Machine + Man: A Field Experiment on the Role of Discretion in Augmenting AI-Based Lending Models

Abstract

Does human discretion improve or diminish lending outcomes? We assess this question in the context of a randomized, controlled experiment using a large group of lenders that rely on machine-generated credit scoring models provided by a third party to make monthly credit decisions. Working with the credit scoring company, we design a new feature for their platform – the slider feature – which allows lenders to incorporate their discretion into the credit score. We randomly assign half of the lenders to the treatment group that gets the slider; the control group does not get the slider feature and thus makes credit decisions based primarily on the machine-generated model. Consistent with discretion aiding in loan decisions, we find that the treatment group’s credit model adjustments are predictive of forward looking portfolio performance. However, we find that discretion is not useful in all cases. In fact, the control group does just as well as the treatment group in predicting credit risk for borrowers that have been traditionally classified as opaque. Our study highlights the growing prominence of AI-based lending models in crowding out some of the human’s role.

Keywords: Relationship Lending, Discretion, Machine-Learning, Fintech

JEL Classification: G2, G21, G32, O33