Large language model-based multi-agent systems for financial markets simulation: a survey

Abstract

Traditional financial modeling approaches, such as econometric methods and dynamic stochastic general equilibrium models, exhibit inherent limitations in capturing the highly interactive and strategic nature of modern financial systems. To overcome these limitations, multi-agent systems (MASs) have increasingly been adopted as a computational paradigm well suited for simulating the complex behaviors and interactions of participants in financial environments. Within this paradigm, the emerging class of large language model (LLM)-based MASs has demonstrated unprecedented potential for modeling intricate financial interactions. This study provides a systematic review of LLM-based MAS applications in financial markets. We first present the motivation for adopting LLM-based MASs and highlight their key characteristics, including heterogeneity, autonomy, adaptability, and bounded rationality, which render them particularly effective for representing complex financial ecosystems. Then, we conduct a technical analysis of financial agents, providing the enabling techniques for LLM-based agents, a taxonomy of their foundational models, and the underlying mechanisms that facilitate the emergence of collective intelligence in LLM-based MASs. Furthermore, we survey some representative applications of LLM-based MASs in financial markets, including market dynamics simulation, systemic risk analysis, policy evaluation, algorithmic trading, and sentiment analysis, while critically discussing the associated technical and regulatory challenges. Through this comprehensive synthesis, we aim to provide a state-of-the-art reference for intelligent agent-based approaches in computational finance and to identify key opportunities and limitations that characterize this rapidly evolving research field.