ESG Sentiment from Large Language Models and Its Predictive Power for Portfolio Risk

Abstract

This research will compare the accuracy of the sentiment derived from ESG data using LLMs to that of traditional financial data alone in predicting the risk of a portfolio. As investors rely more on ESG information, textual reporting, ESG news and sustainability stories and public communication are important sources of risk-relevant signals. The research investigates how LLM-derived ESG sentiment scores can capture both sentiment and tone in corporate sustainability information and how it relates to portfolio volatility, downside risk, value-at-risk and drawdown behavior. The results indicate that negative ESG sentiment is correlated with higher portfolio risk and positive ESG sentiment is correlated with lower portfolio volatility and a more stable portfolio performance. The results also indicate that the negative sentiment and weak sentiment portfolios have a higher downside exposure than the positive sentiment portfolios. The predictive models also reveal that the sentiment of ESG data, derived from language models, improves over regular market indicators’ ability to estimate risk. Finally, the paper highlights the power of LLM sentiment analysis tools for ESG as a tool that can assist portfolio managers, risk analysts, and sustainable investors as a decision support tool. Prior to applying sentiment ESG signals to real investment it is important to validate the model, be transparent, and control for bias.