Resource title

Combination of forecasts of the M3-competition

Resource image

image for OpenScout resource :: Combination of forecasts of the M3-competition

Resource description

The authors study combinations of forecasting methods using 14 individual methods from the M3-competition. They first replicate previous work on combinations of forecasts, fixed over series, done with the data of the M-competition. They then perform a variation which shows that if one chooses a different combination of forecasts for each time series independently, then, on average across the time series: a) it is no longer true that the worst combination of a few - even up to 11 - methods is better than an average single method; and b) the best possible single method in this case is better than any best possible combination. They also investigate if the different methods selected in the best possible combinations for each time series are more or less frequently used than others. They then explore the question whether choosing a good combination is easier than choosing a good single method. To this purpose the authors use a simple criterion for selecting the best choice among different possible combinations or methods. The experiments show that this simple criterion leads to combinations that have performance similar to, and often better than that of the best single method chosen using the same criterion. This indicates that even though the best possible single method per series is better than the best possible combination, it is easier to choose a good combination that a good single method. This may be because, as the first part of this paper shows in agreement to previous work, there is larger variance in the performance of single methods than in that of combinations.

Resource author

Resource publisher

Resource publish date

Resource language


Resource content type


Resource resource URL

Resource license

Copyright INSEAD. All rights reserved