Research analyst at Citizens covering consumer finance and fintech, and it's a pleasure to have CFO Sanjay Datta here again with us from Upstart Holdings. And, you know, really one of the sort of the pioneers in digital lending. And this is the part where I just turn it over to you so you can have your own commercial. Take a minute or so to introduce yourself and the company, and then we'll kind of jump into the Q&A.
Sure. Well, to start, thank you for having me and us. My IR person gave me this statement I have to read, but you had a better version. What do I say? For a safe harbor statement, refer to our IR materials, our latest earnings presentation. Is that good enough, Sonia? Okay, that'll do. So great to be here. So yeah, Upstart, I guess maybe just a quick background for those of you who are less familiar. We are essentially a company that believes at the forefront of applying machine learning models to the lending industry, essentially. We've been doing that since 2014, long before the term AI became in vogue. We apply these models to credit underwriting, to fraud and verification, to servicing operations, to borrower acquisition.
By virtue of that, the borrowers that work with us, on the one hand, have lower APRs because they're subsidizing less default, but they also have a lot more automation and a lot less process friction in a process that's typically stressful and unpleasant. And we do this on behalf of the lenders and the institutional investors who provide the money on the platform, to whom we typically charge a fee. And so the combination of those things creates a business model that, on the one hand, is relatively low capital intensity, on the other hand, allows us to both grow quickly and be healthily profitable, which is a trifecta that I think has been a bit elusive in the world of lending. So I think it's unique. And yeah, we're coming into the year with a lot of momentum.
We've guided a strong 2025 after a couple of years, which clearly were very challenging in the lending space between inflation and high interest rates. And so we're optimistic about the strength and the direction of the core business and just taking this playbook and applying it to newer segments of lending, all of which we think will ultimately be able to benefit from these types of model advances.
Great, great. So listen, it's a tech conference. So before we get into all the fun about funding and credit and whatnot, let's kick off with AI and a little bit of a leading question, or maybe it's a devil's advocate question. So I've covered, you know, there have been some companies.
Right off the bat.
There's some companies like LendingClub. They've been around 12, 15 plus years. And the terms machine learning and alternative data beyond just a FICO score have been discussed within the industry for a number of years. Can you talk a little bit more about what AI really means, the context when you talk about you being an AI-driven lender as well as fraud and whatever? What exactly does that mean for an investor?
Well, yeah, I mean, I guess AI has become such a broadly used term now that it's sort of, well, it hasn't lost. I mean, I think it's broad by its nature. It's a little bit like using the word intelligence as it applies to humans. It can mean a lot of different things. And the question, I guess, now isn't anymore a binary question of whether you ask a human if they have intelligence or not. It's like, what's the nature of your intelligence? How intelligent are you? What flavors of intelligence do you have? So I think with respect to models and with respect to lending in particular, as an investor, I think you'd want to sort of ask that on a couple of different dimensions. Like, first of all, what models are you using? And there's a wide range of things.
I mean, I think technically, semantically, a linear regression is a form of artificial intelligence. It's creating a prediction that would otherwise require human intelligence. But it's not that powerful because you have to make a lot of limiting assumptions. And there's a variety of different types of models now. There are learning models that use technologies like neural nets and gradient boosting. There's language processors. There's computer vision. There's all these different types of models and different sophistications of models. And so what types of models are you using? The vast majority of our company's history has been spent building non-proprietary models or taking off-the-shelf technology and adapting it to the time series problem, which is credit. So credit's not just a classification problem. It's got a heavy time series element that's a bit unique.
And so a lot of the R&D we've done in our lifetime has been creating non-linear, very powerful models. And so what models are you using is an important question. Second, what data are you feeding it, which is a little bit like asking a human what books they're reading, right? What are you feeding into the model and how predictive is that? And when I say that, when you talk about data and machine learning, it's not just how long a history of lending do you have, but through that period of time, what did you know about your borrowers at inception such that you can now use that as pricing risk factors? So do you know about their education and where they went to work and what industry and role and function they play? These are all important things in assessing risk.
