Upstart Holdings, Inc. (NASDAQ:UPST) Q3 2025 Earnings Call Transcript

Upstart Holdings, Inc. (NASDAQ:UPST) Q3 2025 Earnings Call Transcript November 5, 2025

Operator: Good afternoon, and welcome to the Upstart Third Quarter 2025 Earnings Call. [Operator Instructions] As a reminder, this conference call is being recorded. I would now like to turn the call over to Sonya Banerjee, Head of Investor Relations. Sonya, please go ahead.

Sonya Banerjee: Thank you. Welcome to the Upstart Earnings Call for the third quarter of 2025. With me on today’s call are Dave Girouard, our Co-Founder and CEO; Paul Gu, our Co-Founder and CTO; and Sanjay Datta, our CFO. During today’s call, we will make forward-looking statements, which include statements about our outlook and business strategy. These statements are based on our expectations and beliefs as of today, which are subject to a variety of risks, uncertainties, and assumptions and 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 a 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 earnings materials, which are available on our IR website. And with that, Dave, over to you.

David Girouard: Thanks, Sonya. Good afternoon, everyone, and thank you for joining us. To kick things off, I’ll share my perspective on our business. Upstart today is a dramatically stronger company than it was just a few years ago. Our technology, our business, and our teams have never been better. The opportunity for AI and credit is unimaginably large, and there’s no one better positioned than Upstart to lead this $1 trillion industry to this exciting and inevitable direction. Now turning to Q3. Upstart continued to execute on its 2025 game plan of rapid growth, profits, and AI leadership, all under the auspices of exceptional credit performance and precise macro handling. In addition to 80% year-on-year growth in transaction volume and 71% revenue growth, we were nicely profitable once again.

In fact, Q3 GAAP net income grew by a factor of 6 over the prior quarter. Consumer demand for Upstart continued to grow rapidly with more than 2 million applications submitted in Q3, up over 30% from Q2 and reaching the highest level in more than 3 years. Despite this awesome demand, transaction volume on our platform was less than we anticipated. Our risk models responded to macroeconomic signals they observed by moderately reducing approvals and increasing interest rates. This drove a reduction in our conversion rate from 23.9% in Q2 to 20.6% in Q3. If you follow the Upstart Macro Index, you would have seen that this macro indicator ticked up modestly in July and August, which is essentially what our model responded to over the course of the quarter.

We believe this to be nothing more than a speed bump with UMI reverting to lower numbers since. To be clear, we see no material deterioration in consumer credit strength. And in fact, we’ve seen recent signs of improvement. You can and should expect that our models will always do their best to price prevailing risk appropriately. Precise and rapid tuning to changing economic conditions is a foundational capability of Upstart AI, and we’re confident this precision is winning hearts and minds for Upstart in the credit market right now. The system is behaving exactly as it was designed. At a minimum, our Q3 results should give you confidence that we don’t sacrifice credit performance to achieve transaction volume targets. Turning to our newer products, which include small-dollar loans, auto, and home.

These offerings continue to improve and mature, accounting for almost 12% of originations and 22% of new borrowers in Q3. Transaction volume for auto, home, and small-dollar each grew in the range of 300% year-on-year. Our auto retail business, in particular, has really begun to accelerate. We more than doubled the number of live lending rooftops on Upstart in Q3 compared to the prior quarter. Transaction volume for auto retail also grew more than 70% sequentially. We expanded to 4 new states in Q3 and made some significant improvements to our software. This is really a breakout business for us. Additionally, we’ve been quietly working on a hybrid product called an auto secured personal loan that’s beginning to gain traction. As it relates to our home business, beyond continued process innovation, our unique partnerships with banks and credit unions mean we offer the best rates to the primest borrowers compared to other fintechs by as much as 300 basis points.

Best rates and best processes are what we’re all about. Our continued process and automation breakthroughs in our secured products, meaning home and auto, give us confidence that they will be real growth drivers for Upstart in 2026. Finally, with respect to funding on the Upstart platform, we’re in an exceptionally strong position in our core business with significant excess capacity. On the bank and credit union side, we added 7 new partners, our best quarter for new logos this year, and we reached a new all-time high in monthly available funding from these partners in Q3. On the capital market side, we continue to have exceptionally strong execution with our institutional partners. Having signed our first agreement in 2023, we now have 10 active partners.

