LLM-Based Dynamic Multi-Agent Systems for Evaluation of Financial Trading Mechanism

This paper addresses the challenge of evaluating trading mechanisms in complex financial markets. It introduces a multi-agent system, powered by LLMs, to simulate diverse market participants, including rational and emotional traders. This approach overcomes limitations of traditional quantitative models by capturing complex behaviors. The system incorporates roles like fundamental analysis, sentiment analysis, and risk management. Empirical results demonstrate that LLM-based strategies exhibit superior risk control. This research provides a novel framework for analyzing trading mechanisms, offering new avenues for future investigation. This simulation system includes multiple agents, each responsible for a specific role, such as fundamental analysis, news sentiment analysis, technical analysis, industrial policy research, simulated quantitative trading, risk management, and trading early warning mechanisms. Empirical results show that agents based on LLM significantly outperform traditional strategies (Buy & Hold, CR: 30.12%) in terms of cumulative return rate (CR: 42.00%), annualized return rate (ARR: 92.13%), and Sharpe ratio (SR: 2.38), and demonstrate stronger risk control capabilities (MDD: -11.23% vs. -16.42%).