Researchers introduced EvolveTrade, a framework for self-evolving LLM trading agents. It treats the system prompt as a text-parameterized policy. A Policy Agent revises this policy after each update interval using decision traces and portfolio feedback. The backbone LLM remains unchanged during the process. Experiments show improved Sharpe Ratio and Cumulative Return over fixed-policy baselines. Self-evolved policies increase code-mediated analysis and activate regime-relevant computations. Policy changes directly contribute to realized return differences. The paper suggests adapting tool-use procedures is key for robust trading agents.
What changed
LLM trading agents used static hand-written tool-use policies fixed before deployment.
What it unlocks
Trading agents can now refine their information-gathering and portfolio-construction procedures over time.
- 2 LLM backbones tested
- Sharpe Ratio improved in most settings
- Cumulative Return improved in most settings
Sources