Machine learning credit models help lenders solve an ever-elusive problem: how do you increase the size of your underwriting business without adding risk? The answer: By finding new borrowers overlooked by current credit scoring techniques. They’re out there, especially among populations with a thin credit file or a once-blemished credit history. Just because you have a low FICO score doesn’t mean you’re a bad risk. (The converse is also true: There are plenty of bad borrowers with “good” FICO scores.)
Technology is rewriting the rules of financial services, but the future, as they say, is not evenly distributed. In China, nonbank upstarts AliPay and WeChat have rolled out a torrent of mobile innovations that have put cash and traditional payment methods on the road to obsolescence.
We’re now at peak hand-wringing over artificial intelligence, which is more or less what you’d expect of a technology which in influence and hype has outstripped people’s understanding of how it works. Whose dark and stormy prediction do you buy into? Bill Gates doesn’t think we should panic. Elon Musk thinks we should. So does Henry Kissinger. To get ahead of the issue, Google posted a set of ethical principles it promises to live up to as it pours billions of dollars into its decision-making software systems.
Mobile check deposits, robo-advisors, video tellers. So many innovations profound and mundane have swept through the financial services industry in the last decade. One arena that remains relatively unswept, however, is credit risk modeling. The science of predicting good borrowers from bad still leans heavily on the FICO score, a decades-old logistic regression technique that combines a few dozen variables in clever but fairly simple math from the 1950s.
“The math is not the hard part of machine learning,” says Douglas Merrill, CEO of ZestFinance. “The hard part is getting that advanced math into production.”
Douglas puts his finger on the central problem facing lenders who want to compete using 21st century technology rather than math techniques hashed out in the 1950s. If you’re in need of a good summary of the challenges and profit opportunities of underwriting with AI and machine learning, check out this recent American Banker webinar.
The regulatory landscape for data handling changed substantially on May 25, 2018, when the EU’s General Data Protection Regulation, or GDPR, went into effect. The goal of the GDPR is to protect personal data and promote fair and transparent personal data use practices in a world of increasing data collection. This new regulation broadly affects organizations, government agencies, and companies established within the EU, as well as companies outside the EU that collect, use, or share personal data of EU persons.
At ZestFinance, I've worked with many industry leaders who successfully introduced machine learning (ML) into their businesses. Most of them already have an idea of ML’s potential impact but don't know how to navigate the corporate bureaucracy to get to a yes. Fundamentally, all stakeholders in your lending business want the same thing—positive business impact—but how do you get ML into production to start generating that impact? Here are the top three things I’ve learned that might help you influence your organization to adopt ML underwriting.
Compliance departments often get tagged with the (unfair) reputation that all they do is check boxes, but the reality is that compliance professionals play a critical role in delivering innovation. We are responsible for clearing paths for new products and technology to meet regulatory standards so that they can be responsibly introduced to the market. Compliance teams have always stepped up to solve banking’s big innovation challenges, from ATMs to automated models to depositing checks by phone.
Lenders’ traditional Model Risk Management (MRM) process was set up for conventional models based on logistic regression. Model verification, fair lending analysis, and model risk management documentation are already onerous for these traditional models. Machine learning (ML), which can use thousands of variables and multiple modeling methods, is likely to present a significant challenge for your organization’s established manual MRM process. Do the same principles of MRM apply for ML models? Does the MRM process for ML models need to be as time-consuming and laborious as it has been in the past?