Understanding Firm Environmental Violations: A Theory‐Informed Machine Learning Study
TL;DR
Using dataset of Chinese A‐share listed companies from 2010 to 2022, a nonlinear prediction framework is developed and machine learning models significantly outperform traditional linear approaches, with Random Forest delivering the best performance.
Abstract
This study investigates the risk structure of firm environmental violations by integrating the GONE framework—Need, Opportunity, Greed, and Exposure—with explainable machine learning methods. Using dataset of Chinese A‐share listed companies from 2010 to 2022, we develop a nonlinear prediction framework and show that machine learning models significantly outperform traditional linear approaches, with Random Forest delivering the best performance. SHAP‐based decomposition reveals a clear NOEG ordering in the importance of risk dimensions, indicating that environmental violation risk prediction primarily relies on information capturing firms' operational pressures and resource constraints. This information structure remains stable across a series of robustness checks. Further nonlinear analysis shows that key risk drivers exhibit pronounced threshold and nonlinear effects. Heterogeneity analysis further demonstrates that the information structure of environmental violation risk undergoes reconfiguration across ownership types and industry pollution intensity, revealing the strong context dependence of violation risk mechanisms. This study provides theoretical support for understanding and preventing firm environmental violations, while expanding the practical application of integrating theory with interpretable machine learning.