How Lenders Can Find Their Next Billion Through Machine Learning

Subscribe to Our Blog

There are tens of millions of Americans -- and hundreds of millions of consumers around the world -- who deserve credit but can’t get it because they lack sufficient data in their credit file for legacy credit-scoring methods to deliver a good prediction. The flip-side is also true: Legacy credit-scoring methods often fail to detect bad borrowers that looked good on paper. U.S. banks charge off some $100 billion a year making bad loans.

Machine learning-based credit scoring is well suited to meet these opportunities, as ML uses more data and sophisticated math to make better predictions of borrower risk. Zest has been advancing the art and science of ML credit scoring since we were founded in 2009. In March of 2019, we announced that Discover Financial Services would be launching one of the largest AI-based credit scoring solutions in the financial services industry using the Zest Automated Machine Learning (ZAML) platform to improve its credit underwriting with fully interpretable and more accurate lending decisions.

At the LendIt Fintech USA conference in San Francisco on April 8, Douglas Merrill, CEO of ZestFinance, and Roger Hochschild, CEO and President of Discover, sat down for a keynote interview with Selina Wang of Bloomberg to delve into how this partnership can provide a roadmap to increasing credit for another billion people. We've posted the complete video and a lightly edited transcript below.



Selina Wang: How’s that partnership working out?

Roger Hochschild: You know, it's working great. As you think about AI and machine learning with financial services, it will transform every part of the business. And so, if you are a bank or in financial services, you need to make sure you're working with the best and the brightest and that's what brought us to work with ZestFinance.

If you are a bank or in financial services, you need to make sure you're working with the best and the brightest and that's what brought us to work with ZestFinance.

Douglas Merrill: Of course, you probably wouldn't have answered if you thought we stink.

Roger: That's right. So we'll see how it goes.

Selina: I heard there's some dispute around this, but how exactly did the relationship come about?

Douglas: I think the dispute is where it started — our first conversation with Discover was when I rudely interrupted Roger's predecessor who was having a cup of coffee and I kind of broke in and introduced myself. It was extraordinarily awkward because I'm anything if not conversationally smooth. And then two years went by and nothing happened until we finally got reconnected. I think Discover has a different perspective of how it started.

Roger: No, I think that might have been the first start. But Douglas met Carlos Minetti, who runs our non-card lending division. We started to build the relationship. We had Douglas speak to our board of directors and then the partnership has continued to build.

Selina: Douglas, let's dig down into this a little bit. What data points are you looking at that Discover didn't already have before, and in test trials what exactly are you learning?

Douglas: I'm probably going to be a little bit careful talking about Discover's data because it's not my data — and just to make sure we’re in sync — I don't actually sell models and I don't sell data. ZestFinance sells a set of tools to help organizations build their own models and then figure out how to deal with the regulators and their internal model validation folks. We sell a toolkit. And in Discover’s case, we also helped build their first generation of models.

It's not always a matter of looking at new kinds of data or going out and buying this new data, although that sometimes happens. It's often the case that you can look at trending data or interactional data that are a little hard to do with traditional mathematics but that you can do with the mathematics our tools support.

Selina: Roger, how are you expecting this to expand not only the bottom line but also customer opportunities. According to some of the most recent publicly available data, customer personal loans were one of the most challenging areas for Discover?

Roger: I think personal loan underwriting is extraordinarily challenging. Part of it is the credit side. Part of it is that what shows up as credit losses can frequently be fraud: manufactured identities, account takeovers, that type of thing. That's one of the first areas to address because of the challenges from traditional modeling techniques. That's one of the first areas we wanted to focus on and build our partnership around. It's a business we've been in since the mid-2000s. It's an important part of our diversification beyond credit cards. But in the credit card business, you can manage the credit line on an ongoing basis. In a personal loan, you’ve just got that one decision up front. The money goes out the door — so you have to make sure you get it right.

Selina: Prior to working with ZestFinance, was Discover already applying AI and machine learning to some of the underwriting capabilities?

Roger: We were on our AI and machine learning journey prior to working with Zest. One of the things that particularly attracted us to Zest was, in the credit area, some IP they have around specific regulatory issues in explainability and fair lending. That's one of the reasons that we focused our partnership on the credit side first.

