AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors

• Stock trend forecasting is difficult because stock data have low signal-to-noise ratios and considerable noise.Source: Introduction · p. 1
• Formulaic alpha mining seeks simple formulas, while complex alphas generated by deep learning often lack interpretability.Source: Introduction · p. 1
• Fixed factor subsets and weights can become ineffective or reverse as alpha factors’ stock-selection capabilities change over time.Source: Introduction · pp. 1–2
• AlphaForge uses a generative-predictive neural network to mine factors and a composite model to select and combine them dynamically.Source: Introduction · p. 2
• The generator learns a surrogate score distribution and uses gradients to find high-scoring factors in sparse score spaces, while adjusting its objective to reduce correlation with previously mined factors.Source: Introduction · p. 2
• The combination model reassesses factor performance each trading day, selects factors for the next day, and adjusts their weights; experiments and real investment provide evidence that factor timing can be profitable.Source: Introduction · p. 2

Alpha Factor Definition

• An alpha factor maps a day’s original feature matrix for n stocks to one factor value per stock, using open, high, close, low, volume, and vwap features.Source: Alpha Factor Definition · p. 2

Alpha Factor Metrics

• IC is the time-series average Pearson correlation between factor values and subsequent returns, while RankIC complements it because Pearson correlation is unstable.Source: Alpha Factor Metrics · p. 2

Formulaic Alpha

• Formulaic alphas are mathematical expressions over feature and optional constant operands, using unary, binary, and time-series operators, represented as Reverse Polish Notation for deep-learning networks.Source: Formulaic Alpha · p. 2

Factor Mining Model

• The factor-mining network pairs a predictor that models factor fitness with a generator trained to produce formulas with high predicted fitness.The predictor is trained on evaluated factors in a sample library; the generator's output is transformed into a one-hot formula representation.Source: Methodology · p. 2, Factor Mining Model · p. 3
• The generator maps noise through a differentiable masking and Gumbel-softmax step to produce one-hot factor formulas.The predictor is frozen when generator training shifts to maximizing its estimated fitness.Source: Factor Mining Model · p. 3
• To limit premature convergence, generator training penalizes correlation between factors from two noise samples and screens factors for fitness and overlap with the existing zoo.The screening criteria include IC, ICIR, and correlation with existing factors.Source: Factor Mining Model · p. 3

Alpha Combination

• Each trading day, the combination model ranks factors by recent performance, selects up to N, and fits a linear model to predict the current combination.Linear models are favored because nonlinear combinations are susceptible to overfitting in financial datasets.Source: Alpha Combination · p. 4
• Recent reassessment lets the combination adjust factor membership and weights as factor performance changes, while maintaining explainability.The paper identifies fixed weights as inadequate for promptly responding to changes in factor effectiveness.Source: Alpha Combination · p. 4

Overview

• The framework first mines a diverse set of high-quality, low-correlation factors into a fixed Factor Zoo, then uses updated history to form each day's Mega-Alpha.The Factor Zoo remains unchanged during inference and trading; the combination model reassesses its factors at each time step.Source: Overview · pp. 4–5

Experiments

• The experiments ask whether AlphaForge outperforms prior formulaic methods, how pool size affects performance, whether its components help, and how it performs in trading.The pool-size limits tested are 1, 10, 20, 50, and 100 factors.Source: Experiments · p. 5, Additional Experiments · p. 5

Data

• Testing on CSI300 and CSI500 covers 2018–2022, with annual retraining and the preceding year used for validation.The datasets' constituent stocks cover most of China's A-share market capitalization.Source: Data · p. 5

Compared Methods

• The benchmarks include GP, DSO, and RL for formulaic factor generation, plus XGBoost, LightGBM, and MLP.Each model was run five times to reduce random-seed effects.Source: Compared Methods · p. 5

Additional Experiments

• Additional evaluation removes dynamic combination for comparison and simulates trading over a continuous five-year dataset; industry survey and other supplementary experiments are in the supplement.Source: Additional Experiments · p. 5

Main Results

• AlphaForge outperforms all compared methods on CSI300 and CSI500.The comparison includes both formulaic and non-formulaic methods.Source: Main Results · p. 5, Table 1 · p. 5
• The reported comparison covers stock-selection indicators IC and RankIC.Source: Main Results · p. 5

