Transcription:
Penny Crosman (00:03):
Welcome to the American Banker Podcast. I'm Penny Crosman. A fintech startup named Upstart was one of the first online lenders to use AI models to make lending decisions using so-called alternative data such as cash flows, education and occupation data to approve or decline the loans. Today, Upstart's marketplace supports personal loans, auto loans, and home equity lines of credit. The company recently announced a plan to apply for a national bank charter from the OCC and it also announced that CEO Dave Girard will become executive chairman and Paul Gu, the company's chief technology officer, would become its CEO in May. We're here today with Paul Gu to hear more about these recent developments and where Upstart is going. Welcome, Paul.
Paul Gu (00:48):
Hey, Penny. Great to be here.
Penny Crosman (00:49):
Thanks for coming. So when Upstart first launched, I think you and your co-founders had a very strong point of view that traditional banks were leaving consumers behind. They were not making loans to some people who were creditworthy but didn't meet their traditional standard of FICO score plus debt to income ratio, etc. Do you think that credit access is still a problem in this country or do you think it's gotten better?
Paul Gu (01:19):
Credit access is absolutely a problem in this country. I think if anything, we have been surprised by. ... We thought maybe when we started this company and started doing unsecured loans back in 2014, that within a few years we'd see this huge wave of companies following in the same direction, taking AI and applying it to credit. But that really hasn't happened. And I think over the years we've learned that there are a bunch of structural reasons why companies I think are stuck in a certain paradigm of how they do credit. A lot of entrenched systems between the committees that companies have to manage their own risks internally, the sort of financial incentives, innovator's dilemma type incentives, and then the sort of structures of things like rating agencies or third-party funding sources for how capital flows through to end loans. You put all that together and I think you get a system that is really, really slow moving.
(02:17):
And unfortunately the result of that is that too many Americans don't have access to bank-quality credit or they do have access, but it takes a sort of long and arduous process to get there. And in some sense, I think, credit access is a little better, but it's pretty much better by the amount of customers that I think Upstart has served over the last decade or so, which is about 4 million customers. And we really haven't yet seen the kind of total industry transformation that we were expecting, but we still think it's inevitable, just it's maybe a little slower than we would've guessed back in 2014.
Penny Crosman (02:56):
To what extent has Upstart been able to get loans to people who probably wouldn't have gotten one at a bank? Do you have any numbers around that or do you have some examples you could share?
Paul Gu (03:07):
Yeah. When we started the company, our entire focus was on borrowers that didn't have traditionally strong credit profiles and we really started with a focus on people with shorter credit histories. And over time, what we figured out was that actually it's not just these people who maybe in their 20s have recently finished school, recently gotten into the workforce that have this problem. This problem extends to people who maybe have a lower credit score for reasons that are temporary or maybe reasons that don't even make sense. One thing that I remember learning early on is that you could have a low credit score simply because you have not opened too many credit cards and so you don't have that much open credit availability and then you use it once and then suddenly your credit utilization is through the roof. And basically over the years, we just found more and more pockets of people that were not being properly served by the traditional credit system.
(04:04):
And as the scale of the business grew and our ability to get efficiencies in every part of the process improved, including in the cost of capital, including in the operational costs, including in the customer acquisition costs, we found that we had enough efficiencies that we were able to start offering lower rates to not just people that you would think of as traditionally underserved, but really the vast majority of Americans fall into this bucket of what we think of as they have access to credit, but their access to credit really could be much better. It could be at much lower rates. It could be with a much easier process. And so gradually our mission really became about radically reducing the cost and complexity of credit for all Americans. And today we think unlike the Upstart of 2014, today we think we have a product that's really best in class across the whole spectrum and we've added product offerings like HELOC and auto and small dollar to really be able to serve people across that entire spectrum, whether you've got a credit history, you've got a home, you're sort of established, or you're really just getting started or getting reset and everyone in that entire spectrum we think there's an opportunity for them to have better access to credit.
