A Capital Calculator: Upstart Credits AI with Advancing Loans

With two early hits and the promise of more to come, it feels like a whole new ballgame in lending for Grant Schneider.

The AI models he helped create as vice president of machine learning for Upstart are approving more personal loans at lower interest rates than the rules traditional banks use to gauge credit worthiness.

What’s more, he’s helping the Silicon Valley startup, now one of the newest public companies in the U.S., pioneer a successful new hub of AI development in Columbus, Ohio.

A Mentor in the Midwest

Schneider’s career has ridden an AI rocket courtesy of two simple twists of fate.

“In the 2009 downturn, I was about to graduate from Ohio State in finance and there were no finance jobs, but a mentor convinced me to take some classes in statistics,” he said.

He wound up getting a minor, a master’s and then a Ph.D. in the field in 2014, just as machine learning was emerging as the hottest thing in computing.

“Then I read about Upstart in a random news article, sent them a cold email and got a response — I was blown away by the team,” he said.

A Breakthrough with Big Data

Schneider signed on as a data scientist, experimenting with ways to process online loan requests from the company’s website. He trained AI models on publicly available datasets while the startup slowly curated its own private trove of data.

The breakthrough came with the first experiment training a model on Upstart’s own data. “Overnight our approval rates nearly doubled … and over time it became clear we were actually moving the needle in improving access to credit,” he said.

As the business grew, Upstart gathered more data. That data helped make models more accurate so it could extend credit to more borrowers at lower rates. And that attracted more business.

Riding the Virtuous Cycle of AI

The startup found itself on a flywheel it calls the virtuous cycle of AI.

“One of the coolest parts of working on AI models is they directly drive the interest rates we can offer, so as we get better at modeling we extend access to credit — that’s a powerful motivator for the team,” he said.

Borrowers like it, too. More than 620,000 of them were approved by Upstart’s models to get a total $7.8 billion in personal loans so far, about 27 percent more than would’ve been approved by traditional credit models, at interest rates 16 percent below average, according to a study from the U.S. Consumer Financial Protection Bureau.

The figures span all demographic groups, regardless of age, race or ethnicity. “Our AI models are getting closer to the truth of credit worthiness than traditional methods, and that means there should be less bias,” Schneider said.

Betting on the Buckeyes

As it grew, the Silicon Valley company sought a second location where it could expand its R&D team. A study showed the home of Schneider’s alma mater could be a good source of tech talent, so the Ohio State grad boomeranged back to the Midwest.

Columbus exceeded expectations even for a bullish Schneider. What was going to be a 140-person office in a few years has already hit nearly 250 people primarily in AI, software engineering and operations with plans to double to 500 soon.

“Having seen the company when it was 20 people in a room below a dentist’s office, that’s quite a change,” Schneider said.

GPUs Slash Test Time

Upstart has experience with nearly a dozen AI modeling techniques and nearly as many use cases. These days neural networks and gradient-boosted trees are driving most of the gains.

The models track as many as 1,600 variables across data from millions of transactions. So Upstart can use billions of data points to test competing models.

“At one point, these comparisons took more than a day to run on a CPU, but our research found we could cut that down by a factor of five by porting the work to GPUs,” Schneider said.

These days, Upstart trains and evaluates new machine-learning models in a few hours instead of days.

The Power of Two

Looking ahead, the company’s researchers are experimenting with NVIDIA RAPIDS, libraries that quickly move data science jobs to GPUs.

Schneider gives a glowing report of the “customer support on steroids” his team gets from solution architects at NVIDIA.

“It’s so nice for our research team to have experts helping us solve our problems. Having a proactive partner who understands the technology’s inner workings frees us up to focus on interesting business problems and turn around model improvements that affect our end users,” he said.

Early Innings for AI Banking

As a startup, the company built and tested models on GPU-powered laptops. These days it uses the cloud to handle its scaled up AI work, but Schneider sees the potential for another boomerang in the future with some work hosted on the company’s own systems.

Despite its successful IPO in December, it’s still early innings for Upstart. For example, the company started offering auto loans in September.

Going public amid a global pandemic “was a very surreal and exciting experience and a nice milestone validating years of work we’ve put in, but were still early in this company’s lifecycle and the most exciting things are still ahead of us,” he said. “We’re still far from perfectly predicting the future but that’s what we’re aiming at,” he added.

Visit NVIDIA’s financial services industry page to learn more.