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

Hey, thanks, everyone, for coming in -- for sitting in down to the last stretch of our conference. I -- thanks, everyone, for coming in last couple of meetings. And Sanjay, thanks for flying. I know it's not an easy flight from San Francisco. Someone was telling me there's no -- nonstops, you guys like do a connection.

I think you are nonstop. It feels good to be live again. Okay.

Yes. No issues. Perfect.

Just to get us started, I have some questions specific to the business. But just at a high level, if you can kind of start with Upstart as a business and a business model for folks who are less familiar with the name.

Yes. Sure. I guess let's see, the foundation of our company really lies in the theory that -- well, as you guys know, credit is largely an exercise of prediction. You're trying to predict who's going to pay your loan back. Our observation is that a lot of methods used to make that predictions are pretty dated in terms of underlying algorithms. Most of us come from technology background. The consumer Internet world has been working on predictive technologies for a while now and has made a lot of great advances as well on computation or power we have. And our simple thesis was to apply those credit prediction algorithms to the science of trying to predict credit outcomes. So that's sort of been the journey of our company. I think that the progress of our company and the underlying value of it lies in so far as we've been sort of systematically improving our ability to make credit predictions over the 8 years of our existence, and we've gotten systematically better at them. I think we have got pretty good proof points that versus a traditional model or a traditional metric like a credit score, we can do a much better job of ranking and doing what we call separating risk. And so everything's built on the back of that. We can get into the business model, et cetera, but that's sort of the core idea.

Terrific. Upstart, you're certainly innovating in a market that's been around for a very long time. It's a very large market, $700 billion consumer and auto market. When you combine the 2, are there certain aspects of Upstart that enable you to win against competitors who have been doing this for much longer time?

Yes. No, it is a very old market. Thousands of years. I think it comes down to what I just described, which is, just to put a point on it, there has been -- I would say particularly in the last 5 to 10 years pretty dramatic advances unlocked in the science of prediction. And I think we are one of the very few companies that are applying those technologies to prediction in the U.S. unsecured consumer space. And so to the extent we have an advantage, and we believe we do, we believe it comes down to the ability to predict credit outcomes better. Now it turns out that when you're lending someone money, the single most important thing you can do is predict who's going to pay you back and not -- the ability to do that better allows you to better avoid the people who will not pay you back, and therefore, too for the people that are left -- are remaining to give them better rates because they're subsidizing less people who will not pay them back. So the better you are predicting those things precisely, the better the outcomes for the lender, for the borrower who's paying lower rates or getting higher approvals. So it really comes down to our -- the 2 pillars we have. One is just science of credit prediction where we think we are doing a lot of innovation. And then the other one is the sort of [ converse ] of that, which is taking friction out of the process, which -- the way you do that is similar. There's only friction in credit process because when you give someone information on which they underwrite you, then they have to verify that it's true. And if you don't have a good ability to predict whether they're lying to you or not, you have to ask for a lot of documentation like paystub or driver's license or whatever. So the better you are at making these estimates, the more you can sort of automate with confidence and then, of course, the better outcome for the lender and the borrower. So it's all underpinned by the same thing, which is this technological underpinning that really has been made possible by the advances in the last couple of years.

Yes. And so can we dig into that a little bit further? The certain aspects of like AI or machine learning that makes it well sorted to kind of remove the friction? I mean I think almost 70% of your loans -- or over 70% of your loans are basically automated like there's no kind of human things. The automation of loans and also kind of just making sure you have better outcomes from those loans, what is it about like AI and machine learning that enables that to happen?

