Source-linked AI summary
EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents
Sehee Kim, Yumin Choi, Minki Kang, Sung Ju Hwang
TL;DR
LLM trading agents commonly rely on fixed tool-use policies that cannot adapt evidence gathering and risk management to changing market regimes. EvolveTrade updates the system-prompt policy from decision traces and realized portfolio feedback while keeping the backbone LLM fixed. Across multiple regimes and two backbones, it often improves Sharpe Ratio and Cumulative Return and changes analysis toward regime-relevant computations.
Problem
Existing LLM trading agents typically use static policies, limiting adaptation of evidence gathering, tool invocation, signal verification, and risk management under changing regimes.
Method
EvolveTrade treats the system prompt as a text-parameterized policy and periodically revises it from decision traces and realized portfolio feedback without retraining the LLM or modifying tools.
Results
Across multiple market regimes and two LLM backbones, EvolveTrade often improves Sharpe Ratio and Cumulative Return over fixed-policy LLM baselines, with best SR and CR in three model-regime pairs.
Takeaways & Limitations
Adapting the reusable procedure governing tool use can improve risk-adjusted trading performance and activate regime-relevant analysis in LLM agents.
Takeaways & Limitations
Policy-update intervals remain fixed within each run, and transaction-cost estimates omit slippage, market impact, and liquidity constraints.
Abstract
from arXiv · showhide
Large language model (LLM) trading agents can combine market data, news, and executable analysis, but their behavior is often controlled by static hand-written tool-use policies that are fixed before deployment. This limits their ability to adapt how they gather evidence, invoke tools, verify signals, and manage risk under changing market regimes. We introduce EvolveTrade, a self-evolving framework that treats the system prompt of a tool-using trading agent as a text-parameterized policy. After each update interval, a Policy Agent revises this policy using accumulated decision traces and realized portfolio feedback, while keeping the backbone LLM fixed. The updated policy is then used for the next batch of trading decisions, enabling the agent to refine its information-acquisition and portfolio-construction procedure over time. Experiments across multiple market regimes and two LLM backbones show that EvolveTrade often improves Sharpe Ratio and Cumulative Return over fixed-policy LLM baselines, achieving the improved SR and CR in most evaluated settings. Behavioral analyses further show that self-evolved policies increase code-mediated analysis and activate regime-relevant computations; case-level policy-to-return attributions trace how policy-induced allocation changes contribute to realized return differences. These results suggest that adapting the reusable procedure governing tool use is a key direction for building more robust LLM trading agents.
1 Introduction
LLM trading agents must repeatedly gather, verify, and act on noisy information under delayed feedback, yet existing tool-use procedures are typically fixed before deployment. EvolveTrade instead refines the policy text from realized trading experience, often improving risk-adjusted performance while changing tool-use behavior.
- LLM trading agents must acquire relevant information, verify noisy signals, manage portfolio risk, and act under delayed feedback in non-stationary markets.
- Existing trading approaches generally fix either the observation channel or the tool-use behavior at development time.This leaves the information-acquisition procedure unchanged after deployment.
- EvolveTrade refines a tool-using trading policy from realized trades and per-asset reasoning while keeping the underlying LLM and tools fixed.A Policy Agent updates the policy text between trading intervals, changing tool priorities, signal cross-validation, or exposure adjustments.
- Across two backbones and six post-cutoff market regimes, EvolveTrade achieves the best Sharpe Ratio and Cumulative Return among LLM methods in three model-regime pairs.Fixed-policy agents remain stronger in the other two reported pairs.
- Self-evolution increases code-mediated analysis and activates regime-relevant computations such as VaR, signal normalization, EMA, RSI, and SMA.A January 2025 NVDA drawdown attribution links a refined sizing policy to a +1.33 percentage-point relative daily return difference.
2 Related Work
Related work has expanded LLM trading agents with memory, tools, multimodal signals, and deliberation, while benchmarks increasingly test realistic financial conditions. However, the operational procedures governing evidence acquisition, verification, and allocation generally remain static, motivating self-evolving agents.
- Recent LLM trading systems incorporate memory, external tools, multimodal inputs, real-time data, and reflective or multi-agent decision making.
- Financial benchmarks increasingly evaluate LLM agents under realistic information and market conditions, including noisy-source trust, ticker memorization, and temporal leakage concerns.
- Despite these advances, observation channels, prompts, role structures, and tool workflows are typically specified before deployment.The resulting operational policy for acquiring information, verifying evidence, and committing to allocations remains largely static.
- Self-evolving-agent research shifts tool-using agents from fixed tool access toward post-deployment adaptation using execution traces, feedback, or task outcomes.
3 Problem Setup: Tool-Using Trading Agent
The trading agent operates sequentially with a day-specific natural-language tool-use policy, prior portfolio state, and fixed toolset. It actively retrieves information, reasons through tool calls, and outputs a feasible allocation with an asset-level rationale.
- On each trading day, the agent is initialized with the current policy, previous portfolio state, and available toolset.
