While the COVID-19 pandemic has led to a drop in originations for auto lenders and a decline in both new- and used-vehicle inventory, Tricolor Auto Acceptance has experienced a 25% year-over-year increase in its origination volume. As many auto lenders have tightened their credit box amid high unemployment and economic uncertainty, the financier has thrived, largely due to its artificial intelligence and machine learning model, Chief Executive Daniel Chu told Auto Finance News.
The Dallas-based company uses AI to lend to thin- and no-file Hispanic consumers who wouldn’t normally qualify for credit with more traditional lenders, Chu noted. Despite the crisis facing the industry, Tricolor’s portfolio has performed for the last three months at pre-pandemic levels, even with Tricolor tightening its credit box in March, he said.
Chu shared an in-depth look at Tricolor’s machine learning model and the results it brings. What follows is an edited version of Chu’s conversation with AFN.
Auto Finance News: Can you give us a little insight as to how your artificial intelligence and machine learning technology work?
Daniel Chu: When people talk about machine learning, or neural network, they’re talking about the technology enabling you to identify patterns across an enormous amount of data.
In the old days, we would say, “Okay, how many years has this guy lived at the residence, and across all of our data, how does that correlate with default?” We would have one variable and its relationship with default; we would call that linear regression.
Today with neural network, we’re looking at patterns of across not just three or four, but literally hundreds of attributes, and how those attributes correspond to the default. So, for example, we are overlaying national unemployment data and we’re saying, as unemployment data moves around, how are our payment performances moving around? We’re looking at all of the attributes that we collect from our customers. And those new data points, like household size or the amount of time that elapsed between their last two jobs, allow us to see how those have correlated with loan performance during the pandemic.
AFN: How does the machine model provide helpful data, especially during a pandemic?
DC: These artificial intelligence models continue to get smarter, even with data that we collect during the pandemic. We continue to believe that our model benefits from what we would call to some degree a “network effect.”
In other words, as we do more transactions, as we collect more borrower payments, our model is learning and enabling us to continue to underwrite and originate with the confidence that we have the precision and the ability to predict that performance. And we like the fact that we have some technology in place that allows us to, on a real-time basis, continue to learn based on the consumer behavior that we’re observing in the current environment. I think that’s a big advantage. We were fortunate ahead of the pandemic to build those data science and decision science competencies that allow us to be able to execute on that.
AFN: What is the benefit to having access to a robust set of data?
DC: We learn very much in real time what we are observing and then [the algorithm] can adjust how it’s weighing. As you can imagine, during the pandemic, we’re seeing performance look much stronger in large households. Why? Because there’s multiple incomes that are driving that motor vehicle payment. In the event one person has some disruption in their employment, we know in those large households there’s other incomes that can potentially pick up the slack. We’re seeing payments looking much more stable in these large households, and our model learns from that.
If a guy comes in and he has a 780 credit score, for example, you don’t care how many people live in this house, right? We’re going to weigh that [credit score] heavily. Now, because we know in this environment household size is really important, we know when a borrower comes in and he has a lot of incomes, we are going to weigh that more heavily than we maybe previously did.
The more loans you originate, the smarter the algorithm gets and you’re going to be able to continue to modify your underwriting box through something as disruptive as a pandemic. Even though we used to originate a lot of loans to a certain profile, the algorithm might say, “you know what, this profile doesn’t look very good anymore,” so that’s where we trim off that part of the box.
Editors Note: This feature first appeared in the October issue of Auto Finance News, available now.