In August, we renewed one of our largest partners for the second time. And importantly, Upstart has 100% retention of all private credit partners to date. We believe that we have the industry’s best AI for responding rapidly and precisely to changes in the environment, and this is a central reason why our partners have confidence in us. In September, we also issued a securitization with strong demand, leading to significant oversubscription of all classes despite upsizing and tightening of spreads. This ABS deal involved 30 investors, including 7 first-timers, demonstrating the strength of Upstart’s reputation in the market. We’ve also continued to make progress securing third-party funding to support our newer products. We’ve signed 17 partner agreements this year, including 9 signed in Q3 alone, and expect to ramp these partners into production this quarter and next.

All in all, we’re all systems go to finish the year strong and get ready for what we think will be an amazing 2026 for Upstart. To wrap things up, we’re making rapid progress as the leader in AI-powered credit. The somewhat complicated macro economy we all see today is, in my view, the perfect opportunity to demonstrate the strength of our AI platform, and we’re doing just that. While legacy financial services execs continue to ponder the use of AI and credit, the Upstart platform has now generated more than $50 billion in AI-powered loans since inception. Unlike other AI platforms, we generate our own training data with more than 98 million borrower repayment events to date, with about 105,000 more repayments due each day, driving improved separation and model accuracy.

This is enabling us to build quickly toward a future of always-on credit, where every American is persistently and precisely underwritten, providing them with the best rate anywhere, 24/7 credit access right from their phone, with little to no process. That is a proposition and a future that we are betting on 100%. With respect to the investor community, I feel more than ever that those who stay will be champions. With that, I’ll turn things over to Paul Gu, my Co-Founder and Upstart’s Chief Technology Officer. Paul?

Paul Gu: Thanks, Dave. I’ll start by addressing the model conservatism we experienced in parts of Q3. Over the past few years, one of our biggest advances has been our model’s ability to respond with speed and precision to changes in macro conditions. This progress stems from a suite of techniques that are proprietary and critical to resilience through a diversity of economic environments. A few months ago, that led the model to tighten on credit while certain risk signals were elevated before recently normalizing. That model behavior partially reflects irreducible volatility in the outside world, but is also a function of our model design and sampling variance, both of which continue to improve. Since the start of Q3, the improvements we’ve made to our calibration methodology are expected to cut unwanted month-to-month volatility in model calibration-driven conversion changes by about 50%.

As we continue to innovate on our model calibration techniques, we’ll increasingly be able to minimize conversion volatility in the business while delivering on target credit performance. Beyond calibration, Q3 was a productive foundation-building quarter with a number of technology improvements that will power our next phase of growth. I’ll start with personal loans. First, we took another leap forward in our evergreen engineering quest to lower latency in pricing loans. By parallelizing another major portion of loan pricing, we reduced end-to-end latency by as much as 30% and are now rolling it out platform-wide. Reduced latency unlocks the ability to build larger and more complex models as well as make use of the ever-growing data set of 98 million repayment events that Dave mentioned.

Next, we launched a true machine learning model to optimize take rates. It is our intention to capture value in relation to the value that we create for our borrowers. We expect that this framework, over time, will unlock a significant improvement in our ability to monetize model wins that benefit borrowers who are already vested for as well as increase our competitiveness in new segments where we’re still establishing our edge. In the domain of customer acquisition, our ability to utilize digital partnership channels that relied on cloud environments or APIs to target offers was historically limited by the complexity of our underwriting models. This quarter, we built a programming language agnostic framework for data transformation that makes it much faster to translate our models to work with any partner ecosystem.

We also worked with a key partner to enable larger model sizes in their cloud environment, allowing more of Upstart’s unique underwriting algorithms to be used in targeting. On our direct marketing channels, we developed a proprietary technique to target marketing spend based on causal impact to conversion. Compared to our prior, more textbook technique, early results show a 50% uplift in incremental originations from the same level of spend. Advanced AI and underwriting ultimately need equally advanced AI and acquisition, and the successes this quarter were a big step towards that. I’ll wrap up my remarks with a few technological highlights that are driving growth in newer products. We’ve made rapid progress automating the process of getting a HELOC.

