Machine Learning for Credit risk: three successful Case Histories

Abstract

As the financial services landscape witnesses an unprecedented change, banks can use machine learning (“ML”) to expand their databases through alternative sources providing unstructured and semi-structured information, such as transaction data and digital footprint data. However, ML algorithms also suffer from several potential shortcomings, as they may overfit sample data and prove unstable over time, they may quickly become obsolete and need re-estimation, and they may prove hard to interpret. This paper joins the debate on ML in banks by providing three case studies that highlight the benefits of machine learning, while showing how its drawbacks can be minimised: a rating model developed within the IRB framework, a challenger model used to validate a bank’s main model for retail PDs, and an early warning system based on transaction data.