Yes. I mean that's sort of the general class of these technologies is, of course, referred to as AI or machine learning. And I think they [ sensify ] things a little bit when we describe it, but it's really quite basic at the end of the day. Like these are just next-generation prediction technologies. And all that's really happens, like if you sort of rewind to the '80s where there was obviously less compute power than we have available today, in order to make a prediction, you had to make very simplifying assumptions like. So a classic prediction algorithm that we used back in the '80s was a regression. A regression model is to predict an outcome. You have a bunch of input variables and an objective function that you're trying to predict what will happen given a new sort of input condition. And in order to make those models work, you had to, for example, you had to assume the world was linear, that relationships in life are linear. And you had to assume that, well, if there's 6 different variables that I think are useful to predicting the outcome of my objective function, I have to assume that they're all independent. And I have to assume that I can sort of examine each of them in isolation of the other. And these are dramatic assumptions -- simplifying sort of assumptions. And of course, we had to -- we used to have to do it because we didn't have the compute power to handle things more complex. But it turns out when you make simplifying assumptions like that, you dramatically degrade your ability to make a prediction. And all machine learning really is at the end of the day is it is an amount of compute that's allowed us to relax those assumptions. So I no longer have to assume a relationship between an input and output is linear because life is not linear. The extent to which I care about and the rate at which I'm caring about a specific input variable is dependent on its very state. And the extent to which I care about an input variable is dependent on the state of other variables, like how much I care about your income will be governed by where you live, for example. And so machine learning is essentially something which allows us to relax assumptions of linearity because the world is not linear. It allows us to relax assumptions about independence of variables because variables do, in fact, interact in very complex ways. And it creates a massive surface area to explore a pattern. And machine learning is simply the compute power that allows us to explore that. And when you can, you unlock all these nuanced patterns that a linear regression can't find. Now it turns out the majority of the industry is still using the equivalent of a linear regression in trying to predict things. And it's not surprising that compute power that allows you to relax those assumptions. I mean this is -- you can look from an output perspective, these are technologies that are training computers to be just grand masters or training cars to drive themselves. Like they are observably better in terms of predictive outcomes than traditional methods. And so it shouldn't be surprising that they're better in trying to predict when and how a borrower may default based on all the things we can know about.

Yes. That makes sense. And sort of going back to what all these modern means, you've had -- your loan performance in 2020, 2021, like loan performance has arguably been a bit better than, I think, what you've talked about given what your expectations are. Can you talk about like -- have you underestimated how good your models are? Or do you just get, like, a COVID benefit where your kind of borrow behavior had vastly improved? Because of -- again, I'm just trying to dimension how much of is it your models versus just kind of a borrower environment?

Yes. You can sort of decompose the exercise of underwriting a loan to a consumer into 2 sort of notionally separate exercises. One is the evaluation of the borrower themselves. You're trying to figure out, based on someone's educational profile, their deployment profile, maybe some statistics about where they live and what they do with their background is with the cash flows, through their bank accounts. And you're trying to sort of evaluate that borrowers sort of tendency to repay versus default. And so this is a relative exercise. You do it relative to other borrowers. So this is what we call a ranking problem. And this is what is, I think, a very good candidate for machine learning problem and something that I think we do very well. We believe we can rank and relatively order risk in a way that's far superior than a traditional credit score, for example. The second thing you need to do in this exercise implicitly or explicitly is you need to take a point of view on how the environment and the macro will affect all of those borrowers. They can all be higher risk or lower risk depending on what's going on in the world. And this is not a good machine problem. It never will be. Each recession is unique. I don't think -- there's not enough training data to train a machine to predict the next cycle, at least that's what we're aware of. Maybe someone can do it, and they'll be far wealthier than we will. But at best we can aspire to react most quickly to macro events as they occur. And importantly, more precisely, as macro events impact the entire world differently, and each event does it differently -- just as an example, the current cycle we're going through lower affluent borrowers have been dramatically impacted from a default perspective, more prime borrowers have not. The exact opposite was true in 2008. So reacting precisely and quickly to macro events is something machines can do. But all that to say, I think in the first thing I described, which is looking at borrowers, understanding them and understanding their characteristics and understanding them from a relative default standpoint versus one another is something that we believe we do well. We believe the edge we have over a traditional metric is growing systematically, and it has not changed during this past cycle. And that's where we believe a lot of the alpha in our business is. The second thing is something you described, which is, okay, well, in terms of once you underwrite the loan, how do they perform? Well, they're affected by the macro. And I think if you go back in time and you look at how we deliver returns to investors versus what we were targeting -- I mean in 2017, 2018, even early 2019, they were roughly on target slightly above. Just to give numerical parameters, we historically targeted at 7%, 8% gross return, and we would deliver an 8% to 9% return. In 2020, that went to like versus the same target we're delivering a 12% return against it. So that was all attributable to stimulus. Now our ability to -- that either average numbers across the portfolio, our ability to get the ordering right, which is important under the hood was much stronger than it would have been had we been using a credit score. But everyone over delivered. And of course, in the last, I would say, 3 quarters or so, there's been a macro impact where there's been a divergence, and prime borrowers are roughly back on target, and less affluent borrowers are under-delivering. But it's not because we got them wrong from a relative risk standpoint, because some macro events has come along that we couldn't predict. In this case, it was the stimulus-ing and destimulus-ing of the economy and the less affluent borrower as a result is upside down. So there's a macro component that will always be noise. Our job is to react as quickly and as precisely to it as possible. But that is sort of the residual risk that exists even if you're perfect at understanding the relative ranking of risk among borrowers.

