Research map / Label-efficient and robust learning / Recognition and long-tail
AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors
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.
Proposes AlphaForge, a two-stage framework that mines formulaic alpha factors with a generative-predictive network and combines them into a dynamically weighted linear "Mega-Alpha" that is reselected and refit each trading day.
Problem
Formulaic alpha factor mining methods produce inconsistent factors and use fixed combination weights that fail to adapt to time-varying factor performance in financial markets.
Approach
A generator trained against a learned surrogate predictor of factor fitness (with a diversity loss) mines a factor zoo; a linear combination model then re-ranks and selects factors by recent IC/ICIR each day and refits its weights, so factor composition and weights vary over time.
Key results
- On CSI300 the method reaches IC 4.40% and RankIC 5.89%, versus 2.55% and 3.88% for the best baseline (DSO).
- On CSI500 it reaches IC 2.84% and RankIC 5.57%, versus 2.05% and 4.48% for the Static variant.
- In a real money trade on CSI500 over about 9 months it reports a 21.68% excess return over CSI500.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AlphaForge instead of DSO | Formulaic alpha factor mining on CSI300 · CSI300 | IC (%): 2.55(0.69) -> 4.40(0.56) | Helped |
| use AlphaForge instead of DSO | Formulaic alpha factor mining on CSI300 · CSI300 | RankIC (%): 3.88(1.12) -> 5.89(0.69) | Helped |
| use AlphaForge instead of Static combination (ablation) | AlphaForge on CSI300 · CSI300 | IC (%): 2.43 (0.57) -> 4.40(0.56) | Helped |
| use AlphaForge instead of Static combination (ablation) | AlphaForge on CSI300 · CSI300 | RankIC (%): 3.67(0.46) -> 5.89(0.69) | Helped |
| use AlphaForge instead of Static combination (ablation) | AlphaForge on CSI500 · CSI500 | IC (%): 2.05 (0.29) -> 2.84(0.58) | Helped |
| use AlphaForge instead of Static combination (ablation) | AlphaForge on CSI500 · CSI500 | RankIC (%): 4.48 (0.46) -> 5.57(0.58) | Helped |
| set the factor pool size to 10 rather than the other tested pool sizes (ablation) | AlphaForge · CSI300 | IC | Helped |
Surprising findings
- Performance is non-monotonic in factor pool size, peaking at a pool size of 10 rather than the largest pool.
Limitations
- The generator's training objective remains the IC of individual factors, not the combined IC of the factor batch.
- Results are reported on Chinese A-share indices only and the real-money result covers about 9 months.
Experimental setup
Datasets: CSI300 CSI500
Benchmarks: CSI300 CSI500
Metrics: IC RankIC ICIR Rank ICIR annual excess return
Baselines: GP DSO RL XGBoost LightGBM MLP Static
Builds on: reinforcement learning alpha mining deep symbolic optimization Qlib
Closest papers by meaning
- AlphaPROBE: Alpha Mining via principled retriveval and on-graph biased evolution (NeurIPS 2026 · similarity 0.91)
- AlphaQCM: Alpha Discovery in Finance with Distributional Reinforcement Learning (ICML 2025 · similarity 0.88)
- R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization (NeurIPS 2025 · similarity 0.86)
- AlphaPareto: Formulaic Alpha Discovery with LLM-Guided Multi-Objective Reinforcement Learning (NeurIPS 2026 · similarity 0.86)
- FactorGCL: A Hypergraph-Based Factor Model with Temporal Residual Contrastive Learning for Stock Returns Prediction (AAAI 2025 · similarity 0.85)
- Market-GAN: Adding Control to Financial Market Data Generation with Semantic Context (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.