Long-term Time Series Forecasting (LTSF) faces a fundamental challenge: capturing both local fluctuations and global trends across extended horizons. Existing frequency-based methods apply single-scale transformations globally, missing critical scale-dependent patterns that vary temporally in real-world data. We introduce MMFNet, which addresses this limitation through Multi-scale Masked Frequency Transformation (MMFT) - a novel approach that decomposes time series into multiple temporal scales and applies learnable frequency masks to adaptively filter relevant spectral components. Our method combines Discrete Cosine Transform (DCT)-based multi-scale decomposition with scale-specific adaptive masking, enabling the model to capture fine-grained patterns in short segments while preserving long-term dependencies in extended windows. Extensive evaluation across seven benchmark datasets demonstrates MMFNet's effectiveness: it achieves state-of-the-art performance on benchmark datasets, with up to 6.0% Mean Squared Error (MSE) reduction over existing methods, while maintaining computational efficiency comparable to lightweight linear models. The success of learnable spectral filtering over fixed frequency selection provides new insights for adaptive temporal modeling beyond traditional forecasting approaches.
