Please welcome VP of Investor Relations, Sonya Banerjee.
Good morning. Good morning. Can you hear me? Just making sure. Welcome to Upstart AI Day. My name is Sonya. I'm the Head of Investor Relations for Upstart. On behalf of the entire team, we are so glad that you were able to make it today. If you're here in the room with us or joining us on the webcast, we appreciate your engagement and your time. We have a fantastic morning planned for you. We're going to start things off by talking about the significant opportunity in front of us. We're going to tell you a little bit about why Upstart AI is uniquely valuable. We'll take a short break, and when we come back, we'll talk about some foundational topics, like why Upstart is a category of one business, excuse me, and all of the steps we've taken to drive profitability and resilience.
We'll wrap things up with an executive Q&A that will feature our Chief Risk Officer, Annie Delgado, as well. The presentation for today will be available on the IRA website, as will a replay. Okay, before we get started, it's the good stuff. I need to share our legal disclaimers. We will make forward-looking statements related to our business strategy and performance, future financial results, and guidance. These statements are based on our expectations and beliefs as of today, which are subject to a variety of risks, uncertainties, and assumptions. Excuse me. They should not be viewed as a guarantee of future performance. Actual results may differ materially as a result of various risk factors that have been described in our SEC filings. We assume no obligation to update any forward-looking statements as the result of new information or future events, except as required by law.
Our discussion will include non-GAAP financial measures, which are not a substitute for our GAAP results. Reconciliations of our historical GAAP to non-GAAP results can be found in our, excuse me, in our AI Day slide deck, which will be available on the IR website. With that, we're ready to get started. Please join me in welcoming our co-founder and CEO to the stage, Dave Girouard.
Good morning. Good morning, everybody. Welcome. I'm excited. I'm really excited for this. We've been excited for this for some time. Hopefully, put a lot of work into this. A lot of fun we're going to have today. Maybe the first issue that is worth talking about is why we're doing this. You know, we're 13 years in as a company. We're four and a half years in as a public company. This is our first Investor Day. Why did we decide to do this now? I think, you know, there's probably a couple of reasons. One is that I think Upstart continues to be, in many ways, a misunderstood company in terms of what we're doing, how we do it, where we think we're taking it. Partially, that's, you know, on us to be better at describing what we do.
I think it's also a bit of a function of the industry that we're in. Lending is this 5,000-year-old industry. I just think there will always remain a bit of skepticism that any sort of technology, whether it be AI or anything else, can actually fundamentally change the economics and the fundamentals of lending. I think we're here to share our view on that and tell you something very differently. Anyway, we're excited. Hopefully, we're going to make really good use of your time. I appreciate that those in the room here today have dedicated a lot of their day to be here with us and those online as well. We'll hopefully make this fun and informative. Really only have, I'm going to try to watch the right side of the stage here, three goals for today.
First, to help you understand, you know, the size of the opportunity, hopefully to provide what I think is a sort of simple and intuitive framework, because we think the opportunity, again, is probably misunderstood in how big and how awesome the potential is to bring AI to the domain of credit and lending. The second, to understand why we think Upstart will win, why we will be a very large participant in this transformation of this very large industry. We really believe the position we have in the market today, along with the technology that we have and that we're building, the business model that we have chosen to deploy in, and more importantly than all of it, really the team. I think we are the right organization to be a big winner in this very, very important market. Finally, this is an investor conference, after all.
We would like to convince you to buy our stock or hold our stock, hopefully not sell our stock. You know, I think if you're going to do that, if you're going to be part of this thing, we really want to have you have a thorough understanding of who we are and what the company's doing. I think when I think about this day and how I hope it plays out in the future, for many of you, I hope this is the day that the light bulb goes off. They kind of go, "Oh, I get this. I've been sort of following this thing, but maybe I didn't quite get it. Today, I get it." That's what I hope we'll do.
Whether you buy or you sell or you hold or you upgrade or downgrade, whatever it is you do, I want this to be the day you will remember for many, many years from now. That is a high bar. I appreciate it. Let's get started. You are going to hear a lot about our technology and our business model. That is kind of the core of what we are doing here today. I want to start with something different. I want to start with what is in our heart. If I am going to ask you to buy our stock and put it in a vault and hold it forever, you certainly should know who you are getting into business with, what you are becoming an owner of, and how you can expect us to behave and act as a company in the future.
It really starts with, you know, why we do this. I just have a few things I want you to sort of understand fundamentally about Upstart. Number one is we did not start this as an AI company. It was not a word we used or a thing we thought about. We simply thought about a problem we were aiming to solve, which was improving access to credit. A system we see that we saw was enormously important and, in many ways, archaic and dated and in underserving those it means to serve. Along the way, of course, AI became the tool of choice and the way that we see a path to a radically better place. Also at the same time, we were not a, you know, a 2022 November 2022 bandwagon jumper, if you will.
That was not us either, who just sort of attached AI to their name. We started using the term AI and machine learning to describe what we do in 2017. I think it was because at that time we had reached a level of sophistication. We felt that it was warranted. It made sense to use these terms. We were also very cautious. We partner with banks and credit unions, and we have a lot of very serious regulators. In some sense, you know, it was a leap to start using words like that in such a conservative industry. We felt even at that time in 2017 that we needed to lead, and we needed to be the ones to advocate for this technology because we saw, even at that relatively early phase, what the potential was.
But important to say, first of all, you know, we started, we came here to solve a problem, not to start an AI company. You know, we do not use the word mission-driven. You know, this is a word you see in a lot of websites. I kind of, these words just do not flow off my tongue really easily. It is not that I do not want to be mission-driven per se, but I kind of feel like if you are using those words, in some sense, maybe you are not. I do not know. It is just, it does not work for me. But I do like to think about alignment. And this statement here really describes alignment.
If you, when I think about Upstart, what I like to think about is in our most incredible successful version of us, everything we show you today about our vision and where we want to go becomes true times 10. Would that be a fundamentally good thing for the world? What I can say sort of with purity of heart and clarity in my mind is, yes, it would be. It would be a much better world if credit was radically less expensive, radically easier to get. I do not think you can say that for a lot of companies. I actually think there are a lot of parts of our economy where it is more problematic than that. Think of credit card companies. Is it really better that every single one of us is carrying a large credit card balance month to month and that we are paying off the minimum?
Now, that's the final state of a credit card business. Or, you know, think about the way the food economy is working. You know, cheaper calories that are empty and hypertension and diabetes and all the things that's gone on in the American health sector. Or think of social media. Like, honestly, some really good things have come from social media. So I'm not really here to bash it. Do we really want to spend more and more time on our phones connecting to people digitally, wearing things on our faces and doing all sorts of, you know, I'm not sure. I'm not sure. I feel when I think about Upstart in terms of alignment, I feel super good about that. The other thing you have to know, and I do use the words here, mission-driven.
I think after 13 years, you could also say founder-led, because at this time, they're the same thing. At this point, when you're a founder-led company at this long, it is a mission-driven company, of course. These companies will stop at nothing to achieve their goals. That is what you really need to know about us. We are not deterred when things get hard or difficult. We don't quit when there's like a shiny new thing that comes along that looks really attractive. We aren't sort of pushed back by the doubters. The doubters don't bother us. There are plenty of them out there. If anything, it fires us up. That's the nature of a company like Upstart, is we're not going to quit. We've chosen to do something extremely difficult. There's no recipe for it. There's no precedent for it as far as we're concerned.
We're going to find our way along that road. But we're doing something enormous and important, and we will not stop doing it. I want to talk about the market size for a moment. I mean, maybe it goes without saying that lending and credit as an industry is enormous. It's this huge part, this huge sort of substrate of our entire economy. It's a very meaningful part of all of our lives. It's enormous. You know, across all the different sort of sectors that you'll see lending in, this isn't all of them. This is just some of the largest. I think our view is the sort of machinery underneath all of this industry in the next 10 years will be replaced and upgraded by AI.
All the origination and servicing and collections and all the pieces and parts of the machinery that make this business go. $25 trillion, I think, per year originated. I think that's probably an understatement. We're just trying to put a sort of relative size to it. Enormous. It will all change, in our view, probably within a decade. How much of an opportunity is that for the company that powers that, that provides that machinery? I mean, we're not going to do fancy, fancy math here, but we think a reasonable take rate, you know, it's different by credit product. It depends whether you are sourcing, you know, acquiring the borrower, whether you're originating the product, whether you're servicing, collecting. It can vary.
But we think sort of a 5% take rate on the entire industry is what is sort of available to the platform, in our view, an AI platform that powers it. We're talking about in excess of a trillion dollars in revenue. We do not have to, like, you know, fine-tune that $1.1 billion-$2 billion, who knows? It's an insane amount of money. It's an incredible opportunity for that company or companies that can build the platforms that power it. Just to speak to AI for a moment, I mean, if you've been following this, and of course, you cannot, unless you have your head in the sand, you cannot not follow it. The progress of AI, really, as it's become sort of in the public sphere, in the public conscience, starting in 2022, the progress is just astounding.
Every day, every week, there is something that they go, "Wow, I thought that would be really hard to do. I thought that would take five years, 10 years." Boom, it shows up. We're not all the way there yet. I think if you just sort of follow the technology, it's kind of crazy. It's pretty obvious from a commercial sense, the way that work gets done is going to be radically different in pretty much every industry. I don't know that we could name an industry. I mean, this is, if you sort of look at what we have here, this is most of the industry of the world. There's not that many that are actually missing here, and they could be added to the slide. The interesting thing is, you know, what is the common thread? What is AI doing?
Everything from education, media and entertainment, defense and warfare, manufacturing, financial services, all these things. What is the common thread? AI is making radically better products at radically lower cost. That is the common thread. It is just an incredible enabler of a transformation of virtually every industry. What it means and how it plays out, not totally clear. Today, of course, we are going to focus a little more. We are going to talk about financial services, a very specific part of the market. Also, what I would say is a very important part of financial services. Lending and credit is the greatest source of revenue and profits in financial services. It is an enormous part of an enormous industry.
Now, credit and lending, I kind of use both terms because it depends whether you're talking about from the consumer side or the lender side, is a first-tier candidate for AI disruption. Now, why is that? First of all, you can think of it this way. Every time one entity advances credit to another, they are predicting the future. They are predicting what happens next. This happens millions of times every day. Somebody is advancing money with some assumption or some model or some thought or some belief about how it will come back. You know, what's at risk? At least $25 trillion every year is at risk because of that. It's just an almost like picture-perfect thing that AI does, which is better at predictions.
We will also say that there is almost unlimited amount and types of data and approaches that can be used to make those predictions more accurate. That is sort of, in our view, kind of a given. You put that together and it says, "Hey, wow, this is an incredible opportunity." You do not, as you sit there today watching us, you do not have to trust me. I did what any student of the world does, any college student. I thought I would say, "What does ChatGPT think about this? List the global industries that are likely to be most impacted by AI." This was a single query. I did not do any kind of vibe coding. I literally did this in about 10 seconds. This is what I got. Number one, healthcare.
This is truly incredible when you think about healthcare and the state of the world and the state of where it could be. Diagnostic, drug discovery, personalized medicine. It is not crazy to think that in a decade, the poorest people in the poorest parts of the world are going to have better access to healthcare than the richest people in the richest parts of the world today. That sounds totally crazy, but I think it is entirely within reach, even within a decade. Nothing against healthcare. We do not need to compete. Number two, financial services. In particular, if you sort of look at, get right down to it, fraud detection, underwriting, customer service. I think the why is just as important too, because it intuitively lines up with how we think.
I don't know if there's such a thing as a world consensus on anything like this. I think ChatGPT is about as close as it gets. I don't think they're hallucinating here. Data-rich, regulated, already tech-heavy, and AI boosts speed and precision. Speed and precision. Remember those two words. Before we talk any more about AI, I want to talk about the problem, because when I say it's this giant opportunity, what problem are we solving? What are we trying to fix? When it comes to loans of any type, I don't care whether it's commercial, consumer, mortgage, student, it doesn't matter. There's really only two functions or features of a loan that matter: the price and the process by which I can get that loan. That's kind of it. It's a simple product, right? There's not all these other weird things you have to think about.
There's only two things that matter: the price and the process by which I get it. Let's talk about the first one. Let's talk about the price. Why do loans cost what they do? Why do they cost so much? What is all that interest paying for? This is a real simple view on it, but I think it might be helpful. First of all, customer acquisition. Now, what does that mean? It could be a bank branch you built 10 years ago that you're amortizing to this day and has staff and all that. That's customer acquisition. It could be digital ads you're running on Google or on Meta or anywhere else. It could be billboards outdoors. It could be advertisements on television. You have to pay for all that. Of course, everyone you advertise to doesn't get a loan.
In fact, even retailers have a problem with advertising because they do not know who converts and what works. It is even worse in credit because a lot of the people you advertise that like your product, you will not be able to approve them. You can just see how challenging customer acquisition in the world of credit is. Second, people and data. This is the hard costs of originating credit, right? You have to pull credit scores and fraud scores and this and that. You have to have people reviewing documents and dealing with it. There are bad actors out there that have to be kept away. There are very hard costs associated with every loan that goes out there one way or another. Finally, and maybe most obviously, credit losses. Turns out everybody does not pay back their loan.
The important thing to grasp here is, guess who pays for the people that don't pay back their loan? The ones that do, right? If you're paying back a loan, you are by definition subsidizing those who did not pay back their loan. That's it. That's what it's all about. In the end, there's profits. There's a profit need of the entity that's putting forward the capital, right? There's a sort of return on capital. There's sort of a risk premium, and there's a duration premium and other things. Those all define the expectation of profit of whoever's putting that money out there in the first place. That's it. That's why they cost so much. There's really nothing else to it. The question is, what can—I’ll move this way—what can AI do to impact the price of a loan?
I'm just going to show you our experiment thus far and what it does, because it's pretty incredible to see. Customer acquisition. Conversion improvements reduce CAC. That's how CAC goes down by converting more people that hear about this into borrowers, in our case. What we've seen since 2019 is 5x growth and a 50% reduction in acquisition cost per loan. Now, that's pretty astounding. Most businesses, as they grow, they struggle to keep their acquisition costs flat because the marginal customer always costs a little bit more. Conversion improvements, improving models change that dynamic. 50% reduction. How about people and data? Automation, of course, takes out a lot of that cost. Not just the people cost, which is obvious, but the data cost. Why? Because you spend an enormous amount of both people and data costs on people that don't convert.
The fewer of those that there are, the less it costs per loan. We have seen a 66% reduction in the people costs associated with originating a loan since 2022, just three years ago. 66%. Maybe the one that is most obvious, we have talked more about this in our past, is credit losses. Better models, as you will hear a lot about today, are better at avoiding credit losses. When that happens, everybody else gets a bit lower rate. More people are approved. A smarter model reduces credit losses, all else being equal. Enormous reduction in the subsidy that has to be carried by the payers. The one thing I want to say is you see the little yellow text there in the upper right. This is not the final state. This is not all that can be done. This is results to date.
There is no reason. If you have seen us, if you know our automation numbers, how they have just kind of gone and gone and gone, there is no reason all these numbers cannot go and go and go as models get better, as AI gets better, as the compute power gets better. These are not a final state. Very importantly, wanted to label this results to date. Let us talk about, oh, I should also say, sorry, I am sort of leaving this on the bottom. When all these costs go down, there are all these economics that are freed up. You know, do they end up in the lender's hands? Do they end up with the credit investors? Or do they end up at the consumer's hands? Ultimately, that depends how competitive the market is. In a very competitive market, most of that value flows to the consumer.
