FactorMAD: A Multi-Agent Debate Framework Based on Large Language Models for Interpretable Stock Alpha Factor Mining

Abstract

In quantitative investment, alpha factor mining plays a crucial role in predicting stock returns. Traditional approaches rely on human experts to design factors based on financial intuition. To enhance the efficiency of factor mining, recent machine learning (ML) methods have driven a shift toward automated factor mining. However, these ML-based approaches often suffer from a lack of interpretability or are restricted by predefined mathematical operators, limiting their ability to mine novel and effective factors. In this paper, we propose FactorMAD, a multi-agent debate framework based on large language models (LLMs) for interpretable alpha factor mining. Unlike existing methods, our framework employs two specialized LLM agents that iteratively refine factors through structured debate, leveraging diverse prior perspectives and critiques. To enhance factor flexibility and expressiveness, we develop a code-based factor development pipeline that allows LLMs to directly produce implementable factor code instead of being restricted to predefined operator sets. Empirically, we conduct extensive experiments on real-world stock market data, demonstrating that FactorMAD significantly outperforms existing automated factor mining methods in predictive performance and trading profitability. Our findings highlight the potential of LLM-driven multi-agent debate as a powerful paradigm for mining high-quality alpha factors in financial markets.