EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents
Sehee Kim, Yumin Choi, Minki Kang, Sung Ju Hwang
cs.AI, cs.CL
2026-09-15
EvolveTrade online-rewrites a frozen LLM trading agent's system prompt; GPT-5-mini posts 6.84% CR in Sep 2025 and wins LLM Sharpe in three of six regimes.
LLM trading agents can already read prices, fetch news, and run code. What stays frozen is the procedure: when to call tools, how to cross-check signals, how to size risk. That procedure usually lives in a system prompt written before deployment. KAIST ran GPT-5-mini with a fixed policy across a sideways month, a drawdown month, and an uptrend month. The code interpreter computed the same five metrics almost every day: Sharpe, 20-day return, 60-day return, annualized volatility, drawdown, on 90 to 100% of sessions. RSI, EMA, VaR, and signal normalization never appeared. The April 2025 −9.8% drawdown and the September uptrend shared the same analytic vocabulary.
A fixed prompt locks which slice of the model's skill gets used, and nothing in the loop lets realized trades rewrite that procedure.
EvolveTrade treats the system prompt as a text-parameterized tool-use policy. The backbone LLM and the tool APIs stay frozen. Each day the trading agent actively gathers evidence with a price tool, a news tool, and a Python interpreter, then emits a long-only allocation plus per-asset rationales. The simulator rebalances 15 US blue chips plus cash at the close. Retrieval tools share a temporal cutoff against look-ahead.
Every N trading days (default 5) a Policy Agent reads the batch of allocations, rationales, and realized portfolio feedback, then rewrites the full policy text for the next batch. Edits can change tool priority, query wording, cross-validation, risk rules, and output format. They do not retrain weights. Ablations isolate the moving parts: EvolveStrategy evolves high-level strategy text while tool protocols stay fixed; EvolveBase evolves the prompt without tools.
Two backbones, three one-month windows after each model's cutoff, LLM rows averaged over three runs.
GPT-5-mini:
| Method | Jan 2025 SR / CR | Apr 2025 SR / CR | Sep 2025 SR / CR |
| Static Tool-Calling | 2.87 / 3.30% | −1.54 / −4.71% | 6.45 / 4.86% |
| EvolveStrategy | 3.86 / 4.70% | −3.09 / −6.85% | 7.76 / 4.78% |
| EvolveTrade | 5.12 / 5.10% | −2.53 / −6.60% | 8.43 / 6.84% |
January and September are the best Sharpe and cumulative return among LLM agents; September's 6.84% also beats every rule baseline in that window. In the April drawdown the static tool agent loses less. With Gemini-2.5-Flash, EvolveTrade's November bear-market SR −1.03 and CR −1.07% slightly lead the LLM group; February 2026 SR 2.75 and CR 2.92% beat all rule baselines, while EvolveStrategy is ahead at 3.92 / 3.52%; in the April 2026 bull month the tool-free Static Base wins clearly (SR 9.07, CR 11.64%). Across six model-regime pairs, EvolveTrade takes the best LLM Sharpe and return in three; fixed policies win two others.
Behavior matches the returns. Daily code-tool calls rise from about 1.0 to 2.6–4.5; news search does not increase. VaR and signal normalization appear in the April drawdown; EMA, RSI, and SMA appear in the September uptrend; the static tool agent never invoked those in any of the three months. On 24 Jan 2025 the static agent held 10.7% NVDA versus 2.9% under EvolveTrade's revised sizing. NVDA fell 17.0% the next session; portfolio returns were −0.03% versus −1.11%, a +1.33 point gap.
Over 50 trading days (2 Sep to 10 Nov 2025) EvolveTrade still leads on SR 4.00 and CR 10.56%. Policy text grows and periodically contracts rather than inflating forever. Daily updates are worse: average SR about 1.92 and CR 0.16% versus a stronger default five-day interval.
For teams already running tool-using trading agents, this is online adaptation without weight updates: the object that changes is the reusable procedure, not a memory store and not an offline-learned orchestrator. The claim is "better risk-adjusted numbers in most windows," not "wins every regime." April 2025 and April 2026 are the counterexamples.
The update interval is fixed for a whole run; it cannot speed up when volatility jumps. Trading friction is proportional commissions only, no slippage, impact, or liquidity. Windows are one-month scale with three seeds. In the April drawdown the static tool policy lost less, so evolution can write noise into the prompt. Portfolio feedback is delayed and noisy; the Policy Agent can turn luck into a rule. The universe is 15 blue chips.