An Empirical Evaluation of Effort Prediction Models Based on Functional Size Measures

Software development effort estimation is among the most interesting issues for project managers, since reliable estimates are at the base of good planning and project control. Several different techniques have been proposed for effort estimation, and practitioners need evidence, based on which they can choose accurate estimation methods. The work reported here aims at evaluating the accuracy of software development effort estimates that can be obtained via popular techniques, such as those using regression models and those based on analogy. The functional size and the development effort of twenty software development projects were measured, and the resulting dataset was used to derive effort estimation models and evaluate their accuracy. Our data analysis shows that estimation based on the closest analogues provides better results for most models, but very bad estimates in a few cases. To mitigate this behavior, the correction of regression toward the mean proved effective. According to the results of our analysis, it is advisable that regression to the mean correction is used when the estimates are based on closest analogues. Once corrected, the accuracy of analogy-based estimation is not substantially different from the accuracy of regression based models.