On Soring Systems with Binary Input Variables

摘要

An approach to scoring systems is presented that avoids the usual ad hoc methods or assumptions on underlying probability distributions when assigning points to customer characteristics. Instead a perceptron learning algorithm is used for binary input variables. An application of affine scaling transformations leads to an improvement of the network topology. Finally a perceptron learning algorithm is applied to compound characteristics to obtain a default probability using a linear discriminant under comparatively weak assumptions on the underlying probability distributions.