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

I'm Nat Schindler, I'm very happy to have Sanjay Datta, CFO of Upstart.

You guys all have probably seen Upstart, and I want to jump right into this question. How does it feel being probably one of the most exciting stocks in all of Wall Street?

It depends on the direction of the excitement.

Yes. And so it seems like a flippant question, but actually, it has implications. So we are seeing a lot of Silicon Valley companies that have been affected. Your stock in 18 months that you've been public went from $20 to $400 to, whatever, $45-ish, which is actually a very good performing IPO right now, in that cohort $20 to $45. That's pretty good. But one of the things that Silicon Valley companies have been noticing when you have these really fast movements in stocks across all of them is the impact on hiring and holding and retention. So how is this impacting you? And I'm just fascinated to know what's been going on in the Valley as prices of stocks have moved?

You mean appraisingly?

Well, more in the area of hiring.

Yes, it's a hot topic. A lot of the currency in the Valley is in equity. And I don't know, I understand, on the one hand, you sign up for a risk contract when you join a technology company in some regard and you enjoy the upside. So the CFO in me is like, "Well, you should just live with the downside as well." But if you think about people's rationale, incentives right now, if you have a certain market value and some of that was paid in equity and that equity disappeared or significantly reduced, you can walk across the street and that will get reset. And so I think the thing that is on everyone's mind right now is retention, unsurprisingly. And a lot of companies are trying to figure out how to reset internally so that they don't have to suffer a lot of attrition. But I think a lot of companies also have announced layoffs, so we're thinking about workforce reduction anyway. So there's just a broader equation.

So there's one company that hasn't announced layoffs and has a wee bit of cash on its balance sheet. That happens to be your former employer who's kind of famous for just scalping everybody in the Valley with cash offers. Have you seen this yet occur? And how do you combat it?

Well, not really significantly. And there's a lot of variables in this equation as well. I don't want to get off track, but I see you're referring to Google.

Yes, sorry, I didn't mention. Yes, it was at Google.

I think the average person who is working with us or in fintech is not at Google or was not at Google for a reason. It may be that some people were interested in building and creating upside and now have a flight to safety, but I don't think that's predominantly happening. I think most of the people, certainly in our business, continue to remain as convicted as ever about the opportunity and the upside. And so we're going through a bit of a challenging time with the market. But I don't think anyone views that as reflective of the fundamentals of our business. I know everyone can say that, but I actually do believe that. And then there's the whole dimension of what bigger companies versus smaller companies are, how they're approaching the return to work. And I think that a lot of people that have gotten used to distributed working are not eager to go back to sort of 3 days a week in the office, which is what Google is mandating. So I think between how companies are coming back to the office post pandemic and how people are seeing the relative sort of trade-off of security and cash versus upside, nothing's changed substantially, to be honest. We haven't seen a lot of retention or attrition yet.

Okay. So I'll move off from that segue and go back to real Upstart as a business. So you guys are a 2-sided marketplace with probably both sides being some of the most talked about macroeconomic movers. So can you walk through -- I know this could be a big question for a little bit of time, but it's very relevant -- for the last 2 years, the weirdness and then as we're moving forward and what you're seeing on really both sides, the funding sources and the consumers? And what does it mean to your marketplace?

