Credit Risk Modeling: A Benchmark of Machine Learning Methods (joint CORMSIS and CRR seminar) Event

Event details

Credit scoring involves the development of empirical models to support decision making in the retail credit business. In particular, financial institutions use predictive models, called scorecards, to decide upon new credit applications or to manage existing loan agreements. Such scorecards estimate the probability of a borrower to default. The paper presents results of a comprehensive benchmark experiment in which we contrast the performance of 41 alternative forecasting methods across eight real-world credit scoring data sets. The objective of the study is to examine the potential of recent advancements in the field of predictive analytics for credit scoring. More specifically, we consider several novel learning methods that are for the first time assessed in a credit scoring context. Furthermore, using the principles of cost-sensitive learning, we shed light on the link between the (statistical) accuracy of scorecard predictions and a scorecard's business value. Finally, we examine the extent to which different indicators of predictive accuracy agree in their scorecard assessment and thereby contribute to a recent debate concerning the appropriateness of receiver operating characteristics. Our study provides valuable insight for professionals and academics in credit scoring. It helps practitioners to stay abreast of technical advancements in predictive modeling. From an academic point of view, the study provides an independent assessment of recent prediction methods and offers a new baseline to which future models can be compared.