It creates a far better product, a radically better product, which is what I described to you earlier. Let's talk about the other side. You know, our view generally is the world is somewhat evenly split between people that really value a penny and will do anything to get a slightly lower rate and people who really want the easiest process, willing to pay a bit more for it. We do not know exactly, no, but our heuristics have always been the world's relatively evenly split between them. I already talked about the price. What about the experience? Why does getting a loan suck so much? Now, I'm pretty sure everybody here, almost everybody, has gotten a loan of some sort. A car loan, mortgage, student loan, boat loan, motorcycle loan. I do not know. I mean, I'd be surprised there's many people here who haven't gotten a loan.
If you haven't, bless you, you're very fortunate. But we've all experienced this. Why does getting a loan suck so much? There are a lot of reasons. This is just some of them. I kind of like this graphic because it looks like they're standing in line, which is sort of a perfect metaphor for getting a loan. See, I'm going to get at the end of the line. You just get a feeling for what it's like in today's world to get a loan. Endless applications. If you've ever applied for a mortgage, student loan, bless you. It's incredible. Documentation requests, right? Driver's license, passports, proof of residency, pay stubs, tax documents. Pick your pick. Doesn't matter. This could be a business loan, could be a mortgage. Doesn't matter. Employer verifications.
I mean, who doesn't want somebody calling your boss to tell them that you're getting a loan? Isn't that the most pleasurable part of this? Waiting. Waiting for a decision. If there's any sort of word synonymous with getting a loan, it would be waiting because that's central to everyone's experience. Then, of course, that all-too-convenient in-person closing. Need to come in there. Sign the documents. Anyway, there's other parts, but this is sort of the heart of why getting a loan is never a good thing. Why it's somewhere between, I don't know, going to the dentist and going to the Department of Motor Vehicles for an appointment. Somewhere in between those. The question again is, how does AI impact the experience, the borrowing experience? Again, we have our experiment to date, and we'll tell you what we're seeing.
90%, roughly, of Upstart loans have none of these. Not one of them. What you can start to think about is AI is taking us to a world, you're going to hear more about this later today, of always on credit. Credit is no longer a process. It's an attribute. It's something available to you with zero process, hopefully perfectly priced, but with no process whatsoever for every American. That's a different world. That's what we're building toward because that's all that really matters in lending and in credit. How are we going to do this? That sounds like a tall order, kind of persistently underwriting the entire United States, zero process, de minimis fraud, all these things to pull off. How do you do that? You do that with a foundation model.
That means a model that is state of the art, getting better rapidly, and in our case, extremely unique. Now, what is a foundation model? You probably know in the generative AI world about foundation models. OpenAI, Anthropic, and Meta all have built foundation models, X.AI. There are a lot of companies. I probably should not say Google, should I? I worked there for eight years. A lot of companies are building these models. Now, interestingly, there are probably eight going on 10 companies that are credibly building foundation models for generative AI. The interesting question is, you know, is it going to get commoditized? I mean, there is just so much going on and it is moving so rapidly. In the end, they all train on mostly the same data, publicly available data, and they are using somewhat similar techniques. They are differentiated, and they will fight it out.
It is really a slugfest. In building the foundation model for credit, there's just us. Our data is entirely proprietary. Our model is built and trained entirely on data that is produced from our system itself. There is no shortcut to building what we're doing. Paul's going to talk a lot about this foundation model and how it creates the differentiation that we're all here for today. I just want to sort of introduce the basic ideas behind it. Number one, separation. Separation basically means being able to separate good borrowers from bad borrowers and understand the relative risk of an entire population. This is the sweet spot. This is where we started. This is what we're known for. Sometimes people fail to appreciate how impactful this is to what we're building and what we're doing. Separation is the heart of it.
Paul's going to talk quite a bit about this as it deserves a lot of attention. Automation. I sort of talked to you about this. You know, when we started way back, 2014 or so, 2015, the average borrower had to upload three, four, maybe five documents, and they had to have a phone call with somebody. They had to go through these processes that sounded like that picture I had a bit ago. We started experiments to say, what if our models could do this better? What we found, by the way, is every document you asked somebody for, you lost 15% of the borrowers or thereabouts. This was not perfect science. We were a little startup.
It was roughly true that every time you added another document that they had to upload and share and go find and put in there, another 15% would disappear. Conversely, the fewer documents, of course, the better the conversion. Automation is very central to creating the experience that makes us tick. Calibration. Separation is not the only thing. You also have to measure and understand the impact that the macroeconomic climate is having on borrowers as a whole. This is something I wish we started sooner. We have put in the last two or three years incredible energy toward very precise and very timely handling of macroeconomic changes. You are going to hear about the impact of those updates and what we are doing. Calibration, fundamentally important to the foundation model. Generalization.
Now, you could sort of build a great model for one type of credit, one type of loan. I think that would be great, but it would not be, in my mind, world-changing, right? It would be sort of like an LLM that just spits out images, right? Very cool. A year ago, that was neat or two years ago. Ultimately, we need something that can handle all forms of credit. Moving toward generalization is a big part of our future. We're kind of doing it step by step. Our Small Dollar loan and our personal loan now use the same models. We've kind of taken a lot of the automation and sort of income verification kind of stuff from our personal loan product, moved over to our HELOC product. They're still independent models. A unified future is really a very different place.
With the advantages and the scale of a very generalized model can really begin to shine. The last thing I'll say about this is personalization. This is the new avant-garde for us. This is what we're just starting on. It's something I'm very excited about. You can think of this as Upstart going from predicting outcomes, which is all we've been doing, trying to predict, predict, predict, to actually shaping the outcomes and turning more of them positive. The obvious place we start with this is when we do servicing and collections. We're trying to help people get through the process, make good on their credit, make the whole experience good for them. There's another 100 times more we can do to personalize the experience to help in small ways, either a business or a consumer make a bit smarter decision along the process.
Just getting started now, we fundamentally believe that generative AI and agent-based technologies are going to be the centerpiece of this. You should expect to hear a lot from us on this. The last thing I'll say, which is maybe obvious, is, you know, why does this matter so much? Because every person that you can take who might have defaulted, might have made a bad decision, and get them into a better place, it's of course great for them. But guess what? It makes the rest of the system a little bit smarter and a little bit better for everybody else. That's the beauty of this sort of feedback loop that we have. I'm doing okay on time, but I'm taking more than I thought I would. We're okay. 2025 parties.
I first laid these out in February when we had our earnings call, and I kind of gave an update just last, just look at our earnings. I wanted to just share with you, I think it's important for us where we are to tell you what is the most important things we're doing and working on and how are we doing against those. These are what I said. 10x our advantage in AI. AI is the engine that makes us work, and it is sort of the differentiation. I believe we are miles ahead of anybody in this regard. We do not believe there are others trying to build a foundation model for credit. For us, it comes down to both sort of the process and the substance getting better. Process-wise, we've never shipped models faster and more efficiently.
Substance-wise, we just continue to land new innovations. We talked about embeddings just last week. Paul's going to talk more about that. This is what it means. It's how quickly you're innovating, and are you nailing the right things, and does the train keep rolling? I feel very confident that this is happening this year. Second, prepare a funding supply for rapid growth. We've gone through what I think is a wholesale remake of our funding supply in the last three years. I'm just going to talk a lot about this. We fundamentally went from sort of an at-will marketplace, lots of funding that can come and go, to more of a supply chain with locked-in partners that have alignment, that have business models that work really well with ours.
This is a very big change, and we're going to continue to build that toward our future with many products and lots of these really important partners, particularly in private credit. Gap profits, returning to profitability in the second half. You know, this is obviously fundamental to us. We were profitable as a private company. We were profitable for quite a few quarters when we first came public. That environment was very challenging for us, so it's been urgent for us to get back to being a profitable company. I think we're on the path, and I also think we're excited to kind of demonstrate the leverage in our business as the core product is regrowing very, very quickly, and then these newer products are also growing. We're a low-fixed cost business that has a ton of leverage in it. Finally, I've talked a lot about this recently.
Inside the company, we hear about this a lot. Giant leaps toward best rates and best process for all. You know, a few months ago, toward the beginning of this year, I said to every general manager, product general manager on our team that by the end of 2025, I would like you to see whether you can have the best rates for every borrower segment that you serve. By best rates, I would say the highest win rate among competition for the borrower segments that you serve. That's been the challenge. I am very happy to report, and I want to share with you some data now. In our core product, the personal loan, we've achieved that. What you see here is a lot of third-party data mixed together and some internal analysis to see how we do versus the next best competitor in three major segments.
Super Prime over here, sort of Mid Prime here, and Non-Prime there. What you see is by win rate, we today are already the best. There are a few things that are important to say about this. First of all, it is a snapshot in time. It is not guaranteed to last. We need to strengthen it. We need to reinforce it. We need to get better at it. This is not a permanent state of the world. This is what it looked like roughly in the last month, just to give you a sense of what this is. I will also say that these three peers that are in yellow, each are different companies. I think that is a really important thing to realize. We are playing a very different game than anybody else out there in the field.
So I also want to say, I think it's easy to be skeptical or to say, wow, Upstart is just willy-nilly lowering its rates. It's taken on more risk. It's maybe just giving up profitable, giving up its profits, doing unprofitable loans. None of those things are true. What you're going to learn today, and I'm about to turn it to Paul, is that this is actually the completely predictable and logical end game when you have superior AI and very strong business execution. This is the game we're playing. With that, I want to say thank you, and I want to introduce Paul Gu, who's my co-founder and Upstart's Chief Technology Officer.
Thanks, Dave. If alarm bells aren't going off in your head right now, they should be. Because lending is actually a terrible business, just awful. You know, if you're like a tech analyst, go away. Like, you don't want to be here. The reason it's so terrible is this, what we have come to think of as sort of the trifecta of things that you want in a business that you actually have historically almost never been able to get in credit. Did it disappear?
Okay. There were three things on the screen. Maybe we can get them back even. Maybe we can go back a slide. You're really out of a credit business. You want to be able to, like any business, you want to be able to grow fast. You want to be able to make money. Interestingly, in credit, you need to make loans that actually will pay you back. You need the credit to perform. Historically, it's become almost a truism that in lending, you actually can't get all three of these things at one time. There are some businesses that have managed to get one, some businesses that have even managed to get two out of the three. Getting all three at a time has been almost impossible. The reason that's true is the reason that Dave stated earlier.
Lending is a really old business. It's been around for 5,000 years. It's as old as human civilization. There are thousands of players in the market competing in an almost perfectly competitive way. It's almost a textbook business where you would expect it to not be possible to get profits without sacrificing something else that usually comes in the form of growth or credit. Usually, something has to give. That is as close to a fundamental law of the universe as there can be in a business, except for one thing. If you can actually change the underlying technology assumption in the business, then you can actually release the constraint that makes it impossible to have all three. Today, I hope to show you that what we've built is a large and growing edge in AI that makes this possible.
To understand how that could be, I want to start by giving you a little bit of a history of how consumer credit is done and some of the things that we've done to improve upon it. You see, underwriting credit is a very mature discipline, something that's been done again for a very long time. There are actually a ton of textbooks on it. It's something you could just go on Amazon, and you could essentially get an entire download of how this is done and how you should be doing this. There are very strong industry norms that are reinforced by a network of rating agencies, bank credit committees, loan buyers, regulators. They all tell you to do very roughly the same thing. For that, we can just open up one of these textbooks and see what you do to build a credit business.
Of course, the first thing you need to do is you need to figure out how you're going to underwrite your consumers. The very first step of building an underwriting model is called the preliminary scorecard step. For the preliminary scorecard step, you want to go and find 8- 15 variables. Not like 20 or 100 or 1,000, that would be too many. You want 8- 15 variables that you think are likely to predict someone's credit risk. Usually, these characteristics have some kind of intuitive relationship with credit. You could easily describe to your friend who knows nothing about this why they're being underwritten this way.
You are going to identify this set of variables. Then you will put those variables inside what is known as a logistic regression, essentially a linear model that when you stick the variables in, gives you sort of one weight for each variable. It assumes that there is a linear and independent effect between each of your, call it, 10 variables and the outcome of interest, usually whether the borrower pays the loan back. You might think you could just use that linear model, but no, because while you can read the text here, while it is technically possible to just use this linear model to make decisions, usually for most lenders, that would be operationally too complex.
You want to simplify that linear model down into what's known as a scorecard. Prior to working in this industry, I'd never heard the term, but apparently a very common term in the business. A scorecard looks something like this. A scorecard is essentially a lookup table. It tells you things like, if someone has had zero to one delinquencies on their credit file, then you would add 40 points to the amount of risk that you perceive in this loan. Maybe this person has had no inquiries in the last six months. That's really good. You're going to give them 50 points of credit for that. When you're done with this, you end up with a table of scores and a table of how likely you think the person is to pay back the loan.
On that basis, you will decide whether to give them a loan and what APR to charge for that. You will go ahead and verify their income, their identity, their employment. Most commonly, like Dave said, you will do that using documents and phone calls. That is the industry standard. That is what we set out to innovate from back when we got started in this business over 10 years ago. Let us see how that timeline played out. In 2013, that is sort of roughly the state of the world, how credit is done. It is actually very fortunate that we knew so little about the credit business when we got started. We were in many ways very young, very naive, because we at that time had zero data. We did not have any of our own training data.
We did not know that getting to what we now call a calibrated model, meaning loans that perform as you expect them to perform on average, and having a model with enough separation to drive that trifecta of growth, profits, and credit performance was going to be a very long journey. It took us over five years to get there. If we knew that at the time, we might have started a different kind of business. Here we are. I am going to take you through a quick history of the technology. I will just quickly highlight one thing that we did each year between then and now. I will do a couple of deep dives to hopefully give you an intuition for how it is that we have created something so different. In 2014, like I said, no training data, did not have anything, just had to get started.
The first thing we did was we looked at third-party academic research for things that might be predictive of credit that were different than sort of what was traditionally used. We happened upon education data, things like where someone went to school, what they studied, their highest level of education. In order to use this data, we had to come up with a whole bunch of clever sort of statistical tricks because, again, no training data, which meant we had to take all these third-party data sets, find ways to join them together, and subtract out the covariance. We made it work. We got a sort of V1 model out the door. It actually turned out that if you could only choose to have one variable in a model, it would not be FICO.
You would actually want to choose something like someone's highest degree or where they went to school. It's actually more predictive of credit outcome than FICO on a one-to-one basis, which is really, really interesting. Because of that, we were able to build a model right out the gate that achieved some very nice separation, at least for a segment of borrowers. Again, with very little training data, our first vintages of loans did not get properly calibrated, and it would take us many more years to get there. The next year, 2015, we started having just a little bit of our own training data. Importantly, that training data came with these alternative columns, these columns of alternative data that we were uniquely collecting. We began to invest in using that data a little more intelligently.
Instead of running a single linear model, as was the market standard, we created what's called an ensemble of models. That basically meant we built a whole bunch of different models. Some were linear, some were not linear, and each had their own strengths and weaknesses and things they were good at. We took a weighted average of these things based on what each model was good at and how good they were. That, of course, improved the model's predictiveness. In 2016, you heard Dave talk a lot about our ability to instantly approve borrowers without a whole process. We said, let's take some of the same techniques that we've done now in underwriting and apply them to the sort of automation of the process.