Over the last couple of years, yes, I mean, I guess, in sort of normal pre-pandemic times, throughout our history, the majority of the time -- so we're an ecosystem that brings together prospective borrowers and lending capital, either from banks or from institutional markets. Almost the entirety of our history, we've been borrower constrained. So there's been funding availability at some sort of market rate, and our growth model has been about finding more borrowers and converting them economically. So it's the classic sort of consumer growth marketing hacking kind of activity. There's now been 3 occasions in our history where that equation is reversed and we've been funding constrained. The first one was back in 2016 when we were still somewhat early as an industry and a bunch of things happened that I won't belabor. The second time was basically in March of 2020 when the pandemic first hit and everybody was extremely nervous about the direction of unemployment and the economy. That pretty quickly reversed in mid-2020 for us and mainly to do with the stimulus that was injected in the economy. And so if you think about the equation, it's the balance of borrowers and funding. I think historically, it's been borrower constrained. Funding availability essentially went to 0 for a couple of months post pandemic. It then very quickly reverted. And what happened was the funding availability became tremendous, and it's because there was an extra $5 trillion of stimulus in the economy. And the borrower demand itself was very suppressed, and it's because there is very little credit card debt to refi and everybody had pretty flush balance sheets and personal savings accounts. So I'd say for the past 12 to 18 months, there has been a significant excess capital and a real dearth of borrower demand. And that has once again been flipped on its head. And so we're now in a world where demand on the borrower side is going up pretty significantly. I think a lot of the stimulus has now been flushed out of the economy or even spent to where it's been put in the equity markets and thus disappeared. And their credit card balances are going up. Their expenses are going up with inflation. And so demand for credit on the borrower side has gone up. And of course, the funding availability has gotten quite skittish just with the uncertainty around inflation and the potential for recession and war, et cetera. So we're once again, for the third time in our history, in a world where we are funding constrained versus borrower constrained. I don't know if that answered your question.

No, that is a good way to look at it. And can you break that down into what that actually meant in how the world worked, in something that the funders would really care about default rates; what you saw over the last pre-pandemic, pandemic and now coming out. And where are we sitting now? And how are you planning for the future?

Sure, yes. Default rates, if I would just kind of stylistically describe them, pre-pandemic, they were relatively stable. And I think that our lending history, pre-pandemic, which I guess you could think of as 2014 to 2019 roughly, on the balance, was a benign period in the broader history of credit. So loss rates were pretty stable, pretty good as well. Loss rates in the wake of the pandemic, so in early 2020, when the stimulus was introduced, they pretty quickly went to half of their historical level. So again, the numerical example I'd give is if 10% loss rate for a given borrower kind of long term -- that is a made-up number but imagine it was 10% annualized loss rate -- I think pre-pandemic was probably around 9%, so 2019, and then before it was a benign period. Those numbers went to 5% pretty quickly once the stimulus showed up the economy and stayed there for the better part of 1.25 years. So they are historically low. In terms of historical magnitude, that's a very large magnitude change, obviously. Starting in October and November, those things reverted pretty quickly and went back to 10-ish, maybe a little bit above 10%. And that rapid increase probably happened between October and February. And since February, they've been stable roughly around 10%, okay? So it's sort of been hovering, February. March, April and May has been sort of stable again. So that's the direction of things. I think that investors don't just react to those actuals. They also have anxiety about what they will look like in the future. And the thing that may obviously change in the future is if there's a significant change in the employment levels.

So any change in the inflation patterns have not had a direct effect beyond what it was prior to the pandemic. So the only real worry that you think to really change default rates from here, the major cause, would be just employment changes?

Yes. And I can sort of double-click on that a little bit. We've obviously never seen an inflation at this level for decades, right? So there's no digital lending history we can proxy. But you can create retros to proxy this as changes over time in expense levels between borrowers. And you could sort of ask the question like, from a cash flow modeling perspective, if someone's expense base has gone up by 10% or even 15% or whatever, what does that do to probability of default. And the answer is there's not a very significant, meaningful relationship there. Like a 10% shock to your expense patterns in the context of your propensity to repay is not a huge variable. Now if inflation tips us into a recession and unemployment goes up, there's a very strong relationship between employability and propensity to repay. It's almost a one-to-one relationship. And so that's certainly what we're watching for. Now is there a world in which inflation somehow maybe there's a disproportionate increase in required expenses for a segment of the population such that the change in their expense base is much higher and maybe it has some sort of disproportionate impact on defaults, it's possible. We are in somewhat unprecedented times at least in the last 50 years, but it hasn't shown up in our data yet. And I don't think we have a theoretical basis for worrying about it right now.

