ENBIS-17 in Naples

9 – 14 September 2017; Naples (Italy) Abstract submission: 21 November 2016 – 10 May 2017

Robustness of the Classical and Recently Developed Exponential Smoothing Methods

13 September 2017, 10:50 – 11:10


Submitted by
Ales Toman
Aleš Toman (University of Ljubljana, Faculty of Economics)
Exponential smoothing methods are powerful tools for decomposing and smoothing time series and predicting their future values. Distinguished by their simplicity and computational stability, their forecasts are comparable to the forecasts of more complex statistical time series models. Classical smoothing methods can handle trend and seasonality in both, additive and multiplicative form, and the methods we developed recently (e.g. modified and extended Holt-Winters method) are adjusted even to substantial noise component and repeated zero values in the data.

In this presentation we compare the smoothing and forecasting performance of the classical and recently developed smoothing methods when time series contain irregularities such as symmetric or asymmetric outliers, level shifts or structural breaks. We define and justify a new symmetric relative smoothing (or forecasting) efficiency measure (SREM) that allows us to evaluate the performance of different smoothing methods based on unequal-sized groups of time series.

We start the analysis simulating time series with known trend and seasonal patterns and applying several smoothing methods to the series before and after we add the pre-specified irregularities to the data. The magnitude of changes in SREM values indicates the robustness of different methods with the methods undergoing largest changes being the most sensitive. Additionally we use real seasonal time series from the M3-Competition in place of simulated series to analyze the robustness of different methods when the true series pattern is not known. We conclude that robustness should in general not be overlooked by researchers when proposing and advocating new forecasting methods.
View paper

Return to programme