A Sneak Peek into Machine Learning Methods for ESG Factor Score Computation

Machine learning models, when applied to Environmental, Social, and Governance (ESG) data, can serve as an essential guide in assessing various aspects of sustainability for investing, financing, insurance, and even policymaking. They can be used for computing ESG thematic scores, carbon scores, water scores, etc. This chapter provides a “sneak peek” into some of the challenges related to the application of machine learning methods to compute an ESG thematic score based on a set of ESG parameters for a given entity. We start with a general presentation of how ESG thematic scores are computed, the data that we use, and the preprocessing steps that we suggest applying to the underlying data. Then, we explore how to generate ESG themes in an unsupervised manner via clustering algorithms and how to summarize the information contained in such a theme via dimensionality reduction techniques. Finally, we observe that traditional parametric models do not allow one to generalize a given ESG score to a large universe of entities. Our evaluation of these methodologies highlights the complexity of building ESG scores without prior supervision and the difficulty of generalizing available ESG scores to a broader universe.