ETSformer: Exponential Smoothing Transformers for Time-Series Forecasting
ETSformer (Exponential Smoothing Transformer) · Also known as: Exponential Smoothing Transformer, ETS Transformer, ETSformer forecasting model, Üstel Düzleştirme Transformatörü
ETSformer is a deep learning architecture for time-series forecasting introduced by Woo et al. in 2022. It integrates classical exponential smoothing principles directly into the Transformer framework by replacing standard self-attention with an exponential smoothing attention mechanism. The model decomposes a time series into level, growth (trend), and seasonal components, allowing it to leverage both the long-range dependency modeling of Transformers and the interpretable structure of statistical ETS models.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
ETSformer is well suited for univariate and multivariate time-series forecasting tasks where both long-range dependencies and structured trend-seasonal patterns are present, such as energy load, weather, traffic, or financial series. It assumes the series can be meaningfully decomposed into level, growth, and seasonal components. It may underperform relative to simpler linear baselines on datasets dominated by noise or lacking clear periodicity. Alternatives include Autoformer, PatchTST, or DLinear for different trade-offs between model complexity and accuracy.
Strengths & limitations
- Combines the representational power of Transformers with the interpretable structure of classical ETS decomposition
- Exponential smoothing attention provides a strong inductive bias for recency without sacrificing long-range modeling
- Frequency-domain seasonal modeling learns multiple periodicities in a parameter-efficient way
- End-to-end training allows all decomposition parameters to be optimized jointly with the forecasting objective
- Computational cost of the attention mechanism can be prohibitive for very long input sequences
- Decomposition assumptions may not hold for irregular, non-stationary, or heavily noisy series
- Subsequent benchmarks have shown that simpler linear models can outperform ETSformer on some standard datasets
- Hyperparameter choices (number of seasonal frequencies K, smoothing coefficients) can significantly affect performance and require careful tuning
Frequently asked
How does ETSformer differ from the standard Transformer?
The standard Transformer uses uniform-weight self-attention across all positions. ETSformer replaces this with an exponentially decaying attention that gives higher weight to recent time steps, and it adds an explicit structural decomposition into level, growth, and seasonal components. This inductive bias toward classical time-series structure is the defining architectural difference.
Can ETSformer handle multivariate time series?
Yes, the architecture can be extended to multivariate settings by applying the decomposition and attention mechanisms across multiple input channels. However, like many Transformer-based forecasting models, its advantage over simpler baselines in the multivariate setting depends heavily on the inter-variable correlations present in the specific dataset.
Is ETSformer still state-of-the-art?
Since its publication in 2022, several studies have shown that simpler linear or MLP-based models can match or exceed ETSformer on standard long-sequence forecasting benchmarks. It remains a significant architectural contribution for demonstrating how classical statistical priors can be embedded in Transformers, but it is not regarded as the universally best choice today.
Sources
- Woo, G., Liu, C., Sahoo, D., Kumar, A., & Hoi, S. (2022). ETSformer: Exponential smoothing transformers for time-series forecasting. arXiv preprint. link ↗
How to cite this page
ScholarGate. (2026, June 2). ETSformer (Exponential Smoothing Transformer). ScholarGate. https://scholargate.app/en/deep-learning/etsformer
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- AutoformerDeep learning↔ compare
- ETS ModelEconometrics↔ compare