Large language models (LLMs) like ChatGPT and Google Gemini are becoming valuable tools for individual investors, providing a way to sift through vast amounts of company data to glean market insights. Researchers from George Mason University, alongside collaborators from the University of Florida and the University of Massachusetts Boston, are at the forefront of this exploration. Their new working paper, published in the SSRN Electronic Journal, investigates how LLMs can help investors identify "peer firms"—or product market competitors—within industries.
Yi Cao, assistant professor of accounting at George Mason University, compares this process to the real estate market, where the value of a home is often determined by comparable properties. "In the capital market, a firm's value is partially determined by the value of its peers," Cao explains. The study seeks to leverage the capabilities of LLMs to automate and enhance the traditionally labor-intensive task of identifying these peers.
Cao and his colleague Long Chen, who is the associate professor and area chair of accounting at George Mason, utilized Google’s LLM, Bard—now known as "Gemini"—for their research. They chose Gemini for its extensive pre-training data, which Cao argues is more comprehensive than that of ChatGPT.
The researchers focused on large, publicly listed companies between 1981 and 2023, tasking the LLM with generating peer firms based on a defined concept of product market competition. On average, the LLM produced seven peer firms per focal company, a figure in line with recommendations from the U.S. Securities and Exchange Commission (SEC).
When compared to peer lists generated by human experts for leading software companies, the LLM’s performance was notable, with an overlap of over 40%. Additionally, the LLM’s output was compared to traditional methods like the Standard Industrial Classification (SIC) codes and the Text-based Network Industry Classification (TNIC). The results showed that the LLM’s peer identification was generally more accurate, particularly in aligning with monthly stock returns of the focal firms.
However, the study also found limitations. TNIC outperformed the LLM in identifying peers for mid-sized companies, suggesting that while LLMs are powerful, they are not yet superior in all contexts.
Cao and Chen emphasize that while LLMs offer unparalleled efficiency and accessibility, especially for individual investors, the technology is not without its challenges. "There are always costs and benefits associated with using generative AI," Chen notes. He stresses the importance of understanding these trade-offs, especially as AI tools continue to evolve rapidly.
The researchers advocate for embracing AI in investing but caution that the technology is still in development. "Our findings might just represent the lower bound of the effectiveness of the technology," Chen concludes, hinting at the potential for even greater advancements in the future.