Well, so that unprecedented time thing is very fascinating because you are an AI-based lender. And AIs are giant spreadsheets, in a sense, that find data correlations over time, as you have described it in the past. And going back, you said from 2014 to 2019, you were in basically a very benign credit environment. In 2020 through '21, you were in a super benign credit environment. And now you're in, call it, a normal but with a worry that -- well, who knows what we are in the future? But how does an AI react and how quickly can it react if it has never seen?

That's a good question. So maybe first with just at least teasing apart, so in any prediction of whether someone is going to default on a loan or not, there's 2 implicit components. One is the borrowers' fundamental propensity to repay based on their characteristics and their affordability. So in some sense, if you just abstracted away from the macro, you could look at a group of borrowers and just rank them, right, in terms of relative risk. And then there's a macro thing that happens to all of them, maybe somewhat disproportionately, but right now, the macro is affecting all delinquency rates directionally in the same way. And so the unprecedented thing that's happening right now is the macro thing, right? What we're good at, what we claim to be good at, is the first one, which is the relative ranking and separation of risk. If you take a FICO bucket, the spread of risk within that FICO bucket can be like 10x. There's a dramatic spread of risk within something that, from a traditional lens, will look the same. And what we're good at is finding the good and avoiding the bad within that. And our ability to do that doesn't change in any way in a dislocation like this. And you made the claim that we've been benign since 2014. That's almost true. It's not strictly true. There was a period of a couple of months in early 2020 when the pandemic hit, the unemployment went from 3% to 14% and stimulus did not show up. And when it did, everything was fine. But there was a period of time where hardship and forbearance and delinquency went off the charts. And so we were able to sort of at least understand through that stress and through this one. Has our ability to rank and separate relative borrower risk changed? Even though the entire grid has changed because of the macro. And the answer is generally no. In fact, if anything, it gets better because if you just think about it like, if you're going through a macro shock and you're trying to relatively rank borrower risk, would you rather know their FICO score, one thing, or a lot of stuff about where they work and what they studied and what their function in the industry are like? You'd rather have a richer data set than a more limited one. And it turns out, that's the more impactful exercise. Like if you look at historical recessions and what they've done to default rates, if you go back to '08 and you look at what happened to credit card defaults, they roughly went up by 2x, like 100%, and loss rates doubled. If you take, as I said, a FICO bucket and you actually look at the actual performance of risk across the distribution, there's sort of like a 10x spread there. So you're creating much more alpha by figuring out what are the good and bad relative borrowers regardless of the macro environment than you are by trying to time the macro cycle. Now macro is important and it's going to affect default rates and returns and we will sort of go through that, but I guess the very long answer that I'm providing to your question is, our AI models, yes, they're untested through macro cycles, but they're not meant to predict macro things. And in a sense, you could think of this as we try to tell investors like, "We will rank the borrowers for you. We'll show you the good ones, and we'll help you avoid the bad ones. You take your point of view on what the macro holds. And if you think there's a chance of a recession or a dislocation, ask for higher returns. And you can do that. Instead of a 7%, you can ask for a 12%," or something like that. And so that's how the dynamic works. We try to do the job of selecting relative risk. And the lenders are the ones who are, in some cases, better positioned to take a view on the macro risk in the future than our AI models because a macro event, it's a very poor use case for an AI model. As you said, there's no historical data.

Yes. So it can't predict a black swan, and it shouldn't.

Exactly. And each black swan is different, right, I guess by definition.

Yes. Exactly.

But the next recession is not going to look like this one, and it certainly doesn't look like the mortgage prices of '08. Like they're all going to behave differently. So our general philosophy is that our models do the job of ranking the risk, which is frankly the more important one over time. And then if you are worried about a recession in 6 months or 12 or 18, then you should figure out what returns you need in order to accommodate for that.

Well, that actually brings up a really an interesting question. So if your real skill set is ranking borrower 1 through 10 in the relative order, you're basically a piece of software that creates credit models. Why be the lending marketplace itself? Why not be a credit model as a service? Why not be the underlying part that funds these funding sources that have 0 cost of capital or very, very low like, I don't know, BofA?

