Quantifying legal risk with Large Language Models: A text-based investment signal

This study developed a novel, text-based investment signal by applying Large Language Model (LLM) technology to quantify changes in legal risk disclosed in twin Securities and Exchange Commission (SEC) 10-Q and 10-K filings. Using the OpenAI GPT-4o LLM, the method compared the Legal Proceedings sections and associated footnotes from period-adjacent filings to score directional changes in legal risk. All firms with non-zero scores were incorporated into monthly long-short sub-portfolios, where negative scores were ‘longed’ and positive scores were shorted. Positions were dynamically weighted using a hyperbolic tangent transformation based on each firm’s historical volatility and magnitude of legal risk change score. This approach offers a novel method for extracting firm-specific legal risk signals that have been historically underutilized in asset pricing models. It has potential applications in alpha generation, risk management, litigation monitoring, ESG compliance, and research automation. Ordinary Least Squares (OLS) regression was used to test the portfolio’s returns against the standard Fama-French five-factor model plus momentum. The long-short strategy produced a statistically significant mean monthly alpha of 0.471% (annualized to 5.80%) for each sub-portfolio, indicating excess return generation beyond traditional risk factors, though with exposure to a large-to-moderate drawdown. All results were unlevered. Given the strategy’s measured volatility and delta exposure, returns could be further scaled according to risk tolerance via leverage. Controlling for standard risk factors, the strategy’s returns were partially explained by exposure to both growth and conservative-investment firms, but a meaningful portion of alpha remained unexplained by market risk, size, profitability, or momentum, highlighting the distinct contribution of the legal risk signal. An LLM-only, equal-weight mirror test—portfolios traded solely on the LLM legal-risk score (no position sizing or additional parameters)—was also executed: the normal equal-weight portfolio (long risk decreases / short risk increases) decisively outperformed its mirror (same trades but with reversed longs/shorts) over both control and winsorized iterations. These findings suggest that LLMs can uncover priced legal risk shifts and generate predictive signals beyond traditional textual analysis techniques.