AlphaSAGE: Structure-Aware Alpha Mining via GFlowNets for Robust Exploration

Battle Count: 82

Directly addresses the core challenge of automated alpha mining in quantitative finance. Proposes a complete pipeline from alpha discovery to portfolio combination. Evaluated on major equity indices with standard quant metrics (IC, ICIR, Sharpe, drawdown). The diversity-seeking GFlowNet approach directly addresses the practical need for non-correlated alpha portfolios. Highly relevant for systematic trading strategy development.

Requires implementing GFlowNets with Trajectory Balance loss, RGCN encoder on ASTs, multi-component reward computation (IC, structure-aware, novelty), early stopping mechanism, and dynamic alpha pool management. The combination of graph neural networks with generative flow networks adds complexity. However, pseudocode is provided and code is open-sourced. Hyperparameter tuning involves multiple reward weights and scheduling functions.

Code is publicly available at GitHub. Detailed hyperparameter settings, pseudocode (Algorithm 1), feature/operator definitions, and evaluation metric formulas are provided. Multiple random seed experiments confirm stability. However, specific training infrastructure details and exact data preprocessing steps could be more explicit.