StockAgent: A Multi-Agent System Using Large Models to Simulate Real Stock Market Trading Behaviors

StockAgent is an LLM-based multi-agent system for simulating investors' trading behaviors in a real stock market environment. This system addresses the test set leakage problem in existing AI trading systems and can evaluate the impact of external factors such as macroeconomics, policy changes, company fundamentals, and global events on stock trading, providing a new tool for quantitative investment and financial market research.

Traditional quantitative trading models struggle to fully capture the impact of multi-dimensional external factors, while machine learning prediction models face risks of data leakage and overfitting. The strong reasoning and decision-making capabilities of large language models provide the possibility to simulate complex human behaviors. The research team proposes using LLMs to build agents to study the impact of external factors on trading behaviors, which has both academic value and practical significance for quantitative investment.

The StockAgent system consists of four stages: the initial stage sets market parameters and agents' investment styles; the trading stage uses LLMs to parse information and generate trading instructions; the post-trading stage handles daily/quarterly events; the special event stage simulates sudden major events to test the agents' response capabilities.

Existing AI trading simulation systems have test set leakage flaws, leading to distorted evaluation results. Through its design, StockAgent ensures that prior knowledge of test data cannot be used during simulation, creating a fair evaluation environment and accurately reflecting the model's generalization ability.

The research team uses StockAgent to evaluate mainstream LLMs such as the GPT series and Gemini, simulating real trading environments to examine the impact of external factors on trading behaviors and stock price fluctuations, with a focus on studying the gaps in agents' free trading when there is no prior market data.

The simulation results of StockAgent provide insights for LLMs in the field of investment advice, helping to understand market decision-making logic, identify key factors of stock price fluctuations, develop robust strategies, and assess risks. The open-source implementation provides a reproducible platform for academia and industry, promoting the development of LLM applications in the financial field.

StockAgent represents an important progress in the intersection of AI and finance. In the future, it is expected to play a greater role in investment strategy backtesting and optimization, market risk management, investor behavior research, financial education and training, etc. This research has been accepted by the ACM TIST journal and recognized by the academic community.