EMAformer: Enhancing Transformer through Embedding Armor for Time Series Forecasting

The introduction of EMAformer marks a significant advancement in multivariate time series forecasting, particularly in enhancing the Transformer architecture, which has previously struggled against MLP-based models. By addressing unstable inter-channel relationships, EMAformer incorporates three key inductive biases: global stability, phase sensitivity, and cross-axis specificity. These innovations have led to state-of-the-art performance on 12 real-world benchmarks, with a notable reduction in forecasting errors—2.73% in MSE and 5.15% in MAE. This progress not only highlights the potential of Transformer-based approaches but also underscores the importance of continuous improvement in forecasting models, which are crucial across various sectors. The code for EMAformer is publicly available on GitHub, encouraging further exploration and application of this model in real-world scenarios.