All right, we'll go ahead and get started. Thanks for joining us, everybody. I'm James Fausett, Senior FinTech Research Analyst here at Morgan Stanley. Really appreciative of Sanjay Datta of Upstart being here with us today. Before I get started with him and really kind of put your feet to the fire, I do have a quick disclosure to read. 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 those forward-looking statements. Please refer to the company's filings with the SEC and the IR website for additional information and disclosures. The team at Upstart, I want to make sure I read that. For my own compliance, for important disclosures, please see the Morgan Stanley Research Disclosure website at morganstanley.com/researchdisclose.
Also, the taking of photographs and use of recording devices is not allowed. If you have any questions, please reach out to your Morgan Stanley representative. With the legalese out of the way, maybe, Sanjay, great to have you here at the Morgan Stanley 2025 US Financials Conference. Maybe for those that aren't familiar, and it's hard for me to imagine that's the case, but just in case, could you give us a brief overview of your business? What problems are you trying to solve, and how do you fit into the broader ecosystem?
Sure, yeah. It's great to be here. Thanks for having us. Let's see. Our company is called Upstart. We've been around roughly since 2014 in our current incarnation. We are a platform for consumer credit. We bring together borrowers requiring credit and funding sources seeking yield. The core of what we do is we try to create the technology in the middle that largely has to do with risk modeling. The underwriting of the credit, the modeling of the fraud. I guess from a first principles perspective, we've always believed that the existing space of consumer credit, on the one hand, is the most advanced market in the world. On the other hand, the risk models just empirically are not that good. As a result, a lot of people are left out from having access to credit who are otherwise perfectly good credit.
Many of those who do, a large percentage of their APRS is paying default subsidies for other people who will default they've never met. If you have risk models, the math is pretty compelling for how you can include more people, approve more people, and reduce the default subsidies everyone is paying, and thereby reducing APRS. We had this vision of trying to improve approval rates and reduce APRS through the simple sort of application of modern technology to the prediction models and credit.
Got it. No, that's really helpful. It's interesting because I think the way you fully described it, a lot of times it's shortcut as just AI-based underwriting and that kind of thing. We've heard that's a topic that comes up obviously now with every senior management, et cetera. We'll get more into how you're able to differentiate yourself there both from an underwriting as well as a process standpoint. I want to ask just quickly on recent conditions, what's the trajectory been like through May? Maybe you can help paint a picture for us in terms of what you've been seeing in the recent couple of months of April and May as it relates to origination volume growth and delinquency trends. In particular, those factors trending inline or better than your expectations, what's happening?
Can you talk a little bit about the macro index that you use and how that's been evolving?
Sure, yeah. I guess maybe for both understanding volumes and delinquencies, it's useful to sort of understand the context of the macro environment, or at least specifically what we care about, which is consumer repayment patterns and default patterns. If you sort of take a step back and you look at those more broadly, and you sort of take maybe a reference period of pre-COVID, for many years pre-COVID, the world was kind of stable both in terms of interest rates and default trends. The world got really good in about 2021, and default rates, apples to apples, got really low. Of course, that was largely an outcome of all the stimulus that was flooding the economy, and the money supply got really big, and liquidity was flush.
Starting in about 2022, and basically throughout 2022 and 2023, those default rates went from very good up into the right to very bad. In rough terms, if pre-COVID was like, call it a 1.0, 2021 was maybe a 0.6. So 40% lower default rates than pre-COVID. That number sort of went up into the right and peaked probably somewhere around 1.6. So 60% worse than pre-COVID. It was a really tough two years in which defaultiness, if that's a thing, was just kind of consistently getting worse month over month for the better part of two years. That sort of thing stabilized coming into 2024, was stable for much of 2024. Coming into 2025, it moderated a little bit. It's probably now down to 1.4 to 1.5.
That's sort of like a writ large maybe a view of how consumer default trends have sort of evolved. Maybe separately, if we want to, we can go into why all of that has happened, because it's not obvious at a surface level why the world should have done that in a world where there's been very little unemployment. Nevertheless, it has. That sort of is reflected in origination levels and delinquencies. If you look at origination levels, as default rates, as default risks started going up into the right, originations in the industry plummeted. We were on the front end of that for some reasons that had to do with phasing and how that worked its way through the various segments within the borrower base. I think that contraction ran for the better part of two years, sort of stabilized when default trends peaked.
