Machine Learning Is the Future of Underwriting, But Startups Won’t be Driving It

Machine learning is not new, it is just another level of system automation that financial institutions have been working on for ages.

But when it comes to successfully deploying machine learning for underwriting, banks are better positioned than startups, according to Sandeep Sood, vice president of software engineering at Capital One.

“Machine learning is absolutely the future of underwriting,” he told Bank Innovation. “But the startups that are doing that today, they take, what I call a fairly shallow data — like social graphs — and make these wild claims that those superficial elements are really the best predictive pieces to figure out if someone should be qualified or not.”

Banks can learn a lot from those startups, Sood said, but the foundation that major FIs have is “much stronger” for building the most accurate underwriting.

“I would argue, and I think Capital One would argue, that 25 years of solid credit card data is the best way to predict the future,” he added.

Regulation is yet another obstacle that startups have to overcome.

“The model that you build for underwriting based on data has to basically predict the future, the degree to which it can predict the future is the value that machine learning provides,” Sood explained. “When you come up with the model, there are risks that the model has a bias, which can be demographic, or racial, or some other.” Even an unintentional bias can subject any financial institution to regulatory scrutiny, and startups may (in many cases) not have the needed resources to test their technology enough, before running it.

But there are exceptions.

Online lender Upstart raised $32.5 million last month in order to “continues to grow its business,” which is underwriting millennials using machine learning algorithms. 

The company announced that it will begin licensing its technology to banks.

“We built models based on modern machine learning techniques that allow us to use more types of data, and use that data in much more complex, higher order ways,” Paul Gu, co-founder of Upstart told Bank Innovation. “We use many variables, including education and employment-related, but also use web behavioral data to catch the subtle behavioral aspects that can predict if a person will default or not.”

In 10 years, every lender will be using machine learning for underwriting processes, Gu said. “Because if you don’t, you are going to have worse models than your competitors, and eventually you’ll be adversely selected,” he added. “A lot of the banks and credit unions don’t already have teams in place to be able to do that, so there is an opportunity for us to help them with our technology.”

The wealth of data from banks, plus the “alternative” technologies from fintechs — sounds like a winning model, which even CapOne may be pursuing. “We’ll be looking at a variety of models,” including building the technology, as well as partnering up, CapOne’s Sandoop said.