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

Good afternoon, and welcome to Citi Fintech 11. My name is Pete Christiansen. I'm on Citi's U.S. Payments Processors and IT services team. And this session, we have a fantastic session. We're hosting Upstart, which has been an interesting story certainly in the last year or so. And we have today Dave Girouard, CEO and Co-founder of Upstart; as well as Paul Gu, who's co-founder and SVP of Products and data science. Gentlemen, welcome to Citi's Fintech 11, great to have you.

Thanks, Pete. Great to be here.

So we're going to break this conversation down into 3 parts, I think. First, I'd love to talk, Dave, a bit more about the efforts in expanding the auto-lending business, which is a new endeavor for Upstart. Then maybe we'll pivot to Paul talk a little bit more about the AI machine and some of the underlying mechanics, which may help some investors understand how the model works. And then we'll go back to Dave, and we'll talk a little bit more about product expansion and some of the new marketing initiatives that were talked about on the last call. So -- but before we go there, I think it's a super interesting story in terms of the genesis of Upstart. A lot of its foundations rooted in Google, a lot of employees coming from there. But Dave, I was just wondering if you could just give us a sense of what was that a-ha moment like when I guess, maybe Paul and yourself got together and figured out, "hey, there's a solution here. We can help improve upon offering credit to borrowers who would not otherwise qualify under traditional scoring". What was that moment like?

Yes. It's a good question, Peter. I mean, I think it really started with the very basic intuitive sense that the way credit is issued and who has access to it and on what terms was something inherently, kind of, archaic and how it worked in limiting. And we definitely started from a consumer perspective first. I'd say we still, to this day, think about that as credit is a pretty fundamental ingredient in our life. It's a cornerstone of the economy, yet the systems used to decide who and where it goes and at what price really haven't changed in decades, even while a lot of it was brought online and things of that nature, sort of, Internet enabled, the core of credit really hadn't changed. And so that was, sort of, the motivating factor. And I think we didn't really know until we got started. I mean, I did come from Google and some of the idea was there's some real amazing technology and data science that has been built and is still being built at Google. And could some of these ideas, some of the concepts we brought into this 5,000 year old industry and really improve outcomes. First, for consumers and then as the business developed and we thought more about it, is something that would be better for consumers but also for banks and other types of lenders. And that's the heart of what we got to as you, sort of, fast forward it all today. It's this new area of technology, which collectively is referred to as artificial intelligence. And this very enormous trillion-dollar credit industry. And the intersection of those is really what Upstart is about, it's leveraging AI to improve credit origination.

That's helpful. Now I think, would love to talk about the auto opportunity, auto-lending opportunity. And Upstart has already begun to offer auto loan refinancing and has recently announced the launch of Upstart Auto retail software, which includes AI-enabled financing. From a software perspective, how is Upstart's approach going to differ from some of the more established players in this area? I'm thinking like the dealer tracks and the Route1s of the world. How is your auto software solution different?

Yes, sure. So the auto -- the car buying process is not something that's terribly pleasing to almost anybody. And that means to American consumers, anyone who's bought a car in the last ever has probably had an experience they think could be improved. And dealerships themselves, of course, struggle with it as well. It's just not a really well-constructed process today. The Upstart retail -- auto retail software is really a front-end to the process. Think of it as an iPad application that really, sort of, presents a modern interface, both to the consumer and to the dealer, employees of the dealer to just smooth the process and, sort of, put it into the modern era. Now that's a company we acquired at the beginning of the year called Prodigy, rebranded now and to Upstart auto retail, and that's quickly moving out into dealership. So it's just an industry that has resisted change or hasn't really changed all that much, but it's quickly, sort of, modernizing to what consumers expect when they buy things in 2021. So we like to call it Shopify of the auto dealership market. And this is just getting to the point where Upstart loans could be available through there. So that the real big win for us and the motivation for us, not just to fix the buying process, but suddenly had Upstart's AI-enabled loans available to consumers in ways that will make car buying easier, car loans won't cost as much. And for dealerships, they'll be able to move more cars, make more profit. So sort of a win-win opportunity. But for us, there's just a lot of areas to fix in the auto industry, and we're just getting started there.