Effect of the Pool Size

• The pool-size experiment tests [1,10,20,50,100], with the number of factors in Mega-Alpha capped by the pool size as its composition varies over time.Source: Effect of the Pool Size · p. 5, Figure 2 · p. 6

Ablation Study

• The predictive-generative mining method outperforms the prior state-of-the-art algorithm, and the full Dynamic model outperforms the Static variant.Static uses the mining model with RL-style composition; Dynamic is the full model.Source: Ablation Study · p. 6

Case Study

• The case study generates 100 factors and sets the composite model’s factor-pool limit to 10.Source: A case study · p. 6

A case study

• Only five of Day 1’s ten selected factors are used on Day 2, while Factor 3’s weight changes from -0.00014 to 0.00168.Tables 2 and 3 present the factors used on the two days.Source: A case study · p. 6, Table 2 · p. 6, Table 3 · p. 6

• Factor 1 represents whether high-price and volume trends match over the past 20 days; its negative weight is linked to a possible contrarian buying logic.Source: Interpret-ability of Alpha Factors · p. 6
• The factor -1*ts mean(volume,20) is inversely related to average 20-day volume and correlates strongly with market capitalization, so excessive weight can create small-cap exposure.The passage describes that exposure as dangerous and says managers may reduce the weight or use other methods to avoid it.Source: Interpret-ability of Alpha Factors · p. 6
• Managers can analyze factors’ meanings and return sources, then make adjustments and attributions based on their market insights.Source: Interpret-ability of Alpha Factors · p. 6

• The simulation uses the CSI300 stock pool and daily equal-weighted holdings of the 50 highest-scoring stocks, with at most five daily changes.Source: Simulated Trading and Real Money… · p. 7
• Over five simulated years, AlphaForge achieves the best net account value and strongest performance among the comparative models.Source: Figure 3 · p. 7, Simulated Trading and Real Money… · p. 7
• 21.68% higher excess return than CSI500 after about nine months is reported for the 3 million RMB real account.Source: Simulated Trading and Real Money… · p. 7

Formulaic Alpha Factors

• Genetic programming, mutual-information objectives, similarity filtering, and factor improvement preceded reinforcement learning for mining collaborative Alpha factors.The cited approaches include GPLearn, mutual information, IC filtering (Zhang et al. 2020a), and AlphaEvolv (Cui et al. 2021a).Source: Related Work · p. 7
• Current formulaic methods often use fixed combination weights and immutable mining objectives, without accounting for time-varying factor effects.Source: Related Work · p. 7

Machine Learning-based Alpha Factors

• Stock-return machine-learning methods range from time-series and tree models to HIST and FactorVAE, but often lack formulaic Alpha's interpretability.Source: Related Work · p. 7

Symbolic regression

• Symbolic regression work has advanced from genetic-programming variants to methods combining neural networks with symbolic regression.Examples include lexicase selection (La Cava, Spector, and Danai 2016), age-layered populations, and neural-network frameworks (Champion et al. 2019).Source: Related Work · p. 7

Discussion

• AlphaForge is focused primarily on stock investment, and its generator optimizes individual-factor IC rather than combined IC across a factor batch.The authors suggest exploring combined IC as a possible way to improve mining efficiency.Source: Related Work · p. 7

• AlphaForge uses deep-learning search and a surrogate fitness model to generate formulaic Alpha factors in sparse search spaces.Source: Conclusion · p. 8
• Its composite model adjusts component factors and weights at each time slice to adapt Mega-Alpha to market fluctuations.Source: Conclusion · p. 8
• Extensive experiments report better performance than previous formulaic-alpha methods, while realistic trading tests report higher profits and real-money excess returns.Source: Conclusion · p. 8

• The checklist marks conceptual method or pseudocode descriptions, pedagogical references, and distinctions between opinion, hypothesis, speculation, and objective results as present.Source: Check List · p. 10
• The checklist says dataset selection is motivated and existing datasets are cited and public; novel-dataset appendix and release items are marked not applicable.Source: Check List · p. 10
• It marks computational experiments, preprocessing code, metric descriptions and rationale, run counts, distributional analysis, and significance testing as documented.Source: Check List · p. 10
• It marks source code, planned public release, implementation comments, random-seed methods, infrastructure, and final and explored hyperparameters as documented.Source: Check List · p. 10
• It also marks parameter search ranges and selection criteria as reported.Source: Check List · p. 10