Of course, that had a lot of advantages, namely that you get a much faster feedback loop to find out if someone is a fraudster, if someone is lying about things on their application, because those applicants tend to default a whole lot sooner. You do not have to wait a full five years on a loan to find out if someone is a fraudster, because for better and worse, if they are a fraudster, they tend to not pay you back pretty quickly. It also was really nice for us that we started with alternative data because that alternative data turned out to be really useful in building models that could decide whether someone should be approved instantaneously. Consider something like you are trying to figure out if you need to go through the process of verifying somebody's income.
It turns out to be really useful to know that a person actually has a verified computer science degree from Carnegie Mellon if what their claim is is that they work at Google as a software engineer. Those are highly, highly correlated facts. Having one verified makes the other a whole lot more or less likely. In 2017, finally, and this is the year we started using the terms AI and machine learning more publicly, we finally had enough data to start using modern machine learning methods. We immediately ran into some problems. Those problems were that almost all of the modern machine learning algorithms, especially at the time, but even now, were designed around problems with complete data. By that, I mean the classic machine learning problem is a problem like classifying images of dogs or cats.
If you're classifying images of dogs or cats, one nice property is that you already know the correct answers. You usually have your data labeled so you know which ones are dogs, which ones are cats. You do not, importantly, have to wait five years to find out if a picture is a dog or a cat. Now, lending does not, unfortunately, share that property. In lending, you make loans, and it takes a really long time to find out if they were good loans or bad loans. We ended up developing a proprietary, now patented technique for adapting the major machine learning algorithms that were out in the open source, things like XGBoost, so that they could deal with this incomplete data problem. We started underwriting with it. I'll come back to this problem more in a moment.
2018, we took a lot of these machine learning techniques that we were now using in both underwriting and in the automation of the process. We started applying it to what we call revenue science. We started to optimize the trade-off between conversion and pricing. That became important because as our models delivered increasing separation of risk, meaning we were able to identify borrowers that we thought were much more or much less risky than traditional models thought, there started to be a large variance in what we call individual rate sensitivity, rate expectations, and how much value Upstart was creating for the consumer compared to their likely alternatives. It just became that for some consumers, we were so much better than anything else they could see in the market. For other borrowers, we were only modestly better.
Yet for other borrowers, we were still not the best option. The right amount of value to try to capture there is very different in each case. Some borrowers came with very different expectations depending on which marketing channels they came from and what kinds of marketing materials they had seen. We applied machine learning to this problem. All right. In 2019, this is a really, really important year for us. We built a framework that was very important and comes back to this problem of some of the unique challenges of dealing with lending in the machine learning context. I'm going to take you into this one a little deeper. We will step out of the timeline. Okay.
2019, if you go back to that time, that's sort of around the time when machine learning as a discipline and then sort of to this terminology of AI starts to really take off. This graph here is just a graph of the number of academic papers that are sort of ML or AI-centric that are getting published by year. You can see that the years that our business is getting built coincide with this sort of exponential takeoff in the amount of research going on. I think now it's sort of obvious that this is happening. At the time, we were around here. It was really, really starting to take off. You would think that this would have provided a massive and automatic tailwind for a business like ours that was using this kind of technology to try to reinvent something like lending.
There actually were a lot of challenges. You could not use almost any of these things out of the box. That is because of this problem that I described earlier, where loans have this long-term survival problem. That basically means you make these loans. It is a little different if you are making loans that are maybe like a week long or two weeks long. There is actually a good amount of innovation in these ultra short-term payday style loans because of that. When you get to these longer loans, you are confronted by this problem that your choices are either you could just throw out all the recent data. You could say, "We will only train models on loans made five years ago." That would solve the problem.
Now you've gotten rid of all of the recent years of data, which is probably, one, most of your data, and two, most of the most relevant, most recent data. That would be pretty awful. You'd have very stale models. Or you end up having to make some really simplifying assumptions. You could assume that if you've observed a certain amount of defaults at, say, month six, that there would be twice as many at month 12 and three times as many by the end of the loan's life. In fact, that is actually the industry standard. It's called you create a default timing curve, which basically means you assume some fixed ratio between when the defaults actually happen. You can train models on highly incomplete data. Now, earlier, two years before this, I mentioned we'd started to deal with this problem.
At that time, we solved some math problems to just mechanically unlock us using some of these algorithms. But we were not able to yet grasp all of their advantages because of some of these sort of core survival censoring feedback loop problems. In 2019, we finally did. We created something that we called the loan month model. A loan month model is basically something that instead of modeling a single terminal state of the loan, like do you repay the loan at all, we built an architecture that separately predicts defaults and prepayment in every individual month of a loan's life. That basically meant that instead of having sort of a single prediction we were making, we were now making for a five-year loan 120 distinct predictions.
That meant that we could suddenly start applying all the modern machine learning algorithms to this problem because each problem was sort of a distinct and completely discrete problem that would be deterministic in what its sort of resolution state was. Now, you might think, "Oh, okay, that's cute. Does that really matter?" The answer is yes, it matters massively. I'll just give you a specific example to illustrate the point. Let's take two borrowers. Borrower A has a lifetime chance of defaulting on their loan of 12.65%. Another borrower, Borrower B, exactly the same lifetime default probability. What you see in this graph are two curves. The blue curve representing Borrower A shows you their default probability in each month of a 60-month loan and the same for Borrower B.
You can see that despite ending up at the same place, this is a cumulative curve, the trajectories they take to get there are radically different. Specifically, they are different in the way that Borrower A is a fair bit more likely to default in the first year of the loan's life. Borrower B becomes a fair bit more likely to default in the second half of the loan's life. This matters a lot, not surprisingly. Now, if you use a traditional approach and you assume a fixed curve, then you are stuck sort of taking something in between. That is what would happen. You might have an average timing curve as represented by this gray line for your population. You just assume that is the ratio. That is when I expect that defaults actually happen without having a model that has the ability to sort of make these individualized predictions.
You would get the underwriting terribly wrong because the actuarially correct APRs for these borrowers turn out to be really, really different despite having the same probability of paying back the loan. Borrower A's optimal APR, meaning like the actuarially correct thing that solves for your target return with precision, is 16.7%. Borrower B is at 13.09%. That difference, 361 basis points, is a very large number. If you're taking a traditional approach here, you're essentially forced to be hundreds of basis points wrong in one direction or the other. With the loan month model framework, we were able to solve that problem. Okay. Coming back to this timeline, we're in 2019. We solved some problems. 2020, of course, you see our training data is increasing kind of exponentially now.
With the help of all that sort of background research and background technology development happening in AI, we start to be able to use more and more types of machine learning algorithms. In 2020, we started using neural networks. I think at this point, well understood the value of neural networks. We were able to start using them after having solved many of the same problems, having enough training data to actually make use of these things. One of the things about more powerful machine learning algorithms is that they're pretty useless and, in fact, quite harmful if you don't have enough training data to support them. Enough training data, enough background research, enough sort of proprietary development of how to make these things work and be compatible with the sort of lending problem, we were able to start using these.
We added them to our ensemble of models. In 2021, another area that we started applying AI to was compliance. That probably sounds really odd, a bit counterintuitive. Consumer credit is one of the most heavily regulated industries. That regulation, especially in recent years, but really forever, spans policy concerns from fair lending to explainability to safety and soundness in the banking system. There is a whole bunch of different regulators that care about each of these pieces. That body of regulation has been a pretty significant, major deterrent to players venturing beyond the traditional safe confines of the traditional credit scoring system. You know that if you stick with a traditional credit scoring system, that has already been vetted. It has been used for decades.
You're not going to be challenged on whether you are uniquely being unfair to some group or whether people understand what a credit score is. This has sort of been proven out. As soon as you venture beyond that, you're confronted by all of these problems. If you venture beyond that and you use these very new, very complicated models like the ones we were using, you have another problem, which is that almost all of the traditional techniques for managing these kinds of compliance risks did not immediately work out of the box. What we had to do was we developed a whole set of new methods, essentially machine learning-based methods to do compliance around things like fair lending and explainability. We matched the sort of complexity of our underwriting models with equal complexity and sophistication in our compliance systems.
And LDAs and SHAP are just two acronyms for some of these systems that we created. Okay. 2022. This thing is one of my personal favorites. I'm going to do a quick detour to talk about a thing we did there. Now, it's something that Dave thought you all wouldn't appreciate very much because it's just too nerdy. I think you're going to like it. We're going to spend some time on this. All right. This is actually not the cool thing. This is the boring thing. All machine learning models have something called the loss function. A loss function is just a fancy term for it's the sort of it's their goal. It's the thing that they're trying to get as low as possible. They're trying to minimize some amount of error.
This is a very generic loss function, something that anybody could just do. It is present in many of the open-source machine learning models. It is called log loss. This basically is just a loss function that lets you represent sort of good and bad outcomes as ones and zeros. Its sort of key property is that, well, you only get ones and zeros in this thing. If you predict something is a one and it is a one, it is really good. If you predict something is like a zero and it is a one, it is really bad. That is kind of what this loss function measures. It measures your sort of average error in sort of a log way. This thing is very nice. It has some really nice mathematical properties. It makes machine learning models really fast to run.
It has a lot of limitations when you apply it to lending. It comes back to some of the issues we were talking about earlier. For one thing, it does not have a nice way to let you separate the instances where a loan defaults really early and a loan defaults really late. They are kind of just treated the same, again, ones and zeros. It only has one type of one and zero. You are either stuck saying prepayments do not matter at all, or you have to say prepayments are just as bad as defaults. Obviously, they are not.
The third is it does not let you distinguish between the cases where you predicted that a loan would default really early and it actually defaults really late from a case where you predicted a loan will default really early and it defaults just a little bit later, meaning the sort of amount you are off in time does not get well represented, again, when you are in a world of only ones and zeros. Now, these all sound maybe like, okay, kind of like interesting. Again, how much does it matter? I am going to show you. Okay. I have to explain this graph a little bit. This is a graph of, it is a heat map.
What the colors are showing you is a measure of the sort of real error in your financial objectives, specifically the net present value difference of cash flows between what you expected and what you actually got. This color gradient here, the lower the number means the less error that you've got. The higher the number, more error. We've color-coded it. Blue means you're cold. Red means you're hot. You've gotten to a good place with less error. Of course, you want as little error as possible. The x-axis is the predicted default in month one of the loan. The y-axis is the predicted default in month 60. Remember, our models have to make 120 predictions for every underwriting. In reality, what's happening in the model is this thing, but you're on a 120-dimensional sort of thing. We can't visualize that, of course.
We have simplified it to two for purposes of the illustration. The point is, if you get to the red, then you have less error in the sort of true financial metric, the thing that considers all of the things like a default in month one costs you a lot more than a default in month 60. Getting it wrong by one month is a lot less bad than getting it wrong by 50 months. And sort of prepayment costs you a little bit, but not nearly as much as a default does. Okay. What happens if you use that generic loss function, that log loss thing? This is what happens. A machine learning model, as I said, follows a loss function. It follows it along a gradient.
It does something called gradient descent, where it's trying to find the values for the model that minimize the loss function. It follows a gradient to do that. This is actually what happens on the generic loss function here: the model starts out here. Its first guess of what the correct parameters are, meaning it's like, "Okay, hey, you said a predicted default in month one of 5% and in month 60 of 10%." That's its sort of first guess of what the "right answer" for this borrower is. Then it sort of says like, "No, no, no, that's not right." It sort of keeps on looking, keeps on looking, keeps on looking. It ends up over here. Okay. Pretty good. You definitely got from bluer to redder. You'll notice that its optimization path is just subtly wrong.
It doesn't actually get where you really want to go. The reddest location is over there. You can imagine that when you expand this to 120 dimensions, there's a lot more space to be right or wrong than in just two. Now, the proprietary loss function that we developed, and this was sort of quite a project. It was a bunch of math and also a bunch of engineering. This actually optimizes along this path. You can see this gets you a whole lot closer to the place you really want to be than this does, a very significant difference in both the amount of error you end up with and the actual predictions that the model makes. This piece of work, as I said, a very significant piece of work.
It took sort of both sort of what we think of as math work and engineering work. You have to sort of first figure out how to represent the sort of correct loss function as a loss function. The second is you have to figure out how to make that thing compatible with gradient descent and sort of fast optimization in a machine learning algorithm. That is work we did back in 2022. All right. In 2023, we focused on a bit of a different problem. That was the challenge of applying machine learning algorithms to a business where you have large macroeconomic fluctuations. Therefore, the patterns of repayment tend to change over time. Again, you do not really have this problem in other types of machine learning. Dogs always look like dogs, and cats always look like cats.
In lending, the sorts of people that default and the aggregate level of defaults tends to change quite a bit over time. That was something that machine learning models naturally did not handle well because they treated time as just an unimportant dimension. It was something that you could just not worry about because all history was equal. We invested in making our systems adapt dynamically to changing macro conditions.
That meant that you would have a system that if there was a change in the world, whether it happened in the aggregate, like twice as many people default this year as last year, or something that happens at some kind of arbitrary segment level, like government employees are sort of under threat, or people who studied English in college are suddenly unable to find jobs, whatever it might be, any of those changes are now things that the model could natively and dynamically pick up without any kind of sort of need for Upstart staff to get involved. That essentially means the moment there is a statistically significant trend in the data, our models dynamically respond. Later, I am going to come back and show you how big of a difference that would have made if we had had that a couple of years ago. Okay.
2024, one of the forever problems in lending is the problem of adverse selection. The people that want your money are always worse than the people that do not want your money. This is just always true, kind of unfortunate, but it is what you deal with. In 2024, we came up with a way to mitigate this problem. We called that APR as a feature, meaning we literally gave our models APR as a feature to the model. Essentially, that meant that the output of the model, the APR that you want to assign the borrower, is now also an input to the model. The reason that is a partial solution to this problem is that one of the main avenues of adverse selection is when people are willing to take loans at APRs higher than they should be given their circumstances.
For example, if you have someone who seems to have a really clean credit profile, but they're super happy to take a high APR loan, well, there's something suspicious about that. That probably implies that there's some kind of hidden information that only they have that makes them worse than your model really thinks. The problem technically with doing this is that this kind of creates an infinite loop. If your output is an input, then the problem can go on forever. To solve this, we ended up running many parallel instances of our model at once so that we could actually make the problem converge and resolve in a finite enough time to satisfy our borrower while they're waiting.
In trying to do that, we had to make a very significant investment in low-level optimizations that we had not had to worry too much about before. Suddenly, we had to worry about things like latency, compute, memory, things like that. In fact, increasingly, as our training data size has grown and the complexity of our models has grown, these lower-level bottlenecks have become increasingly the thing that we have had to solve for. I am going to give you an example of one of the things we did in low-level optimization now. This is sort of a simplified example of roughly what our training data looks like. You have all these borrowers. They have some attributes from the time that you gave them the loan, things like their credit score and education.
Because of this loan month framework we have, for all of these borrowers, you have one copy for every month of their loan's life going up to 60 months. What that means is that there is actually a lot of repetitive data in the training data matrix because the data from the time of underwriting does not change at all. If we knew that you had a 714 credit score on the day you applied, that 714 never changes. We have up to 60 rows with exactly the same columns of data stored in them, while we have only part of the matrix that has any unique data. Each payment month, of course, you have a different sort of outcome of whether you pay back or not that is unique.