So you're asking why our platform, like why don't we just provide the APIs?

Yes. Your real skill set is being able to rank those things that people can't rank. And everybody else is going out there and finding funding, that seems a complicated way to make money.

So we do have an aspect of our business which is essentially describing a credit API. We could go to your employer and say, "Hey, look, I've got this great API. Plug in the data, and we'll price it for you." And we do, do that. It's available. So if you think about it, well, the real difference is in areas where we're the actual platform, where we are at the point of distribution, the real value of that is that we can control what questions are asked for, right? So if all we did was give Bank of America our APIs but didn't change their actual lending process, you would basically pull their credit file. And our APIs would be limited. What we would like to do and what we would, do in an example like that with Bank of America, is we'd say, "Let's take over your lending program." If someone goes to the Bank of America website, they would see a Bank of America-branded version of our sort of application process. But in that application process, we would seek to understand what they studied and where they went to, what industry they're in and what their title is. And we would observe things about how they interacted with our application. We would say, "Okay, will you tell us what your FICO score is?" And it turns out people who know what their FICO score is, that's a signal. Like they're paying attention. And there's other people who don't know when we check, and there's like, "How much do you ask for?" If someone asks for $10,000 versus someone who asks for like $9,830, there's like a signal in that. And these are subtle things, but they're actually very meaningful. They're much more meaningful than the classic. Like people think, "Oh, do you scrape social media?" Like we don't find any signal in that. But like how people actually respond to and what they know about their own financial situation is very meaningful. And if we didn't have access to any of that information, our credit models, our APIs, would ingest. And our models would be probably more than half as powerful but certainly not 100% as powerful as the world where we control the application process. So the way to think about this from a business model perspective is, where we can be at the point of distribution, either by getting the traffic to upstart.com or white labeling it in a bank partner or by being at the point of sale in the dealership for a car sale, that's always best because we control the collection of data. In areas where we think we'll probably never be at the point of sale or at the point of distribution, we will expose our APIs. And lenders will do better with those APIs than with the traditional credit model. But they want to unleash the full power of AI because they don't have the alternative data sets that we like.

Fascinating. I kind of want to know, is someone who knows their credit score a better credit risk or worse?

Better.

Okay, just wanted to know because I don't know my credit score all the time. I kind of ignore it.

Yes. I mean the answer changes with where you are in the spectrum, of course. Like some folks like yourself may not need to care. But there's folks who have to look at this constantly and worry about it. And folks who are versus are not generally have a different repayment profile.

That makes sense. Okay. Going into other questions. So every question, probably 90% of the questions I get about your company focus on only one side of your marketplace right now, the funding sources. Can you walk through the funding sources you have and kind of their current behavior patterns?

Sure. Again, it's not a discrete spectrum. Everyone is a little bit different. But if I had to sort of characterize it, on the one hand of the spectrum, you have banks that are using our technology to lend under their own sort of licenses and charters for their own balance sheet. So in the example you gave, we're not working with Bank of America but we could sort of provide our software to Bank of America. They can use that to lend to their customer base and then also take referrals from upstart.com, and they would be lending to the consumer, to the borrower, that would look like a Bank of America loan. It would be branded Bank of America, and they would get a Bank of America loan. And so that's some amount, maybe call it 1/4 of our platform is in sort of what we call our lending bank partner or a lending partner referral network. So that's one. So that tends to be a balance sheet source of capital. It tends to be hold to maturity. It tends to be very low-cost cost of capital because they're typically using deposits. And the risk aperture of that capital tends to be on the primary side because banks have regulatory reasons for not wanting to have loss rates, which are too high. It becomes very punitive. So that's one source of capital. And then if it's not being balance sheeted by a bank, in general, it's being sort of absorbed by the institutional world. And institutions could mean anything from pension funds, investment banks, Act 40 funds, hedge funds, credit funds. And there's a broad spectrum of buyers in there. Some of those are similarly hold to maturity type buyers that don't use a lot of leverage. On the other end of the spectrum, you have folks who lever up a lot and rely on liquidity from the ABS markets. And so if I were to sort of describe the resiliency of the capital, the banks have the most stable and the cheapest capital. In the institutional world, anyone who's holding the maturity and not using a lot of leverage tends to be not quite to the level of the bank but sort of relatively cheap capital and relatively resilient. Anyone on the other end of the spectrum, that's more like you got a brand of a trading fund that is buying, trying to sell for a gain in the ABS markets, looking for liquidity, does not have a huge capacity for holding large sums to maturity. They tend to be easier capital to get when you're growing, and it's less resilient capital when there's sort of uncertainty in the macro.