It certainly makes sense, dealers are going to be able to move more goods. And that certainly helps them. Maybe you could talk a little bit about how you've been driving adoption across new dealership clients or partners, I guess, you should call them or what you call as your rooftops. How is the initiative going there? Maybe talk a little bit about your acquisition strategy in terms of getting -- on-boarding some of these dealer networks onto the Upstart platform? And then maybe I'll ask one more question on top of that. Do you see the potential to add, I don't know, other online marketplaces or even captive financing networks that are associated with the OEMs?

So first, today, it's kind of like what -- if you're in the, sort of, software-as-a-service type of world, you think of as a small business selling process, driven by e-mail and over the phone and lightweight touch, you try to generate leads, you convert leads. We focus on dealer groups where you can sell to one entity and get several rooftops. So it's working incredibly well. I mean, just a few months back, it was literally people you could count on one hand that were selling this and making good progress. But we've invested a lot. So certainly, it's ramping up. A new -- I've seen a case where somebody new to the team is closing a 5 rooftop deal within a couple of weeks of joining. So it is definitely accelerating, and it's a process that we want to refine. We would like it so that a dealership should get started just by downloading the software, sign the agreement and get started without talking to us whatsoever. That would be the touchless sales process that every small business provider wants to have. And I think we're actually not that far from that, and we'll get there. So that's moving very quickly. To the other part of your question, everybody in the auto lending ecosystem today is potentially a partner of ours. There's really not a competition per se. It's -- whether it's the captives, the OEMs that have their own financing arms, we can help them make better loans to the right people just by having better risk models underlying what they do. So that's important or even some of the banks that are in the industry today, we certainly would hope to work with them so that they can actually have better performance on loans, approve more people. So when you have fundamentally better risk models, there's a lot of wind to go around, and we don't have to necessarily knock anybody out of the industry but just upgrade the different ways that people play in the industry today.

That's helpful. Good color. So the auto opportunity, I think you discussed, it's 6x, 7x more than the personal loan opportunity out there. But generally, how should investors think about how this business should scale maybe over the next 1 to 2 years? And perhaps also, how should we think about the unit economics for Upstart? Maybe the best exercise there is to compare versus your current unit economics in the personal loan space?

Yes. I think any time we're entering a new space -- and this is really our second, is there's a period of R&D and getting the grounding in place. And so it will move at a pace. And then certainly, there's inflection points where the conversion rates start to move up, the model start to get better and that certainly was our history in personal loans, and I would expect it to repeat without question in auto and probably on a more compressed time frame. We're not really starting from scratch, but we are entering a new market. So as you said, it's a very large market. It's probably 6x or 7x larger than the personal loan market. But there's a lot of problems we have to solve. So I think we're just getting in there. We expect to become material for our business in 2020 -- sorry, 2022. And so we're, kind of, planning on that. And I think the teams are making fast progress, both on the refinance product and on the purchase product. I would say our expectation on unit economics over time will be not dissimilar to our personal loan product. We create as much value in the sense of having smarter pricing and a better process. And we don't necessarily want to extract more value than we need to. So I think you can expect over the long haul, that they are larger loan sizes generally than a personal loan, but the percent take maybe less. So it's hard to say in the long run. But generally speaking, I would say, we feel very confident that our value-add in this ecosystem is at least as good as it is in personal lending. So there's every reason to believe the unit economics will be not dissimilar.

That's helpful. And Paul, I'll turn to you and thank you so much for joining. It's great to have you. I think investors have been hearing from a number of outside players out there who are adopting machine learning and AI techniques in credit more broadly. So I was just hoping maybe you can help us understand where Upstart really differentiates here. And maybe one of the first areas we could talk about is, how has the model, I guess, evolved in the last few years from a high level?