- The policy is a natural-language system prompt governing tool use, evidence verification, risk control, and output requirements, while the LLM and tool interfaces remain fixed.Static baselines hold the policy constant, whereas EvolveTrade varies it across days.
- The agent actively retrieves pre-decision information through price, news, and Python code tools rather than receiving market information pre-injected into the prompt.All retrieval tools enforce a common temporal cutoff against look-ahead leakage.
- The allocation is executed after decision time and evaluated over the next holding interval within the portfolio constraint set.Experiments use a long-only allocation simplex over tradable assets and cash unless otherwise specified.
- The frozen LLM interleaves reasoning with tool invocations and terminates by emitting a feasible portfolio allocation and per-asset decision rationale.That recorded decision trace is passed to the policy-refinement step.
4 Motivation: Static Policy Locks the LLM into a Fixed Analytic Coverage
Across distinct market regimes, the fixed policy produces a nearly invariant analytic vocabulary and repeated retrieval patterns, leaving regime-relevant computations unused. This lack of adaptation motivates revising the tool-use procedure from trajectories and post-trade feedback.
- Observed Static-Policy Behavior: Figure 2 maps daily metric usage by regime, with frequency-ordered rows and right-side counts showing persistent repetition of the same analytic metrics.Regime-relevant metrics occupy the nearly unused bottom rows.
- Experimental Scope: The three diagnostic periods cover January 2025 sideways conditions, April 2025 drawdown with V-shaped recovery, and September 2025 steady growth.
- Observed Static-Policy Behavior: The static agent uses the same five metrics on 90–100% of days across sideways, drawdown-recovery, and uptrend regimes, while RSI, EMA, VaR, and signal normalization never appear.News queries remain per-asset boilerplate and price lookbacks stay near six months.
- Why Self-Evolution Is Needed: The fixed prompt applies the same analytical scope to a −9.8% drawdown and a steady uptrend despite different informative risk and trend indicators.
- Why Self-Evolution Is Needed: Because the prompt is fixed before market data is observed, earlier trades cannot change which tools are invoked or which metrics are computed.A self-evolving procedure is therefore designed to revise tool use between trading days from trajectories and post-trade feedback.
5 EvolveTrade
EvolveTrade refines a tool-use policy between trading batches using decision traces and realized portfolio feedback. A fixed-period Policy Agent updates the full policy text, enabling later policies to incorporate experience into tool selection, evidence verification, risk control, and output structure.
- 5.1 Trading Experience as Policy Feedback: EvolveTrade combines each decision trace with post-trade feedback to identify procedural weaknesses and reusable safeguards for policy revision.The trace contains the emitted allocation and per-asset rationale, while feedback provides ex-post performance context; feedback alone is ambiguous because gains or losses may reflect noise or market shocks.
- 5.2 Online Policy Self-Evolution: At a fixed interval of N trading days, the policy remains constant within each batch and is updated only between consecutive batches.The resulting later-batch policy has been shaped by realized experience from earlier batches rather than remaining the initial prompt.
- 5.2 Online Policy Self-Evolution: A separate Policy Agent analyzes batch records and rewrites the full policy text, linking concrete performance observations to specific prompt lines.The update instruction favors grounded edits over generic rewrites.
- 5.2 Online Policy Self-Evolution: Refinement can change tool selection, query formulation, signal interpretation, evidence verification, risk control, and final-output structure.For example, repeated action on unconfirmed news signals may produce a cross-validation rule before increasing exposure.
- 5. EvolveTrade: Iterating these updates produces a sequence of self-evolved batch policies whose performance under repeated online refinement is evaluated over the trading horizon.The framework treats policy text as an editable variable updated from feedback.
6 Experiment
Across multiple regimes and two backbone models, EvolveTrade generally improves risk-return performance over static LLM policies while changing how agents use analytical tools. Longer-horizon results, regime-specific computations, and an NVDA case study link policy refinement to adaptive analysis and allocation.
- Main Result: EvolveTrade attains the best result for many metric-regime pairs, with GPT-5-mini strongest in January and September and Gemini-2.5-Flash strongest in November.These results indicate benefits beyond a single backbone model, although fixed-policy agents remain stronger in some evaluated settings.
- Main Result: EvolveTrade reaches a 6.84% CR with GPT-5-mini in September 2025 and, under Gemini-2.5-Flash in February 2026, achieves SR 2.75 and CR 2.92%, exceeding rule-based baselines in those comparisons.The reported comparisons extend beyond LLM baselines to traditional rule-based methods.
- Tool-Use Behavior: EvolveTrade activates regime-relevant metrics absent from the Static Tool-Calling baseline, including VaR and signal normalization during the April drawdown and SMA, EMA, and RSI during the September uptrend.It also drops some repetitive baseline metric calls when they are no longer informative.
- Tool-Use Behavior: Code execution rises from about 1.0 to 2.6–4.5 daily calls under EvolveTrade, while price retrieval stays near one call and news search does not increase overall.Together with the metric-level evidence, this indicates a shift toward executable analysis rather than merely broader tool access.
