Alternative Models for Probability of Default Prediction in Banking

Due to the tightening competition within the banking sector, banks need to consider customer requests as quickly as possible while at the same time steadily reducing the requirements for a potential borrower to ensure competitive advantages. This trend leads to increased credit risks and requires new quality standards from scoring systems, but the limitations of regulators exclude the possibility of using modern approaches to modeling, which show better results, because of their uninterpretability. The purpose of this paper is to construct an effective method for interpreting black-box models, in particular, a family of gradient boost algorithms. To build the solution, the existing practices of interpreting both the arbitrary algorithm and the interpretation of tree ensembles were studied and the new method based on decision trees was proposed. The results of the new method are demonstrated on 4 different data sets. Also, the paper proposes a method of transition to a global interpretation through a local interpretation.