Research map / Graph learning / Graph algorithms and clustering

AlphaPROBE: Alpha Mining via principled retriveval and on-graph biased evolution

NeurIPS 2026 · Graph learning › Graph algorithms and clustering · approach: Continuous DAG structure learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Frames alpha factor mining as navigation of a DAG of factor lineage, pairing a Bayesian factor retriever (prior + likelihood over diversity and child-sparsity) with an LLM multi-agent DAG-aware generator that uses ancestral traces, evaluated on CSI 300/500/1000 against 8 baselines.

Problem

Automated alpha factor mining lacks a global structural view: decoupled generation treats factors as isolated, and iterative evolution only refines local parent-child chains, leading to redundant search and limited diversity.

Approach

Model the factor pool as a DAG of lineage edges; a Bayesian retriever scores candidate parents by quality with depth and retrieval penalties times a likelihood of novelty and child sparsity, and an LLM Analyst/Execution/Validator generator conditions on the full ancestral trace to produce non-redundant descendants.

Key results

• On CSI 300, AlphaPROBE reaches IC 5.84 and ICIR 39.02 versus the best baseline AlphaSAGE IC 5.02 / ICIR 35.82.
• On CSI 500 portfolio construction, AlphaPROBE achieves AR 17.45% and SR 0.8262 versus R&D-Agent(Q) AR 12.36% and SR 0.5220 for the next best.
• Ablations show removing the Bayesian prior or likelihood, or replacing the retriever with random selection, lowers IC and ICIR across all three indices.

Surprising findings

• Removing the topology penalty raises RIC/RICIR on CSI 500 (e.g., RICIR 69.24 vs 73.18 full), suggesting the penalty trades off some rank stability.

Limitations

• Evaluated only on daily-frequency Chinese stock pools; high-frequency and fundamental factors left for future work.
• Relies on a single backbone LLM (DeepSeek V3.1), so generality across LLMs is not shown.
• Hyperparameters gamma and omega were set by sensitivity analysis on CSI 300 only.

Experimental setup

Datasets: CSI 300 CSI 500 CSI 1000

Benchmarks: CSI 300 CSI 500 CSI 1000

Metrics: IC ICIR RIC RICIR AR MDD SR

Baselines: Alpha158 GP AlphaGen AlphaForge AlphaQCM AlphaSAGE AlphaAgent R&D-Agent(Q)

Builds on: AlphaForge AlphaGen AlphaSAGE AlphaAgent R&D-Agent(Quant)

Closest papers by meaning

• AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors (AAAI 2025 · similarity 0.91)
• AlphaQCM: Alpha Discovery in Finance with Distributional Reinforcement Learning (ICML 2025 · similarity 0.86)
• AlphaPareto: Formulaic Alpha Discovery with LLM-Guided Multi-Objective Reinforcement Learning (NeurIPS 2026 · similarity 0.85)
• R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization (NeurIPS 2025 · similarity 0.85)
• FactorGCL: A Hypergraph-Based Factor Model with Temporal Residual Contrastive Learning for Stock Returns Prediction (AAAI 2025 · similarity 0.84)
• FinSearchComp: Towards a Realistic, Expert-Level Evaluation of Financial Search and Reasoning (ICLR 2026 · similarity 0.82)