Penny Crosman (05:13):
And I believe you have been accommodating some banks that want to lend to more prime customers, but just going back to people who lack access to traditional credit from traditional banks, what are some of those buckets? Are they young people just getting started who have a good income, so-called Henry's? Are they renters who maybe have never had a mortgage? Are they gig workers who don't have as regular of an income stream? What are some of the typical customers that Upstart serves?
Paul Gu (05:54):
It's all of the above. So all of the types of people you talked about and as I was describing how we got from where we started, which was really a focus on people with shorter credit histories who are new to credit, didn't have that thick file. I mean, that certainly was, I think if you looked at our borrower base in 2014, that was an overwhelming proportion of who we were making loans to. But if you look at the borrower distribution today, to my point, I mean, it looks like a cross section of America because there are people up and down the credit spectrum with long work histories, short work histories, long credit histories, short credit histories who aren't getting the right access to credit. And that's either because of cost or because of complexity, or in most cases because of both. And so we end up serving Americans across the whole spectrum.
(06:42):
And I think that's probably the most unique thing about Upstart is I think you'd be really hard pressed to find almost any other business in the lending industry, whether they're a traditional bank or fintech, that serves that full spectrum. I think most players are either focused on more upmarket, other players are more focused down market. And I think we've kind of uniquely figured out is that some of the problems that we figured out that there are some core problems in terms of cost and complexity that are common across the whole range of Americans.
Penny Crosman (07:15):
OK. So let's talk a little bit about the models that you use. I believe you use a combination of machine learning, neural networks, other types of artificial intelligence to build out these models that make these decisions or that are used in making the decisions. How did you originally develop and train it? Did you take a whole bunch of loan decisions that had been made and feed that into your model to help it understand what to look for?
Paul Gu (07:44):
Yeah, all of our models are what we would call machine learning models. So they're all algorithms that are fairly compute intensive. They're nontraditional in the sense that they aren't straight line models that you could draw on a piece of paper and explain in a simple way. And you're right that one of the big challenges in developing models like these, especially ones that use alternative columns of data, which is a big part of how we get so much predictive accuracy lift in our models, has to do, a big challenge with doing that is figuring out how you actually get started. When we first started making loans, we had to take our best shot at building a model. We did the first version using a bunch of third-party datasets that we had kind of strung together with either finding, I would say, tenuous ways to join data together across different data sets that were imperfect, had to make a lot of assumptions, or by doing certain statistical techniques that also made a bunch of assumptions about the independence of the causal effect of one variable and another variable, we were able to string these different datasets together to build the first version of a model.
(08:55):
And that one was workable. It had certain advantages over a traditional credit model, but it also suffered from lots of disadvantages resulting from all these assumptions. And so not surprisingly, that first model wasn't perfect. It got a lot of loans wrong on both sides, but it was just good enough that we were able to get off the ground and get to the next version of the model with now some borrowers that had been originated in our own ecosystem where we could observe all of the nontraditional variables in addition to all of the credit outcomes. And iteration by iteration of the model, we basically built up the training dataset that we've now got and that process just took a long time. I think it's frankly one of the reasons why we haven't seen, as we were talking about earlier, why we haven't seen more companies rush into applying AI to the core underwriting of credit is just that it's one of those problems that you really can't shortcut.
(09:52):
You have to go through the painful process of building multiple generations of models, waiting the years required to collect the training data, see how they perform. And in the process, risk both financial losses and credit losses that might impact your reputation or other things. And so I think that makes it really challenging to get going. And we certainly had a lot of years of going through that journey early on.
Penny Crosman (10:18):
Sure. And I believe your annual report said that your models analyze over 2,500 data points to assess creditworthiness. What are some of those data points? What are some of the types of data you use that the traditional lending world generally doesn't?
