AlphaPROBE: Alpha Mining via Principled Retrieval and On-graph biased evolution

In one sentence

Summary

AlphaPROBE addresses the limitations of existing automated alpha mining methods, which typically treat factor discovery as either isolated generation events or local iterative refinements. These approaches lack a global structural view, leading to redundant searches and limited diversity in the factor pool. The paper reframes alpha mining as the strategic navigation of a Directed Acyclic Graph (DAG), where factors are nodes and evolutionary links are edges. This topological perspective allows the system to treat the factor pool as a dynamic, interconnected ecosystem rather than a static collection, enabling more efficient and robust discovery of predictive signals. The framework consists of two core components: a Bayesian Factor Retriever and a DAG-aware Factor Generator. The Retriever balances exploitation and exploration by calculating a posterior probability for each factor, considering its quality, depth in the graph, and retrieval frequency, while also assessing its potential contribution to the pool's diversity. The Generator uses an LLM-based multi-agent workflow (Analyst, Execution, Validator) that leverages the full ancestral trace of selected parent factors to produce context-aware, non-redundant optimizations. Experiments are conducted on three major Chinese stock market datasets (CSI 300, 500, 1000) using Deepseek V3.1 as the backbone LLM, comparing AlphaPROBE against eight baselines including expert-designed factors, reinforcement learning methods, and other LLM-based agents. Results demonstrate that AlphaPROBE significantly outperforms baselines in predictive accuracy (IC, RIC) and portfolio construction metrics (Sharpe Ratio, Maximum Drawdown) across all three datasets. The method shows enhanced stability against market regime shifts and faster convergence during training. Ablation studies confirm the necessity of the Bayesian retrieval mechanism and the DAG-aware generation process. The paper highlights that leveraging global evolutionary topology is essential for efficient alpha discovery, reducing redundant mutations and encouraging diverse factor expressions. The implementation is open-sourced, providing a reproducible pipeline for automated factor mining.

Research problem

The paper addresses the challenge of automated alpha factor mining in quantitative finance, specifically the inefficiency and redundancy inherent in current methods. Existing paradigms, such as Decoupled Factor Generation and Iterative Factor Evolution, fail to capture the global structural relationships between factors, treating them as isolated events or local chains. This lack of a global view leads to limited diversity and suboptimal search efficiency in discovering predictive signals.

Core contributions

• Introduction of AlphaPROBE, a framework that models alpha factor evolution as a Directed Acyclic Graph (DAG) to capture global structural relationships.
• Development of a Bayesian Factor Retriever that balances exploitation and exploration by evaluating factors based on quality, topological position, and potential contribution to pool diversity.
• Design of a DAG-aware Factor Generator using a multi-agent LLM workflow that leverages full ancestral traces to produce context-aware, non-redundant factor optimizations.
• Extensive empirical validation on three Chinese stock market datasets (CSI 300, 500, 1000) demonstrating superior predictive accuracy and return stability compared to eight competitive baselines.
• Open-sourcing of the AlphaPROBE implementation to facilitate reproducibility and further research in automated alpha discovery.

Method

• Modeling alpha mining as navigation on a Directed Acyclic Graph (DAG) where factors are nodes and evolutionary links are edges.
• Bayesian Factor Retriever calculates posterior probability for each factor, balancing quality (prior) and diversity/contribution (likelihood).
• Prior term includes normalized quality, depth penalty (to avoid over-optimization), and retrieval penalty (to encourage exploration).
• Likelihood term assesses value diversity (value, semantic, syntactic) for leaf factors and parent-child/inter-child sparsity for non-leaf factors.
• DAG-aware Factor Generator uses a three-stage multi-agent workflow: Analyst (strategy formulation), Execution (expression synthesis), and Validator (constraint checking).
• Generator leverages the full ancestral trace of the selected parent factor to ensure context-aware and non-redundant optimizations.
• Dynamic Factor Integrator re-selects recently effective factors to form a 'Mega' factor for portfolio construction.
• Evaluation using IC, ICIR, RIC, RICIR for predictive power and AR, MDD, SR for portfolio performance.

Results & findings

• AlphaPROBE achieves the highest IC (5.84%) and ICIR (39.02) on CSI 300, outperforming the best baseline (AlphaSAGE) by significant margins.
• On CSI 500, AlphaPROBE achieves an IC of 6.26% and ICIR of 52.39, surpassing all baselines.
• On CSI 1000, AlphaPROBE reaches an IC of 9.04% and ICIR of 70.49, demonstrating strong performance in small-cap stocks.
• Portfolio construction results show AlphaPROBE yields the highest Sharpe Ratio (0.4411) and lowest Maximum Drawdown (22.25%) on CSI 300.
• Backtesting reveals superior resilience during market stress, such as the 2023-2024 bear market and April 2025 tariff turmoil.
• Ablation studies confirm that removing the Bayesian retriever or DAG-aware generator leads to significant performance degradation.
• AlphaPROBE converges faster than LLM-based baselines, requiring fewer training iterations to achieve high predictive power.
• The method effectively reduces redundant factor mutations by leveraging ancestral traces in the DAG.

Data & experiments

• Datasets: CSI 300 (large cap), CSI 500 (mid cap), and CSI 1000 (small cap) Chinese stock markets.
• Time periods: Training (2010-2020), Validation (2021-2022), Test (2022-2025).
• Baselines: Expert-designed pool (Alpha158), DFG methods (AlphaGen, AlphaForge, AlphaQCM, AlphaSAGE), and IFE methods (GP, AlphaAgent, R&D-Agent(Q)).
• Metrics: Predictive power (IC, ICIR, RIC, RICIR) and Portfolio construction (Annualized Return, Maximum Drawdown, Sharpe Ratio).
• Backtesting: Cumulative return curves and drawdown analysis under market stress events.
• Ablation studies: Evaluating the impact of different retriever strategies and generator components.

Limitations & caveats

• Experiments are limited to Chinese stock markets (CSI 300, 500, 1000), limiting generalizability to other markets.
• The method relies on a specific set of base features and operators, which may not capture all possible alpha sources.
• The DAG structure and retrieval mechanism may become computationally expensive as the factor pool grows very large.
• The LLM-based generator's performance is dependent on the quality of the backbone LLM (Deepseek V3.1).
• The dynamic factor integrator's effectiveness in real-time trading environments requires further validation.
• The paper does not extensively discuss the interpretability of the generated factors beyond their mathematical expressions.

Curator notes

Discovery status: institutional_trading_round3_promoted. Source links: AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay; Can Large Language Models Mine Interpretable Financial Factors More Effectively? A Neural-Symbolic Factor Mining Agent Model; Navigating the Alpha Jungle: An LLM-Powered MCTS Framework for Formulaic Factor Mining