Exploring the Predictive Capacity of ESG Sentiment on Official Ratings: A Few-Shot Learning Perspective

Environmental, social, and governance (ESG) criteria are increasingly central to corporate reporting. This study applies natural language processing (NLP) techniques, specifically a RoBERTa-based few-shot model, to conduct aspect-based sentiment analysis (ABSA). Our analysis targets ESG-related entities and their sentiments within EUROSTOXX 50 company reports to assess their impact on ESG ratings. The ratings data are sourced from established providers, including Refinitiv, S&P, and Bloomberg. Furthermore, to explore the potential reciprocal influences on these variables, we employ a vector auto-regressive (VAR) model, which facilitates the modeling of bidirectional interactions. This combination of advanced NLP methods and comprehensive data integration aims to provide detailed insights into the dynamics between company disclosures and rating providers’ ESG scores. The results of our study indicate that in general there is no discernible relationship between the ESG sentiment as reflected in company reports on the EUROSTOXX50 and the ESG ratings provided by the rating agencies. Nevertheless, our tool can provide an alternative, fine-grained measure of companies’ own views on ESG-related matters.