UPST Stock: Upstart's Co-Founder on 92% Loan Automation
Summary
Company Overview: Upstart is an AI-driven lending company focused on providing consumers with better rates and processes across various credit products, leveraging AI to automate 92% of loans.
Market Opportunity: The company addresses a significant market gap where 80% of Americans repay loans on time, yet only 50% are considered prime credit, aiming to improve access to bank-quality credit for more consumers.
AI Advancements: Upstart continuously invests in enhancing its AI models, which include deeper and wider data integration, to improve credit decision accuracy, reduce APRs, and increase loan approval rates.
Fraud Prevention: The company employs AI to detect and manage fraud, maintaining a small percentage of manual loan reviews to handle potential fraud cases effectively.
Expansion Plans: Upstart is expanding its product suite to include more significant markets like auto and home loans, with a focus on automating processes and improving consumer access to credit.
Competitive Advantage: Upstart's competitive edge lies in its advanced AI capabilities and data-driven approach, which allow for better separation of creditworthy borrowers and more efficient lending processes.
Future Outlook: The company aims to transform the lending landscape by integrating AI into larger credit markets, potentially increasing consumer access to credit and reshaping industry standards.
Key Challenges: Upstart faces challenges in scaling its operations across different lending markets, requiring deep industry knowledge and adaptation to specific market dynamics.
Transcript
Yeah. So, Upstart is an AI lending company at heart. Our mission is to offer consumers the best rates and best process uh for every category of consumer product. Fundamentally, something like 80ish% of Americans actually repay their loans on time with no problems. But only about 50% of Americans are considered prime credit based on their credit scores and traditional credit reports. It's like, you know, you want credit when you need it. No matter what you think about AI, whether you're an enthusiast or not enthusiast, I think you're just going to have to look at the numbers and, you know, come come to terms with reality that this is the new way lending is done. >> Hello everyone. Today we're joined by Paul Goo, co-founder and CTO of Upstart, the AI lending marketplace behind more than 100 bank and credit union partners. Upstart just posted tripledigit revenue growth and uh return to profitability in Q2 with 92% of loans fully automated and conversion up to approximately 24%. Paul, welcome to the show. >> Thanks, Ed. Great to be here. >> Um are you are you calling from your HQ at the moment? Where where's the is this the background? >> No, no, this is my home office. Um built built for the digital world. Uh >> whereabouts are you based? America. >> Uh yeah, I spend most of my time right outside Salt Lake in the US. >> Okay. Oh, nice. That's cool. And uh the team spread out over the country or >> Yeah, we've got people in almost all 50 states. Um really over the course of COVID, we discovered that you could get incredible talent um by hiring all over. And I think for companies that hit their growth spurt around 2021, 2022, um you just ended up with uh with a great spread of talent everywhere. >> Nice. So you've you've maintained that going forward. This is your structure for how the company operates basically. >> Yeah, we we have a working model that we call digital first where uh we do hire people everywhere across the country and we think the best talent can be all over the place. But we also really believe in in-person work. So we bring teams together on an almost once a month basis to spend time together in person. So people do launches together, you know, they can do kind of like intense collaborations or just spending time getting to know your your co-workers. doing that's really important and then the math actually pencils if you think about the there's travel cost versus office cost and especially you know if the office cost was going to be in a really expensive place um the math actually works just great. >> Yeah. Okay. And so when they come to meet up do you hire like an event spaces because you don't have an office or >> Yeah. So we actually do have some office spaces that pretty regularly because teams are always gathering on a regular cadence. The offices are always being used um but it's it's not like every team at the same time. So, we've got offices sort of in a few places across the country. Um, but yeah, we also do event spaces for larger gatherings. Um, and you know, once a year we bring the whole company together and then a couple times a year, you know, big chunks of the company. >> Amazing. And can you just touch on um your background as well because I I believe Yale and then pretty much straight into Upstart. >> Yeah, that's right. Um, I don't have a terribly long life story to share before Upstart. Um, but yeah, I I was at Yale. I was doing a double in computer science and uh in econ and um I dropped out to do Peter Teal's 2020 program uh the so-called teal fellowship. Um so the first year of that program and um >> so I dropped out I spent a little bit of time um at this quant fund called DE Shaw and then um pretty quickly pivoted to thinking about um problems really at the intersection of uh of machine learning and problems that would have an impact on uh on real people. Uh, you know, I always felt like there was just an incredible amount of uh brain power and compute power um at uh at Dshaw and places like it that were just being applied to this super specific problem of arbitrageing securities prices. And it just seemed to me like if you could bring the same amount of intelligence to bear on a problem that would actually affect people's lives in some real way that it would be pretty uh incredible. And so ended up thinking about what's the sort of um adjacent problem that uh that regular people have. and you pretty quickly get to this problem of access to money. Um and uh and with a couple um steps and skips, we ended up um finding our way to the business that we're in now, which is uh consumer credit and and the problem of uh how you make it so that everyone can access consumer credit and access it uh instantly, easily uh in uh in sort of both a a good way and uh ultimately with a good rate. And so can you touch on um like briefly just the product suite upstart and you know what you offer the consumers? >> Yeah. >> Yeah. So Upstart is an AI lending company at heart. Um so we see ourselves first and foremost as a technology company. Uh but of course we actually are also importantly a consumerf facing company and that's because consumers go to upstart.com uh to look for credit and our mission is to offer consumers the best rates and best process uh for every category of consumer product. Today uh we have uh unsecured consumer loans uh usually called personal loans. We have auto refi. We have auto purchase uh and we have heliloc. Um, and in the near future, we expect to round out and complete the suite of products that consumers generally care about for credit. >> And the main reason that um consumers will come to you is because you're using these AI models to hopefully deliver lower costs for them for accessing credit. Is that right? >> Yeah, that's right. So, the the motivation behind the company and and the reason this works, you might think, well, how is it that you could just give everybody lower rates? And of course, the precise answer is we can't give every literally everybody a lower rate. um because then you would just end up with underperforming loans. Um and of course, you know, there are people out there who will try to do that once in a while. It never ends well. >> Yeah. >> Uh the thing that we're actually doing is what we call better separation. Uh and what that means is uh there is always uh there's always going to be uh some some people that you didn't think would pay back that pay back and there's going to be some people that you thought would pay back that don't pay you back, right? And there's errors in in both directions. And all kinds of models have uh these kinds of errors. But in lending, the thing that we discovered uh 10 or so years ago when we started this was that there's an incredible amount of error. The vast vast majority of decisions are erroneous um in both directions. And um and that's because fundamentally it turns out to be the case that something like 80ish% of Americans, if you actually were to do kind of a pure back test, actually repay their loans on time with no problems. M >> but only about 50% of Americans are considered prime credit based on their credit scores and traditional credit reports. So if you're to pull a bureau report and use traditional scoring, whether that's a FICO or something else, you end up thinking, okay, only half of Americans should qualify for bank quality credit and there's a whole another third of the country that would have been able to pay you back, but you you would sort of um erroneously disqualify. Now, the people that have uh access to bank quality credit, generally people with high credit scores, long credit histories, etc., etc. Actually, it turns out that there's a small fraction of them that also still can't pay back those loans. And so, you've got people that uh banks are giving really low rates to who are nonetheless defaulting. And you've got people that banks are completely turning away or have to go to, you know, payday or other really high interest rate options that um that are actually would have been perfectly good borrowers. And so it's our sort of mission to solve both problems. And when you get better separation that way in both populations, you can actually get big gains. So in the population of people that banks are approving, you can actually get to lower APRs for the people that actually will repay because you're pulling the sort of defaulters out of the pool. >> And then of course for the people that are just erroneously um rejected, those people can be pulled into the approval set. And so you can get dramatically higher approval rates and lower APRs. And um and that's so that's that's that's what we offer the end consumer and it's also what we offer uh through banks. So a bank can come to us and work with our technology and offer all of those benefits to their customers and to future ones. >> Amazing. And um are you able to give a like a a recent real life example of of how um your AI has helped someone get credit that they wouldn't have got before? Is there some sort of story you can give there? Um yeah, people always always love specific um borrower stories. Um you know, I I think we I don't have a good one right off hand, but there are like literally every day like hundreds of people write uh write their story on on Trust Pilot. You could just as I think it's like 60,000 of them. Um and probably like 20,000 of them will talk about their experience where, you know, they either their their next best option was some kind of payday loan. They they tried three banks. even a bank that you know they've been using as a customer for the last 10 years turned them down or the bank doesn't um offer the kind of credit that they need or uh that they they were nominally able to get approved for a loan. But then there they were faced with sort of this crazy process that involved uploading all sorts of documents and it seemed to be sort of a neverending labyrinth of requests to verify, you know, if their income is a little non-traditional or um they have something complicated in their past credit history that just turn into these sort of really long processes. And you know, people are are when they need credit are generally looking to spend uh many days or weeks kind of filling out paperwork. It's like, you know, you want credit when you need it. >> Yeah. And um so as I mentioned, you you work with 100 banks and uh and credit unions or whatever. Um and they obviously come to you because you help them save money. Um with more accurate uh loans um how do you maintain that competitive advantage moving forward? Like how do you your values in your AI and and and how accurately you can um do this and how do you do that going forward? >> Yeah. Uh well number one thing is we continue investing in the AI to make it better every day. Uh it gets better in a few different ways. One is that it gets you can think of it as it gets deeper and it gets wider. the the deeper it just means uh like any kind of model company like nowadays people have good familiarity with LLM based companies and of course they're always releasing new generation versions of the models that have more training data uh sort of more uh more sophisticated sort of model architectures