DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks

Trained an autoregressive LSTM network across thousands of related time series, predicting parameters of parametric predictive probability distributions.

Amazon Web Services flagship forecasting engine deployed in Amazon Forecast and GluonTS for retail inventory planning.

Assumptions

• Shared non-linear dynamics exist across diverse time series that can be pooled into a single global neural model with scale normalization

Limitations

• Auto-regressive rollouts during inference can be slow; error accumulation on long forecast horizons