Upstart Holdings, Inc. (UPST) Earnings Call Transcript & Summary

Up next, we have Dave Girouard, CEO and Co-Founder of Upstart. Dave, thank you for being here. I want to start. I'll do a little housekeeping first. Today's discussion may contain forward-looking statements that relate to future results and events, which are based on Upstart's information available today and are subject to risks and uncertainties. Actual results may differ materially from these forward-looking statements. The discussion may also include non-GAAP financial metrics, which are not a substitute for GAAP metrics. Please refer to the company's filings with the SEC and its IR website for additional information, including GAAP and non-GAAP reconciliations, along with other disclosures.

Thanks, Will.

Thanks for being here. Okay. So maybe we could start with big picture. Upstart is a lender that uses advanced machine learning techniques to make credit decisions. So can you talk about the importance of AI in your models and how you think about the extent to which you can improve upon the credit outcomes of other lenders using more traditional methods?

Sure. First of all, I wouldn't describe us as a lender. I would say we're a lending marketplace. There are hundreds of lenders and private credit institutions, et cetera. We're more a market maker in the middle. But the whole premise of our business is that lending and credit can be completely modernized in ways that take advantage of technology like machine learning and AI to improve access to credit from the borrower perspective and improve the product from a lender perspective. And we started the company 13-plus years ago, really with the idea that these technologies were coming to a place of maturity. Credit and lending in the consumer segment really hadn't changed much since the FICO score came about in the late '80s. And I was at Google, and the idea generally was the types of technologies we were building at Google for a very different reason could be applied to this very old and important industry. And essentially, what that means is we've built technology and models that are much better at assessing risk and understanding who is likely to default and who's not likely of default, et cetera. And also just kind of automating the process so that the experience of the borrower is much, much better. Today, in excess of 90% of the loans on our platform are approved in a moment. There's no human involved, completely automated by machine learning techniques. And also what amounts to the risk separation and being able to price the right person correctly. And so this is really something that just accrues value and gets more accurate as we have more data, more sophisticated AI models, more training, et cetera. And in our view, at least, if you looked ahead a decade to us, it's inconceivable that lending and credit around the world wouldn't be AI-powered in almost every location you see it because the economics are so much better, the opportunity to create a better product is so obvious. And I think the world is now, in the last few years, really begun to open up to the possibilities of AI.

Yes. So I want to talk about maybe the trajectory of the business longer term. I think you said at AI that you want to have the best offer for all Americans, which is a bit of an extension from the TAM where you currently operate. How do you think about competing in some of the more competitive areas like prime lending, for instance, in the U.S.? And then how close are you to having the most competitive rate outside of some of the near prime personal loan categories?

Yes. So I think it's to say the way lending has worked for many decades is every bank or credit union or other lender would kind of spin their own model, they would get a score like a FICO score or something else. And it was -- some were a little better or a little worse than others, but it was a place where most of the technology was just homegrown. And with the advent of machine learning and AI models, that sort of changes things pretty dramatically. It's not realistic for a lender to just spin up a model and hope that it's competitive. So what we basically came to the conclusion a few years ago is lending won't just be this little thing that appears at your bank or at your local or in a neobank, et cetera. It's a very, very sophisticated technology opportunity that requires intense focus and investment. And so we are building toward a destination for credit that our hope and expectation is 100% of Americans will be persistently underwritten, meaning once we've seen you, once we know you, you are permanently underwritten. There's no process after that. And we can provide access to the very best forms of credit that you might need, whether that be a home loan, a personal loan, a car loan, a revolving credit, et cetera, all different flavors of credit. And at some point, we would like to be able to guarantee the best rate as well. And this is kind of what AI does. It separates and allows you to get to the vast majority of Americans that are extremely creditworthy and make a better product for them. So we've just started a year ago, as you referenced. Internally, we say we want the best rate and the best process for everybody because that's how to build a lasting brand in a huge enterprise if you can have that sort of brand value. And we're working toward that. I would say, in our personal loan, we feel very good. We're on the cusp of really having the best rates for everybody. And there's a few subtleties to having all the right capital behind that, but we're making great progress. And in our home product and in our auto products, I think we're also on a very good trajectory. And that would be the goal is to have a trusted brand where you know you will get a phenomenal rate. The process will be effectively no process whatsoever. And as we kind of expand on the product set, and this is our chance to be the largest player in AI. I think we are the only company building a foundation model for credit in AI and the proceeds of that will be coming in the next decade.

