Machine-Learning-Guided Control Selection in ESG–Performance Models: Evidence from a Global Firm Panel

This study examines the relationship between environmental, social, and governance (ESG) performance and firm financial performance while considering the moderating role of institutional quality and the sensitivity of the results to control-variable selection. The analysis uses 68,244 firm-year observations for 6204 listed firms across developed and developing economies between 2011 and 2021. Firm performance is measured using Tobin’s Q, return on assets (ROA), and return on equity (ROE), while ESG is examined through its aggregate score and its environmental, social, and governance pillars. We use Random Forest to rank candidate controls, and we combine its rankings with multicollinearity diagnostics. Conventional full-control and Random Forest-informed specifications are then compared using OLS and quantile regressions. The results show that control selection materially affects the magnitude, direction, and statistical significance of the estimated ESG coefficients. Although there are clear relationships between ESG and performance variables in the baseline model, they become statistically insignificant when using the firm-fixed effects models, implying that the findings arise mainly from the persistent difference between firms rather than within-firm ESG performance. In addition, market responses are more selective: the Social pillar is positively associated with Tobin’s Q, whereas the Environmental pillar is negatively valued in several specifications. Institutional moderation also varies across ESG pillars, governance dimensions, performance measures, and quantiles. Government Effectiveness and Rule of Law provide the most consistent evidence of positive moderation for the aggregate ESG–Tobin’s Q relationship, while the evidence is weaker or less stable for ROA and ROE. Our findings demonstrate that the financial relevance of ESG is context-dependent and sensitive to empirical model specification.