Itzhak Ben-David, Mark J. Johnson, René M. Stulz
Review of Finance, Volume 29, Issue 3, May 2025, Pages 711–745, https://doi.org/10.1093/rof/rfaf009
The paper examines the resilience of data-driven lending during the unprecedented economic upheaval caused by the COVID-19 crisis, focusing specifically on small business fintech lending in March 2020. Traditional lenders often use a combination of hard and soft information in underwriting loans. Hard information includes verifiable data such as FICO scores and bank statements, while soft information encompasses qualitative insights gathered over time. Fintech lenders, however, rely strictly on hard data fed into statistical models to make lending decisions, enabling them to process applications efficiently and provide quick approvals.
Despite its efficiencies, data-driven lending has a significant vulnerability known as model risk—the risk that the models used are not suitable for new or drastically changed economic conditions. Since these models are trained on historical data, they may fail when the future deviates substantially from the past. The COVID-19 pandemic presented such a scenario, with economic conditions that had no precedent in modern data. As a result, fintech lenders’ models became unreliable.
The study utilizes daily data from a fintech platform connecting small businesses with online lenders, capturing loan applications, approval decisions, and offer terms. In March 2020, there was a surge in loan applications—more than doubling at their peak compared to the pre-pandemic period—as financially constrained small businesses sought emergency funding. Interestingly, the credit quality of applicants improved during this time. However, despite the increased demand and improved applicant profiles, the supply of fintech loans collapsed dramatically in the latter half of March. The number of loan offers plummeted to less than a tenth of the peak, and many lenders ceased operations on the platform. Further, we show that the terms of loans offered to applicants were only marginally affected by the applicants’ exposure to the COVID-19 shock. Finally, there is no evidence that lenders shifted credit from riskier low-FICO score applicants to high-FICO score applicants. The analysis suggests that fintech lenders continued to rely on their existing models without adjusting for the pandemic’s impact until they recognized the models’ inadequacy and halted lending.
The paper highlights a key contrast with traditional banks, which continued lending during the crisis. Banks often use both hard and soft information and maintain relationships with borrowers, so that they are better able to assess risk in unusual conditions than fintech lenders. Relying solely on hard data and statistical models left fintech lenders particularly vulnerable when confronted with an economic environment vastly different from their training data.
In conclusion, the COVID-19 crisis exposed a fundamental weakness in data-driven lending: the inability of models trained on historical data to adapt to unprecedented events. This lack of resilience led to a significant withdrawal of credit at a time when small businesses needed it most. The findings contribute to the literature on the implications of data technologies in credit markets, the dynamics of fintech lending to small firms, and the broader understanding of financial systems’ responses to systemic shocks.