Regime-Aware LLM Agents for Cryptocurrency Markets

Large Language Models have been applied to financial markets successfully in recent years and have shown remarkable capabilities regarding decision making and portfolio management. However, these LLM-based trading agents are mostly based on a static approach which makes them one-size-fits-all type. These do not take into account the dynamic nature of the market and different market regimes. Especially, cryptocurrency markets and prices are characterized by extreme volatility, regime shifts and co-dependency among the coins. Static prompts are not suitable for them, and result to inadequate performance. This document proposes a Regime-Aware LLM Agent (RALA) which is a novel framework that first detects regimes using Maximum Mean Discrepancy (MMD) and then routes the decision making process through specialized LLM prompts for different regimes and specific market conditions. We have also taken into account diverse market conditions like the FTX collapse and ETF approval events which causes a sudden change in the market regime and also a trend shift. Our results demonstrate that this regime-aware strategy given an improvement of 34\% in risk-adjusted returns when compared to the state of the art static prompt LLM model.