more compute and there's exactly the same dynamic uh in uh in lending where our models every generation are upgrading across all of those dimensions and those dimensions actually have a lot of synergies with each other where you can't really just have more columns of data without more rows. You can't just like throw more compute at the problem unless you have a more uh you know more data or a more powerful architecture to take advantage of the compute. But conversely, if you don't have the compute, you just you know thing things don't quite work. Uh in our world because we have a very process sensitive consumer that comes to upstart.com, they are expecting to get a loan uh in real time. It's not like you can, you know, take a really long time to process a decision. So you've got latency constraints and um and so basically all of these kind of like technology problems and data problems together create a pretty um pretty rich space for upgrading the models and so that's something we're constantly doing and and actually it drives the majority of growth uh in in the core business. So that's that's a very important dimension for us is making the AI deeper and better. Um then there's this other dimension which is applying AI to virtually everything that we do. When we started the company, we really only applied it to one area and that was the core credit decision, which was how do we approve more people at lower APRs. Uh we did a good job of that, but over the years, we've realized that there's actually enormous potential to apply AI to the whole rest of the lending process. And so we started with uh the the actual process of verification of a loan. Today, it's about 92% of all our loans are what we call fully automated. that means that neither uh manual work had to be done on the borrower side or on upstart side. And so obviously that has both cost advantages which can be passed through to the borrower but also it's just a way better experience for the consumer who's actually getting a loan. >> Um we've started applying it to the servicing of loans uh which can help people sort of structure their payments and timing and all that to be kind of right place right time. Uh and the second order effect of that is you get much higher repayment rates and of course higher repayment rates leads to uh lower default rates which means again lower uh a lower loss rates that you have to pass through to the APRs of the consumer. Um we've of course uh invested in applying AI to the actual marketing and targeting of loans so that you can reach the right people who actually need the product at the right time. And in our case, that's a really extra interesting problem because if you're going to do differentiated uh decisioning, then you kind of need an equally sophisticated way to do targeting. Otherwise, you can't find the people who would benefit the most from uh the differentiated targeting. And so, we've had to invest uh a pretty serious amount of effort in our machine learning teams on uh on the targeting marketing side. So, all of those I would say are are um uh newer areas for us. Some of them are, you know, a few years old, some of them are just getting started. um but but a lot of rich surface area and then you know on the other side of uh of the ledger is like what's the rest of the market doing and I I think for for better and worse um I think there continues to be I think a fair amount of reticence to move quickly and decisively and I do think that you know the with the pace of AI it's one of these things where if you're not going to be uh both feet in going 100% uh you're basically just not going to get anywhere and I think just the other day we saw on CNBC the the CEO of FICO, a really really successful company, you know, much much larger than us certainly, but um you know, he he was on CNBC just saying that you can't use um AI in uh in underwriting because it's a quote unquote black box. >> Yeah. >> And um you know, agree or disagreement, we obviously disagree, but like you know, when people ask this like how do you know that everyone else isn't doing the same thing? It's just like well they're literally telling you you're not allowed to do it and and so um that's how we know. So when do you think uh ALI will be used in underwriting properly? >> Yeah, I I think I think it's going to come pretty quickly here over probably the next five years or so. I think a few few tailwinds are combining to make that happen. But I think ultimately, you know, the thing that really is going to get make it proven out is just like >> it's we we've had this great success doing this in unsecured consumer loans. Um but unsecured consumer loans, frankly, are only a very small part of the market. I mean, if you were to ask a large bank, they would say that is an irrelevant part of their business. It's an irrelevant part of the portfolio. Maybe they don't even offer the product. They don't care, right? Because in the grand scheme of consumer credit, you think about like auto and homes and cards and all the products like uh personal loans are are are sort of a rounding error. And that's um that's where we spent the first uh almost decade of our uh our business focused on. And really only in the last couple years we started to build out the sort of products that uh the vast vast majority of the sort of consumer credit ecosystem is in. But essentially they've got all the same problems um where you know you you either need to apply AI to get wins in credit decisioning or you need to apply it to get wins in verification and automating the process. And the moment that in, you know, in auto you're suddenly able to double approval rates, uh, or, uh, in home, you're able to have a 90% rate of loans being fully automated, I mean, it's so transformative in lending that I think it's it, no matter what you think about AI, whether you're an enthusiast or not enthusiast, I think you're just going to have to look at the numbers and, you know, come come to terms with reality that this is the new way lending is done. >> Yeah. And I did, so how established are you in those new markets now? Uh so home equity loans for example. >> Yeah. So each each one is a little different. Uh home is uh is probably our newest product that uh is is in market. So it's it's relatively nent for us today. We have a heliloc product but we are not yet in purchase mortgage which is of course the sort of the big part of the market. We're not yet in refi which is sometimes really big sometimes less big depending on interest rates. Um, but HELOC is a really good uh testing ground for us because HELOC uh is in many ways, you know, a halfway step between a personal loan and a purchase mortgage, but it has a lot of the same parts of the process that you have to automate in terms of the you think about titles and leans, um, managing co-barorrowers, like all the things that you would have to do on a purchase mortgage. Um, but you get to do it sort of on a loan where you get training data a little bit faster, um, a little bit more more efficiently because the loan aren't quite as longdated, they aren't quite as big. um and uh and and of course there is um you don't have to deal with quite all of the problems and so it's a really good sort of first step for us um and then we'll build out towards the whole rest of the home market. Uh auto we have a little bit more time on. So we started first in uh in auto refi and then we uh we added an auto purchase product and so if you were to go to um uh one of uh many hundreds of dealerships today you could find uh an upstart powered loan. And the uh the differentiated thing about um what we're doing in auto is that it's not like with uh with any kind of lending product that you could find at the car dealership where you find it through the traditional market ecosystem. We've actually sort of built out our own lending platform that dealerships use. And that platform while it's a bit of a sort of longer journey to build up market share because you can't immediately borrow from the distribution network of the of the sort of existing uh ecosystem. uh you get to sort of I think reimagine the process from scratch in a way that could allow for the possibility of a dramatically simpler and faster process. One where you could gather more signal on the borrower and therefore make a much smarter, more accurate uh underwriting decision and again therefore get to higher approval, lower APR and ultimately you know that means more uh more space for the dealer to be able to sell cars and and sell the products that they need to sell. >> Yeah, that's quite amazing. I'd never um realized that how big the total addressable market is for for loans and how small a section you've started on. It's quite an exciting sort of potential trajectory there. >> Yeah, absolutely. Lending is like >> lending is is an enormous part of the market. We think it's in some sense the the entire source of profits for the whole uh financial sector. >> Yeah. Um, just coming back to something else you mentioned earlier, uh, 90% of loans are fully automated with AI. What's in the last 8% that still needs people? Would you eventually think they'll be automated or there'll always be a percentage that have some human element to it? >> The last 8% is basically fraud. >> Okay. >> Basically fraud. Um, so of course, uh, there are always, um, uh, sorry, I I I I I guess I mean, um, in a more serious way, I mean, of course, there are like always people that are trying to defraud you. If you're giving away money, then someone always wants your money. And a big part of what you have to do is is combat that. And um, and so why is it that we can't have um 100% of all loans be processed fully automatically? Well, the the the sort of um the fundamental limiting constraint is that you have a nonzero number of fraudsters that are coming to attack your system. And when you have a fraudster, you don't want to send them of course through a fully automated process. You may want to uh fully automatically with the p benefit of you know all the machine learning models know that this person is or this applicant is highly likely to be a fraudster. Um, but then once you've identified them as highly likely to be a fraudster, you want to send them through really as uh as cumbersome a review process as you possibly could as opposed to the opposite. Um, and so that that really is what's driving uh what's always going to drive a small residual of manual loans. Um, uh, in theory, of course, if you got uh if you got a perfect model accuracy where you never had a single false positive flag, then you would never have a single uh person that ultimately could get a loan that goes through that. Um, but of course, you know, because fraud is such a rare event, you're going to have at least some false positives because you're sort of always looking for needles in a hay stack. And so, I think that 8% manual number of loans will will go down, but as a percent of applications, there's there's just always going to be some uh applications we we send through the kind of like intensive fraud fraud review channel. And and that's a that's a good thing. Um, yeah. And is there a um a particular type of fraud that's the hardest to sort of stop? Uh yeah, for sure. Um so you know when you think fraud, you often think of kind of like egregious explicit fraud. Um and that really is um what generally in the industry we call third party fraud. And that basically means like someone else has come and stolen your identity. Maybe there's a sophisticated fraud ring. Maybe you know these people are going out and um kind of like tricking uh seniors who who don't quite use the internet as much. You know all sorts of things that like this that could be happening. In some sense, these are easier to detect because they have common patterns to their behavior where it's a fraud ring. You know, there's there's sort of multiple applications that h happen at the same time or there's data out there on the internet. Maybe they're kind of discussing their strategies in forums on the dark web and there sort of like actual uh data about about it and patterns about it that you can detect. And of course, that is what AI is very good at at doing. Then there's this kind of softer kind of fraud that I think you could even debate is it really fraud? um what what we sometimes call first party fraud. And that basically means that a person has decided that they want to strategically trash their credit score uh in exchange for more money. And um uh and that is that's very hard in some ca in some cases it's really easy because there will be what we think of as signs of uh signs of malicious intent. Uh, and actually I think a lot of the non-traditional signals in our underwriting models and our verification models are really good at