ABSTRACT Forecasting financial time series requires models that can simultaneously filter noise, align to multiscale temporal structure, and exploit latent cyclical regularities. Classical econometric kernels lack the nonlinear expressivity to satisfy these demands, while recent neural predictors often treat all features and lags homogeneously, blurring the separation between fast order‐flow shocks, slow momentum drifts, and periodic rhythms. We introduce the dual‐branch spectral‐trend attention network (DB‐STAN), an end‐to‐end architecture that tackles these challenges along three dimensions. (i) A gated component‐wise attention mechanism assigns adaptive importance weights to individual indicators, shrinking estimation variance in feature‐rich, noise‐dominated regimes. (ii) A flux–momentum decomposition routes instantaneous flow variables and slower momentum cues through distinct convolutional encoders, preserving their heterogeneous temporal spectra prior to fusion. (iii) A hybrid temporal‐spectral module couples multiresolution convolutions with frequency‐domain filtering, reconciling abrupt shocks with longer horizon cycles and trends. Extensive experiments on daily equities (S&P 500 and Dow Jones) and minute‐level cryptocurrency data (Bitcoin) show that DB‐STAN cuts mean absolute percentage error by 33%–48% and mean absolute error by 40%–46% relative to the strongest deep learning and classical baselines, while boosting directional accuracy to 74.6% on the S&P 500 at a 60‐bps threshold. Paired Diebold–Mariano tests confirm all improvements at the level. Ablation studies confirm the orthogonal benefits of each module, while a theoretical analysis establishes Lipschitz continuity of the gating‐attention operator and derives generalization‐error bounds under adversarial noise. DB‐STAN thus narrows the gap between statistical efficiency and market realism, offering a transparent, modular blueprint for next‐generation deep financial forecasting.
Raman et al. (Tue,) studied this question.