Thank you. Thank you, everyone, for joining. This is the Upstart session, so welcome to the Upstart session. I'm Mahir Bhatia. I cover consumer finance and payment companies here at Bank of America Research. For your tech investors, you can also vote in the consumer finance or the payments categories in II, so please vote for us. Before we get started, I do have to read some of these disclosures. Today's discussion may contain forward-looking statements that relate to future results and events, which are based on Upstart's information available as of 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 measures, which are not a substitute for the GAAP results.
Please refer to the company's filings with the SEC and its IR website for additional information, including GAAP to non-GAAP reconciliations along with other disclosures. With that out of the way, it's my privilege to be hosting Sanjay and Paul. Sanjay is the Chief Financial Officer at Upstart. Paul is the Chief Technology Officer. Thank you both for joining us. Really appreciate you guys taking the time to come in today.
Thanks for having us.
Now, usually when we have these discussions, we start with the macro backdrop and then start getting down to company-level questions. You guys just hosted an AI Investor Day. We are at a technology conference, so I'm going to try to do things a little different. We'll start with a couple of questions around AI and Upstart. The first question, the first thing I think we should acknowledge is that Upstart has been talking about AI for much longer than the last 24 months when it has become much more of a buzzword, right?
Yes.
You've been using it, machine learning techniques, some AI to a certain degree for many years. The question really is, how do you distinguish yourself or your models from other lenders, right? When I talk to other lenders, they tell me, hey, we use AI too. We have machine learning techniques we've been using. And we're talking large, sophisticated companies, the Capital Ones of the world, the Discovers, large companies that have been using these for a while. So what is different about the way Upstart uses AI, and what's the sustainable advantage?
Yeah. Yeah. I mean, we get this question all the time. We've been really getting it since we started the company all the way back in 2012 when we first got started. Maybe I'll start sort of from a zoomed-out answer, and then I'll give some of the more specific details. Whenever I talk to my team, I think one thing I like to say is that ultimately, in the long run, the only actual competitive differentiator is speed. I believe that very deeply in the sense that there is nothing fundamentally that would stop any player existing or new from attempting to build what Upstart has built. Anybody could pick up and decide to start. In order to actually start, they would need to do a few things.
The first is that they would have to invest in a fairly expensive technology team composed of machine learning researchers. We have a team today of about 70 machine learning researchers. You would need to build a team like that that can build this kind of technology. You would need to empower that team to be able to actually do work that gets to production. This is something that turns out to be quite challenging at legacy financial institutions that have very particular kinds of risk controls, risk committees. You would need to start gathering alternative data, which is something that has been pretty important to us. You would need to gather alternative data about actual borrowers that you can connect with repayment outcomes on consumer loans, which typically have a life cycle of multiple years. You have sort of this multi-year data gathering process.
You need to actually build the actual machine learning algorithms. Now that you've got the team, you've got the data, you need to sort of build models that are capable of making use of these things. It turns out that a lot of the work that needs to be done is not like you just take sort of off-the-shelf algorithms that are open source and apply that to a problem. You have to do a whole bunch of modification to make them work in the sort of specific case of loans, which have a bunch of technical properties that are quite different than sort of other machine learning domains like image classification or text prediction or what have you.
You need to sustain this effort through, one, regulatory pressure and, two, the sort of very significant likelihood that your early models will very significantly underperform when you do not yet have enough training data. You sort of incur a fairly significant amount of financial and reputational and regulatory risk in order to have possibly the outcome that five to ten years from now, you will have a model that will be sort of in the neighborhood of Upstart's. I do think people could decide to go and do that. We have not, in our time, seen a lot of activity in that direction. In theory, I think anybody could decide to start.
Of course, then the question of, would we still have sort of differentiated enterprise value in that world where, let's say, five people decide to start doing this today is like, what will we accomplish in those 5 to 10 years that they're pursuing the work that we did over the last 10? The answer to that, of course, is just like, what does our team get done during that time? What's our current rate of progress? The thing that has been really consistent throughout our history is that there's been essentially no diminishing of the sort of returns to research investment. We've basically been able to continuously compound sort of model accuracy with sort of quarter by quarter over our history.