picking up stuff like this. If you think about stuff like you know just like how fast you're doing things um is kind of like there's some unnatural speed that you might do things if you just don't care about repayment. Like you know why do you even care what the monthly payment is if you don't intend to ever make a monthly payment, right? Um and and so there's there's all these sort of like little signals that help with this problem. Um, but I think in a very uh fundamental sense, like a very uh well uh sort of planned um kind of firstparty fraudster that is intent on uh trashing their own credit uh credit in order to like get some money and like sort of knows the whole game and knows how to play it. Um that is that's something that uh that at the limit I think is is very hard to completely eliminate. and um assume that this is another reason why lenders choose your platform because you help them to mitigate using AI against fraud as well. >> Yeah. >> Yeah. Yeah. For sure. I mean I think the history of traditional banks moving into online uh online products where it's not a customer they've had a long relationship with. It's what it's like internet walk up traffic. I mean that is a history where you talk to any bank and they'll tell you that there are just so many cases where that ends in these you know super high fraud rates. You know we could be talking about like 5% of all the loans you've made turn out to be to fraudsters. And of course that's just crushing from an ultimate financial perspective. And so I think a lot of the reason that um traditional banks are reluctant to make uh new unsecured loans on the internet to strangers that they don't have a relationship with is there's just such a high chance that something something goes wrong with that and um and we give them the ability to do that. Um which is is a big thing because of course you want new customers and new customers today are so much more likely to want to come via the internet instead of walking in the branch. >> Amazing. Um, and we'll swing back round to the Q2 numbers because they're really great. Uh, revenue up, loans up, um, more automation. What changed in your strategy? Where does where's this come about from? Um, yeah. Could you just give us more details about that? >> Yeah. Um, so as I said before, the single biggest source of growth for the company comes from wins in AI. And um one of the big wins that we had in this past quarter was what we call the uh an improvement to the meta model architecture. So our core underwriting model is a um is actually an ensemble of different machine learning models. That means is there are different algorithms that are used to uh to learn from training data and make predictions about whether a particular applicant is going to repay or not. in fact uh whether they're going to uh default or prepay in in each time period of the entire loan's life and we use uh these different kinds of learning algorithms and then we uh blend the results together to get the final output. Now when we do the blending we actually are passing the outputs of the first layer of models into a second layer that we call the meta layer. The meta layer has also machine learning algorithms and those algorithms are responsible just for figuring out the best way to blend the outputs of the sort of individual base models and previously we had had neural networks as one type of model uh algorithm in the base layer but we did not have it in the meta layer and that's uh for various reasons but I think the most general one you can think of as uh the more sophisticated your model architecture the exponentially more training data that you need in order to make it work. >> And previously we just we didn't have quite enough training data and we didn't have quite enough uh compute or latency tolerance in order to uh to have neural networks which are one of the most powerful types of learning algorithms in both the base and meta layer. Uh and in Q uh in the last quarter we were able to change that. So now we have neural networks operating at both layers of the model. And that architectural change uh created a very large boost to model accuracy and that boost to model accuracy unlocked these separation gains that allowed us to have higher approvals at lower APRs and that uh drove a bunch of growth. >> Okay. So it's it's mainly it's mainly from people who were coming to your site asking for loans but were getting rejected and this actually included them in the segment where they would be accepted. Yeah. >> Uh some of them uh getting previously were rejected. Some of them previously were getting an APR that was too high for them to want to take out. >> Yeah. Got you. Uh which is really interesting. So it was coming from traffic and demand that you were already receiving. You just Yeah, like you said improving the model. So um how much you me mentioned earlier because you you spend fair bit on marketing now. How is that a large percentage of of revenue and is that increasing and how much is that becoming part of your business to generate demand as well? >> Yeah, we we certainly um we certainly spend on marketing though one of the cool facts about the company is that our CAC the the cost of acquiring customer has actually been uh loosely speaking trending down uh over uh over over the years. And that's um kind of the opposite of generally what happens in these businesses where people think if you're a business where you're just kind of like skimming off the top. You found some little niche of of people that were underserved before >> and you reach those people and then it's like you want to reach any more people, your CAC's generally going to go up because it gets sort of harder on the margin to find the sort of incremental that person because you've already picked off the lowhanging fruit. And we actually have experienced the opposite dynamic where we've been able to grow the kind of pie enough over time that I think the rate at which we're growing the the population of people that we can be the best rate and best process for is actually happening faster than the rate at which we're kind of depleting the lowhanging fruit from a targeting and marketing perspective. And so we've been able to actually bring the um uh the the sort of CAC of our business down um and uh and all while sort of increasing our level of investment there. So I would expect that to to continue. Um I I think we don't have actually an aspiration for like we don't need from a business model standpoint for cap to be any lower. We're perfectly happy to be where it is or even higher is fine too. But um but we want it to be a we want to be able to keep investing more and more dollars in marketing and reach larger and larger share of the population as we have the right uh and best products for uh for more people in more situations. Mhm. And um I'm just jumping around sorry because I I remember things from earlier in the conversation but uh we were talking about the AI algorithms again uh the models. Um how do we make how do you make sure that it's being fair between two similar borrowers so they get treated the same way? >> Yeah. >> Well, my puny answer is that only humans are capable of treating two similar people different ways. It's not something that like a computer or an algorithm can do. you know, computers are like really good at just kind of like um following the data. And if you know, the data says two people are the same, then you know, the the model's going to treat them the same. That's, you know, all all cares about doing. >> Um but I think in a more serious sense, I think there's a really interesting debate to be had about, you know, what what does it mean for two borrowers to be similar? And I think this is the essence of um of the question that I think you have to answer. And um one of the one of the moes if you will around our business is that we've spent so many years and so much uh research time on on this particular question we've done it in partnership with major regulators like the CFPB. We've done it with uh major researchers at universities who are some of the kind of leading uh academics that publish papers on questions like what is fairness? How do you measure it? How do you test it in the context of a machine learning model of um these sort of like neural network models? And um and I think the gist of it ultimately that we get down to is the the similar the dimension of similarity that really matters is willingness and ability to repay a loan. Um at the end of the day, you know, when you say two borrowers are similar, is it fair? What you really mean is if two people are both similarly able to repay a loan and you treated them differently, then then there's something wrong. That's not a thing that I think we as a society want. But I think sometimes people have confused the input and the output or the X and the Y on this question where they say, well, what if two people with similar FICO scores get treated differently? And my answer is they should absolutely be treated differently because if unless the FICO score is the beall endall of whether you're willing and able to pay back a loan, it should not be the dimension on which we measure similarity for purposes of fairness because ultimately what we want to know is are are you going to be able to pay back this loan? And I don't think it's in anyone's best interest for us to give loans to people who can't pay them back and ultimately, you know, going to end up in bankruptcy or another kind of bad financial situation. So then the question is okay well if what we really care about is finding similarity with respect to uh ability to repay a loan then what we actually need to first have is the most accurate possible model for knowing if someone is going to be able to pay back a loan. And so you really don't want to benchmark against a really simplistic uh sort of traditional model. You want to benchmark against the most accurate available model that you've got and use that as the source of truth for likelihood of repayment and then make sure that you're treating uh those borrowers uh in in a similar way. >> And slightly related but um if we look at general macro conditions uh which affects people's ability to pay back loans or you know and vice versa when it's you know in different kinds of macro conditions. How do you um how do you prepare for you know or mitigate against loan delinquencies in bad macro conditions and then you know change that when it when things get better? Is that is that all part of the model that does it does it analyze economic conditions? >> Yeah. So one of the things we've done over the last couple years partially in response to some of the learnings we had from 2022 or or thereabouts is that macro is really important and rather than ignore it we should figure out what part of the problem is tractable um and solvable uh by AI and be the best in the world at solving that particular part of the problem. Now the reason I I describe it as what particular part of the problem is solvable is that uh in the broadest sense I think macroeconomic conditions are of course very subject to kind of idiosyncratic singular events that don't tend to have a long history of repetition. So meaning like if you think about like COVID as a macro shock, well there's never been a macro shock like COVID before and u of course all kind of AI machine learning are based on training data that have sort of historical bases and you you know you're just not going to be particularly good at predicting that. Now I don't think that's a weakness of AI per se because I don't think there are like human experts that uh you know and if you ask them in 2018 we're going to tell you that in 2020 we're going to have the COVID pandemic, right? just like that was a kind of um black swan type event. But unfortunately, of course, those things do happen in macro. And so what we eventually figured out was that like you're not going to be able to predict all of the kind of idiosyncratic black swan events, but you can be the most precise and the most quick to respond uh to to any particular set of macro conditions and have it properly represented and priced in the model. And that actually is a pretty big edge because it turns out that most of the ways that people measure macro are delayed by a fairly significant amount of time. We think of it as often with a three to six month lag. And that's because if you think about what are the indicators of macro? Well, the Fed publishes data. The Fed publishes data on the past quarter in the middle of the next quarter. >> Yeah. >> And oftent times they end up revising that number for the next two to three quarters. >> Yeah. So to get like a real kind of estimate of what's going on in the macro, like you probably need to wait n months before you really