Ford Motor Credit tests AI’s ability to spot overlooked borrowers
Jim Moynes, vice president of risk management at Ford Motor Credit in Dearborn, Mich., first became interested in using machine learning to improve car loan underwriting several years ago.
“We were watching what others were working on,” he said. “We like to be innovative and try to stay up with what’s going on.”
The company recently ran an experiment to see if machine learning could help its underwriters better understand the loan applications it receives.
It was a champion vs. challenger test: Moynes’ team took several years of loan data, removed all personally identifiable information from it, and gave it to ZestFinance, a provider of machine-learning-based online lending software, and its own modeling team, which creates logistic regression models to predict potential borrowers’ creditworthiness.
Each team ran the loan application data through its models and predicted the future performance of the loans. Moynes then compared the actual performance of those accounts over the past several years to the two teams’ predictions.
The machine learning software won.
“What we discovered in this initial test is this more accurately places people on the scale from superprime down to subprime,” Moynes said. “It does a better job than the tools we’ve been using today.”
However, Ford Motor Credit will continue to test the ZestFinance software.
“It’s going to take us a long time to move forward,” Moynes said. “As we develop these models using machine learning, we’ll continue to test them side by side with our existing model, and only after we go through that entire process over several years, checking the accuracies to make sure they hold up over time,” will the company consider taking it live. That will take at least two years, he said.
“We’re prudent lenders, we make sure we’re thoughtful about anything we do, but the results that came out of the test were interesting to us, and we’re going to study this as we go forward,” Moynes said. “If we get to the end of the train and find out it’s not performing any better, we won’t make the change. But based on what we saw, it looks like a very exciting possibility for us.”
What ZestFinance’s software does
Douglas Merrill founded ZestFinance in 2009, the year after he left Google.
“I wanted to see if we could apply the kind of math we had used at Google to build web pages to the problem of credit for thin-file and no-file borrowers — is there a way to have a financial inclusion play that allows lenders to make more loans but not increase their risk by doing so?” Merrill said. “I was thinking there’s a huge market inequity in that people who are pretty good credit risks pay unfair credit rates. Banks are also hurting to grow their credit, and that’s kind of a strange configuration of events.”
According to the Consumer Financial Protection Bureau, 26 million American adults, or about one in 10, have no credit record, making them difficult to underwrite using traditional methods. This includes millions of millennials who are also starting to buy cars. Last year, new vehicles purchased by millennials represented 29% of all U.S. sales, and that number is expected to grow to 40% by 2020.
Traditional underwriting, Merrill said, does not work well for millennials because it only draws “a few handfuls” of data from credit bureaus and plugs it into logistic regression models.
“If any of that data is missing or wrong, you end up with a score that’s not actually indicative of your true credit,” Merrill said.
Google had built mathematical models to correct for missing and erroneous information on web pages, such as misspellings and missing or wrong tags.
Merrill’s team at ZestFinance took the same basic technology to build the Zest Automated Machine Learning Platform, which tries to compensate for errors and missing information in credit histories.
“It was an unproven hypothesis,” Merrill said. “We’ve come a long way and millions of lines of code to demonstrate that it works in a variety of companies, it works in subprime loans, near prime [and] thin file loans and credit cards,” Merrill said. “And it works on auto loans.”
Other vendors, including CrowdProcess, Upstart and Enova, offer similar software.
Credit standards untouched
Ford Motor Credit shares Merrill’s ideal of helping people with thin credit files, such as millennials and the underbanked, get access to credit, to a point.
“We’re not going to change our risk appetite,” Moynes said.
If someone did not meet the company’s underwriting guidelines because of a poor payment history on installment loans before, they probably would not be approved in the new platform, he said.
However, the new platform could potentially let Ford Motor Credit approve people who have limited credit histories, especially young people.
For investors in Ford Credit's asset-backed bond sales securitized by its auto loan and lease originations, that could mean more data points in assessing the overall credit risk of a portfolio. Ford Credit currently identifies the weighted average FICO range of prime borrowers in its loan pools.
“Everybody who has a 700 or 800 FICO score today started with a first piece of credit, they started somewhere,” Moynes said. “For people who may be just establishing their credit history, this additional data might give us an earlier insight into their creditworthiness, and they may fit into our existing risk appetite with these new insights.”
Some of these new insights already exist in credit bureau data, but Ford Motor Credit's underwriting models do not consider them.
For instance, if a credit applicant consistently uses the same phone number in applications to Ford Motor Credit and other creditors, that can indicate stability, a positive factor in the assessment of credit. People who change their phone number every time they apply for credit could be unstable or a fraud threat.
And occupational credentials, such as medical degrees or technical licenses, can be a positive indicator that Ford Motor Credit’s models have not looked at in the past but that ZestFinance’s software mines.
“Those kinds of things are not that hard to get,” Merrill said. “We’re not off crawling your Facebook page or your Instagram feed. Our argument is you don’t have to do that” to get useful information about potential borrowers.
It will take time for Ford Motor Credit’s legacy loan systems to be adapted to pull in all of the additional data ZestFinance’s program calls for, Moynes acknowledged.
Moynes also pointed out Ford Motor Credit does not approve or deny loans based on a score, so the machine learning platform will not make credit decisions on its own.
“We have a team of highly trained credit analysts, to whom we provide these tools,” Moynes said. “Another feature of machine learning is these tools can learn over time and suggest things. We’re still a long way away from that.”