Paul Gu (10:37):
Yeah, there are a few buckets of them. One big bucket is around data that describes someone's, what we think of broadly as their educational and work experience history. So we go a lot into, today I think something that's really relevant is the different kinds of occupations and subsectors that somebody could work in. Link that all the way back through their educational background if they went to college or got graduate degree, what they studied, how that kind of academic journey progressed. So that's one category of what we think of as nontraditional variables. A second whole category has to do with what you can get from nontraditional credit reporting agencies. So instead of your standard credit report, which talks about what you've done on credit, there's lots of data out there that describes someone's typically other sort of financial adjacent interactions with the system that could have to do with debit transactions, that could have to do with rent and utilities transactions.
(11:35):
And there's sort of lots of datasets that describe your usage of financial products or your performance and financial contracts. And then the third bucket is data that gets generated from your interactions directly with Upstart. So that could be more behavioral data points about how you're applying for a loan, exactly what kinds of choices you're making as you go along with the journey. And the thing about almost all these variables is that because the models are not traditional, they're not standard bank type scorecard models. It's not like we're saying, oh, only people that are applying on iPads can get loans with Upstart. There's nothing so simple like that and there's not even any really strong effects across the 2,500 variables that we use. All the effects are really nuanced. They're all found in interactions and combinations of variables. And so the individual contribution from any one variable when you look at something through a diagnostic tool like SHAP values or anything else, almost all the variables have very, very little importance and it is really just by combining them together and running them through one of these machine learning algorithms that has the ability to natively find more of these interaction effects, more of the underlying subtleties of how these variables are related to credit outcomes, that you're actually able to make use of this.
(13:01):
And I think if you were just trying to take some of the nontraditional variables we use and put them into a standard scorecard model, it just wouldn't do anything for you.
Penny Crosman (13:11):
Now one objection I've heard over the years is that the traditional scorecard method you're mentioning, which often incorporates the FICO score and debt to income ratio and credit repayment history, some lenders will say, well, there's no better way to determine somebody's creditworthiness than their past history of paying back loans. What would you say to that if someone said that to you?
Paul Gu (13:39):
Well, I think that would be almost comically misinformed because even if you look at something like what's inside the FICO score, you find that only a part of it has to do with your past performance on credit. A lot of it is just correlated observations about someone's financial life. So the example we talked about earlier, one of the ways in which someone could fall through the cracks of the credit system in a way where if you were to just think about that, you would think like, oh, well, that doesn't make any sense. One of those is like, hey, it actually helps your FICO score if you just open up tons of credit cards and don't use them. And if you were just thinking about advising your friend like, "Hey, here's the path to leading a good financial life." And you didn't know anything about the credit scoring system, I think you would never in a million years think to be like, "Hey, the best thing you can do for yourself is just go and open up 10 new credit cards right now and just keep them in your wallet." That would be crazy.
(14:30):
No one would suggest that. And yet that is a part of the scoring mechanism. And I'm not saying it's wrong empirically, I'm sure they've done great work. It is correlated to credit outcomes, but it is not of course inherently causally related. Having more credit cards is not the thing that causes someone to pay back their loan. The reason that that works in the FICO score and in those traditional models is that it just happens to be the case that this thing is correlated with people paying back and it's correlated probably for these indirect reasons like the fact that somebody has all this sort of open available credit but isn't using it is sort of a good sign. It means that they're not financially desperate. I mean, you could tell sort of a whole story about why that is, but the point is that it's not like in anybody's model, they're only using what you would think of as fundamentally causal first order relationship variables.
(15:23):
Everybody has built their models off a bunch of things that are just correlated, not causal. And once you're in that world, then you should start asking yourself, look, well, why do we stop at just the set of things that traditionally were used in FICA, which was something developed several decades ago? And really there's not a great answer to that other than that's what's been done. We know that works. So the safe thing to do is to stick with that. And of course, I think that's an OK perspective to take, but the implication is that we're never going to get beyond the small set of people who've got traditionally good access to credit.
Penny Crosman (15:59):
And you do use a FICO score when it's available, right?
Paul Gu (16:03):
Absolutely. Yeah. Our philosophy is we want to use every data point we possibly can and the more data that we've got, the more we're going to be able to make an accurate determination of who will be able to pay back. And generally that's going to mean more people get approved at lower rates.
