Can machine learning uncover ESG alpha in the Chinese A-share market? An 'alpha illusion' case study

This study investigates whether Environmental, Social, and Governance (ESG) factors can generate alpha in the Chinese A-share market, a market characterized by nascent ESG integration and distinct investor behavior. Using data from 2010 to 2023, we find that raw ESG scores fail to predict stock returns. In contrast, an ESG factor constructed using machine learning (ML), particularly XGBoost, exhibits a highly significant Fama-MacBeth coefficient of 2.122 (t = 11.98). However, this seemingly robust alpha is an illusion. Interpretability analysis reveals that the ML model’s predictive power stems almost entirely from its non-linear capture of traditional factors like illiquidity, size, and value, with the ESG signal contributing virtually nothing. The strategy's profitability relies on high-frequency trading and short-selling, making it unfeasible in practice. Causal inference based on China's carbon market launch confirms that policy shocks failed to activate a genuine ESG pricing mechanism. We conclude that the "ESG alpha" identified by ML is a repackaging of known anomalies—an "alpha illusion"—highlighting the critical distinction between statistical significance and true economic value.