Maintainability of classes in terms of bug prediction

Measuring software product maintainability is a central issue in software engineering which led to a number of different practical quality models. Besides system level assessments it is also desirable that these models provide technical quality information at source code element level (e.g. classes, methods) to aid the improvement of the software. Although many existing models give an ordered list of source code elements that should be improved, it is unclear how these elements are affected by other important quality indicators of the system, e.g. bug density. In this paper we empirically investigate the bug prediction capabilities of the class level maintainability measures of our ColumbusQM probabilistic quality model using open-access PROMSIE bug dataset. We show that in terms of correctness and completeness, ColumbusQM competes with statistical and machine learning prediction models especially trained on the bug data using product metrics as predictors. This is a great achievement in the light of that our model needs no training and its purpose is different (e.g. to estimate testability, or development costs) than those of the bug prediction models.