Penny Crosman (16:21):
Other objections that I've gotten over the years to the idea of AI-based lending is that banks can't have black boxes. They have to be able to show how they made decisions. They have to be able to explain and provide good adverse action codes. AI models could potentially pick up bias as we've seen in some famous examples, or large language models can hallucinate and make errors. What do you say to those kinds of objections?
Paul Gu (16:53):
Yeah, two things. The first is that it's certainly true that applying AI to credit creates a whole new set of technology problems that have to be solved, find no objection for me on that. And a lot of the work we've done over the years is exactly in that direction, figuring out how do you build technology that will help you generate adverse action notices out of a model that is a more complex machine learning model? And it's not an easy problem, but it is just a technology problem that can be solved and we've done a lot of work to do exactly that. The same is true on the fair lending questions. It is certainly the case that it is theoretically possible that any model, whether it's a machine learning model or a scorecard model or a human just deciding based on their intuition that you could get bias.
(17:44):
And the correct question to ask is not, is it a machine learning model or is it a simple model or is it a human deciding? The question to ask is, how do you actually objectively in a rigorous and automated way figure out whether the decision-making procedure you're using is producing biased outcomes or not? And so we did a ton of work over the years in the early years in partnership with the CFPB and the later years expanding on that work with some of the leading researchers in the sort of academic context on questions of fairness, just figuring out how to build the actual infrastructure and set of tests around your models that will let you rigorously and precisely answer that question. And then once you've done that, you can actually do a second level of things, which is pretty cool actually, is you can actually build that into your model training process so that if you do discover that in the training, you've got a model that could be both more fair and more accurate that you just build that model instead.
(18:52):
And so we've built a lot of custom tooling into our model training infrastructure that allows us to natively consider questions of fairness and incorporate it directly into the results. And the end outcome of that is we think we can make significantly more loans at low prices to borrowers from every demographic group and that compared to more traditional ways of doing things where you're looking at just a few variables and those variables are very strongly correlated with things like race and gender. I mean, just think about something like incomes, employment, credit scores, historical patterns of default. I mean, these things are enormously different across the different demographic groups. And if you have a model that strongly just says, "Hey, we're only going to consider people with a certain level of income, a certain DTI, a certain track record of successfully repaying credit, certain FICO score, you're going to inevitably end up with pretty large disparities in which groups are getting credit.
(19:57):
And even worse, in absolute terms, you're just not going to be approving very many people in certain demographic categories. And the beauty of what we do is that we can really improve things for borrowers in every category regardless of their race, regardless of their gender. And I think that is ultimately something that's a great win for everyone.
Penny Crosman (20:17):
So let's switch gears to what's new. Upstart recently applied for a national bank charter. Why did it do this from your point of view? What are you going to get out of that?
Paul Gu (20:28):
Yeah, the national bank charter we think is a pretty unique opportunity to up-level the efficiency of everything we do here. It's not a fundamental change in strategy for us. I think we've always wanted to be a technology-first company, really focused on how do we develop better and better AI and apply it to credit so that more borrowers can win in the system. And this isn't a departure from that in any way. For us, this is really about in today's model, we originate loans through hundreds of different partner financial institutions. Each of those institutions has slightly different rules, slightly different processes. They have a different relationship to their particular state and federal regulators and all of those complexities get passed along to us, which then get passed along to our borrowers, either as financial cost or as operational cost. And we think there's an opportunity to streamline that enormously by building a direct relationship with the OCC as one streamlined federal regulator.
(21:31):
And we think that can produce better outcomes really on all the fronts that we care about. I think it can clarify a lot of things in the minds of the regulator so that they can understand exactly what we're doing and we can really be the right face and voice of what it means to do AI lending. And it is also just better for borrowers because there's going to be less cost and complexity that is getting passed through over to them. So a really nice opportunity for better efficiency, not really for us a fundamental change in strategy.
