ContestTrade is a multi-agent trading system that addresses the high sensitivity of LLM-based agents to market noise by introducing an internal contest mechanism. The system comprises two teams: the Data Team, which processes raw market data into concise textual factors, and the Research Team, which uses a Deep Research methodology with specialized financial tools to generate trading signals. The core innovation is a real-time evaluation and ranking mechanism within each team, where agents are scored based on market feedback and only top-performing outputs are selected. This adaptive framework enhances robustness against noise and improves trading performance.

Experiments on A-share market data from January to June 2025 show that ContestTrade significantly outperforms traditional methods (MACD, RSI&KDJ), machine learning (LGBM, LSTM), deep reinforcement learning (A2C, PPO), and other multi-agent systems (MASS) across cumulative return, Sharpe ratio, and maximum drawdown. The internal contest mechanisms are validated by high Rank IC and ICIR values, indicating effective prediction of factor and signal quality. Ablation studies confirm that each component—LLM Judge, contest mechanisms, and Deep Research—is crucial for the system's success.

The paper's contributions include a dynamic multi-agent architecture with internal contests, adaptation of Deep Research to finance, and robust information denoising. Future work aims to scale simulations, apply to broader markets, and integrate diverse data sources.