And if you are to just rank, so about 1/4 of your capital usually is from the bank partner's resilient capital, 3/4 from these various institutional money. And how would you distinguish between the kind of faster money versus the hold to loan in dollar amount?

Of the 75, probably half. Half and half roughly. This is more fluid. I think the bank percentage steadily grows. The comings and goings of institutional money between people who were holding unlevered and trading is on some level of function of the situation in the markets, right? Like in 2021, as an example, you could buy a loan at par from us and immediately flip it for 108 cents on the dollar in the ABS markets. That's how constructive they were. Pension funds were like, "I'm taking that deal." Like even folks who want to hold yield to maturity were like, "Let's take that deal and recycle the cash." Now those folks are not doing it because you can't get that kind of a gain in the ABS market. So it's somewhat fluid depending on the state of the markets. But I think if I had to like roughly characterize the entities themselves, it would probably be like half and half. Or at least that's what it was maybe a few months ago. Now anybody who relies in the ABS markets is a lot more fickle than they were.

Yes. So that's interesting. So as the world evolves and fear and uncertainty and doubt gets bigger, obviously, of a spectrum starting with banks over here and fast credit hedge funds over here, those guys are affected the most, right? So you're funding it.

Yes.

But those guys are the ones who came on when you grew outrageously during the pandemic, I mean hundreds and hundreds of percent a year. In one quarter, I think you had quadruple-digit growth. That was a new one for me, and I'm an Internet guy. These guys are relatively stable. So it must have been coming mostly from this side, this faster money credit money, right?

Not quite. I mean we've been adding banks that are pretty -- I mean, if you remember, we IPO-ed with 8 partner banks. We've gotten north of 50 now. So that's been growing. Like it's been keeping up. Banks have been keeping up to growing as a percentage of our platform even when it was quadrupling or whatever. So yes, I mean a lot of the -- when you have some growth, when you grow quickly, it certainly is the case that it's easiest to get the most expensive capital into the system. But then you gradually replace it as you grow. And so banks did absolutely keep up as a percentage for a platform even through that growth because we were both bringing new ones on. And the ones we had, they tend to start more tepidly. And then they see the credit for 6 months, and then they start to grow even within their own accounts.

Okay. So well, we have 8 minutes left, and I'll open it up to questions. We have a lot in here. These guys are a fascinating company. And there's a lot of nuance here to jump on. So raise your hands at any time. So walk me through a little bit, as the macroeconomic changes -- or even if it doesn't, if the funding source is changing, so the cost of capital is changing, that affects lots of different things. And also on the other side, the consumer is getting easier to fund but default rates are going up, so you have a lot of moving things coming in here. Walk me through the impact to you, for investors and yield for you and then also cost of customer acquisition and how that changes -- cost of, I guess, borrower acquisition -- I want to just make sure I'm clear on words -- and approval rates and how it relates.

