How to overcome modelling and model risk management challenges with artificial intelligence and machine learning

Abstract

This paper gives an overview of existing applications of artificial intelligence/machine learning (AI/ML) and selects an example, credit card fraud detection, to illustrate the application of modelling methods to AI/ML. Specifically, tests of modelling assumptions and assessment of model performance and stability are explained, and opaque ‘black box’ model outputs analysed to identify the most important drivers of the output. As these testing and opacity considerations, as well as the short or even real-time development cycles are important challenges to meet model risk regulations as SR 11-7, solutions for these new considerations are proposed here. In conclusion, with modifications, all methods used for conventional models can also be applied to AI/ML techniques.