Selina: You may not know this about ZestFinance but the company works with a variety of industries, whether it's telecommunications, autos, large banks. You've been doing this for several years so how have you seen the receptivity change to what you're doing, when it comes to these traditional industries?

Douglas: I think, to first order, machine learning has had zero impact on financial services, whether in classic financial services or aliases to financial services, for example, telecom which makes a lot of money off financing handsets. I believe that’s largely because no one has had a regulatory strategy. In the last year or two we've started to see more people engaging in conversations. Part of the reason we were so excited to work with Discover is that Roger and his team are extremely forward-looking in a way that most folks are not. And now they are beginning to use machine learning beyond just fraud to do more complicated underwriting, particularly focused on synthetic identities and things like that. Obviously, for my company’s sake, I hope everyone joins the bandwagon, but we'll see.

Selina: Talk to us about why you chose to work with ZestFinance and how the culture of a small Los Angeles-based startup meshes with that of a large traditional bank.

Roger: As we think about talent, not everyone wants to come work for Discover, shockingly, so partnerships are a core part of our strategy. We focus on building our own capabilities but we want to make sure that we're working with the leading-edge companies and frequently, they're out here on the West Coast. I grew up in San Francisco. I spent my summers in high school coding for a startup way way way long time ago. So I'm familiar with that small company culture. I think it's really been about having a shared set of values and that's where we've been very aligned. We have a great team internally focused on building our AI/machine learning capabilities. We have a team of 200 in Shanghai with advanced math degrees but we know that, to succeed, you've got to partner with the best and the brightest. And it's really been very easy and seamless working with Douglas and the team.

To succeed, you've got to partner with the best and the brightest. And it's really been very easy and seamless working with Douglas and the ZestFinance team.

Selina: Is part of that driven by the fact that it's hard to recruit top talent when you're fighting with all the big tech companies out here?

Roger: It is hard to recruit the top talent no matter how good you are. Maybe Google or Amazon have the scale to try and do everything themselves. I think very few of the rest of us do. We have great talent at Discover but we don't have all the great talent and that's where working in partnerships is — in my view — how to drive our business forward.

Selina: Douglas, ZestFinance was founded about 10 years ago and in that span of time, there have been major developments in AI and machine learning. So how has your technology changed and the data points that you're using changed?

Douglas: So we only use data from our clients. We sometimes recommend that clients buy additional third-party data but our general finding has been that I think none of our clients have ever added additional data for the first-generation model. From a data perspective, we're using what you already have more effectively, which is a big cost win. As algorithms get revised you add them into the toolkit but that’s kind of not that hard. That’s just work. Where we're devoting a bunch of our time is, over time, how regulators and model validation folks think about the processes of validation and the processes of understanding whether this model is safe or not, while they're doing their job of assuring safety and soundness.

Selina: There's been a lot of discussion around the explainability of an algorithm and whether Zest’s data can explain why a customer from Discover was rejected or not. How does Zest make it so that the model is explainable and so that it's not a black box?

Douglas: The short answer is that it's hard and it's right at the center of our IP. There have been a lot of developments in explainability, particularly in academic research over the past couple of years that have done some really interesting pieces of work. And that’s super helpful for anyone trying to do explainability. In our case, what we want to do is make it transparent to the modeler and transparent to the user. So we have an AI model that inspects the model that you're building to figure out the various things that should be explained. Obviously, you have to do it right. You have to get adverse action right, which means if I deny you, I have to tell you the five reasons why I denied you and they ought to be correct. And if you're doing things like mortgages — we're working on one of the GSEs — you have to do a really good job of fair lending, which is to say, make sure that you're not even accidentally discriminating. And that requires another degree of explainability. All-in-all, that is the game. The modeling’s, relatively speaking, the easy part. The regulatory stuff and explainability stuff is the game.

Selina: Roger, what are your discussions like with regulators right now? How are you seeing them react to the use of AI in a very high-stakes area like credit underwriting.

Roger: It is top of mind for all of our regulators, whether it is the Fed, the FDIC, the CFPB, and they are on a rapid learning curve. They bring it up at every meeting. We've done briefings for the regulators. They know it's coming. I think they don't necessarily have the capabilities in-house to understand it the way they did some traditional modeling techniques. On one hand, they're excited about the power of these models to open up credit for more people. But they have the traditional concerns around fair lending and discrimination, so I'd say there’s uncertainty and a real desire to learn more as quickly as possible.

