Reliable Confidence Intervals for Software Effort Estimation

This paper deals with the problem of software effort estimation through the use of a new machine learning technique for producing reliable confidence measures in predictions. More specifically, we propose the use of Conformal Predictors (CPs), a novel type of prediction algorithms, as a means for providing effort estimations for software projects in the form of predictive intervals according to a specified confidence level. Our approach is based on the well-known Ridge Regression technique, but instead of the simple effort estimates produced by the original method, it produces predictive intervals that satisfy a given confidence level. The results obtained using the proposed algorithm on the COCOMO, Desharnais and ISBSG datasets suggest a quite successful performance obtaining reliable predictive intervals which are narrow enough to be useful in practice.