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
2018-01 – 2020-12 · S&P500 · equities
- CSI300 market-prices unknown
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
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
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
Finds 3
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.
6 papers in the archive : AllocFlow: End-to-End Portfolio Trading Action Generation … ; FactorBench: A Portfolio-Aware Benchmark for Automated … ; AlphaPADI: Formulaic Alpha Discovery via Pool-Aware …
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)
- Learning gflownets from partial episodes for improved convergence and stability (2023)
- A theory of continuous generative flow networks (2023)
- Let the flows tell: Solving graph combinatorial problems with gflownets (2023)
- Biological sequence design with gflownets (2022)
- Markov chain monte carlo and variational inference: Bridging the gap (2015)
- A tutorial on energy-based learning (2006)
- Open graph benchmark: Datasets for machine learning on graphs (2020)
- DistDGL: Distributed graph neural network training for billion-scale graphs (2020)
and 50 more
- Cluster-gcn: An efficient algorithm for training deep and large graph convolutional networks (2019)
- Distance encoding: Design provably more powerful neural networks for graph representation learning (2020)
- Geometric deep learning: Grids, groups, graphs, geodesics, and gauges (2021) — link
- Neural message passing for quantum chemistry (2017)
- How Powerful are Graph Neural Networks?
- Graph attention networks (2017) — link
- Inductive representation learning on large graphs (2017)
- Multilayer perceptrons for classification and regression (1991)
- Graph convolutional networks for text classification (2019)
- The graph neural network model (2008)
- Flow network based generative models for non-iterative diverse candidate generation (2021)
- Trajectory balance: Improved credit assignment in gflownets (2022)
- Gflownet foundations (2023)
- Modeling relational data with graph convolutional networks (2018)
- Learning from Expert Factors: Trajectory-level Reward Shaping for Formulaic Alpha Mining — → 2507.20263
- LightGBM : A Highly Efficient Gradient Boosting Decision Tree (2017) — link
- XGBoost : A Scalable Tree Boosting System (2016) — link
- Generating Synergistic Formulaic Alpha Collections via Reinforcement Learning — → 2306.12964
- AlphaEvolve : A Learning Framework to Discover Novel Alphas in Quantitative Investment (2021) — link
- AutoAlpha : an Efficient Hierarchical Evolutionary Algorithm for Mining Alpha Factors in Quantitative Investment (2020) — link
- Alphaforge: A framework to mine and dynamically combine formulaic alpha factors — → 2406.18394
- AlphaAgent : LLM -Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay (2025) — link
- Can Large Language Models Mine Interpretable Financial Factors More Effectively? A Neural-Symbolic Factor Mining Agent Model (2024) — link
- AlphaQCM: Alpha Discovery in Finance with Distributional Reinforcement Learning (2025)
- A five-factor asset pricing model (2015)
- An empirical study of financial factor mining based on gene expression programming (2021)
- 101 formulaic alphas (2016)
- Qlib: An AI-oriented Quantitative Investment Platform (2020) — link
- $ \textbackslash text\ Alpha \ \textasciicircum 2$: Discovering Logical Formulaic Alphas using Deep Reinforcement Learning (2024) — link
- RiskMiner : Discovering Formulaic Alphas via Risk Seeking Monte Carlo Tree Search — → 2402.07080
- QuantFactor REINFORCE : Mining Steady Formulaic Alpha Factors with Variance -bounded REINFORCE — → 2409.05144
- Beyond the Black Box: Interpretability of LLMs in Finance — → 2505.24650
- Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead (2019) — link
- Explainable Artificial Intelligence (XAI) approaches for transparency and accountability in financial decision-making (2023)
- Explainable AI (XAI) for Credit Scoring and Loan Approvals (2025)
- Enhancing explainability in real-world scenarios: towards a robust stability measure for local interpretability (2025) — link
- Can i trust the explanations? investigating explainable machine learning methods for monotonic models — → 2309.13246
- Cumulated gain-based evaluation of IR techniques (2002)
- A Theoretical Analysis of NDCG Type Ranking Measures (2013) — link
- QuantBench: Benchmarking AI Methods for Quantitative Investment — → 2504.18600
- 101 Formulaic Alphas (2016) — link
- R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization — → 2505.15155
- AlphaEval: A Comprehensive and Efficient Evaluation Framework for Formula Alpha Mining — → 2508.13174
- Navigating the Alpha Jungle: An LLM-Powered MCTS Framework for Formulaic Factor Mining (2025) — link
- Alphaagent: Llm-driven alpha mining with regularized exploration to counteract alpha decay (2025)
- Chain-of-Alpha: Unleashing the Power of Large Language Models for Alpha Mining in Quantitative Trading (2025) — link
- Generating Synergistic Formulaic Alpha Collections via Reinforcement Learning — → 2306.12964
- AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors — → 2406.18394
- An empirical study of financial factor mining based on gene expression programming (2021)
- 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.