Multi-Agent LLM Trading Arena: Let Large Language Models Compete in a Virtual Market

This article introduces the innovative multi-agent trading system LLMTradingAgents, which allows different large language models to compete in a virtual market through a three-layer architecture of strategist, risk manager, and fixer. The system integrates multi-dimensional information such as market data aggregation, technical analysis, fundamental analysis, and news sentiment, providing a standardized platform for evaluating LLMs' complex decision-making capabilities.

The capabilities of large language models are expanding, but the financial trading field is full of challenges due to complex dynamics and uncertainties. Instead of pursuing a single "trading holy grail", LLMTradingAgents builds a multi-agent competition framework that allows different models to compete in the same virtual market, revealing differences in models' risk perception, strategy formulation, and execution discipline, and providing a standardized evaluation platform.

Three-Layer Agent Architecture: 1. Strategist Agent: Receives market data (technical indicators, fundamentals, news sentiment) and outputs structured proposals including direction, confidence level, and reasoning; 2. Risk Manager Agent: Verifies whether the proposal complies with risk control rules such as capital, position, and short-selling restrictions, and converts it into an executable trading plan; 3. Fixer Agent: Corrects format errors and improves system robustness.

Data Layer: A multi-source aggregation mechanism covering price history (yfinance, CCXT), technical indicators, fundamentals (SEC reports, CoinGecko), and news sentiment (Alpha Vantage) to ensure data quality and reliability.

Trading execution is handled by SimBroker, a simulated brokerage, which manages virtual cash, positions, and orders, supporting market orders, slippage, and transaction fees to closely mimic real scenarios. Trading status is stored in an SQLite database.

The frontend is built as a static single-page application using React 19, Vite, and Tailwind CSS v4. The serverless architecture simplifies deployment, and CI/CD processes (running trading sessions, generating data, deploying to Vercel/Netlify) are implemented via GitHub Actions.

The system is developed with Python 3.10+, enabling high customization through YAML configuration files: configurable competitor lists (different providers and models), position limits, and simulation parameters (initial capital, slippage, etc.). The CLI entry provides operational interfaces for real-time trading, dry-run mode, manual data export, and more.

As a standardized evaluation platform, LLMTradingAgents can assess models' comprehensive ability to process multi-modal financial data, behavioral patterns under risk constraints, and strategy evolution in multi-round interactions. The framework can be extended to scenarios such as multi-asset portfolio optimization, cross-market arbitrage, and event-driven trading, potentially becoming a standard tool for evaluating AI's complex decision-making capabilities.

LLMTradingAgents demonstrates the potential of multi-agent architecture in complex decision-making tasks. By splitting the trading process into specialized roles, it achieves a balance between functional modularity and robustness. The design idea of "letting models compete instead of relying on a single one" can provide reference for more high-risk AI application scenarios.