Machine Learning Is Changing the Game for the Next Generation of Consumer Borrowers
Ken Michlitsch, AIP Alternative Lending Group Portfolio Manager, interviews Upstart CEO David Girouard about the increasingly important role machine learning plays in the consumer lending process.
Machine learning has been embraced by many industries, from online shopping to autonomous cars, but finance has been a relatively slow adopter. With the rise of alternative lending over the last decade, that has started to change to the benefit of consumer borrowers.
With alternative lending, online platforms use a marketplace model to bring together borrowers and lenders. These platforms think about all aspects of the underwriting process through a different lens than traditional lenders do. While traditional lenders focus on the FICO score, a backward-looking consumer credit risk assessment metric, alternative lending platforms use a much broader set of data that can help predict a borrower’s ability to repay a loan in the future based on educational background, current employment, city of residence and bank account details. For a young borrower with a credit history too short for a meaningful FICO score, alternative lending platforms can mean access to credit not otherwise available and at rates more attractive than those offered by credit card companies.
Underpinning this marketplace model is data—access to lots of data and the ability to analyze it quickly. That’s where machine learning comes in. Machine learning facilitates real-time processing of huge volumes of data about loans and ever improving, ever “smarter” models to make sense of it all. In short, the success of alternative lending has been driven by the same boom in data that we've seen disrupt industry after industry.
What prompted your decision to launch Upstart?
The traditional lending model hasn’t changed in many decades. Banks continue to use the same backward-looking metrics and models that they’ve used for years, making credit available to the same types of borrowers at the same terms. That works well for some borrowers, but not all. We realized the types of technologies we were building at Google could make a positive impact on how credit is accessed by more types of borrowers. The tools we were working on could provide a more predictive model that could understand who an applicant is, his propensity to pay something back and his capacity to pay it back. In short, we saw the opportunity to dramatically improve access to credit for people. Far more people in this country are judged as “risky” than actually are.
What type of borrower stands to benefit from the alternative lending model?
Younger borrowers come to mind. We think of ideal candidates as “future prime” borrowers: they don't have blemished credit, but they don't have a lot of history either. They are transforming themselves from being net debtors, often because of student loans or medical bills, to being wealth creators. Machine learning helps us see the trajectory of their incomes.
Can you compare and contrast alternative lending underwriting with traditional bank underwriting?
Traditional models tend to be FICO-centric and rules-based. For example, a borrower must be above a certain FICO score and can't have had any bankruptcies within a certain number of years in order to qualify for a loan. The rate a bank assigns to a borrower who meets those criteria depends on the perceived credit-worthiness assigned by their model. A borrower who falls short on any of these measures is simply denied credit. Traditional lenders have come to use a larger number of variables, relying on linear regressions to inform their views on which variables matter most, but their models have remained fairly static over time.
Alternative lenders analyze thousands of variables, including borrower bank account transactions, monthly repayments, prepayments, delinquencies, and their models look for patterns in the way those variables interact. Much more quickly than a human being ever could, machine learning enables models to “learn” from data as soon as it’s introduced, becoming better and better able to make predictions about likelihood of defaults. In other words, over time models can predict that if ten specific things are true, an eleventh is likely also true. For example, knowing where someone went to school, what he studied, what his tax burden is, where he works, and what his monthly bank transactions are allows a model to extrapolate other things that increase a lender’s confidence about that person’s willingness and ability to repay a loan in the future.
How does machine learning impact other parts of the alternative lending process?
Machine learning has a role to play at every stage of the process: this includes customer acquisition, credit scoring, loan servicing and the mechanism through which loans are repaid.
With customer acquisition, the challenge is identifying the right borrowers, people who need a loan and are likely to qualify for one. Lenders target potential borrowers digitally, but somewhat surprisingly also achieve good results through direct mail campaigns.
Machine learning facilitates a degree of automation in the application process as well, reducing the amount of human oversight and time required to process a loan. For example, if an applicant provides access to his bank account, a lender can immediately verify his education, income and current employer without requiring documentation for each of those three things. This means less paperwork for the applicant to gather, fewer documents for the lender to review and quicker turnaround time for both parties. Over time, this can translate into lower costs being passed through to the consumer.
Loan servicing can also be done in an automated way. Alternative lenders generally use an automated clearinghouse to pull payments directly from a borrower's bank account; no bill is sent and no paperwork is filled out on a monthly basis.
Do you think there are broader applications for machine learning in the lending space?
In a decade we believe that most types of lending around the world will rely on machine learning as opposed to traditional models because the difference in economics is so dramatic. We have fairly reliable data that shows for a given set of borrowers in certain FICO ranges our loss rates are less than half of what we see in our industry. That is an enormous difference, and it's not random. It's a fundamentally different system, one that is transferable to small business, real estate and auto lending.
Dave was formerly President of Google Enterprise and built Google's billion-dollar cloud apps business worldwide, including product development, sales, marketing, and customer support. He started in Silicon Valley as a Product Manager at Apple and was an associate in Booz Allen's Information Technology practice. Dave's career began in software development with the Boston office of Accenture. He graduated from Dartmouth College with an AB in Engineering Sciences and a BE in Computer Engineering. Dave also holds an MBA from the University of Michigan with High Distinction.
Ken Michlitsch is an Executive Director of Morgan Stanley Investment Management (MSIM) and Portfolio Manager for the AIP Alternative Lending Group. He joined MSIM in 2011 and has 21 years of professional experience. Prior to joining the firm, Mr. Michlitsch was a Portfolio Manager and Financials Analyst at a multi-asset class hedge fund. He also co-founded and served as CEO of an online marketplace utilizing a novel price discovery method. Mr. Michlitsch previously was part of the founding team at multiple venture capital-backed medical technology companies, and he contributed to the earliest stage development at medtech companies that were acquired for over $1 billion in aggregate. Earlier in his career, Mr. Michlitsch served as Director of Product Strategy & Intellectual Property at Jomed Inc., as Technical Advisor at Fish & Neave LLP, and as Mechanical Engineer at Lawrence Livermore National Laboratory. Mr. Michlitsch is an inventor on over 50 U.S. patents. He received SB and SM degrees in Mechanical Engineering from MIT, as well as an M.B.A. with Honors in Finance from the Wharton School of the University of Pennsylvania.
A credit score first created by a company called Fair Isaac Co., which later changed its name to FICO, that attempts to predict a consumer’s credit worthiness.FICO score:
The statements herein reflect the views and opinions of one or more third parties. MSIM believes these statements to be an accurate summary of such views and opinions as relayed to MSIM by such third parties. However, MSIM has not independently verified the accuracy or completeness of any such views or opinions, or of any factual statements or other information forming the basis of or contained in any such views or opinions, and accordingly MSIM makes no representation or warranty, express or implied, with respect thereto. MSIM and its affiliates disclaim any and all liability arising from or in connection with your use of or reliance upon any such views, opinions, statements or other information.
This communication is only intended for and will only be distributed to persons resident in jurisdictions where such distribution or availability would not be contrary to local laws or regulations.
A separately managed account may not be suitable for all investors. Separate accounts managed according to the Strategy include a number of securities and will not necessarily track the performance of any index. Please consider the investment objectives, risks and fees of the Strategy carefully before investing. A minimum asset level is required. For important information about the investment manager, please refer to Form ADV Part 2.