Yes, yes. And the macro, we have seen you'll react to kind of these -- kind of the credit markets kind of having lower appetite for loans. But before I get to that question, when you think of lower guidance that you had last month, how much of it was because of kind of funding constraints versus borrowers' appetite for loans?

Well, the 2 are not additive, of course. They're overlapping. So when you have a contraction, it's the greater of the 2. That's governing it. What's governing our current volume is funding constraints. Now the fact of the matter is even if funding had been ubiquitous and infinite, it would be a contraction in our business for the simple fact of the matter that loss rates -- risk has gone up in the economy. Loss rates are higher. All of us are defaulting at higher rates, whether you're a prime borrower or a lower prime borrower. All of us are defaulting high rates than we were in mid-2021. And the fact of loss rates going up, as long as you're originating to an APR cap, like most responsible institutions do, your losses go up, you're going to push people out of the approval box. And so it is, in fact, a feature, not a bug of a system that lends less relatively when there's higher risk in the economy. But our view is that the funding markets have -- the credit markets have reacted disproportionately to the risk. And so the bigger constraint on our business right now is the sort of anxiety of the funding markets and their willingness to fund loans.

Yes, yes. And of course, you have talked about now basically kind of filling that gap in -- by using your own balance sheet. And if I can ask like, certainly, when -- initially, when you were going public, you mentioned really not using the balance sheet, and you've had a couple of turns. What is the current view on kind of the use of balance sheet?

