LLM-Enabled Agent-Based Modeling for Stock-Market Simulation: An Analysis of Trading-Style Switching in Agents,SIM

Zeping Li, Guancheng Wan, Keyang Chen, Zhenfei Yin, Yu Chen, Guangnan Ye, Philip Torr, Hongfeng Chai

Keywords: Agent-based modeling, LLM-based simulation, Behavioral finance, Style switching

Abstract: Stock-market simulation is a natural micro-to-macro testbed with abundant disclosures and price-volume records, yet most prior systems fix an agent's trading style, precluding switches between fundamental and technical strategies. We ask a core behavioral-fidelity question: Under realistic information and constraints, do LLM agents switch between fundamental and technical strategies in accordance with financial theories, and thus qualify as credible components in market simulation? We propose an evaluation framework that operationalizes four behavioral-finance drivers---loss aversion tendency, herding tendency, wealth differentiation sensitivity, and price misalignment sensitivity---as personality parameters set via prompting and stored as long-term traits. In year-long 2024 simulations, agents ingest daily price–volume data, compute standard indicators, trade under a designated style while a counterfactual ledger tallies foregone P\&L for the alternative, and reassess their style every 10 trading days with self-reported rationales. This paper introduces four alignment metrics and uses Mann–Whitney U tests to evaluate whether LLM agents’ style switching aligns with financial theory. The experiments show that LLM agents’ style-switching behavior aligns with financial theory in part, but their comparatively rational, short-term focus also yields mismatches with behavioral-finance predictions.

Area: Modelling and Simluation of Societies (SIM)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 1794