FDM-Net: A frequency-decoupled network with adaptive masking for time series forecasting

Multivariate time series forecasting is essential for applications such as financial analysis and environmental monitoring. However, achieving robust forecasting remains challenging due to two fundamental issues: (1) real-world time series exhibit heterogeneous temporal patterns across different frequency bands, where low-frequency components contain long-range structures while high-frequency components are dominated by local variations and noise; (2) inter-variable dependencies are easily distorted by spurious correlations introduced by noisy or unreliable channels. To address these challenges, we propose FDM-Net, a frequency-decoupled forecasting architecture that assigns specialized modeling strategies to different frequency bands. The low-frequency branch captures global temporal dependencies and stable channel interactions, whereas the high-frequency branch focuses on local pattern extraction and noise suppression. To further enhance variable-dependency modeling, FDM-Net incorporates an adaptive masking mechanism that filters unreliable channel relationships before attention aggregation. In addition, we introduce a hybrid wavelet–time loss that enforces consistency between the frequency-domain structure and time-domain predictions. Extensive experiments on multiple real-world benchmarks demonstrate that FDM-Net achieves state-of-the-art forecasting performance while maintaining high computational efficiency, making it a practical and robust solution for real-world deployments.