Multinomial Logistic Regression Applied on Software Productivity Prediction

In software cost estimation various methods have been proposed to yield a prediction of the productivity of a software project. Most of the methods produce point estimates. However, in practice it is more realistic and useful to have a method providing interval predictions. Although some methods accompany a point estimate with a prediction interval, it is also reasonable to use a method predicting the interval in which the cost will fall. In this paper, we consider a method called Multinomial Logistic Regression using as dependent variable the predefined cost intervals and as predictor variables the attributes, similar to the ones characterizing completed projects of the available data set. The method builds a model, which classifies any new software project, according to estimated probabilities, in one of the predefined intervals. The proposed method was applied to a well-known data set and was validated with respect to its fitting and predictive accuracy.