AlphaSAS: Structure-Aware Neural-Symbolic Search for Automated Alpha Discovery

The automated mining of predictive signals, or alphas, is a central challenge in quantitative finance. While Reinforcement Learning (RL) and neural-symbolic methods have enabled automated alpha generation, existing frameworks face three key limitations. First, large-scale exploration over expression spaces introduces multiple hypothesis testing concerns, causing factors with strong historical performance to exhibit limited out-of-sample reliability. Second, existing approaches rely on sequential representations that inadequately capture the structural relationships governing mathematical expressions and factor interactions. Third, current methods do not effectively incorporate previous search outcomes, resulting in inefficient exploration of redundant or unstable alpha structures. To address these challenges, we introduce AlphaSAS (Structure-Aware Neural-Symbolic Search for Automated Alpha Discovery), a framework built upon three innovations: (1) a structure-aware neural-symbolic representation that captures mathematical expression properties and factor relationships; (2) an adaptive candidate evaluation mechanism that incorporates previous discovery outcomes during search; and (3) an exploration strategy that promotes complementary alpha generation while reducing redundant factor discovery. Experiments across multiple equity markets demonstrate that AlphaSAS discovers more robust, diverse, and persistent alpha portfolios, establishing a systematic framework for reliable automated alpha discovery.