Alpha Discovery via Grammar-Guided Learning and Search

Battle Count: 84

This paper is directly and centrally relevant to quantitative trading. It addresses the core problem of alpha factor discovery, which is fundamental to quantitative investment strategies. The framework produces interpretable, formulaic alpha factors used for stock ranking, portfolio construction, and trading signal generation. The evaluation includes realistic backtesting with top-k/drop-n trading strategies, Sharpe ratios, and maximum drawdown metrics. The method is benchmarked against state-of-the-art factor mining approaches (AlphaGen, AlphaQCM) and ML-based prediction models.

The implementation involves multiple complex components: (1) designing and implementing a context-free grammar with syntactic and semantic constraints, (2) building a Tree-LSTM encoder with both N-ary and Child-Sum variants for different operator types, (3) implementing grammar-aware MCTS with adaptive PUCT selection, (4) training policy and value networks jointly with reinforcement learning, (5) implementing the search-learn-search loop with replay buffers, (6) computing tree similarity for diversity-aware value targets, and (7) the linear factor combination optimization. The interplay between grammar constraints, tree search, and neural network training adds significant engineering complexity.

Source code is publicly available on GitHub (https://github.com/HanYang544/AlphaCFG). Detailed hyperparameters, algorithm pseudocode, and evaluation protocols are provided in the appendix. However, the specific data preprocessing pipeline and some implementation details of the Tree-LSTM variants would require careful reading of the supplementary material. The use of standard financial datasets (CSI 300, S&P 500) aids reproducibility.