EvoAlpha: Evolutionary Alpha Factor Discovery with Large Language Models

Alpha factor discovery is a central challenge in quantitative finance, traditionally addressed by human experts or automated search methods such as genetic programming and evolutionary algorithms. These approaches often lack semantic guidance, leading to inefficient search and fragile results. We propose a language-model-guided evolutionary framework, where large language models (LLMs) act as intelligent operators to guide mutation, crossover, and selection of candidate factors. By embedding evolutionary instructions into prompts, the LLM leverages domain knowledge and backtesting feedback to generate interpretable and high-quality signals. We first validate the approach through static factor searching, showing that LLMs can iteratively refine factors in a controlled setting. We then evaluate the framework in sparse portfolio optimization, where LLM-generated factors are used to rank assets and construct portfolios under $\ell_0$ constraints. Experiments on multiple real-market datasets demonstrate consistent improvements in portfolio performance over traditional baselines, highlighting the promise of combining LLMs with evolutionary search for systematic factor discovery.