EvolveTrade Refines LLM Trading Policies Through Experience

Summary

EvolveTrade addresses a limitation of LLM trading agents whose tool-use policies are hand-written and fixed before deployment. It treats the agent's system prompt as a text-parameterized policy that governs evidence gathering, tool invocation, signal verification, and portfolio construction. At each update interval, a separate Policy Agent uses accumulated decision traces and realized portfolio feedback to revise that policy, while the backbone LLM remains unchanged. The revised policy controls the next batch of trading decisions. Experiments across multiple market regimes and two LLM backbones found that the approach often improved Sharpe Ratio and Cumulative Return over fixed-policy LLM baselines, with higher values for both metrics in most evaluated settings. Behavioral analysis indicated that evolved policies increased code-mediated analysis and activated computations relevant to the current regime. Case-level attributions linked policy-driven allocation changes to differences in realized returns. The authors present adaptive tool-use procedures as a potential route to more robust LLM trading agents, while the reported evidence is limited to the evaluated settings.