PatchTST
Also known as: PatchTST — Yama Tabanlı Zaman Serisi Transformer, patch-based time series transformer, channel-independent transformer
PatchTST is a patch-based Transformer architecture for time series forecasting, introduced by Nie and colleagues in 2023, that cuts each series into overlapping patches treated as tokens and processes channels independently. It balances computational efficiency with strong accuracy on long-horizon forecasting.
Key highlights
- Strong accuracy on long-horizon forecasting while staying computationally efficient.
- Patching shortens the token sequence, cutting the cost of self-attention versus point-wise Transformers.
- Channel-independent modelling keeps the network compact and curbs overfitting on multivariate data.
- Assumption-light: does not require normally distributed data and is well suited to long forecast horizons.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
A good fit for long-horizon forecasting and prediction on continuous time series with a reasonably long history — the source recommends at least about 200 observations and is comfortable with long forecast horizons. Patch length and overlap must be tuned, and a GPU is recommended for training. On short series (below about 200 points) the patch-based attention becomes meaningless and classical methods such as ARIMA are preferable; below about 500 points it tends to overfit, where Random Forest is a safer fallback.
Strengths & limitations
- Strong accuracy on long-horizon forecasting while staying computationally efficient.
- Patching shortens the token sequence, cutting the cost of self-attention versus point-wise Transformers.
- Channel-independent modelling keeps the network compact and curbs overfitting on multivariate data.
- Assumption-light: does not require normally distributed data and is well suited to long forecast horizons.
- Needs a fairly long series (at least about 200 points) and overfits below roughly 500.
- Patch length and overlap are sensitive hyperparameters that must be tuned.
- A GPU is recommended; training cost is higher than for classical forecasting models.
- As a deep model it offers little direct interpretability compared with ARIMA-style coefficients.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What is a patch in PatchTST?
A patch is a short contiguous segment of the series — a small window of consecutive time steps — that is embedded as a single token. Using patches instead of individual points shortens the sequence the attention mechanism handles and gives each token more local context.
Why model channels independently?
In multivariate data each variable (channel) is forecast with the same shared Transformer backbone rather than being mixed together. This keeps the model compact, reduces overfitting, and lets patterns learned on one channel transfer to others.
How much data do I need?
The source recommends at least about 200 observations. Below roughly 500 points the model tends to overfit, and below about 200 the patch-based attention becomes meaningless, so classical methods like ARIMA are preferable on short series.
Do I need a GPU?
A GPU is recommended for training. As a deep Transformer, PatchTST is more compute-intensive than classical forecasting models, and patch length and overlap should be tuned to the data.
Sources
- 1.Nie, Y., Nguyen, N. H., Sinthong, P. & Kalagnanam, J. (2023). A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. ICLR.
- 2.Zhou, T., Ma, Z., Wen, Q., Wang, X., Sun, L. & Jin, R. (2022). FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting. ICML.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 1). PatchTST. ScholarGate. https://scholargate.app/deep-learning/patchtst