Risk classifiers and generalized perceptrons

Risk classification of financial institutions is analyzed in the case where several risk classes are present. It is shown that the classical scoring systems (simple perceptrons) may no longer provide sufficient storage capacity in this case (no matter which learning algorithm is employed). Hence, a generalized version of the perceptron, which increases the storage capacity, is considered. It is argued that the resulting advantages outweigh the disadvantages. Preliminary experimental results are described which indicate that the proposed method, in spite of its theoretical shortcomings, is eminently suitable for practical purposes. More extensive tests with realistic data (which are hard to come by) are proposed to verify the claim.