Dynamic Spectral Decoupling for Time-Series Forecasting

Abstract

While frequency-domain approaches have advanced time-series forecasting, many existing methods process spectral components through a unified pathway, despite substantial energy disparities across frequencies. High-energy components can dominate the learning process, making it difficult to capture weaker residual dynamics that may contain both predictable structure and stochastic fluctuations. To address this issue, we propose Dynamic Spectral Decoupling (DSD), a model-agnostic module that separates frequency components based on instance-wise energy concentration. DSD uses a Lorenz-curve-based adaptive mechanism to determine instance-wise separation boundaries, separates dominant components from the residual spectrum, and assigns the resulting components to separate prediction pathways. This design reduces interference between energy scales without assuming that all low-energy components are equally useful. Extensive experiments on five real-world datasets show that DSD improves the average performance of diverse forecasting backbones and outperforms existing model-agnostic methods. Our code is available at https://github.com/anthi7-google/DSD.