And I thought AI did a really good job showcasing how Upstart is ahead of the pack in terms of leveraging some true machine learning techniques in real time, constantly innovating on that and deploying it into the market. It's clear that a lot of the banks are not doing this. When you look at some of the more modern competitors in the space, like do you see signs that others are beginning to apply some of these more sophisticated techniques? And how do you think about the risk longer term that competitors' underwriting models begin to converge and produce kind of consistently different underwriting results than the traditional credit models.

For better or worse, I don't think there's like a wide array of competitors trying to do what we're doing. I think there's a few, if any. If you look at people that are in the kind of buy now, pay later space, people like Affirm, Klarna, et cetera, they tend to have a lot of focus on fraud, given what they are really kind of payments first. And the credit they do is very short term. So it's a little different. And so they have, I would say, some genetic comparability to us, but they're not really pursuing the products that we have. So in that sense, there's others a little bit like us. But if you think about longer-term credit, 2 years, 5 years, 10 years, 30 years, the players in that market, in our view, are not AI native, and they're not really building AI native products. They aren't building products with instant approvals. They're still based on someone going into a branch, sitting across from a loan officer many times. You can't realistically get a home equity loan from your local bank in a matter of a day or 2, and you can't do it from your home. You're going into an office. So I just think that the industry at large is not building AI-centric models. There are a few -- I think there are a few maybe copycats, if you will, but there's not a lot of. So I don't see a convergence. I mean, certainly, some others will step up. I think the opportunity is so vast that it will become more apparent to people. But you can't spin this up in a matter of weeks or even years. I mean we've been building models that the data set that it's trained on has been growing for more than a decade. Every loan, every repayment, every late payment on our platform becomes part of the training set. And you can't replicate that overnight.

Yes. Makes sense. Maybe just switching gears to the customer value prop. It clearly beneficial for a customer that can't get access from a traditional provider to gain access to credit because of the underwriting. How do you think about the ongoing relationship with that customer after that transaction in that first loan? And then what do you track on repeat usage that maybe shows the outlook of Upstart kind of being more of a customer acquisition vehicle than kind of like an originator?

Well, first of all, I'd say the idea that you come to Upstart because you can't get the product from your sort of core banking provider would be a very old view of the industry. Banks don't hold their customers tight that way. Credit is -- essentially has been shopped for a decade or more now. If you're doing any form of accessing credit, whether it's for a card, home, whatever, you are generally shopping. And so if you deliver the best rate and you have a brand behind that, you will be a first look, not a second look. And I think that's where we're quickly evolving to. So I think -- another transition we're making is from someone that just gets a transaction with us, gets a loan once. Generally, they're paying it back in an automated fashion. They don't have to talk to us again. I think you will see more and more from us. About 1/3 of our loans today are from repeat borrowers and people that had great experience came back for another product. We are very rapidly moving into a place where someone might come for a small unsecured loan. We discover we can offer them a better auto loan than the one they have today, and we are able to cross-sell that or somebody comes looking for a personal loan, again, our core product, and they discover they can get a larger and maybe less expensive home equity loan. So cross-selling products and having that long-term relationship, I think you will see us more and more have products that people interact with more frequently. But today, we've had phenomenal unit economics and lowering acquisition costs through our entire history. If you just sort of went all the way back to the pre-public days, we've been extremely good at improving conversion, having very -- decreasing cost of acquisition. So that's all been on a great trajectory for a long time.

Yes. Makes sense. Maybe we just hit on the macro environment, the kind of obligatory macro question, the upstart macro index. Wondering how you've used some of the more recent revisions to the job data, particularly weaker unemployment environment. How do these data points inform your particular machine learning approach to underwriting standards to the extent that you're feeding that in?

