Exponential smoothing in the telecommunications data
Exponential smoothing methods gave poor forecast accuracy in Fildes et al.'s study of telecommunications time series. We reexamine this study and show that the accuracy of the Holt and damped trend methods can be improved by trimming the time series to eliminate irrelevant early data, fitting the methods to minimize the MAD rather than the MSE, and optimizing the parameters. Contrary to Fildes et al., we show that the damped trend is more accurate than Holt's method. Because most of the telecommunications series display steady trends, we test the Theta method of forecasting and a closely related method, simple exponential smoothing with drift. The Theta method proves disappointing, but simple exponential smoothing with drift is the best smoothing method for this data, giving about the same accuracy as the robust trend.