Making accurate credit risk predictions with cost-sensitive MLP neural networks

Abstract

In practical applications to credit risk evaluation, most prediction models often make inaccurate decisions because of the lack of sufficient default data. The challenging issue of highly skewed class distribution between defaulter and nondefaulters is here faced by means of an algorithmic solution based on cost-sensitive learning. The present study is conducted on the popular Multilayer Perceptron neural network using three misclassification cost functions, which are incorporated into the training process. The experimental results on real-life credit data sets show that the proposed cost functions to train such a neural network are quite effective to improve the prediction of examples belonging to the defaulter (minority) class. © Springer International Publishing Switzerland 2013.