We aren't per se feeding in Fed data. I mean, honestly, most of the data you would get publicly would be so dated, it would not be of interest to us. I think that what we are trained on is just literally the tens of thousands of repayments that come in or could be delinquent any particular day. That is far more real-time accurate view of what is going on in the real consumer economy. But I should say, overall, the consumer has been in a stressed place for a long time now, probably a year or 2 at least as they came out of COVID and the stimulus went away and consumers or families have been like overspending relative to what they earned for some time. If you look at every category of credit today, student loans, cards, auto loans, et cetera, default rates are much higher than the long-run average and much higher than 2019 before COVID happened. So that's not anything to do with us. That's the way the world is right now. That has been priced into our loans for a long time. Of course, interest rates are higher, too. So we have grown really quickly. We had triple-digit growth rates in Q2. But frankly, that is against a headwind, which is both higher interest rates, which is part of the cost of funding as well as a higher risk premium in the market. And at some point, that will unwind. But our view generally is the consumer has already been stretched very fully. And I would generally say we're not macro forecasters, but we are not terribly worried about unemployment. I just generally -- our view is there's a sort of secular lack of workers in this country. And if anything, reduced immigration, et cetera, is only kind of stressing that even further. So I don't -- of the many risks, and we tend to worry about everything, at least somewhat. But I would generally say the risk of sort of super high levels of unemployment feels like pretty remote to us.

Got it. Yes. So I guess, summary then would be you have more frequent data that's already sort of priced into the model. There's not like a management overlay or anything that you would do around underwriting to react to some of the recent data that's been coming out.

No. The model is generally targeted to be somewhat buffered against expectations so that if there was some worsening in the future, we'd be fine. But generally speaking, we're using real-time repayment data as the basis for what we would call calibration.

Got it. Okay. So sticking with sort of a macro theme, the overall funding environment, there's been a lot of capital flowing into the consumer credit space in general, the rise of private credit, alternative asset managers looking to increase their allocations to the space. You and a lot of the consumer lenders out there in the space that I cover, a firm, SoFi, et cetera, have had significant access to, call it, alternative capital supply, off-balance sheet supply. How do you think about the sustainability of this capital through a credit cycle? And what are you kind of watching for signs that some of the fixed income investors out there are getting more nervous about consumer credit allocations?

I view it as, first of all, it's just the early stages. We really created the first partnership with private credit in consumer lending in 2023. So it's only been a couple of years. But it's important to say these partnerships were created and structured purposely to survive through credit cycles. If you looked at the way funding and sort of online lending, et cetera, in the prior era, it was very much dependent on hedge funds. It was at low funding. It was very ABS dependent. If the securitization markets were fluid and pricing well, there was more money than you know what to do with. But of course, that could change, and it did several times change. These are long-term partnership structures that there is alignment, there's co-investment. And they're designed essentially to work through credit cycles. Private credit companies have long-term locked up capital. They're looking for a certain mid or high teens ROE on that capital. And we are very well aligned to do that. So as long as we're co-invested and the structures are set up so that through good times and bad times, we can continue to deliver on that yield, then I think they're great partnerships. They've also begun to get longer. The first partnerships we signed were single year, moved to 18 months. They're moving to 2 years. So I do think there's great alignment. This is not like the subprime mortgage crisis with originate to distribute and no verification of anything. These are highly aligned business models. So I don't necessarily -- it doesn't mean everybody will perform because our skin is in the game. And we think that's the right way it should be. How can anybody commit long-term capital to you if you don't have any skin in the game. But at the same time, we need that long-term capital. If you look at what we filed when we share in our sort of earnings deck quarterly, we've returned something in the range of 9%-ish ROA on average, if you invested since in our platforms, I think, since like 2018. So that ROA very easily translates into the ROE that these firms need. And that's the way to think about it. Over a long haul through whatever cycle, you can deliver that. And it's a little insurance like in the sense that there could be worse quarters and better quarters. And collectively, it's -- there's actually great alignment.

