Can large language models (LLM) simulate market-scale, real-time business sentiment from the Chief Financial Officer (CFO) perspective? In their June 2026 paper entitled "CFOs Meet LLMs", John Graham, Campbell Harvey and Manish Jha use an LLM (GPT-5.4) to simulate CFOs of specific firms. Simulations involve two prompts for each of 6,075 actual responses to the quarterly Duke-Federal Reserve CFO Survey during 2002 through 2025:
- System prompt - establishes context, persona and analytical instructions, including instructions to collect from the web CFO interviews and analyst reports, price targets and consensus revenue estimates. Instructions include a temporal restriction to ensure that information collected was available before the matched survey response date.
- User prompt - provides firm profile (industry, revenue, headcount and geographic footprint), respondent history (prior CFO optimism scores when available and the LLM's own prior forecasts) and the key survey question: "Rate your optimism about the overall U.S. economy on a scale from 0–100, with 0 being the least optimistic and 100 being the most optimistic." Again, a temporal restriction imposes an information cutoff date matching the actual survey date.
They run each pair of prompts three times a few seconds apart and average outputs to suppress LLM statistical variation. Finally they match every LLM average output to the response of the actual CFO at the same firm in the same quarter. Using past CFO Survey results and associated firm data spanning 2002 through 2025, they find that:
Subscribe to Keep Reading
Get the research edge serious investors rely on.
- 1,200+ research articles
- Monthly strategy signals
- 20+ years of backtested analysis
Cancel anytime