And so collecting that data and feeding it into the models is the second thing. Third, where are you using this in your business? A lot of people have the types of people who can build machine models and AI, but use it as almost a research function. There's some lab coats in the basement, and they go, and their job is to find insights and bring that to the credit committee. And if the credit committee likes it, they sort of get implemented as rules in a model. And we believe the better way for this to be unleashed is as a production model. So our production model is a machine model. It doesn't have versions per se. It's something that is updating as it learns, and it is the thing making the decisions.
If a new model form or an updated algorithm is better, it gets implemented, and it's the new decision maker in the company, and then the last thing I would say is you'd want to, as an investor, just understand the outcomes. After you do all of that, is anything different? Do you have a bunch of collateral that performs as you could predict through their credit score, or is it something that is a bit confounding and hard to understand by the rating agencies who use more traditional methods to try to proxy things, so those are all questions I would ask, and I think they probably have different flavors and different answers depending on the company you're talking about.
So I want to follow up on that. You and I have known each other for a while. So when we first met, just prior to your IPO, which, boy, was it like four years ago, five years ago?
Late 2020, yeah.
Yeah. AI was not a term that was thrown around a lot in financial services, certainly not consumer lending. And I guarantee if you walk into every room on this floor, there's probably a slide right now, no matter what business they're in that has AI on it.
Yes
But just given the advances in generative AI and other things, do you think differently about your business and your credit analytics now versus just four or five years ago, both in terms of, A, the accuracy or how proprietary you may think it is, and B, how should investors think about the competitive landscape just versus four or five years ago? Are advances in generative AI allowing somebody like me to go out and hire some programmers and raise some capital and make up some ground that I couldn't do four or five years ago?
Yeah, it's a great question. I mean, when we were out on the road five years ago, we were using these terms a lot, and I think they were almost just kind of ignored. And now it's captured the imagination of the public. The short answer is Gen AI has not really changed the way we think about the business. Generative AI uses a lot of the same underpinnings that we use, but it's a very different application. Generative AI is generating content. It's ingesting the internet, and it's making sense of it, or it's trying to sort of distill it. We're using similar type underlying machine models to make predictions. And all the innovation that's happened around Gen AI has not really certainly impacted what we're doing internally. We're doing a very specific thing. We're ingesting variables about borrowers.
We're using that to predict who and when they will either pay us back or default or prepay us, and so it's not really changed the internal equation for us, and I guess as an investor, the important thing is what's unique about what we're doing to the extent we're sort of differentiated is two things. One is you're taking models that are built for general purpose things like classification, and you're adapting them to the problem of credit, and it is a bit of a unique, wonky problem in that a lot of your training data and training data in lending is the loans you've given out. I mean, if you were to classify them, you'd want to know which ones are defaulted and which ones have repaid you, but of course, at any given point in time, they're all in flight.
You don't know necessarily which ones are going to end up paying you back or not. And so you have to do a lot of math to adapt for the fact that your training data is in flight in a time series problem. And you don't just care about whether a photo is a duck or a horse. You care about when they may have some event like a default or a prepayment. And if that happens at month 36 versus month 3, it's like a massive difference in the economics. So you have to have models that are adapted for this problem. And I don't really know of any or many folks in the American market who are doing those types of adapting right now. ChatGPT is not set up to handle this problem algorithmically.
And the other thing is you need the data to feed the models in order for them to make their prediction. And that data is not widely available. The way you get it is you collect a lot of data about a borrower, and then you give them money. And then you have to wait to see who paid you back and who didn't and when. And then you can say, okay, well, just by way of example, someone studied at Georgetown, and maybe they studied art history, and now they're working as a hedge fund analyst in New York. And this person went to Chico State but did computer science, and now is working at a startup in Silicon Valley and all the things in between. You have to know what every single one of those things means for the probability of default or repayment.
You can only learn that's not on the internet, right? You just have to lend and observe. We've been doing that for 11 years now. So someone to come down this path would have to start doing those things today, start collecting that data, putting capital at risk with a V0 model that probably won't be very intelligent. They'll have to put a lot of maybe 10 years of work. Maybe they can do it faster than us. Maybe they do it in five years. They'll have to do a lot of adapting of existing off-the-shelf models to be able to sort of handle the time series problem of credit, et cetera. All that to say, I think it's definitely generated a lot of buzz, and we've gotten a lot of attention for it because we've been talking about this stuff for five years.
