FactorMoE: a framework with mixture-of-experts and attention to dynamically combine alpha factors in quantitative investment

Financial markets are characterized by high volatility and low signal-to-noise ratios, placing stringent demands on quantitative investment methods. Formulaic alpha factors are valued for their interpretability but often lack adaptivity, whereas machine learning-based alpha factors deliver stronger predictive performance at the expense of transparency. Here we introduce FactorMoE, a deep-learning framework for the dynamic combination of formulaic alpha factors that reconciles this interpretability-performance trade-off. FactorMoE conditions learnable gating networks on market regime indicators and recent factor performance to dynamically activate and weight candidate alpha factors. A chained Mixture-of-Experts architecture coupled with Multi-Head Attention enables sequential expert refinement and the modelling of nonlinear, synergistic interactions among factors, allowing the framework to prioritize robust factor combinations as market conditions evolve. Empirical evaluation on recent real market data shows that the proposed framework outperforms contemporary state-of-the-art and baseline methods on both factor-selection and stock-picking tasks, validating its effectiveness and superiority. The code is available at the https://github.com/ZCJ66q/FactorMoE repository.