Autoformer: Decomposition Transformer for Long-Term Time-Series Forecasting
Autoformer (Auto-Correlation Decomposition Transformer) · Also known as: Auto-Correlation Transformer, Decomposition Transformer, Series Decomposition Forecaster, Oto-Korelasyon Ayrışım Transformer
Autoformer is a deep learning architecture for long-term time-series forecasting, introduced by Wu et al. from Tsinghua University at NeurIPS 2021. It replaces the standard self-attention mechanism with an Auto-Correlation mechanism that exploits periodic dependencies in the frequency domain, and embeds a progressive series decomposition block throughout the encoder and decoder to separately model trend and seasonal components.
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
Autoformer is well-suited for long-horizon forecasting tasks (e.g., 336–720 steps) on time series that exhibit clear periodic structure, such as weather, electricity demand, or traffic flow. It assumes the series can be meaningfully decomposed into trend and seasonal parts. It may underperform on short-horizon tasks or highly irregular series. Consider Informer for extreme-length inputs, FEDformer for frequency-domain mixing, or TimesNet for 2-D temporal pattern modeling as alternatives.
Strengths & limitations
- Progressive series decomposition enables the model to separately refine trend and seasonal components at every layer
- Auto-Correlation mechanism reduces complexity from O(L²) to O(L log L) while capturing periodic dependencies
- Aggregation over top-K lags is more interpretable than standard softmax attention over individual time steps
- Encoder-decoder design with cross-series Auto-Correlation supports effective information transfer for multi-step forecasting
- Moving-average decomposition with a fixed or heuristic kernel may not cleanly separate trend and seasonality in all series
- Auto-Correlation assumes periodic recurrence; series with irregular or non-periodic dynamics may not benefit from lag-based aggregation
- The model remains computationally heavier than simple linear baselines that have since shown competitive accuracy on benchmark datasets
- Long-horizon accuracy can degrade on real-world datasets where the periodicity assumption is only partially valid
Frequently asked
How does Auto-Correlation differ from standard self-attention?
Standard self-attention computes pairwise similarity between every time step, scaling as O(L²). Auto-Correlation instead computes the cross-correlation between query and key sequences in the frequency domain using FFT, identifying which periodic lags are most informative and aggregating values at those lags. This reduces complexity to O(L log L) and explicitly exploits periodicity rather than arbitrary pairwise relationships.
Why is series decomposition embedded throughout the network rather than applied once at the input?
Applying decomposition only at the input produces a one-shot separation that cannot be corrected as the representation evolves. Embedding decomposition at every sublayer allows the network to progressively refine the separation: as each layer learns richer representations, residual trend signals that were missed earlier can be extracted, leading to cleaner seasonal and trend streams at the output.
Is Autoformer still competitive with more recent models?
Autoformer was a strong baseline at its release but subsequent work (PatchTST, DLinear, TimesNet) demonstrated that simpler architectures can match or exceed it on standard benchmarks. It remains widely cited as a foundational decomposition-based forecaster and is often included in comparative studies, but practitioners should evaluate it alongside recent lightweight alternatives before deployment.
Sources
- Wu, H., Xu, J., Wang, J., & Long, M. (2021). Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. NeurIPS, 34. link ↗
How to cite this page
ScholarGate. (2026, June 2). Autoformer (Auto-Correlation Decomposition Transformer). ScholarGate. https://scholargate.app/en/deep-learning/autoformer
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.
- ARIMAEconometrics↔ compare
- FEDformerDeep learning↔ compare
- InformerDeep learning↔ compare
- TimesNetDeep learning↔ compare