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

Good morning, everyone. Appreciate it. So I'm Andrew Boone. I cover [ Internet ] here at JMP. Very pleased to welcome Sanjay Datta, CFO, since 2016. I'd to look this up through the IPO process, and now we're here today.

Happy to be here.

So let's kick it off. I think macro has been this overarching story across the last year across more companies than just Upstart, but for you guys in particular. Can you talk about UMI, talk about the macro index, you guys on the last call said it was stabilizing? More or less a month later, how does the overall macro environment looked? And then let's bring this back to what everyone is talking about. How has that led to any changes in lending partners that are now coming back to the table?

Sure. Well, maybe just a quick blurb about what UMI is, it stands for the Upstart Macro Index. And really what that is, is in our repayment data trying to isolate what are the moves or the changes in repayment patterns that are due to exogenous factors. So you can imagine if you have a few hundred thousand loans all of different vintages at different places on the timing curve. Each loan has its own story. There's a very broad aperture of borrowers, profiles. They're going through their own life, each have their own sort of path. But when all of them at all different points and the timing curve sort of move in a similar way, at the same time, you can imagine that something exogenous is happening. And so we do a lot of math to try and pinpoint what is sort of borrower level patterns and noise and what things are sort of viewed as common. And that sort of isolation of exogenous impact we call the Upstart Macro Index. And it's really just meant to be our best representation of what's changing repayment and default patterns in consumer credit due to things that are not to do with our internal changing underwriting models changing borrower mix, et cetera, over time. And just to sort of briefly chart that index, it's sort of historically, you could imagine a sort of a 1.0 being a long-term normal. It went down to about almost 0.5, 0.6 in the wake of COVID when stimulus was plentiful. Going down is good. It means there's a lot of defaults in general happening. So a 0.5 would mean you're seeing half the defaults you normally expect in a long-term normal environment. And then as the stimulus waned that sort of went up into the right over the course of the past 12 to 18 months, and sort of peaked at a level close to 2. So more recently, you're seeing defaults that are twice as high as you might expect and what we would think of as a long-term macro normal index. So that's sort of the shape of things. We did signal on our call in February that, that sort of march up into the right, that's been happening pretty systematically for, as I said, 12 to 18 months now, has stabilized, and there's even some signs that it's coming back down. What it's related to, in our view, in the broader macro, we care a lot about the balance between income and consumption. We believe that, that's really underpins a lot of the increase in this index and in defaults in general. Consumption has kind of gotten out of whack with income and stimulus is what allowed that to happen initially. When the stimulus went away, they didn't really come back into line, and that's slowly now happening. You're looking at the residual of income and consumption is obviously savings and the savings rates in the economy. Personal savings rates went down to a level that we've not seen in decades, sort of historically at an 8% to 9% level, they fell to a 2%. I think the latest print was now like 4.7% or something close to 5%. So savings rates are recovering. Consumption is finally moderating. People are trickling back into the workforce participation numbers. And those are being reflected in what we've seen in the default patterns.

Is that playing out in terms of your lender partners? Are they starting to turn stuff back on, or how does it reflect?

So the lending partners, which to us, basically, are banks and credit unions. So we work with 2 funding sources predominantly in the credit tax decision on our platform. You've got the institutional capital markets and then you've got the sort of world of banks and credit unions. Banks and credit unions tend to be a lagging indicator, they tend to move most conservatively. And so I don't -- I would not say that they are as of right now, rushing back to fund, if anything, the institutional markets are the ones that are much more precisely trying to calculate timing. I think banks will sit back and want to ensure that the world is, in fact, going in the direction they wanted to before they start opening up the throttle.

Makes sense. I want to shift to permanent capital, right? So big picture, right? So understood that we're at a conference. Let's go big picture though. What should we expect in terms of permanent capital becoming a bigger part of the funding mix? How should we think about permanent capital in terms of the funding mix in a more normalized environment over time? How do you guys envision that as you guys are having conversations and talking to long-term capital partners?

Well, it's a bit hard to know what to expect. I could maybe outline our ambitions. Obviously, in this market, doing long-term committed deals is a challenge, but historically, as I've said, we've sort of worked with, on the one hand, banks and credit unions; on the other hand, in the institutional world. We sort of worked on an at-will basis, and that's worked fine up until a year ago. And by at will, I mean every month, they could come and tell us how much credit they want to fund, and we would deliver it. And historically, there's been a shortage of [indiscernible] where it's not capital. So it's worked very well. Of course, in the last year, that equation has kind of been flipped on its head and probably the challenge we've had in the last year is that there's not been a resiliency in the capital base. Our business has contracted more than you might expect if you were just looking at the default patterns in the consumer base. Obviously, the institutions tend to work as a herd in the credit world. And so coming out of this and in preparing for maybe what that next cycle will be, we have an ambition to have a greater percentage of our capital base before we committed. And if I could maybe roughly, if I could [indiscernible], I think I could imagine like 1/3 of our capital base being banks and credit unions, 1/3 of it being long-term committed institutional capital and 1/3 being what you might think of as the spot market or something like that. Maybe we would even like to lean a little bit more heavily into committed capital. But sort of roughly those proportions, I think, would form a nice model.

