> ML_ALGORITHM // DEEPAR-AUTOREGRESSIVE-RECURRENT-FORECASTING_v1.0

DeepAR (Autoregressive Recurrent Networks)

Pioneering Amazon deep learning method for probabilistic forecasting that trains an autoregressive LSTM network across large cross-sectional collections of time series.

Deep Probabilistic Forecastingtime-series-forecastingblack-boxlarge (>100k)

Back to All Algorithms

Computational Complexity

Training Complexity:O(epochs * series * steps * LSTM_pass)

Inference Complexity:O(samples * horizon * LSTM_pass)

Hardware Profile

CPU Friendly:No

Requires GPU:Yes

Memory Footprint:moderate

Interpretability & Data

Interpretability Tier:black-box

Training Data Needs:large (>100k)

Interpretability Assessment

Outputs full Monte Carlo sample paths for probabilistic prediction intervals (p10, p50, p90).

Suitable Tasks & Supported Modalities

Suitable Tasks:

large scale demand forecastingprobabilistic forecasting

Supported Modalities:

time-series

Implementing Libraries

GluonTSAmazon Web Services (AWS) · v0.15.1

View Spec

pytorch-forecasting

Foundational Literature

DeepAR: Probabilistic Forecasting with Autoregressive Recurrent NetworksDavid Salinas, Valentin Flunkert (2020) · International Journal of Forecasting

Common Pitfalls & Warnings

Series with radically different scales cause gradient instability if per-series scaling heuristic (mean level normalization) is bypassed