Real-world credit scoring: a comparative study of statistical and artificial intelligent methods

Abstract

Credit scoring is an integral and crucial part of any lending process that any little development in it can reduce huge potential losses of financial organisations. The assessment of model performance varies because of different performance measures under a variety of circumstances on different nature of datasets. Therefore, this study employed six well-known classification approaches on six real-world credit datasets for comprehensive assessment by combining ten representative performance criterions. The experimental outcomes, statistical significance test and the estimated cost of prediction error confirm the marginal superiority of logistic regression (LR) and TreeNet over CART and MARS, being more robust compared to other two approaches LASSO and RF.