You've got now this interesting thing where you have this big matrix of training data. Some of it has a bunch of repetition in it. It gave us an opportunity to compress and save on memory. What we did was we created a new data structure. We call it a block sparse matrix format. Essentially, it involves compressing this data in a way where you only store the unique data or you only store data that is unique. The data that repeats, you only store once. You can see that things have been slightly shifted in the data structure, more data represented with zeros. We rewrote all of the sort of core machine learning algorithms to support this data structure natively. You can think of it as they recognize this data structure now.
That way, when you ingest this thing with all these zeros, it knows to sort of use the sort of unique data in all the places it needs to despite it not sitting in the matrix. By doing this, we immediately reduced memory consumption by a factor of 20 today, 20 today because it happens to be about the sort of average age of a lot of these loans, 20 months deep. Over time, that actually will grow to a memory savings of a factor of 60 because most of our loans are 60-month loans. Pretty cool win there. All right. We're almost to the present day, 2025 this year. We shared just a couple of weeks back. We've gotten really excited this year about our work with something called embeddings. We talked about that at earnings, so I won't belabor it.
Essentially, embeddings are a technique for finding meaning in unstructured data. They open up the door to using many, many more kinds of data than we previously used. Very excited about that. Okay. I've spoken a lot about what we did. I want to turn my attention to the question of whether it matters. Does it matter that we have all of these sort of machine learning techniques that we've developed over the years? Let's say for a moment you believe that we've built at least something a little different than what's classically done. To answer this question, does it matter, I'm going to refer back to this framework that Dave laid out, the sort of five ways that we apply AI to credit. I'll start with separation. Separation is the thing we've worked on longest.
It is the sort of figuring out which borrowers are more or less likely on a relative basis to default. I want to first give you a sense of how much separation actually is happening. The short answer is a lot. This graph uses some of the same third-party data that Dave referenced earlier. It is a random sample. Just so we could put it on a screen, we randomly sampled some 20-some borrowers. For each one, we took the range of rates that were observed by industry players, meaning online sort of lending platforms. We basically plotted in the gray bar the min and the max APR that this particular borrower saw. In the teal dot, we showed what Upstart offered them.
The thing I want to point your attention to on this graph is how often it is really surprising, even to me, how often the Upstart dot is the lowest or highest dot, very often outside the range entirely on this graph, which means that the rates that we are offering to people look nothing like the rates that they see anywhere else. Sometimes they are way higher, that we think this person is way riskier than anybody else thinks. Sometimes we think they are way less risky than anybody else thinks. It is very obvious that a large degree of separation is happening. I think it is about 70% of the time in this particular sample. We are offering very different rates at least. Of course, it does not tell you yet if that is good different or bad different, but at least very different.
Now, to find out if it's good or bad, we've got to start looking at some of the outcomes. We'll start with a single illustrative example of a borrower, and then we'll show you it with more data. A specific person five years ago applied for credit. These were some of the facts about them that a traditional model saw and what the traditional benchmark model that we hold internally for calibration would have said. This person had a mid sort of 600s FICO score, kind of on the border between sort of non-prime and kind of mid-prime. They only had about five years of credit history, and they only had about $60,000 of income, not a particularly prime borrower. This person, a traditional model, very likely turns down. Our version of a traditional model said their likelihood of defaulting is about 32%.
That's a model built using the textbook techniques that I described at the beginning, pretty likely to not pay back on a five-year large installment loan. Here's what the Upstart model saw, something radically different. We saw that this person had a JD from the University of Pittsburgh. They were working as a lawyer for the city of Philadelphia. Even though their income wasn't super high, they had a fair bit of stability to it. Their debt use was extremely low relative to other people in similar circumstances. We took, of course, those facts, ran it through all those models, and we ended up with predictions for, again, every month of a 60-month term for default and prepayment.
In aggregate, the model said, we thought, in contrast to that 32% likelihood of defaulting in the traditional model, we thought there was only a 2% chance that this person would default. Not only should we give this person a loan, we should give them a really good loan. What happened? This was our set of predictions for each month of default and prepayment, the black line being predicted defaults by month, the teal line being predicted prepayments by month. This borrower actually made 21 on-time payments before fully paying off the loan, prepaying in month 22, right around the time that our sort of by-month prepayment curve was going up. A success. This was a good loan to have had separation on. Of course, that's just one loan. Let's now repeat that exercise with a bunch of loans.
We've sort of randomly sampled a couple hundred loans back from 2020 so that they're fully played out. What we've done is we've gone ahead and ranked them using a traditional model up on top and an Upstart model on bottom. These little icons here are real people that applied for loans back in 2020. This means that this person was the top-ranked person by the Upstart model. That person was the top-ranked person by the traditional model. The colors tell you whether those people ended up actually paying back their loans or not. Gray means they paid back. Red means they did not pay back. With this set of risk rankings, the amount of separation that you've got, if you're using these models as a lender, you have some choices to make about where you set your approval rate.
Let's say that you do a simple thing. You say, "I want to improve the better half of people," and you improve half the population. In that case, with a traditional model, you have an approval rate of 50%, and you end up with a default rate of 9%. If you were to use the Upstart model instead, you have an approval rate of 50%, and you end up with a default rate of 5%. It's just sort of basic math because if you have better separation, then at the same approval rate, you can solve for a way lower loss rate. Now, usually what happens, though, is that it does not exactly play out like this because people want to solve for a certain acceptable default rate in their business. Let's say you wanted to normalize these.
Maybe as sort of a traditional player, you want to get your default rate down a little bit. To get it down, say you want to get it down to 7.5%, you have to cut your approval rate because mathematically, the only way to do that, given this is the best ranking you can produce, is to just pull that yellow approval box in until your approval rate gets down to 7.5% or sorry, your default rate gets down to 7.5%. To do that, you end up here with a 27% approval rate. A very significant reduction in your approval rate. By contrast, when you're running the Upstart model, because the degree of separation in your risk ranking is so much better, to solve for that same 7.5% default rate, you can bring your approval rate all the way up to 67%.
Again, this is just illustrative, but it is a random sample, so fairly representative, a very large difference in approval, approvability of the population at the same level of risk because of that better separation. Now, automation. This is something we've talked a lot about, but we started doing instant approval back in 2016, back in the timeline. Within a couple of years, it became extremely important to us. In 2018, it was 40% of our loans were fully automated. That number would keep going up to the number that you now see, which is about 90%. When you look under the hood, the reason that became so important to us is that there's a massive difference in the conversion rate of loans that are automatically approved versus manually reviewed.
There's probably no single thing that you could do that's more important to whether a borrower is likely to actually take your product than to automatically approve them. The conversion rate is about three times the level compared to if you send them through manual review. That's great news. The reason it's even better news is that one thing you might think about this 90% metric is that it's sort of done going up. It can't really go any higher. The sort of secret, weird, counterintuitive good news about having a really high conversion delta between people that are fully automated and people that are manually reviewed is that when we have 90% of our loans fully automated, it actually means that only about 75% of the applications are fully automated.
That is because they convert so much better that they end up being a larger portion of the end loans. It means that the opportunity is not to go up from 90%. It is to go up from 75%. We expect that this number will continue to rise as we pour more ML attention into automation. Okay. Calibration. This is a pretty juicy one, and I think a topic that is very top of mind because of the events of the last few years. Calibration is, by the way, just the term we use to refer to whether you get the aggregate level of defaults predicted right. It is sort of what if you were a loan buyer or someone that was holding the sort of end risk of the loans, what you really care about.
If we look back at what happened in the industry in the last few years, we see something really stark, a reminder that credit is a pretty terrible business. Because one of the terrible things about credit is that it's very subject to macro forces. It's very volatile, and you get these sort of times when everything's great and times when everything becomes terrible. That's really what happened over the last few years. If you rewind back to early 2021, we had a relatively low level of defaults. This is data for the whole industry, by the way, not Upstart-specific in any way. This comes from TransUnion. It's just on the entirety of what TransUnion sees for personal loans, credit cards, and auto loans.
In the last few years, we saw a 50% rise, industry aggregate rise in personal loan delinquencies, 40% rise in auto loan delinquencies, 70% rise in credit card delinquencies across the entire industry really in a relatively short time. This measure's start to end, but if you just measured even from 2021 to 2022, you would see that same rise. We got hit by that too. We got hit in starting early 2022 when that curve really starts to go up. The way that manifested is that our loans coming out of the Upstart ecosystem in Q1 2022 had a very large what we call loss variance. That means that the realized defaults divided by the predicted defaults on that vintage of loans was 59% in excess of what we predicted.
You can see that this loan underperformance persists for a whole bunch of quarters because as that graph continues to go up and persists, and our machine learning models take time to calibrate, it says, "Well, all history is created equal, so we're not quite sure. Is this the new normal? Maybe we should sort of take a weighted average of the new period and the old period." That is the sort of thing that machine learning models naturally do. It takes one, two, three, four, five, six, seven, eight, nine quarters until you get a vintage of loans that is fully, fully calibrated.
Now, in response to that, we started investing, as I said, in what we call dynamic macro handling, giving our models the ability to, in real time, respond to macro changes and giving the model what we call time as a feature, meaning telling our models that not all years are created equal and that time is something you can actually use in combination with any arbitrary characteristic of loans to understand what's changing about the world. By doing that and building certain techniques to make our models train and update much faster, we're able to now backtest what would have happened if we had our state-of-the-art systems back when this all played out. With those new systems in place, we would, in Q1 of 2022, have immediately started seeing our models respond to the training data.
They still would think that it's not totally obvious if it's here to stay. As those curves are really steep, there would be two quarters where you would still have underperformance. You have eliminated the majority of the underperformance immediately, but you would still have some. Within just three quarters, you start getting very, very significant reductions in the loss variance. By the end of the very first year of the turn in the credit cycle, we go from a sort of real history where we had 41% of excess loss variance to right where we want to be, which is right around or below 0%, meaning that there are actually just slightly fewer realized defaults than we predicted, which is where we like to sit. That effect persists through the entire rest of the time series here. Really powerful stuff.
It makes us very optimistic that in the ability to handle sort of generalized turns of the credit cycle, because there is nothing about what we put in place here that was specific to 2022 or the specific way that the end of stimulus effect played out. It is all just sort of native machine learning stuff, and all it sees is actually just data. All right. How does this generalize to our new products? This is a pretty important question because increasingly we are getting away from just one product into the sort of entire set of consumer credit products. The list of things that I described in the timeline, again, those are just a sample of things, but I think, again, for purposes of understanding how far along we are on the journey to generalization, it is useful.
Here I've got the sort of four new products that we have today: Small Dollar Relief Lending, Auto Refi, Auto Retail, and HELOC. I am showing you where we are in the journey of generalizing all of the work that we've done to apply to these new areas. In Small Dollar Lending, we've got basically everything. That is because Small Dollar Lending has a lot of training data because those loans are small, but they are very numerous and are contributing a very significant number of new borrowers to our ecosystem today. It also means it has enough training data to support virtually all of these techniques. Auto Refi is our oldest new product and has been around the longest, so it has picked up most of these techniques, but it still needs a little more training data before it can usefully use all of these things.
Auto Retail and HELOC are relatively newer. They're still a couple of steps away, but they will get there. The great news about this setup is that essentially, unlike in personal loans, where each time we had to develop brand new novel techniques to be able to move the models forward, here we already have all of them. We're just sort of waiting for you need to have 100,000 HELOC loans before you can use certain techniques and maybe a million before you can use other techniques. There's essentially a long runway of proven techniques that will increase the separation, improve the calibration, increase the automation of each of these products as they hit enough scale to use them. Lastly, personalization. This is kind of the new frontier for us.
It's just a new frontier, so I'll be brief and just give you a bit of a glimpse into the kinds of things that we're working on here. The general fact that motivates this whole area of research is that there is a very clear relationship between reducing charge-offs and increasing revenue in our business. That's maybe not surprising because we're in the business, of course, of facilitating credit. When there are less charge-offs, more people can be approved for loans at lower rates. More people take the loans. That relationship is very roughly a sort of 1% to 1% relative basis. In order to go after that, there are a whole bunch of different things that we can do, starting with the way that we service loans that are facing some kind of delinquency.
There are things that every player in credit does today, loan modifications and debt settlement, that are done with essentially fixed rules, meaning when borrowers have sort of circumstances X, Y, and Z, we allow them to get a loan modification, or when they are a certain number of days late, we will settle the debt at certain terms. Those modifications and settlements are not individualized in any way, meaning they're not targeted. That means that there is both the person who just lost their job but otherwise was an entirely good faith borrower who's getting a loan modification and the person who's opportunistically using it as a way to defer making payments so that they can maybe buy another round of expensive goods that they want to buy. A similar story in debt settlement.
In spite of this, debt settlement and loan modifications are very positive to the returns of loans. That is why we and everybody else have them in our set of tools. They are positive despite the fact that you are netting out a population for whom it is very useful and a population for whom it is actively harmful. Of course, the promise of machine learning is separation, that we could separate these two populations and just give modifications and debt settlements to the people that would really actually benefit from it and avoid giving it to those who would not. You can go one step beyond that, and you can do it for the specific types of modifications and settlements, the specific amounts, and so on and so forth.
Very easy to imagine you could get a very meaningful reduction in charge-offs from this since essentially every borrower that ends up defaulting will pass through some of these processes. Then there is the sort of standard interventions you do. Interventions are things like at the low end, you have phone calls, emails, texts, send people mail. At the high-end touch, you could imagine some people actually need debt counseling. They have too much debt. They need help with sort of how to manage their finances, how much spending, budgeting, all of this stuff. You do not want to spend that amount of money on everybody. It just would make servicing not work. Again, targeting is the name of the game. You can optimize payment due dates so that you can sort of draw money when it is most likely to be there in relation to payday, etc., etc.
You can match borrowers and agents so that the best people to talk to a particular borrower are, in fact, the people that will talk to them. The sort of holy grail of it all is that you can actually build AI agent servicing to dramatically change the cost structure and, again, not really in the service of cutting costs, which is sort of nice, but really so that you can use those dollars towards high-cost interventions on people where it will really make a difference in reducing the number of defaults. That is where the high-leverage financial play is.
With all that, with all sort of five areas of AI, we believe that within just a couple of years here, AI is going to enable us to have achieved all three legs of the credit triangle and to have done so across all the major categories of consumer credit across their whole life cycle. That means to consumers, we'll have the best rates and the best process for every type of credit. The gap between us and the next best option is going to grow rapidly because all this is supported by a uniquely profitable business model that, again, is very hard to achieve in lending given the dynamics of the market. We've got these compounding technology wins in algorithms and data, and we'll have a full lifecycle consumer ecosystem.
With that, we will go to a break, and afterwards, Chantal will be back to tell you about our consumer brand. Thank you.
Please welcome Chief Marketing Officer and SVP of Growth, Chantal Rapport.
Hello, everyone. Welcome back from the break. I'm Chantal Rapport. I'm our SVP of Growth and Chief Marketing Officer here at Upstart. Paul talked about our technology and how uniquely differentiated it really is. Dave talked about the opportunity we have in front of us. What I want to talk to you about is why we believe we are a category of one business and how we plan to take our technology and have the largest impact possible. First, let's start with an important fact, which is that credit is incredibly important. Dave talked about it, but I really do not think it can be said enough.