And even when we were doing private fundraisers, we used to talk about it. But it hasn't really changed the equation internally, nor do I think it really changes the sort of opportunities and threats in the market.
Got it. So last question on, call it your time series. You need data. You need performance to learn. So I think you said 11 years. For the first eight of your 11 years, the company and so many other startups in financial services had never experienced Fed funds rate over zero. They'd never experienced inflation like 1970s level. And I mean, those two things alone, just disruptions in borrowing costs, NIM, funding backdrop, as well as at least one consumer tailwind that hadn't existed in a meaningful way for probably 20-25 years. How do you gauge how well your models have adapted and learned from the effect of those inputs?
I'm not sure if you can speak to it, but do you get a sense that companies who you perceive to be direct competitors, whether based on how they're pricing loans, whether they're learning as well from those developments?
The macro in general. Yeah, I guess the macro, obviously a very current topic, has been for a while. But I guess how is this? The core thing we care about, the main thing we care about is if you were to give us 100 borrowers or applicants that look identical from a traditional standard, they have the same credit score, the exact same credit score. But we know there's a big spread in risk across those 100 borrowers. What we care about is our ability to separate the risk borrower by borrower. What can I learn about those 100 borrowers so that I could rank them one to 100 in terms of their propensity to repay a loan and the timing with which they'll do it? And that's sort of very much a borrower-level exercise. All 100 of those borrowers will get impacted by the next recession somehow.
In absolute terms, that entire bucket, their risk may go up the next time we have a recession. And they may get impacted with different timing, right? A recession may roll its way through different industries at different times, and there will be phasing differences in terms of their stress. Now, it turns out after that macro event rolls through, it doesn't necessarily change the relative rankings of those borrowers very much. And it is also unfortunately true that whatever we may have learned about this past macro hiccup, and just to be very clear about what it is, because it's been a bit of a weird one, right?
Normally in lending, you care about the stress of unemployment. That has not existed. But what has happened, to be very clear, is a pandemic happened, and that caused the government to increase the money supply by 40%.
It caused consumers to spend a year very flush on their balance sheets, and they changed their consumption patterns relative to their savings, and then when the stimulus went away, that perpetuated, and everyone got upside down in a way that caused stress in lending that was almost on par with the relative stress of the GFC, and yet there is no unemployment, so that's sort of the very short description of what's happened from a lender's perspective over the last couple of years, and we've learned a lot in going through it, but none of it will help us with the next thing because we don't know what that's going to be, and recessions tend to be an N of one, like what did we learn from the GFC, well we learned subprime mortgage risk exposure was very risky, well that didn't help us in this one.
And in 2000, we may have learned that exposure to the tech industry and having tech assets or sort of stock market assets was risky. And before that, it was probably an oil crisis or something. So recessions, I think the best you can do, our view is there's a lot of value in being able to rank those borrowers from one to 100. That's where the biggest spread of risk is. And the best you can do for whatever macro thing happens next is detect it faster than anyone and react to it more precisely than anyone. And I think we've created the tools to be able to do that. We did not have them, candidly, in 2022 when all of this started to happen. And I wish we did. I think we would have navigated it far more artfully.
But all that to say, I think that was our big learning. It's like we need to be better at having our finger on the pulse of the macro. And it tends to take the lending industry weeks to months to actually realize what's happening because things have to get to a certain level of delinquency, and then you have to analyze them, and then you have to worry about controlling for mixed changes and other things, and then you have to discuss them in the committee, and then you worry and you take action. And we want to take action in weeks versus months. And I guess that's my answer to your question. I don't think we're going to be able to predict the next thing in advance, unfortunately. I don't think any of our competitors will be able to.
Maybe they will act more conservatively today in advance of it, but you could argue that then we will all be acting overly conservatively, which is, I think, the default sort of.
There's an argument that's happening now.
There's an argument that's happening right now.
Let's stay on topic then. I mean, in terms of the macro, what exactly is Upstart's view of, well, actually, number one, if you could remind people who you view your target borrower as. Now, even though the world still thinks in terms of FICOs, we speak in FICOs, right?
