Predicting software errors, during development, using nonlinear regression models: a comparative study

Accurately predicting the number of faults in program modules is a major problem in quality control of a large software system. The authors´ technique is to fit a nonlinear regression model to the number of faults in a program module (dependent variable) in terms of appropriate software metrics. This model is to be used at the beginning of the test phase of software development. The aim is not to build a definitive model, but to investigate and evaluate the performance of four estimation techniques used to determine the model parameters. Two empirical examples are presented. Results from average relative error (ARE) values suggest that relative least squares (RLS) and minimum relative error (MRE) procedures possess good properties from the standpoint of predictive capability. Moreover, sufficient conditions are given to ensure that these estimation procedures demonstrate strong consistency in parameter estimation for nonlinear models. Whenever the data are approximately normally distributed, least squares may possess superior predictive quality. However. in most practical applications there are important departures from normality: thus RLS and MRE appear to be more robust