DeepAR
DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks · Also known as: DeepAR — Olasılıksal RNN Tahmini, probabilistic autoregressive RNN forecasting, Amazon DeepAR
DeepAR is Amazon's industrial forecasting model, introduced by Salinas, Flunkert and Gasthaus (2017; published 2020), that uses an autoregressive recurrent neural network to estimate the parameters of a probability distribution at each step, producing a confidence interval rather than a single point forecast. It can model many related time series jointly within one model.
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
A good fit for forecasting and prediction on continuous time-series or panel data with at least about 100 observations, and it is especially strong when you have many related series to model together. An appropriate likelihood (such as Gaussian or Negative Binomial) must be chosen to match the data, and a GPU is recommended for training. Below roughly 500 observations the distribution parameters cannot be estimated reliably and a model such as Random Forest is safer; below about 100 observations the autoregressive RNN is meaningless and a classical method such as ARIMA is preferable.
Strengths & limitations
- Produces full probabilistic forecasts with confidence intervals, not just point estimates.
- Models many related time series jointly in one model, sharing strength across series.
- Does not require normally distributed data and supports flexible likelihoods (Gaussian, Negative Binomial).
- Handles the cold-start problem better than per-series models by borrowing patterns from data-rich series.
- Needs a fairly large dataset; below about 500 observations the distribution parameters are estimated unreliably.
- Requires choosing an appropriate probabilistic likelihood up front to match the data.
- A GPU is recommended, so it is more compute-heavy than classical forecasters.
- On very short series (below about 100 observations) the autoregressive RNN gives meaningless forecasts.
Frequently asked
How is DeepAR different from ARIMA?
ARIMA fits one model to a single series and returns a point forecast (optionally with intervals from its own assumptions). DeepAR trains one recurrent network jointly across many related series and outputs the parameters of a probability distribution at each step, giving probabilistic forecasts. On a single short series ARIMA is usually the better choice.
How much data does DeepAR need?
It is data-hungry. With fewer than about 500 observations the distribution parameters cannot be estimated reliably, and below about 100 observations the autoregressive RNN is essentially meaningless; in those cases a method like Random Forest or ARIMA is safer.
Which probability distribution should I choose?
Match the likelihood to the data: a Gaussian distribution for real-valued series and a Negative Binomial for count data are the common choices. Picking a mismatched likelihood degrades the forecast.
Do I need a GPU?
A GPU is recommended. As a recurrent neural network trained over many series, DeepAR is more compute-intensive than classical forecasters, and GPU training makes it practical at scale.
Sources
- Salinas, D., Flunkert, V., Gasthaus, J. & Januschowski, T. (2020). DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks. International Journal of Forecasting, 36(3), 1181–1191. DOI: 10.1016/j.ijforecast.2019.07.001 ↗
- Salinas, D., Flunkert, V. & Gasthaus, J. (2017). DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks. arXiv:1704.04110. link ↗
How to cite this page
ScholarGate. (2026, June 1). DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks. ScholarGate. https://scholargate.app/en/deep-learning/deepar
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
- Conformal Prediction (Time Series)Econometrics↔ compare
- N-HiTSDeep learning↔ compare
- PatchTSTDeep learning↔ compare
- Random ForestMachine learning↔ compare