Evaluating Accuracy (or Error) Measures

This paper surveys all major accuracy measures found in the field of forecasting and evaluates them according to two statistical and two user oriented criteria. It is established that all accuracy measures are unique and that no single measure is superior to all others. Instead there are trade-offs in the various criteria that must be considered when selecting an accuracy measure for reporting the results of forecasting methods and/or comparing the performance of such methods. It is concluded that symmetric MAPE and Mean Square Error are to be preferred for reporting or using the results of a specific forecasting method while the difference between the MAPE of NAIVE 2 minus that of a specific method is a preferable way of evaluating some specific method to some appropriate benchmark.