Disentangled Dual-Granularity Learning for Market-Adaptive Stock Return Forecasting under Non-Stationary Environments

Deep learning has emerged as a prominent paradigm in stock return forecasting, demonstrating remarkable capabilities in extracting non-linear patterns from historical stock data. However, existing approaches often process stock features as a monolithic input with fixed temporal receptive fields. This structural inflexibility fails to distinguish transient microstructure fluctuations and persistent long-term trends, resulting in entangled representations that impede adaptation to non-stationary environments. To address these issues, we propose the Disentangled Dual-Granularity Learning Network (DDGL-Net) to decouple and integrate diverse temporal signals. Specifically, we first construct a Disentangled Dual-Granularity Encoder (DDGE) to isolate fine-grained short-term and coarse-grained long-term features via parallel streams. To fuse dual-granularity temporal representations, we introduce a Market-Adaptive Contextual Modulator (MACM) as a regime-aware filter, which leverages global market context to dynamically reweight granularity contributions. Recognizing that input modulation alone is insufficient for the non-stationary latent-to-return mapping, we devise a Hybrid Residual Mixture of Experts (HR-MoE) predictor to decompose forecasts into stable consensus and adaptive residuals, ensuring resilient performance across shifting regimes. Extensive experiments confirm that DDGL-Net outperforms state-of-the-art baselines on real-world stock datasets, validating the effectiveness of disentangled modeling and regime-aware adaptation.