We've been able to make the models better by continuing to sort of innovate on the algorithm design, getting more columns of variables, getting more rows of data just as repayments come in, which unlock more complexity in the models, and then investing more recently in the sort of compute speed and memory layer that unlocks the ability to support larger models.
Good. One of the things you mentioned was about.
Is Paul's mic on?
Oh, it's on.
Oh, OK.
Can I speak a little louder?
Yeah. Hi.
Is that coming through the PA?
I think that's just my voice.
Can everyone hear Paul? I guess I don't know for the webcast. You can? Oh, sorry. They can't hear Paul. Can we check his mic? Pull it higher.
Yeah. Hi.
Can you just tap it?
Hey, it's not.
Is he? No. There you go.
There you go.
There we go.
Oh, much better.
Better.
The hold-on button. One of the things you mentioned, Paul, was your continuous improvement in the model. I think you've talked on calls about 2%-3% a month. I understand it's not linear, but talk a little bit more about that. How does the model improve? What are you doing every month to make this model improve?
Yeah. The areas of improvement are divided into the four categories I referenced earlier. The most common one is investment in model architecture. This is essentially work in algorithm design where we might introduce different types of algorithms into the ensemble, like a different kind of neural network or a different way that the models get ensembled together or a different way that we modify the loss function. All of the inner workings of a machine learning algorithm. This is where probably the plurality of our research time goes. We have gotten more wins from this area than any other. The second is just getting new types of data about consumers. There are a lot of third-party data vendors. There are a lot of ways to collect data directly from the consumer.
We're always experimenting with new ones and seeing which ones add incremental value on top of the existing ones. Historically, every couple of quarters, we've found something new that sticks. The third bucket is investments in compute and memory. This basically unlocks the ability for us to support larger, more complex models in production without running into the latency constraints of your model runs too slowly and the consumer leaves because they're tired of waiting or the training cost is too high to be able to regularly retrain this model. The last thing that kind of just happens in the background is getting more rows of repayment data. Rows of repayment data basically are sort of the hard constraint on how many variables you can have in the model and how sort of fancy your algorithm design can get.
As you get more repayment data, you basically can just unlock more variables or more complexity in algorithm design.
Another area we've talked previously before about model improvements has been calibration and the macro calibration. I think one of the statistics mentioned at the AI Day was 55% of excess defaults that you experienced in 2022 you wouldn't experience if you had a similar event now. Talk a little bit about why that is the case. What gives you confidence? Obviously, the next macro crisis might be different. It's not going to look like 2021, hopefully. We don't have a pandemic and similar stimulus. What gives you confidence that the model is much more macro resilient today?
Yeah. Yeah. This one's actually pretty straightforward, unlike a lot of other things that we do. Historically, we basically didn't care about macro at all in our models. We viewed our models as having one job, and that was just to separate risk, rank risk, figure out which borrowers are relatively more or less likely to pay back their loans than others. What that meant was that the models are essentially implicitly calibrated to the entire period of observed training data, which you can think of as the past 10 years or so. After the events of 2022, we realized that it was actually quite valuable and important for our models to care about the sort of current macro conditions and not just the average macro conditions of the last 10 years.
We introduced a number of things into the model that had the effect of making the model time-aware. By time-aware, I mean it became aware of what calendar month any given payment due date was in. It was able to isolate out from that first a sort of aggregate macro effect, which we've subsequently published as the UMI or Upstart Macro Index, which basically is just a statistical measure of the sort of likelihood of defaulting when your payment is due in a particular calendar month when you hold all the other characteristics of your borrower population constant. The second thing we were able to do was let the model dynamically interact that time variable with any of the other borrower-level characteristics.
