Research on Feature Selection Method for Enterprise Financial Crisis Early Warning Based on LRRF-MIC Integrated Screening Approach

In capital markets, Special Treatment (ST) designation for listed companies stems from multiple complex factors. To enable early ST risk identification, this study explores integrated machine learning methods combining feature selection and predictive models, focusing on comparative evaluation of feature selection techniques through empirical analysis. Using a sample of 430 Chinese A-share listed firms with financial/non-financial indicators, 35 key indicators distinguishing distressed vs. healthy firms are first identified via statistical screening.We propose an innovative "LRRF-MIC Integrated Screening Method" integrating Lasso regression, Recursive Feature Elimination (RFE), Random Forest (RF), and Maximal Information Coefficient (MIC). This hybrid framework generates multi-dimensional feature evaluations and visual analytics, systematically selecting 14 core predictive indicators based on established criteria. Empirical validation uses these features as inputs for MLP and gcForest models, compared with benchmark models using raw data. Results show the LRRF-MIC framework outperforms single methods (RF, RFE, MIC, Lasso) by over 8% in prediction accuracy on average and benchmark models by 13%, demonstrating the efficacy and innovation of the proposed integrated approach.