An LLM-Driven Multi-Agent Framework for Behaviorbased Stock Market Simulation

Most existing applications of Large Language Models (LLMs) in financial markets focus on price prediction and often suffer from information leakage and weak causal reasoning. To address these limitations, this paper proposes a behavior-oriented, multi-agent stock market simulation framework that moves beyond direct price forecasting. The framework models market dynamics through interactions among knowledge-isolated agents operating under strict time-bounded information constraints, ensuring that decisions are made solely based on currently available market context. The market is represented as a real-time stochastic environment in which heterogeneous agents adapt their trading behavior in response to evolving prices and interactions. To enhance interpretability, agents generate LLM-inspired natural-language reasoning traces that explain their decisions at each simulation step. Experimental results show that the proposed framework reproduces key market phenomena such as price formation, volatility clustering, and herding behavior while preventing unintended access to future information. Overall, the framework provides an interpretable and flexible platform for analyzing market dynamics, evaluating financial theories, and studying market stability under uncertainty.