New ideas and emerging research: evaluating prediction system accuracy

Abstract

BACKGROUND: Prediction e.g. of project cost is an important concern in software engineering. PROBLEM: Although many empirical validations of software engineering prediction systems have been published, no one approach dominates and sense-making of conflicting empirical results is proving challenging. METHOD: We propose a new approach to evaluating competing prediction systems based upon an unbiased statistic (Standardised Accuracy), analysis of results relative to the baseline technique of guessing and calculation of effect sizes. RESULTS: Two empirical studies are revisited and the published results are shown to be misleading when re-analysed using our new approach. CONCLUSION: Biased statistics such as MMRE are deprecated. By contrast our approach leads to valid results. Such steps will greatly assist in performing future meta-analyses.