Credit Scoring Using Deep Learning Driven by Optimization Algorithms

Abstract

Credit scoring as good or bad has a significant importance for financial institutions. This paper presents a Deep Learning approach for credit scoring. We have defined the credit score as a cost obtained by a weighted sum of the number of false negative errors (FN) and the number of false positive errors (FP). Our objective was to obtain the lowest possible score. The largest weight is allocated to the indicator FN (this corresponds to the prediction of bad credit as good credit). As a consequence, our best credit score corresponds to minimization of Miss Alarm Rate (MAR) for a given sum of total errors. The proposed model of cost minimization uses state of the art mathematical algorithms and deep learning techniques. In our work, we use optimization algorithms for selecting a deep learning neural network architecture and for finding the optimum hyperparameters. The method is tested using the German credit dataset. The best result leads to a MAR of 3%.