Scoring systems, classifiers, default probabilities, and kernel methods

Perceptron learning is discussed in the context of so-called scoring systems. It is argued that in conjunction with maximum likelihood methods this is particularly suitable for such a banking application. Several practical reasons are given why in this context it should be preferred to support vector machines. The interpretation of the perceptron output as a posteriori probability using a prior from the exponential family is explained. Encouraging experimental results concerning an anonymous but otherwise genuine substantial data set are presented. Finally it is shown that the well-known "kernel trick" employed for support vector machines may equally well be utilized for perceptrons.