The Role of Country- and Firm-Level Factors in Determining Firms’ Environmental, Social, and Governance (ESG) Performance: A Machine Learning Approach

The objective of this study is to apply machine learning techniques to predict firms’ environmental, social, and governance (ESG) performance. We employed ten supervised machine learning models—decision tree, stochastic gradient descent, random forest, adaptive boosting, extra trees, extreme gradient boosting, k-nearest neighbors, multiple linear regression, transformer-based regression, and multi-layer perceptron—and evaluated their effectiveness in ESG prediction. The analysis covers global firms and incorporates seven firm-specific characteristics and country-level macroeconomic variables. Using a sample of 19,476 observations from 2010 to 2018 from publicly listed firms across 53 countries, our findings indicate that random forest regressor exhibit the highest predictive power, with the lowest root mean square error. For the random forests regressor, the coefficient of determination (R2) was 30% and the mean absolute error (MAE) was 1.52. The second-best predictive model was the extra-trees regressor with an R2 value of 27% and an MAE of 1.56. In our analysis, the multiple regression model had the lowest R2 value of 14%, with an MAE value of 1.70. We also analyzed the feature importance of the random forests regressor, identifying firm size, capital expenditure, and cash holdings as the top three predictors of ESG performance. Additionally, we used the double-debiased machine learning technique to assess causal inference in feature importance, revealing capital expenditure, political stability, and leverage as the most influential factors in ESG predictions. Although the ranking of feature importance varies between the causal inference and predictive models, the key factors identified by both techniques largely align and are consistent with the significance of the coefficients from the multiple regression model.