Construction of a k-nearest-neighbour credit-scoring system

This paper describes the construction of a credit-scoring system using the k-nearest-neighbour method, a standard technique in pattern recognition and non-parametric statistics. An important part of our analysis is the selection of distance metrics. We propose using an adjusted version of the euclidean distance metric, which incorporates an estimate of underlying equiprobability contours for class membership. This approach was used to classify a sample of applications for mail-order credit from the Littlewoods Organisation. We describe the results of comparison experiments with a range of discrimination techniques, including logistic regression, projection-pursuit regression, and decision trees.