Fedformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting

Recent Transformer-based approaches have greatly advanced long-term time series forecasting, yet they suffer from two major limitations: high computational cost and difficulty in modeling the overall global structure (such as long-term trends). To overcome these issues, we introduce a framework that integrates Transformer architectures with seasonal–trend decomposition. The decomposition component extracts the global profile of the sequence, while the Transformer focuses on capturing finer local patterns. Moreover, recognizing that time series often have sparse representations in classical bases like the Fourier transform, we design a frequency-enhanced Transformer that exploits this property to strengthen long-horizon predictions. The resulting model, named Frequency Enhanced Decomposed Transformer (FEDformer), achieves both improved accuracy and linear complexity with respect to sequence length, making it more efficient than conventional Transformers. Experiments on six widely used benchmarks demonstrate that FEDformer lowers forecasting errors by 14.8% for multivariate series and 22.6% for univariate series compared with state-of-the-art baselines.