Yes. Sure. On the first point, which is the returns to the investor and even, I guess, downfield from that is the coupon to the borrower. So 2 things are happening to the borrowers' coupons. The borrower's coupon is the sum of the sort of net return to the investor after losses and the loss rates themselves. So if you're defaulting at a 10% annualized rate and the investor providing the loan wants a 7% return, your coupon is 17%. Both those things are going up right now. Loss rates are going up because the model is reacting to the trends in default rates as they evolve. That hasn't changed, but your propensity to default is changing based on what's going on in the macro. And so the models are reacting to that change in the environment. So maybe when the stimulus was at its peak in the economy, the model would say, well, that has a 7% chance of defaulting based on all the things I can tell. Now that's become 9% or 10% because suddenly, you had a huge equity account that is worth less. And maybe on the balance of being employed to Bank of America is more sort of risky than it was, I don't think that's true, but like hypothetically, so it's changing the estimates of loss rates. And then on top of that, the investors are saying like, "7% is not good enough anymore. I need a 10% because I'm worried about a recession." So it's like a double impact on the coupons, which serves to reduce -- it sort of serves to constrain our conversion funnel, right? That makes it less likely that the Board will take a loan essentially. And it makes us less likely that we will approve it because we can only approve up to a certain APR cap. And if the increase in your coupons has put you over that cap, we're not going to approve you anymore, okay? So there's been an increase to loss rates. There have been an increase to returns, which is your question. Investors are absolutely saying like, "I need more now. I need more because I'm worried about a recession. I need more because I rely on the ABS markets and they are not as constructive anymore, so I need more spread in order to make that trade work," or whatever, banks are saying like credit committee is being more conservative, so that just pushes out. Great. So that's unsurprising because it sort of mirrors what's going on with the fundamental interest rate curves in the economy. Now when you say what does that do to our yield, I wouldn't call -- our business model doesn't really have any yield in it. It has a take rate. So the way our business model work is we transact the loan, the borrower gets funded by the funding source and we just take a transaction. It's not in any way related to the sort of upside or the downside of the credit. When those loans are being sold at $1.08 into the ABS markets, we didn't get any of that. And so we just take a transaction. It's a take rate. It's a take rate we charge to the banks in order to transact the loans. So they're our customer in the business model. And then the question is, well, does any of this impact that. And the answer is not in any exogenous way. We can charge whatever price we want and the banks can choose to sort of pay that or not. Ultimately, they pass it on to the consumer. So it's a question of whether the APR will work for the consumer or not. But I will say that, in a more specific sense, our take rates are generally under-optimized to our P&L, meaning they tend to be, even in good times, just abstract away from the current environment. It was the case historically that we could have raised our take rates. It would have lowered origination volumes because banks would have passed that on to borrowers. But it would have increased our revenue and our profitability in calendar P&L. And the reason we had under-optimized take rates is because there is a value to us of the origination volume even at lower revenue because it accelerates the learning of the models. So more volume flowing through the models makes the models learn faster. And in addition, there's also a lifetime value to that loan, which is there is some probability that we'll take a second loan 2 years from now. And so we were, in a sense, investing forward in model learning and lifetime value of the customer in a way that in the sort of current environment of macro strife that you're describing, we're typically incented to raise those take rates. So what you'll generally see, what you saw in March of 2020, when funding became very scarce, is that our take rates and our contribution margins went up. And that's just us shoring up the P&L as we go through a rocky macro environment.

Now if funding sources dried up, you just say we'll have a higher take rate and your P&L is fine. Well, going to cost of customer acquisition, that's related to approval rates, which is related to coupon rate and potentially default rates. But it's also related to competitive environment and who's out there buying the leads and also what they're charging. So can you walk us through what the competitors are doing and how they've been reacting to the changes in the macro and stimulus coming off?

I mean the short answer is I don't know. I think we only see lagging indicators from the broader industry. So I hear things anecdotally, probably nothing that I can really speak to with precision, other than to say, I think that you generally have 2 models emerging in a more general sense. One is maybe sort of more where you are a marketplace like us and one is where you're using balance sheet capital. And I guess I would say, in an environment like this, a marketplace is subject to the market forces. If the required rates of return go up from funding sources, you have to react to that and price accordingly. If you're using your own balance sheet, you don't have to do that. And so in some sense, you have a bit more degree of freedom if it's your own balance sheet. Whether it's the right economic decision to maybe not follow market prices in how you use your own balance sheet or not, it's not really a question we face so I don't think a lot about it. But that's the one dynamic I'd point to. Different players in the industry held different standards of whether or not they need to follow the market interest rates or not.

