EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents
Sehee Kim, Yumin Choi, Minki Kang, Sung Ju Hwang
cs.AI · cs.CL
Quick verdict
Self-rewriting system prompts yield suggestive LLM trading gains (best in 3/6 regimes), but evidence is thin: 3 runs, 20-day windows, no significance tests.
Summary
EvolveTrade treats the system prompt of a tool-using LLM trading agent as a text-parameterized policy. Every five trading days, a Policy Agent (same frozen backbone) reads accumulated decision traces plus realized portfolio feedback and rewrites the full policy; tools (price retrieval, news search, Python interpreter) and LLM weights stay fixed. A motivating study shows a static policy locks the agent into a fixed analytic vocabulary (five metrics used daily; RSI/EMA/VaR never invoked) across regimes. On a LiveTradeBench-style daily-close simulator with 15 US blue chips, across six post-cutoff monthly windows and two backbones (GPT-5-mini, Gemini-2.5-Flash), EvolveTrade attains the best LLM-method Sharpe and cumulative return in three of six settings (Jan/Sep 2025, Nov 2025), loses both April drawdown windows to static baselines, and trails EvolveStrategy in Feb 2026. Behavioral analyses show code-interpreter calls rising from ~1 to 2.6-4.5/day, regime-relevant metric activation, and a traceable NVDA case (10.7% to 2.9% exposure before a -17% day). Robustness checks include 3-run standard deviations, 10bps transaction costs, a 50-day horizon, and update-interval sensitivity. Main limitations: three runs with no significance tests, ~20-day windows making annualized Sharpe estimates very noisy, no code-availability statement, and execution at the same close used for decision inputs.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Revising a tool-using trading agent's system prompt online from its own decision traces and realized portfolio feedback improves Sharpe Ratio and cumulative return over fixed-policy LLM agents in a subset of post-cutoff regimes (best LLM method in 3 of 6 model-regime pairs).
- Static tool-use policies lock the LLM into a regime-invariant analytic vocabulary, whereas self-evolution activates regime-relevant computations (VaR and signal normalization in drawdowns; EMA/RSI/SMA in uptrends) and raises code-interpreter use from ~1 to 2.6-4.5 calls per day.
- Policy refinement produces traceable allocation changes linked to realized returns, e.g., cutting NVDA exposure from 10.7% to 2.9% before its -17% single-day drop in January 2025.
- The benefit persists over a 50-trading-day horizon with non-monotonic policy length, survives a 10bps transaction-cost adjustment at lower turnover than evolving baselines, and a moderate 5-day update interval works best.
Strengths
Clean conceptual framing with well-isolated controls
Sec 5 formalizes the prompt-as-policy update (Eqs. 3-5) precisely, and Sec 6.1's five-configuration ladder (Static Base, Static TC, EvolveBase, EvolveStrategy, EvolveTrade) separates the effect of tool access from the effect and scope of policy evolution, including an ATLAS-adapted strategy-only updater as a comparison.
Empirically demonstrates the claimed bottleneck before proposing the fix
Sec 4 / Fig 2 shows the static agent invokes the same five metrics (Sharpe, 20d/60d return, ann. vol, drawdown) on 90-100% of days in all three regimes while RSI, EMA, VaR and normalization are never called - concrete evidence that a fixed prompt caps the analytic vocabulary the LLM actually uses.
Mechanism-level behavioral evidence, not just a leaderboard
Unusual transparency and robustness appendices
Weaknesses
No statistical significance testing
Very short windows make headline metrics unreliable
Wins are partial and the framing softens this
Same-close execution embeds mild look-ahead
No ablation isolating feedback-driven refinement; reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is EvolveTrade?
A framework where an LLM trading agent improves itself by rewriting its own system prompt (its 'policy') every five trading days, using logs of its past decisions plus the actual portfolio returns those decisions produced. The underlying LLM and tools stay frozen; only the instruction text evolves.
Does EvolveTrade beat the market or rule-based strategies?
Not consistently. Across six one-month test windows and two LLM backbones it was the best LLM-based method in three windows, and in a couple of windows it beat all rule-based baselines on return metrics (e.g., GPT-5-mini, Sep 2025). But it lost to simpler static baselines in both April drawdown windows and to another evolving baseline in February 2026, so it should not be read as reliable alpha.
Is code or data available for EvolveTrade?
The reviewed text contains no code or data release statement. The paper does fully publish its initial prompt, update prompt, and evolved policies in appendices, and builds on the public LiveTradeBench evaluation framework with tools from AI-Trader, but the news source and decoding parameters are unspecified, limiting exact replication.
Why does EvolveTrade underperform in April drawdown periods?
The authors' own analysis (Appendix F) shows the evolved policy raised average cash allocation well above the baselines (+10.1 percentage points in Apr 2025 and +35.9pp in Apr 2026), suggesting overly defensive positioning that missed the subsequent recovery or rally - evidence that online refinement can overfit to recent negative feedback.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The source is a text extraction of the PDF: tables are linearized, so table values were transcribed carefully but column misalignment cannot be fully ruled out.
- Figures 2-6 are images; only their captions and in-text descriptions were reviewed, not the underlying plots.
- No code/data availability statement appears in the extracted text; one could exist in a portion or footnote not captured by extraction.
- Some numbers in the abstract (e.g., '+1.33 percentage-point') do not exactly match the figure-level values visible in the text; flagged as a question rather than resolved.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.