In this research, parametric software cost estimation models and their related calibration methods have been analyzed, especially for the COCOMO model and the Bayesian calibration approach. This research combines machine learning techniques and statistical techniques. With this approach, the prediction powers of the COCOMO parametric software cost model are shown to be significantly improved while the variability is decreased with respect to the dataset being analyzed. This research studies not only the accuracy but also the variances of the model and the variables. It can improve the confidence of people who use software cost estimation models, show the prediction power of software cost estimation models after calibration, and make it easier and better to perform software data collection and analysis. However, the research also identifies risks in using the approach, such as dropping parameters that will vary on future projects. This research provides methods that can help an organization to reason about the relationship between the characteristics of the organization and its projects' software development costs and schedules. The methods can thus help the organization to make more cost-effective development decisions and investment decisions.

The research also provides new insights on how to combine calibration, stratification, hold-out, and machine learning techniques to produce more accurate parametric models for particular organizations or situations.