Intel / article

EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents

Source: arxiv.org
Author: Sehee Kim, Yumin Choi, Minki Kang, Sung Ju Hwang
Date: Published Sep 17, 2026

Context

Researchers built EvolveTrade, a framework where an trading 's , the instructions governing how it gathers evidence, calls tools, verifies signals, and manages risk, is treated as an updatable policy rather than fixed in advance (arxiv.org). After each trading interval, a separate 'Policy Agent' revises that prompt using the accumulated decision traces and realized portfolio feedback from the interval just finished, while the underlying model itself stays unchanged; the updated prompt then governs the next batch of trading decisions. Across multiple simulated market regimes and two different model backbones, this self-revising policy typically improved Sharpe ratio and cumulative return compared to a fixed-policy baseline agent. The authors also found the self-evolved policies shifted toward more code-mediated analysis and toward computations relevant to the current regime. This is a backtested finding about one experimental setup, not a live-market result, and nothing here constitutes trading advice.

Key takeaways · AI-distilled

The backbone LLM stays completely fixed in EvolveTrade — only the system-prompt policy is rewritten, by a separate Policy Agent using accumulated decision traces and realized portfolio outcomes, so the adaptation happens entirely in text rather than model weights.
Behavioral analysis shows self-evolved policies increase code-mediated analysis and activate more regime-relevant computations — the agent doesn't just get better returns, it changes how it reasons depending on market conditions.
The authors trace case-level policy-to-return attributions, identifying specific policy-induced allocation changes that produced measurable return differences, rather than reporting only aggregate performance.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
system prompt — The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.

Read the source arxiv.org