Faizaan Kisat will be presenting in person. Viewers may also participate via Zoom
This paper compares the efficiency and equity of algorithmic versus human screening methods in reviewing digital loan applications in developing economy financial markets. In partnership with a digital credit provider, I conducted an experiment in Pakistan that randomly provides borrower identities to both traditional loan officers and a machine learning algorithm. Conditional on reviewing the same set of de-identified loan applications, the algorithm achieves a 17% reduction in loan default relative to the loan officers. The officers exhibit a strong gender equity preference and approve more women once they observe gender. Conversely, an algorithm additionally trained on applicant identities approves 49% fewer women while reporting negligible efficiency gains. I estimate large and positive levels of algorithmic discrimination against women across all algorithms. The results show that the use of machine learning models trained on conventional borrower data may exacerbate existing gender gaps in financial access in emerging markets.