Alpha Discovery via Grammar-Guided Learning and Search

Han Yang, Dong Hao, Zhuohan Wang, Qi Shi, Xingtong Li

Keywords: Quantitative Finance; Formulaic Alpha Factors; Reinforcement Learning; Representation Learning

TL;DR: A novel framework for discovering syntactically valid, semantically interpretable and finacially profitable alpha factors

Abstract:

Automatically discovering formulaic alpha factors is a central problem in quantitative finance. Existing methods often ignore syntactic and semantic constraints, relying on exhaustive search over a unbounded and unstructured space that limits performance and interpretability.
We present AlphaCFG, the first framework for defining and discovering alpha factors that are syntactically valid, financially interpretable, and computationally efficient. First, AlphaCFG defines an alpha-oriented Context-Free Grammar to construct a tree-structured, size-controlled search space of human-interpretable alpha expressions, enabling grammar-tailored search and learning. We then formulate the search of high-performance alphas in this space as a very large, tree-structured linguistic Markov Decision Process (TSL-MDP), where each state is an alpha expression with its information coefficient as reward. To efficiently navigate the TSL-MDP, we develop syntax-similarity-based representation learning method to estimate alpha expression performance (value network) and grammar production rule probabilities (policy network), and integrate it into a grammar-aware Monte Carlo Tree Search. Experiments on China and US stock markets' datasets show that AlphaCFG outperforms state-of-the-art baselines in both search efficiency and trading profitability. AlphaCFG also provides an easy-to-use approach for refining and improving existing formulaic alpha factors.

Primary Area: reinforcement learning

Submission Number: 10538