Benchmarking alternative interpretable machine learning models for corporate probability of default

Abstract

<abstract> <p>In this study we investigate alternative <italic>interpretable machine learning</italic> ("IML") models in the context of <italic>probability of default</italic> ("PD") modeling for the large corporate asset class. IML models have become increasingly prominent in highly regulated industries where there are concerns over the unintended consequences of deploying black box models that may be deemed conceptually unsound. In the context of banking and in wholesale portfolios, there are challenges around using models where the outcomes may not be explainable, both in terms of the business use case as well as meeting model validation standards. We compare various IML models (deep neural networks and explainable boosting machines), including standard approaches such as logistic regression, using a long and robust history of corporate borrowers. We find that there are material differences between the approaches in terms of dimensions such as model predictive performance and the importance or robustness of risk factors in driving outcomes, including conflicting conclusions depending upon the IML model and the benchmarking measure considered. These findings call into question the value of the modest pickup in performance with the IML models relative to a more traditional technique, especially if these models are to be applied in contexts that must meet supervisory and model validation standards.</p> </abstract>