A comprehensible SOM-based scoring system

Abstract

The significant growth of consumer credit has resulted in a wide range of statistical and non-statistical methods for classifying applicants in 'good' and 'bad' risk categories. Traditionally, (logistic) regression used to be one of the most popular methods for this task, but recently some newer techniques like neural networks and support vector machines have shown excellent classification performance. Self-organizing maps (SOMs) have existed for decades and although they have been used in various application areas, only little research has been done to investigate their appropriateness for credit scoring. In this paper, it is shown how a trained SOM can be used for classification and how the basic SOM-algorithm can be integrated with supervised techniques like the multi-layered perceptron. Classification accuracy of the models is benchmarked with results reported previously. © Springer-Verlag Berlin Heidelberg 2005.