Time series forecasting with DeepAR

This is an improved (fixed) implementation of the timeseries forecasting algorithm DeepAR by Amazon. I am providing a clear implementation in a Jupyter Notebook and clean Cython 3, without requiring SageMaker.

• Original paper: https://arxiv.org/pdf/1704.04110
• Documentation by Amazon: https://docs.aws.amazon.com/sagemaker/latest/dg/deepar_how-it-works.html

What is DeepAR?

DeepAR is an algorithm developed by Amazon Research producing accurate probabilistic timeseries forecasts, based on training an auto regressive recurrent network model on a large number of related time series.

Know issues with selected python module versions

• Be careful with your pandas version as freq parameter has been recently deprecated for df.Timestamp()
• If you are using Google Collab to run the code with a TPU it might fail with error TypeError: cannot pickle 'generator' object, due to PyTorch's generator handling. Therefore, you might need to run in via a CPU instance itself