The M4 competition: Conclusions

By Spyros Makridakis, Fotios Petropoulos. 2020

26 citations recorded by the source corpus.

There have been exactly 40 years since the publication of the predecessor of the M Competitions, the Makridakis and Hibon (MH) study (1979), with its surprising and highly contested finding that simple statistical methods were at least as accurate as complex and statistically sophisticated ones. More specifically, this study found that the Mean Absolute Percentage Error (MAPE) of Naïve 2 was, on average, 1.4% more accurate than that of the Box-Jenkins methodology to ARIMA models, while Single Exponential Smoothing (SES) was correspondingly 2.61% more accurate. Clearly, this was unwelcome evidence for the most popular approach to statistical forecasting at that time and was not received kindly by the statistical academic community (see Hyndman’s paper, this issue, “A brief history of forecasting competitions” for details). Nevertheless, the MH study started a long forecasting w inter that continued with the M1 (Makridakis et al., 1982) and M2 Competitions (Makridakis et al., 1993) that provided similar findings. Such findings were partially lifted with the M3 Competition (Makridakis et al., 2000), when the sMAPE of Theta, still a simple method, was 2.46% more accurate than Naïve 2, 1.31% more accurate than SES and 1% more accurate than the Box-Jenkins method. The forecasting spring began with the M4 Competition where the top method, a complex hybrid approach

Subfields: Management Information Systems, Management Science and Operations Research