Paper: arXiv 2509.25055

Authors: Binqi Chen, Hongjun Ding, Ning Shen, Jinsheng Huang, Taian Guo, Luchen Liu, Ming Zhang

Abstract

The automated mining of predictive signals, or alphas, is a central challenge in quantitative finance. While Reinforcement Learning (RL) has emerged as a promising paradigm for generating formulaic alphas, existing frameworks are fundamentally hampered by a triad of interconnected issues. First, they suffer from reward sparsity, where meaningful feedback is only available upon the completion of a full formula, leading to inefficient and unstable exploration. Second, they rely on semantically inadequate sequential representations of mathematical expressions, failing to capture the structure that determine an alpha’s behavior. Third, the standard RL objective of maximizing expected returns inherently drives policies towards a single optimal mode, directly contradicting the practical need for a diverse portfolio of non-correlated alphas. To overcome these challenges, we introduce AlphaSAGE (Structure-Aware Alpha Mining via Generative Flow Networks for Robust Exploration), a novel framework is built upon three cornerstone innovations: (1) a structure-aware encoder based on Relational Graph Convolutional Network (RGCN); (2) a new framework with Generative Flow Networks (GFlowNets); and (3) a dense, multi-faceted reward structure. Empirical results demonstrate that AlphaSAGE outperforms existing baselines in mining a more diverse, novel, and highly predictive portfolio of alphas, thereby proposing a new paradigm for automated alpha mining. Our code is available at https://github.com/BerkinChen/AlphaSAGE.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; reported Sharpe ratio of 6.32 is implausibly high for daily or lower-frequency data; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)

Why this score: The paper introduces advanced concepts like Relational Graph Convolutional Networks (RGCNs) and Generative Flow Networks (GFlowNets), requiring deep mathematical understanding. Empirical results are backed by real-world data from Chinese and U.S. stock markets, and the code is publicly available on GitHub, indicating high implementation readiness.

Research Flowchart

  flowchart TD
  A["Research Goal: Automate mining of predictive,<br>diverse 'alpha' signals in finance"] --> B["Data & Inputs:<br>Equity market data &<br>pre-defined operations"]
  B --> C["Problem: Traditional RL fails<br>due to reward sparsity,<br>poor structure, & lack of diversity"]
  C --> D["Solution: AlphaSAGE Framework<br>Integrates: RGCN, GFlowNets, &<br>Multi-faceted Rewards"]
  D --> E["Process: Generate candidate formulas<br>using Structure-Aware GFlowNets<br>for robust exploration"]
  E --> F["Outcome: Robust Alpha Portfolio<br>High diversity, novelty, & predictive<br>power (vs. baselines)"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes AlphaSAGE, a structure‑aware, diversity‑seeking framework that combines a relational graph convolutional encoder, a GFlowNet generator, and a dense multi‑faceted reward to mine formulaic alphas.

Source sentence
we introduce AlphaSAGE (Structure-Aware Alpha Mining via Generative Flow Networks for Robust Exploration), a novel framework is built upon three cornerstone innovations: (1) a structure-aware encoder based on Relational Graph Convolutional Network (RGCN); (2) a new framework with Generative Flow Networks (GFlowNets); a

Assumes

No explicit assumptions found in the text.

Data

Backtest sample: 2018-01 – 2020-12 · S&P500 · equities

  • CSI300 market-prices unknown
    2010-01-01 to 2024-12-31 (train/val/test · Chinese market constituents
    Source sentence
    market ... The split of the training set/validation set/test set for the Chinese market is defined as follows: 2010-01-01 to 2020-12-31 / 2021-01-01 to 2021-12-31 / 2022-01-01 to 2024-12-31.
  • CSI500 market-prices unknown
    2010-01-01 to 2024-12-31 (train/val/test · Chinese market constituents
    Source sentence
    market ... The split of the training set/validation set/test set for the Chinese market is defined as follows: 2010-01-01 to 2020-12-31 / 2021-01-01 to 2021-12-31 / 2022-01-01 to 2024-12-31.
  • S&P500 market-prices unknown
    2010-01-01 to 2020-12-31 (train/val/test · U.S. market constituents
    Source sentence
    For the US market: 2010-01-01 to 2016-12-31 / 2017-01-01 to 2017-12-31 / 2018-01-01 to 2020-12-31.

Costs

Not modeled A trading or execution result is reported without costs.

Finds 3

Reported performance (AlphaSAGE): Sharpe 6.32 · annual return 19.5% · max drawdown 4.2%

  • AlphaSAGE achieves the highest correlation metrics (IC 0.079, ICIR 0.496, RIC 0.094, RICIR 0.583) and the best portfolio outcomes (annualized return 7.62%, max drawdown -17.3%, Sharpe 1.71) across CSI300, CSI500, and S&P500. other
    Source sentence
    AlphaSAGE ranks first on all correlation metrics, with notably higher ICIR/RICIR, and these gains translate into the best portfolio outcomes (highest annualized return, lowest drawdown, highest Sharpe).
  • Ablation studies show that replacing the sequence encoder with a GNN provides the largest single lift across correlation and risk metrics, highlighting the value of structure‑aware representations. other
    Source sentence
    Replacing the sequence encoder with a GNN provides the largest single lift across correlation and risk metrics, underscoring the value of structure-aware representations.
  • Adding the novelty reward improves signal quality and tradability by reducing redundancy among factors, while the entropy regularizer yields higher IC/RIC, AR, and Sharpe with controlled MDD. other
    Source sentence
    the novelty reward (NOV) raises both signal quality and trad‑ability by reducing redundancy among factors. Finally, the entropy regularizer (ENT) yields the best overall results—higher IC/RIC, AR, and Sharpe with controlled MDD.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 6 papers in the archive (6 by independent authors): AllocFlow: End-to-End Portfolio Trading Action Generation … (used as a baseline); FactorBench: A Portfolio-Aware Benchmark for Automated … (uses its method or data); AlphaPADI: Formulaic Alpha Discovery via Pool-Aware … (uses its method or data)