When we launched instant property valuations back in June, we automatically approved less than 1% of home loans. Since then, automatic home loan approvals have grown to 10% in September and about 20% in October. While we’d love to simply automate away almost all the documents like we have in personal loans, the world of home loans is just less digital, less standardized, and there are more requirements. So for the next leg of improving the HELOC funnel, we’ve begun using multimodal AI to do the work of human document reviewers in real time. Our rapid pace of process improvements makes me optimistic that we’re on a path to an industry-leading home equity product. Additionally, our small-dollar relief loans continue to make rapid progress. In September, we launched Instant Funding for the first time.

Most borrowers who qualify for instant funding see funds in their bank account within around 90 seconds of approval. While small installment loans at bank-friendly APRs are a wonderful innovation, you should expect to see a lot more from Upstart in this area in the coming months. Thanks to the team’s work this quarter, I’m more excited than ever about our upcoming pipeline of technology wins. With that, I’ll turn it over to Sanjay. Sanjay?

Sanjay Datta: Thanks, Paul, and thanks to all of our participants for sharing some of your time with us today. I’ll now spend a bit of time reviewing our Q3 numbers. At a headline level, we were pleased to finish the quarter with healthy annual and sequential revenue growth as well as extend our run back to profitability. Within that, our transaction revenue this past quarter was marginally short of expectations as our models expressed some temporary conservatism in piloting the current environmental dynamics, but this was largely offset by growth in interest income from the strong return performance of our balance sheet. Margins and take rates have remained steady, and credit performance continues to land right on target. We are carrying a larger-than-normal loan balance on our books as we work towards closing a number of deals across all of our new product areas, which will both reduce R&D carrying balances and flow new volume directly to our lenders and investors.

We remain pleased with the progress of those various conversations and expect to have tangible outcomes on this front by the end of the year. More broadly, third-party capital in our core unsecured lending segment remains readily accessible, handily outstripping our borrower supply, and is currently not in any way an impediment to growth. Spreads on our third-party capital continue to compress, partially a result of the competitive funding environment and partially as an expression of investor confidence in the steadfast performance of our credit. With respect to borrower approvability, our model has exhibited some recent caution in response to a UMI run-up of almost 0.2 points that happened over the course of the past quarter before more recently subsiding as well as to a rising trend in repayment speeds, which is generally an encouraging longer-term signal for credit, but in the near term, limits interest income from current loans and requires higher coupons to compensate.

In all of this, we, as always, care, first and foremost, about getting credit performance right, which will always result in the best long-term outcome for our business. We have an inherent belief that AI models are better suited to navigating a complex and changing environment than human intuition, and we have demonstrated the discipline to heat them even when they express a bias toward moderation as now. If the currently observed higher repayment speeds and easing consumption growth are indeed indicators of imminent credit improvement, these could represent the long-anticipated tailwinds that could accelerate growth prospects heading into next year. In the meantime, we continue to be guided by the North Star of prudence in the underwriting of risk on behalf of our lenders and investors.

With this as context, here are some of the financial highlights from Q3 of 2025. Total revenue for Q3 came in at roughly $277 million, up 71% year-on-year and 8% sequentially. This overall number included revenue from fees of approximately $259 million, which was up 54% year-on-year, but short of our internal expectations by roughly 6%, mainly for the model-related reasons previously mentioned. Within fee revenues, our servicing revenue stream continued its steady growth clip at a 10% sequential rate. Much of the shortfall in expected fees was counterbalanced by higher-than-expected net interest income of approximately $19 million, resulting from continuing strong return performance on a loan balance that remains temporarily elevated. To reiterate, we are aiming to enter into a phase of reducing our R&D-related balance sheet holdings, which we anticipate will gain steam in Q4 and continue into 2026, and we would expect this revenue item to moderate as we are successful.

The volume of loan transactions across our platform was approximately 428,000, up 128% from the prior year and 15% sequentially, and representing approximately 300,000 new borrowers. The average loan size of approximately $6,670 was 12% lower than the prior quarter from a combination of borrowers requesting lower loan amounts, a model exercising increased caution in improving loan sizes, and a mix shift towards smaller loan products and risk rights. Our contribution margin, a non-GAAP metric, which we define as revenue from fees minus variable costs for borrower acquisition, verification, and servicing as a percentage of revenue from fees came in at 57% in Q3, down approximately 1 percentage point from the prior quarter and versus guidance as lower conversion rates created some mild upward pressure on both acquisition and onboarding unit costs.

