Neural network based software effort estimation & evaluation criterion MMRE

Abstract

Evaluation of any prediction model is the most important part for comparison. Mean Magnitude Relative Error (MMRE) is one of the traditional evaluation criterion. In this paper we demonstrate the application of MMRE for model for predicting software development effort estimation and show that MMRE is not a suitable evaluation criterion for the software development estimation models. The argument is supported by series of simulations. In this paper we also suggest changes needed in MMRE calculations and propose Modified MMRE algorithm for the effort estimation evaluation criterion.