A study in which researchers tested alternative credit scoring models using artificial intelligence found that lower-income families and minorities have a problem: the predictive tools are between 5-10% less accurate than those for higher-income groups and non-minorities.
The algorithms that determine credit scores are not biased against disadvantaged borrowers. It’s not that credit score algorithms are biased against underprivileged borrowers.
Lenders prefer to see more information than less. It also means that a few minor mistakes, like past delinquent debt, can significantly impact a person’s credit score.
We’re dealing with flawed data due to various historical factors, says Laura Blattner, an assistant professor of Finance at Stanford Graduate School of Business. She coauthored this new study with Scott Nelson and opened a new window of the University of Chicago Booth School of Business. If you only have one credit card and have never had a home mortgage, you will have less information to determine whether you will default. “If you defaulted once several years ago, it may not be a good indicator of the future.”
Root Cause of the Problem
Researchers used artificial intelligence to analyze the problem and a large volume of consumer data to test various credit scoring models.
First, they had to determine if credit scores were accurate for all demographic groups. The researchers used AI to analyze anonymized credit data from 50 million consumers, which a significant credit score company provided. The team also used a large marketing dataset to identify borrowers based on income, race, and ethnicity.
The challenge was determining whether those rejected for home loans were likely to default if they’d been approved. The AI models examined how applicants who were denied mortgages compared to other types of car loans. These loans are closely related to the likelihood that a person will default on a home loan.
Credit scores were less accurate than those for minority and low-income borrowers.
Blattner and Nelson conclude that the credit scores of people who live in low-income and minority households contain a lot more “noise.” The researchers found that minority scores are 5% less accurate at predicting default than scores from non-minority borrowers. Scores for those in the lowest fifth of income range are also about 10% less accurate than scores for higher-income borrowers.
Why? Are the algorithms biased because they can’t detect the distinct patterns of certain demographic groups?
Blattner and Nelson tested different scoring models, which were fine-tuned for minority and low-income borrowers. The results did not change much. The scores for these borrowers were still lower.
The problem lay in the data. Creditworthiness was harder to determine for people with limited credit histories, who have taken out fewer loans and hold fewer if not any, credit cards. This was particularly true for those with one or two blemishes in their credit records. Minority borrowers and those with low incomes were more likely than others to have a thin or spotty record, which made their credit scores less accurate.
Blattner and Nelson claim that by making credit report data richer, for example, compiling thicker, more diverse files, it is possible to eliminate half of the accuracy disparity.
Misallocation of Credit
The findings have several important implications. The study shows that people in disadvantaged communities get denied mortgages more than they should, but some borrowers are approved even when they shouldn’t. This leads to a misallocation of credit. This can also perpetuate inequalities, as people who cannot get mortgages will have fewer opportunities to establish a solid record. Then, they lose out on a meaningful way to build wealth.
Experts have claimed that financial institutions cannot incorporate factors such as ethnicity, race, or gender into their models. These prohibitions were intended to stop discrimination. However, they can also prevent lenders from recognizing differences that could improve some borrowers’ scores.
This is not a concern, according to the study. Even though Blattner and Nelson developed scoring models specifically for minority and low-income borrowers, their scores were less accurate.
Blattner says there is no easy solution. One possible strategy is for financial companies to run experiments where they approve loans to individuals with low credit scores.
Blattner says that if you were a bank, you could lend money to people and then see who paid. “Some fintech companies do exactly that: they give loans, and then learn.”