An academic paper studying anonymized administrative data from fintech lender Upstart Network found its alternative credit-decisioning model could widen the pool of potential borrowers more than traditional credit models.

“Invisible Primes: Fintech Lending with Alternative Data” found that a traditional counterfactual model, developed in coordination with the Consumer Financial Protection Bureau (CFPB) and used for regulatory reporting purposes, resulted in a 70% higher probability of credit applicants being rejected and higher interest rates for those approved by the alternative lending model used by Upstart. Originally issued as a working paper in 2021, the revised version, published online May 28, 2022, offers new details.

The paper was co-written by Marco Di Maggio, associate professor at Harvard Business School and faculty research fellow at the National Bureau of Economic Research (NBER); Dimuthu Ratnadiwakara, assistant professor of finance at Louisiana State University; and Don Carmichael, Ph.D. candidate at C.T. Bauer College of Business at the University of Houston. The research has been widely presented at academic conferences and institutions, according to Maggio.

“The analysis shows that FICO is a very backward-looking and coarse measure, especially for a certain segment of the population, e.g., those with short credit history, minorities and young professionals,” Maggio told Bank Automation News. “Incorporating additional data into a credit model significantly improves the chances of identifying creditworthy individuals who are underserved by traditional lenders. Furthermore, the type of data that our study focuses on — education and employment history — could potentially be easily incorporated by other lenders.”

Those most affected by alternative models are “borrowers with low credit scores and short credit histories, but also a low propensity to default,” the analysis found.

“We show that funding loans to these borrowers leads to better economic outcomes for the borrowers and higher returns for the fintech platform.” Under traditional models, “Consumers with high credit scores have reaped the benefits of having multiple low-rate credit options,” and “Underscored individuals who are equally creditworthy face limited and expensive options,” according to the paper.

Alternative data expands path to credit

Fintechs can change that with the use of alternative data and algorithmic underwriting models, potentially opening up credit to millions of additional consumers, the report notes.

“According to Fair Isaac Corporation, a leading provider of credit scores, 28 million Americans have files with insufficient data to generate credit scores, and 25 million Americans have no credit file at all,” the paper states.

Still, since such alternative credit models are proprietary, there is also risk.

“Using information about education, utility bills, or bank transactions could inadvertently reduce credit access for some households, and the extent to which new data fed into a model is correlated with information that could result in discriminatory practices is largely unknown,” the report said.

Upstart, which provided access to its anonymized administrative data for the report, originated more than $3 billion in personal loans from April 2019 to March 2020, and uses alternative data, including education and job history, in its underwriting process.

After examining personal loans funded by Upstart, the authors found no relationship between the probability of defaulting and credit scores below 700.

“This comparison suggests that, for some individuals, credit score may not paint an accurate picture of future creditworthiness,” they note in the paper.

Upstart’s AI model outperforms credit score

Additionally, after studying Upstart’s model, the authors discovered its new credit model outperformed the traditional credit score in predicting delinquencies and charge-offs.

“The new model is able to identify significant differences in creditworthiness, even among borrowers with similar credit scores,” the report concluded.

The main variables that account for Upstart’s improved predictability include non-traditional information, such as education and employment history, that supplement traditional credit report variables, it noted.

Bank Automation Summit Fall 2022, taking place Sept. 19-20 in Seattle, is a crucial event on automation and automation technology in banking. Learn more and register for Bank Automation Summit Fall 2022.