In defense of ARIMA modeling

A number of empirical studies published in the forecasting literature in the 1970’s and 1980’s have come to the conclusion that univariate ARIMA time series modeling (Box-Jenkins) is not a more accurate univariate time series forecasting method than some simpler and older alternatives, including various exponential smoothing methods. One specific study published in the International Journal of Forecasting in 1986 is examined in this paper. It is suggested that the conclusion of that study that ‘exponential smoothing is usually a better time series method than Box-Jenkins in forecasting department store sales’ is unjustified because of the very limited sample size in the forecasting comparison. It is further suggested that the study fails to illustrate one of the important strengths of the ARIMA modeling approach, the ability to go beyond the basic univariate model by considering interventions, calendar variation, outliers, or other real aspects of typically observed time series.