The Effect on Forecasting Accuracy of the Holt-Winters Method When Using the Incorrect Model on a Non-Stationary Time Series

Abstract

The Holt-Winters method is one of the most popular forecasting techniques for time series, particularly with trend and seasonal components. There are two variations of the Holt-Winters method depending on the nature or type of the seasonal component: additive and multiplicative, and the type of seasonality is required to select the appropriate one. Unfortunately, time-series data are sometimes ambiguous, which can lead to incorrect identification of the model resulting in erroneous predicted values. In this study, the effect on forecasting accuracy when using the incorrect seasonal model in the Holt-Winters method was considered. Ten simulated datasets, five of which contained additive seasonality and the other five multiplicative seasonality, were used to study the effect of using the incorrect model on the forecasting accuracy. Five real datasets, in which it was difficult to distinguish the type of seasonal component, were used in the experimental study. Each dataset was examined using both additive and multiplicative models while varying the three smoothing parameters of the Holt-Winters method from 0.1 to 1 in increments of 0.1. The forecasting accuracy was evaluated in terms of the mean-absolute-percentage error and the root-mean-squared error. The results confirm the significance of the correct identification of the type of seasonality. For the ambiguous time-series data in which identifying whether to apply the additive or multiplicative model is not simple, the results show that utilizing the multiplicative model achieved significantly higher accuracy than utilizing the additive model.