The credit industry has a serious problem. Banks want to reach new customers but are wary of lending to borrowers they view as risky. Specifically, I’m talking about people of color, people without a college degree, people who’ve immigrated to the U.S., and millennials with no history of using credit, who are often the last to get a loan.
These underserved communities are often deemed “credit invisible,” meaning they don’t have a traditional credit score or have very little information in their credit file. Research by the Consumer Financial Protection Bureau shows that as many as 15% of African Americans and Hispanics are credit invisible and another 13% of Blacks and 12% of Hispanics have records that are so thin they’re treated as unscorable by commercially available credit scoring models. These numbers are close to double the percentages for white credit applicants. Invisibility affects young people in similar ways. The population under 25 has three times more people missing information in their credit file when compared to people in the 30-50 age range.
Machine learning has the power to reduce discrimination in credit. For one mortgage lender, ZAML was able to generate a fairer model that shrank the approval rate gap between white and African American borrowers by 32%.
The credit invisible find themselves stuck in a vicious cycle through no fault of their own. The fact that they lack a credit history doesn’t mean that they are necessarily riskier borrowers than someone with a robust file. What often hurts these borrowers are the flaws built into traditional credit scoring, which has traditionally been limited to a couple of dozen factors such as credit score, income and current debt outstanding. Limiting the factors ignores a good deal of information that can greatly impact a lender’s decision to approve a loan.
This is where artificial intelligence (AI) and machine learning (ML) can help. ML models use up to 100 times more data points and more sophisticated math to generate a better risk prediction in a few seconds. ML credit models can fold in more indicators of creditworthiness and surface subtle connections among pieces of information that paint a clearer picture of whether someone is a good credit risk. ML can increase access to credit for minority and low-income borrowers who have been left out of mainstream lending.
Our work building ML models for dozens of lenders has shown time and again that borrowers with no credit score can often be just as likely to repay a loan as those with a score. In our work with one major auto lender, we were able to safely increase approval rates among applicants with low or no credit scores by 8% to 10%. For one subprime auto lender, ML was able to increase approval rates for millennials by 25% and for people under 20 by 200%.
Zest Automated Machine Learning (ZAML) has built-in tools to identify and remove bias in credit models, even traditional logistic regression models. Lenders using ZAML have been able to increase loans consistently to protected classes, including African American and Hispanic borrowers, without seeing an increase in defaults. In a test with one client, ZAML was able to generate a fairer model within minutes, shrinking the approval rate gap between white and African American borrowers by 32% and between white and Hispanic borrowers by 29% while at the same time saving the lender $1.2 million annually compared to their benchmark traditional model, because the fairer ML model results in fewer defaults.
For one subprime auto lender, ML was able to increase approval rates for millennials by 25% and for people under 20 by 200%.
It’s crucial that lenders who adopt ML to improve fairness work with the right vendors. Regulators are highly concerned about lenders using ML methodologies that can’t identify specific reasons why credit was denied to someone. ML runs that risk if a lender’s model is a black box where they know the information that goes in and comes out but not why the model denied or approved any given loan.
At Zest, we built real explainability into ZAML so that every variable that affects a credit decision can be examined and discriminatory variables can be removed or have their impact muted as needed to comply with fair lending regulations. It can also give the lender a clear explanation of why a loan was denied and provide all the required risk documentation at the push of a button.
ML is a powerful tool in the fight to make lending fairer. When used the right way, ML can help build a world that gives more people the opportunity that comes with credit.