In total, GAAP operating expenses were around $253 million in Q3, roughly flat to Q2. Expenses that are considered variable relating to borrower acquisition, verification, and servicing were up 11% sequentially relative to the 15% increase in volume of loan transactions. Fixed expenses were actually down 7% quarter-on-quarter, largely due to a reduction in compensation-related accruals. Q3 GAAP net income was approximately positive $32 million, well ahead of expectations and reflecting outperformance on net interest income, reduced fixed costs, and a $7.2 million gain on our convertible debt repurchase. GAAP earnings per share were $0.23 based on a diluted weighted average share count of 110 million. Adjusted EBITDA was roughly $71 million, also corresponding ahead of expectations.

Adjusted earnings per share were $0.52 based on a diluted weighted average share count of 125 million. We ended Q3 with approximately $1.2 billion of loans held directly on our balance sheet, up from just over $1 billion in Q2. As shared last quarter, we have multiple new products simultaneously exiting R&D status and entering the scale-up phase. And our business development efforts this past quarter have been aimed at putting in place the third-party capital arrangements that will enable us to shift away from balance sheet funding on these emerging products and release back our invested capital. We are very pleased with the progress of these efforts and believe that we are on a path to putting multiple agreements in place across all of these new product lines, which will set them up to further scale in 2026.

Exact deal timing is, of course, not perfectly predictable, and it is important for us to do the right deals with the right partners. So we will take the necessary time to ensure we are well set up on this front for next year. In the meantime, returns from our balance sheet holdings continue to be strong, delivering healthy spreads above market base rates, as can be seen in the data on Page 23 of our earnings presentation. As we look to Q4, the broader economic backdrop for credit remains favorable in our estimation. Decelerating personal consumption growth is a signal of improving credit health, if perhaps counterintuitively so. Against this, we perceive a labor market that has remained at full employment since lockdown, meaning there are as many open jobs as job seekers in the economy, as well as a muted impact of the recent tariff policies on inflation and a gradual easing of the monetary climate.

In this scenario, we once again assume a stable UMI as well as holiday seasonality typical of Q4, which tends to serve as a mild headwind. We expect the impact of any further rate cuts this year to both improve consumer financial health and lower investor return requirements. But at this stage, any such effects would not be felt until the new year. In this environment, we will continue to produce model and targeting accuracy gains as well as automation wins to grow our top line. Our net interest income will start to benefit from the returns on our committed capital investments that were made in prior years. Now that our P&L is once again back to profitability, we will plan to begin dialing up our forward investment into customer lifetime value by slightly moderating take rates in exchange for higher origination volumes and higher repeat transactions in the future.

And as usual, we will expect to continue our fixed expense discipline in how we manage the cost side of our business. With this context, for Q4 of 2025, we are expecting total revenues of approximately $288 million, consisting of revenue from fees of approximately $262 million and total net interest income of approximately $26 million. Contribution margin of approximately 53%, GAAP net income of approximately $17 million, adjusted net income of approximately $52 million, adjusted EBITDA of approximately $63 million, with a basic weighted average share count of approximately 98 million shares and a diluted weighted average share count of approximately 111 million shares. For the full year of 2025, we now expect total revenues of approximately $1.035 billion, consisting of revenue from fees of approximately $946 million and net interest income of approximately $89 million.

Adjusted EBITDA margin of approximately 22%, and we expect GAAP net income of approximately $50 million. Before we move to Q&A, I will take the opportunity to thank all of the various teams across Upstart for their hard work and continuing dedication to our mission. And with that, operator, over to you.

Q&A Session

Operator: [Operator Instructions] Our first question comes from Dan Dolev with Mizuho.

Dan Dolev: Just wanted to ask a quick question on the application demand. It seems very strong quarter-over-quarter. Maybe, Sanjay, if you can comment on the strong demand in the third quarter, and then maybe just like tie it all into the guidance, which was a little bit below what we were expecting and below the guidance in 2Q. So, how do you square these 2 things together?

David Girouard: Dan, this is Dave. Yes, as we said in the remarks, we grew applications about 30% quarter-on-quarter, which was ahead of the origination, the transaction volume quarter-on-quarter. And really, a lot of things came together in terms of just marketing programs and cross-selling and all these things. So the application growth is certainly great to see. I think what it really highlights is that our model took a step towards conservatism during the third quarter, just based on seeing macro factors. And I think that is just a natural thing we might expect. As we said, it’s since reverted, but it was a period of time where it saw signals, and it was moving quickly. I think maybe overreacting. I think in some sense, having a model that overreacts is better than having ones that underreact because it did revert.

