Application Scorecard Modelling with Artificial Neural Networks

Credit scoring models, currently used for classifying new credit applicants, does often not have satisfactory predictive power and it is of high interest to find better models. With the recent surge of machine learning methods, artificial neural networks especially, many credit institutions are now curious of testing these methods in their fields. This thesis evaluates the practical use of artificial neural networks for credit score modelling. The practises in current credit scoring models used, the theory of artificial neural networks and a suitable development approach is discussed. It is found that artificial neural networks can outperform both current credit scoring models and other machine learning methods, such as the random forest. The main contribution of this thesis is that the networks are found to be reasonably transparent after applying a white-boxing method, but perhaps not transparent enough to comply with credit regulations. It is suggested that credit institutions can use artificial neural networks in some internal extent