TradingMoE Improves LLM Trading Performance in Evolving Markets.

Large language models (LLMs) show promise in financial analysis, but direct trading presents challenges due to varying predictive needs across assets and market conditions. Existing LLM-based trading systems often struggle with effectively coordinating experts or adapting to market changes. This paper introduces TradingMoE, a trading-oriented sparse Mixture-of-Experts (MoE) system. It augments a frozen dense LLM with lightweight residual experts and features a novel Query-Key router that matches token-specific expertise requirements with learnable expert keys. A key innovation is a sparse expert selection update mechanism that samples inactive experts to dynamically adjust routing as market conditions evolve, maintaining computational efficiency. Experiments across stock and cryptocurrency markets demonstrated that TradingMoE significantly outperformed 22 baselines, achieving over 30% higher cumulative returns. Rolling paper-trading simulations further confirmed its sustained advantage in forward-only deployment, highlighting its ability to route the most suitable experts in dynamic financial environments.