We all use credit every single day for small purchases, for big purchases. As a business, we're acutely aware of our cost of capital. A change of even a fraction of a percent, we would understand its impacts. Why is something that is so fundamental, so important to economic mobility, to the American economy, had not been innovated on in decades? Does not make a lot of sense. A lot of it is because of all the things that Paul just talked about. Lending is extremely difficult to do it well. To do it in a business that is resilient, that is profitable, it takes focus. It takes relentless focus and effort. Importantly, it takes a business model that was designed to solve for the problem of credit, which is what Dave talked about. That is how we started.
We built an entire business, an entire business model just around the problem of credit that is huge. It's a really big problem that we want to solve. While I dive into our business model, first I want to talk a little bit about some of the new realities of lending, we'll call them, new trends that we're seeing in the market. The first is that banks are losing market share. This is probably not a surprise to many of you. They've been losing market share in mortgages for decades. Now we're also seeing banks lose market share in new categories of credit. A lot of the reason is because consumers are shopping for loans, just like they're shopping for any other consumer good. The internet has taught us well. We can go online. We can compare things easily. It's really easy to find the best.
The same thing will happen for lending and loans. No longer are consumers captive to their banks because they can find the best rate, the best process, as Dave talked about. We believe that credit, too, will be unbundled from the banks and from the so-called super apps that are trying to just add credit as a feature. It is too important to be a feature. It is not just too important to be a feature, but it is also really hard because to do it well, you need AI. The other reality is that AI is a commitment. We walked through almost a decade with Paul of everything we have done every year to improve our AI. It is not just a feature you can take off the shelf and plug into your models or your business model like you might an LLM model and insert it into customer service.
It just doesn't work like that. It's extremely complicated. A lot of banks and even fintechs just don't have the time or the focus or the expertise to do it really well. For our banks, that's honestly probably not their fault because it's a highly regulated industry, which means it's really hard to have really, really great innovation in this space. The last thing is that our capital markets are evolving. The banks are pulling back, as we talked about, becoming increasingly conservative as the macro environment is shifting. We're seeing this increase in private credit coming into the industry. Knowing these trends, we believe that our business model has really been built to win because it encompasses our technology. That's what we're here mostly to talk about. It's where we start and end, the core of our business.
But it's not the only thing that's important to our business. You also need to have great capital, an intelligent capital marketplace, which we'll talk about. Lastly, you need to have a great brand. You need to win the hearts and minds of consumers again and again and again. Let's dive into each of these things. First, our leading foundation model for lending, core to who we are. Paul went into this in great depth. How does it actually impact our business model? When we normally think about our AI and our technology and how good are we or not at it, we often ask ourselves, what would be the impact of great technology on the business? We think it would be a few things. The first is pricing power because differentiated value should lead to differentiated ability to price in certain segments.
We have certainly seen that as our take rates have been elevated over the last few years. The next is what we call positive selection. Now, Paul just presented this chart where we looked at a sample of consumers and just how different our rates are from some of the industry. Now, what happens when Upstart starts to just take some of these best borrowers? Because we can offer the lowest rates, because unlike others, we do not have to average them out. We know who is going to pay us back and who is not. We have great risk separation, about eight times better than a traditional model. What happens when we start to pick off these best borrowers? Our competitors are left with an unbalanced applicant pool. Over time, default rates may rise, which drives prices up. The result is this increased difference between us.
Our advantage starts to compound where consumers continue to choose us again and again. Now, the next thing is that AI is not just our mode, but it is also a constant accelerant for our business. Every win in accuracy, verifications, servicing that Paul talked about is not just a win for the model, but it also drives business efficiency. It lowers costs. It raises margins, drives conversion, as we have talked about. In fact, for every 1% improvement in accuracy, we see a 13% improvement in conversion, which is an incredible return on our accuracy gains. Lastly, what is it all for? Dave talked about building a radically better product. That radically better product means best rates. It means best process. It means a product that is available to everyone, to all Americans.
This might sound pretty obvious, but it's actually very far from what exists today, something that has the best rates, that is cheaper, that is easier to access, because it's really only unlocked by great technology. That's been our focus today. While it's important, it's not the only thing. We are a lending business, after all. Let's talk about capital and how we differentiate in our capital section. Most fintechs and lenders have one capital source, a single source of capital that is quite limiting. Maybe there are a couple, but often they're really focused on the safest tranches of borrowers. When market conditions start to shift, they may pull back. There is no one left in the place. We've seen this time and time again.
The reality of the market is that we have a diverse set of consumers and we have a diverse set of investors, all with different price sensitivities, different risk appetites. It is impossible to serve a market as large as the one that we hope to serve with a model as simple as this. What we did is we built something and designed for it. We built a marketplace for our capital and many-to-many marketplace. On one side, we have a diverse set of capital providers, hundreds from banks to asset managers, each with different goals. On the other side, we have millions of consumers across the credit spectrum. All of our lenders are effectively competing for the borrower in real time to offer them the best rates while still maintaining control over their programs.
We as Upstart sit in the middle, intelligently matching capital in real time. This provides us a few key advantages. The first is that it's scalable. We can serve more borrowers across the spectrum. That's incredibly important. Like we talked about, we have a huge market we want to serve. The second is that it's efficient because we can optimize our rates dynamically, offering the best rate available to that user in that moment in time. The third is it's flexible. Because we have so many different capital sources, one steps out, another steps in. We can be resilient and flexible over time. The free markets in this system help optimize both lender returns and borrower outcomes, which reinforces our value as a marketplace. Technology and capital are very important for our lending product, and they'll build a great lending product.
We talked about how consumers shop. What will ensure that consumers choose us again and again? We talked about those hearts and minds. How do we win them? We believe that building a great brand does not come from just having a great Super Bowl commercial. Awareness is important. We will also do that as we continue to build our brand. Ultimately, we believe a great brand, a generational brand, comes from having a fundamentally better product. You think about some of your favorite brands that you are using generation after generation. That is because they did something radically different. That is what we are building Upstart into, the brand for credit, one that serves all Americans across every credit product that matters and will change the way that we think about credit in the future. What makes us stand out to consumers?
The first is that we offer something for everyone. We have four credit products today, more on the way. We virtually conserve most of the market, which is very unique, actually, and gives us a large addressable market. Whether you're building credit or you're optimizing it, we can serve you. As Dave mentioned, we're not just competitive in our non-prime segments, but we also have started to become very competitive in super prime too, which is a really exciting new journey for us. Not only does having multiple products give us the ability to reach out to multiple customers and increase that breadth in consumers, but it also allows us to serve our consumers through a lifetime because users change over time. They have different needs. Their needs evolve. With Upstart, they do not just come to us once.
They can come to us for multiple things as their needs change. Every new product that we launch strengthens our relationship with the consumer. Not only does it give us more data, it feeds our models, but it also builds that relationship with the consumer. We already see that potential for lifetime value in our existing user base, many of which came to us through our flagship product in personal loans. These are users that have not even yet been optimized for the world that we are building today now with multiple products. Over 75% of them have a car. 40% of them have homes. 30% of them have already shown that they are likely to take out a second personal loan. The path to lifetime value in our user base is extremely clear. This is really just the beginning.
When users come to us, they might come for an availability of a product. They might stay because we have multiple things or for the best rate. What they really love about us, interestingly, is the process. Dave talked about how this is important. Our AI in verifications has really changed the game for lending. We're not just efficient in our loans, but borrowing with Upstart can feel like magic. It's simple. It's personalized. It's very, very fast. A few stats. Dave shared this one, but 90% of our personal loans approved automatically. No paperwork. Dave mentioned all those things. No waiting in line. If you've ever done the loan process, you know what we're talking about. People were so confused when they used to go through our process.
They're like, "Why is this so fast?" We actually had to name it because we were getting reviews of saying, "Did that just happen? Is my money actually coming into my account? I was only there for five minutes. I'm confused. We are an online lender. You're putting in all your information online." We had to tell people, "You're on the Upstart fast track." It turns out 90% of people are on the fast track and do this incredibly quickly. Importantly, we didn't just do it in personal loans, but we're applying it to the rest of our products as well. As far as we know, we have the only instant auto refinance product on the market. This is a product that typically takes weeks to close, involves tons of paperwork. Applying our technology to the process, we now have the first instant auto refi.
Same thing with home lending. Our home lending product is fairly new, but already we are twice as fast as the industry standard. This does not just save consumers a few minutes or even a few days. We are talking weeks in the case of a home lending product. It is radically different already. The impact of this culminates in our reviews. We have tens of thousands of five-star reviews. I would encourage you to go read them on Trustpilot because our consumers really are why we do this, the heart and soul of what we are doing. It leads to not just satisfaction, but advocacy. You can see again and again people talking about how incredibly fast the process was, how low the rates were, how competitive they were, and how they would recommend Upstart to absolutely anyone, to all their friends and family.
Importantly, we're showing the ability to repeat this across credit products. These reviews, these are just from our product today, our product of yesterday. Using our technology, our potential is really unlimited. There's so much we can do. We're early in applying all of our impacts of our technology into the consumer experience. You could imagine a future, maybe one not so far away, where credit for Upstart members is always on, always available at the lowest cost of capital. For any loan you need, $50, home loan, doesn't matter. It's always available to you as an Upstart member in your pocket and always priced with precision. Make it as simple as it is to use a credit card. Dave mentioned how credit cards are so accessible to all of us today.
We don't see any reason why amazing credit can't be that easy, that accessible in your pocket at all times. I'll take it back to today. This is a very exciting vision. What are we doing today? I want to show you a tactical, tangible example of a consumer's experience today and how we drive not just revenue and profit, but also consumer loyalty. To do that, we'll talk about the auto refi market. Dave, when we launched auto refi, famously internally said this was a $0 billion market. Why did he say that? Auto is obviously a very large market. I didn't see it up there in his $25 trillion slide for market size. It's because when we launched auto refi, we realized very quickly that this is actually a product that no one knows about.
No one really understands that they can refi their car. If they do think it's an option for them, it's only if you're in this very specific situation and it's really hard. We had consumers thinking they had to go to the DMV to refi their car. We had to do a ton of education and awareness in our marketing. Honestly, it just wasn't that effective. It wasn't the same margin profile in our marketing that I'm used to. We took a step back and we pivoted. We took all of the dollars that we were spending on marketing, we put them into the product, passed the savings on to the consumer, and decided to focus on our existing consumers and users that were right in front of us.
The result was much better economics, of course, but also higher volumes and, importantly, a long-term relationship with our consumers. What does this look like? Here is a real example of a customer. Her name is Fran. Fran came to us one afternoon looking to pay off her credit card debt. She came to us just before noon, 11:53 A.M., and she applied for a personal loan. About 10 minutes later, she had completed her entire loan process. Nothing new or more needed from Fran. It is about $3,000. It was fast, easy, kind of the process that all of our borrowers and applicants come to expect from Upstart. Her journey did not end there because we know Fran. We know her credit profile. We know her information. We also knew that Fran had a car that we could refi and give her a much better rate on.
We offered it to her in a personalized offer. Fran looks at it and she clicks on it, 12:05 P.M. Just nine minutes later, she's completely refinanced her car, walking away with a better loan term, saving her $420 a year. This is meaningful to someone who's looking to save just a little bit every month. She did all of that from what I can see in less time it would take to eat lunch, probably on her lunch break, refied her credit cards, got a better loan. Out of it, we got a very loyal customer. She left us this feedback. She's very grateful to us. That's very nice. We have tens of thousands of these, like I've said, on TrustPilot. The bigger point is that most people actually couldn't do this in the market.
They couldn't do something so fast, so seamless, so personalized because typically an auto refi loan would take weeks to close and to process. It's high CAC. It's very low margin. It's very hard to scale. With Fran, we were able to do two things. The first is originate a second loan at 0% CAC. The second is we deepened our relationship with Fran, ensuring and positioning Upstart as the go-to platform for credit for the rest of her life. Fran isn't an exception or just one example. She really is our blueprint. We can keep doing this around our emerging products. In fact, by shifting our model to focusing on product-led growth in our existing users for auto refi, we saw almost 20x growth year over year.
We've seen incredible success in this already and are continuing to apply it to all of our new products with the products and the platform to really build lifetime value at scale. I'll wrap with where we are today. While we're a lending platform that is trusted by millions of people already, we have proprietary AI that delivers eight times better risk separation than a traditional model. We have a flexible capital model with over 100 bank partners. We have the product breadth and the infrastructure now to unlock efficient growth in new categories. This is where we are today. It's a pretty good business, I would say. What we're really excited about, and what we actually don't share publicly that often, but we're going to talk about it today, is our vision and where we're going.
It's the idea that we're building a place where the best borrowers don't search for credit. They simply have it. Credit that is always on, guaranteed at the best rates, zero friction that Dave mentioned, and all delivered through a single, intelligent, and deeply defensible, importantly, platform. That is why we often say that Upstart is in this category of one, because yes, we have lots of different competitors, but we actually don't have any competitors that compete with us on all three fronts of our business model that have amazing AI and deep technology and this incredible moat that Paul talked about, that have a productized marketplace for capital that's efficient and resilient, and that's taken all of that and then wrapped this amazing consumer brand around it that millions already trust and millions more to come in a brand that's really scalable for consumers.
No one has all three. What we're building today is really a platform that's meant to win the market today, but also to redefine the process of consumer credit and the industry at large. We're really, really excited about this. Happy to answer questions on it. I could talk all day about our vision and where we're going. It's what keeps us every day motivated to continue to win this business. I know soon I'm going to stand in between you and lunch. With that, I will introduce you to Sanjay, our Chief Financial Officer, who will talk about how this model translates into profitability and how we're building more resilience in our capital structure. Thank you.
Hey. Thank you, Chantal. Thanks to all of you for being here. I know there's a lot going on right now. It's great to see many of you in the flesh. Hopefully, you got a bit of a sense or a bit of a lens into how we think about our technology and our business model and our opportunity. I'm going to bring us home this morning. I'm going to talk a little bit about our financial profile, some of the ways in which technology shows up in our financial profile, some of the ways in which our financial profile is unique, and also talk a little bit about our risk profile as a business and how it's been evolving over time. Before I get into that, I'm just going to reiterate a couple of key messages you've already heard this morning that I think are important.
The first one you heard from Dave, that this is not a small opportunity. You all already know how vast the space of credit is. Mispricing is ubiquitous in credit. As Dave said, every single borrower who paid back their loan is paying a tax to subsidize a bunch of strangers who did not pay back their loan. The majority of borrowers do pay back their loan, by the way. That means that the majority of the mispricing is overpricing. This is not a 50/50 thing. Very often, that default subsidy, by the way, is the single biggest component of the APR, which means this is not insignificant overpricing. It is significant overpricing to a majority of the applicants. We have all just accepted this as normal because this is the way it has always been, but we should not because it is a consequence of poor risk models.
Now, from a process standpoint, any of you that have accessed credit recently, as Dave said, I think probably viscerally knows how excruciating the experience can be. Why? We are all paying a process tax because of a small minority of applicants who are going to attempt to defraud the lender. And because we do not know who that small minority is, we all pay a pretty heavy process tax for it. By the way, pricing and ease of process, as Dave said, are the two single most important buying criteria in this consumer space. Other than maybe pioneering digital distribution, I think that our industry, frankly, has not done too much to innovate on these two core value principles, these two value propositions. I think that anyone who can and will is going to make a lot of money.
