A Virtual Study of Moving Windows for Software Effort Estimation Using Finnish Datasets

CONTEXT: Studies have shown contradictory results on the effectiveness of using a moving window of only the most recent projects for effort estimation, compared to using the full history of past data. Moving windows improved the accuracy of effort estimates for a single-company subset of the ISBSG dataset (www.isbsg.org), but not for three single-company subsets of the Finnish dataset (www.4sumpartners.com). The contradiction may be caused by different characteristics of the data sets: in particular, they differ noticeably in heterogeneity of industry sector. GOAL: To investigate the effect on estimation accuracy of differences in the characteristics of the data sets. METHOD: Conduct an experiment with a virtual data set, composed from the three subsets of the Finnish dataset. The composite data set is similar to the ISBSG subset in that it includes data from multiple industry sectors; the largest group of projects in both data sets comes from the same industry sector; and in both data sets the projects are concentrated in a similar number of years. RESULTS: The conclusions is the same as in the past study using the individual Finnish subsets: in the composite data set, moving windows are of no help. CONCLUSIONS: In this instance, increased heterogeneity of projects does not explain the contradiction. It is still not clear when windows may be helpful. Practitioners and researchers should not assume automatically that only the most recent data is best for effort estimation.