Selina: Do you think that the public-facing rhetoric of regulators is different than what's being said behind closed doors? At least publicly, it doesn't seem clear whether they're endorsing or not endorsing the use of algorithms in making lending decisions.

Roger: I think they're realistic. They know that these techniques are so powerful and that they are coming. And that they will take over financial services. So it's really about how can they adjust their models, their ways of doing business to cope with this new world and accelerate that learning curve.

Douglas: It is expected that fintech folks, like the folks wearing hoodies, are going to insult the regulators and say that they're terrible. It's been my experience with all the regulators that I've talked to and tried to help educate that they’ve been trying to do their duty to the system and recognize, as Roger said, that this stuff is coming so they need to find a way to get in front of it. That's an amazingly positive perspective.

Selina: Are you finding yourself playing a role in terms of educating the regulators?

Douglas: I'm in D.C. roughly twice a month training or working to educate regulators and folks on the Hill, but mostly the regulatory agencies. I spend time with all the major agencies.

Selina: Do you think there's a quiet push to prevent the technology from being used widely because there are so many concerns, as you mentioned, around unintentional discrimination? There have been cases where the AI algorithms have been applied and it hasn't been explainable.

Douglas: So I have not noticed a sub rosa push to stop the development. Pushing against the tide is kind of a hard task. I do think that there is a legitimate question of ‘Do you have explainability and is it correct?’ And I think we've seen lots of examples where people asked the explainability question very narrowly and so the answer they got was ‘yes.’ But, in fact, in practice it wasn't true. You see that a lot in image search. So I don't think anything is a slam dunk yet, but no, I haven't seen a sub rosa argument.

Selina: As machine learning and AI continue to advance in their development, how do fintech companies balance innovation with this explainability that we were talking about. Neural networks are, for example, incredibly complex. How do we make sure we don't get to a point where it's not unexplainable?

Douglas: I think the answer is: We just don't. If people decide that it matters to be explainable, they will find ways. For example, neural networks actually are explainable. And they're not materially harder to explain than a support vector machine. There's lots of hard math out there. But there have been researchers who have done a great job of figuring out how you open those black boxes. It's just hard. And I think what some people will find out is, I think I’m putting words in Roger's mouth, that it's not worth finding it out for yourself when you could find a couple of people who already know how to do it and license what they do.

Selina: Roger, how would you respond to that when you were thinking about this partnership and the use of this technology?

Roger: I would say first to the question about the regulators that the explainability, the discrimination testing, those are absolutes. I mean there are downsides to being in a heavily regulated industry. I actually have started to come to embrace the upsides. I think that the heavy regulation around parts of financial service sales are barriers to entry and are keeping some of us old-line companies from being disintermediated as you've seen happen to countless industries. So those are absolutes and I don't think they're going to change. But that's why you need to make sure you are working with those few who really understand how to balance the regulatory absolutes with creating the potential for massive amounts of business value by making better lending decisions enabled by the new models and the new technology. And that's what got us focused on this partnership.

You need to make sure you are working with those few who really understand how to balance the regulatory absolutes with creating the potential for massive amounts of business value by making better lending decisions enabled by the new models and the new technology. And that's what got us focused on this partnership.

Selina: And shifting gears a little bit. I mean what do you think is the future of the credit score in America?

Roger: I think credit scores will always be out there. I think sometimes the predictive nature of alternative data is a bit overstated. One of the things that we focused on is really pushing the credit scores out to consumers and making it so that they understand and they can start working on managing their credit scores themselves. That's why we pioneered putting it on the statement for every cardholder. We then shifted so that anyone can get their credit score from Discover, even if you're not a customer and understand what those things are that you need to change to move your credit score higher.

But I think scores will change more slowly given their importance in all the underwriting models. You hear a lot of buzz about new scores. My experience with traditional underwriters is they're very slow to make changes. And my experience with some of the newer underwriters is that brand new model that's supposed to be better than anything out there can frequently end in tears for some of these newer underwriters.

Selina: Douglas, how do you respond to that? What do you think is the future of the credit score?