But I think it is useful to point out that the application volume was quite strong, our strongest in 3 years, and grew quite a lot. And I think that’s a very healthy statement for the business, even if it didn’t in Q3 transfer to as much volume as we expected.

Operator: [Operator Instructions] We’ll go to our next question from Kyle Peterson with Needham.

Kyle Peterson: I wanted to ask specifically in auto, obviously, there’s been some high-profile bankruptcies and kind of negative credit events in the space. Have any of the headlines or news impacted your expansion plans or how conversations with customers are going? Or I guess just how has the recent news and events impacted how you guys are viewing things and progressing in auto right now?

David Girouard: Yes. None of that has had a direct impact on us for sure. We have not seen that type of couple of examples that were out there of fraud zone activity. So, I don’t think it’s anything that we would describe as widespread. It’s not something from our perspective that is widespread. I think when you have examples like that, it does create a little bit of caution in the market. So banks or others providing senior financing probably do a bit more diligence, et cetera. But I don’t think there’s any wholesale change in the market, but that is the nature of it. A couple of the larger banks got bitten on that particular auto lender. But we’ve been pretty rigorous about building processes to make sure we’re effectively underwriting the dealership themselves and mitigating risks against dealer activity that’s not what we want.

So, this is an area that I think we’re handling well. We have not seen any major issues. But again, I think whenever you read headlines, it does cause a little caution in terms of increasing amounts of diligence or questions that need to be asked, et cetera, but that’s part of the course

Kyle Peterson: I guess as a follow-up, I wanted to specifically ask about what you guys are seeing in the super prime segment. I guess, looking at the originations, it was down a little bit sequentially. So, I guess, was that where the model tightness that you guys called out? Did you see a little more in that 720-plus FICO score versus the core product? Or is there more competition there? I guess, just like what are you guys seeing? And is any of that concentrated more in the super prime? Just trying to think how we should square that with some of the positive commentary on demand and funding capacity from your bank partners.

Sanjay Datta: Kyle, this is Sanjay. I think it’s a combination of things. I mean, the thing you pointed out is definitely a factor, meaning our models reacted generally to some macro signals as Dave described. I think that was true of the primary segments as well. In fact, if you look at the segmentation of our UMI, the subprime consumer is actually at a relatively low UMI, probably somewhere around 1.2, 1.3. And if you start to look into the segments in the low to mid-700s, it’s quite a bit higher. So, there was definitely a model impact. I think it’s fair to say that it’s also a very competitive segment, and we see other growth numbers in that segment, and they’re healthy. So there’s a price impact or an aspect of competition as well.

Operator: [Operator Instructions] We’ll take our next question from Pete Christiansen with Citi.

Peter Christiansen: I want to follow up on some of the earlier questions. As it relates to the improvements that you made with your marketing channels, which sounds pretty exciting and is obviously illustrated by the higher number of applications. Is there a way to at least get your sense for the quality of these leads? I know the AI system was a bit conservative this quarter. So, taking that into account, do you think that the quality of applications has remained the same or maybe improved or what have you with these new capabilities?

Paul Gu: Peter, this is Paul. Yes. So I spoke in my prepared remarks nice wins we had in applying AI to customer acquisition. And the way you can think about those wins is ultimately, at the point of customer acquisition, we are somewhat indifferent between selecting for people who have a high propensity to apply and people who have a high propensity to convert or be approved. Ultimately, it’s the product of those 2 things that we’re solving for, of course. And so the improvements we make can help, one, both, or ultimately just the product of those things. So obviously, we did have a larger increase in applications relative to where the final originations count ended up. So I think mechanically, you can infer from that change in the likelihood to convert through the funnel.

Of course, the conversion rates are lower. Now that’s in large part, as we already said, because we were knowingly making a choice with our model to be a little bit more conservative on the credit side in earlier parts of the quarter. So relative to that model, of course, we did end up marketing to people who were a little less likely to be approved or a little less likely to convert, but that’s not necessarily a chosen strategy.

Peter Christiansen: Then my second question, non-prime auto has had elevated delinquencies even before some of the more noticeable news events that have been happening in the space for a couple of months now. Dave, I’m just curious, if we were to see an improvement in that specific category, would that be a needle mover for Upstart’s auto originations?

