Advancing Fairness in Lending Through Machine Learning
More advanced models may improve predictions for each group, but they don’t address the disparities between groups.
Why Does Model Improvement Matter?
For any group of loan applicants, there will be those who will pay back their loan and those who will not.
Lenders don’t know beforehand whether an applicant will default on a loan. Typically, a lender will use an applicant’s credit score — an output of a mathematical model that uses information, such as the applicant’s payment history and current loans — to predict their likelihood of repaying a loan.1, 2 Some lenders may set a credit score threshold: Those with a credit score at or above a specific threshold are granted a loan; those with a credit score below the threshold are not (e.g., “You need a credit score of at least 600 to get this credit card.”)
Credit modeling and resulting credit scores for applicants are useful but never perfect. The interactive graphic below showcases a set of applicants with a range of credit scores. Try interacting with the data by setting different thresholds: Notice how any threshold will exclude some applicants who would have repaid and/or lend to some applicants who would have defaulted.
Applicants with credit scores below this threshold will not receive loans.
One goal of improving any predictive model is to better align its predictions with reality. With credit scoring, an improved model should result in credit scores that better separate applicants who will repay from those who will default (see below).
Are More Advanced Models Fairer?
Over the past two decades, innovations in machine learning (ML) have helped create modeling techniques that can improve credit default predictions and expand credit access. At the same time, concerns have emerged that the gains from more advanced ML models could accrue unequally between demographic groups. For example, recent research finds that Black and Hispanic mortgage borrowers benefit less than White borrowers from more sophisticated models (Fuster et al., 2021). Other research finds that credit scores are less accurate for low-income and minority borrowers, in part because these borrowers tend to have shorter credit histories and fewer lines of credit (Blattner and Nelson, 2021). Our research expands on previous work by focusing on applicants living inside and outside of (LMI) areas.
Our results show that more advanced models, in comparison to standard models, improve the accuracy of predictions for both LMI and non-LMI applicant groups. However, more advanced models do not reduce existing disparities between the LMI and non-LMI groups (see below). That is, predictions are worse for applicants living in LMI areas compared with applicants in non-LMI areas, even when more sophisticated models are used. This gap translates to a lower proportion of applicants living in LMI areas gaining access to credit compared with applicants in non-LMI areas. More advanced models are not necessarily fairer.
- Current regulation aims to reduce discrimination by prohibiting lenders from using protected classes, such as race, gender, or age. This also means that lenders may not use sensitive information that is highly correlated with these classes, such as a person’s neighborhood, when developing credit score models and making lending decisions. This research presents an alternative approach that would use neighborhood for setting fairness goals and making a lending decision but still does not use neighborhood or other sensitive information in model development.
- This credit score can be a score that lenders request from a credit bureau or a custom score built using data based on prior borrowing behavior.