Got it. I wanted to talk maybe around pricing then. I think during the quarter, you mentioned you did some testing on pricing elasticity. And so I was wondering if you could just expand a bit on what that means and how you think about the pricing opportunity in the business. But I'll lump it in with maybe a bigger picture question just on spreads in general. So several years back, you saw a big increase in industry loss rates. Upstart had to put some loans on the balance sheet. Since then, funding markets have come back, obviously, in a big way, there's not an issue anymore. I guess fixed income markets tend to be more backward looking than forward looking sometimes when it relates to kind of lender performance and track records. Do you feel like you're getting credit for the improvement in underwriting models over the last several years? Or do you think lumping it back with the pricing question. Is there more room to go to see improved spreads on some of the loans that you distribute?

I think we're getting credit where we need it. I mean the market is not what it was. I mean, today, the influx of private credit dollars isn't remotely like hedge funds buying, no risk retention by the originators, et cetera. That world does not exist today. Even ABS markets have improved a lot. The bonds have -- the yields have tightened and the spreads have tightened, but there's really not a fluid market for the equity or the residuals. So like I don't view this as like a crazy market of just money flying all over the place. These are very carefully structured co-investment type markets. But I think, first of all, our credit has been performing really well for some time. I think the partners we have would probably gladly take more from us, which is really what led them, I think, to go to other platforms. We'll see how everybody performs on that. But right now, we're very confident in the performance of our models. They calibrate really quickly now. They -- our separation continues to get better. And I think this is just a natural -- this is an AI shaped problem. And we feel very good about that. I think the private credit coming into this market is a realization that depository credit is just one form of low-cost dollars. There's plenty of other dollars that need yield over time. The relationships between private credit firms and insurance companies means you can have a blend of capital to solve these types of problems. So I think it's actually a natural evolution. And I love that our platform has a combination of depository capital funding for the primmest of borrowers and then private credit to have more structured style funding for less prime parts of the market. Our product set is now going from home equity, which has super low loss rates, 1% or less, all the way to higher risk smaller dollar products. But that's as it should be. That's what a true credit market looks like.

Yes. That makes sense. Okay. On one of the near-term questions, specifically on approval rates, they stepped up meaningfully in the most recent quarter. I was wondering if you could talk a little bit about what drove that? And then bigger picture, talk about how you view the normal cadence of model improvements on an ongoing basis.

Sure. So our models were sort of continually doing research and making improvements to our models. They can come in the form of -- or come from new sources of data or which have some ability to improve the models or improved algorithms. So this is just like OpenAI is trying to work on the next version of their ChatGPT or some other product. We are constantly doing research and then periodically, we actually release new models quite regularly, probably in the order of every 2 or 3 weeks. But those can just be some of point releases and then maybe more of a once-a-quarter cadence of kind of a primary release. Those releases essentially drive better separation. Like that's the whole goal is to better separate good risk from bad risk. When you do that, effectively, you, in the end, are getting some likely credit losses out of the pool, and that allows you to lower the rates to everybody else, which improves conversion. So calibration neutral, meaning if the economy is not changing a bit, then each new model is generally going to improve conversion, which is how we grow. That is the sweet spot of how we've grown over years is just improving conversion rates. I would just say that is in contrast to the entire world of credit and lending, which generally has very static models. They might upgrade them every year, every 1.5 years. They aren't generally looking for a model win to improve conversion. They're more in an exercise of loosening or tightening based on perceived performance of their credit. This is more structural, a structural change, which says a newer model is trained by more data, has more sophisticated algorithms behind it, can separate good risk from bad risk better and results in better conversion. We're at the early stages of this, but I think it's a new world that is completely different from the history of credit and lending.

So I want to pivot a bit and talk about some of the newer verticals. You've been testing this for some time now. This quarter, your commentary seemed to signal kind of moving to later stages of operationalizing them. I guess where are you in terms of testing and tweaking the underwriting models? Is there anything left to do before pursuing some of the funding arrangements that you brought up on the call were kind of slated to go live later this year.

