Informer
Also known as: Informer — Uzun Dizi Transformer Tahmini, Informer transformer, ProbSparse attention forecaster
Informer is a Transformer-based model introduced by Zhou et al. in 2021 for long-sequence time-series forecasting, using a ProbSparse self-attention mechanism that lowers the computational complexity of the standard Transformer to O(L log L). It is built for problems that demand predictions across thousands of future steps.
Key highlights
- Reduces Transformer attention complexity to O(L log L), making very long sequences tractable.
- Designed specifically for long forecasting horizons beyond 96 steps.
- Handles multivariate continuous time series.
- Does not require normally distributed data.
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 Informer when you need long-horizon forecasting of a continuous time series — typically a prediction window beyond about 96 steps — and you have a long history of at least roughly 500 observations. It is well suited to multivariate series and applications such as energy, weather, and traffic, and it normally requires a GPU. It does not assume normally distributed data. On short series (below about 200 observations) it overfits, and below about 500 the ProbSparse attention cannot find reliable patterns, so simpler models such as ARIMA or Random Forest are safer.
Strengths & limitations
- Reduces Transformer attention complexity to O(L log L), making very long sequences tractable.
- Designed specifically for long forecasting horizons beyond 96 steps.
- Handles multivariate continuous time series.
- Does not require normally distributed data.
- Needs a long history — around 500 observations or more — to learn reliably.
- Requires a GPU for practical training.
- On short series (below about 200 observations) it overfits badly.
- Below about 500 observations the ProbSparse attention fails to find stable patterns.
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 is Informer different from a standard Transformer?
It replaces full self-attention with ProbSparse self-attention, which computes attention only for the queries that dominate the distribution. This lowers the complexity from quadratic to O(L log L), making very long input and output sequences feasible.
How much data do I need?
Informer is a deep model aimed at long sequences and expects a long history — roughly 500 observations or more. Below about 500 the attention mechanism cannot find reliable patterns, and below about 200 it overfits, so simpler models are preferable.
Do I need a GPU?
Yes. Like other deep forecasting Transformers, Informer is built to run on a GPU; training on CPU alone is impractical for the long sequences it targets.
When should I not use Informer?
Avoid it for short time series. Below about 500 observations consider Random Forest, and for very short series below about 200 observations ARIMA is a safer choice.
Sources
- 1.Zhou, H. et al. (2021). Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. AAAI.
- 2.Wu, H., Xu, J., Wang, J. & Long, M. (2021). Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting. NeurIPS 34.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 1). Informer. ScholarGate. https://scholargate.app/deep-learning/informer