Determining relevant training data for effort estimation using Window-based COCOMO calibration

Abstract Context A software estimation model is often built using historical project data. As software development practices change over time, however, a model based on past data may not make accurate predictions for a new project. Objectives We investigate the use of moving windows to determine relevant training data for COCOMO calibration. Method We present a windowing calibration approach to calibrating COCOMO and assess performance of effort estimation models calibrated using windows and all data. Results Our results show that calibrating COCOMO using small windows of the most recently completed projects generates superior estimates than using all available historical projects. Large windows tend to produce worse estimates. Conclusions This study provides empirical evidence to support the use of small windows of projects completed so far to calibrate models when COCOMO-like data is available. Additionally, when the change in software development over time is rapid, the use of windows is more justifiable for improving estimation accuracy.