Loan Officers, Algorithms, & Credit Outcomes: Experimental Evidence from Pakistan

Abstract

This paper studies how loan officers and machine learning algorithms differentially respond to revealed demographics of loan applicants in a developing country. I conducted an experiment in Pakistan involving 30 loan officers and 5,500 digitally submitted loans. The intervention assigned loans to either the officers or a machine learning algorithm and provided applicant identities for a subset of loans to each decision-maker. The loan officers exhibit a gender equity preference and approve more women once they observe gender without raising overall loan default. When trained on an anonymized applicant pool, the algorithm achieves a 21% reduction in default relative to the loan officers and approves a similar fraction of minority borrowers. I quantify the level of both human and algorithmic discrimination and show that revealing applicant identities has opposing effects on each agent’s level of gender-based discrimination. Specifically, while discrimination declines for loan officers, it increases for the algorithm. The results show that blinding algorithms to applicant demographic characteristics may boost efficiency and ensure equity in developing economy credit markets.

Keywords: Machine learning, loan officers, discrimination, informal credit

JEL Classification: G51, J15, J16, O16