EvolveTrade: A Strategy Optimization Framework for Self-Evolving LLM Trading Agents

The Problem with Static Strategies

LLM-based trading agents have demonstrated unique advantages in recent years: they can integrate market data, news, and executable code analysis into a complete decision-making pipeline. However, these agents have a long-overlooked weakness — their behavior is almost entirely governed by manually crafted, fixed tool-use policies written before deployment.

This means that regardless of whether the market enters a bull run, a bear market, or a highly volatile sideways period, the agent's approach to gathering evidence, invoking tools, validating signals, and managing risk remains unchanged. In the fast-moving world of financial markets, this rigid behavioral pattern clearly limits an agent's adaptability. The EvolveTrade framework, proposed in a recent arXiv paper, directly addresses this pain point.

Treating the System Prompt as an Evolvable Strategy

EvolveTrade's core innovation lies in a clever shift in perspective: it treats the system prompt of a tool-use trading agent as a text-parameterized policy.

Traditional approaches either rely on manual, iterative prompt tuning or require expensive model fine-tuning. EvolveTrade takes a lighter and more sustainable path — keeping the underlying LLM backbone entirely unchanged while allowing the strategy itself to evolve over time.

Here's how the mechanism works: at the end of each update cycle, a dedicated "Policy Agent" steps in. It uses the decision traces accumulated during that period along with realized portfolio feedback to revise the existing strategy. The revised strategy is then applied to the next batch of trading decisions.

Through this loop, the agent continuously refines its information-gathering and portfolio construction processes, achieving truly experience-driven self-evolution — without ever touching the model weights.

System prompts are instruction texts injected into an LLM application at deployment time, defining the model's role, behavioral guidelines, tool-calling conventions, and other core parameters — essentially the agent's "operating manual." Unlike model weights, system prompts can be modified at any time without retraining the model, giving them natural plasticity. Treating them as "strategies" rather than "configuration files" means they can be iteratively optimized using optimization algorithms — and that's exactly what EvolveTrade does, transforming an engineering practice problem into a formally defined policy optimization problem. This paradigm, often called "prompt optimization," has attracted growing academic interest, though most work focuses on static tasks. Applying it to dynamic financial decision-making is one of EvolveTrade's key innovations.

Experimental Results: Outperforming Fixed-Strategy Baselines in Most Settings

The research team evaluated EvolveTrade across multiple market regimes and on two different LLM backbones. Results show that in most evaluated settings, EvolveTrade outperforms fixed-strategy LLM baselines on two key metrics:

• Sharpe Ratio (SR): A core metric for risk-adjusted returns, reflecting the excess return earned per unit of risk taken.
• Cumulative Return (CR): A direct measure of the strategy's overall profitability.

The paper emphasizes that these improvements are not isolated cases but appear consistently across "most evaluated settings," suggesting the self-evolution mechanism has a degree of generalizability rather than relying on the idiosyncrasies of a specific market or model.

The Sharpe Ratio, developed by Nobel laureate William Sharpe, is calculated as the strategy's excess return (over the risk-free rate) divided by the standard deviation of returns. It is one of the most widely used risk-adjusted performance metrics in quantitative finance: a higher Sharpe Ratio means more return earned per unit of risk. Comparing cumulative returns alone tends to ignore volatility — a high-return but wildly volatile strategy is often difficult to execute in practice, as drawdowns may trigger forced liquidations or emotional stop-losses. Evaluating both Sharpe Ratio and cumulative return simultaneously assesses a strategy's ability to "make money" and its "smoothness," providing a more comprehensive standard for judging whether a trading strategy is truly excellent.

Behavioral Analysis: Evolved Strategies Are Better at Computation

Beyond return metrics, the researchers also broke down the behavioral patterns of evolved strategies and uncovered several interesting findings.

Greater Preference for Code-Driven Analysis

Self-evolved strategies show a marked increase in code-mediated analysis. In other words, the agent learns to handle quantitative tasks more frequently by writing and executing code rather than relying on the language model's intuitive reasoning. This is especially important in finance, a domain that depends heavily on precise computation.

Code-mediated analysis refers to an LLM generating and executing code (e.g., Python scripts) during its reasoning process to complete computation-intensive tasks, rather than producing conclusions directly in natural language. This capability has been demonstrated in tool-augmented LLM frameworks such as ReAct and ToolFormer. In financial contexts, pure language reasoning is prone to numerical errors when handling technical indicator calculations, statistical tests, or portfolio optimization — whereas code execution guarantees computational precision and verifiability. The fact that evolved strategies proactively increase code invocation suggests the agent has "learned" from historical feedback that precise computation is important for improving decision quality — a meta-level self-optimization about how to analyze, not just what to analyze.

Activating Market-Regime-Specific Computation

Evolved strategies also "activate" specific computational logic relevant to the current market regime. This means the agent is not just learning general techniques — it is adapting its analytical approach to different market states, which is precisely the kind of adaptability that static strategies cannot achieve.

Traceable Policy-to-Return Attribution

The paper provides case-level policy-to-return attribution analysis, tracing how strategy-induced changes in asset allocation contribute to final return differences. This interpretability is critical for financial applications — it makes the "black box" decisions of the agent auditable and understandable.

Significance and Outlook

The core message of EvolveTrade is clear: rather than repeatedly fine-tuning a massive model, letting the reusable process that governs tool use evolve on its own may be a key direction for building more robust LLM trading agents.

This insight extends beyond trading. For any agent system that relies on tool invocation, treating prompts/policies as optimizable objects and continuously iterating via real-world feedback loops offers a low-cost, highly flexible evolutionary paradigm. Of course, as a newly published research result, its robustness in live-money environments, its behavior during extreme market conditions, and the potential overfitting risks introduced by strategy evolution all await further validation.

Regardless, EvolveTrade delivers a noteworthy answer to the frontier question of "self-evolving agents" — tested on the challenging, high-feedback battlefield of financial trading.