Paper: arXiv 2406.18394

Authors: Hao Shi, Weili Song, Xinting Zhang, Jiahe Shi, Cuicui Luo, Xiang Ao, Hamid Arian, Luis Seco

Abstract

The complexity of financial data, characterized by its variability and low signal-to-noise ratio, necessitates advanced methods in quantitative investment that prioritize both performance and interpretability.Transitioning from early manual extraction to genetic programming, the most advanced approach in the alpha factor mining domain currently employs reinforcement learning to mine a set of combination factors with fixed weights. However, the performance of resultant alpha factors exhibits inconsistency, and the inflexibility of fixed factor weights proves insufficient in adapting to the dynamic nature of financial markets. To address this issue, this paper proposes a two-stage formulaic alpha generating framework AlphaForge, for alpha factor mining and factor combination. This framework employs a generative-predictive neural network to generate factors, leveraging the robust spatial exploration capabilities inherent in deep learning while concurrently preserving diversity. The combination model within the framework incorporates the temporal performance of factors for selection and dynamically adjusts the weights assigned to each component alpha factor. Experiments conducted on real-world datasets demonstrate that our proposed model outperforms contemporary benchmarks in formulaic alpha factor mining. Furthermore, our model exhibits a notable enhancement in portfolio returns within the realm of quantitative investment and real money investment.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces advanced neural network architectures (generative-predictive models) and dynamic weighting schemes with some mathematical notation, but the core challenge is implementation-heavy. The rigorous validation via real-world datasets, portfolio simulations, and financial metrics like IC/ICIR positions it for practical application.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Alpha Factor Mining & Combination"] --> B["Input: Financial Market Data"]
  B --> C["Stage 1: Factor Mining<br>Generative-Predictive Neural Network"]
  C --> D["Stage 2: Dynamic Factor Combination<br>Time-Aware Weight Adjustment"]
  D --> E["Output: Adaptive Formulaic Alphas"]
  E --> F["Outcomes<br>Higher Returns & Better Adaptability"]

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 AlphaForge, a two‑stage framework that mines formulaic alpha factors with a generative‑predictive neural network and dynamically combines them with time‑varying weights.

Source sentence
this paper proposes a two-stage formulaic alpha generating framework AlphaForge, for alpha factor mining and factor combination.

Assumes 3

  • Models are retrained annually with the preceding year as validation data. model
    Source sentence
    The model was retrained annually with updated data, using the year preceding the test year as the validation dataset, resulting in a total of five training sessions.
  • The label used for training is Ref(VWAP, -21)/Ref(VWAP, -1) - 1, assumed to reflect real‑world scenarios. data
    Source sentence
    We use ’Ref(VWAP, -21)/Ref(VWAP, -1) - 1’ as the label because it more closely reflects real‑world scenarios, although it may lead to changes in the metrics.
  • Trading is limited to changing at most 5 stocks per day to avoid excessive trading costs. execution
    Source sentence
    Additionally, a daily limit of changing a maximum of 5 stocks was imposed to avoid excessive trading costs.

Data

Backtest sample: 2018-01 – 2022-12 · CSI300 stocks · equities · daily

  • CSI300 · Qlib market-prices public
    2018-01-01 to 2022-12-31 · CSI300 constituents
    Source sentence
    We choose the CSI300 and CSI500 dataset because the constituent stocks of these two indices cover the majority of the market capitalization in China’s A‑share market.
  • CSI500 · Qlib market-prices public
    2018-01-01 to 2022-12-31 · CSI500 constituents
    Source sentence
    A real account was used with an investment of 3 million RMB in CSI500.
  • Training/validation splits synthetic unknown
    2010-01-01 to 2016-12-31 (train), 2017-0
    Source sentence
    The first training set, validation set and test set are respectively (2010-01-01 to 2016-12-31), (2017-01-01 to 2017-12-31) and (2018-01-01 to

Recognised sources: Qlib datasets

Costs

Partial Costs appear only as a robustness check, a flat fee without spread or impact, or outside the headline numbers.

A daily limit of changing at most 5 stocks is imposed to avoid excessive trading costs; no explicit cost model is reported.

Source sentence
Additionally, a daily limit of changing a maximum of 5 stocks was imposed to avoid excessive trading costs.

Finds 3

  • AlphaForge outperforms prior formulaic alpha approaches across metrics such as IC and RankIC. performance
    Source sentence
    As shown in Table 1, our method demonstrates superior performance across various metrics, including stock selection ability indicators such as IC and RankIC.
  • The dynamic combination model yields higher returns than the static version, confirming the benefit of time‑varying weights. performance
    Source sentence
    Our findings indicate that our predictive‑generative alpha factor mining method achieves superior results compared to the previous state‑of‑the‑art algorithm. Furthermore, the superior performance of the ”Dynamic” model over the ”Static” model confirms the efficacy of our dynamic factor combining approach.
  • In real‑money trading, the model generated a 21.68% higher excess return versus CSI500 over about nine months. performance
    Source sentence
    A real account was used with an investment of 3 million RMB in CSI500. Untill now, after approximately 9 months of investment, it has generated a 21.68% higher excess return compared to CSI500.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 17 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 12 papers in the archive (11 by independent authors): AlphaLogics: A Market Logic-Driven Multi-Agent System for … (used as a baseline); FactorBench: A Portfolio-Aware Benchmark for Automated … (uses its method or data); Alpha-R1: Alpha Screening with LLM Reasoning via … (uses its method or data); AlphaRLM: Aligning Language Models for Literature-Grounded … (uses its method or data); EFS: Evolutionary Factor Searching for Sparse Portfolio … (uses its method or data)