Yes, happy to be here and speak to that. So I think the way to really understand what's happened to our model over the last few years actually is also really the answer to this question of how we differentiate and why it will be difficult for somebody else to get started in this area, not impossible, of course, but challenging. And it boils down to, you can think of, kind of, the core fundamentals of the AI system ours or any other as really being 3 things. There's really rows of data, there's columns of data and then there's learning algorithms. Those are kind of the 3 fundamental building blocks of the system. Now you would think that you can kind of just go and start working on one of these things and then work on the next one and then the one after that and eventually, you get there. But one of the very difficult things about going from the thought, "hey, it would be nice to build an AI system" to actually realizing one that does something useful for your business is you actually need to do these 3 things in concert because if you only move 1 or 2 of these blocks, the other one very quickly becomes a limiting factor. So just as an example, if you get tons of training data, but your algorithms are very weak, well you're not going to be able to actually get any value from all the effort you put into collecting that data. Similarly, if you have very powerful algorithms, but you have either very limited columns or very limited rows, the algorithms are not going to tell you anything interesting. In fact, they may give you results that are worse than, kind of, a standard model that is used to dealing with a limited number of columns or a limited number of rows. You're going to [ overstate ] your data, all sorts of [ promised ] results. And I think institutionally, it ends up being difficult for people to manage improvements at a similar pace across all 3 of these, each of which are individually hard to gather, but to do them at the same time in a way where none of them bottlenecks you and you can show proof points back to the mother ship that justifies the very large financial and time investments in getting these. That's why it's hard. Now what have we been doing over the last few years? Essentially, because we had, by, call it, 2015, 2016, really built the foundation for this. Now a lot of the pieces were simple. But essentially, the foundation a few years ago was we had a -- you could think of as a business that was -- a product that was like a human specified rules-based system wrapped around a series of ML models. That was what we had a few years ago. And what's happened in the last few years is we've just iteratively gone to each part of that originally human design system and taking each part that was originally human design and said, "Hey, this thing that really was a rules-based system should be entirely replaced with first a model and then have the standard model be replaced with proprietary AI". So that's basically what we do. We just go to each place where there was an assumption and replace it with a model, and then we take the model and we replace it with proprietary custom-built AI. So as an example of what that means. I think Dave had alluded to this in an earnings call, maybe 1 or 2 quarters ago, where we talked about how it used to be the case that we had multiple ML models that were used in decisioning of loans. But essentially, the results just came out of those models. You've got outputs from those models, and they just got averaged together. That used to be what happened. Now increasingly, what happens is -- and starting a couple of quarters ago, what we started doing is we said, "well, actually, that step where you're taking the outputs of multiple ML models and bringing them together shouldn't be human specified logic, it should actually be learning from the data directly". And so that became this optimized weighted averaging, which is really more of a traditional model. But then that, in turn, was thrown out and said, well, actually, there really should be no reason that humans are specifying the shape of how you're bringing these things together, right? The fact that you're waiting them in this linear or otherwise fashion, that's a human specific assumption that should get thrown out. And in fact, the entire structure of it itself should be another ML model. So now you've got this stack of ML models. We've got multiple models feeding into multiple models, and each layer of the process of what you're doing is itself a proprietary AI built system. And so that's what we've been doing, and we still have a lot more to do in that regard. So I won't pretend that every corner of the code base is run intelligently, but it can be and it should be. And I think we're fundamentally architected in a way where we just start from this core place where we had some ML and we just basically branch outwards to do more and more abstraction of the whole system so that it can make the most intelligent decisions every step of the way.

That's fascinating. So really, I should think about this as a series of machine learning portions, supervised AI, moving to more of an unsupervised environment? Is that how we should characterize it? Just getting more optimized and learning on its own?

