An intelligent soft–voting based early warning model for loan risk

Abstract

: Against the backdrop of fintech driven intelligent risk control, traditional manual loan review, hindered by low efficiency and high subjectivity, struggles to meet the massive demand for credit risk assessment. This paper proposes a multi–model fusion framework based on soft–voting, integrating Feedforward Neural Network (FNN), LightGBM, and CatBoost, and combining with scorecard technology to achieve precise prediction of loan default risks and their operational implementation. Using more than 1 million loan records from Alibaba Cloud Tianchi, the study employs data cleaning, clustering binning (10 bins), standardization, and construction of 10 composite features (e.g., ratio of income to loan amount). The core features are selected using chi–square test (p), Population Stability Index (PSI), and Information Value (IV). Through exhaustive experiments, the soft–voting mechanism determines optimal weights (0.1:0.1:0.8), allowing the integrated model to achieve an AUC of 0.739, significantly improving the generalization while maintaining high accuracy. The scorecard model converts probabilities into credit scores, classifying customers into four tiers (A to D). This study transcends the limitations of single models, achieving deep integration of technical interpretability and business implementation through the dual mechanism of model fusion and scorecard. It provides financial institutions with end–to– end solutions from risk quantification to strategy execution, facilitating dynamic monitoring and management of in–lending behavioral risks.