Machine learning in credit risk: measuring the dilemma between prediction and supervisory cost

New reports show that the financial sector is increasingly adopting machine learning (ML) tools to manage credit risk. In this environment, supervisors face the challenge of allowing credit institutions to benefit from technological progress and financial innovation, while at the same ensuring compatibility with regulatory requirements and that technological neutrality is observed. We propose a new framework for supervisors to measure the costs and benefits of evaluating ML models, aiming to shed more light on this technology’s alignment with the regulation. We follow three steps. First, we identify the benefits by reviewing the literature. We observe that ML delivers predictive gains of up to 20?% in default classification compared with traditional statistical models. Second, we use the process for validating internal ratings-based (IRB) systems for regulatory capital to detect ML’s limitations in credit risk mangement. We identify up to 13 factors that might constitute a supervisory cost. Finally, we propose a methodology for evaluating these costs. For illustrative purposes, we compute the benefits by estimating the predictive gains of six ML models using a public database on credit default. We then calculate a supervisory cost function through a scorecard in which we assign weights to each factor for each ML model, based on how the model is used by the financial institution and the supervisor’s risk tolerance. From a supervisory standpoint, having a structured methodology for assessing ML models could increase transparency and remove an obstacle to innovation in the financial industry.