Rob J Hyndman

Twitter Gplus RSS

Another look at measures of forecast accuracy

Published on 16 November 2006 in Refereed papers

Inter­na­tional Journal of Fore­cast­ing (2006). 22(4), 679–688.

Rob J. Hyndman1 and Anne B. Koehler2

  1. Depart­ment of Eco­no­met­rics and Busi­ness Stat­ist­ics, Mon­ash Uni­ver­sity, VIC 3800, Australia.
  2. Depart­ment of Decision Sci­ences and Man­age­ment Inform­a­tion Sys­tems, Miami Uni­ver­sity, Oxford, OH 45056, USA.

Abstract: We dis­cuss and com­pare meas­ures of accur­acy of uni­vari­ate time series fore­casts. The meth­ods used in the M-​​competition and the M3-​​competition, and many of the meas­ures recom­men­ded by pre­vi­ous authors on this topic, are found to be degen­er­ate in com­monly occur­ring situ­ations. Instead, we pro­pose that the mean abso­lute scaled error become the stand­ard meas­ure for com­par­ing fore­cast accur­acy across mul­tiple time series.

Keywords: fore­cast accur­acy, fore­cast eval­u­ation, fore­cast error meas­ures, M-​​competition, mean abso­lute scaled error.

Sample cal­cu­la­tions: Excel spread­sheet show­ing MASE cal­cu­la­tion for the “product C” series.

Data: Data used in examples.

Online art­icle