Business & Finance Investment ESG Sentiment from Large Language Models and Its Predictive Power for Portfolio Risk

Abstract

The Large Language Models (LLMs) are also using sentiment analysis to generate corporate report sentiment data, articles sentiment data, earnings call sentiment data, sustainability report sentiment data, and social media sentiment data. This paper investigates if sentiment written by a large language model (LLM) can predict portfolio risk. It investigates the relationship between positive and negative signals and the risk metrics such as volatility, downside risk, drawdown exposure and portfolio stability. The results indicate that sentiment derived from LLM models can offer valuable forward-looking insights into ESG evaluations, in addition to conventional financial signals. At each of these firm levels, consistently positive ESG sentiment was associated with reduced perceived risk, increased investor confidence, and lower volatility of the portfolio, while negative sentiment was associated with increased downside exposure, reputational risk, and volatility. Another key strength the study finds is that LLMs are better able to understand nuanced language associated with the ESG when compared with the basic keyword-based approach, which will enhance the quality of sentiment measurement. Yet, there are still risks and issues, such as model bias, the risk of hallucination, the use of inconsistent ESG terminology and reliance on source document quality. Overall, LLM-generated ESG sentiment and financial data and human validation, and responsible model governance can be harnessed for use in portfolio risk assessment, the paper concludes.