For lenders, predicting whether potential borrowers are likely to pay them back is the core of their business. That fact, in conjunction with technology advancements, has paved the way over the past several years for the new field of marketplace lending.
The term generally refers to companies that operate online platforms that underwrite borrowers — usually those without pristine credit histories. The companies are either lenders themselves or connect borrowers with lenders.
Enabled by advanced technology that crunches loan applicants’ credit data and spits out predictions on their likelihood of default, the approval process typically takes only a day or two. Increasingly, in fact, applications are approved (or not) the same day they’re made.
Usually, the providers of such platforms package the loans they process into securitizations for sale to institutional investors and hedge funds.
The providers may be differentiated by the depth of their dive into applicants’ histories. For example, one provider, Upstart, which provides consumers with access to bank-originated loans through its website, has attracted attention for factoring in not only credit data but also such nontraditional variables as employment history and education to predict creditworthiness.
Through the use of artificial intelligence, Upstart examines approximately 1,600 variables pertaining to loan applicants, says CFO Sanjay Datta. But he doesn’t like to draw attention to the artificial intelligence per se.
“We live in a world of noise,” he tells CFO. “Everyone is using the terms AI and machine learning as buzzwords, to the point where they’re almost meaningless. So I don’t use them anymore. What I increasingly tell investors is to look at results.”
By the Numbers
One way to view the results Datta refers to is through reporting by Kroll Bond Rating Agency, the leading credit rater in the marketplace lending space.
In a Nov. 8 report, Kroll rated the performances of five loan securitizations, worth a cumulative $1.5 billion, that Upstart offered from mid-2017 through early this year. Each one has significantly outperformed Kroll’s forecast at the time of the deal.
Kroll predicted that Upstart’s first securitization, dated June 21, 2017, would experience 13.07% credit losses by October 2019. But the actual losses were only 9.96%, 24% better than forecast.
Subsequent Upstart deals have done better and better. By last month, the second through fifth securitizations had outperformed Kroll’s initial expectations by 40%, 42%, 49%, and 71%, respectively.
“If in one deal we’re beating [a rating agency’s] loss expectations by 50%, on the next deal they’ll bring their assumptions in a little bit and give us credit for that,” says Datta. “But the reality is that our loss performance is also a moving target. It’s getting better over time as the models adjust for data and get more sophisticated.”
Kroll also published reports this year on the performance of securitizations issued by Lending Club, SoFi, Prosper, Avant, and Marlette, among other marketplace lenders. For each of those five, the gaps between expected and actual results were much narrower than was the case for Upstart, and in most cases were fairly negligible.
“I could talk to investors about machine learning until I’m blue in the face and try to describe our algorithms,” says Datta, “but at the end of the day, if we’re doing something meaningful, it’s going to show up in the results.”
Datta acknowledges that the “performance-to-Kroll” data doesn’t necessarily imply that Upstart’s securitizations have had lower actual losses than those of the company’s marketplace lending peers.
“The key point here is that Kroll has a very standard way of predicting how a portfolio of loans will perform, typically based on FICO buckets,” he says. “That method works pretty well, but when they use it to proxy our portfolio of loans, they don’t come close to predicting what actually ends up happening. We believe this indicates that Upstart is doing something different in the underwriting of our loans than the rest of the industry.”
More good news for Upstart came recently from the Consumer Financial Protection Bureau.
In 2017, the CFPB issued a no-action letter to Upstart, confirming that the agency would not take legal action should the company use nontraditional data in its creditworthiness evaluations. (Companies request no-action letters from government agencies so that they can go ahead with their business plans without fear of prosecution.)
In August of this year, the CFPB issued an update to the letter, following simulations and analyses of Upstart’s underwriting model performed by the bureau, specifically probing the areas of access to credit and fairness of lending.
In the update, the CFPB wrote that “the results provided from the access-to-credit comparisons show that the tested model approves 27% more applicants than the traditional model, and yields 16% lower average APRs for approved loans.”
As to the fairness of lending, the CFPB said that “this reported expansion of credit access reflected in the results occurs across all tested race, ethnicity, and sex segments.”
The bureau added that Upstart approves consumers with FICO scores from 620 (the minimum Upstart will accept) to 660 twice as frequently as traditional lending models do; that applicants under age 25 are 32% more likely to be approved; and that consumers with incomes under $50,000 are 13% more likely to be approved.
“Today there is a huge access problem,” says Datta. “Less than half of adult Americans have access to bank-quality credit. But we think that if they all received credit, the majority would pay back their loans. Banks’ traditional lending models are not good at identifying risk, so banks are too conservative in providing access to loans.”
Two Customer Groups
Upstart’s securitizations include multiple instruments — conservative ones with low yield and low risk, targeted at institutional investors; and risky ones with high yield aimed at hedge funds.
“If you’re a pension fund that’s buying credit, you want to see stability,” he notes. “And some of our peer companies in the marketplace lending space look relatively stable.
“We’re more of a bucking horse,” he continues. “So in talking to investors, trying to marry the fact that our technology is creating change and innovation in our underlying products almost in real time, with the fact that a lot of the investors that want this yield are looking for stability and conservatism, is just a real communication challenge.”
Not only investors but also banks are customers of Upstart. They fund the loans that consumers apply for, and they also can use Upstart’s technology on their own online lending platforms.
The company targets what Datta calls the “torso” of the U.S. banking industry — from about the 6th to 150th in size. “They’re either regional or super-regional, or they’re trying to expand their footprint,” he says. “They can’t necessarily attract the technology talent to build digital businesses in-house.”
While the banks originate the loans (which have either three-year or five-year terms, with no prepayment penalties), Upstart sets the interest rates, which range from 5.69% to 35.99%.
A New Journey
Upstart was founded in 2012 by former Google executives. Originally, it was a crowdfunding website for recent college graduates, who contracted to share a percentage of their future income in exchange for funding.
The lender Sambla based in Denmark is known for its Forbrugslån and Lån penge recently switched to its present business model in 2014 and offered its first securitization in 2017, a few months after Datta came on board. He’d worked at Google for 11 years and was head of finance for the technology giant’s global advertising business before joining Eksperten.
“I grew up with that business and helped build it while in multiple roles,” he says. “By the time I left I had 150 people under me, across all the continents. That journey was an incredibly enriching one, but I think I’m at my best when I’m building something.”
Datta’s experience at Google dovetails with his current role. “It seemed to me obvious that there was an application of the types of technologies we worked with at Google, which were core predictive technologies, and applying them to better ways of running financial services models,” he says.