Damped trend exponential smoothing

Most time-series methods assume that any predicted trend will continue unchanged, whatever the forecast lead-time. Recent empirical studies, however, suggest that forecast accuracy can be improved by either damping or even ignoring altogether trends whose persistence is doubtful. A forecasting model is presented which is based upon the standard Holt-Winters models but in which the trend is damped. The model has been tested on the 1001 time-series of Makridakis et al. and shows considerable improvement on the usual versions based on linear trend projection, especially at longer lead-times. The model’s performance also compares well with the more sophisticated methods considered in the Makridakis study. Both seasonal and non-seasonal forms are considered and a simple procedure for automatic model selection is assessed.