Code

https://github.com/BerkinChen/AlphaSAGE — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (58)

  1. Learning gflownets from partial episodes for improved convergence and stability (2023)
  2. A theory of continuous generative flow networks (2023)
  3. Let the flows tell: Solving graph combinatorial problems with gflownets (2023)
  4. Biological sequence design with gflownets (2022)
  5. Markov chain monte carlo and variational inference: Bridging the gap (2015)
  6. A tutorial on energy-based learning (2006)
  7. Open graph benchmark: Datasets for machine learning on graphs (2020)
  8. DistDGL: Distributed graph neural network training for billion-scale graphs (2020)
and 50 more
  1. Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks (2019)
  2. Distance encoding: Design provably more powerful neural networks for graph representation learning (2020)
  3. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges (2021) — link
  4. Neural message passing for quantum chemistry (2017)
  5. How Powerful are Graph Neural Networks?
  6. Graph attention networks (2017) — link
  7. Inductive representation learning on large graphs (2017)
  8. Multilayer perceptrons for classification and regression (1991)
  9. Graph convolutional networks for text classification (2019)
  10. The graph neural network model (2008)
  11. Flow network based generative models for non-iterative diverse candidate generation (2021)
  12. Trajectory balance: Improved credit assignment in gflownets (2022)
  13. Gflownet foundations (2023)
  14. Modeling relational data with graph convolutional networks (2018)
  15. Learning from Expert Factors: Trajectory-level Reward Shaping for Formulaic Alpha Mining — → 2507.20263
  16. LightGBM : A Highly Efficient Gradient Boosting Decision Tree (2017) — link
  17. XGBoost : A Scalable Tree Boosting System (2016) — link
  18. Generating Synergistic Formulaic Alpha Collections via Reinforcement Learning — → 2306.12964
  19. AlphaEvolve : A Learning Framework to Discover Novel Alphas in Quantitative Investment (2021) — link
  20. AutoAlpha : an Efficient Hierarchical Evolutionary Algorithm for Mining Alpha Factors in Quantitative Investment (2020) — link
  21. Alphaforge: A framework to mine and dynamically combine formulaic alpha factors — → 2406.18394
  22. AlphaAgent : LLM -Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay (2025) — link
  23. Can Large Language Models Mine Interpretable Financial Factors More Effectively? A Neural-Symbolic Factor Mining Agent Model (2024) — link
  24. AlphaQCM: Alpha Discovery in Finance with Distributional Reinforcement Learning (2025)
  25. A five-factor asset pricing model (2015)
  26. An empirical study of financial factor mining based on gene expression programming (2021)
  27. 101 formulaic alphas (2016)
  28. Qlib: An AI-oriented Quantitative Investment Platform (2020) — link
  29. $ \textbackslash text\ Alpha \ \textasciicircum 2$: Discovering Logical Formulaic Alphas using Deep Reinforcement Learning (2024) — link
  30. RiskMiner : Discovering Formulaic Alphas via Risk Seeking Monte Carlo Tree Search — → 2402.07080
  31. QuantFactor REINFORCE : Mining Steady Formulaic Alpha Factors with Variance -bounded REINFORCE — → 2409.05144
  32. Beyond the Black Box: Interpretability of LLMs in Finance — → 2505.24650
  33. Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead (2019) — link
  34. Explainable Artificial Intelligence (XAI) approaches for transparency and accountability in financial decision-making (2023)
  35. Explainable AI (XAI) for Credit Scoring and Loan Approvals (2025)
  36. Enhancing explainability in real-world scenarios: towards a robust stability measure for local interpretability (2025) — link
  37. Can i trust the explanations? investigating explainable machine learning methods for monotonic models — → 2309.13246
  38. Cumulated gain-based evaluation of IR techniques (2002)
  39. A Theoretical Analysis of NDCG Type Ranking Measures (2013) — link
  40. QuantBench: Benchmarking AI Methods for Quantitative Investment — → 2504.18600
  41. 101 Formulaic Alphas (2016) — link
  42. R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization — → 2505.15155
  43. AlphaEval: A Comprehensive and Efficient Evaluation Framework for Formula Alpha Mining — → 2508.13174
  44. Navigating the Alpha Jungle: An LLM-Powered MCTS Framework for Formulaic Factor Mining (2025) — link
  45. Alphaagent: Llm-driven alpha mining with regularized exploration to counteract alpha decay (2025)
  46. Chain-of-Alpha: Unleashing the Power of Large Language Models for Alpha Mining in Quantitative Trading (2025) — link
  47. Generating Synergistic Formulaic Alpha Collections via Reinforcement Learning — → 2306.12964
  48. AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors — → 2406.18394
  49. An empirical study of financial factor mining based on gene expression programming (2021)
  50. From Linear to Hierarchical : Evolving Tree -structured Thoughts for Efficient Alpha Mining (2025) — link

12 of 58 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.