StockAgent: An Innovative Framework for Simulating Realistic Stock Trading Environments Using Large Model Agents

Original Author/Maintainer: MingyuJ666 Source Platform: GitHub Original Title: Stockagent Original Link: https://github.com/MingyuJ666/Stockagent Source Publication/Update Date: 2026-06-08

StockAgent is a large language model (LLM)-based multi-agent system designed to simulate realistic stock trading environments and study the impact of external factors on trading behavior. The core innovation of this system lies in addressing the test set leakage issue in traditional AI trading simulations, providing a new perspective for understanding the application of LLMs in complex financial decision-making scenarios.

In the financial field, stock trading is influenced by various external factors such as macroeconomics, policy changes, company fundamentals, and global events. Traditional quantitative models struggle to fully capture the interactions of these dynamic factors, while rule-based simulation systems lack flexibility and realism.

In recent years, the strong reasoning capabilities of LLMs have made it possible to build more intelligent trading simulation systems. However, existing AI agent simulations face the problem of test set leakage—models may use knowledge from test data in pre-training to make decisions, leading to distorted evaluation results.

StockAgent is a multi-agent AI system that uses LLM-driven agents to simulate real investor trading behaviors, build a near-real trading environment, and avoid test set leakage.

The system simulation process is divided into four stages:

1. Initial Stage: Initialize the trading environment, market parameters, and agents' initial states (including investment goals and risk preferences);
2. Trading Stage: Agents make buy/sell decisions based on market status, news, and their own strategies;
3. Post-Trading Stage: Process daily events (closing price adjustments, news releases) and quarterly events (financial report releases);
4. Special Event Stage: Simulate random breaking news, policy changes, and other events to test agents' response capabilities.

Multi-agent Collaboration Mechanism

Adopts a multi-agent architecture representing different types of investors (value investors, trend followers, etc.) to reflect market diversity and complexity.

External Factor Modeling

Focuses on the impact of four types of factors: macroeconomic indicators (GDP, inflation rate, etc.), policy changes (monetary policy, regulatory updates), company fundamentals (profitability, financial status), and global events (geopolitical conflicts, pandemics, etc.).

Test Set Leakage Prevention

Ensures authentic evaluation through three mechanisms:

1. Time window isolation: Restricts the range of historical data accessible to the model;
2. Information shielding: Blocks future information input during testing;
3. Dynamic scenario generation: Creates market environments unseen by the model.

The research team evaluated StockAgent on different LLMs with experimental settings close to real market conditions. Evaluation metrics include:

• Trading behavior analysis: Trading frequency, holding period, asset allocation;
• Profitability assessment: Return rate, risk-adjusted return, maximum drawdown;
• Market impact research: Impact of trading on price volatility.

Experimental results reveal the influence of key external factors on trading behavior and price volatility, especially the pattern characteristics shown by agents' free trading when they have no prior knowledge of market data.

The application value of StockAgent includes:

1. Risk Management: Simulate different market scenarios to help understand potential risks;
2. Strategy Validation: Test the effectiveness of investment strategies before actual deployment;
3. Market Mechanism Research: Provide a tool for academia to study market microstructure;
4. Regulatory Policy Evaluation: Simulate the impact of new policies on market participants.

In addition, the system is developed based on Python, supports mainstream LLMs such as GPT and Gemini, and has a modular design for easy expansion.

Limitations

• Market simplification: Cannot fully replicate all complexities of the real market;
• Agent homogenization: Agents based on the same LLM may exhibit similar behaviors;
• Data dependence: Simulation quality depends on the quality and coverage of input data.

Future Directions

• Introduce more complex market microstructure;
• Support more types of financial assets;
• Explore the impact of multi-modal inputs (images + text) on trading decisions.