That means that maybe there is an aggregate relationship, but maybe there's a sort of different aggregate relationship in just borrowers with higher or lower credit scores or white-collar versus blue-collar workers or people with more or less types of education or maybe people in a particular industry like nursing or people who study computer science. If there's a particular sort of sector-specific shock, the model is able to dynamically pick that up. To your question of why do we think this is sort of something that will work in the next macro shock, which is certainly going to be different than the last one, the answer is that there's nothing we did here that was specific to the 2022 macro shock at all. There's actually nothing, sort of, there's no kind of economic macro forecasting that's going into this model.
It's sort of just a fully general machine learning solution that just has its core innovation is just making time an operable feature in the model that can interact with any borrower-level characteristic. This will work as long as it's the case that the next shock happens in a way that is related to any of the thousands of borrower-level characteristics that we observe, which is a really very broad set of things. It could be happening for certain occupations or certain skill sets or certain sort of socioeconomic strata, certain education levels. All of that stuff would work. Now, obviously, of course, in theory, it could happen in some way that isn't reflected. I think for the vast majority of the things that people talk about when they think about macro shocks, those are things that are reflected in the borrower-level variables.
Now, what that would mean, assuming that that plays out as I described, is that it does not mean that there is no effect to the business. The business will still be impacted because we will not be able to approve as many people for loans. The credit performance will be in a much better place because it means that new loans that get made will very rapidly be issued with a model that is properly calibrated to the new time environment that is observed.
I think it used to be eight quarters it took to adjust the model. How much would you say it is now? Is it immediate? Is it?
Yeah. In the actual history of the sort of 2022 shock or the end-of-stimulus shock, it took eight quarters for new loans getting issued to be properly calibrated to the new environment. That was just because, again, the model was basically training on the whole sort of history of what we'd ever observed. The past 10 years, it was calibrating to that. Whereas with this sort of new setup, the model is able to start picking up changes within the sort of very first quarter. That's why we have that factoid about you get half the loss reduction or half the excess loss reduction almost immediately. Within just a couple of quarters, you basically get to fully calibrated. Instead of eight quarters, you can think of it as being two.
Even within those two, you're already cutting the excess losses very materially right away.
Got it. Now, Sanjay, maybe turning to you, we talked about the underwriting model improving, becoming much more resilient. Where do you see that impacting the business today or in the financials? I understand in a future recession, future macro turbulence, the model will adjust fast. Do you feel the effects of that in the business today, whether it's from borrowers being more willing or loan funders or loan buyers being more open to buying an Upstart loan, whether it's moving up? Just in general, are you seeing any positive impacts from that in the business today?
You're asking specifically what Paul described?
Yeah, like just including the macro calibration and the model improvements.
I mean, everything in our business over time flows from credit performance. If you cared about one thing, if there's one KPI I had to track, it would be the credit performance. That gives confidence to the capital markets and the investor side of the equation over time. It allows us to more confidently sort of approve and underwrite the borrower side. It will, I mean, immediately show up as capital resilience, if you will. The whole business model we have is predicated on us being smart at credit. This is a huge upgrade in our ability to do that. To me, it's hard to, there's not a line item on the P&L where it shows up.
Right. No, and that's why I guess maybe lift the hood a little bit. How do these forward flow agreements work? Like a big private fund comes to you and they're like, how much testing do they do? How much do they actually dig in on your historical underwriting or what the underwriting model? How does that actually work? What is that process like? Where are you seeing the benefit coming?
Extensively is the answer. You could imagine that these folks who are there, I mean, they're in the business of credit and consumer credit. And they go very deep in their diligence. They kick a lot of tires. Sometimes they'll even, before getting to what you described as a forward flow agreement, which is essentially a commitment to purchase sort of flows of loans in the future, first they'll buy an existing portfolio and just watch it, get sort of acquainted with it. There's a lot of tire kicking on us as a team. And it's one thing to sort of take a point of view on how the credit looks. But when you're making a forward commitment that's on the time horizon that we're talking about, you have to worry about how your counterparty is going to show up when you get into economic stress.