Now that it's not going up into the right anymore, and in fact, it's normalized a little bit, our originations growth has picked up. That's maybe another piece of the puzzle that's important to understand about our particular business model, because we have a unique growth model in credit. In credit, if your underwriting is relatively static, your growth levers are limited. You can either spend more on marketing to acquire more, or you can loosen your box, if you will, and have looser credit standards, neither of which are particularly palatable. If you can improve your underwriting over time, what that means is you're sort of on the margin doing a better job of avoiding some of the defaulters in the system. Everyone else is paying a lower default subsidy. APRS go down, acceptance rates go up, everything gets better.
That is something that happens consistently for our business. It's been like a consistent pace of model improvement over time. That got overwhelmed by the negative environment that existed. Now that the environment stabilized, we can sort of grow through model improvements again. That is what you've seen from us in the last couple of quarters. Now that we're no longer facing a stiff headwind, every month, every quarter, we have a model that's a little bit smarter, it's a little bit better on the margin at avoiding an incremental defaulter. Everyone's APRS come down, we can underwrite better for the good borrowers, and that creates growth. To get back to your original question, originations were down and sort of troughed for a while. I would say in the last two or three quarters, we're back on a growth trajectory.
We're obviously guiding growth for the rest of this year. The assumption behind that growth is that the macro environment will not get worse. It doesn't need to get better, but it just needs to be sort of stably high. Delinquency trends will reflect that as well. Delinquencies, like our loan cohorts, underperformed a lot when default trends went up into the right. Obviously, we've sort of recalibrated to that new world. Now that that world is stable, I think delinquencies are relatively consistent, they're flat, and they're in line with how we've calibrated the models. Of course, everyone's asking, well, what's happened since Liberation Day? The answer is, in our repayment data, nothing yet, really. None of that has trickled through into real-world behaviors on credit repayment or default.
Got it. Got it. And just as part of that, one of the things that you typically see is, and we talk a lot about this internally with the econ team, et cetera, is that a lot of times when the market's moving around, maybe the lower-income, less wealth-exposed customer, maybe they do not respond, but the higher-income or at least higher-wealth customers will. But you have not seen any of that either.
Sorry, responding to what?
Just responding to changes in market values, stock market, indices levels, et cetera.
Yeah, yeah, nothing really. I mean, in terms of asset values, I mean, the market's gone down, but it's kind of come back up. There's been a lot of uncertainty and volatility, but it's not clear what direction it's all going in.
Right. So that volatility doesn't seem to have created any impact across the borrower base that you have.
Not on the borrower side. You can imagine it creating some funding challenges because those funding sources do not love volatility and uncertainty. On the borrower side, I think the impact to sort of someone's day-to-day life has not necessarily yet materialized from all of this stuff in the news.
Right. Right. Let's talk about on the funding side. To your point, that's usually where I think about there being stress emerging first, especially in volatile market environments. How does that stress manifest itself for you on the funding side, especially in a situation where you guys have been working very hard for a while to have more secure sources of funding and more predictable sources of funding?
Yeah. Traditionally, as markets become more uncertain, spreads get wider, ABS markets become harder to navigate, then the funding sources tend to become more scarce. I think in our particular case, we spent a lot of the last couple of years trying to do deals with counterparties in structures that are sort of designed to survive a cycle. They are meant to harvest some overperformance in benign periods. Certainly, there will be periods or vintages where there are macro surprises, and that leads to underperformance. A lot of these structures are sort of predicated on how performance works over the duration of that cycle. These are with counterparties that are themselves more resilient than what we have worked with in the past. They have got LP bases and funding sources that are a bit more durable. They are as yet untested.
In theory, I think we've got more resiliency on the LP base of these capital providers. We've got contractual commitments that are bidirectional. We've got mechanisms to figure out or to try and make these partnerships whole over the duration of the cycle and not a specific vintage that requires on ABS trading and not with counterparties that are historically maybe a bit more fickle when it comes to redemptions and LP sort of resiliency and things like that. We put a lot of things in place. I guess at some point it'll be tested.
Got it. Got it. Appreciate that. Let's go back to the underlying technology, though, model differentiation. This was a particular focus at the investor day, especially I think the team did a really good job on illustrating some of the improvements you've made with respect to calibration and handling the dynamics of the macro environment. Perhaps explain what those dynamics are first for the audience. What are the things that you track in the macro and the impact on loss variance? I'll just set it up for you so that you can then explain why you feel so confident now that the model is appropriately primed to react to changes such that you can, by extension, be more confident in your ability to deliver stronger credit outcomes.