Now, hopefully, you've gotten a sense for the fact that as a company, we are technologists at heart. Paul talked obviously a lot about how we think about these problems. To us, the clear answer to these problem statements in this vast space has to do with better models. Better modeling directly addresses the two core value propositions of this product. A better model will allow you to reduce more of the default subsidy everyone is paying by doing a better job of avoiding the defaulters. Because a majority of the borrowers are overpriced in this market, a majority of them will benefit from this. Better models allow you to dispense more money with less friction. Why? Because it turns out these models do a pretty good job of guessing who the minority of the people are who will attempt to defraud you.
It allows us to hyper-concentrate the friction on that minority of applicants, and it allows the rest of us to run fast and free. There is a lot of hand-wringing in AI circles these days about monetization models. Who is going to make money from these interesting new technologies, and how are they going to make money? We think we already have a pretty good answer. We have an already existing very large industry with a lot of inefficiency, and better predictions directly improve the two single most important value propositions of the product. AI for us is not a talking point. I hope you understand that now. It is not a flavor of the month. It is not an R&D project on the side that is attempting to generate insights that will then get brought to some credit committee so it can then get coded into a rules-based system.
This is at the root of our DNA. We've been laser-focused on this mission since the earliest days as a private company. We've spent the time and the effort and the money and the R&D to build what we have, and it's foundational to our production systems. It's not an exaggeration to say that our credit committee is an AI model. We don't believe you can do this thing half-baked. That's philosophical to us, and this is how we intend to approach these problem statements in this industry. You heard from Chantal about our vision for how we want to take this technology and create a very unique business with it. A business that is the destination to go for all things credit, that's relevant to all borrowers across the entire spectrum.
A brand that's broadly recognized, a set of consumer relationships that start highly engaged and follow you throughout your lifetime. An ability to underwrite that's always going. This is a vision for a business that I do not believe exists today in fintech, and that's the opportunity we're pursuing. Now, from a unique business, you would expect a unique financial profile. There is a lot of ways in which I do believe our financial profile has unique aspects, and a lot of those arise from our specific use of technology. We have a unique growth model in lending. It's a direct result of how our technology evolves. Technology gives us a relatively uncommon degree of control over our pricing and our take rates. Technology allows us to control and reduce variable costs and, in what is otherwise a very competitive market, carve out healthy and expanding margins.
Like any good software business, you expect good leverage and productivity from fixed resources out of technology. If you put all of that together, look, ultimately, what we all care about, what you and I care about, is that as we execute, this is a business that throws off profits. I want to touch on each one of these and talk a little bit about how technology impacts these things. Okay. We have a unique growth model in lending. As Paul said, you would be right to be skeptical of rapid growth in lending. Why? Because at some steady state with static underwriting, historically, the only ways to meaningfully grow rapidly in credit are either by ramping your marketing costs and risking your unit economics or by opening your credit box and risking your credit performance. Those are not particularly sustainable ways to grow a business.
However, if your underwriting is dynamic, you do not have to make that terrible choice. Why? Because dynamic underwriting, as your model gets better and better and more accurate, you do a better and better job of reducing the default subsidy. As you do a better job of reducing the default subsidy, your APRs improve. The immediate thing that happens when your APRs improve is that your conversion rates grow. There have been a lot of periods in our recent history where we have demonstrated very rapid growth, and they tend to be very correlated to periods in which we are achieving conversion gains. It is not the destiny of conversion rates to go up and to the right ad infinitum. At some point of very high conversion, rationally, you should spend more money to increase the top of your funnel because of how productive that funnel has become.
Over time, model gains start to take the form of more marketing spend and more volume at flat to declining acquisition costs. That is a pretty magical model as well. Regardless, this is a unique model in lending because historically, there is not a long history of people who have dynamic underwriting which improves at consistent rates. Importantly, what I want you to remember, when we next go through a period of very rapid growth and we fully intend to, this is not necessarily a sign that something is wrong. It is possible to grow quickly in lending if you have dynamic underwriting. Remember this. A second unique aspect of our financial profile that arises from technology is that we have a relatively high degree of control over our take rates.
If your underwriting is commoditized, you will generally end up playing in a highly competed segment of the market along with a number of other competitors who probably have somewhat similar underwriting. You will generally end up as a price taker. Generally, you will originate loans at something close to breakeven, and you will earn the money to run your business off of the market yield of your assets. However, if your underwriting is differentiated, you're not resigned to this fate. If your underwriting is differentiated, you can perceive hidden value in the borrower base that the market does not appreciate. If you can perceive unique value in the borrower base, you will enjoy a degree of price and elasticity in the demand for your product. The better and more accurate your models get, the higher the degree of price and elasticity you will enjoy.
You guys saw the chart that Paul put up. It showed how we sort of price a random sample of borrowers relative to the market, and they had dots where we price people and ranges where the market pricing was coming in. Hopefully, what you noticed is in a lot of that chart, and certainly in the borrowers that we like a lot, there is a lot of room on that chart. It means that we can raise our take rates meaningfully for those borrowers and still be the best rates in the market. That is where the inelasticity comes from. That is what has allowed us to raise take rates very meaningfully. In our case, in recent history, anything from 40%-60% and raise them accretively to the business. Now, like conversion rates, it is not the destiny of take rates to forever go up into the right.
At some level of high take rate, you have the luxury of exploring the trade-off between less current economics and more future volume and lifetime value. You will rationally, at some point, moderate your take rate, and you will invest some of that in the ability to cross-sell a higher future volume in increasing number of increasingly long consumer relationships and monetize them via a growing product portfolio, the vision of cross-sell that Chantal talked about a little bit. All of that is sort of beside the point. This is all internal calculus and optimization. The key point here is, in our core business, because of our technology, we are generally not price takers. It is generally true that our take rates do not come under pressure from outside competition.
In addition to rising take rates, technology gives us the ability to control variable costs and expand margins. One of the two large variable cost components of originating a loan is the operations cost you incur when onboarding a new loan. As Paul talked about, if you have very good verification technology and good fraud modeling, you can take a lot of that cost out of the system. As you have seen how our automation rates grow, we have taken a lot of that cost out of the system. At the same time, the second large variable cost you incur in the origination of a loan is the acquisition cost. As Dave said, in general, acquisition costs tend to grow as your business scales because the next marginal consumer is harder to find.
Because of the growth model I described, we've demonstrated an ability to scale our business at flat to declining acquisition costs. Now, when you combine increasing take rates, decreasing operations costs, and flat to declining acquisition costs, you get margin expansion. You've seen our contribution margins expand from the mid-40% to as high as the mid-60% over time in what is otherwise a very competed space. Again, I'll make the disclaimer. It is not the goal of our company to increase contribution margins up into the right. It is the goal of our company to maximize the number of contribution dollars we are earning. We will decide to launch and incubate new products that will be dilutive to the contribution margin but accretive to the contribution dollars that are falling to the P&L.
We will make the appropriate trade-offs between current economics and more lifetime volume and value as makes sense. Again, the key point here is that in what is otherwise a heavily competed space, technology has given us the room to create very healthy contribution margins and a unit economic profile that technology can improve over time. We think that's relatively unique. The next thing you'd expect if you invest heavily in technology is a certain level of leverage or productivity from your fixed resources. Of the salaried headcount at our company, more than half of them sit in the technology organization. That's a rule of thumb that we keep track of, and that's important to us. Why? Because it means we're investing a certain appropriate amount, reinvesting a certain appropriate amount into innovation and R&D. What's the outcome of that?
One of them is, for example, if we were to execute on the guidance that we've laid out for you for this year, we will once again reapproach a ratio of roughly $1 million of annualized revenue per salaried headcount. That's a heuristic that we think is a good rule of thumb for a software business. It means you're getting good productivity out of your resources. It means that as we have grown the top line of our business over the past couple of quarters, a significant fraction of our cost base is steady, and it's creating the room for EBITDA margins to grow. That's what you've seen with our business over the past couple of quarters. In a sense, operating leverage allows the top line dollars to fall very efficiently to the bottom line.
Speaking of the bottom line, ultimately, what we care about, of course, is GAAP profitability. We were a bit of a strange bird as a private company in that we were GAAP profitable for a large chunk of it. Of course, back then, it was kind of held against us, frankly. Back in those days, as a private company, the fact that we were GAAP profitable, the venture capitalist community just basically thought that meant that we were not very serious about growth. It was weird to be GAAP profitable. Now, of course, as you know, we're very serious about growth. We think that in lending, there is a good way to grow and a bad way to grow. We decided that growing with GAAP profits would be a signal that we were growing the right way. Of course, that carried on into public life.
In early public life, we were sort of GAAP profitable for the first six quarters post-IPO. In the rough sort of cohort of fintech IPOs from that era, that was relatively rare as well. Now, obviously, our sort of goal of remaining GAAP profitable has been challenging as we've navigated the elevated default risk environment of the last couple of years, unquestionably. We've maintained a priority in returning to GAAP profitability. That's been very important to us. We've guided that we will be getting back there in the second half of this year, which I think is safe to say is ahead of most people's expected schedules that I've talked to. As you can see, we're well on track to accomplishing that. Okay.
You put together a model which has a unique growth profile, strong, high degree of control over take rates, an ability to carve out healthy margins and reduce variable costs, operating leverage, ultimately profitability. I think it's worth reiterating, our intention is to do all of this in a model of low capital intensity. We're not trying to do this off the back of net interest income. There's nothing wrong with net interest income. It's a perfectly good revenue stream. It's not the one we're going for. We continue to have no intention of becoming a bank. I think we may be one of the last platforms standing in that regard. There's nothing wrong with a bank. It's a perfectly good business model, not for us. Why?
Because the ambitions we have around scale and velocity are such that we do not want them to be limited by the need to furnish all of the capital ourselves, frankly. We have spent a lot of time and talked a lot about the fact that we have put in place a network of partners who are bringing third-party capital to the platform. Frankly, we think that is working very well. We do have a balance sheet, and that is an important tool for our business. It is important to say. It allows us to launch and incubate new products at velocity. It allows us, frankly, to co-invest in some of those partnerships and achieve a level of co-commitment and resilience that is important to our capital base.
This is not a no-capital model, and it's not a no-risk model, but it is a low-intensity capital model that will scale well and allow us to be very nimble as we push the further frontiers of risk. Now, all of the things I just talked about—growth, profitability, capital light—are all great. It is worth saying all of these things—well, on the one hand, I think we've demonstrated versions of this at points in our past across the board. These are not leaps of faith. They're things that have been demonstrated. On the other hand, they are all things that have come under strain in the past couple of years as we've navigated this high-risk environment that we're in currently. I think it's natural to ask, how are all of these things going to perform the next time there's a macro downturn?
Are they all just going to kind of fall away? I think that's a very reasonable question. It's a question we spent a lot of time on. We spent a lot of time thinking about it and debating about it and working on it. I think there's a bunch of things about our business that have evolved and that will continue to evolve such that the risk profile of our business is going to look very different as we sort of prepare to navigate future macro disturbances in the force. Some of those things have to do with diversification along various dimensions of our business. Some of those things have to do with resiliency on the funding side of the business. Some of those things have to do with some of the macro tooling that Paul has talked about. I'm going to quickly touch on some of these.
Okay. On the borrower side of our business, the first axis of diversification that is important to us is on the borrower side of the business. Historically, you may have thought of us as a business and a model that competed disproportionately well in certain segments of the risk spectrum and, as a result, having had certain concentrations of borrower pockets in certain areas of the risk spectrum. I believe that is an old mental model. Dave and Chantal explain why increasingly we believe we are relevant to borrowers across the entire borrower base. That is sort of playing out in the mix of borrowers that we have and how that's evolving. It's leading to a greater breadth of distribution across the borrower base, more balance across the various segments, a greater diversity of borrower types.
It means that when we encounter the next macro sort of bump in the road, we'll have more stability and more diversification in the borrower book. We will be less exposed to segment-level and sector-level shocks. We'll have less pockets of concentration in the borrower base, particularly those near this 36% rate cap, which is really the only hard constraint we've placed on our business. Overall, a more resilient and stable borrower base as we grow. The second vector of diversity is along the product axis. Historically, originally, we have chosen to train our risk models on unsecured installment loans. Because this product is at the bottom of the payment waterfall, in some sense, it is the purest expression of borrower risk and therefore the best way to learn. It has served our model training very well.
The downside, of course, is that you're at the bottom of the payment waterfall. Now you're starting to see the momentum we have in new emerging products. That momentum is early, but it's real. These are products, on the one hand, that have more security. They're higher up on the payment waterfall. In a macro disturbance, because they tend to be secured by cars and homes, they are more stable. On the other hand, you have another set of products that are shorter duration in nature, shorter term. A smaller duration means you have less exposure to future macro events. You take these together, and as it grows, you're going to see a product portfolio that has increasing amounts of security.
It's at different points in the payment waterfall, and maybe a sort of increasingly short duration will limit some of the macro volatility along the product dimension. The third point is something we've talked about at length, and I won't belabor it, but it has to do with the resiliency of our supply chain on the capital side. Today, more than 65% of our capital base is in committed form. That's up from zero not very long ago. These commitments are with capital partners who have durable capital. They themselves are typically in the business of raising and deploying money across larger time horizons and across cycles. These agreements are designed to disperse across the cycle. We are co-invested with them, and we have skin in the game. I'm very confident that this will be a more robust capital model than the one we had a few years ago.
Frankly, we may be getting a mini test of it as we speak. Ever since Liberation Day and notwithstanding the recent agreement with the Chinese, you may have noticed that the markets are skittish and treasuries are up and spreads are up. I think if we had been in our old capital model, I don't doubt that we would be fielding calls today requesting pullbacks in May and in June just because we're a little nervous. As it stands with our existing partners, we are in the boat together. We're navigating the environment. We are aligned and sharing in the risk, and everything's working exactly as designed. So far, so good. The fourth point is the point that Paul raised, so I also won't belabor it. Suffice to say, our original obsession as a company was with borrower selection.
In the last few years, as you heard, we've pivoted a non-trivial amount of our resources to becoming good at measuring the state of the macro precisely and reacting to it quickly. We now think we have what is some pretty unique tooling in the industry. If we had had that tooling a couple of years ago, as Paul said, of the underperformance that our loans have experienced in this period of turbulence and high default, we would have avoided well north of half of it. That is a real source of value we are now delivering to lenders and investors. I think it is a very unique set of tooling, and it means that the next time we enter some macro turbulence, we're going to be able to navigate it much more artfully, much more precisely.
There's one more thing that I didn't really put in this framework, but I am convicted in it, and I wanted to sort of at least throw it out for consideration. It has to do with our revenue model itself. Today, our revenue model is highly transaction fee-centric, obviously. That is, in many ways, a good thing. In fact, it was certainly useful to us as a young, underfunded private company because it means that you get all your revenue and, more importantly, all of your cash upfront. It's a very efficient way to build a business. Now, of course, the downside is that it's very sensitive to in-period transaction volumes, and we've seen the effects of that. There is one revenue line item that we have that was much more resilient over the past couple of years. It is our servicing fee revenue.
The reason is because that's a revenue stream that's paid radically over the life of the loan. Therefore, it's much more related to your outstanding book of loans that you've originated over the prior years and much less about your in-period transaction volume. I think you're going to start to see more revenue models like this. The obvious sort of near-term example is in auto. As that business scales, I believe we will see more and more of the economics materialize in the form of servicing fee revenue that will be paid radically over the life of the loan. It'll be a more resilient, slower to grow, but more resilient stream of revenue. Some of the other things we haven't talked about as publicly, but maybe you can intuit from Chantal's vision.
