Software metrics: successes, failures and new directions

The history of software metrics is almost as old as the history of software engineering. Yet, the extensive research and literature on the subject has had little impact on industrial practice. This is worrying given that the major rationale for using metrics is to improve the software engineering de cision making process from a managerial and technical perspective. Industri al metrics activity is invariably based around metrics that have been aroun d for nearly 30 years (notably Lines of Code or similar size counts, and de fects counts). While such metrics can be considered as massively successful given their popularity, their limitations are well known, and mis-applicat ions are still common. The major problem is in using such metrics in isolat ion. We argue that it is possible to provide genuinely improved management decision support systems based on such simplistic metrics, but only by adop ting a less isolationist approach. Specifically, we feel it is important to explicitly model: (a) cause and effect relationships and (b) uncertainty a nd combination of evidence. Our approach uses Bayesian Belief nets, which a re increasingly seen as the best means of handling decisionmaking under unc ertainty. The approach is already having an impact in Europe. (C) 1999 Else vier Science Inc. All rights reserved.