DeepAR is a probabilistic forecasting method based on autoregressive recurrent networks, designed to produce accurate forecasts for time series data. The method leverages a large number of related time series to train a model that can handle varying magnitudes and skewed distributions, which are common in real-world forecasting scenarios. DeepAR uses a negative binomial likelihood for count data and incorporates special handling for time series with widely varying magnitudes. It is trained on historical data of all time series in the dataset and can generate probabilistic forecasts through Monte Carlo sampling, enabling accurate quantile estimates across different prediction ranges. The model also allows for forecasting items with little or no historical data, a challenge for traditional single-item forecasting methods. DeepAR outperforms existing methods in accuracy, particularly in datasets with power-law distributed scales, and demonstrates the effectiveness of deep learning in probabilistic forecasting. The model is trained using a weighted sampling approach to address the imbalance in data distribution and can handle both real-valued and count data through appropriate likelihood functions. DeepAR's ability to learn complex patterns such as seasonality and uncertainty growth over time makes it a powerful tool for forecasting in various domains, including retail and energy consumption. The method has been evaluated on several real-world datasets, showing significant improvements in forecast accuracy compared to state-of-the-art approaches.DeepAR is a probabilistic forecasting method based on autoregressive recurrent networks, designed to produce accurate forecasts for time series data. The method leverages a large number of related time series to train a model that can handle varying magnitudes and skewed distributions, which are common in real-world forecasting scenarios. DeepAR uses a negative binomial likelihood for count data and incorporates special handling for time series with widely varying magnitudes. It is trained on historical data of all time series in the dataset and can generate probabilistic forecasts through Monte Carlo sampling, enabling accurate quantile estimates across different prediction ranges. The model also allows for forecasting items with little or no historical data, a challenge for traditional single-item forecasting methods. DeepAR outperforms existing methods in accuracy, particularly in datasets with power-law distributed scales, and demonstrates the effectiveness of deep learning in probabilistic forecasting. The model is trained using a weighted sampling approach to address the imbalance in data distribution and can handle both real-valued and count data through appropriate likelihood functions. DeepAR's ability to learn complex patterns such as seasonality and uncertainty growth over time makes it a powerful tool for forecasting in various domains, including retail and energy consumption. The method has been evaluated on several real-world datasets, showing significant improvements in forecast accuracy compared to state-of-the-art approaches.