Is there anything out there that's analogous to this today that would give us an idea of what would happen in terms of unit economics as you guys are at long-term capital?

I don't think so. I think the simplest way is -- we're sort of determining it as we go. But I think the simplest way to think about it is that in exchange for commitment, they'll commence some premium. And we think that premium, I don't know, you might want to model it somewhere on the order of 50 to 100 basis points in exchange for durable forward commitment.

Okay. Shifting a little bit to pricing, right, so the transaction fee take rate on our math stepped up about 1 point in 4Q. Can you talk about taking economics in terms of what you guys are doing there? And then is this just some sort of component of the current kind of credit environment? Or is this something that is more long term in nature as we think about that step-up?

Sure. Yes. Just to maybe describe our take economics. So we charge banks who originate the credit. We charge them to use our technology and, in some instances, for referring the borrower as well. For many of those banks that are balance sheeting or using their own balance sheet, they will typically absorb the fee, and they will earn it back in interest over time. For all the loans that flow into the institutional world, it's typically passed on to the borrower. So in that instance, we have a lot of ability to change the take rate. And really what governs it is the underlying, I think, it was the elasticity or inelasticity of the borrower demand. And so what's happened in the last year is fundamental demand for credit right now is very high. And so it's very inelastic. And so you can charge higher fees and not suffer any drop in volume compared to 2 years ago when credit was more plentiful and everyone it's stimulus. So that has allowed us to increase our take rates. I think it's probably natural to assume that as the world normalizes, as default rates come down, as the sort of macro stress abates a little bit that inelasticity would reduce and our fees would commensurately -- I don't think they'll go all the way back down to where they were pre-COVID for the simple fact that in going through, we've been during the last year, we're much better at the [ massive ] fee optimization. And I think we were sort of happy to be under-optimized pre-COVID because we're growing fast, and there was a benign environment out there, but I think we have a much better idea of how to measure elasticity now and how we might optimize. That said, I don't think they'll remain at the rates that they are currently.

Is there a structural change? I can just imagine that repeat borrowers would be more captive to your program, they may not shop around as much. Is there anything else structural that may be taking place?

I don't know that I've ever believed that. I know that, that's sort of the mean that once you have the consumer that they're sort of yours. And that's definitely like -- it's definitely a very typical way of thinking in general consumer internet. In financial services, it's a very [ shocked ] environment and price is a very big deal, right, compared to think of any other sort of consumer products where brand and service and all these other things. Financial service at the end of the day, like alone, you care mostly about the price. So I think our rough philosophy is you're going to have to give them the best product each time. I don't think you can like capture the consumer and then overcharge them in a way that's durable. So I don't think that's really the thing we would factor in.

Okay. All right, balance sheet, right? So loans on the balance sheet, kind of close to $1 billion kind of this last quarter. Is there a time line in terms of thinking about how to reduce those? Is that just market dependent? How are you guys managing that portfolio?

Yes. So historically, we've not run as much of a balance sheet at all. We've typically been third-party funded. In the turbulence of the past year, our balance sheet has ramped a little bit. It's sort of what I would consider to be at the sort of the maximum level we would be comfortable with. We're not in any hurry to reduce it. I think we're always active sort of soliciting price. I think there's a bit of an asymmetry right now in so far as -- with the rapid change in default and repayment profiles over the last year, our model has recalibrated pretty significantly, pretty dynamically. I think if you're trying to judge current production of loans based on how last year's vintage performed, you'd be out of date. And so we obviously are much more real-time understanding of how current loans are being modeled. And that creates an asymmetry such that I think anyone trying to price the loans based on like last year's production are not going to meet our price, and what will be required is a bit of time for the curves to play out. And as they do, hopefully, the prices will -- we'll get to our clearing price. But if they don't, we're happy to sort of hold the loans and earn some interest income in the meantime. It's not a core business model of ours, but since CFO, I'm not going to turn away the extra dollars these days rather than sell the loan at a price that I think don't make sense. So it's sort of an ongoing non-region sort of exercise for us.

Is there a way for us to think about what would happen in terms of the clearing price and maybe what you guys might take in terms of a loss to the P&L as you guys think about ultimately selling those again, to the point is, hey, the loans are now different than they were a year ago? So what would that look like in a sale as you guys [ trying ] to get there?

