DeepAR is a methodology for producing accurate probabilistic forecasts using autoregressive recurrent networks. The authors propose a model that learns a global model from historical data of all time series in the dataset, leveraging the data from related time series to fit more complex models without overfitting. The model addresses the challenge of widely varying magnitudes in time series data by incorporating a negative Binomial likelihood for count data and special treatment for varying magnitudes. Extensive empirical evaluation on several real-world forecasting datasets shows that DeepAR achieves accuracy improvements of around 15% compared to state-of-the-art methods. The main contributions of the paper include an RNN architecture for probabilistic forecasting and empirical demonstrations of its effectiveness. DeepAR offers advantages over classical forecasting approaches, such as minimal manual feature engineering, the ability to produce calibrated forecast distributions, and the capability to handle items with little or no history. The model is trained using stochastic gradient descent and can handle missing observations by sampling from the conditional predictive distribution.DeepAR is a methodology for producing accurate probabilistic forecasts using autoregressive recurrent networks. The authors propose a model that learns a global model from historical data of all time series in the dataset, leveraging the data from related time series to fit more complex models without overfitting. The model addresses the challenge of widely varying magnitudes in time series data by incorporating a negative Binomial likelihood for count data and special treatment for varying magnitudes. Extensive empirical evaluation on several real-world forecasting datasets shows that DeepAR achieves accuracy improvements of around 15% compared to state-of-the-art methods. The main contributions of the paper include an RNN architecture for probabilistic forecasting and empirical demonstrations of its effectiveness. DeepAR offers advantages over classical forecasting approaches, such as minimal manual feature engineering, the ability to produce calibrated forecast distributions, and the capability to handle items with little or no history. The model is trained using stochastic gradient descent and can handle missing observations by sampling from the conditional predictive distribution.