Why didn't Box-Jenkins win (again)
This paper focuses on the forecasting performance of the Box-Jenkins methodology applied to the 111 time series of the Makridakis competition. It considers the influence of the following factors: (1) time series length, (2) time-series information (autocorrelation) content, (3) time-series outliers or structural changes, (4) averaging results over time series, and (5) forecast time origin choice. It is found that the 111 time series contain substantial numbers of very short series, series with obvious structural change, and series whose histories are relatively uninformative. If these series are typical of those that one must face in practice, the real message of the competition is that univariate time series extrapolations will frequently fail regardless of the methodology employed to produce them.
- Research Organization:
- Oak Ridge National Lab., TN (USA)
- DOE Contract Number:
- W-7405-ENG-26
- OSTI ID:
- 5905453
- Report Number(s):
- CONF-830682-2; ON: DE83014043
- Resource Relation:
- Conference: 3. international symposium on forecasting, Philadelphia, PA, USA, 5 Jun 1983; Other Information: Portions are illegible in microfiche products
- Country of Publication:
- United States
- Language:
- English
Similar Records
Price of gasoline: forecasting comparisons. [Box-Jenkins, econometric, and regression methods]
Working papers: applicability of Box Jenkins techniques to gasoline consumption forecasting