Confidence in software cost estimation results based on MMRE and PRED

  • Authors:
  • Marcel Korte;Dan Port

  • Affiliations:
  • University of Applied Sciences and Arts Dortmund, Dortmund, Germany;University of Hawaii at Manoa, Honolulu, HI, USA

  • Venue:
  • Proceedings of the 4th international workshop on Predictor models in software engineering
  • Year:
  • 2008

Quantified Score

Hi-index 0.00

Visualization

Abstract

Bootstrapping is used to approximate the standard error and 95% confidence intervals of MMRE and PRED for a number of COCOMO I model variations applied to four PROMISE data sets. This is used to illustrate a lack of confidence in numerous published cost estimation research results based on MMRE and PRED comparisons such as model selection. We show that many such results are of questionable significance due to large possible variations resulting from population sampling error and suggest that a number of inconsistent and contradictory results may be explained by this. By using more standard statistical approaches that account for standard error, we may reduce the incidence of this and obtain greater confidence cost estimation in research results.