- Case Analysis: During the January 2025 NVDA drawdown, EvolveTrade reduced NVDA exposure from 10.7% to 2.9%, limiting the next-day loss to -0.03% versus -1.11% for the Static Tool-Calling Agent after NVDA fell 17.0%.The refined policy prioritized longer-horizon risk-adjusted evidence and portfolio-improvement and turnover checks when sizing the volatile position.
- Long-Horizon and Sensitivity Analysis: Over 50 trading days, EvolveTrade outperforms all baselines in SR and CR while its policy length grows and periodically contracts, consistent with continued adaptation without indefinite expansion.Update-interval sensitivity is non-monotonic: moderate intervals perform best on average SR and CR, whereas daily updates produce the lowest average SR and CR.
7 Conclusion
EvolveTrade treats an LLM trading agent’s system prompt as a policy for evidence gathering, tool use, risk management, and portfolio allocation. Periodic revision from tool-use trajectories and realized portfolio feedback adapts this procedure across market conditions.
- EvolveTrade periodically revises the system-prompt policy from tool-use trajectories and realized portfolio feedback as market conditions unfold.
- Across multiple market regimes and two LLM backbones, EvolveTrade improves risk-adjusted performance over static baselines.
- The evolved policies change tool-use behavior through more code-mediated signal checks and a case-level link between allocation changes and realized performance.
- The results suggest that robust LLM trading agents should continually refine how they act on evidence while adapting to changing markets.
Limitations
The evaluation uses fixed policy-update intervals and simplified execution-cost assumptions. Baselines and active strategies also operate under defined information, universe, and position-allocation constraints.
- The selected policy-update interval remains fixed throughout each run, although dynamically scheduled updates in response to market changes remain future work.
- Transaction-cost estimates use proportional costs but omit slippage, market impact, and liquidity constraints.
- Rule-based baselines use only historical prices and exclude news, LLM reasoning, tool selection, and policy refinement.
- Active strategies cap each asset at 1/15 and enter or exit equal-weight positions when signals change.
B Evaluation Metric Definitions
The evaluation reports daily-return-based measures of risk-adjusted performance, growth, drawdown, win rate, and volatility. These metrics characterize both portfolio outcomes and their variability over each evaluation window.
- Daily portfolio return is defined from the change in portfolio value between consecutive trading days.
- Sharpe Ratio measures risk-adjusted performance using mean daily return and return standard deviation without a risk-free-rate adjustment.
- Cumulative return measures total portfolio growth over the evaluation window.
- Maximum drawdown measures the largest percentage decline from a previous portfolio peak.
- Win rate is the fraction of trading days with positive portfolio return.
- Daily volatility is the sample standard deviation of daily returns, reported in percentage points.
C Main Results with Standard Deviation
Repeated three-run experiments preserve EvolveTrade’s main performance pattern across selected regimes and backbones. Its strongest average SR and CR occur in January and September with GPT-5-mini and November with Gemini-2.5-Flash.
- EvolveTrade achieves the strongest average SR and CR among LLM-based methods in January and September with GPT-5-mini and November with Gemini-2.5-Flash.The main results tables report mean±std across three runs for each backbone.
- Its standard deviations are generally comparable to those of other policy-evolving agents, with no systematic increase in run-to-run variability.
- EvolveTrade preserves its main performance pattern across repeated independent runs.
D Main Results with Trading Friction
After applying 10 bps transaction costs, EvolveTrade retains its strongest LLM-based performance rankings in the same settings as the main results. A single-day example illustrates how its code-mediated allocation procedure produces a return advantage.
- After 10 bps transaction costs, EvolveTrade retains the highest SR and CR among LLM-based methods in January and September with GPT-5-mini and November with Gemini-2.5-Flash.It also achieves the second-best SR and CR in February.
- The September 10 example evaluates both agents at the September 11 close and is designed to show the intermediate mechanism behind the realized return gap.The example does not claim access to September 11 information.
- EvolveTrade uses code for metric computation, signal calibration, target-weight construction, and validation rather than only as descriptive evidence.
F Underperformance in the April Regimes
EvolveTrade underperforms its strongest static baselines in both April windows, where substantially higher cash allocations coincide with weaker SR and CR. The results suggest that reduced equity exposure may have caused missed gains during April recoveries and bullish conditions.
- EvolveTrade underperforms the strongest static baseline in both April windows: GPT-5-mini in April 2025 and Gemini-2.5-Flash in April 2026.
- EvolveTrade’s average cash weight exceeds the selected baseline by 10.1 and 35.9 percentage points in the two April windows, versus at most 6.4 percentage points elsewhere.Table 7 compares EvolveTrade with Static Tool-Calling for GPT-5-mini and Static Base for Gemini-2.5-Flash.
- The April windows combine elevated cash allocation with weaker SR and CR, suggesting reduced equity exposure may have contributed to missed gains during the April 2025 recovery and April 2026 bullish period.