The study discussed in this paper attempts to take advantage of the strengths of the Bayesian model, such as its ability to generate a point forecast (or probability bounds for a forecast), and integrate it into a new predicting software development effort methodology. The approach is intended to augment similar products, based on the Bayesian statistical model or by providing useful instruments in several stages or aspects of software development.

The causal model for the estimation of software development effort contains three variables as predictors of software development effort: software development methodology, software development computer-aided software engineering (CASE) tools, and programmer CASE tool experience. It is based on several earlier research studies. The authors use Bayesian networks to learn about joint probability distribution. They also use a belief updating procedure to complement the naive Bayesian network in modeling uncertainty.

The objectives are threefold. First, the authors review different software effort estimation models, and propose a probabilistic model. Second, they weigh the Bayesian approach against the artificial neural network and regression tree algorithms. They use a real-world data set, consisting of 33 software projects obtained from two US major companies, to assess the performances of the three techniques. The results show reasonable performances for Bayesian modeling as compared to the other two approaches. Finally, the authors show how managerial or other new information can be integrated into a Bayesian software effort estimation model, and suggest future improvements.