In a certain way. Now those terms have technical meaning in the machine learning world that I think don't match perfectly to I think the intuition of what they might mean. So -- but setting the technical pieces aside, I think directionally, what I'm describing is a world where if you think about -- if you start with, hey, we have this one ML model or one AI system. Well, it produces this something. It produces an output, you have to decide what to do with the output. And if all you have is a single isolated model where you get some outputs and then guess what, a person has to decide what to do with the output or a person has to tell a computer what to do with that output. Now we want to replace that segue and say, "well, hey, there's really now a human involved in the process of either deciding or writing code that takes the output deciphers and does something with that. We want to make that step itself into another AI system. And then you -- as you look across how your entire product works, like inevitably every place where there's a ML model produces an output. That output needs to get used. And when it gets used, that's an opportunity to introduce yet another ML system. And that's progressively what we're doing.

That's super helpful. And apologies for being the laymen here. I think a lot of us investors are trying to understand how the mechanics of this works. One of the questions that comes up often, though, is the use of alternative data versus the algorithms and how each one of those plays a role in producing better outcomes. Maybe -- can you help us explain the interaction between those 2 elements of the decisioning model just generally and how they contribute to producing some of these better outcomes?

Yes. So both the algorithms and the data are entirely necessary for the system to get to the place it's ultimately getting to. And this gets back to this idea that if you think of rows, columns and algorithms as the three fundamental building blocks, you need all of them and you need them in reasonable proportions to each other or else just one becomes your limiting factor in a pretty hard way. And so the role of alternative data in our system is that if you just stuck with a limited set of traditional credit data, right? You take someone's credit score and their income, their debt-to-income ratio, maybe like a handful of things like that. And then you tried to throw, say, a neural network at that. Well, it's not going to do very much for you because there just isn't enough information in that information space for you to exploit with the learning algorithm no matter how powerful the algorithm is. And so to really give these more high-powered algorithm something useful to do, you need much more granular data, first of all, so in every category of data, even if it's traditional credit data, you want to have not tens, but really almost arbitrarily flexible data structures that you can exploit. And then you also want to have really orthogonal types of data, data that are very lightly correlated with the data that you already have. And that's why we've got alternative types of data that are things about someone's employment history or their educational history or the way they're interacting with their application. These things bring in what we think of as orthogonal data sources that expand the total information space. And you think of it as depth and breadth of data. You want more kinds of data and in each data, you need really great granularity. And then when you have that data, you can then pair it with really powerful learning algorithms that can actually then -- the two are well matched to each other, and that's when it all works really well.

That's helpful. And the orthogonal reference actually makes a lot of sense, familiar with that. For those who are involved in quantitative finance, short selling, relative trading. That's helpful. As we think about the business is right now, I think, 70% fully automated. Is there a limitation, and maybe this is a question more for Dave. Is there a limitation of how high the automation level can go? And how should investors think about the degree of manual review that's impacting overall throughput?

Yes. I can answer that. The limit is 1 minus the number of true fraudsters or the percent of true fraudsters, right? That's the limit. Because, of course, we don't want to give a fully automated loan to someone who is -- who's a bad actor in any way, whether they're misrepresenting themselves or being somebody -- trying to be something they're not. But the good news about that is, except for a short burst of coordinated attacks, almost all of the time, true fraudsters are a tiny fraction of the overall population of applicants. Most people are trying to do the right thing, trying to legitimately get a loan there. They may be rounding their incomes up a little bit, but they are not fundamentally trying to misrepresent who they are. And so as long as that's the case, that's what we see as the limit, right? And we -- our mission statement is around getting to true risk. And in this area, that's exactly what it means. True risk here are the people who are the true fraudsters. And so you're just always looking at those -- the rates at which you're getting it wrong. And our model gets a lot of people wrong. And we see that as an opportunity. I think it's -- we look at a lot of metrics, and they all tend to tell you the same story, which is that a relatively small minority, I don't need this to be a precise number, but on the order of 1/10 of all the risk that -- of the error that exists in the system has been taken up by all the work we've done over most of the past decade. But the vast majority of it remains to be solved. And as you solve that, you can get that rate way up. And it's worth noting that 70% sounds like, oh, well, how could that be when you're only solving such a small fraction of error? And it's because, remember that the people who get that fully automated experience are converting twice as much as the people who don't. And so there's a -- the rate at which you're getting people the fully automated experience is going to be at the front of the funnel is a lower number than the back, which means as the opportunity is much bigger than it would look like from that 70% number.

