AlphaMaster: Dual-Chain Feedback for Scalable and Diverse Alpha Factor Discovery

Alpha factor mining aims to discover predictive and interpretable signals for cross-sectional returns, yet automated pipelines based on genetic programming, reinforcement learning, and large language models often converge to a narrow expression family, causing factor homogenization that produces redundant alphas, accelerates decay, and limits scalability. We propose AlphaMaster, an artificial intelligence–driven scientific discovery framework that formulates alpha mining as a closed loop of hypothesis generation, experimental evaluation, and feedback-guided refinement. AlphaMaster integrates reinforcement learning, Generative Flow Network sampling, large language model–based symbolic synthesis, and genetic programming in a dual-chain design (generation for diverse proposals; optimization for multi-objective refinement). Crucially, it enforces correlation-aware diversity to suppress redundancy while improving predictive strength and stability. On CSI 500/CSI 1000 (2010–2025) under realistic trading constraints, AlphaMaster achieves the best overall performance, ranking first on five of six metrics for CSI 500 and on all metrics for CSI 1000: CSI 500 information coefficient (IC) 0.050, annualized return (AR) 0.136, information ratio (IR) 1.464; CSI 1000 IC 0.069, rank information coefficient (RankIC) 0.091, AR 0.150, IR 1.410, demonstrating scalable and reliable factor discovery.