A pragmatic view of accuracy measurement in forecasting

Accuracy measurement in forecasting is always a subject of debate because of its importance. An adequate metric is necessary to properly select a forecasting method for a specific application. Competitions to determine the best method have helped the practitioner. The criteria for selection have not received as much attention. Of the two kinds of measurement statistics--relative and absolute--the former may present problems for the user if zeros or near zero values appear. This is more a practitioner problem because artificially generated time series do not usually have zeros. The relative and absolute measures are discussed and a solution for the existence of zeros in the data is given. If symmetry of the errors is a problem solutions are discussed. Managers will select the metric depending on the application and their management style. Once the metric has been selected the decision as to which forecasting method to select in a given situation becomes a less difficult problem.