Yes. So we are in the process of becoming a 4 or 5 product company and expect we will be within a few months. And I'd say that we've had products that we offer to consumers for some time, as Will mentioned. But really to get them ready for third-party funding, you have to have, first of all, a significant enough volume that credit is performing, and you can sort of prove that on a statistical basis. Unit economics are good. You're originating at a return level that is acceptable to the market, which we've been doing for some time. So all those sort of check boxes are done. We're not needing to like tweak those. These products are ready for market. We're essentially in the final stage, which is you have to have the agreements in place, you have to have the operational abilities in place to transfer loans, et cetera. So I would just say our goals and our expectation is the majority of these products will be third-party funded by the end of the year. So we're in this last phase. It was, I would say, unplanned and a little unusual that we have 3 products, all of which are going through this final step at the same time. As I've said, it's like Monday is the first day of school and the triplets are going to all start on Monday. We wouldn't have exactly planned it this way, but it turned out this way. But I think the opportunity really is that our expectation is by next year, these products will be vast majority third-party funded. And our balance sheet, I expect, will begin to turn into more cash than anything else. And so this is just that final step. There's nothing holding us back, and I think we're on target to have this done by end of the year.

Got it. Makes sense. Let's -- maybe let's dive into a little bit on each of these. First, on T-Prime, could you talk about how you view the opportunity in T-Prime loans? Prime loans, obviously, a much larger part of the market, more competitive, more bank involvement. So do you see the same inefficiencies and opportunities in the prime segment as you see in kind of the traditional near prime personal loan space?

Yes. I guess the context -- T-Prime is a term we use to market to banks and credit unions. The opportunity to come and get very, very prime borrowers who are, of course, attractive to that type of institution to cross-sell into other products. So a bit over a year ago, we basically came to the conclusion we were known for and very experienced at serving non-less prime, near-prime borrowers and really did not have good offers for someone who would be an obvious prime person. Think of somebody with a 750-plus credit score and solid income, et cetera, a traditionally prime person. And we kind of felt this was not long term in the best interest of the company. It's hard to have a brand just to say we might have an amazing rate for you, but we're not sure compared to what you could get elsewhere. So we basically said we want to get to a place where we can have the best rate for everybody. And that really meant bringing more depository capital in that can serve the lowest rates to the best borrowers. And that effort has really gone incredibly well. Our market share in the super prime end of the market, which has moved up very rapidly. You can see in our earnings deck, it is in our core product, maybe 1/3-ish of our borrowers already in just a year. And again, it's that best rates, best process for all, we really needed to go after the super prime part of the market. I would also say that it's interesting that we don't think that has to be served entirely by depository capital. Private credit, along with their kind of insurance partners really can create very, very compelling sources of funding for the whole market. So that's who we're becoming very quickly. And of course, home equity is itself a very prime product. So you will see from Upstart over time, our goal to be best rate for everybody and have the best experience for everybody. And T-Prime was really the kickoff of that effort way back in mid-2024.

And it seems like a good time to be getting into home equity loans. The industry has been in long-term decline into the last couple of years, and I'm sure some of the elements on interest rates and home prices have something to do with that. So what do you kind of see in the market opportunity right now, consumer interest for home equity loans? And then just how are you approaching that opportunity?

Well, yes, I mean, home equity loans, in some sense they're popular when you want to tap the equity in your home, but often in a high interest rate environment, refinancing your mortgage doesn't make a lot of sense. There's no opportunity there. So they're just a way to tap into home equity for whatever expense you might have. That could be an improvement to your home, but it could be a wedding, it could be paying off some other debt. So there's a lot of reasons why home equity loans are popular. Traditionally, they've been the type of loan that you go to your bank or you go to your credit union, you will take 5 to 6 weeks on average to close a loan. So to us, that's like as that exists, not a very interesting category. What we're building toward is a product that is really very close to our personal loan, if not instantly approved within a very short period of time. The hard part, of course, is you have liens and titles to deal with and state regulations and such. But we're moving very quickly toward a place where a consumer can come in. If they want an unsecured loan, they can have it in a moment. If they want a little bit more effort, they can have a home equity loan. And we're not talking about 5 or 6 weeks, we're talking about a few days. And hopefully, at some point, less than a day, we are getting into the first automatically improved home equity loans. So it's a category that there's a couple of online digital types and then there's -- it's a very fragmented market. There's a lot of banks and credit unions who have cheap capital, but they have a process that's archaic and slow. So in our view, this is a category where we will be the most AI-native participant. We can grow market share and hope to grow market share for a very long period of time. And we hope -- we would expect over time to have a whole suite of home lending-related products.

