Summary

In this chapter, we briefly looked at univariate time series forecasting techniques, such as ARIMA and exponential smoothing. However, as demand varies by multiple variables, it becomes important to model multi-variate series. DeepAR enables modeling of multi-variate series, along with providing probabilistic forecasting. While point estimates may work in some situations, probabilistic estimates provide better data for improved decision making. The algorithm works by generating a global model that is trained across a large number of time series. Each item or product across several stores and departments will have its own weekly sales. The trained model accounts for newly introduced items, missing sales per item, and multiple predictors that explain sales. With the LSTM network and Gaussian likelihood, DeepAR in SageMaker provides a flexible approach to demand forecasting...