ESG: All Bark and No Bite? Exploring the utility of environmental, social and governance variables in empirical asset pricing via machine learning

Abstract

In this thesis I investigate the impact of including environmental, social and governance (ESG) variables in explaining the cross section of expected stock returns. Using three machine learning frameworks applied to a broad dataset of firm characteristics, macroeconomic predictors and ESG-related variables, I find that ESG contributes to a small but statistically significant increase in explanatory power. The governance category appears to be most important, followed by the environmental category. The social category is not found to contribute significant explanatory power, but does impact predicted excess returns comparably to the other categories. Governance variables contribute to a 4.54% increase in out-of-sample R2 on average, whilst environmental variables contribute to a 1.44% increase. Including all ESG variables increases explanatory power by around 3.87% on average, but results are highly dependent on model selection, with some models yielding as much as 13.22%. Large firms experience the biggest increase in explanatory power from the inclusion of ESG variables. Finally, I expand on some recent findings in the literature such as the risk premium for CO2 emissions. Using neural network bivariate marginal effects, I find that premiums for younger firms are steeper and more sensitive to CO2 intensity.