So I'll object to the term filling the gap because it's definitely not that we can or want to do. But it is worth dwelling on this a little bit because we admittedly been a bit schizophrenic on this narrative. And I guess I would describe it as follows. We said from the beginning, and it continues to be the case, that we will not be a bank, we will not be a balance sheet business. We will not be a business whose point is to earn net interest income or to earn a spread or a yield. It's not what our model is built for. I think originally as a consequence of that, we're almost religious in saying loans shall not touch the balance sheet other than for one use case, which is -- we've always said like one of the reasons we went public and raised money was to accelerate our R&D. From the life cycle of a product, you want to do it for a couple of months on balance sheet, develop some loss curves, some history, then you can take it to the funding markets, whether it's a bank or a credit fund. You can show them the history, and then they'll adopt it. And at some point, the product becomes ready for the capital markets for the banking world and then its third-party capital, not ours. And we've historically said, look, once something is ready for the markets, then we're out of the game. And we were almost like a bit kind of [ zealous ] about that. Now what's changed or, I guess, maybe our realization in the last 12 months, it's sort of related to what I described, which is, in our view -- and maybe this was a bit naive on our part in understanding this before, but in a dislocation like what we've had -- and to be clear, the dislocation really exists. To the extent there's 2 different consumers in the U.S. right now, there's a more affluent and a less affluent one. There's been a big dislocation for the less affluent consumer. And I don't think this is really -- but my view is this is not really coming through in the public narrative. People just talk about the U.S. consumers. Most of the financial machinery that's public in our economy is geared at serving the prime consumer. We're somewhat -- I don't say unique, but we're in the minority in public companies that the borrower that's at the heart of our mission is a less affluent borrower. So there's been a dislocation with the less affluent borrower in the U.S. over the last 12 months. In our view, in my view, the funding markets have disproportionately panicked. Like if you just look at the sort of actual risk as it's materialized and you sort of talk through the current conservatism in the underwriting, you do some reasonable risk scenarios, the current state of the credit markets -- and this is maybe unsurprising, but it's -- the decision-making is more emotional and fact-based. And so in a dislocation like this, I think it's been our observation that, look, we're trying to convince or at least give transparency to funding partners or credit institutions about how the models have recalibrated to the new normal and whether -- how to think about the upside versus the downside case given the forward macro and how conservative our assumptions are about the macro. In many instances, their response is, hey, look, like if you believe that, like I have to trust you, and there's asymmetric knowledge. I can see how the models have recalibrated, I think you can trust my word. So in many instances, they're like, well, can you put your money where your mouth is? And being able to do some confidence signaling with our balance sheet, it's occurred to us it's useful in a world where there is a dislocation. And everyone candidly is on the sidelines right now trying to figure out when to jump back in, but nobody wants to be first, right? There's a bit of a herd mentality that takes over when you get emotional and a bit panicky. And there's a use for us, a very surgical one, where in a dislocation -- where we have asymmetric knowledge about our model, and people want to see signs of confidence in order to take to their credit committees, we can play that role. We cannot fill the gap, right, versus where we were and where we are today. And we want to because, as I said, legitimately, when risk is higher in the macro, you want volume to contract. It's reasonable. But this is the difference. So the summary of it is we're not -- we haven't changed our strategy, our business model. And we're not going down the path to becoming a bank. We historically have been religious about not touching the balance sheet. It has made us realize that there is a use case for our balance sheet. Even in our core business, where in a dislocation, we can sort of lead the way a little bit. And we thought long and hard about changing the signaling to the public markets. But in the end, it's sort of what we decided is best for the business. And so we need to now socialize a somewhat changed narrative. I know you're very sort of acutely aware of the pain of doing that, and so am I, but it was the right thing to do for the business. So...

Yes. And by leveraging your balance sheet, have you seen change in behavior among the lenders and your banking partners?

I'll just say it's very early days. Very early days. But yes, everyone is eager to see. Everyone is like -- I think you get to a point in a dislocation where everyone is on the sidelines, and nobody wants to be the first back in, and nobody wants to be the last back in. And so the #1 question I get from credit investors is asking me what the other credit investors are doing because they're all trying to figure out what the right timing is here, and they're all taking their own views on macro.

And what's your answer to that when they're asking how the others are doing?

I tell them it's none of their business.

Okay. Perfect. I'm going to open it up for questions in a second, but let me get one more in before that. So given that you're using this AI and ML and it's more database and it's -- arguably, what you've talked about is it's a better way of evaluating risk. And with the macro getting pressured, when we come out of this, would you say that you'd probably end up with more market share because then -- kind of the last couple of years, pretty much -- and anyone could basically do this and be fine. But like in this market where you just really need to find out who the good borrowers are or the bad and really predict like kind of borrower outcomes. Would you say that you'll actually take market share over the next couple of years? Is it a better environment for you all to take market share?

In the current environment?

Yes.