If you think about a set of consumer relationships that are increasingly durable, that are increasingly cross-sold across many products, where underwriting is sort of more and more persistent over the duration of that relationship, it may not surprise you to know that we're starting to experiment with things like revolving forms of credit and subscription models, obviously much more reoccurrent in nature as revenue streams. Those are things that I suspect we will have more to talk to you about in the coming quarters. One thing that's maybe farther out there because, frankly, there's no active energy going towards this at all in terms of business model, but I'm actually very convicted in this, so I want to put it on the table.
I sort of, at the highest level, view us to be in the business of applying AI to all of the different components of the loan origination process to improve them, all today in the service of loan originations on our platform. If we're very successful at that, I think there's every opportunity to carve out specific components of that process and to offer them as standalone businesses or services. The one that really comes to mind that's obvious to me is in the servicing space. Paul talked a little bit about the efforts we're making. We spent the last two or so years really focused on deploying our technology chops to making the service sort of processes that we have better and, in some ways, quite unique.
If you think that servicing is a business that could benefit from AI models, and if we're successful in being at the forefront of that transition, I think we have every opportunity to open that up to the broader industry and to make those services available, maybe servicing as a service to the broader space. That's an example of something that I think is an exciting opportunity. It would obviously be much more correlated or very decorrelated to our own loan sort of platform portfolio volumes. I'm painting a picture of sort of increasing diversification across the borrower base, increasing portfolio of products that are at different points and stages of the payment waterfall, an increasing resilience in our capital base, tooling we have that is unique that's going to allow us to navigate twists and turns in the macro.
I think over time, maybe a bunch of emerging revenue models themselves that are decorrelated from our transaction volume, all that will create a bit of a different risk profile for our business as we navigate future macro. It is not going to decouple us from the macro, obviously. We are, at the end of the day, in the business of applying technology to credit. And there is a macro component. I think in terms of the sensitivity and the volatility in our business that we have observed in the past, it is going to look very different as we scale and mature as a business compared to when we were a one-product borrower-concentrated youngling with relatively understanding of the macro around us and an at-will capital base that we were still learning to manage. I am going to leave you with that thought and conclude my remarks.
I'm going to reiterate what Sonya said at the beginning of the day. We're very grateful to have your time. This is a group of people. I know most of you. You are amongst the busiest people I've ever met. The sheer amount of information you guys have to get your head around in your day-to-day, not just with us, but all of the things that you cover, is mind-boggling to me. I know that taking an entire morning to go deep on a specific technology is a luxury that you guys do not often have. The fact that you guys are happy to spend your time with us this morning is very meaningful. On behalf of the entire leadership team, thank you. We're going to pivot now to more of a Q&A session, so you can start bringing up some of the chairs.
Bear with us. We're going to call the leadership team back up here. We're going to introduce you to Annie, who's a very important part of our leadership team as well. We will be happy to take questions on anything you've seen today or anything else. Afterwards, we're going to have some breakouts, and you can come find us and grab us one-on-one or in small groups and talk about whatever's on your mind. We're happy to have you here, and we're at your disposal. I guess with that, in terms of prepared remarks, thank you.
We'll take questions in just a second. Marissa, did you want to provide the backup mic, or you're all good? You're all set? Okay, great. We've got two runners in the room, so we're going to do this open Q&A style. Pete, because your hand went up first, you go first. Assuming there's a mic over there.
Thank you. Great sessions today. Very helpful. Thanks for doing this. I want to talk about macro sensitivity. It's the number one conversation we have with our client investors. It seems like the model, Paul, you talked about how the model is really you've spent a lot of time on working on calibration and adapting to various macro environments and things like that. On the other hand, Sanjay, the risk retention of the business model has changed quite a bit. How should we think about now with exposure to things like residuals, so on and so forth? How should investors think about the economic sensitivity of Upstart's model today versus perhaps how it was in the previous cycle with those two kind of factors juxtaposing themselves? Maybe a faster reacting underwriting model versus a different risk retention model. Thank you.
Before we jump in, it's my mistake. I should have introduced our esteemed Chief Risk Officer, Annie Delgado, who's joined us for this session. We're very excited to have her up here with us. I just wanted to make sure everyone knew who was on stage.
Sure. Thanks for your question, Pete. I can maybe take an opening crack at that. I think one very important component of this and of our evolution, as you say, the risk profile has evolved, is the investments we've made in calibration and the tooling that we have and the fact that the models now understand time. They're not predictive of the next macro event, but I think they're very quickly reactive. With that, we've set up a bunch of arrangements. The equation, from our perspective, roughly looks like this. We are going to have vintages in benign periods. There will be some vintages in challenging periods. I believe our vintages in benign periods are designed to overperform. There will be vintages in challenging periods that underperform.
Our goal and the way these agreements are set up with our counterparties is that they're meant to harvest the overperformance and put that in play to sort of cover for underperformance. Over longer time frames, over longer periods of time, the equation we need to get right, and the one we feel very confident in, given the tooling you've heard about, is we need to get the aggregate sort of performance right over the cycle. That means, in some sense, this is a bit of almost a macro insurance layer. If we can get the performance roughly correct over the cycle, these things will work well. If we can't, then I think we have to look at ourselves and figure out why not. That's the fundamental bet we've made and the one that we feel really good about.
I would just add that I think if you rewind back to the beginning of 2022 and you look at what happened, I mean, the sort of most fundamental thing that shocked the business was that loss variance chart we showed earlier, that the sort of actual losses were 50% higher than they were modeled to be. In that circumstance, a lot of the sort of sources of capital that were in the system did not want to continue funding in that system. I think there is a world of difference right out the gate between 50% and 20%, and then, of course, between being in that state for nine quarters versus being in that state for three quarters. I think those are the things that are unlocked by the sort of technology change, which we feel very good about.
On the business model side, as Sanjay said, of course, we have different agreements with the nature of that capital. If you forget about the sort of any contractual things, you just think about what the fundamental thing is, it's like investors in good times are maybe looking for a certain return. Let's just, for the sake of argument, say it's like a 9% sort of annualized return on the credit. If that 9% goes to 7%, they might be not thrilled about that, but it's really not a catastrophic event in the way that maybe if 9% goes to 0% or it goes below 0%, that might be. In our view, again, really a world of difference between 50% versus 20%, nine quarters versus three quarters.
Sure, Ramsey. Go ahead.
The lights are blinding up there.
Yep, a little bit.
Ramsey El-Assal from Barclays, thanks so much for taking my question, and thanks for a super insightful day. A lot of new information to digest, which is greatly appreciated. Both Paul and Sanjay, and maybe following up on the prior question, you mentioned these new model inputs or tooling on the macro side that in back test showed better performance in 2022. In layman's terms, if possible, could you just help us think through what that means? Is this new data you're pulling in, or are these new variables you've identified? What gives you what is the secret sauce there in terms of that macro improvement in macro predictability?
Yeah, good question. Just to take a step back, I want to remind everyone about these sort of two terms that we throw around a lot now, separation and calibration. It's important because they actually have very different functions, and the story of what happened a couple of years ago really depends on understanding these two things. Separation is our ability to accurately rank the relative risk of borrowers, knowing which borrowers are more or less risky than others. This is the thing that's most heavily influenced by things like how many facts you know about the person.
If we add in cash flow data, if we add in sort of the stuff I talked about with embeddings, or now we know about the sort of value of eating at Chipotle, all this stuff, we can add, we put tons and tons of work here, but it primarily benefits us in separating risk. On a relative basis, we know that one borrower is 2.3 x as likely as another borrower to default. We know that borrower A is more risky than borrower B. It does not tell us very much about the aggregate level of defaults, and that is the thing we call calibration. Up until 2022, we never really gave a second thought towards investing machine learning time towards the calibration problem. We focused at almost 100% on the separation problem. When 2022 rolled around, the model's separation actually held up great.
There were no problems with the sort of model's ability to rank the relative risk. When we said that one borrower was more risky than another, we were getting that just right. What happened was across the entire industry, every category of credit and every type of consumer basically got riskier from January 2021 through January 2023, almost eight consecutive quarters of increasing likelihood of default across all observable segments of risk. Nothing you would have done in separating the risk was going to help you. What we did was we said, "Okay, what would we do if we wanted to be the very fastest at reading and reacting to macro changes?" We introduced something to the model that we call time as a feature.
Time as a feature basically means that unlike sort of the normal way you would train a model, where you just take all of your training data and you pour it in and you tell the model to treat all of this training data the same, we instead told the model that one of the facts you can have about every data point is the day it happened. If you know that the day it happened and you're allowed to dynamically interact that variable with any of the other variables, suddenly you can tease out these interaction effects. You can see things like, "Hey, it looks like people with certain FICO scores, the sort of effect of FICO score changes across the months and quarters and years that are observed in the training data." Of course, you could see that in aggregate.
You can see that, "Hey, if it's 2024, the same borrower with exactly the same characteristics is twice as likely to default as that borrower if they were in 2020." You can also see trends like a borrower who has a sort of low level of income, for example, who was getting a lot of stimulus was a whole lot less likely to default in 2021 when there was a lot of stimulus than in 2023 after the stimulus ran out. You would pick that effect up almost immediately as soon as it happens, which is the thing that causes the sort of reductions in loss variance. Because basically, as soon as those sort of stimulus changes run out, it's not that the model is going and reading CNN and knowing about the news.
It's just that we almost immediately see in the delinquency data, in the early delinquency data, you'll start to see that borrowers in certain categories will see higher rates of delinquency than they used to see. As soon as that effect becomes statistically significant, because the model has time as a feature interacted with all of the borrower characteristics, it can pick up at an arbitrary subsegment level these changes in trend. As soon as they're significant, it says, "Oh, this is a new thing going on in the world," and then it prices it into the new risk. That's why within the same first 2022 Q1 quarter, you go from that 59% level of loss variance down to 22%, and then within three quarters, it's entirely accounted for.
Does that make sense?
Hey, guys. Kyle Peterson from Needham. Thanks for doing this, but I wanted to touch on funding. Obviously, the commitment to capital slide with the progress you guys have made with how much, I think it's from 0% to 60%, 65% or something like that in the last few years. I guess what's the right mix? I know there's some co-investment there, but I guess how much of the funding stack do you want to leave open for some of that at-will channel, and how much are you guys willing to push up the commitment to capital as potentially the overall funding stack? I'm sure it makes it easier for you guys to plan growth and marketing and such. Yeah, anything there on optimal funding mix and how you guys see that evolving over the next year plus would be really helpful. Thank you.
Yeah, I would say it's not an exact science. We know it's not 100%, obviously, because when there are changes in the broader risk environment, you want some flexibility, and you can't be 100% committed. I think we definitely know it's more than half. The next time there is macro turbulence, we want well north of half of our capital to be sort of very firm. Where we are right now feels pretty comfortable. Some of the trick is as we grow, you have to put agreements in place to sort of scale with the platform. That also is not an exact science from a timing perspective. I think the percentage will jump around, and it'll be somewhere sort of well north of 50%, but certainly not 100. That's not a precise answer, but I think operationally, that's probably the best we can give.
Thanks, guys. Great job today. Henry, [Henry Invests on X]. You talked a lot about new product verticals and new launches. Curious if purchase mortgage is something you guys are considering. On partnering with that, you announced a recent deal with Walmart to sort of co-brand your personal loan product. How should we think about that for auto? Maybe a deal with Carvana, or if you ever get to purchase mortgage, maybe something with Zillow. Thank you.
Hey, Henry, thanks. Thanks for the question. I think we label that category for us not HELOC, but home, because it is our intention to play very broadly across it. We do not have anything imminent. I think for sure, if you looked at the nature of what we have said in several things, which is we are building models for all forms of credit, we think there are enormous wins that AI brings of different types to each form of credit. What Chantal Rapport said about having a permanent relationship with somebody as they go through parts of their life, to me, it seems fairly obvious. You should expect purchase mortgages to be part of our future, as well as refinance mortgages and probably some other categories. Our view is we will get there.
We're going to get to all the forms of credit that really matter to consumers, and I believe at some point businesses as well. That sort of magic of permanently underwritten, guaranteed rates, et cetera, I think is very, very powerful across all flavors of credit. It is really for us a question of when, realizing that we have very modest, very, very modest market share in the products that we're in, so enormous growth opportunity. We always have to balance the trade-offs of investing in the products that we're in versus when we feel comfortable enough spinning up a new team to take us into another category. You will see more categories this year. You will see more categories. I would expect every year you should. Yeah, we'll get there at the right time.
Sorry, I can't see who's got the mic.
On this side, right side.
Go ahead.
Hey, Vincent Caintic, BTIG. Thanks for the time. This has been super helpful. Wanted to ask actually about your industry perspective. I think you highlighted how much more wins you have, your separation. I think that's all been great. Competition, though, does talk a lot about AI and all these great things. I'm wondering from your view, when you think or how much separation there is between where you are and the industry, how long is it going to take for the industry to catch up where you are now, and what are you working on so that you're always ahead and how much investment that might take? Thank you.
I mean, I would say we do not see a lot of evidence of the industry taking on the types of problems that we are taking on. I mean, we certainly are pretty open and we share a lot. Maybe there are some amazing competitors doing things of the nature of what we are doing, and they are just very quiet about it. Of course, it is possible, but it is a fairly small industry. The types of people you need to hire to build the types of things we are building is a fairly tight group, if you will. A lot of them got their PhDs together in this and that. It is, for us, hard to imagine that there is some incredible R&D in the areas that we described happening in this industry that we are entirely unaware of. I would just say, look, our view is it has taken us a while.
If you've been watching Upstart for a few years, you've been watching the construction of this sort of foundation model we're describing, and it's coming into its own in ways that we feel very optimistic about. I mean, that's what we're trying to share with you today. We're at the point where we can compete across products with the best product out there, price, experience. I mean, I guess our perspective is we see no evidence that anybody's closing the gap. Quite the opposite. We feel like we're kind of competing with ourselves. What we're really just trying to do is strengthen the business, strengthen the models. We feel confident you're going to see that expressed in our financial results in the future because it's just sort of an inevitable outcome.
Not to say we won't make mistakes and things will always go perfectly, but we just don't see another company attempting to do what we're doing. I will just say this also. There's a thesis behind which you sort of heard, at least indirectly today, around who we are and what we do. That is that each bank, each neobank, each fintech building their own lending as a sort of add-on to their main thing, in our view, will not be a viable approach. I mean, not to say that it would be terrible and awful, but we are a lending unbundled play. That is that consumers will shop and that we can build the destination they will trust, they will go to first, maybe last. That intense focus on this very, very hard problem is going to yield incredible results.
That's our play in the market. I don't know who else, I mean, I couldn't list the company that is probably trying to do something most like this because, as Chantal said, we're kind of a category of one.
Thanks. Hi, just like to echo that. It's been a fantastic day so far. Thank you very much. Simon Clinch from Redburn. Just want to follow on that sort of competitive discussion now. Trying to think about how the competitive landscape is going to look in the future. We've got a lot of large banks trying to invest in technology with third-party partners, trying to digitalize, modernize at the moment, leveraging large language models where they can. Then you have a unique company like yourselves that are effectively trying to provide a platform for those banks to outsource and do things. I just wonder how that conflict is going to look, certainly in your vision, in the next decade plus.
Sure. I would just say, you know, if you went back to the time we went public and how we imagined the world evolving, I think we saw our business as just as much a private label, white label, distributed by large banks and really kind of having this balance. The reality that's played out is much more is our brand and our presence. Just to be frank, large banks have sat on the sideline of this evolution for the most part today. That may change with product mix. They certainly like secured products more than unsecured products.