know, you know, what was the employment rate back in uh November of 2024. Like we're probably just really getting a clear picture of that today. >> Yeah. And um you know this there's a similar uh but less extreme story even if you look at credit data where maybe you can go to credit bureaus but generally like people are looking at things like 60-day delinquencies as a key metric or 90-day delinquencies and well guess what you have to wait 60 or 90 days to measure 60 or 90day delinquencies. And so you're you're just always lagged by 3, six or nine months uh when you're trying evaluating macro. And so we invested in building what we call the Upstar macro index. And actually behind the scenes is this whole suite of other tools uh that you know will figure out the right ways to surface over time. But what we are um invested in doing is figuring out how we can make it so that we are just the fastest uh to respond and the most precise in responding. And our models today now have something that we internally think of as sort of this individualized um macro component where we can look at any arbitrary set of characteristics that you've got. Maybe you're a high-income person. Maybe you're a low-inccome person. Maybe you're a person who works in nursing. Maybe you're a person who works in the government. Maybe you're a person who works in a tariffsensitive industry. Maybe you're a person who doesn't work in a tariffsensitive industry. And the model can pick up uh trends that are individualized to that particular set of characteristics that you've got where if we see for example that maybe tariffs sensitive industries are being more negatively impacted and uh defaults are just starting to creep up there. When I say defaults, I actually mean you know the first signs of delinquency is starting to creep up. Um the model can immediately pick that up and respond uh in real time and it doesn't require uh people to specify like oh we're interested in you know bucket A and bucket B and bucket C because you just don't know what the next macroshock is going to affect precisely and you want that to be something that uh the model can figure out for itself. And so that's that's how we've architected the model and we think it's a very powerful solution for being the the fastest to respond, most precise to respond. But again, we're not we don't pretend that we're going to predict the next co so to speak. >> Yeah. Yeah. That's amazing. That's pretty um it's pretty incredible how many areas uh there are to do improvements in the way it used to work. Um I mean it just makes sense, right, that you should be quick at reacting to things like this. >> Kind of incredible. So, as you're picking up in all these new areas as you get deeper into the the loans industry, um, what are you most excited about over the next 5 10 years in lending? Most excited about over the next 5 to 10 years. Well, you know, I think in um uh I think in probably as we were talking about earlier, probably in 5 years time, I think we'll have successfully brought uh AI lending to the parts of the consumer credit market that actually are big and matter to people. Yeah. So that's, you know, really I'd say home and auto and uh and some flavor of revolving credit are the three big categories. And I think when that happens, I think there's going to be a seismic move in uh in how people perceive the relevance of AI and lending. Today, it's this weird niche thing that's just like maybe this this cute little company Upstart does. Maybe they'll go away. Who knows? Or maybe it'll stay in there a little. Um but I just really think that the perception around it is going to completely change uh when you bring it to these these much larger uh markets. And u and I think when that happens, I think everybody is going to care about this. And the end result of that is I think you're actually going to asmmptoically approach uh the real levels of access to credit that should exist in this country which as I said earlier is something like 80% of Americans should have access to bank quality credit. They should have it on demand with virtually no work required of them. And the other 20% should have a clear and defined path of how to get from where they are to where they want to be. >> And what's stopping up start doing this attacking those areas earlier? Uh we're we're working on it. Uh as I said, it was you know started in the home world on on Heloc. Of course, each of these you know you you have uh there's there's a lot of work to to do. Um and some ways each market has uh work that's specific to that market. So you think about like how do you how do you actually become relevant in the auto business? Well, you have to be in car dealerships because that's where almost all auto loans happen. But car dealerships are, you know, their own uh very specific industry with the way they work. Um, and frankly, I think, you know, we've had to spend a couple years like learning the ropes of like, okay, uh, we have to really understand things like how do dealer markups work? How do, um, how do add-on products work? How does like gap insurance work? Like the things that dealerships really care about. Those were not things that, you know, certainly I I didn't know anything about them when I started this company. And I didn't know anything about them three years ago when we started in the auto lending business because none of those questions were relevant in personal loans. But if you want to be a relevant player in a car dealership, you need to really understand those things. Your product needs to be very good at supporting those. And uh and so th those are, you know, those are things we've kind of figured out and iterated towards and um and and gotten a lot of great wins on. Same in in home. I think a different set of challenges that are home specific um but that we're working through and again I said earlier Heliloc has been a great sort of learning and testing ground for us to really figure out like all the mechanics of how you can most uh most quickly and most efficiently process figure out leans and titles and uh co-barorrowers all of the pieces that you really need on a home loan. >> Awesome. And so do you think Upstart will it will sort of exist as a mortgage broker sort of way like it will be one of the bigger players in that space and just you know I think that they they're the ones facilitating demand into different uh lending platforms at the moment or would it work differently than that? >> Uh we certainly expect to be a big player in in home loans. I mean home uh home loans I think in some sense are when people think about credit uh why why do they care about credit? It's because one day they hope to be a homeowner. They hope to be able to, you know, buy buy a house when when they need it and probably they hope to be able to get the loan quick enough that they can actually close on the home that they're trying to trying to buy instead of, you know, running out the clock or someone else getting grabbing the house. >> Um, so, uh, in that I guess, uh, in that sense, I think it is inevitable that home is going to be a central piece of what we care about here at Upstart. It's just just so big, so fundamentally important uh, for the consumer. Now within that ecosystem, you know, there's a whole bunch of different ways that you can technically slot into the ecosystem from, you know, are you are you are you broker, are you correspondent, are you technology provider, are you uh the originator? Um and um and I think we're not particularly uh dogmatic about uh which particular sort of um way uh technically we're engaging, but ultimately we want it to be the case that if you if you come to Upstart uh that you you have a guaranteed way to ultimately get uh get uh the home loan that that you need. Uh, and I say when I say guaranteed ultimately I mean either you can qualify today, you can get it instantly, you can get it with high predictability really fast, or if you can't qualify today, we can uh we can help you figure out what needs to be true so that you can qualify and so that you can get uh the home loan that that you need to sort of make make your life happen. And um as long as that's true for the consumer, you know, exactly who's doing the originating uh and um you know, how how it interacts with the GSC's and all that, that can be figured out on on the back end. >> And do you um foresee moving into other regions in the future? I believe you're just US at the moment. Um or is that I mean the US is such a gigantic market obviously a lot of companies stay here for a very long time just because it's big enough to grow very big without going somewhere else. Yeah, we we are exclusively uh US-based today. Um I I I would say um uh we never say never about it. Uh we certainly don't have any immediate plans to be outside the US. It is such just the such large part of uh the world economy, but also I think in credit in particular, consumer credit, consumer credit is so big in the US compared to outside the US. It just sort of really makes sense to start focus there. But um but I do think there are really interesting credit access problems outside the US that if you're thinking of the problem as where in the world is there the very worst access to credit where there's the most amount of kind of undervalued consumers that aren't being served right today that could be served. Um you know certainly there's a lot of interesting places outside the US and you know we could be there one day. >> Awesome. Um I've got a quick lightning round if that's all right where I'm just going to ask a few questions and >> Okay. >> A very quick answer to it. Um, what is the bigger bottleneck for uh improving AI lending? Better data or more compute? Um, in the short to medium term, better data. Uh, in the long run, it's going to be latency. Um, so a version of compute, not exactly compute, but kind of uh kind of like compute. >> Yeah. Okay. Um, one input you wish every borrower had to share with you. Um, I would take I would take everyone's driving records. >> Okay. Uh, by 2027, which LLM Oh, will LLMs be part of underwriting? We sort of discussed. >> Yeah. >> Um, what's Upstart's biggest moat in 2030? Data or distribution? >> Data. >> By 2030, what percentage of loan decisions will be made by AI? I think we're the wider market here. >> 100% of consumer. >> Okay. Um, one job in finance that AI will create not kill or in tech you can widen it. You know, I think that I think there's a real space for uh for more um for more financial counseling. Um I think that that is something people want with an emotional uh human touch to help them work through >> uh their plans, their life plans, how it relates to money. I think money is a, you know, it's a it's a it's a stressful thing for people and I think we can uh make that process uh we can make it quick, we can make it easy, we can make it delightful, we can make it, you know, your rates low. Um but uh but I actually think that if you take away all of that stuff and make it so instant, so easy and you know at at great prices for people at the end of the day, you still need to decide how you want to live your life and you know what kind of lifestyle you want to support. Do you want to spend more now, spend more later? And uh and I think ultimately those are human questions. >> Yeah. And last question, one non-obvious book or paper that shaped your approach to business. >> Yeah, there's this um there's this really long super nerdy um uh paper written by this guy Cliff Aznes who runs a fund called AQR. And um I I I don't quite remember the name of this paper but he the author of it and it is a paper on uh the ways in which markets are efficient and inefficient. And the beauty of the paper is just that you really see like if you want to think rigorously about efficient markets and uh therefore inefficient markets. um you have to think really really hard um because the um some of the smartest people in the world uh have worked on this problem and um if you want to find opportunities to generate alpha to do sort of do better than market do better than sort of uh the status quo uh it's not something that you know you can just um you can just you could just sort of do on a whim it's something you need to take really really seriously and I think the ethos of that paper has stuck with me for all these years of building upstart >> awesome thanks Paul um It's been a pleasure to have you on the show and really interesting to dig into Upstart. It's definitely one of the companies to watch and I'm sure everyone will value this interview a lot. Um, thank you so much. I don't know if there's anything else you'd like to to leave with our community. >> No. Awesome. It's been great to be here. >> Cheers. Thanks, Paul. Have a great day. >> Hey, you. Bye.