Yeah. I guess maybe, again, part of the context is a very good machine learning problem is feeding it a lot of information about the borrower and asking your model to take a point of view on the borrower from a relative standpoint. That's always been our bread and butter. A less good machine learning problem is trying to get it to anticipate what's going to happen in the macro, because every macro event is a bit of a sort of unique event. We’ve poured a lot of effort into making our models not predictive of the next macro event, but much more reactive to it.
Got it. Yeah.
Because we collect a lot of alternative data upfront when we originate loans, our models can actually do this in a very nuanced way now. For example, if tomorrow, I do not know, maybe you might have a hypothesis that government employees are suddenly a higher risk now because of DOGE.
DOGE, right, whatever.
I mean, that's probably a level of assumption that, well, maybe that one is obvious, but there are many that are subtle. Because we collect a lot of employment information, we know in our borrower base employment profiles and such. If something were to happen such that that population were to become more risky, imagine they inflect with their default rates, these models will pick it up, and they'll pick it up instantly and react to it. The precision with which we can detect things happening in the macro and the speed at which we can sort of react to them is much greater than it was maybe even in 2022. Those have also given us the tools to sort of not anticipate these things, but maybe manage with conservatism through them.
If the world is in a place where it's defaulting at a 50% higher rate than pre-COVID, these are now tools which we can use to say to the models, "Look, I want you to assume that it's going to be 60% going forward." We can create a bit of conservatism in how we think about the environment, give it some buffer to get worse. Those are the tools by which we manage these committed capital deals, because we do need to create some upside in the deals in benign periods to pay for downside. There are a lot of subtle things in the models themselves now that they're not going to allow us to avoid underperformance in whatever happens next in the macro. I think our models, compared to 2021, will react much more quickly and much more precisely to it.
It would have dramatically limited the underperformance that we did see in those vintages if we had the tools that we do now.
Got it. So that's underwriting and underwriting performance. But another part of the work that you've done that I consistently am impressed by is the level of automation for the loans that you're underwriting. You've talked about automation rates reaching an all-time high of 92% for unsecured loans. How do you see this impacting your ability to serve different borrower segments, especially as you expand into the prime and super prime markets? How important is this?
It's very important. In fact, there's sort of two different sides of the same coin. If you think about risk and friction are related, if you want to lower your risk, you can put a lot of friction into the process. You can ask for a lot of documentation to make sure that you can verify every fact they've given you. On the other hand, if you want to reduce friction by not asking for that documentation, you can introduce more risk. There is always a trade-off frontier. What our modeling allows us to do is push that frontier out. How do you reduce friction and documentation, but not take on increased risk? The answer is by being better at fraud modeling. There are kind of two different manifestations of the same thing.
If you're working with high-risk borrowers, the value is probably much more in reducing price, because default subsidies are very high. If you're already working with very prime borrowers, there's not as much loss or price to take out of the system. You can manifest that goodness as sort of instant process. That's much more valuable to the prime borrowers. You talked about the fact that we are becoming a little bit more aggressive in competing for prime borrowers. Historically, that's not our sweet spot. It's largely a game of cost of capital. The technology contribution we're making is by making it a very frictionless process, because those borrowers that are very prime have a lot of alternatives. The less friction you put in front of them, the more positive the selection lies.
When you think about your relationships with your bank and capital partners and those people that you're originating loans on behalf of, et cetera, how do you think they weight the advantages and benefits of underwriting versus automation? Like you said, it's part of the same coin, but it seems like depending on what they're trying to do, maybe it's just as simple, "Hey, if I want to increase my portfolio exposure to prime and super prime, I like automation more." If I want more subprime, perhaps with a higher return, maybe I weight the model more. I'd love to hear from you.
With banks and credit unions specifically, there is a very specific thing going on, which is we're not selling them assets. What we're doing is we're giving them technology. They are the brand. They are underwriting and originating the loan themselves. Sorry. They're using our technology to underwrite the risk, but they are originating the loan as an entity. That looks like a consumer transaction to them. They care a lot about the experience. I think the financial services family famously has relatively poor NPS scores. When you can give them a process that's very seamless that the consumer loves, that's a big deal to them. They do tend to originate borrowers that are primer, that there's not a lot of pricing to take out of the system.
They want a borrower that's relatively prime with little adverse selection and a very seamless process is really what matters most to them.
Got it. So is that something that you can continue to push, is that seamless experience and the automation? And where do you think you can get to from that 92%?
