Another look at forecast-accuracy metrics for intermittent demand

 

Some of the proposed measures of forecast accuracy for intermittent demand can give infinite or undefined values. This makes them unsuitable for general use. I summarize the various measures and demonstrate what can go wrong. Then I describe a new measure (the mean absolute scaled error) which does not have these flaws. I believe it should become the standard measure for comparing forecast accuracy for multiple intermittent-demand series.

Errata: In the series shown in Table 2, the sixth value should be 11 (not 1). Similarly, the fifth error value should be 11 (not 1).

Sample calculations: Excel spreadsheet showing MASE calculation for the example series.

Data: Data used in examples.

Cite

@article{Hyndman06,
  author  = {Rob J Hyndman},
  journal = {Foresight: the International J Applied Forecasting},
  pages   = {43--46},
  title   = {Another look at forecast-accuracy metrics for intermittent demand},
  volume  = {4},
  year    = 2006
}