Innovations in information technology are changing ground rules across many industries. As technological tools become increasingly robust, they are being implemented in innovative ways, providing consumers — and the companies that cater to them — an expanding array of benefits, some of which were unimaginable even a decade ago.

In some industries, the pace of technological progress is provoking new ways of thinking about long-standing perceptions. Entrenched notions are being tested. In the market for consumer loans, for instance, developments in machine learning (ML) and artificial intelligence (AI) are prompting questions such as, Can sophisticated technology be used as part of an effort to expand access to credit so that more creditworthy applicants — including those that live in lower-income neighborhoods — receive loans?

In their working paper, "One Threshold Doesn't Fit All: Tailoring Machine Learning Predictions of Consumer Default for Lower-Income Areas," Vitaly Meursault, Dan Moulton, Larry Santucci, and Nathan Schor1 examine what transpires when fairness considerations and advanced ML credit models are combined in the credit-granting decision, looking specifically at what happens when credit score requirements are lowered for applicants who live in low- and moderate-income (LMI) neighborhoods.2

By melding the two components of their framework — combining ML technology with fairness goals — the authors find that it's possible to achieve both greater overall fairness and greater profitability (relative to current standards).

Their novel idea for improving credit access is grounded in how the two parts work side by side: ML technology is used to build state-of-the-art risk models (which help lenders sharpen how accurately they assess the creditworthiness of loan applicants), while group-specific criteria (in the form of lower credit thresholds for residents of LMI communities) are simultaneously employed. The purpose of using lower thresholds for residents of LMI communities is to equalize — as much as feasibly possible — the true positive rate in those communities. In other words, the goal is to ensure that two individuals who live in different areas but have a statistically similar probability of repaying their debts are treated the same in terms of the credit decision. Because models are noisier in LMI neighborhoods, attaining that goal requires a different threshold for those neighborhoods.

Taken together, these two innovations are shown — in quantifiable and measurable ways — to significantly enhance how creditworthiness is assessed while fostering an environment in which (1) applicants with similar creditworthiness are treated the same even though they may live in poorer or richer neighborhoods and (2) lender profitability is at least as high when compared to using existing models without explicitly considering fairness.3

While documenting the welfare improvements shown in their study, the authors stress that reductions in credit score thresholds come at a cost because loans will be granted to certain lower-scoring borrowers who ultimately fail to repay them. The cost of these defaults for lenders is softened, however, by the gains realized from the adoption of sophisticated ML technology. This is because ML-enhanced underwriting is generally more accurate than legacy underwriting models, which helps lenders reduce the overall risk levels within their loan portfolios. This improvement naturally translates into higher profits, and when combined with fairness goals, it brings about a scenario in which "a lender that simultaneously adopts ... sophisticated modeling technology and group-specific thresholds could experience increases in both fairness and profit."

This research demonstrates a constructive and nuanced way to address the issue of fairness in consumer lending while maintaining bank profitability. The odds of achieving this balance are greatly improved by ML and AI models, which continue to evolve and become increasingly accurate. In "One Threshold Doesn't Fit All: Tailoring Machine Learning Predictions of Consumer Default for Lower-Income Areas," the authors introduce a new dimension to this evolution, unearthing evidence that the benefits accrue to borrowers and lenders alike.

  1. The views expressed here are solely those of the author and do not necessarily reflect the views of the Federal Reserve Bank of Philadelphia or the Federal Reserve System.
  2. Meursault is a senior machine learning economist at the Federal Reserve Bank of Philadelphia. Moulton is an advisor and senior manager of data science and engineering at the Bank. Santucci is a senior advisor and research fellow at the Bank. At the time of the study, Schor was a research analyst at the Bank.
  3. In an example of this type of reduction, lenders might reduce their credit score minimum from, say, 700 to 650 for applicants in LMI neighborhoods.
  4. The authors based their study on real-world consumer attributes derived from the Federal Reserve Bank of New York Consumer Credit Panel/Equifax dataset. They augmented these consumer characteristics with demographic information produced by the U.S. Census Bureau and the Federal Financial Institutions Examination Council. This information helped the authors determine each borrower’s income level (or, more specifically, the income level within each borrower’s census tract).