David Girouard: Yes. We’ve seen very good credit performance in auto. So we do feel good that our models are working that become calibrated. To the extent others are having issues or what have you, maybe some are withdrawing from the market; those can be good things. Maybe it suggests a transition or an inflection point in the market. So for us, it is just really important. We get calibration, we get more separation. We bring partners on. We keep refining the processes. And I think it’s going really well on all fronts there. So I think in 2026, we do feel very optimistic that the auto business as a whole is going to be a contributor. Again, a little disruption or a little noise in the market when you’re new like us to it, can be a good thing. It means there’s an opportunity when things are shifting.

Operator: We’ll move to our next question from Simon Clinch with Rothschild & Company, Redburn.

Simon Alistair Clinch: I wanted to just jump back to the first question, really, about the application volume growth that you saw. And just, Sanjay, if you could just remind us what you said about what’s implied in the fourth quarter? Because it sounds like you’re assuming that the conservatism in the model is going to continue in the fourth quarter, despite your comments around the UMI actually starting to show some signs of improvement. Is that correct?

Sanjay Datta: Simon, I guess I would note that the improvements in UMIs are materializing. As usual, we are conservative and want to watch them bake, and we’re already past the month of October. So some amount of Q4 was impacted by that UMI rise as well; even though it is now subsiding, we will, of course, follow it with some lag. So, I think that what we described is the model impact in Q3, even though it appears to be abating, will impact Q4 as well.

Simon Alistair Clinch: Just as a follow-up, then, when we look at the broad demand for personal loan growth, I mean, the kind of view I’ve had through most of this year, and I think is consensus view is that there’s a lot of demand for just refinancing credit card debt. Is that still very much the case that’s really driving that personal loan demand? Or are we seeing that demand broaden out into other drivers?

Sanjay Datta: I think refinancing debt really continues to be the dominant use case for personal loans. But it is very much the duct tape of credit. It’s useful for so many things. And so there’s a very long tail of ways that people use personal loans. I think in some cases, because the process is so much simpler and the rates can be quite competitive, that it does compete at some places with secured loans, whether that would be to buy a used car off of a website or what have you, places where you might otherwise, or home improvement, where you don’t want to get a HELOC. So I think an unsecured loan, if it’s fast, easy, and the rate is competitive, will always be very, very broadly useful to the consumer.

Operator: We’ll go next to Patrick Moley with Piper Sandler.

Patrick Moley: I just have one on the balance sheet expansion you saw in the quarter. Just wondering how conversations with some of the potential funding partners of the R&D products have trended recently? And then you touched earlier on the auto, some of the credit issues we’ve seen recently in auto, and how that’s impacted the consumer. But has there been any contraction in demand from any of your private credit partners there? And then I understand that they’re waiting to see how the portfolios season in some of those R&D products. Is there anything you can share with us there on how they’re feeling about that and how those conversations have gone?

Sanjay Datta: Patrick, this is Sanjay. Yes, as we said in the remarks, we’re very pleased with the direction of all of those conversations. We’re obviously having them across a number of different new product areas right now. I think appetite is good. These are large deals, a multiyear time frame, and a large check size. So there’s a lot of diligence in these conversations and these processes, and they are not perfectly predictable in terms of timeline. But with respect to progress, I think it’s all going well. We’re excited about all of it. On the auto side, in particular, I don’t think there are any concerns about credit, to be honest, at least in the loans that we’re producing, I think the credit performance is pretty clear.

As Dave mentioned, there’s some broader noise about fraud in the space, and I do believe that has probably lengthened timelines in terms of these processes. Everyone’s diligence lists have sort of doubled and tripled in size. And so that’s sort of a component. With respect to timing, I don’t think it’s really changed the motivation or the appetite at all with respect to the specific conversations we’re having. I do believe we now have enough seasoning in our portfolio for people to look at our loan cates and get a really good sense for calibration and for how the credit is performing. So it’s really just deal processes, legal processes, getting in place financing, bank relationships, et cetera. So they’re heavy lifts, but I think we’re very happy with how they’re going.

We’re very excited about the partners we’re talking with. And as we said, we hope to have some tangible outcomes for you guys to digest pretty soon.

Operator: And our next question comes from Mihir Bhatia with Bank of America.