A lot of that is just evaluating the counterparty and just making sure that's who you want to be in a partnership with. That whole process is extensive. I think things along the lines of what Paul described are helpful. Ultimately, the proof needs to be in the pudding in credit. It's not like some of the equity-style investing, particularly at the venture stage, is a bit more of an art. In credit, it's not. In credit, there's deep science. We could talk about what Paul talked about thematically, and it would be interesting to them. They would want to see how that looks in the next cycle. They'll put up the appropriate guardrails between now and then to ensure that they're getting into something that is controlled risk for them.
Maybe switching gears a little bit. One of the changes at Upstart maybe compared to 2022, other than just the credit model changing, is the diversity of products. Talk a little bit about that, the expansion more fully into auto, HELOC, Prime. What's driving that? What are you seeing? Is that from your partners, the funding partners? Is that more driven by them and the demand from them? What made you choose to do those products and why do those?
Yeah, I can start. I think ultimately, our product strategy is just designed around what do borrowers need. The first product we came to market with, the sort of unsecured personal loan, was really the product where it was so sort of blindingly obvious that there was a hole in the market, especially in the sort of non-traditional prime sector where there were many very significantly under-approved and mispriced borrowers. I think that is to some extent true in some of these adjacent markets and motivated our entry. To some of them, we certainly think that there are a lot more people that you could approve for good bank-quality credit in the auto world on both the refi and the purchase side. We think to some extent that's true in HELOC and in some other categories.
That's sort of like the first kind of motivating reason we get into a new product. The second one is that in some of these areas, maybe there is sort of only a limited amount of mispricing of the credit, but there's a significant opportunity for a reduction of the sort of process required to get the credit. I think they're always being the sort of frontier of trade-offs between the risk you can underwrite to and the process that you put the borrower through. Sort of at the limit, if you put someone through enough diligence, even with no AI of any kind, you could get your risk probably to zero. No one would sign up for a process that looks like that.
On the other extreme, even with the smartest AI, if you did not collect a single bit of data about the person, you would have very high losses. It is always about sort of getting to a higher frontier on that trade-off curve. In some products, like in maybe the prime HELOC category, for example, there is more of an opportunity to reduce the sort of process of getting credit, holding risk constant. Finally, the last thing I would say is we have also been increasingly motivated by the recognition that our borrowers are going to be borrowers not just for one product, but over their life cycles for multiple products. That has increasingly played out in a way where we see that.
Now, once someone's in the ecosystem and they've had a really good experience getting an Upstart loan, and when we have more data about them than anybody else has, we have sort of a competitive advantage in underwriting them correctly for all sorts of different credit products. We want to be able to serve them over their whole sort of consumer credit life cycle. That means offering the full coverage of products.
The one thing I'll add, maybe at a higher level here, is that, as Paul said, we've always believed these would be obvious opportunities of expansion from our core modeling technology. Back in 2022-ish, when the world got really stressed for us, at that time, we had sort of four little bets in incubation. We had auto lending, the sort of early instance of home lending, which is a HELOC for us now. We had small business lending at the very sort of small proprietor end and this small dollar short-term duration loan. When stuff got really stressed and many of the tech companies, the value are going through rifts and layoffs, we had a decision to make. It's like, how much do we protect versus sacrifice these things that were taking resources at that time.
We made the call to essentially sacrifice one of them, but protect three of them. The one we sort of put on ice was the small business lending. We protected these other three. We went through the pain of we reduced our fixed cost base. We increased our margins and our take rates. We hunkered down. For the last couple of years, we've been quietly incubating these bets with the idea that one day we would be happy. We would be thankful that we kept them. I think we are getting to the point now, maybe between now and the end of this year, coming into early next, where we're going to really start to see the fruits of that decision. They all look like they're emerging now. They are. They all have great momentum. We're very happy about the trajectories they're on.
The default environment is such that they're now, I think, ready to sort of scale. It didn't come out of nowhere. It's because we've been quietly working on this and sort of tooling it under the hood for the last couple of years. I think it's a very exciting time from the perspective of these new products.