Douglas: So when I look back at arguably the most central innovation of the 20th century, it's two dudes, one named Fair and another one named Isaac, who figured out how to use this credit bureau thing and this new piece of math called logistic regression to move the world from totally bespoke human interaction where I would go to a bank and I would talk to you across the desk and would try to convince you give me a loan. These dudes, Fair and Isaac, figured out how to use a score and yield automated underwriting, which I believe is one of the most important innovations of the 20th century because all of a sudden, over that period, credit goes way up. Losses go down. You have to love that.

The question is that in about 1980 credit availability plateaus and losses plateau. Maybe you've exhausted a lot of the value there or you've gotten to a place where people are just churning too much. The question isn't to me, ‘Will there be a credit score?’ The answer is, I hope, yes. Because it's an important signal to people of how their credit is doing and an important prop to get someone moving to improve it. And the question is, like, what's in that score. I 100% agree with Roger when he asserted that a lot of times scores based on internal data are more hype than reality and more smoke than mirrors. Because a lot of times it is really hard to build credit models and it's really easy to do it poorly and when you do it poorly, something bad happens at some point in the future.

Selina: I’m curious what you make of the innovations in some other countries around the world, countries where they didn't have a traditional FICO score like we have. For instance, in China there's a social credit system being built. Some people call it dystopian. Other people say that it will bring millions of people into the financial system that weren't there before. So I'm just curious what you make of these other global innovations?

Douglas: So there are roughly three alternative credit scores in China. One’s from AliPay, which is the dystopian one. One’s provided by Baidu and one is provided by JD Finance. Baidu and JD Finance use our technology to provide those scores, so I'm a little bit required to say that they are awesome. [Laughter]

There are a lot of people in all countries — 80 percent of Chinese but also 40 million people in the U.S. — who are poorly served by their current bank and some number of those people actually are good credit risks who have been marked as bad credit risks. And there is a moral beauty in finding people who are being mistreated by the system and fixing that. So my hope is that it will continue to crank.

Selina: And on that point, as you see this partnership play out, how are you forecasting the new customers you can drive?

Roger: Just to circle back, I find the concept of the social credit score or the social score very interesting. I would point out that it's already here in the U.S. It's just distributed. You know what someone's rating is on eBay. You know what their rating is on Uber. You know what their rating is on Airbnb. It isn't too big a leap of faith to see those starting to come together. When you're transacting with people you don't know, when you're no longer borrowing face-to-face with someone who's been in your community forever, you do need some way of knowing who am I dealing with and are they a trusted counterparty in that transaction. It'll be interesting to see how it evolves here in the U.S. as well. As we look at our lending and leveraging the capabilities that we're building in working with Zest. I think the easiest way to think about it is, it just helps you swap in and swap out.

There are people that we are approving in our traditional models that we shouldn't be approving. There are some people we are declining that we shouldn't be declining. There are more people that we were declining so the net result is actually an increase in credit availability, but we're very excited about the early potential we're seeing.

Selina: Discover has been pretty open about working with Amazon for moving some of its capabilities to the cloud. So how much of your business has actually been shifted to the cloud and has that helped partnerships with companies like this?

Roger: I would say similar to most old-time companies, it is a journey to migrate your applications to the cloud. But from a data and analytics standpoint, everything we're doing is cloud-based including all of the work we're doing on the partnership side and quite frankly, there are no alternatives. You cannot get the speed, you can't get the performance. You don't have the tool kit and you just can't get the scale to store the vast amount of information — the vast amount of data that we need — and that we're leveraging working with Zest.

Selina: And how are you in the industry looking at this potential concentration risk? For instance, there's only really four players you can go to for migrating your cloud services. Is that a concern?

Roger: I think over time companies will have a multicloud strategy so you see a lot of the migrations and Douglas is much more of an expert than I am. People will pick a single cloud provider for the transition and that's the fastest path to value over time. Over time you may not want that reliance so you'll have a multicloud solution. I'm not sure there's any more risk than if your data was concentrated in your own two data centers. Most large companies don't have more than four data centers that they were using for their traditional approach. I would view it as — actually given the redundancy that's built in — less risky than where you were before migration.

Selina: Douglas, would you agree with that? Do you think regulators view it similarly?

Douglas: I think it's still early days. I think the regulators are still finding their way around the cloud. But overall, it is so much safer than running your own data systems and data services that everyone is going to get there. The cost also is just so radically different. Shouting against the tide doesn't help you. Eventually everyone is going to be on some variant of cloud.

