The Promise and Peril of Generative AI: Evidence from GPT-4 as Sell-Side Analysts

Abstract

We investigate how advanced large language models (LLMs), specifically GPT-4, process corporate disclosures to forecast earnings. Using earnings press releases issued around GPT-4’s knowledge cutoff date, we address two questions: (1) Do GPT-generated earnings forecasts outperform analysts in accuracy? (2) How is GPT’s performance related to its processing of textual and quantitative information? Our findings suggest that GPT forecasts are significantly less accurate than those of analysts. This underperformance can be traced to GPT's distinct textual and quantitative approaches: its textual processing follows a consistent, generalized pattern across firms, highlighting its strengths in language tasks. In contrast, its quantitative processing capabilities vary significantly across firms, revealing limitations tied to the uneven availability of domain-specific training data. Additionally, there is some evidence that GPT’s forecast accuracy diminishes beyond its knowledge cutoff, underscoring the need to evaluate LLMs under hindsight-free conditions. Overall, this study provides a novel exploration of the “black box” of GPT-4’s information processing, offering insights into LLMs’ potential and challenges in financial applications.

JEL Classification: G10, G14, M40