EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents
Sehee Kim, Yumin Choi, Minki Kang, Sung Ju Hwang
KAIST · DeepAuto.ai
Published September 15, 2026 · on the feed September 17, 2026 · ▲ 39 · arXiv 2609.17632
The abstract
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.
The Brief
2:40A short podcast summary of the paper. Written and voiced by AI.
Transcript
EvolveTrade. Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents. By Sehee Kim and three others, from KAIST and DeepAuto dot ai. Published on the fifteenth of September. No code repository is linked to it.
Trading agents built on large language models can combine market data, news, and executable analysis. But their behaviour is usually governed by static tool-use policies, written by hand and fixed before deployment. That limits how well they can adapt when the market regime changes: how they gather evidence, invoke tools, verify signals, and manage risk all stay as they were first written. A market that shifts calls for a different way of working, and a policy written once has no way to follow it.
EvolveTrade is a self-evolving framework built on one reframing. The system prompt of a tool-using trading agent is treated as a policy parameterized by text. A policy that is text can be revised, and the backbone model underneath stays fixed while it is. The model's weights are never touched. Only the words that tell it how to behave are learned, and they are learned from the agent's own experience of trading.
After each update interval, a Policy Agent revises that policy. It draws on the decision traces accumulated so far, and on feedback from the portfolio's realized results. The updated policy then governs the next batch of trading decisions. Over time, the agent refines both how it acquires information and how it constructs its portfolio. The loop is the familiar one of learning from results, carried out in language. Trade under a policy, see what the portfolio did, rewrite the policy, and trade again under the new version. The revisions are aimed at the procedure itself, which steps to take and which tools to call, rather than at any single decision.
Experiments cover multiple market regimes and two LLM backbones. EvolveTrade often improves the Sharpe Ratio and the Cumulative Return over baselines whose policies are fixed, reaching better values of both in most of the settings evaluated. The word the authors use is often, not always: the improvement holds in most of the evaluated settings rather than all of them.
Behavioural analyses show what changes. The self-evolved policies make more use of analysis carried out through code, and they activate computations relevant to the current regime. Attributions at the level of individual cases trace how allocation changes induced by the policy contribute to the differences in realized return.
The results point the authors to one conclusion: the reusable procedure that governs tool use is a key thing to adapt if LLM trading agents are to become more robust. Rather than retraining the model, EvolveTrade rewrites the instructions it trades by, and lets the market's feedback decide the next draft.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A trading agent with a fixed rulebook meets a market that will not sit still.
You run a trading agent built on a language model. It can pull prices, search news and run code, guided by a long instruction prompt you wrote before deployment.
The road ahead
- 1 · The problem ~6 min A trading agent guided by a prompt written before deployment keeps working the same way when the market turns, and never learns from its trades.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Treat the agent's system prompt as a learnable text policy, rewritten by a separate Policy Agent from trading evidence while the model stays frozen.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Trade a batch under fixed policy text, collect traces and results, let a Policy Agent rewrite the text, and use it for the next batch.
- 6 · The experiment ~6 min You’ll guess: Across market regimes and two LLM backbones, how consistently did EvolveTrade beat fixed-policy baselines on Sharpe Ratio and Cumulative Return?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.