Yes, yes.
Near prime's a nebulous term, but maybe you can kind of reintroduce us to who you're actually targeting. And then there were a wide swath of kind of opinions from managements we encountered during this earnings season on how confident they were in terms of opening up their credit box. So maybe you can kind of talk about what you're seeing.
The first question is, what's our target segment? I guess target's the wrong word. We are determined to serve as broad an applicant base as possible. We want to be ultimately the lowest price for everyone in America who needs credit. Now, as you get to the riskier borrower segments, or in FICO parlance, maybe like, I don't know, below 680, 680 and south, what is more and more important is how good you are at navigating risk.
And as you get to the prime segments, what's more important is your cost of capital. And of course, our historical edge, if you will, has been in having advanced modeling to navigate risk. I think historically we've disproportionately competed well in the torso of the borrowing base. That's maybe on the riskier side, not because we targeted it, just because that's where we were very good.
I think our ability to generate very essentially depository capital, because that's what you need in order to compete for the very prime borrowers, because it's the cheapest capital, our path to accumulating that, because again, we don't ever want to be the balance sheet in our business. We want to predominantly decision third-party capital. And getting third-party capital from banks and credit unions that have that depository capital, it's been a bit of a different journey. But that's, I think, gathering steam now as well. So the answer to your first question is we want to be broad across the spectrum, but we have different strengths and weaknesses as you sort of go from less prime to more prime.
Which, by the way, sorry to interrupt, it's probably a broader, well, I'll use the word target than maybe at the time you went public. I mean, I think both on your product suite and underwriting, it seems like you arguably are going after a larger TAM, maybe, of lending volume than you were five years ago.
I think that's true. I think at that time we were sort of very focused on our core strengths, and I think we're so good in that market now that there's not a lot of alternative, and I think we've also been pleased at our ability to partner with these credit unions and banks and get them to sort of target very prime borrowers, where let's be honest, there's narrower margins and less NIM in those segments, but I think we've got partners now that are interested in those borrowers, so yeah, it's been an evolution, but.
And their health. So how are they doing?
Macro. I mean, I think it's a fool's game to try to predict macro these days, so we honestly don't. I think I'll say this. I think we're very good at measuring where the macro is right now, and I think that may sound to a tech audience a bit basic, but it's actually not. At any given time, you don't know exactly what's going on in the macro, and I think at least as how it impacts the propensity to default in the environment, all else equal, we have a very good measure of it right now. It's precise as anything. It's probably fresh as of three, four weeks, whereas normally you've got to wait two months, three months, so I think we've got a very good feel for where the macro is at any given time.
And I have confidence that we are underwriting loans with some buffer to that. So we've got some amount of safety mechanism. But that's true. Normally when you say that means, well, we're just going to do this with prime borrowers. I don't view prime borrowers to be any less risky than less prime borrowers right now on a relative basis. So we're trying to underwrite all of them, but with some amount of buffer.
We'll call it a tariff buffer.
Okay.
I like that term. You know what? I had a bunch of other questions we didn't have time to get to, obviously, on the product and funding side. But if there are any questions that anybody has for Sanjay, we've got time for probably one. Okay. Well, I'm going to then close it out with a quick follow-up on the product side, because as you said, you tend to focus on kind of your core personal loan product in the past. I mean, you've moved into what, auto?
Yeah.
Indirect auto, HELOC, which is interesting, secured and super prime in some cases, small business, small dollar. Could you rank those in terms of if we did have a crystal ball three, five years from now? I mean, what's really going to be the big number two asset class for the company besides personal loans?
Yeah. I mean, we basically have three new emerging bets right now. One is auto, one is HELOC, and one is small dollar, and we were just saying, actually, there's a bet going on on the leadership team, like what horse do you have in this race?
They're all sort of at roughly equal stages of nascence. There's no consensus on ours. I think there's different things to like about all those. I think we're excited about all of them. I think the near-term bet, this is my view, not Upstart's view, because my boss may disagree with me, but in the very near term, I think HELOC is perfect for this market. We're very bullish on it. In the medium.