DeepAR

DeepAR (Salinas, Flunkert, Gasthaus, Januschowski, 2020, International Journal of Forecasting) is Amazon's probabilistic forecasting model built on autoregressive LSTM/GRU networks. Instead of predicting point values, it learns a full predictive distribution per step: hidden states are mapped to distribution parameters (Gaussian, negative binomial, and others) and forecasts are generated by Monte Carlo sampling of trajectories, yielding quantiles and prediction intervals out of the box. Its signature idea is cross-learning: one model trains jointly across many related series (SKUs, stores, regions) with shared parameters plus static and dynamic covariates like price, promotions, and holidays.

DeepAR is the flagship model of Amazon's GluonTS library and powers a large share of production demand-forecasting pipelines on AWS. It is Apache 2.0 licensed and also reimplemented in Nixtla's NeuralForecast. The sequential RNN design is its main cost: training and sampling are slower than parallel transformer and MLP alternatives, and the learned representations are hard to interpret.

Author-reported: outperforms ARIMA/ETS and other classical baselines on Amazon-scale retail and electricity datasets in the 2020 paper; independent Sundial-benchmark evaluation ranks it mid-pack among neural forecasters, with probabilistic scores (CRPS) competitive but point accuracy behind newer transformer models.