Code

https://github.com/DulyHao/AlphaForge — 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 (62)

  1. Blackboard Systems (1986)
  2. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education (1983)
  3. Classification Problem Solving (1984)
  4. New Ways to Make Microcircuits Smaller (1980) — link
  5. New Ways to Make Microcircuits Smaller—Duplicate Entry (1980)
  6. Strategic explanations for a diagnostic consultation system (1984) — link
  7. Strategic Explanations in Consultation—Duplicate (1983)
  8. Poligon: A System for Parallel Problem Solving (1986)
and 54 more
  1. Transfer of Rule-Based Expertise through a Tutorial Dialogue (1979)
  2. The Engineering of Qualitative Models (2021)
  3. Attention Is All You Need (2017) — link
  4. Pluto: The ‘Other’ Red Planet (2015)
  5. 101 formulaic alphas (2016)
  6. A generative neural network for maximizing fitness and diversity of synthetic DNA and protein sequences (2020)
  7. The cross-section of expected stock returns (1992)
  8. Factorvae: A probabilistic dynamic factor model based on variational autoencoder for predicting cross-sectional stock returns (2022)
  9. Generating Synergistic Formulaic Alpha Collections via Reinforcement Learning — → 2306.12964
  10. Alpha-GPT: Human-AI Interactive Alpha Mining for Quantitative Investment — → 2308.00016
  11. AutoAlpha: An efficient hierarchical evolutionary algorithm for mining alpha factors in quantitative investment (2020) — link
  12. AlphaEvolve: A Learning Framework to Discover Novel Alphas in Quantitative Investment (2021) — link
  13. Online detection of stealthy false data injection attacks in power system state estimation (2016)
  14. Xgboost: A scalable tree boosting system (2016)
  15. AlphaEvolve: A Learning Framework to Discover Novel Alphas in Quantitative Investment (2021)
  16. A review on time series forecasting techniques for building energy consumption (2017)
  17. Long short-term memory (1997)
  18. A closer look at invalid action masking in policy gradient algorithms (2020) — link
  19. 101 formulaic alphas (2016)
  20. Factor Momentum and the Momentum Factor (2022) — link
  21. Lightgbm: A highly efficient gradient boosting decision tree (2017)
  22. Research on Product Sales Forecasting Based on Online Reviews and Data Mining (2023) — link
  23. Sales forecasting for Chemical Products by Using SARIMA Model (2022) — link
  24. Spatio-temporal self-supervised learning for traffic flow prediction (2023)
  25. Customer churn prediction using composite deep learning technique (2023)
  26. Revisiting Stock Alpha Mining Based On Genetic Algorithm (2019)
  27. Revisiting Stock Alpha Mining Based On Genetic Algorithm (2019)
  28. You impress me: Dialogue generation via mutual persona perception (2020) — link
  29. Symbolic regression via neural-guided genetic programming population seeding (2021) — link
  30. Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients (2021) — link
  31. Quantitative equity portfolio management: modern techniques and applications (2007)
  32. Learning equations for extrapolation and control (2018)
  33. Mastering atari, go, chess and shogi by planning with a learned model (2020)
  34. Proximal policy optimization algorithms (2017) — link
  35. How are you feeling?: A personalized methodology for predicting mental states from temporally observable physical and behavioral information (2017)
  36. Finding Alphas: A quantitative approach to building trading strategies (2019)
  37. Symbolicgpt: A generative transformer model for symbolic regression (2021) — link
  38. Attention is all you need (2017)
  39. DeepTrader: a deep reinforcement learning approach for risk-return balanced portfolio management with market conditions Embedding (2021)
  40. Hist: A graph-based framework for stock trend forecasting via mining concept-oriented shared information (2021) — link
  41. Rest: Relational event-driven stock trend forecasting (2021)
  42. Qlib: An ai-oriented quantitative investment platform (2020) — link
  43. Stock price prediction via discovering multi-frequency trading patterns (2017)
  44. AutoAlpha: An efficient hierarchical evolutionary algorithm for mining alpha factors in quantitative investment (2020) — link
  45. A unified framework for deep symbolic regression (2022)
  46. Neural symbolic regression that scales (2021)
  47. Data-driven discovery of coordinates and governing equations (2019)
  48. Age-fitness pareto optimization (2010)
  49. Geometric semantic genetic programming (2012)
  50. Epsilon-lexicase selection for regression (2016)
  51. Multiple regression genetic programming (2014)
  52. Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients (2021) — link
  53. Where are we now? A large benchmark study of recent symbolic regression methods (2018)
  54. A Unified Framework for Deep Symbolic Regression (2022) — link

2 of 62 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.