Evaluating the Performances of Software Cost Estimation Models through Prediction Intervals

Abstract

The task of predicting accurately the cost required for the completion of a new software project is a challenging issue in the Software Cost Estimation area, since it is closely related with the activities of project management and the wise decision-making of organizations in order to bid, plan and budget a forthcoming system. However, the accurate prediction of the cost is often obtained with great uncertainty and for this reason there has been noted a lack of convergence in experimental studies. The main reason for the discrepancy can be derived from the inherent characteristic of prediction methodologies, since they produce point estimates without taking into account the risk covering the whole process. In this study, we propose a statistical framework, so as to focus on the construction of Prediction Intervals which provide an “optimistic” and a “pessimistic” guess for the true magnitude of the cost. The proposed framework that incorporates different accuracy indicators, formal hypothesis testing and graphical inspection of the predictive performance is applied on a dataset with real software projects.