Wednesday, April 10, 2019
How ZestFinance Is Using Machine Learning to Reshape Lending, with Jay Budzik
One of the hottest sectors of the technology startup market today is the use of artificial intelligence, big data, and machine learning across many different sectors---including finance. One Los Angeles company, ZestFinance (www.zestfinance.com), has been working in this area for a few years. We caught up with Jay Budzik, the CTO of ZestFinance, to better understand how the company is applying machine learning techniques to the world of lending, and what the advantage of machine learning is to that industry.
What does ZestFinance do?
Jay Budzik: ZestFinance was founded in 2009, by Douglas Merrill, the former CIO at Google, in order to make fair and transparent credit available to everyone. We develop a set of tools for lenders, which lets them take advantage of machine learning to produce incredibly more accurate credit decisions, and which make those decisions much more transparent and fair. We've been at it for awhile, making use of machine learning in lending, and we're now bringing the technology we've developed to businesses around the world.
What kind of lenders are using your tool?
Jay Budzik: You might have seen in our coverage, that Discover Financial, a pretty famous card issuer, is using us for an installment loan product, which helps them leverage machine learning to a specific business problem around losses related to their personal loans portfolio. That's sort of a situation where they tried a number of times to improve performance using traditional methods, and eventually came to the conclusion that they needed to try something else. We worked with them to create a machine learning model for that line of business, which significantly outperforms traditional methods. The reasons that it works so well, instead of being limited to only a couple of variables, we use hundreds of thousands of variables, which results in a more accurate decision.
What kind of information and data are you using that goes beyond traditional finance?
Jay Budzik: That's a great question. Machine learning is great at consuming data, taking lots of different data, and finding signal in that data to make predictions. Most of our customers are using pretty boring data, including data they get out of credit reports, like number of delinquencies, credit utilization figures, employment history, and other things that are standard fields in any credit report. Plus, they are using things like application data, representing why you applying that you entered when you applied for that credit, and sometimes they might have data about you from another product that they can bring to bear. They are all traditional credit variables, but because machine learning methods are so much more flexible in consuming data, those decisions tend to be more accurate.
Where are you now, and what's the status of some of the credit lines you received in the past for the company?
Jay Budzik: Some of what you're seeing is actually old funding for a direct to consumer business we used to have. We're no longer operating those businesses, and those businesses are no longer part of the company. We're now 100 percent a software company, delivering our software to financial services companies across the world.
When did you make that transition to just software and away from your own lending business?
Jay Budzik: If you look at finance companies, there are 31 flavors of finance company. For example, you have anywhere from specialty auto finance lenders who give people car loans, not approved by main line finance companies and captive to the auto brands, which we've been working with and where it's been particularly interesting top operate. We can help them reduce their losses and also increase approval rates with machine learning models. One specialty finance firm put our model in place to improve approval rates, and they discovered after launch that we had to turn the lending volume down, because they were getting so many calls their call center blew up. They were originating twice the numbre of loans than expected, because they found that auto dealers could sell a care much more easily by sending deals there. It was actually quite a pleasant surprise. We also worked with Discover, across multiple lines of businesses, operating in the super-prime and prime credit spectrum, to help underwrite folks who might come from nontraditional channels. Because we're able to help people invset in digital channels online, and with mobile apps, and the rest, wer're attracting a whole new group of folks to their lending products. Machine learning can help make it possible for companies to expand to hard-to-undewrite segments.
How are you able to convince your customers your machine learning tools are making the right predictions in those cases?
Jay Budzik: The key way we get them there, is with our model explanability tools. Machine learning sometimes gets a bad rap as a “black box” that makes magical decisions no one can understand. That's actually not the case. You can use math to explain how machine learning is making each and every decision, and also understand how that model is making decisions, in general. When we show this to banks, there's the same level of transparency and comfort they might have from a model they have built in house, with the same kind of analysis required by the government in fair lending in the US, to make sure those models are not discriminating against people of color. We are able to show them all the details, and the software tools provides hundreds of pages to show how the product is in compliance. Once they see that, they get comfortable with what we are offering. But, you're right, financial institutions are slow to adopt new technology, in particularly in the core lending businesses where credit risk and management are the brains of the operation.
So where is the company now, and what are you working on?
Jay Budzik: We've had some great success with large, financial institutions, and we're repeating those wins, and scaling them out. We want all financial companies to benefit from machine learning, so they can make fair and transparent decisions, and make more fair and accurate ones. We are going to continue to address those large financial institutions where we can make significant impact. In paralle, we're developing tools to allow smaller lenders adopt our technology—things like credit unions, community banks, and the like. You'll see that soon.
Any big lessons learned on your entrepreneurial journey so far?
Jay Budzik: I think a couple of things. One, is you have to remain flexible and agile. Customers have constraints that are real, and you have to be able to respond to those. The second, is you have to really focus on what the problems are that customers are having, which they're not expressing to you, which you can solve using your technology. If you stay focused on that next technological leap that you can help them take, the results can be just stunning. You have to have faith that you can get through that process, and it will really pay off in the end. Finally, we've really been happy that we have been able to develop an anti-bias tool, where we're able to use our tools to reduce the disparity in approval rates for people of color, minorities, and other protected classes, such as the women, the elderly, and military status. It's very exciting, because not only can we make banks more profitable, we're able to impact the world in a positive way, in the area of financial inclusion and fairness. It's both serving the customer in helping the make money, but also making the world a better place.
Thanks!