No for the reason that -- so the dynamic is as follows: precrisis, if you go even back to 2018, 2019, it's a fairly efficient market in the sense that there is sort of return required from unsecured credit. And in that world, we were systematically growing market share because the only really moving part is everyone has more static models, and ours is getting better. We're improving our ability to cherrypick the better consumers. And of course, the underlying subtext is when I say cherrypick, the majority -- it is a truism in credit that the majority of the people are overpriced, and a minority are underpriced because at the end of the day, if you had perfect hindsight, it's a minority of people who didn't pay you back, and everyone else should have gotten the risk-free rate. So everyone else is paying too much. The majority of people are paying too much because we're subsidizing a minority of defaulters who are in the system that the model couldn't keep out. So as we have gotten better and better at that, we would gain the sort of market share because it would be better. Avoiding the defaulters, pricing the remaining people closer to the risk-free rate, and that was really the only moving part. In a market like today, there's many moving parts. And the predominant one is the required rates of return are all over the ballpark right now. And then sort of in particular, if you're using your own balance sheet to originate loans on a scale that we don't, you have billions of dollars on your balance sheet, you are a bank using your own technology, you can return any number you want, right? And so you have a lot more flexibility in taking decisions that, at least in the short term, may be good for origination. In the medium term, they may or may not be good for credit. So whereas us, because we are predominantly a model that governs third-party decisions, third-party capital, we're sort of -- we are in the service of what that capital demand as a return. So in the current environment, there's just a lot of distortion. Now I think that under the hood, our model is systematically getting better. And when the world snaps back to something that's sort of more long-term normal and efficient, I believe we'll come out of it looking quite good because our ability to price the borrowers in 2023 will be even better than it was in 2019 and in 2020 or 2021. So all that to say, there are good things that continue to happen under the hood. But of course, when you're applying technology to a market like lending, and this is where I think some of the investors out there who are more technology-centric are still sort of getting their head around this. The underlying market itself has macro volatility. And in our model, it won't ever materialize to a meaningful extent in the form of credit risk. But there's legitimate volume risk in lending because lending comes and goes with the risk in the economy. And the risk in the economy is cyclical. So I think that if you look at the long-term trajectory of the business, absolutely, there continues to be gains in market share under the hood, but there'll be distortions on both sides as macros are sort of [indiscernible] increase.

[indiscernible] Any questions from the audience?

[indiscernible].

It's not -- it's a bit optical. The true risk is the one we've discovered, which I'll describe. So if you think about our funding channels -- there's really 2 channels. The first channel is we give our technology to banks and credit unions. And they use the technology in order to originate for their own balance sheet. And that historically has been a 1/4 to 1/3 of our business, okay? If we have a borrower that is not wanted by those banks, right, because it's out of their risk profile or it's just above their aggregate balance sheet capacity, then it goes to a special bank that exists for the -- we call it a conduit bank. It's for the sole purpose of originating the loan and instantaneously selling it into the credit markets into the institutional world. Now from an optics perspective, that bank, and there's 2 of them that play that function, are large concentrations of revenue because they're the ones who pay our fees. The reality is the actual underlying risk is the funding party. I'm used to say, well, we've got a lot of funding parties behind those 2 banks that appear to be large concentrations. And it's a very diverse world. Now what we've discovered is -- was all of them act in concert at the same time when there's risk. So we have actually suffered, I would say, a contraction in funding availability, but it's not because we had a large concentration that went away. If one of those banks disappeared, we just replace it. There's other banks that provide that service. The true risk that we've discovered is we had very outwill capital relationships with the funders that were behind those banks, and we need more resiliency in that capital base. And that's the learning that we've had as we've gone through the cycle that we need to set up for the next cycle. But the true risk is the one that materialized, which is -- we were 2x oversubscribed with capital in 2021, and that felt pretty good until we realize that, well, we react the same way to the same event. So that's maybe a long way of saying that's optical concentration risk, but there's actual risk that has materialized in the sense that we need more forward committed capital.

I see. Can we maybe take it off-line? Because we [indiscernible], but thank you for sitting in or anyone else who has questions, thanks for coming.

Thank you.

Appreciate your time. Thanks, Sanjay.