AlgoEvolve: LLMs evolve trading strategies with meta-evolution loop

AlgoEvolve, presented by Dhruv Sharma and Gautam Shroff, extends the use of LLMs as semantic mutation operators from static coding benchmarks to the complex domain of algorithmic trading. The framework generates and evolves trading strategies expressed as Python code, evaluating them through rigorous testing protocols. A key innovation is a meta-evolutionary outer loop that evolves the prompts used for inner-loop program synthesis. This outer loop discovers improved search heuristics that balance exploration and exploitation, significantly reducing zero-trade failures.

In experiments, AlgoEvolve consistently outperformed initial human-designed instructions, demonstrating emergent regime-adaptive logic — strategies autonomously shift trading rules in response to market conditions. The work shows that LLM-based semantic evolution provides a viable, scalable approach for continual program synthesis in noisy, non-stationary, and highly discontinuous environments. This could transform how quantitative firms develop and maintain trading algorithms, enabling adaptive strategies that evolve with markets without manual intervention.