To date studies using CBR for Web hypermedia effort prediction have not applied adaptation rules to adjust effort according to a given criterion. In addition, when applying n-fold cross-validation, their analysis has been limited to a maximum of three training sets, which according to recent studies, may lead to untrustworthy results.

This paper has therefore two objectives. The first is to further investigate the use of CBR for Web hypermedia effort prediction by comparing the prediction accuracy of eight CBR techniques, of which three have previously been compared. The second objective is to compare the prediction accuracy of the best CBR technique against stepwise regression, using a twenty-fold cross-validation. All prediction accuracies were measured using Mean Magnitude of Relative Error (MMRE), Median Magnitude of Relative Error, Prediction at level l (l=25%), and boxplots of the residuals.

One dataset was used in the estimation process and, according to all measures of prediction accuracy, stepwise regression showed the best prediction accuracy.