EvolveTrade
A framework that enables LLM-based trading agents to optimize strategies through self-evolution. It treats system prompts as evolvable text-parameterized strategies and iteratively refines them using decision trajectories and portfolio feedback.
Timeline (last 90 days)
EvolveTrade is a strategy optimization framework for LLM trading agents that treats system prompts as policies parameterized by text, enabling them to self-evolve.
EvolveTrade leaves the underlying LLM completely unchanged and only allows the strategy itself to evolve over time, without modifying model weights.
At the end of each update cycle, EvolveTrade uses a dedicated Policy Agent to revise the existing strategy based on decision trajectories and real portfolio feedback.
The research team validated EvolveTrade across multiple market environments and on two different underlying LLMs.
EvolveTrade outperforms fixed-strategy LLM baselines on both Sharpe ratio and cumulative return in most evaluated settings.
The self-evolved strategies significantly increase the proportion of code-mediated analysis, handling quantitative tasks more frequently by writing and executing code.
The evolved strategies activate specific computational logic relevant to the current market environment, adjusting their analytical approach for different market states.
The paper provides case-level policy-to-return attribution analysis, tracing how strategy-driven changes in asset allocation contribute to differences in returns.
The core insight of EvolveTrade is that rather than repeatedly fine-tuning the model itself, it is better to let the reusable processes governing tool use self-evolve.
EvolveTrade's robustness in live-capital environments, performance under extreme market conditions, and the overfitting risk of strategy evolution have yet to be validated.
All Facts (10)
EvolveTrade is a strategy optimization framework for LLM trading agents that treats system prompts as policies parameterized by text, enabling them to self-evolve.
50%UnverifiedEvolveTrade leaves the underlying LLM completely unchanged and only allows the strategy itself to evolve over time, without modifying model weights.
50%UnverifiedAt the end of each update cycle, EvolveTrade uses a dedicated Policy Agent to revise the existing strategy based on decision trajectories and real portfolio feedback.
50%UnverifiedThe research team validated EvolveTrade across multiple market environments and on two different underlying LLMs.
50%UnverifiedEvolveTrade outperforms fixed-strategy LLM baselines on both Sharpe ratio and cumulative return in most evaluated settings.
50%UnverifiedThe self-evolved strategies significantly increase the proportion of code-mediated analysis, handling quantitative tasks more frequently by writing and executing code.
50%UnverifiedThe evolved strategies activate specific computational logic relevant to the current market environment, adjusting their analytical approach for different market states.
50%UnverifiedThe paper provides case-level policy-to-return attribution analysis, tracing how strategy-driven changes in asset allocation contribute to differences in returns.
50%UnverifiedThe core insight of EvolveTrade is that rather than repeatedly fine-tuning the model itself, it is better to let the reusable processes governing tool use self-evolve.
50%UnverifiedEvolveTrade's robustness in live-capital environments, performance under extreme market conditions, and the overfitting risk of strategy evolution have yet to be validated.
50%