In our core business, I guess, I mean, if you think about the theoretical limit, it's one minus the fraud rate, right? Fraud rate you measure in basis points. Imagine it's 100 basis points. Your theoretical limit of automation is 99%. You should give 99% of the people the money and put all the friction on that 1% that you're worried about fraud. That would be an issue. Obviously, that's a theoretical limit. Maybe the other way to think about this is 90-some percent of our loans are fully automated. If you look at it on the basis of the applications, it's only about 70%. Why are those numbers different? It's because the ones that are eligible for automation convert three times better. 70% of the applications represent 90% of the loans.
That means there's 30% of the applications that are not converting well because there's so much friction. There's still a fair amount of process we can take out of the system. I mean, I used 100 basis points, but the reality is fraud is, I think, 30 basis points. So 30% of the applications we're not smart enough to automate yet, but only 30 basis points of them are actually lying to us. That suggests there's still a long way to go.
Got it. Got it. On that point of three times the conversion, is that substantially different between prime and non-prime borrowers in terms of the conversion?
That's the conversion?
Yeah.
It's universally similar.
Universally similar.
Yeah. The moment you ask for a pay stub or something, there's a conversion drop-off that's very clear.
Got it. Sticking with the theme around prime, you've been meaningfully shifting your mix, excuse me, towards the prime segment. This is generally what we think about, 720 FICO scores and above. You mentioned that your March originations were up roughly 125% year over year and now are about 32% of your overall origination mix. What's going to get us to 50%? Are we going to go over 50%, especially since those borrowers tend to be able to borrow more? How should we think about that?
I mean, I think it's important to say it's not necessarily an intention we had, right? We're not targeting some mix. We didn't have an ambition to increase the prime mix. We want to be successful everywhere. We just happen to have more success in the prime segment in the last quarter, partially because we're sort of still kind of more getting started in that segment. And we had a lot of success. I mean, I think a lot of the competition or a lot of the competitive dynamic in prime lending is cost of capital. And we had a lot of success working with banks and credit unions to reduce their target returns, and it gave us a boost. That was like an outcome of goodness that we benefited from. It wasn't as a result of us trying to change the mix.
We would love our mix to be representative of America. I think the rough, the sort of back of the envelope is half the country's prime and half is not prime or something like that. That would be a good outcome. We would be equally competitive across the board.
If you're 50/50, but then by exposure, origination volume, doesn't that mean you're going to be substantially more prime or not necessarily?
Yeah. By construction, it would, because loan sizes are larger in the prime segment, for sure. Yeah. That's probably right.
Got it. We've been going for about 25- minutes here. I want to see if there are any questions from the audience.
Oh, there's a question there.
Oh, here in the back. Let's give you a microphone. Then I have to repeat it, and then I'll mess it up.
Are there any constraints? I mean, one of the things with the new administration, Michelle Bowman gave a speech on Friday talking about community banks. She comes from a banking family. Her family actually owns the oldest bank in the state of Kansas. A lot of her speech was about community banks. Are there any things that this administration can do to help, I do not know, to help some of the pain process of growing in this area of consumer credit? Do you feel like people are—do you feel like bankers are constrained at the community segment?
It's a great question. Yeah. Banks are undergoing a lot of evolution, and they have been since 2023 or maybe even slightly earlier when a lot of the turmoil happened and some of the failures happened. I think that, so banks have definitely pulled back as direct lenders in this economy, for certain. I think there's both regulatory scrutiny and very punitive capital charges that are at work. I don't actually have a prima facie view as to whether those are good or bad things. If you wanted banks to step in more as direct lenders, you would have to relax those regulations and those capital sort of charge constraints. Again, there's pros and cons to doing that. We do want deposit-taking institutions in this country to be very safe, and you don't want them to take on too much risk. Yeah.
I could make an argument, and I have in the past, that it's not clear to me why banks have a history as direct lenders and insurance companies that have a similar problem, right? They've got float and they need capital preservation. And they sort of show up as senior lenders. They buy sort of over-collateralized securities, and banks do direct loans. I don't know. I think there's a good argument to be made that maybe banks should increasingly become financiers instead of direct lenders in this economy or something. I don't know. That's me now just riffing about what the different sort of scenarios are. If you wanted banks to be re-engaged as direct lenders, I think you probably have to allow them to take a bit more risk in ways that I think are very hard for them to do today.
Got it. Got it. I wanted to talk about other parts of the offering. One of the things that Upstart has tried to do is be innovative in its offering to the customer, et cetera. Maybe speaking a little bit to what you were saying is that allowing maybe the traditional banks should be financiers and allow you to kind of figure out what customers are going to be most responsive to. You've talked about experimenting with subscription offerings and revolving products. What would that look like in your mind? Do you view them as an opportunity to compensate for the low level of frequency that we tend to see in personal loans? How do you think about those as relationship builders?