Mihir Bhatia: I wanted to start with the conversion rate. You talked a little bit about the conversion rate being impacted by higher UMI. Is that the primary factor? Are there other factors that maybe we aren’t seeing or you aren’t seeing from the outside inside the models that’s driving it? And then just on conversion rate, Paul, I think, mentioned –touched on limiting variability in the metric going forward. Can you talk about that some more? And if there’s a particular level that should stabilize that? Like is this low 20s percent the right level?

David Girouard: Yes. Dave, I’ll cover the first half of the question, and then Paul can answer the second. No, really, the conservatism in the model is, from our point of view, pretty much the dominant driver of the change in conversion rate. So it comes in the form of a small fraction, fewer people approved, the rates they’re approved at being a little bit higher, which means just marginally less likely to take that load, and then sometimes the approved loan size is a little smaller. So that is the basics of a slightly more conservative twist in the model. So again, we don’t believe this is anything sustainable. And we do think that we’ll get to a model that’s a little less responsive, honestly, and maybe overresponsive in this particular case. But no, there’s no other factor going on, as you saw the application volume is quite strong.

Paul Gu: Then, on the second part of the question, about the reduction in volatility around conversion rates, and specifically around the macro calibration contribution to conversion rates. So, a few things to understand about this. The first is, as Dave said, one of the single largest contributors, almost every quarter, to the overall conversion rate, which is the state of macro conditions. If borrowers are generally financially healthy, that’s going to be helpful. If borrowers are struggling, that’s going to be unhelpful. And that’s just because, of course, approvability is such a big, immutable part of conversion. I want to point out that there is a second component of conversion, which is also not necessarily normative. It’s not good or bad, and that’s the mix of applicants.

We talked a little about this earlier in the question about targeting and what kinds of people come in the door. And there’s always some trade-off between propensity to apply and propensity to be approved or converted. And so there’s always an optimization going on there. And I think that’s neither good nor bad. It’s just that we do what’s optimal for the business. And so that can cause it to move a little bit. But the thing that I was referencing earlier in my prepared remarks about what happened this quarter and the improvements we’ve made that we expect to be durable and lasting with respect to reducing variance on this metric specifically has to do with managing how the model responds to the latest signals in macro. So over the last few years, one of the things that we invested the most heavily in was building our models in such a way that we think they are the fastest, most precise at responding to the latest patterns in borrower repayment, including at the macro level.

So if it’s like federal employees or if it’s like service sector workers or if it’s high primness borrowers or low primness borrowers that are being impacted, or it’s everybody being impacted by a big macro event, we want our models to be the very fastest at responding and respond as precisely as the data allows. And so we’ve made a ton of progress towards that. We’re very proud of the sort of system we’ve designed and built. But one of the side effects of that system is that it can be a little overly responsive to the latest changes. And that, in addition to being responsive, there’s always some kind of sampling and measurement error. You can think about what we have, of course, a large amount of data, but relative to all people in the U.S. or the whole economy, it’s still a relatively small sample.

So there’s a natural statistical sampling error that comes about from that. And we were doing a lot of work this quarter on understanding how much natural error there is in the match between the sample and the actual levels of calibration. Then we devised some techniques to be able to shrink that measurement error by about half, so that we don’t have as much what I call unwanted variance in this metric. We really just want the model to respond to real changes as opposed to changes that are just measurement error, and we were able to reduce that measurement error by a very significant amount this quarter, which means that in future periods, we expect that all else equal, we will see less volatility in our conversion rates as affected by macro.

So that’s good.

Mihir Bhatia: And Sanjay, I think you also called out that repayments have increased in the script. Any theories on what is driving that? Are those folks refi loans away from you at a lower rate? Is that the borrower’s financial? He is just improving, so the people are paying off their loans faster. Can you just talk a little bit about what’s going on there and just the credit implications of that? Have you seen delinquency rates already move because of that?

Sanjay Datta: I mean, as we said, it’s an empirical observation that repayment speeds have increased. It seems pretty broad. I mean, I think there are a lot of theories as to what’s behind it, but we don’t know for sure, obviously. It seems to be broader than just one specific use case, meaning I don’t think it’s just a refi boom. It seems to be happening across both partial and full prepayments, which would imply that it’s something broader than just a spike in refinancing. As we said, in the broader scheme of things, this is typically a good thing. When repayments are happening faster, you’d expect that it’s on some level of reflection of improving underlying consumer health. You’d expect it to be inversely correlated to defaults over time.

