Do arbitrary function approximators make sense as software prediction models?

Predicting software development effort with high precision is still a major challenge for the industry and a major research area in software engineering. A significant share of research on software prediction is devoted to research on arbitrary function approximators, such as estimation by analogy, regression trees and artificial neural networks. This paper questions the use of arbitrary function approximators (AFA's) for software prediction by invoking theory of science and appealing to common sense. We argue that arbitrary function approximators may be useful in exploratory data analysis but we question their value for predictive purposes, and especially for software effort prediction