At the risk of asking, which is your favorite child, which of these three is your favorite?
Oh my goodness.
Of the most excited about from just an opportunity standpoint, maybe?
Oh, yay, yay. I mean, we have debates on this at our leadership. I don't know. I'll let you go first, Paul.
I mean, they're really different. They serve different parts of the market. I think for each of us on the leadership team, there's maybe different versions of the sort of credit problem that are the most exciting to us. We've talked a lot about our expansion into the sort of prime audience. I think if you look at it from that and sort of the fact that we can become relevant to sort of essentially the entire spectrum of the US population, certainly, I think something like HELOC is sort of the most complementary to our existing product from that perspective, where we used to be a solution for kind of people who didn't really have too many net assets. HELOC, of course, is the very opposite sort of end of the spectrum. Having great product there is something that we're really excited about.
We really do care about getting past the point where marketing needs to be so highly targeted that it's like, if you're in this particular sort of cross-section of the population, Upstart is the best for you to get into a place where it's just generally true that if you need credit, then Upstart is the best place for you. I think HELOC is a big step for us in that direction. I think at the opposite end of the extreme, the small dollar product that we have is just the one that sort of goes most directly to sort of, in some sense, the sort of original reason we started this company, which was the observation that there are a lot of people who are very underserved from a credit standpoint. Having five-year, $10,000 personal loans is good.
It's not really the sort of true marginal consumer of credit. That consumer needs something like the small dollar product, and being able to get to a place where approval rates can be super, super high and people can just be very confident that if I need credit, I go to Upstart, I'm going to get credit. The small dollar moves us in that direction, which is in some ways opposite to what the HELOC thing does, but is sort of very exciting there.
Auto sort of has its own case for it, which is that I'd say of all the different credit products, I think auto is in some sense the one that is most sort of universally needed and most universally used, where it's like there's only certain types of people that would ever need a HELOC or a certain type of person that would ever need a small dollar or a personal loan. Essentially, no matter who you are, an auto loan is useful for you, whether you're a super prime person or you're not at all a prime person. Pretty much auto is like, it's actually a thing that is useful for your life. It's not sort of as at risk for being kind of just frivolous. It's relevant to everyone across the whole population.
That was a very fulsome answer. I'll just say it in a very summarized way. I think the Small Dollar Product is our most strategic bet. I think the Auto Product is our most audacious bet in what we're trying to do. If we pull it off, what it could be.
We will want to dig into that. We only have five minutes left. I'm going to switch a little bit to the year and now, maybe. One of your guidance for the year is profitability in the back half of the year.
Yeah.
Walk us through what needs to go right, the pieces. How do you get there? What are the pieces that need to change? Are you just on a glide path there now, or is there some risk to that?
It's very simple. There is some risk to it. The core assumptions are basically two. One, the macro, quote unquote, as we experience it through the default risk levels, will roughly stay static. Two, we will continue to execute on the types of model improvements Paul described at roughly our historical clip of improvement. If those two things happen, our fixed cost base is very steady. Our margin profile is resilient. It's really just those two things. In that, there's execution risk. Maybe we won't drive the technology improvements that we historically have or think we will. I like that risk because it's in our control and we're good at that. The other risk is that the macro environment is not static. It could get worse, and that would be a headwind. It's really down to those two things.
Got it. Maybe another question we get a lot from investors on the co-investment model on these forward flow agreements. What kind of guarantees or what kind of risk are you taking at Upstart? Are you guaranteeing a certain level of performance in the forward flow agreements? If the loans underperform, does it fall back to you? How much of this risk is there on versus off balance sheet? What are you comfortable with on that?