UPST Stock: Upstart's Co-Founder on 92% Loan Automation
Summary
Transcript
Yeah. So, Upstart is an AI lending company at heart. Our mission is to offer consumers the best rates and best process uh for every category of consumer product. Fundamentally, something like 80ish% of Americans actually repay their loans on time with no problems. But only about 50% of Americans are considered prime credit based on their credit scores and traditional credit reports. It's like, you know, you want credit when you need it. No matter what you think about AI, whether you're an enthusiast or not enthusiast, I think you're just going to have to look at the numbers and, you know, come come to terms with reality that this is the new way lending is done. >> Hello everyone. Today we're joined by Paul Goo, co-founder and CTO of Upstart, the AI lending marketplace behind more than 100 bank and credit union partners. Upstart just posted tripledigit revenue growth and uh return to profitability in Q2 with 92% of loans fully automated and conversion up to approximately 24%. Paul, welcome to the show. >> Thanks, Ed. Great to be here. >> Um are you are you calling from your HQ at the moment? Where where's the is this the background? >> No, no, this is my home office. Um built built for the digital world. Uh >> whereabouts are you based? America. >> Uh yeah, I spend most of my time right outside Salt Lake in the US. >> Okay. Oh, nice. That's cool. And uh the team spread out over the country or >> Yeah, we've got people in almost all 50 states. Um really over the course of COVID, we discovered that you could get incredible talent um by hiring all over. And I think for companies that hit their growth spurt around 2021, 2022, um you just ended up with uh with a great spread of talent everywhere. >> Nice. So you've you've maintained that going forward. This is your structure for how the company operates basically. >> Yeah, we we have a working model that we call digital first where uh we do hire people everywhere across the country and we think the best talent can be all over the place. But we also really believe in in-person work. So we bring teams together on an almost once a month basis to spend time together in person. So people do launches together, you know, they can do kind of like intense collaborations or just spending time getting to know your your co-workers. doing that's really important and then the math actually pencils if you think about the there's travel cost versus office cost and especially you know if the office cost was going to be in a really expensive place um the math actually works just great. >> Yeah. Okay. And so when they come to meet up do you hire like an event spaces because you don't have an office or >> Yeah. So we actually do have some office spaces that pretty regularly because teams are always gathering on a regular cadence. The offices are always being used um but it's it's not like every team at the same time. So, we've got offices sort of in a few places across the country. Um, but yeah, we also do event spaces for larger gatherings. Um, and you know, once a year we bring the whole company together and then a couple times a year, you know, big chunks of the company. >> Amazing. And can you just touch on um your background as well because I I believe Yale and then pretty much straight into Upstart. >> Yeah, that's right. Um, I don't have a terribly long life story to share before Upstart. Um, but yeah, I I was at Yale. I was doing a double in computer science and uh in econ and um I dropped out to do Peter Teal's 2020 program uh the so-called teal fellowship. Um so the first year of that program and um >> so I dropped out I spent a little bit of time um at this quant fund called DE Shaw and then um pretty quickly pivoted to thinking about um problems really at the intersection of uh of machine learning and problems that would have an impact on uh on real people. Uh, you know, I always felt like there was just an incredible amount of uh brain power and compute power um at uh at Dshaw and places like it that were just being applied to this super specific problem of arbitrageing securities prices. And it just seemed to me like if you could bring the same amount of intelligence to bear on a problem that would actually affect people's lives in some real way that it would be pretty uh incredible. And so ended up thinking about what's the sort of um adjacent problem that uh that regular people have. and you pretty quickly get to this problem of access to money. Um and uh and with a couple um steps and skips, we ended up um finding our way to the business that we're in now, which is uh consumer credit and and the problem of uh how you make it so that everyone can access consumer credit and access it uh instantly, easily uh in uh in sort of both a a good way and uh ultimately with a good rate. And so can you touch on um like briefly just the product suite upstart and you know what you offer the consumers? >> Yeah. >> Yeah. So Upstart is an AI lending company at heart. Um so we see ourselves first and foremost as a technology company. Uh but of course we actually are also importantly a consumerf facing company and that's because consumers go to upstart.com uh to look for credit and our mission is to offer consumers the best rates and best process uh for every category of consumer product. Today uh we have uh unsecured consumer loans uh usually called personal loans. We have auto refi. We have auto purchase uh and we have heliloc. Um, and in the near future, we expect to round out and complete the suite of products that consumers generally care about for credit. >> And the main reason that um consumers will come to you is because you're using these AI models to hopefully deliver lower costs for them for accessing credit. Is that right? >> Yeah, that's right. So, the the motivation behind the company and and the reason this works, you might think, well, how is it that you could just give everybody lower rates? And of course, the precise answer is we can't give every literally everybody a lower rate. um because then you would just end up with underperforming loans. Um and of course, you know, there are people out there who will try to do that once in a while. It never ends well. >> Yeah. >> Uh the thing that we're actually doing is what we call better separation. Uh and what that means is uh there is always uh there's always going to be uh some some people that you didn't think would pay back that pay back and there's going to be some people that you thought would pay back that don't pay you back, right? And there's errors in in both directions. And all kinds of models have uh these kinds of errors. But in lending, the thing that we discovered uh 10 or so years ago when we started this was that there's an incredible amount of error. The vast vast majority of decisions are erroneous um in both directions. And um and that's because fundamentally it turns out to be the case that something like 80ish% of Americans, if you actually were to do kind of a pure back test, actually repay their loans on time with no problems. M >> but only about 50% of Americans are considered prime credit based on their credit scores and traditional credit reports. So if you're to pull a bureau report and use traditional scoring, whether that's a FICO or something else, you end up thinking, okay, only half of Americans should qualify for bank quality credit and there's a whole another third of the country that would have been able to pay you back, but you you would sort of um erroneously disqualify. Now, the people that have uh access to bank quality credit, generally people with high credit scores, long credit histories, etc., etc. Actually, it turns out that there's a small fraction of them that also still can't pay back those loans. And so, you've got people that uh banks are giving really low rates to who are nonetheless defaulting. And you've got people that banks are completely turning away or have to go to, you know, payday or other really high interest rate options that um that are actually would have been perfectly good borrowers. And so it's our sort of mission to solve both problems. And when you get better separation that way in both populations, you can actually get big gains. So in the population of people that banks are approving, you can actually get to lower APRs for the people that actually will repay because you're pulling the sort of defaulters out of the pool. >> And then of course for the people that are just erroneously um rejected, those people can be pulled into the approval set. And so you can get dramatically higher approval rates and lower APRs. And um and that's so that's that's that's what we offer the end consumer and it's also what we offer uh through banks. So a bank can come to us and work with our technology and offer all of those benefits to their customers and to future ones. >> Amazing. And um are you able to give a like a a recent real life example of of how um your AI has helped someone get credit that they wouldn't have got before? Is there some sort of story you can give there? Um yeah, people always always love specific um borrower stories. Um you know, I I think we I don't have a good one right off hand, but there are like literally every day like hundreds of people write uh write their story on on Trust Pilot. You could just as I think it's like 60,000 of them. Um and probably like 20,000 of them will talk about their experience where, you know, they either their their next best option was some kind of payday loan. They they tried three banks. even a bank that you know they've been using as a customer for the last 10 years turned them down or the bank doesn't um offer the kind of credit that they need or uh that they they were nominally able to get approved for a loan. But then there they were faced with sort of this crazy process that involved uploading all sorts of documents and it seemed to be sort of a neverending labyrinth of requests to verify, you know, if their income is a little non-traditional or um they have something complicated in their past credit history that just turn into these sort of really long processes. And you know, people are are when they need credit are generally looking to spend uh many days or weeks kind of filling out paperwork. It's like, you know, you want credit when you need it. >> Yeah. And um so as I mentioned, you you work with 100 banks and uh and credit unions or whatever. Um and they obviously come to you because you help them save money. Um with more accurate uh loans um how do you maintain that competitive advantage moving forward? Like how do you your values in your AI and and and how accurately you can um do this and how do you do that going forward? >> Yeah. Uh well number one thing is we continue investing in the AI to make it better every day. Uh it gets better in a few different ways. One is that it gets you can think of it as it gets deeper and it gets wider. the the deeper it just means uh like any kind of model company like nowadays people have good familiarity with LLM based companies and of course they're always releasing new generation versions of the models that have more training data uh sort of more uh more sophisticated sort of model architectures more compute and there's exactly the same dynamic uh in uh in lending where our models every generation are upgrading across all of those dimensions and those dimensions actually have a lot of synergies with each other where you can't really just have more columns of data without more rows. You can't just like throw more compute at the problem unless you have a more uh you know more data or a more powerful architecture to take advantage of the compute. But conversely, if you don't have the compute, you just you know thing things don't quite work. Uh in our world because we have a very process sensitive consumer that comes to upstart.com, they are expecting to get a loan uh in real time. It's not like you can, you know, take a really long time to process a decision. So you've got latency constraints and um and so basically all of these kind of like technology problems and data problems together create a pretty um pretty rich space for upgrading the models and so that's something we're constantly doing and and actually it drives the majority of growth uh in in the core business. So that's that's a very important dimension for us is making the AI deeper and