EvolveTrade gives LLM trading agents a self-updating policy
Most LLM trading agents run on a static, hand-written tool-use policy: fixed rules for when to pull market data, check news, run analysis, and manage risk, set before the agent ever trades. EvolveTrade treats that policy differently. It packages the trading agent's system prompt itself as a text-parameterized policy, something that can be rewritten rather than only followed.
A separate Policy Agent does the rewriting. After each update interval, it looks at the accumulated decision traces from the trading agent and the portfolio's realized results, then revises the system-prompt policy accordingly. The backbone LLM that actually makes trading decisions is never retrained or fine-tuned; only the text policy that steers it changes. The updated policy then governs the next batch of trading decisions, so the agent's approach to gathering evidence, using tools, and sizing positions keeps shifting as market conditions do.
The authors tested EvolveTrade across multiple market regimes and two LLM backbones. Against fixed-policy baselines, EvolveTrade often produced higher Sharpe Ratio and Cumulative Return, with the improvement holding in most of the evaluated settings, though the paper reports this as a qualitative pattern rather than giving specific improvement percentages. A behavioral analysis found that the self-evolved policies lean more on code-mediated analysis and switch on computations tied to the current market regime. The authors also built case-level attributions tracing how a specific policy-driven change in allocation shows up in the resulting return, connecting the policy updates to the outcomes they produce rather than leaving the mechanism a black box.
The paper does not name the market regimes or the two backbone models it tested, does not give the length of an update interval, and does not report a code or dataset release alongside the work.
Key facts
- EvolveTrade turns a trading agent's system prompt into a text-parameterized policy that a separate Policy Agent rewrites after each update interval, using accumulated decision traces and realized portfolio feedback, while the backbone LLM stays fixed.
- Tested across multiple market regimes and two LLM backbones, EvolveTrade often beat fixed-policy baselines on Sharpe Ratio and Cumulative Return, with the improvement holding in most evaluated settings.
- Behavioral analysis found self-evolved policies increase code-mediated analysis and activate computations relevant to the current market regime.
- Case-level policy-to-return attributions trace how a specific policy-induced allocation change contributes to the realized difference in return.
- The paper does not name the market regimes or the two LLM backbones tested, does not state the update-interval length, and reports no code or dataset release.
Why it matters
LLM trading agents are usually built on a tool-use policy that is written once and frozen before deployment, which caps how well they can adapt when market conditions change. EvolveTrade's contribution is to make that policy itself an object the system can revise: the agent's system prompt becomes a text-parameterized policy that a Policy Agent updates from trading outcomes, rather than a fixed script. The backbone LLM never changes, only the instructions steering it, which keeps the update cheap compared with retraining a model.
Who it affects
This is a research contribution aimed at people building or studying LLM-driven trading and portfolio-management agents, not a consumer or enterprise product. It speaks most directly to researchers working on tool-using agents more broadly, since the core idea, treating a system prompt as an updatable policy driven by outcome feedback, is not specific to trading.
How to use it
The paper describes the framework and its evaluation but does not mention releasing code, a dataset, or a trained model alongside it, so there is nothing here to install or run directly. The method itself does not require retraining: it works by having a Policy Agent rewrite the trading agent's system prompt at each update interval based on decision traces and portfolio feedback.
How solid is it
The claims rest on experiments across multiple market regimes and two LLM backbones, plus a separate behavioral analysis and case-level attribution study. The paper reports the Sharpe Ratio and Cumulative Return gains as holding in most evaluated settings, but only qualitatively; it does not give the actual magnitudes, so the size of the advantage over fixed-policy baselines cannot be judged from what is stated. The regimes and backbones used are also not named, which limits how far the result can be checked or reproduced from the text alone.
Risks and caveats
The paper withholds several details that would normally support a trading result: which market regimes and which two backbones were used, how long an update interval runs, and any numeric size for the reported improvements. As with any backtested trading strategy, gains measured this way do not automatically carry over to live markets, and the paper's own presentation stops short of the specifics needed to independently verify the claimed edge.
“adapting the reusable procedure governing tool use is a key direction for building more robust LLM trading agents”
— the paper's authors