All right. And so just going into the last quarter, historically, you've taken a very small amount on to your own balance sheet, and you took a slightly higher amount on to your balance sheet. Is that something you see? Does it makes sense to stabilize the market when the funding sources are under FUD? Or do you just say the market's the market?

Well, I'll be honest with you, in Q1, I think we had an idea. It was probably a good use of our balance sheet because it's just a random cross-section of the loans that are being priced at the newer market rate. We weren't using it to somehow underprice loans or anything like that. We were once again in a situation where the equilibrium equation on the platform is that we had more borrowers than funding dollars even at the sort of new market rates. And we thought it would be a good use of our balance sheet to sort of absorb some of that excess forward demand. And then you can always traffic it through secondary markets or whatever. So I think we had that maybe point of view in Q1. Obviously, the market reacted very unhappily to that. And so it's not something I've ever carried strongly enough to sort of hold out on. I think the market really wants to view us as a marketplace that reacts to the vagaries of the market and accepts volume volatility as a result, which is how we largely view ourselves, then I think that the takeaway for us is that we'll just keep our balance sheet out of it. And we'll really sort of reserve it for the pure purpose of R&D, which is sort of the core reason we've raised the money.

Makes sense. We don't have any time, but I really want to ask one more thing. So the CFPB has recently come out with a pretty strong statement against -- or not against but their intention to look at how black box AI lenders work and who they're giving loans to and not. How is this going to affect you? And I mean their question is they want to know why any individual loan is denied. I could understand why the CFPB thinks that way. But on the whole concept of an AI system, then it seems almost impossible. So how do you do that? How do you react to what they're doing? And what are they really doing?

It's a good question. The CFPB is obviously a hot topic right now. You're talking about a topic that's roughly described to as explainability. Insofar as we worry about regulatory, it's actually not explainability that we worry about. Actually, we strongly agree with the CFPB and other sort of AI thought leaders out there that you can't just have a black box model that just fits stuff out. So I disagree with you that explainability is impossible. It's actually not. Almost half our machine learning team spends their time building models, which explain our core model. So you can take our core model and you can run it through different iterations of changing smaller variables, and it's very computationally expensive. But there are ways of creating adverse action notices that have explainability in them that roughly explain what's going on and why. And in fact, we have pioneered this with the CFPB, right? I think for as much as everyone talks about AI in lending, we're the only ones who I think are deeply embedded with the CFPB and trying to figure out what to do about it. And I'm not talking about the political class of the CFPB, I'm talking about the professional class. We've been working with them through the Obama administration, the Trump administration and now the Biden administration. They've been there throughout. And the policy and the science of how to create explainability through AI and AI model, we believe is extremely important. It's extremely hard, and we work very hard at it. And we think it's doable. And I think the CFPB agrees with us. So explainability is something I actually think, if anyone wants to come down this road and start to actually in-depth apply machine learning and AI models to the problem of credit production, they're going to have to do all of this. And we think it's important, and it's not trivial. There's a whole separate topic with the CFPB, which is around fair lending and adverse disparate impact of protected classes of borrowers. And that's a whole other discussion which is in part fact-based and in part political as most discussions are. And I think that's probably where more of our focus is right now, in working with the CFPB, trying to sort of figure out policy. There's a lot of gray area in terms of what is acceptable disparate impact, what the right trade-off is between impact and accuracy and things like that. So it's a bit of a different topic than explainability. But to us, it's maybe the more relevant one at this point.

Makes sense. So one of the things you just said is that you literally have a team of AI scientists who are psychologists for AI.

I think we probably have as much machine learning resource dedicated to the CFPB than we did Upstart.

Yes. Sorry, the explainability.

So yes, they're our psychologists. I'm just trying to figure out...

Just thought it was fascinating.

[ Now you're going to have to be that people ].

Constantly do. Sorry, everyone, we're way over time.

Thanks, Nat.