better. Um then there's this other dimension which is applying AI to virtually everything that we do. When we started the company, we really only applied it to one area and that was the core credit decision, which was how do we approve more people at lower APRs. Uh we did a good job of that, but over the years, we've realized that there's actually enormous potential to apply AI to the whole rest of the lending process. And so we started with uh the the actual process of verification of a loan. Today, it's about 92% of all our loans are what we call fully automated. that means that neither uh manual work had to be done on the borrower side or on upstart side. And so obviously that has both cost advantages which can be passed through to the borrower but also it's just a way better experience for the consumer who's actually getting a loan. >> Um we've started applying it to the servicing of loans uh which can help people sort of structure their payments and timing and all that to be kind of right place right time. Uh and the second order effect of that is you get much higher repayment rates and of course higher repayment rates leads to uh lower default rates which means again lower uh a lower loss rates that you have to pass through to the APRs of the consumer. Um we've of course uh invested in applying AI to the actual marketing and targeting of loans so that you can reach the right people who actually need the product at the right time. And in our case, that's a really extra interesting problem because if you're going to do differentiated uh decisioning, then you kind of need an equally sophisticated way to do targeting. Otherwise, you can't find the people who would benefit the most from uh the differentiated targeting. And so, we've had to invest uh a pretty serious amount of effort in our machine learning teams on uh on the targeting marketing side. So, all of those I would say are are um uh newer areas for us. Some of them are, you know, a few years old, some of them are just getting started. um but but a lot of rich surface area and then you know on the other side of uh of the ledger is like what's the rest of the market doing and I I think for for better and worse um I think there continues to be I think a fair amount of reticence to move quickly and decisively and I do think that you know the with the pace of AI it's one of these things where if you're not going to be uh both feet in going 100% uh you're basically just not going to get anywhere and I think just the other day we saw on CNBC the the CEO of FICO, a really really successful company, you know, much much larger than us certainly, but um you know, he he was on CNBC just saying that you can't use um AI in uh in underwriting because it's a quote unquote black box. >> Yeah. >> And um you know, agree or disagreement, we obviously disagree, but like you know, when people ask this like how do you know that everyone else isn't doing the same thing? It's just like well they're literally telling you you're not allowed to do it and and so um that's how we know. So when do you think uh ALI will be used in underwriting properly? >> Yeah, I I think I think it's going to come pretty quickly here over probably the next five years or so. I think a few few tailwinds are combining to make that happen. But I think ultimately, you know, the thing that really is going to get make it proven out is just like >> it's we we've had this great success doing this in unsecured consumer loans. Um but unsecured consumer loans, frankly, are only a very small part of the market. I mean, if you were to ask a large bank, they would say that is an irrelevant part of their business. It's an irrelevant part of the portfolio. Maybe they don't even offer the product. They don't care, right? Because in the grand scheme of consumer credit, you think about like auto and homes and cards and all the products like uh personal loans are are are sort of a rounding error. And that's um that's where we spent the first uh almost decade of our uh our business focused on. And really only in the last couple years we started to build out the sort of products that uh the vast vast majority of the sort of consumer credit ecosystem is in. But essentially they've got all the same problems um where you know you you either need to apply AI to get wins in credit decisioning or you need to apply it to get wins in verification and automating the process. And the moment that in, you know, in auto you're suddenly able to double approval rates, uh, or, uh, in home, you're able to have a 90% rate of loans being fully automated, I mean, it's so transformative in lending that I think it's it, no matter what you think about AI, whether you're an enthusiast or not enthusiast, I think you're just going to have to look at the numbers and, you know, come come to terms with reality that this is the new way lending is done. >> Yeah. And I did, so how established are you in those new markets now? Uh so home equity loans for example. >> Yeah. So each each one is a little different. Uh home is uh is probably our newest product that uh is is in market. So it's it's relatively nent for us today. We have a heliloc product but we are not yet in purchase mortgage which is of course the sort of the big part of the market. We're not yet in refi which is sometimes really big sometimes less big depending on interest rates. Um, but HELOC is a really good uh testing ground for us because HELOC uh is in many ways, you know, a halfway step between a personal loan and a purchase mortgage, but it has a lot of the same parts of the process that you have to automate in terms of the you think about titles and leans, um, managing co-barorrowers, like all the things that you would have to do on a purchase mortgage. Um, but you get to do it sort of on a loan where you get training data a little bit faster, um, a little bit more more efficiently because the loan aren't quite as longdated, they aren't quite as big. um and uh and and of course there is um you don't have to deal with quite all of the problems and so it's a really good sort of first step for us um and then we'll build out towards the whole rest of the home market. Uh auto we have a little bit more time on. So we started first in uh in auto refi and then we uh we added an auto purchase product and so if you were to go to um uh one of uh many hundreds of dealerships today you could find uh an upstart powered loan. And the uh the differentiated thing about um what we're doing in auto is that it's not like with uh with any kind of lending product that you could find at the car dealership where you find it through the traditional market ecosystem. We've actually sort of built out our own lending platform that dealerships use. And that platform while it's a bit of a sort of longer journey to build up market share because you can't immediately borrow from the distribution network of the of the sort of existing uh ecosystem. uh you get to sort of I think reimagine the process from scratch in a way that could allow for the possibility of a dramatically simpler and faster process. One where you could gather more signal on the borrower and therefore make a much smarter, more accurate uh underwriting decision and again therefore get to higher approval, lower APR and ultimately you know that means more uh more space for the dealer to be able to sell cars and and sell the products that they need to sell. >> Yeah, that's quite amazing. I'd never um realized that how big the total addressable market is for for loans and how small a section you've started on. It's quite an exciting sort of potential trajectory there. >> Yeah, absolutely. Lending is like >> lending is is an enormous part of the market. We think it's in some sense the the entire source of profits for the whole uh financial sector. >> Yeah. Um, just coming back to something else you mentioned earlier, uh, 90% of loans are fully automated with AI. What's in the last 8% that still needs people? Would you eventually think they'll be automated or there'll always be a percentage that have some human element to it? >> The last 8% is basically fraud. >> Okay. >> Basically fraud. Um, so of course, uh, there are always, um, uh, sorry, I I I I I guess I mean, um, in a more serious way, I mean, of course, there are like always people that are trying to defraud you. If you're giving away money, then someone always wants your money. And a big part of what you have to do is is combat that. And um, and so why is it that we can't have um 100% of all loans be processed fully automatically? Well, the the the sort of um the fundamental limiting constraint is that you have a nonzero number of fraudsters that are coming to attack your system. And when you have a fraudster, you don't want to send them of course through a fully automated process. You may want to uh fully automatically with the p benefit of you know all the machine learning models know that this person is or this applicant is highly likely to be a fraudster. Um, but then once you've identified them as highly likely to be a fraudster, you want to send them through really as uh as cumbersome a review process as you possibly could as opposed to the opposite. Um, and so that that really is what's driving uh what's always going to drive a small residual of manual loans. Um, uh, in theory, of course, if you got uh if you got a perfect model accuracy where you never had a single false positive flag, then you would never have a single uh person that ultimately could get a loan that goes through that. Um, but of course, you know, because fraud is such a rare event, you're going to have at least some false positives because you're sort of always looking for needles in a hay stack. And so, I think that 8% manual number of loans will will go down, but as a percent of applications, there's there's just always going to be some uh applications we we send through the kind of like intensive fraud fraud review channel. And and that's a that's a good thing. Um, yeah. And is there a um a particular type of fraud that's the hardest to sort of stop? Uh yeah, for sure. Um so you know when you think fraud, you often think of kind of like egregious explicit fraud. Um and that really is um what generally in the industry we call third party fraud. And that basically means like someone else has come and stolen your identity. Maybe there's a sophisticated fraud ring. Maybe you know these people are going out and um kind of like tricking uh seniors who who don't quite use the internet as much. You know all sorts of things that like this that could be happening. In some sense, these are easier to detect because they have common patterns to their behavior where it's a fraud ring. You know, there's there's sort of multiple applications that h happen at the same time or there's data out there on the internet. Maybe they're kind of discussing their strategies in forums on the dark web and there sort of like actual uh data about about it and patterns about it that you can detect. And of course, that is what AI is very good at at doing. Then there's this kind of softer kind of fraud that I think you could even debate is it really fraud? um what what we sometimes call first party fraud. And that basically means that a person has decided that they want to strategically trash their credit score uh in exchange for more money. And um uh and that is that's very hard in some ca in some cases it's really easy because there will be what we think of as signs of uh signs of malicious intent. Uh, and actually I think a lot of the non-traditional signals in our underwriting models and our verification models are really good at picking up stuff like this. If you think about stuff like you know just like how fast you're doing things um is kind of like there's some unnatural speed that you might do things if you just don't care about repayment. Like you know why do you even care what the monthly payment is if you don't intend to ever make a monthly payment, right? Um and and so there's there's all these sort of like little signals that help with this problem. Um, but I think in a very uh fundamental sense, like a very uh well uh sort of planned um kind of firstparty fraudster that is intent on uh trashing their own credit uh credit in order to like get some money and like sort of knows the whole game and knows how to play it. Um that is that's something that uh that at the limit I think is is very hard