Software fault prediction

Abstract Cost-effective and timely software development methods are essential today as software costs and backlogs escalate while applications are developed in rapidly changing environments. Focusing testing efforts on those portions of the code with the largest number of faults can reduce development costs and time, but requires prediction of the potential location of faults. This study is an attempt to determine whether these predictions can be made by applying neural networks to predict faults in several National Aeronautics and Space Administration software development projects. We found that neither code size nor subjective factors describing the problem, schedule, personnel, process, product, and development environment provided enough information with which to predict the total number of faults in highly fault-prone modules. However, if faults tend to cluster, then identification of fault-prone modules through initial testing could guide subsequent testing efforts that focus on these modules.