So it would depend on the vintage. So on our balance sheet, we've got a variety of the past vintages and 2 things have changed versus a vintage from a while ago, 1 are the loss pattern, so some of them were originated 6 months ago when we didn't know where the world was going and there the losses are higher than we had anticipated. And then, of course, the interest rates have gone up. So I don't know if you were to take a loan originated today with our current model, we think those are over performing, those are being sold at par. And that's where I would expect to clear something in the secondary. If you go back 6 months when some of the loan was originated at a rate when the interest rates were lower and the loss rates were lower and both things have changed, then there would be some discount, but I don't think the magnitude of these discounts is particularly large in the grand scheme of things.

Okay. It's not like a material expectation that we should have out there in kind of maybe the back half of this year, the sale would create some sort of hole on the P&L.

It may create some pressure on the P&L. But it would be done at what we consider to now be a fair value. And in fact, we mark our loans to market. We don't -- most banks use what's called CECL accounting, which involves reserves and sort of book accounting, we mark to market. So we're doing our job right, we should be taking increases in interest rates into the P&L sort of in quasi-real-time.

Yes. Okay. Fair enough. Switching to kind of the model, right? One of the things that has certainly been messaged from you guys is that the model continues to improve as you guys are incorporating more information, macro, everything involved with that. Can you talk about just how the underwriting model has changed over the last maybe year or 2? What is different about it? And then just help quantify kind of the improvements in accuracy.

Sure. And it's probably worth describing the fact that when you're underwriting credit, consumer credit, you're implicitly doing 2 different things. One of the things you're doing is trying to take a point of view at the borrower level on that person's characteristics, their history, their cash flow, their propensity to repay. And that's what we believe, that's where the majority of our historical effort has gone. It's what we believe we do very well. The second thing you have to care about is the fact that, that borrower will be impacted by the macro somehow, and you have to worry about that. So on the first topic, that's what we try to get very good at by using alternative models and alternative data. And that process, as you said, I think that in the past 12 months, we've probably gotten better or more accuracy gains than we had gotten in the prior 2 years -- 2 to 3 years. And there is -- the reason for that is, I think, maybe as of -- maybe a year ago, we were starting to get bogged down in a lot of regulatory process. And we straightened that out with the CFPB, we fixed that. And the velocity of sort of alternative data -- sort of new alternative data, the velocity of model improvement has really accelerated. So that's very exciting to us. I would generally say, in 4 out of 5 years, that's what's driving the enterprise value of our business. Like that's what's making a difference in your ability to take a group of borrowers that otherwise look pretty indistinguishable and your ability to cherry pick the good ones and avoid the bad ones because within a FICO bucket, there's actually a lot of variance in the credit performance. So that's good. Now of course, we're in the fifth year where all of those gains haven't really manifested themselves because everyone is getting impacted by the macro. And on the macro front, we've historically spent less time working on that because it's not a good machine learning problem, right? There's not -- this recession is not giving us good training data for the next one because it probably won't be a pandemic plus stimulus, it will be something else. But we've taken this -- the opportunity in the past year to get much better at what I described at the beginning, which is -- it's actually like quite hard to isolate what the impact of macro is on your portfolio. And I think we're as good as anyone at doing that now at measuring it in what I call predicting the present, like how good are you in real time in understanding what macro is doing to your portfolio, not like understanding it 6 months later, but as it's happening. And my rough view of macro is we're never going to be able to predict it. I don't know that anyone can -- if anyone can for more lucrative ways to deploy that technology. But I think we can aspire to react the quickest to it and measure it the most precisely. And our ability to do that versus a year ago's dramatically different because it's just not something we ever had to spend a lot of time on. And it won't be the defining sort of underpinning of our business going forward. I still think loan level, borrower level, credit decisioning is still by far the more important aspect, but the ability to sort of ebb and flow with the macro, I think, is a significant improvement as well.

Staying on technology, LLMs are kind of older rate right now, right? That's what -- assuming everyone here is going to get a question on this throughout today. Talk about just AI broadly, like is there low-hanging fruit for AI that you guys are now incorporating on the [ generative ] side? Is there any way that, that shows up into models where do you guys incorporating any large language models into your actual AI [ desk tech ]? Like what changes as it relates to the big picture tech with Upstart?

