LLM-Guided Evolutionary Strategy Generation for Quantitative Trading

Di Zhang

Abstract: This paper proposes LLM-GA, a novel framework that integrates large language models (LLMs) with genetic algorithms (GA) for automated trading strategy generation. The system architecture comprises three synergistic modules: 1) a signal generator extracting technical, fundamental, and sentiment indicators; 2) an LLM-enhanced GA core that initializes seed strategies and performs semantically-aware crossover/mutation operations; and 3) an execution module forming a closed-loop adaptive system. Unlike traditional GA that randomly combines signals, our approach leverages LLMs' financial reasoning capability to maintain logical consistency during strategy evolution. Experiments based on historical data of the Chinese stock market in the past five years (2020-2024) show that, LLM-GA achieves superior risk-adjusted returns (Annualized Excess Return (AER)=12.3\%, Maximum Drawdown (MDD)=35.2\%) compared to baseline methods including vanilla GA, PSO, and ensemble learning. Ablation studies reveal that LLM-guided initialization improves starting strategy quality by 215\%, while semantic crossover reduces invalid strategies by 83.5\%. Despite performance gaps against RL methods (2-3\% lower AER), our method provides unique advantages in strategy interpretability and diversity, addressing critical limitations in black-box approaches like reinforcement learning. The work establishes a new paradigm for human-AI collaborative quantitative strategy development.