Thank you, Paul. That was super, super interesting. Great responses there. Dave, I wanted to pivot back to you last call, you talked about developing products, getting into different areas of the credit spectrum. And if I think about it, the value-add is a better credit decisioning. So how should we think about Upstart's ability to compete in the prime category? And does that impact the value proposition in terms of what contribution margin that you can earn on a high credit -- I'm sorry, high-quality borrower versus a lower quality borrower? And then vice versa, you're looking into some of these lower-quality borrower products. How should we think about the system and its value profit, those book ends of the credit spectrum. ?

Yes. It's a good question, Pete. I mean, I think we feel pretty confident that there's a value proposition everywhere. Even if you're looking at the primus end of the segment to really know who is that prime part and there are defaults there. There's also this process of automation we're discussing where who can most quickly identify that, that is a very super prime applicant that you want to get through the process very quickly. So I would generally say our -- on our journey as a company, we believe the real manifestation of our brand and our success will come when we can make the very best offers to the right people across the entire spectrum and across all the categories of credit that they care about. So that's the journey we're on. The sweet spot of where we've been is right in the middle, right, where it's like -- it's not the super prime world but also not people with super low FICO scores. But right in the middle, that it's mispriced and misunderstood and traditional models have a hard time separating the risk. But ultimately, the model has advantages up and down the spectrum and across flavors of credit. And I think, again, what we want to get to is that Upstart will mean the right price, the best price you can get in an air tight simple process. The best process, of course, is no process whatsoever. It literally is just an attribute you can grab for what you need. So that's the journey we're on. The risk models are what enable all of this to have the best product to bring on more selection in the flavor of more banks and credit unions and other types of providers on the back end. And I think that's ultimately where we're headed is more providers with more diverse sets of capital and what they want to do, better risk models that enable a better consumer product and ultimately developing a trustful relationship with the consumer, so they come back to us again and again. And that's all just part of the journey. I think also having the best product for more of the spectrum, meaning for all Americans, if you will, means that we can market our brand more broadly. Suddenly, we're starting to do streaming television, which is something we wouldn't have considered a while back because often our offers are competitive for more people, and that makes it all start to work economically.

So you increased the top line funnel, you increased your network -- your third-party lending network from a product perspective and you also expand your go-to-market reach that -- a lot of sense there in terms of strategy. I know we have to end a little early. I know you have a hard stop. But -- and thank you both for coming. Maybe we'll finish off with this question. Dave, last call, you also talked about introducing an SMB product. This is a new territory, I would think for Upstart, primarily using consumer data, building the platform, all of that. Now looking at SMB data, where there is [ intrinsic ] competitors there. How do you think about Upstart's ability to differentiate and compete in this new area? And then also ramping up in a new data category as well.

Yes, it's a good question. I think small business lending is certainly a challenging area, and there's been many that have tried and failed in it. There are many that are just in the middle of doing it. We do see it and there's different aspects of small business lending. So we don't pretend to necessarily want to be involved in all of them. But I think there is generally, for small business owners, a need for installment loans at reasonable prices with a super fast, largely automated process, the thing we're good at doing. And the risk models are different. Every small business is quite different. The sources of data are different. But trying to make a product that is really appealing to the business owner. At the same time, works for lenders, for banks and credit unions and others is not a small challenge, but it's exactly the kind of thing, I think, Upstart is good at and has an opportunity to bring a skill set and a DNA to the problem that hasn't been brought to bear yet. So we certainly appreciate the hill we have to climb here. It is different than the consumer loan. But of course, when you get into a smaller business, you start to really -- start to look at the individual who owns the business as well as the business itself. And it's just a compelling problem and one we think we have a unique value proposition to solve.

That's great. Well, gentlemen, Dave. Paul, thank you so much for being a part of Citi Fintech 11. Fascinating conversation, certainly looking forward to following the story further. Thank you, both.

Thanks, Pete. Great.

Thanks.