Exponential smoothing algorithms are a class of widely used forecasting methods that were born on the basis of heuristic insight. Originally, they lacked a proper statistical background, unlike the more sophisticated time series models that we outline in Section 11.6. More recently, attempts to justify exponential smoothing have been put forward, but the bottom line is that they proved their value over time. Indeed, there is no consistent evidence that more sophisticated methods have the upper hand, when it comes to real-life applications, at least in demand forecasting. However, heuristic approaches are less well suited to deal with other domains, such as financial markets, which indeed call for more sophistication. Apart from their practical relevance, exponential smoothing algorithms have great pedagogical value in learning the ropes of forecasting. Methods in this class have a nice intuitive appeal and, unlike moving averages, are readily adapted to situations involving trend and seasonality. One weak point that they suffer from is the need for ad hoc approaches to quantify uncertainty in forecasts; always keep in mind that a point forecast has very limited value in robust decision making, and we need to work with prediction intervals or, if possible, an estimate of a full-fledged probability distribution. In the next two sections, we illustrate the basic idea of exponential smoothing in the case of stationary demand. We also point out a fundamental issue with exponential smoothing: initialization. Then, we extend the idea to the cases of trend, multiplicative seasonality, and trend plus seasonality.
Leave a Reply