Artificial intelligence in finance has long been overhyped, but AI is making an impact in lending decisions. Online lender Upstart exemplifies how the set of technologies can expand lenders’ inclusiveness without necessarily taking on any more risk.
Paul Gu, a co-founder at Upstart, joins us on the podcast to talk about his firm’s new AI-powered credit decisioning API that is letting banks and other lenders deliver instant credit decisions for auto, personal, and student loans. We discuss Upstart’s performance during COVID as well.
The evolution of AI
To be honest, we have had the tailwinds behind us. There has been a rapid pace of innovation in AI and ML technologies over the past decade. Some of that is frankly powered by basic innovations in the amount of compute that’s available. It’s getting more cheaply and easily available through cloud technologies. Some of the mathematical work that underpins AI has also been advancing. We’re big consumers of academic literature that demonstrates new ways to do things. It’s a great time for us to be in this.
Banks’ openness to AI
It’s still very early stages. I think traditionally banks haven’t wanted to be on the cutting edge of new technologies because they’re risk averse. The way we built this business was starting out as a sort of R&D lab, proving these technologies in more limited settings, putting more of our own capital at risk — to prove AI can be successfully applied to solve problems in consumer lending and get dramatic gains in reduction of loss rates without impacting approvals.
To give you a sense of that, we can reduce loss rates by about 75% while approving the same number of people, compared to traditional institutions. That’s a mindboggling change in the economics of the business. We encourage our bank partners not to just improve their profit margins by reducing losses, but expand the pool of people they are approving. We spent the first few years demonstrating these claims. With strong proof points in hand, we’ve been getting more banks directly adopting our technologies on to their balance sheets and loan programs.
Performance during Corona
We’ve been proud of the way the model and loans have performed and held up during the crisis. For a business founded in 2012 and operating until today, it was one of the most benign decades you could have been lending in. We got the question a lot about how we would hold up in a downturn.
Lenders spend too much time trying to forecast macro. It’s a really hard thing and they should not be in that business. The data doesn’t support only lending to people with higher credit scores. Even if you think a recession is coming, it’s not a useful response to retreat to people with higher credit scores — it won’t save you. Whether there’s a recession or not, we think you should get better an individual risk assessment — not the macro.
No one expected the timing and cause of this crisis. When you look at how predictive Upstart’s scoring of loans against something like a FICO score, we see that Upstart was 5 times as predictive as traditional FICO. When you look at the results of our platform across all banks, we ended up at the end of May with a 5.8% hardship rate. That’s almost 50% less than industry benchmarks.
Rolling out new products
When we think about what to build next, we think about what our bank customers need next and how we can make a big difference in the financial lives of consumers. The first big candidate was auto. It’s a category of loan that’s ubiquitous and most banks participate in it. It’s not efficient, either. People who deserve low rates aren’t getting them — not by a long shot.
We partner with banks with a white-label solution. Our bank partners brand our product and technology. There are other banks that aren’t quite ready to move to an end-to-end Upstart-built process, but see the value of our models. They just want to plug those models into a consumer-facing flow that they’ve built themselves. We wanted to enable that. We’re in the early stages of partnering with some banks on our new credit decision API.