ARARMA vs ARIMA—A study of the benefits of a new approach to forecasting

Abstract

The ARARMA methodology of time series forecasting introduced by Parzen has compared well with longer established techniques such as Box and Jenkins ARIMA models in the results of a major forecasting competition. The two main differences between these methodologies are the way data is transformed to stationarity, and the interpretation of the concept of parsimonious models. The additional benefits familiarity with the ARARMA approach can bestow on those used to the ARIMA approach are identified. Following a description and demonstration of the ARARMA methodology, a comparison of forecasting performance is made. The comparison is carried out on sets of data for which published ARIMA models are available. To ensure that the comparison is of genuine forecasting ability (rather than fitting ability), a portion of data is saved solely for forecasting performance measurement. The results indicate that there are additional benefits to be gained from the ARARMA approach. The benefits lie in the approach's tendency to avoid over differencing and in the diagnostic tools for identifying the stationary ARIMA model, rather than in the different transformation to stationarity.