Large Language Models, Investor Disagreement, and Trading Volume
Abstract
Differences of opinion are central to theories of trading and volatility, yet existing disagreement proxies are too low-frequency or rely on platform-specific investor tags. We develop a replicable, daily disagreement measure in an unlabeled social-media environment and study how belief heterogeneity maps into trading in China’s A-share market. Using 52.3 million investor comments for CSI 300 constituents from 2016–2025, we implement an LLM stance-simulation design: prompt-engineered Neutral, Bullish, and Bearish agents generate training labels that we scale with three parallel BERT classifiers to construct within-group (information uncertainty) and cross-group (model conflict) disagreement. Within-group disagreement robustly predicts lower abnormal trading volume, whereas cross-group disagreement predicts higher volume. Mechanism tests show the within-group effect strengthens in high policy uncertainty, consistent with a value-of-waiting channel, while cross-group effects are amplified, consistent with gains from trade under competing valuation models. Taken together, our results imply that the trading consequences of disagreement depend on whether it captures within-model noise or across-model conflict, contributing to the heterogeneous-beliefs literature; our stance-simulation design further delivers a reproducible way to construct belief-heterogeneity measures in environments with scarce investor-type labels.