Penny Crosman (22:04):
So you recently became CEO. How do you feel about taking that tough leadership position and what are some of the things you might do with Upstart going forward?
Paul Gu (22:20):
Yeah. Well, Dave [Girouard, former Upstart CEO] and I have been building Upstart side by side for the past 14 years and we've partnered on almost everything as we've done that. So I think unlike for a lot of companies, I don't think of this as some huge change in the direction of the company or some huge change in leadership. Dave is still going to be on the board and very active. And before this transition, Dave and I, we've always worked very closely together on everything. So I think the core vision, the core mission of the company, no change at all to that. We've always known that if you want to do something as disruptive as apply AI to probably the world's oldest business of lending, it's going to take a long time and it's probably going to take more than one chapter of the company. So I think not really a big surprise.
(23:09):
I think it's something we've been thinking and talking about for a really, really long time. I do think that of course in the moment there are always new challenges that the business needs to respond to and new opportunities. And some of that for us is things around, we are in this moment of transitioning from a business that was historically just focused on a single product to a single segment of borrowers. It was mostly personal loans for what we used to call the future prime borrower. And today it's increasingly we have products for every American and those could be people with homes. This could be people buying cars. It could be people who are looking to just bridge a very short-term gap in their finances. And so having the best product for every consumer out there, that's a really big difference. It's a really big change in the opportunity that the company's got.
(24:01):
And I think it's naturally the outcome of a lot of work we've been doing for the past decade plus, but it's now just starting to materialize. So that's a really exciting moment in the company's history. And then I think, of course, we're in this moment with AI and AI being applied to how companies work, how companies build and what kinds of products can be built that touch your consumer. And so we spend a lot of time thinking about that. And I think that's just something that in 2026 is an opportunity that didn't exist five years ago.
Penny Crosman (24:37):
Are you thinking about incorporating large language models and agentic AI into the platform itself or more behind the scenes the way people do their work on an everyday basis?
Paul Gu (24:49):
All of the above. So we already are applying it in a few places, but ultimately it will be all of those things that you said. So if you start from the very innermost layer, of course, like every company, we are thinking really hard about where we can improve the output of our people. And I think in technical rules in particular, there's just already huge gains we're getting from that and we think those are going to get a whole lot bigger. We're very focused on that. I think we have the luxury of being a company that is growing probably in. ... It's one of the faster growing companies in the public markets and we see that continuing for some time. So we are, I think, fortunately in a position where we're not having to deal with questions of, well, what do you do if suddenly every person is able to do so much more and your business opportunity is the same.
(25:46):
In our case, the business opportunity can just be that much bigger and there's so much more to grow into it. So I think we're kind of in an ideal position with respect to bringing AI into the system without it having to raise uncomfortable questions about the role of people and jobs and all that. Then there's a second layer, which we have also been doing for some time and that is bringing large language models or generative models generally into the behind-the-scenes workflow of doing verifications work. So of course, a lot of the work that goes into a loan, especially on some of your secured products, we have to think about things like leans and perfecting things, property records, stuff like that. You have a lot of work that involves people manually reviewing documents and verifying things, sometimes going back and forth with the borrower.
(26:33):
And this is a perfect body of work for AI to be in. And it's not just better because it's more cost efficient, it's better because it's faster. It's something that instead of having to wait for a review, you can get review in real time. And so that's been a big focus area for us and we think it's going to continue to deliver a lot of wins. And then ultimately, I think for user-facing applications, we fully expect to be in market with more user-facing applications of large language models. And there are some early ideas we're working on there that we'll be excited to share more about later this year.
Penny Crosman (27:13):
All right. Well, Paul Gu, thanks for joining us today. And to all of you, thank you for listening to the American Banker Podcast. I produced this episode with audio production by Anna Mints, Adnan Khan and WenWyst Jeanmarie. Special thanks this week to Paul Gu at Upstart. Rate us review us and subscribe to our content at www.americanbanker.com/subscribe. For American Banker, I'm Penny Crosman, and thanks for listening.
