When should we (not) use the mean magnitude of relative error (MMRE) as an error measure in software development effort estimation?

The mean magnitude of relative error (MMRE) is an error measure frequently used to evaluate and compare the estimation performance of prediction models and software professionals. This paper examines conditions for proper use of MMRE in effort estimation contexts. We apply research on scoring functions to identify the type of estimates that minimizes the expected value of the MMRE. We show that the MMRE is a proper error measure for estimates of the most likely (mode) effort, but not for estimates of the median or mean effort, provided that the effort usage is approximately log-normally distributed, which we argue is a reasonable assumption in many software development contexts. The relevance of the findings is demonstrated on real-world software development data. MMRE is not a proper measure of the accuracy of estimates of the median or mean effort, but may be used for the accuracy evaluation of estimates of most likely effort.