Using Large Language Models to Estimate Novel Risk: Impact on Volatility

This article presents an integrated framework to estimate hard to measure (novel) financial risk for volatility forecasting. Recognizing the limitations of traditional models—which often overlook emerging “novel risks”—the article leverages advanced large language models (LLMs) to extract and quantify key risk factors, including ESG, geopolitical, and supply chain disruption risks, from corporate disclosures. These LLM-derived risk scores are then combined with conventional financial indicators such as leverage, beta, and short interest, and incorporated into a long short-term memory (LSTM) neural network to predict firm-specific (idiosyncratic) volatility. Empirical analysis, conducted on over 18,000 regulatory filings spanning 2015 to 2024, demonstrates that the integrated model significantly improves volatility forecasting, as evidenced by enhanced R2 values and reduced mean squared error. Additionally, feature importance analyses confirm the pivotal role of novel risk measures. Overall, the findings underscore the benefits of merging unstructured and hard to quantify data with quantitative models to offer a more nuanced approach to estimation of novel financial risk.

Read full research paper here.