Commentaries on "generalizing about univariate forecasting methods: Further empirical evidence"

J. Scott Armstrong, A. B. Koehler, R. Fildes, M. Hibon, S. Makridakis, N. Meade

Research output: Contribution to journalReview articlepeer-review

Abstract

This set of comments addresses issues raised by Fildes et al. (1998) ["Generalizing about univariate forecasting methods: further empirical evidence", International Journal of Forecasting, 14, 339-358]. Armstrong highlights the issue of the breadth of conditions that should be considered when comparing different (univariate) forecasting methods and the generalizability of the findings. He criticizes Fildes et al. for their failure to include a wider segmentation of the two data bases used: the Makridakis data and the telecoms data. Koehler also criticizes the particular measures used as of limited value. In response, Fildes and his co-authors accept that a wider range of methods by which to segment the comparative error statistics could well prove productive and illustrate the argument by using a measure of the compatibility between the 'long term basic trend' and the 'short term recent trend' proposed by Armstrong. This particular measure, however, proved unhelpful. Fildes et al. conclude with a summary of what they regard as the important practical conclusions deriving from this research: they include the use of multiple time origins, the need to use robust estimates of any trend in the data and the benefits from updating any parameter estimates.

Original languageEnglish
Pages (from-to)359-366
Number of pages8
JournalInternational Journal of Forecasting
Volume14
Issue number3
Publication statusPublished - 1 Sept 1998
Externally publishedYes

Keywords

  • Comparative methods-Time series: univariate
  • Error measures-Evaluation
  • Forecasting competitions
  • Robustness
  • Time series-Exponential smoothing

Fingerprint

Dive into the research topics of 'Commentaries on "generalizing about univariate forecasting methods: Further empirical evidence"'. Together they form a unique fingerprint.

Cite this