Reproduce “EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sehee Kim, Yumin Choi, Minki Kang, Sung Ju Hwang
EvolveTrade treats an LLM trading agent’s system prompt as an adaptable policy that can be revised from realized trades, tool-use traces, and portfolio outcomes without updating model weights. Across selected market regimes and two LLM backbones, the approach changes the agent’s analytical workflow and often improves risk-adjusted performance, although results remain mixed during drawdown recoveries and across update schedules.
Key insights
- Policy evolution changes the agent’s behavior in observable ways: code-interpreter calls increase from about one per day to roughly 2.6–4.5 per day, and indicators such as value-at-risk, signal normalization, SMA, EMA, and RSI become active when relevant.
- The useful adaptation target is the procedure for acquiring and processing evidence, not merely access to more tools or a revised high-level trading objective.
- Moderate five-day update intervals outperform daily revisions in the reported sensitivity analysis, suggesting that accumulating several days of feedback helps limit adaptation to noisy short-term outcomes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.