That makes sense. Can you talk a little bit about just some of the operational differences in that product relative to the unsecured space or in auto, things like collections, foreclosures, just mortgage servicing in general, tends to be a little bit more labor-intensive. Like how are you approaching that? How are you kind of dealing with the cost associated with it?

Well, both home and auto, yes, you have an asset out there that at some point, you can repossess or not repossess. So first of all, it's important to say one of the great things about home in particular, is the default rates are much, much lower. I mean, less than -- typically less than 1%, right? Nobody wants to lose their home. That's why they're right at the top of the repayment list. I think we're just growing into and getting better at servicing in the auto segment. We're not servicing yet in the home segment at all. But generally speaking, our servicing and collections capabilities have gone from, I think, pretty mundane a couple of years ago to quite state-of-the-art as we're beginning to introduce AI and machine learning models into how we do loan servicing. And the opportunity there is much, much larger, I think, than anybody appreciates because the subtleties of when to intervene, how to intervene, what message to deliver, when you might settle a loan versus let it go further or sell it off to a debt buyer. I mean there's so many avenues and paths in terms of how you deal with a delinquent loan and servicing and collections that AI is a very natural fit there. And I would just say we didn't focus on that for a long time. Only in the last couple of years, we're beginning to gear up. And the results have been pretty extraordinary. We hope and expect over time, we will have an independent credit servicing business that we would begin not just to service credit originated on our platform, but everywhere. And I think that's another very large market opportunity for us.

Yes. I mean, since you brought up, I think you mentioned on the call, I think there was some servicing related improvements to the model this past quarter, isn't that right?

Yes. We have very ambitious goals in the range of like 20% reduction in defaults over the course of a year, not by anything to do with the credit decisioning at the front or by the macro economy, but just better servicing and collections of loans. So we've been able to achieve those kinds of improvements. And I think, honestly, we're still at the like earlier stage because we're barely getting the first machine learning models into production. And the opportunity there is very large. I mean it's just hard to imagine -- if you just think about it, like, for example, it's very common that once someone has charged off, meaning you've charged off a loan, then you might pursue a settlement on that loan. But there's no real reason. Maybe you should pursue a settlement earlier if your model tells you that's how you're going to get the maximum dollars back. So you can just see how over time, it's very obvious, there's a lot of different choices you can make, interventions you can make that the dollars back can get better. And when that happens, by the way, that goes directly back to a better price to the next borrower, right? And that's kind of the essence of the model. The more you can collect better, the more -- reduce the cost of origination. All these things accrue to a better product to the consumer. So there's a great flywheel there.

Got it. Okay. In the last minute or so here, I was hoping we could talk through just economics on some of the new products, how you view sort of the stack ranking of sort of like unit economics on dollars of GMV from kind of a top line and a contribution margin perspective.

Well, the products are very different. I mean we literally have a small dollar product that could be a loan for a few hundred dollars all the way to today, a home equity product that could be a $70,000 home equity loan, and we'll have larger loans and we'll have smaller loans. So it's just going to really cover the vast. I think generally, what we are optimizing for as a business is contribution dollars, right? We are not in the business of making loans where there's nothing in it for us or building for some magic future. But at the same time, I mean, the percent can vary a lot by the nature of the product. But when you think about how we originate, what we originate, how we spend on acquisition, we're looking for net additional contribution dollars. And I think that's for us always the thing, along with keeping our fixed expenses extremely low as a business. And you can already see, if you've just looked at our economics in recent quarters, growth goes immediately to the bottom line. We've moved back into GAAP net income profitable. It's our expectation and our plan to stay that way and grow our profits. But that's the nature of our business. It's low headcount. It's super efficient, super automated credit origination. In the last quarter, 92% of the loans on our platform had no human involved in them whatsoever. They are approved in a moment. We're getting to some what we call loans where there might be some other process going on after the fact, but there's not a human involved in it. So as you start to think about all the possibilities of AI, I don't think there's anybody in our industry who's as AI forward as we are, and I think the advantages will continue to accrue.

Great. Well, I think with that, we're out of time, but thanks so much for the conversation. Thanks for joining us.

Thank you, Will.