Selina: Roger, over time how do you see AI being applied more broadly across Discover’s services?

Roger: I see AI and machine learning touching every part of our business. Douglas talked about the earliest applications around the fraud side. I think that's been the fastest path to value for organizations. You've all had the experience of either fraud on your credit card or being declined for a transaction that was you — and you think, ‘Gosh, how could the company not know this was me?’ A lot of potential on the fraud side. A lot of potential in the marketing.

I see AI and machine learning touching every part of our business.

But some of the benefits are more subtle. If you have the experience of calling a call center and that representative somehow really understands your issue, you don't have to get transferred to level 2 or 3 support, you get the help you need, There's probably an AI that's driving that customer interaction and providing the information for the representative. We see it as truly pervasive to every touchpoint we have with the customer. We've done things such as shifting around the 100-200 people we had doing call-quality monitoring and listening to a sample of calls. Leveraging AI we now listen to every single call and cue up the ones that have compliance issues so team leaders can talk to their representatives.

Selina: What advice do you have, Roger, to a startup that's interested in being integrated into a large organization? You mentioned that there were discussions a few years ago and back and forth. What really makes it to the finish line?

Roger: Two things: it's really what you have to sell and who you sell it to. It's really about solving a problem. If you can show convincingly that you have a product or a solution that's going to create value for a big bank, that makes the sale much easier. Part of it is understanding their business model, listening and reading the analyst reports, and knowing what are the pain points for that organization, what's going to help them grow faster, what will help them take out costs. The second thing is figuring out the right person to talk to in the organization and I'll use Douglas's failed sales approach as an example of that. People frequently think that if you can just talk to the CEO or talk to someone really senior that'll solve all your problems. That usually isn't the case. They are probably not making the decisions. Sure, their support would be helpful. But they're not making the buying decisions. It's finding the right person in the organization and that's the advice I'd share.

Selina: Douglas, do you agree with that? Any additional insights?

Douglas: I think that's right. When we found the right person in Discover, things moved. Quite quickly, actually. I think I would give the same advice that he gives any other technology person, which is ultimately, selling is hard. And it's uncomfortable for introverts and many technologists are introverts. And you have to stand by your guns. If you have a solution which provides value, someone's going to write you a check. And if you don't, go solve some other problem.

Selina: So at Bloomberg News at least we've been very focused on the forays of Apple, Amazon and Google and their pushes into financial services more broadly. Of course, these companies are partners. Some of them. But how do you view their deeper pushes into financial services? For instance, Amazon is already in lending; we just had the announcement about the Apple Card.

Roger: Yeah I think that — for better or worse — there is a regulatory moat around large parts of the financial services. Alternative lenders have been very successful getting into personal loans because those are easy to originate, to fund outside of a bank structure. You don't see that in the credit card space. So Apple is launching their card. They are leveraging a traditional credit card network. And leveraging a traditional bank to bring that product to market.

I don't see that regulatory environment changing. The answer is really about those partnerships, and we are out there working with just about any one of those big companies you name looking for ways to either leverage our payments network and provide payments functionality or leveraging our bank charter. There are a lot of companies that think they want to get into banking. Once they understand the margins, once they understand the regulatory overhang and what you have to do, they pretty quickly come around to partnering with a bank as an easier way to get there and that's where come in and where we've seen great success.

Selina: Doug, how would you view the trends of late?

Douglas: I was asked recently, ‘Do I expect to see a flood of fintechs applying for bank charters or a flood of fintechs becoming banks?’ And my response without thinking was ‘Oh my God. I hope not.’

Selina: There are a few out there.

Douglas: Yeah, maybe one. It's really hard to be a bank. It requires a lot of work. One might be willing to try and insult the current banks if one wants to, but when you look at it, it's hard to run a payments network. It's fragile and has to be fast. And it’s really low margin. Everyone looks at banks and says, ‘Oh, my god, they're all these high margin business.’ Yeah, but that's because the cost of capital is basically zero, which helps a lot. My hope is that the fintech side of the fence will busily help the banking side of the fence continue to grow and be more profitable, etc., because that just means more sales targets for us. I cannot imagine a world in which I would want to become a bank.

Selina: Alright, well that wraps it up. Thank you so much for listening, for your time and for the great conversation.