Machine Learning for Business Analytics, 2nd Edition

19 SMOOTHING AND DEEP LEARNING METHODS FOR FORECASTING

In this chapter, we describe a set of popular and flexible methods for forecasting time series that rely on smoothing. Smoothing is based on averaging over multiple periods in order to reduce the noise. We start with two simple smoothers, the moving average and simple exponential smoother, which are suitable for forecasting series that contain no trend or seasonality. In both cases, forecasts are averages of previous values of the series (the length of the series history that is considered and the weights that are used in the averaging differ between the methods). We also show how a moving average can be used, with a slight adaptation, for data visualization. We then proceed to describe smoothing methods that are suitable for forecasting series with a trend and/or seasonality. Smoothing methods are data driven and are able to adapt to changes in the series over time. Although highly automated, the user must specify smoothing constants, which determine how fast the method adapts to new data. We discuss the choice of such constants and their meaning. The different methods are illustrated using the Amtrak ridership series. We then describe basic deep learning methods suitable for forecasting time series: recurrent neural networks (RNN) and Long Short‐Term Memory (LSTM) networks.

Smoothing Methods in JMP: In this chapter, we will use the JMP Time Series platform (Analyze Specialized Modeling Time Series) to apply smoothing ...