The Market's Mirror: Revealing Investor Disagreement with LLMs

Abstract

Large language models (LLMs) can emulate human perspectives. Leveraging this idea, we study how investor disagreement emerges in response to firm news. We endow an LLM with 216 representative investor personas and elicit buy, hold, or sell responses to S&P 500 firm news headlines from 2010-2025. Dispersion in responses yields article-level disagreement for 5.5 million headlines and sheds light on its sources. Disagreement is highest for socially-oriented news and lowest for fundamentals. Persona responses reflect non-pecuniary rationales and align with human-survey benchmarks. Disagreement predicts elevated same- and next-day abnormal trading volume, especially retail, and results persist beyond the LLM’s training cutoff.

Keywords

Tags of Social Finance