to completely eliminate. and um assume that this is another reason why lenders choose your platform because you help them to mitigate using AI against fraud as well. >> Yeah. >> Yeah. Yeah. For sure. I mean I think the history of traditional banks moving into online uh online products where it's not a customer they've had a long relationship with. It's what it's like internet walk up traffic. I mean that is a history where you talk to any bank and they'll tell you that there are just so many cases where that ends in these you know super high fraud rates. You know we could be talking about like 5% of all the loans you've made turn out to be to fraudsters. And of course that's just crushing from an ultimate financial perspective. And so I think a lot of the reason that um traditional banks are reluctant to make uh new unsecured loans on the internet to strangers that they don't have a relationship with is there's just such a high chance that something something goes wrong with that and um and we give them the ability to do that. Um which is is a big thing because of course you want new customers and new customers today are so much more likely to want to come via the internet instead of walking in the branch. >> Amazing. Um, and we'll swing back round to the Q2 numbers because they're really great. Uh, revenue up, loans up, um, more automation. What changed in your strategy? Where does where's this come about from? Um, yeah. Could you just give us more details about that? >> Yeah. Um, so as I said before, the single biggest source of growth for the company comes from wins in AI. And um one of the big wins that we had in this past quarter was what we call the uh an improvement to the meta model architecture. So our core underwriting model is a um is actually an ensemble of different machine learning models. That means is there are different algorithms that are used to uh to learn from training data and make predictions about whether a particular applicant is going to repay or not. in fact uh whether they're going to uh default or prepay in in each time period of the entire loan's life and we use uh these different kinds of learning algorithms and then we uh blend the results together to get the final output. Now when we do the blending we actually are passing the outputs of the first layer of models into a second layer that we call the meta layer. The meta layer has also machine learning algorithms and those algorithms are responsible just for figuring out the best way to blend the outputs of the sort of individual base models and previously we had had neural networks as one type of model uh algorithm in the base layer but we did not have it in the meta layer and that's uh for various reasons but I think the most general one you can think of as uh the more sophisticated your model architecture the exponentially more training data that you need in order to make it work. >> And previously we just we didn't have quite enough training data and we didn't have quite enough uh compute or latency tolerance in order to uh to have neural networks which are one of the most powerful types of learning algorithms in both the base and meta layer. Uh and in Q uh in the last quarter we were able to change that. So now we have neural networks operating at both layers of the model. And that architectural change uh created a very large boost to model accuracy and that boost to model accuracy unlocked these separation gains that allowed us to have higher approvals at lower APRs and that uh drove a bunch of growth. >> Okay. So it's it's mainly it's mainly from people who were coming to your site asking for loans but were getting rejected and this actually included them in the segment where they would be accepted. Yeah. >> Uh some of them uh getting previously were rejected. Some of them previously were getting an APR that was too high for them to want to take out. >> Yeah. Got you. Uh which is really interesting. So it was coming from traffic and demand that you were already receiving. You just Yeah, like you said improving the model. So um how much you me mentioned earlier because you you spend fair bit on marketing now. How is that a large percentage of of revenue and is that increasing and how much is that becoming part of your business to generate demand as well? >> Yeah, we we certainly um we certainly spend on marketing though one of the cool facts about the company is that our CAC the the cost of acquiring customer has actually been uh loosely speaking trending down uh over uh over over the years. And that's um kind of the opposite of generally what happens in these businesses where people think if you're a business where you're just kind of like skimming off the top. You found some little niche of of people that were underserved before >> and you reach those people and then it's like you want to reach any more people, your CAC's generally going to go up because it gets sort of harder on the margin to find the sort of incremental that person because you've already picked off the lowhanging fruit. And we actually have experienced the opposite dynamic where we've been able to grow the kind of pie enough over time that I think the rate at which we're growing the the population of people that we can be the best rate and best process for is actually happening faster than the rate at which we're kind of depleting the lowhanging fruit from a targeting and marketing perspective. And so we've been able to actually bring the um uh the the sort of CAC of our business down um and uh and all while sort of increasing our level of investment there. So I would expect that to to continue. Um I I think we don't have actually an aspiration for like we don't need from a business model standpoint for cap to be any lower. We're perfectly happy to be where it is or even higher is fine too. But um but we want it to be a we want to be able to keep investing more and more dollars in marketing and reach larger and larger share of the population as we have the right uh and best products for uh for more people in more situations. Mhm. And um I'm just jumping around sorry because I I remember things from earlier in the conversation but uh we were talking about the AI algorithms again uh the models. Um how do we make how do you make sure that it's being fair between two similar borrowers so they get treated the same way? >> Yeah. >> Well, my puny answer is that only humans are capable of treating two similar people different ways. It's not something that like a computer or an algorithm can do. you know, computers are like really good at just kind of like um following the data. And if you know, the data says two people are the same, then you know, the the model's going to treat them the same. That's, you know, all all cares about doing. >> Um but I think in a more serious sense, I think there's a really interesting debate to be had about, you know, what what does it mean for two borrowers to be similar? And I think this is the essence of um of the question that I think you have to answer. And um one of the one of the moes if you will around our business is that we've spent so many years and so much uh research time on on this particular question we've done it in partnership with major regulators like the CFPB. We've done it with uh major researchers at universities who are some of the kind of leading uh academics that publish papers on questions like what is fairness? How do you measure it? How do you test it in the context of a machine learning model of um these sort of like neural network models? And um and I think the gist of it ultimately that we get down to is the the similar the dimension of similarity that really matters is willingness and ability to repay a loan. Um at the end of the day, you know, when you say two borrowers are similar, is it fair? What you really mean is if two people are both similarly able to repay a loan and you treated them differently, then then there's something wrong. That's not a thing that I think we as a society want. But I think sometimes people have confused the input and the output or the X and the Y on this question where they say, well, what if two people with similar FICO scores get treated differently? And my answer is they should absolutely be treated differently because if unless the FICO score is the beall endall of whether you're willing and able to pay back a loan, it should not be the dimension on which we measure similarity for purposes of fairness because ultimately what we want to know is are are you going to be able to pay back this loan? And I don't think it's in anyone's best interest for us to give loans to people who can't pay them back and ultimately, you know, going to end up in bankruptcy or another kind of bad financial situation. So then the question is okay well if what we really care about is finding similarity with respect to uh ability to repay a loan then what we actually need to first have is the most accurate possible model for knowing if someone is going to be able to pay back a loan. And so you really don't want to benchmark against a really simplistic uh sort of traditional model. You want to benchmark against the most accurate available model that you've got and use that as the source of truth for likelihood of repayment and then make sure that you're treating uh those borrowers uh in in a similar way. >> And slightly related but um if we look at general macro conditions uh which affects people's ability to pay back loans or you know and vice versa when it's you know in different kinds of macro conditions. How do you um how do you prepare for you know or mitigate against loan delinquencies in bad macro conditions and then you know change that when it when things get better? Is that is that all part of the model that does it does it analyze economic conditions? >> Yeah. So one of the things we've done over the last couple years partially in response to some of the learnings we had from 2022 or or thereabouts is that macro is really important and rather than ignore it we should figure out what part of the problem is tractable um and solvable uh by AI and be the best in the world at solving that particular part of the problem. Now the reason I I describe it as what particular part of the problem is solvable is that uh in the broadest sense I think macroeconomic conditions are of course very subject to kind of idiosyncratic singular events that don't tend to have a long history of repetition. So meaning like if you think about like COVID as a macro shock, well there's never been a macro shock like COVID before and u of course all kind of AI machine learning are based on training data that have sort of historical bases and you you know you're just not going to be particularly good at predicting that. Now I don't think that's a weakness of AI per se because I don't think there are like human experts that uh you know and if you ask them in 2018 we're going to tell you that in 2020 we're going to have the COVID pandemic, right? just like that was a kind of um black swan type event. But unfortunately, of course, those things do happen in macro. And so what we eventually figured out was that like you're not going to be able to predict all of the kind of idiosyncratic black swan events, but you can be the most precise and the most quick to respond uh to to any particular set of macro conditions and have it properly represented and priced in the model. And that actually is a pretty big edge because it turns out that most of the ways that people measure macro are delayed by a fairly significant amount of time. We think of it as often with a three to six month lag. And that's because if you think about what are the indicators of macro? Well, the Fed publishes data. The Fed publishes data on the past quarter in the middle of the next quarter. >> Yeah. >> And oftent times they end up revising that number for the next two to three quarters. >> Yeah. So to get like a real kind of estimate of what's going on in the macro, like you probably need to wait n months before you really know, you know, what was the employment rate back in uh November of 2024. Like we're probably just really getting a clear picture of that today. >> Yeah. And um you know this there's a similar uh but less extreme story even if you look at credit data where maybe you can go to credit bureaus but generally like people are looking at things like 60-day delinquencies as a key metric or 90-day