Yeah. We've talked a little bit about this in the past, I think in a very maybe still ethereal way. I don't think we have any concrete announcements on that. I think in a larger scale perspective, if you think about the credit landscape for consumers, the two big areas we've not really sort of gone after yet are revolving credit and purchase mortgage. I think those will be interesting areas for us to explore at some point. What we like about revolving credit is what you said, which is there's a much more engaged relationship. Today, most of our products are quite transactional. I think there's a lot of advantage in having a more recurring dialogue, both from an acquisition and sort of engagement standpoint, but also an underwriting standpoint, frankly.
I think those features are such that we'll have thoughts or plans on those segments at some point in the future.
Got it. Appreciate that. I want to talk on a couple of things where we get a lot of questions, especially as it relates to thinking about the macro environment and sensitivity of Upstart generally. In the past, you've talked about 35% or just over a third of your borrower base has a student loan debt. You think that around 2%-3% of this group is in some form non-current. How much confidence do you have on the mix that the mix of non-current won't step higher? If it does, how should we be thinking about the impact to your business?
The mix itself, I guess, is just a function of what we're originating, what's coming into our funnel. We are aware of it. I think the more interesting question is, what is the risk of that segment? Today, as we watch it, we've not seen any real difference between their performance and the control. There's a thesis that it might change with the change of the moratorium. I think there are also scenarios in which it doesn't change. This is an example of something that's like, if something changes in that segment, our models will react to it quite quickly. It's aware of these facts. Every time we underwrite a loan, the model is aware of what their status is on their trade lines and on their student loans in particular.
One question might be, why do not you anticipate it and start pricing them as being riskier starting today? With things that have to do with the macro, and this may be a little bit counterintuitive, but I think we probably have a 50/50 hit rate in guessing. Here is a fact about the student loan moratorium. When it was first enacted, it did not help credit performance at all. People did not get a break on their student loans and go start paying their credit cards. They probably bought an iPhone or something. It did not help credit performance. Now the question is, will it hurt credit performance to take off the moratorium? Is that money going to come from excess consumption, or is it going to come from the installment loans that they are now not going to pay? I think there is a 50/50 there.
By the way, when the moratorium was first ended, which was during the Biden administration, we got the same question then. I was like, why do not you start pricing these guys as being riskier? It turns out the Biden administration said, yeah, we do not care if the moratorium is over. We are not going to enforce it. Trying to guess what is going to happen in the macro is a bit fraught. I think you are probably right as often as you are wrong. To us, the better answer is, unless it is a very large exposure, something on the order of 1%-2%-3% of our book is in the rough bucket of macro things that could go one way or the other. We want to react to them as quickly as possible.
If I were to definitively say today, although that's going to be a riskier population, it's not clear. I could be wrong in the other direction.
It's not the best case, right.
If you're wrong in the other direction, you start getting adversely selected. This is sort of the risk is fraught. If it's something on this level, there's a whole basket of macro things that I think are at any given time being juggled. They sort of fall into this bucket of like, okay, are we going to react to them appropriately as they happen?
Got it. Last couple of minutes here. I want to go back and compare where we're headed to history again. Let's talk about the specific factors that you think we would need to see align for you to return to the types of origination volumes that we saw in 2022. How much of that depends on ongoing model improvements that you can manage and control versus macro conditions?
It's a good question. Both of those are paths. They have different timings. If the macro sort of stays roughly as is, and in particular, if that default index is like 1.5, we will get there over time through model improvements, maybe 2%-3% at a time. We're probably at half of what we were in early 2022. I don't know. It'll take us a year or two to slowly work our way back there. If that macro were to drop, we would get there much more quickly, right? Tomorrow, default risk in the environment subsided for all borrowers. It would be an immediate tailwind.
To be clear, your perception of macro and default risk, or as you're measuring it, better said, is that defaults right now are about 50% above where they were pre-COVID, right?
Yeah. We're still in a very high default.
Very high default environment, et cetera. That's great. Really appreciate you spending time with us today. It's a really interesting conversation. I think that it's something that everybody aspires to. What I find most compelling a lot of times with Upstart is that you're not only taking on the hard challenge of underwriting, but we've heard a lot of banks talk about, oh, we're going to use AI to improve things like customer interaction and automation, et cetera. You clearly are already well along that path. I'm looking forward to seeing how it develops on a go-forward basis. Appreciate it.
All right. Thanks for having me.
Thank you very much.