EvolveTrade develops agent trading policies autonomously: reviewing directives raises returns without modifying language models

Financial trading agents that rely on large language models face a structural operational dilemma when moving from simulated environments to live markets, because most of those systems depend on manually written directives and policies that remain fixed after deployment, limiting their ability to adapt to market volatility and changing systems. A new research paper authored by Sehee Kim, Yumin Choi, Min-ki Kang, and Sung Jo Hwang, published on the arXiv platform for computer science and artificial intelligence, described a technical framework called EvolveTrade, which aims to address this shortcoming by enabling the trading agent to autonomously and continuously adjust evidence-gathering policies, tool invocation, and risk management.

The engineering innovation in this framework treats the agent’s core directive, the structured text governing tool usage, as an operational policy with textual parameters that can be refined and periodically swapped.A dedicated policy agent intervenes after each predefined time interval to review these directives based on accumulated decision logs and the realized returns of the investment portfolio.The entire process runs while keeping the underlying language model unchanged, without additional training or weight adjustments, and the revised policy is then used to guide the next batch of investment decisions, giving the system flexibility to refine data-gathering mechanisms and position building over time.

Experimental test results covering multiple market systems and two core language models showed a clear advantage over baseline models with fixed policies.The EvolveTrade framework recorded a tangible improvement in the Sharpe ratio for risk-adjusted performance and in cumulative returns across most evaluated financial environments and trading conditions.Behavioral analyses showed that self-developed policies increased the agent’s reliance on code-driven executable analyses and enabled calculations suited to the prevailing market system, while detailed attribution mechanisms revealed how asset-allocation adjustments driven by the new policies translated into differences in realized returns.

This shift has practical implications for asset managers, quantitative investment funds, and fintech firms in Gulf and Egyptian markets, where institutions are racing to explore language-model capabilities for automated trading. Rather than expending resources on training custom models or relying on rigid directives that falter amid liquidity swings and regional energy market volatility, the findings allow engineering effort to focus on automating decision-log evaluation and self-adjusting system directives. If you are building an agent-based trading system, the operational priority is to equip the model with code-execution environments and to review policies based on actual performance, rather than depending on closed-form textual directives.

The research results demonstrate that developing procedures and governance rules for tool usage is the most effective path to building stable trading agents capable of protecting financial portfolios.Relying on a policy agent that learns from real returns bridges the gap between language-model capabilities and the rapidly increasing market complexity, without incurring the hefty computational costs of continual retraining.