I'll tell you once, years ago, I met with the Chief Risk Officer of one of the largest consumer banks in the U.S., and I said, "Do you realize you're only serving about 10% of Americans when it comes to credit?" He said, "Yes, that's all we want to serve." Our interest in pushing this technology sort of through the industry is there, but at some point, there's enormous value in aggregation because you don't know whether that bank is going to approve you or this bank is going to approve you. That's the whole point, if you can guarantee the experience and the rate through a relationship with us and our network of partners. Another way to put it also is this technology, AI, in the banking world is revolutionary and new.
For all the reasons I think you know, banks are required to move slowly, adopt things slowly, take a very conservative approach. That kind of means you can expect for a long time to come a sparse matrix of adopters, right? Not that it'll be a long time, I believe, before the majority of banks will be fully on board and moving in this direction. That means, again, there's enormous value in us aggregating, bringing the demand together, matching up the right applicant to the right product, the right bank, the right lender. That is the heart of what informs our model, is we want to be the leader in this technology. Leader will come with the largest platform, largest volume, most data, most performant models. In our view, we have this model that we're quite happy with.
We hear? Wait, do you have the mic? Let's see. Thank you.
Thank you. I wanted to go back to the model calibration discussion a little bit, maybe to use an actual example, something like student loans, right? When there's policy changes, how does the model adjust to that? So you have the student role, maybe a two-part question. First on when the policy changed, so hey, student loans, you can forbear those, there's no penalty. People stop paying those, the model starts getting that historical data, see student loans are there, but people aren't making payments on those. Does the model start thinking, "Hey, we never have to make payments on these and we don't consider that in the ability to pay"? And then four months ago, this changed the policy again. Student loans have to be paid now. Today, our model suddenly, like we just got one Q data yesterday.
Delinquency has jumped and potential for wage garnishments, et cetera. Policy changes. How does the model calibrate for things like that? Maybe even just while you're answering that, if someone wants to comment on how you're thinking about this whole student loan repayment starting. Thank you.
I think the first thing I'll say is that personally, on kind of like a judgmental basis, I have no idea what the effect of the student loan thing will be on anything. What we've built from a technology standpoint is really the thing I want to emphasize about is it's sort of a fully general solution in the sense that it doesn't assume any particular sort of relationships in the way that macro variables interact with each other. It doesn't assume that a certain level of unemployment corresponds to some level of defaults or anything like that. What it's really doing, as I explained earlier, is it's taken time as sort of a core feature and it has the ability to dynamically interact that with any of the other thousands of characteristics in the model.
As it relates to the student loan example, what would concretely happen in the model and what is concretely happening is there are borrowers that have student loans. Some of those borrowers have been making payments on their student loans. Other ones have not been making payments on their student loans. We know the sort of dates that all these things have been happening, and we also know whether they're paying back on their Upstart loans. That's the training data sort of universe that we observe.
If it's the case that the sort of recent change in student loan policy that requires people to pay their student loans or else they get their wages garnished is causing these sort of significant amount of a noticeable amount of financial distress to some one pocket of these borrowers, either the people that weren't paying or maybe the people that, well, probably just the people that weren't paying, I guess, then those people, because that is a sort of set of people identified by a set of variables in the model that interacted with time would be observed by the model to be something that has changed. If that change is statistically significant, then it just gets dynamically incorporated into the model as a sort of new risk factor that gets priced into how we price loans.
If that doesn't happen and the model says nothing has happened and it doesn't respond to the change at all. The beauty of something like this is that it works for any kind of change that's happening. The reality of macro is that you could read CNBC every day and have a new fear about something that's going to distress consumer credit or some segment of consumer credit, and you'd be right to because literally all the time there are different segments that are going through ups and downs, and the model is able to find all of those things because it's sort of a model of how it does this is fully general.
As long as any of the sort of thousands of characteristics that we know and gather and have about the individual borrowers are able to capture what those segments are, such as whether you have student loans, whether you're paying on student loans, whether you're a government employee, maybe whether you work in a particular industry that's being affected a certain way, those things get interacted with time and just get dynamically responded to.
I can maybe add a little layer to that, Maher, and maybe provide some numbers. In this particular case, about 35% of our borrower base has a student loan of any kind, federal or not. Of those 35%, about 2%-3% of them are in some form of non-current. That is the magnitude of the risk pool. In general, if the magnitude of a potential risk that we've identified is big and we have some thesis to be concerned, whether we have data on it or not, we can act preemptively. We can tell the models, "Look, we don't have any data to understand, but this is a big risk.
If these conditions are met, please price more conservatively. There is an ability at the macro level or even at a sector level for us to intervene when the magnitude is sort of in the noise of the macro and we do not have a thesis one way or the other, we let the machines handle it. That is the sort of general dynamic we have in dealing with these things.
How often do you intervene? Sorry. How often do you intervene like that with the macro, with adding in some type of overlay to the model?
I mean, in a sense, at the macro level, we consistently intervene because we want loans to overperform in benign macros. So whatever the model thinks the macros are, we're consistently telling it to be in the aggregate a little bit more conservative. That is a human intervention. At the micro level with specific things like this, very rarely, because what you generally.
Almost never.
Almost never. Because what you generally find is human intuition. There's a reason these models are good and human intuition is not.
Hey, Sonya. Hi guys. Rob Wildhack from Autonomous. I echo everyone else. Really appreciate the time and the presentation today. You spent a lot of time highlighting improved and new ability to separate, calibrate, new customer acquisition strategies and channels. You're not funding constrained anymore. With all of that as the backdrop, though, you probably originated maybe half of what you did at the peak in terms of an annual or quarterly run rate. As you see it now, what's kind of the rate limiting factor to getting back to and eventually beyond past level of scales where you were doing $4 billion-$5 billion in originations per quarter?
Yeah, the real factor is today our prices on a relative basis are significantly higher than they were back at that time. They are really for two reasons and kind of in this order. One is that consumers are still at a risk level that is significantly higher. What we call the Upstart macro index today is, I believe, 1.49, something like that. That means compared to a long-term run, default rates expected as a sort of function of the macro environment, 49% higher. That has been priced into our loans for a long time. That means prices are significantly higher. Also, the return expectation of either a lender or a credit investor are a few hundred basis points higher than they would have been at that time in that zero interest rate environment.
Put those together, and that means we're approving fewer people relative to them or the people being approved are paying higher rates. That is just that's what the machine that is the machine doing what it is supposed to do. We do not feel anything is wrong with that. We do believe it's inevitable that you and I will at some point begin to tailor off and head back towards 1.0. That, of course, will be reflected in the prices borrowers see. Where interest rates go, that's obviously out of our control. Those are inputs to the system. Having said that, most of our growth in history happens through model improvements that lead to conversion improvements. We do not need to wait for those to get back to our peak. We have much smarter models than we had.
We have much more automation than we had. We have much more separation than we had. We're now growing additional products. I don't want anybody to take my comments to mean we're going to have to wait for you and I. We're going to have to wait for interest rates to go down. That's not remotely the case. On an equivalent basis, that's the difference. That's why we aren't at the level volume that we were at that time. I think right now we're in a place where we do expect to scale quite rapidly. I mean, I think our guidance for 2025 suggests that. In these new products, obviously, we're being very, very transparent about how rapidly these new products are growing. I feel good that we will be at and beyond our current peak in terms of volume on the platform in short order.
David.
Hi, morning. It's David Scharf at Citizens. I'll echo everybody's thanks. David, you may have partially answered this when you gave your anecdote about talking with a bank risk manager and the benefits of aggregation, but I'm wondering, either you or Chantal, can you explain for investors the economic value of investing in a forward-facing brand and running a marketplace and all the complexities of managing funding risks and vicissitudes versus integrating directly and with a ton of lenders who may want either your underwriting, your customer acquisition, your servicing, à la carte, or all three of them? What are the benefits of investing in this forward-facing marketplace and all the risks entailed versus being sort of a white-labeled behind-the-scenes provider of all these services?
Maybe I'll give my take and then let Chantal add to it. I mean, my take is, again, waiting for lenders to adopt a technology that itself depends on volume and scale to get better seems like the wrong approach, right? Where on the alternative, we can actually control our destiny and attract consumers, have a relationship with consumers, have ability to cross-sell to consumers, have a lifelong relationship where they recognize our brand. Because ultimately, look, in any AI system, I don't care if it's generative AI or what we're doing or in 10 other industries, the models that learn most quickly are going to win. I like taking on the customer acquisition because guess what? That's an enormous part of our value propositions. We actually acquire far less expensively than others, and we do it very constructively.
By the way, acquisition gets better as the conversion gets better. In my mind, sitting behind, even if they were all ready to adopt and aggressive to move forward, sitting behind all of them, we would be a very different company. I think the quickest path to creating an unimaginable value in the industry and an incredible company for the ages is going to happen, at least in my mind, through a process where we control and we can communicate what is special about Upstart and what we're doing. We can tell every American consumer, "Come here first. You will get the best offer, the best process guaranteed." That's the place we're trying to get to. That is not possible as an invisible software provider.
Yeah, I think Dave nailed it. All that I would add is this idea that acquisition is actually quite difficult, and it turns out we're fairly good at it. Part of the reason that it is difficult is when individual lending programs are after very specific needs, products, consumers, you have this very individual approach, and marketing becomes very expensive. It becomes hard to scale. The beauty of our marketplace is because we can aggregate all the capital and we can aggregate the demand, we have a much more powerful message to consumers where they can get any type of credit. We can serve a wide variety of credit. We can have higher approval rates because we've aggregated the credit on the back end. I think that allows us to continue to have a really great consumer brand that continues to pay us dividends.
I think this point about the risk manager is also important to remember. It goes back to this question about the competition and where they're at. I would say the typical bank is not fully comfortable with what we do today. When I say not fully comfortable, I mean there's some version of what we do that I think we can get banks comfortable with, but it looks something like the 2015 version of what we do. Each year we add on, the pace at which we're adding new things is generally faster than the pace at which I think most of the industry is comfortable adapting to these new technologies, especially, again, in a fairly regulated space, very strong sort of network effects.
That means that we would have to constrain the pace of technology rollout to the pace at which sort of traditional financial institutions are willing to adopt them. I think that would pretty much be death for the strategy that we've chosen here.
I'll just make one other point. Beyond the efficiency of the acquisition itself, which we think we do very well, but largely reflects a transaction, we increasingly care about the relationship with that consumer over time and the lifetime value that we can extract from it. That's something that you can't really achieve as a software provider.
Hi, thank you. It's Matt O'Neill from FT Partners. Appreciate all the time as well. Sonya, I believe on your last slide in one of the quadrants there, you had revolving. I was hoping you guys could maybe give a little bit more of the vision on how you would think about rolling out a revolving product. Maybe a two-part question, one around sort of the commercial launch and how you're thinking about that, and then two, specific to the model, how adaptable is the model to thinking about underwriting on a recurring basis versus the current use case of the model? Thanks.
On the first one, I think I'll simply say that we are experimenting with it. That's all we have to really say about that right now. In terms of the extensibility of the models, I don't know, Paul, if you have any thoughts.
Yeah, I shared a bit about sort of how naturally extensible they are today, which is to say partially, the techniques are very extensible, but training data is a limiting factor when you get into new product categories, and training data is really sort of a necessary ingredient for any of this stuff to work. I think there's always this thing where you have kind of the whole runway of technologies that you're going to apply to a new product that are lined up essentially almost copy-pasted from everything we've already done. They just can't get going and can't really yield any value until you hit a certain quantum of training data. What you're going to likely see with that or any other product that we roll out is that initially it rolls out in a fairly small way.
There's a period of time when it's not really that amazing of a product, not as differentiated of a product because it needs to accumulate training data. As it does, it hits some significant unlocks and basically just picks up all the proven wins that we've demonstrated in our existing product set.
Partially touches on what I said earlier about the value of our own balance sheet as well, because this is the kind of thing we can do very quickly using our own balance sheet capital. That is why it's such an important tool for us.
Dan.
Hey, guys.
Very recognizable, thank God.
Hey, guys. Great. Dan Dolev at Mizuho. Great presentation. I just wanted to go back to maybe a little bit of Q1 and sort of the win on Walmart, if that's okay to ask. Maybe just a little more in depth, what did Walmart see, maybe just the background of the conversations and the capability of your AI that made them choose Upstart? Just any color would be great.
Sure, I can take that. Our relationship with OnePay, we're very excited about. I think naturally, as we think about partnerships generally, people often choose us and our partners choose us because we are able to serve a wide variety of their customers. Ultimately, we just have the best product in the market. We have the great rates, and we can show the great process. A lot of people just care about giving their consumers the best. It's pretty simple. I think in many of these partnerships, we actually go head-to-head with competitors sometimes and compete it out, battle it out, and we win. I think that's what happens in this scenario and others, is just we fundamentally can serve more of their consumers at better rates.
I think that type of partnership is something we believe we can repeat where it is really our core product. It can be co-branded and distributed through somebody else. One of the real advantages is when you are that partner and you're somehow or another trying to provide a credit product to your end customer, there's always this question of who can you approve, who can you not approve, will the rates be okay, how long will this take, the myriad of things, because you care about your customer, you care about whether you're sending them down a good path or a bad path or what's going on.
All the things we tell you about our business, the experience, having the best rates for everybody, that puts us right at the front of the line in terms of being the type of partner, if you want to deliver a consumer credit product to your customer, that we can work with. I think that explains that relationship today. I think it's a great footing for us to bring our products to market through others. Again, we always want our brand in the mix. We do think the consumer should know who Upstart is and experience it. These types of partnerships allow us to do that.
Hello. My question is around automation and customer acquisition cost. What happens in that 10% that's not automated? Where do you see the customer acquisition cost leverage level kind of optimizing, especially as you're thinking about these additional revenue sources going forward?
The 10% are really when we do not have, our models do not have sufficient confidence that this is not necessarily that this isn't a bad actor, that some information provided may not make sense. There will always be, you'll never go to 100%. There's just enough bad actors out there that need to be stopped. When you're stopping bad actors, you will always, sadly, stop some good actors as well, that after a bit of a process and human involvement will be approved. That's what that 10% really reflects. It's people who are perfectly good borrowers, but our systems were not yet smart enough to separate them from the bad actors. That will change over time. Customer acquisition. What was that question again?
Oh, I think the way to think about that for us is if we're in a place we're happy with the unit economics, which we certainly are today, we don't really have a desire to drive it down. We would rather pour that into growth. You can think of that as when you're in the happy place in terms of unit economics, if you get more wins and you're converting more, in the end, we're going to end up spending more or in one way or another bringing more people to the place. It's just like we don't want take rates to go through the roof. We don't need acquisition costs to go through the floor. We like them in a really healthy place where our margins are solid, we're a profitable company.
From there on out, we just keep pouring the value back into the product so it's better, or we reach out to more customers. That's really the way to think of it. It's not intended to go to zero.
Okay. Any last questions? All right. We appreciate your engagement. Do you want to say a couple of closing thoughts?
Yeah, just thank you. You gave us great attention today. I know there's a lot of information we threw at you all. I want to thank you for all being here. The management team will be around for another 30 minutes or so. If you want to hang out and speak with any of us, please feel free to do so. Thanks to everybody online who joined. Hope you all have a great rest of the day.