Time series forecasting with Amazon’s DeepAR
We have already covered time series analysis and forecasting in Chapter 6, Time Series Analysis and Forecasting, and Chapter 7, Machine Learning-Based Approaches to Time Series Forecasting. This time, we will have a look at an example of a deep learning approach to time series forecasting. In this recipe, we cover Amazon’s DeepAR model. The model was originally developed as a tool for demand/sales forecasting at the scale of hundreds if not thousands of stock-keeping units (SKUs).
The architecture of DeepAR is beyond the scope of this book. Hence, we will only focus on some of the key characteristics of the model. Those are listed below:
- DeepAR creates a global model used for all the considered time series. It implements LSTM cells in an architecture that allows for training using hundreds or thousands of time series simultaneously. The model also uses an encoder-decoder setup, which is common in sequence-to...
