Temporal Fusion Transformer
Also known as: Temporal Fusion Transformer (TFT), TFT, interpretable multi-horizon forecasting transformer
The Temporal Fusion Transformer (TFT), introduced by Lim, Arık, Loeff and Pfister in 2021, is an interpretable deep learning architecture for multi-horizon time series forecasting. It combines variable selection, gating, multi-horizon attention and quantile outputs, processing static, past and known-future inputs together to produce multi-step forecasts.
Key highlights
- Handles static, observed-past and known-future covariates jointly in one model.
- Produces interpretable outputs: variable-selection weights and attention patterns can be inspected.
- Quantile outputs deliver prediction intervals, not just point forecasts.
- Native multi-horizon forecasting across many steps ahead.
- Assumption-light: no normality requirement, supports mixed continuous and categorical features.
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
Use TFT for multi-step forecasting on time series or panel data when you have well-defined historical and known-future covariates and enough training data — at least about 200 observations, with a GPU recommended. It is assumption-light (no normality requirement) and handles continuous and categorical inputs together. Below roughly 500 observations its variable selection and attention become unreliable; with very short series (under ~100) a Transformer-based model is not meaningful and classical methods such as ARIMA are safer.
Strengths & limitations
- Handles static, observed-past and known-future covariates jointly in one model.
- Produces interpretable outputs: variable-selection weights and attention patterns can be inspected.
- Quantile outputs deliver prediction intervals, not just point forecasts.
- Native multi-horizon forecasting across many steps ahead.
- Assumption-light: no normality requirement, supports mixed continuous and categorical features.
- Needs substantial training data; variable selection and attention are unreliable below about 500 observations.
- On very short series (under ~100) a Transformer-based forecast is meaningless — use ARIMA instead.
- A GPU is recommended, raising compute requirements relative to classical forecasting.
- Requires historical and future covariates to be explicitly defined and supplied.
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
How much data does TFT need?
The method is set up for at least about 200 observations, but its variable selection and attention only work reliably with more data — below roughly 500 observations the model becomes unstable, and below about 100 a Transformer-based forecast is not meaningful.
What makes TFT interpretable?
Its variable-selection networks expose how much each input contributes, and its multi-head attention exposes which past time steps the model relies on, so you can inspect the drivers of a forecast rather than treating it as a pure black box.
What is the difference between past and future covariates?
Past (observed) covariates are only known up to the present, while known-future covariates — such as holidays or planned events — are available in advance. TFT processes these as separate streams alongside static covariates.
Do I need a GPU?
A GPU is recommended. The architecture is computationally heavy, so training on CPU can be impractically slow for realistic series.
Sources
- 1.Lim, B., Arık, S. Ö., Loeff, N. & Pfister, T. (2021). Temporal Fusion Transformers for Interpretable Multi-Horizon Time Series Forecasting. International Journal of Forecasting, 37(4), 1748–1764.
- 2.Lim, B. & Zohren, S. (2021). Time-Series Forecasting with Deep Learning: A Survey. Philosophical Transactions of the Royal Society A, 379(2194), 20200209.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 1). Temporal Fusion Transformer. ScholarGate. https://scholargate.app/deep-learning/temporal-fusion-transformer