Improving long-horizon forecasting of neural networks via spectral features and decomposition

Long-horizon time series forecasting is commonly improved through increasingly specialized neural architectures, making the resulting advances difficult to transfer across forecasting models. We instead propose an architecture-independent representation that combines component-wise forecasting with selective spectral augmentation. A dominant seasonal period is first estimated from training data using a Hann-windowed periodogram, and Seasonal and Trend decomposition using Loess (STL) separates the series into trend, seasonal, and residual components. For each historical window, Fourier amplitude and phase are added as input channels to the trend and seasonal predictors without increasing the sequence length, while the residual predictor remains in the time domain. Three independent copies of the selected forecasting backbone predict the components, and their forecasts are summed in the original scale. We evaluate the representation using MLP, CNN, LSTM, and Transformer backbones on four datasets and five forecast horizons, producing 80 architecture–dataset–horizon settings. The proposed method obtains the lowest MAE and MASE in 68 settings and the lowest RMSE in 69 settings. Its overall mean MASE is 0.6913, corresponding to reductions of 20.1% relative to the Base model, 10.4% relative to Fourier-only augmentation, 12.9% relative to STL-only forecasting, and 18.4% relative to a larger Base model. It also achieves the lowest cross-dataset, cross-architecture mean MASE at every evaluated horizon and the best overall result within each architecture family. Parameter analysis shows that these gains cannot be reproduced by increasing backbone capacity alone, supporting structured input representation as a transferable alternative to architecture-specific redesign.