Predictability of software-reliability models

A two-component predictability measure that characterizes the long-term predictive capability of a model is presented. One component, average error, measures how well a model predicts throughout the testing phase. The other component, average bias, measures the general tendency to overestimate or underestimate the number of faults. Data sets for both large and small projects from diverse sources with various initial fault density ranges have been analyzed. The results show that: (i) the logarithmic model seems to predict well in most data sets, (ii) the inverse polynomial model can be used as the next alternative, and (iii) the delayed S-shaped model, which in some data sets fit well generally performed poorly. The statistical analysis shows that these models have appreciably different predictive capabilities. >