delinquencies and well guess what you have to wait 60 or 90 days to measure 60 or 90day delinquencies. And so you're you're just always lagged by 3, six or nine months uh when you're trying evaluating macro. And so we invested in building what we call the Upstar macro index. And actually behind the scenes is this whole suite of other tools uh that you know will figure out the right ways to surface over time. But what we are um invested in doing is figuring out how we can make it so that we are just the fastest uh to respond and the most precise in responding. And our models today now have something that we internally think of as sort of this individualized um macro component where we can look at any arbitrary set of characteristics that you've got. Maybe you're a high-income person. Maybe you're a low-inccome person. Maybe you're a person who works in nursing. Maybe you're a person who works in the government. Maybe you're a person who works in a tariffsensitive industry. Maybe you're a person who doesn't work in a tariffsensitive industry. And the model can pick up uh trends that are individualized to that particular set of characteristics that you've got where if we see for example that maybe tariffs sensitive industries are being more negatively impacted and uh defaults are just starting to creep up there. When I say defaults, I actually mean you know the first signs of delinquency is starting to creep up. Um the model can immediately pick that up and respond uh in real time and it doesn't require uh people to specify like oh we're interested in you know bucket A and bucket B and bucket C because you just don't know what the next macroshock is going to affect precisely and you want that to be something that uh the model can figure out for itself. And so that's that's how we've architected the model and we think it's a very powerful solution for being the the fastest to respond, most precise to respond. But again, we're not we don't pretend that we're going to predict the next co so to speak. >> Yeah. Yeah. That's amazing. That's pretty um it's pretty incredible how many areas uh there are to do improvements in the way it used to work. Um I mean it just makes sense, right, that you should be quick at reacting to things like this. >> Kind of incredible. So, as you're picking up in all these new areas as you get deeper into the the loans industry, um, what are you most excited about over the next 5 10 years in lending? Most excited about over the next 5 to 10 years. Well, you know, I think in um uh I think in probably as we were talking about earlier, probably in 5 years time, I think we'll have successfully brought uh AI lending to the parts of the consumer credit market that actually are big and matter to people. Yeah. So that's, you know, really I'd say home and auto and uh and some flavor of revolving credit are the three big categories. And I think when that happens, I think there's going to be a seismic move in uh in how people perceive the relevance of AI and lending. Today, it's this weird niche thing that's just like maybe this this cute little company Upstart does. Maybe they'll go away. Who knows? Or maybe it'll stay in there a little. Um but I just really think that the perception around it is going to completely change uh when you bring it to these these much larger uh markets. And u and I think when that happens, I think everybody is going to care about this. And the end result of that is I think you're actually going to asmmptoically approach uh the real levels of access to credit that should exist in this country which as I said earlier is something like 80% of Americans should have access to bank quality credit. They should have it on demand with virtually no work required of them. And the other 20% should have a clear and defined path of how to get from where they are to where they want to be. >> And what's stopping up start doing this attacking those areas earlier? Uh we're we're working on it. Uh as I said, it was you know started in the home world on on Heloc. Of course, each of these you know you you have uh there's there's a lot of work to to do. Um and some ways each market has uh work that's specific to that market. So you think about like how do you how do you actually become relevant in the auto business? Well, you have to be in car dealerships because that's where almost all auto loans happen. But car dealerships are, you know, their own uh very specific industry with the way they work. Um, and frankly, I think, you know, we've had to spend a couple years like learning the ropes of like, okay, uh, we have to really understand things like how do dealer markups work? How do, um, how do add-on products work? How does like gap insurance work? Like the things that dealerships really care about. Those were not things that, you know, certainly I I didn't know anything about them when I started this company. And I didn't know anything about them three years ago when we started in the auto lending business because none of those questions were relevant in personal loans. But if you want to be a relevant player in a car dealership, you need to really understand those things. Your product needs to be very good at supporting those. And uh and so th those are, you know, those are things we've kind of figured out and iterated towards and um and and gotten a lot of great wins on. Same in in home. I think a different set of challenges that are home specific um but that we're working through and again I said earlier Heliloc has been a great sort of learning and testing ground for us to really figure out like all the mechanics of how you can most uh most quickly and most efficiently process figure out leans and titles and uh co-barorrowers all of the pieces that you really need on a home loan. >> Awesome. And so do you think Upstart will it will sort of exist as a mortgage broker sort of way like it will be one of the bigger players in that space and just you know I think that they they're the ones facilitating demand into different uh lending platforms at the moment or would it work differently than that? >> Uh we certainly expect to be a big player in in home loans. I mean home uh home loans I think in some sense are when people think about credit uh why why do they care about credit? It's because one day they hope to be a homeowner. They hope to be able to, you know, buy buy a house when when they need it and probably they hope to be able to get the loan quick enough that they can actually close on the home that they're trying to trying to buy instead of, you know, running out the clock or someone else getting grabbing the house. >> Um, so, uh, in that I guess, uh, in that sense, I think it is inevitable that home is going to be a central piece of what we care about here at Upstart. It's just just so big, so fundamentally important uh, for the consumer. Now within that ecosystem, you know, there's a whole bunch of different ways that you can technically slot into the ecosystem from, you know, are you are you are you broker, are you correspondent, are you technology provider, are you uh the originator? Um and um and I think we're not particularly uh dogmatic about uh which particular sort of um way uh technically we're engaging, but ultimately we want it to be the case that if you if you come to Upstart uh that you you have a guaranteed way to ultimately get uh get uh the home loan that that you need. Uh, and I say when I say guaranteed ultimately I mean either you can qualify today, you can get it instantly, you can get it with high predictability really fast, or if you can't qualify today, we can uh we can help you figure out what needs to be true so that you can qualify and so that you can get uh the home loan that that you need to sort of make make your life happen. And um as long as that's true for the consumer, you know, exactly who's doing the originating uh and um you know, how how it interacts with the GSC's and all that, that can be figured out on on the back end. >> And do you um foresee moving into other regions in the future? I believe you're just US at the moment. Um or is that I mean the US is such a gigantic market obviously a lot of companies stay here for a very long time just because it's big enough to grow very big without going somewhere else. Yeah, we we are exclusively uh US-based today. Um I I I would say um uh we never say never about it. Uh we certainly don't have any immediate plans to be outside the US. It is such just the such large part of uh the world economy, but also I think in credit in particular, consumer credit, consumer credit is so big in the US compared to outside the US. It just sort of really makes sense to start focus there. But um but I do think there are really interesting credit access problems outside the US that if you're thinking of the problem as where in the world is there the very worst access to credit where there's the most amount of kind of undervalued consumers that aren't being served right today that could be served. Um you know certainly there's a lot of interesting places outside the US and you know we could be there one day. >> Awesome. Um I've got a quick lightning round if that's all right where I'm just going to ask a few questions and >> Okay. >> A very quick answer to it. Um, what is the bigger bottleneck for uh improving AI lending? Better data or more compute? Um, in the short to medium term, better data. Uh, in the long run, it's going to be latency. Um, so a version of compute, not exactly compute, but kind of uh kind of like compute. >> Yeah. Okay. Um, one input you wish every borrower had to share with you. Um, I would take I would take everyone's driving records. >> Okay. Uh, by 2027, which LLM Oh, will LLMs be part of underwriting? We sort of discussed. >> Yeah. >> Um, what's Upstart's biggest moat in 2030? Data or distribution? >> Data. >> By 2030, what percentage of loan decisions will be made by AI? I think we're the wider market here. >> 100% of consumer. >> Okay. Um, one job in finance that AI will create not kill or in tech you can widen it. You know, I think that I think there's a real space for uh for more um for more financial counseling. Um I think that that is something people want with an emotional uh human touch to help them work through >> uh their plans, their life plans, how it relates to money. I think money is a, you know, it's a it's a it's a stressful thing for people and I think we can uh make that process uh we can make it quick, we can make it easy, we can make it delightful, we can make it, you know, your rates low. Um but uh but I actually think that if you take away all of that stuff and make it so instant, so easy and you know at at great prices for people at the end of the day, you still need to decide how you want to live your life and you know what kind of lifestyle you want to support. Do you want to spend more now, spend more later? And uh and I think ultimately those are human questions. >> Yeah. And last question, one non-obvious book or paper that shaped your approach to business. >> Yeah, there's this um there's this really long super nerdy um uh paper written by this guy Cliff Aznes who runs a fund called AQR. And um I I I don't quite remember the name of this paper but he the author of it and it is a paper on uh the ways in which markets are efficient and inefficient. And the beauty of the paper is just that you really see like if you want to think rigorously about efficient markets and uh therefore inefficient markets. um you have to think really really hard um because the um some of the smartest people in the world uh have worked on this problem and um if you want to find opportunities to generate alpha to do sort of do better than market do better than sort of uh the status quo uh it's not something that you know you can just um you can just you could just sort of do on a whim it's something you need to take really really seriously and I think the ethos of that paper has stuck with me for all these years of building upstart >> awesome thanks Paul um It's been a pleasure to have you on the show and really interesting to dig into Upstart. It's definitely one of the companies to watch and I'm sure everyone will value this interview a lot. Um, thank you so much. I don't know if there's anything else you'd like to to leave with our community. >> No. Awesome. It's been great to be here. >> Cheers. Thanks, Paul. Have a great day. >> Hey, you. Bye.