“You don’t marginalize by accident,” he says Eric Anderson, professor of marketing at the Kellogg School. “If you think of it as a big funnel, unbanked customers don’t get pulled down the funnel. Then, if they happen to get into the funnel, they get kicked out again because they don’t have credit scores.”
But what if financial institutions could use other, more creative measures to predict who will repay their loans? They may acquire new customers while improving access to credit. This possibility has businesses, scholars and policymakers asking whether behaviors not directly linked to income or credit history—like, for example, how someone shops for groceries—can predict responsible loan repayment habits.
New research by Anderson and his colleagues Jung Youn Lee of Rice University and Joonhyuk Yang of the University of Notre Dame explores this idea. The team overlays data from a grocery store and a credit card company to uncover connections between consumer shopping habits and payments.
Specifically, their study investigates whether “good” or “bad” grocery shopping behaviors correlate with levels of financial health. A person who exercises discipline—shops at predictable times, follows a budget and buys items on sale—might also demonstrate consistency in paying on time with a credit card, the researchers suggested.
The researchers’ instincts proved correct.
“It turned out that most of what we came up with worked to some extent,” Anderson says. “We’re finding a lot of tiny signals that give you only a little bit of information about your credit, but when you put them all together, you get a strong signal.”
Vinegar dressings vs. beef mortadella
The data showed that people with consistent supermarket behavior tended to pay their credit cards on time. These customers were more likely than others to shop at regular times of the day and on weekdays, make repeat purchases, spend roughly the same amount each time, and consistently rely on offers.
Bankrupts, on the other hand, behaved more erratically in the store, visiting at unpredictable times and also looking for a variety of brands and items.
Even after the researchers controlled for credit scores, income and other sociodemographic factors, the association remained between grocery shopping habits and credit risk.
On a fascinating level of findings, it wasn’t just the way people shopped, but the items they bought that correlated with their loan repayment habits. Cigarette purchases were the biggest red flag that someone was going bankrupt, which was defined as missing two consecutive credit card payments. Buying pre-processed foods such as beef mortadella, as well as energy drinks and canned fish, also indicate poor payers.
The single biggest indicator of a non-bankrupt, on the other hand, was spending a significant amount on vinegar salad dressing. “They overestimate being healthy,” Anderson says.
The researchers found that people who did not go bankrupt bought more individual ingredients, such as milk, flour or beans – items that form the basis of home-cooked meals, which require labor and time to prepare. Dry bread, fresh produce and imported snacks also filled their carts.
Between these two extremes of defaulters and non-bankrupts are “sloppy payers,” who may miss a payment occasionally and be back on track the following month. They stood out because they bought easily edible pasta, meat and sausages from the deli counter.
For lenders, the predictive power of this grocery data appears to be similar to that of a traditional credit score — making it especially useful for people with no credit.
When researchers modeled the value of grocery data in determining creditworthiness among credit card applicants without on credit scores, they found that it effectively filters out defaulters, leading to a 1.46 percent increase in earnings per person. For applicants with For credit scores, however, the incremental benefits of adding grocery data were much smaller.
The researchers also modeled the value of grocery data in predicting a default after a card is issued. They found that for new customers who didn’t have a credit score, grocery store data could help predict when a missed payment would lead to bankruptcy. However, the utility of grocery data declined over time as the consumer began to build a credit score and repayment history.
Overlapping data
To reach these conclusions, the researchers assumed the lens of a lender examining credit card applicants.
They gained access to data streams from a multinational conglomerate in Asia that owned both a grocery store and a credit card company. The grocery data came from both in-house retail checkout scanners. Data from the credit card issuer included socio-demographics, credit scores, spending data and payment history. Essentially, researchers could match a person’s credit card data with supermarket loyalty card data.
“We were able to combine the two things and connect the dots and then identify customers that were in the grocery data as well as the credit data,” Anderson says.
Anderson and colleagues then split their data samples into two distinct 15-month time periods, focusing on grocery data from the previous period to see if it could predict credit card data in the next period. The study excluded consumers who, during the first period, had either shopped fewer than five times or had defaulted on their credit card. This ensured they had enough data to characterize consumer buying habits. It also allowed them to remove buying behavior that could be an immediate response to the kind of economic shock that would cause bankruptcy.
This yielded 30,089 consumers. 81 percent were never delinquent in the second period, while 12 percent were defaulters and 7 percent defaulted. Credit scores were not available for 49.7 percent of the final sample.
Anderson and his colleagues applied a well-known machine learning model, XGBoost, to investigate the relationship between grocery shopping behaviors and other financial behaviors. They built a series of algorithms to win consumer loyalty.
The researchers paid independent consumers found through Amazon Mechanical Turk to rate grocery items on levels of healthiness and convenience.
The consequences
What do the findings mean for lenders?
The new story is that habits in one domain show up in another domain, according to Anderson. “We’re not the first to suggest this, but it’s an interesting application where we can see that what a layman considers good shopping behaviors also translates into good financial behaviors and getting your credit card back.”
Ultimately, this kind of information can help lenders distinguish between low-risk and higher-risk customers, he says. And as data collection and artificial intelligence improve, new possibilities emerge for connecting the dots between previously disjointed sets of information.
“That’s the world we’re moving into, which is that your data is going to be fragmented into many different areas and used in ways that you might not have thought about before,” Anderson says.
Signals like these may also help lenders recognize a higher risk of default and try to intervene early. “A lender might say, ‘Oh, we’re seeing a red flag,’ and approach a customer who has an outstanding balance.” says Anderson. “The sooner a lender can step in, the better for both parties.”
However, from a customer’s perspective, the idea that the choices you make while filling your shopping cart could affect other parts of your life in ways you don’t understand or consent to can be disturbing to many. “When you think about why we’re having conversations today about data privacy regulation, it’s right around issues like the ones we’ve looked at in this article,” says Anderson.
However, he remains optimistic about the good this kind of technique can do for consumers as well. Because the status quo for those without credit—limited access to emergency resources, reliance on cost-effective alternatives like high-fee lenders—isn’t ideal.
“There is an ‘unbanked’ problem around the world,” Anderson says. “How do you get a short-term loan or some kind of financial aid and you can’t get it? The information in our paper provides another means to that end.”
