Alpha Discovery via Grammar-Guided Learning and Search
Han Yang, Dong Hao, Zhuohan Wang, Qi Shi, Xingtong Li
TL;DR
AlphaCFG introduces a grammar-guided, efficient search framework for discovering interpretable alpha factors in quantitative finance, outperforming existing methods in efficiency and profitability.
Contribution
The paper presents AlphaCFG, a novel grammar-based approach that ensures syntactic validity and interpretability in alpha factor discovery, using a tree-structured search with Monte Carlo Tree Search.
Findings
AlphaCFG outperforms baselines in search efficiency.
AlphaCFG achieves higher trading profitability.
Framework applicable to various quantitative finance tasks.
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 unstructured and unbounded spaces. We present AlphaCFG, a grammar-based framework for defining and discovering alpha factors that are syntactically valid, financially interpretable, and computationally efficient. AlphaCFG uses an alpha-oriented context-free grammar to define a tree-structured, size-controlled search space, and formulates alpha discovery as a tree-structured linguistic Markov decision process, which is then solved using a grammar-aware Monte Carlo Tree Search guided by syntax-sensitive value and policy networks. Experiments on Chinese and U.S. stock market datasets show that AlphaCFG outperforms state-of-the-art baselines in both search efficiency and trading profitability.…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. Principled, Structured Search via Grammar: By leveraging domain-specific rules—such as operator arity, finance logic, and context-dependent operand constraints, the paper addresses a central limitation in prior factor mining efforts: managing the combinatorial explosion and redundancy of candidate spaces. 2. Integration of Reinforcement Learning and Neural MCTS: The formulation of alpha search as a Tree-Structured Linguistic MDP and the embedding of neural MCTS with Tree-LSTM–based networks o
1. The paper lacks a quantitative report (e.g., a 1% syntax error rate) to conclusively prove that all generated expressions are syntactically valid 2. The paper extends the grammar to α-CFG-Sem (Definition 3) by incorporating domain-specific semantic constraints. The designed constraints are heuristically sound and based on reasonable financial logic. However, the connection between these specific constraints and the formal property of "interpretability" is not rigorously proven. 3. Although S
- The use of a context-free grammar (CFG) grounded in financial semantics effectively constrains the search space, ensuring valid and interpretable formulas while reducing redundancy compared to unrestricted symbolic approaches. Modeling formula generation as a Tree-Structured Linguistic MDP (TSL-MDP) offers a clean theoretical lens for combining RL with symbolic expression search, justifying the use of MCTS and policy/value learning. - Using Tree-LSTM–based policy and value estimators enables p
- Semantic equivalence pruning lacks rigor and scalability analysis. The paper uses subtree-based similarity for pruning equivalent or redundant expressions but does not analyze its computational complexity or describe practical optimizations (e.g., hashing or canonicalization). Given the large candidate pool, this omission is a serious limitation. - Restrictive grammar and operator set. The experiments use only six primitive features and a small operator vocabulary, which raises concerns about
1. The paper introduces Context-Free Grammar (CFG) into the automatic discovery of alpha factors for the first time, combining linguistic grammar generation principles with reinforcement learning and Monte Carlo Tree Search (MCTS) to propose a unified framework with both theoretical innovation and practical significance. 2. By defining syntactic and semantic constraints through CFG, the generated alpha factors exhibit clear financial meaning and structural readability, significantly enhancing th
1. The problem formulation of alpha discovery could be further improved, as the current description may not provide sufficient clarity for readers who are less familiar with this domain. A more detailed and accessible problem definition would enhance the paper’s readability and impact. 2. The experiments involve numerous hyperparameters, whose tuning is inherently complex and requires extensive search to achieve optimal settings. 3. Both the grammar and reward designs are fixed and manually spec
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsStock Market Forecasting Methods · Complex Systems and Time Series Analysis · Financial Markets and Investment Strategies