Studies in Software Cost Model Behavior: Do We Really Understand Cost Model Performance?

While there exists extensive literature on software cost estimation techniques, industry practice continues to rely upon standard regression-based algorithms. These software effort models are typically calibrated or tuned to local conditions using local data. This paper cautions that current approaches to model calibration often produce sub-optimal models because of the large variance problem inherent in cost data and by including far more effort multipliers than the data supports. Building optimal models requires that a wider range of models be considered while correctly calibrating these models requires rejection rules that prune variables and records and use multiple criteria for evaluating model performance. The main contribution of this paper is to document a standard method that integrates formal model identification, estimation, and validation. It also documents what we call the large variance problem that is a leading cause of cost model brittleness or instability.