A Hybrid AI Framework for Investor Sentiment: Causal Inference and Multi-Agent Simulation in China's Stock Market
Year: 2026
Type: conference-paper
Abstract: This paper develops a hybrid AI framework to examine investor sentiment transmission in China's stock market. Using 6.3 million stock forum comments from 2010 to 2024, we construct a daily emotional index via large language model-assisted text classification. We combine difference-in-differences (DID) with multi-agent deep reinforcement learning (DRL) simulation to study aggregate effects and behavioral mechanisms. The DID results, based on the 2019 STAR Market launch as a quasi-experiment, show that sentiment's average treatment effect on market returns is statistically insignificant. Rather than interpreting this as irrelevance, we argue that sentiment operates through behavioral channels. The DRL simulation reveals that sentiment shocks increase trading behavior consistency and amplify market volatility. These findings suggest that macro-level insignificance can coexist with meaningful micro-level behavioral convergence. Methodologically, this study integrates text-based sentiment measurement, causal inference, and behavioral simulation in a unified framework, offering a comprehensive perspective on sentiment transmission in financial markets.
Authors: Qianhui Sun, Jiaqi Wang, Yundi Zhou
Language: English
Subfield: Finance
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7213293100