Yeah. I mean, it's not really a guarantee per se. We do recognize in the context of these agreements with our counterparties that are meant to span cycles. There will be good vintages and maybe some underperforming vintages because of the types of shocks we've talked about, which our models will react very quickly to, but not instantaneously to. The equation that we need to get right over that duration is that in benign periods, these vintages are meant to overperform. That overperformance, we will harvest. Upstart will, as the counterparty. We will put that in the vault. Where there's underperformance, there's a macro shock and there's a few bad vintages. The overperformance needs to be sufficient to pay for the underperformance. In that sense, it's a bit of a we're creating a bit of a macro insurance layer.
Now, if the underperformance is extreme or it's higher than the amount of overperformance we're able to harvest, some of that we're at risk of. It's on balance sheet. I think we disclose it in our investor materials very clearly. There's a certain aggregate amount that's theoretically at risk. And we have a certain valuation of it. Right now, we value it at slightly above par, if you will. And so that's the variable we manage from a risk management perspective.
Right. From your perspective, is there a target? Is there some kind of ratio, something you're looking at in terms of what that dollar amount is?
Yeah. It's sort of a low to mid single digit % of the overall origination volume that we do through those deals. So just for the sake of argument, take a 5% number. That would be our sort of basis in the risk.
Got it. So we have a minute left in case anyone has a question. Go ahead.
From your guys' perspective, I think as a macro investor, we've seen this disconnect between people's feelings about the economy versus the hard data, the hard versus soft data, right?
Yeah.
Consumer confidence, et cetera. How do you feel that reconciles? Do we basically convert back up, or do we have a kind of a cash down? Or how do you feel about that from a tech and AI perspective versus the macro strength and the traditional markets?
You're asking about the sort of the sentiments about the economy versus the hard data on the economy and how they reconcile? Yeah. Some of this is down to averages versus distribution. Some of this is down to what the sort of the mainstream narrative cares about with respect to macro. What I mean by that is, I mean, the two most salient macro things that are talked about are GDP, which is essentially consumption and production in the economy, and unemployment, obviously. GDP has been remarkably resilient. It's been since the period of the stimulus to today, we've always celebrated the strength of the American consumer. That, I think, is evidence that the recession or the economy in aggregate is strong. Of course, the labor market itself is extremely resilient as well.
That is sort of data that suggests that the economy is strong. I think that if you look at it in a more nuanced way, that the GDP that is sort of materializing and the consumption that is underpinning it, you can ask the question how affordable it is. Normally, we had a certain level of consumption and GDP. We were putting away 7%-9% of our national income. That was a savings rate. Today, we are spending higher amounts than before. GDP has grown. We are barely putting any money away. On average, it is about, I do not know, 2%, 3%, 4%. Again, that is an average. I think if you would look at the distribution of that number, there is a significant fraction of this country that is barely getting through the month, right?
There is an increasing reliance on these cash flow products to get through the month. You have this weird disconnect where we are all spending a lot of money. Arguably, we cannot afford it. Clearly, there is no money for a significant fraction of Americans going into the bank at the end of the month. Are they disgruntled about that? Sure. You could say, why do we not spend less? I do not know the answer to that question. We seem to have had our spending habits turbocharged with the stimulus, and they never got unwound. Maybe some of that is at the heart of the disconnect. We have gotten used to a certain way of living. We cannot really afford it. All the base numbers are good. I do not think people feel like there is much of a safety net around them.
The labor market does remain extremely resilient. Frankly, we do not see that really changing anytime soon. From a structural perspective, like most Western countries, and particularly the Asian ones that are leading the charge on this, we are sort of running out of workers every generation. We reached full employment as an economy back in 2019. It was interesting that nobody ever really talked about it because we then went into the pandemic. I think we have, on the one hand, an extremely resilient labor market. It is not producing for some large percentile of Americans the types of incomes that will both support the consumption habits that have materialized and the savings sort of practices that we have typically been used to and that sort of create security for us.
I think the result of that is you've got a lot of consumption and a little unemployment, but very little financial security and a lot of people who feel like they're going backwards.
Right. So with that, we're at time. I only got through about half the questions. We'll have to have you back here next year to continue. Thank you so much for joining us.
Great. Thank you.