72), and lower Brier score (0.158) than logistic regression and survival forest. The study bridges the gap between predictive accuracy and interpretability, offering a reliable tool for financial institutions to assess credit risk.