Sure. So LLMs are large language models, the most vivid example of that in the press right now is ChatGPT that everyone is talking about. There is low-hanging fruit, but not in the way you're thinking about it. To me, I think our rough view of something like ChatGPT is, it's another example of an event that has made the -- it's made the description of AI very vividly tangible in the public eye, right? Other examples of this are when -- so a while back, one of the big first seminal public sort of realizations of AI came when a company called DeepMind, which had been acquired by Google, started beating all of the World masters at a game called Go, which historically is thought of as a very complex game. It's very hard to model in any sort of predictive or algorithmic way. And suddenly, you have these models beating, these Go masters. And then, of course, these cars are starting to drive themselves around. And now you've got this thing that talks to you like it's a robot trying to take over the world. And I think that makes it very vivid and tangible in the public. But like nothing technological has changed, right? This is just exposing this in ways that make it very clear that it's a unique and interesting technology. And they've trained their models in elegant ways. There's no doubt. But like I think there's nothing new in the technology front, right? So we've been saying for a while, look, you could take this technology and apply it to decisioning credit outcome, you can try and predict who's going to pay you back and who won't. And it will do a better job than the FICO score, right? The FICO scores invented in the '80s when you -- where you're basically using linear regression, and you've got these complicated models now. And that's sort of very theoretical and hard to grok. And so ChatGPT been a great marketing thing for us. It's like, oh, wait, these models seem to be able to do interesting things. The low-hanging fruit for us is the following. There is a world out there that continues to pioneer more and more sophisticated models for all different applications. I mean some of the most complex are the self-driving cars, for example. We can't use the cutting-edge models yet for the application of trying to predict credit outcomes because we don't have enough training data. And what that means is, okay, we've spent prior 8, 9 years lending to people collecting -- the training data is whether they pay you back or not. And so we've been observing those outcomes versus all of the variables we have in the borrowers. And now we have a couple of hundred thousand loans out there over 9 years. So it's a couple of million or billion repayment events. And so that's a very powerful set of training data that allows us to use the models we have. If you took all this data and tried to throw it into a linear regression, it would regurgitate. But we still don't have enough training data to harness the power of model that's telling a car how to drive itself, right? That requires trillions of data points, and we have billions. So the very fact of us continuing to grow the corpus of training data over time, lending more and more alternative data means over time, we can sort of continue down the path of using more and more sophisticated model forms because the training data is still our limiting factor. But there is an example of ChatGPT, I don't think have changed anything for us in that journey. They've just made it more vividly sort of tactile to the general public.

Makes sense. All right. Autos, all right, so you've now partnered or you're now powering, excuse me, lending at 27 dealerships. Said you won 42% of loans for those dealerships. Is it fair to get really excited and to extrapolate this under the larger dealership opportunity, right? Like how does that 27 become 200, 2,000 right to that 42% stick?

So directionally, maybe not precisely, but the rough landscape is there's about -- there's probably 15,000 dealers of reasonable scale in the U.S., it's actually like 40,000, but 15,000 of them are like legitimate franchise or some of your independents as well, but some of these dealers are selling 3 cars off the backlog, right? There's like -- I think our target list would say there's probably 15,000 dealers that are interesting opportunities from a lending volume perspective. 800 of those are using our software to sell the car. Nothing to do with lending. That 800 number is growing by 100, 200 every quarter. It's the fast-increasing penetration of the overall footprint. Of that 800-some dealers, you said I think there's maybe 30 of them where we've enabled our loan product. In those 30, we're getting pretty good takes, pretty good share of wallet, if you will. So the first question is, is that share of wallet representative directionally. You might think that we've probably started with the dealers where our loan product was particularly suitable. So there may be some drop off as you sort of increase the penetration. On the other hand, the loan product itself is improving quite rapidly. We're going down the same path of accuracy improvement that we did in personal loans over 8, 9 years. So it may not be exact, but I think that is sort of our expectation. And then the question is, how does 30 increase? The simple answer is capital, right? We could turn 30 into 800 pretty much overnight because we have all of them using our software. We're just not delivering the loans. But the capital is not there for 2 reasons. One, we've not been aggressive at bringing the capital into the system because we're still calibrating the models. So you don't want to do that until your model performs, right? If you go and lend a bunch and the credit doesn't perform that you're going to sort of shoot yourself in the foot. I think our models in auto are now as calibrated as they are in personal loans, maybe as of the last month or so. We're starting to be aggressive on bringing the capital. The other factor, of course, is the capital sources are still nervous about the macro. And so it's going to take some effort to sort of -- they need to become comfortable with the world and then they need to become comfortable with the new product. And as that happens, we'll expand 30 to 800 and then we continue to grow 800 just as a general sort of dealership retail software play, which has been working very well. So there's a lot of dynamics in there, but I think they're all -- the framework is pretty exciting for us because at the end of the day, what we've proven is that the market fit for the loan seems to be really good.

Makes sense. We're over time, but I'd take any audience questions if anyone has one. Okay. Sanjay, thanks so much for the time. I appreciate your coming up.

Happy conference, everyone. Thank you.