Abstract
In this paper, we describe an experiment, which analyzes the relative importance and stability of change metrics for predicting defects for 3 releases of the Eclipse project. The results indicate that out of 18 change metrics 3 metrics contain most information about software defects. Moreover, those 3 metrics remain stable across 3 releases of the Eclipse project. A comparative analysis with the full model shows that the prediction accuracy is not too much affected by using a subset of 3 metrics and the recall even improves.