So that’s what we would like to see. But in isolation, with all else constant, repayment happens faster. In the immediate term, it means there’s a little bit less interest to be earned on the loans to offset the defaults. And so in the immediate term, your model becomes a little bit more conservative on pricing. It puts a bit more coupon into the loans. So that it compensates for the fact that the duration of the loan has become shorter in a sense. So, there’s a bit of an immediate conservatism by the model. But I think in the broader scheme of things, we’re pretty excited to see it because it means that on some level, personal fiscal situations are probably a little bit more stable.

Operator: We’ll move next to Reggie Smith with JPMorgan.

Reginald Smith: The origination number again. But I guess my question is thinking about — obviously, there are 2 components to the conversion rate are what you guys are approving and then the consumer acceptance. I was curious if either had an outsized impact on the conversion rate. And then I was also curious, as you look at your application flow, much of it, do you have a sense of what is shown, I guess, comparison pricing with other loan products? So, like, I don’t know if your loans are showing against LendingClub or SoFi or something like that, if you had a sense of that mix. And the reason I ask both of these questions is that obviously, those 2 companies had very strong origination trends this last quarter. And I’m just trying to figure out if there was a share shift, if you guys were fighting with one hand, time you back because of your model, like just trying to sort through all that stuff.

So anything you can provide there would be helpful. And I have one follow-up.

Paul Gu: Yes. This is Paul. Yes, the conversion changes were predominantly related to our model’s level of conservatism, so reflected in approvals primarily, and that tends to be the single most sensitive metric when you flip someone from approved or denied, you have a 100% decline in their relative conversion rate. So that’s the thing that’s most sensitive and tends to dominate changes in the metric, and that’s what happened in this particular time period.

Reginald Smith: And so, I guess you were like declining super prime. I think you talked about there being some sensitivity in that area. Is that the right way to think about it?

David Girouard: No, declines don’t happen in the super prime area. The rates just move up a little bit for somebody who would be at the end of the spectrum. It’s at the other end of the spectrum where there are a lot of declines. So the combination of those 2 things is what really amounts to a lower conversion.

Reginald Smith: And then, if I could ask one more. Just thinking about the HELOC product, and I know it’s early days, but how should we think about the, I guess, day 1 economics or take rate for that product relative to, I guess, your base corporate average?

Paul Gu: Reggie, let’s see. I mean, I think in the past, we’ve alluded to the fact that take rates will be healthy, but a little bit more modest than in PL, but on much larger loan sizes. So, without quite knowing exact numbers yet, maybe you could think about a take rate that’s maybe roughly half the amount, but a loan size that’s certainly far more than double.

Reginald Smith: And then the last one, I guess, nothing to call out from a credit performance in your book, despite what you guys are seeing in the UMI, just to be clear.

David Girouard: That’s correct. We’ve seen exceptional credit performance, and that’s kind of the whole reason for UMI is to make proper adjustments. Also, Reggie, your question that we didn’t get to was what others are doing in the market and if they’re growing at higher rates at this particular period of time. We obviously don’t know what’s behind their rates. We don’t know what their models look like, how quickly they respond to signals they see. But there’s no question, if you just look at the nature of lending, that there’s always a way to grow. In our view, the model is always right. The model is going to tell us what’s proven at what price, and we don’t overrule the model. So I think that’s probably our way of looking at it.

Reginald Smith: It sounds like, if I’m hearing you right, that the model may have given you guys a false negative, and like a blip, and things are better than what may have been showing up a couple of months ago in the model.

David Girouard: I mean, we don’t know false negatives. It may be something that’s helpful to us down the road, that it saw what it saw and it priced what it priced. So it doesn’t necessarily mean it was, in any sense, a false negative. It’s a constantly learning system.

Paul Gu: No, nothing to add.

Operator: And our next question comes from James Faucette with Morgan Stanley.

James Faucette: Just a couple of quick follow-ups for me. Can you give any specificity to what elements of the model kind of you saw weaken and then subsequently improve, or other indications? Just trying to get a sense of where your systems may have been looking versus the broader market.

Paul Gu: Yes. I think, first of all, we definitely wouldn’t describe the model as weakening. As a reminder, our primary performance metric for the model is model separation, and our separation accuracy metrics are our highest ever. The other metric that we track very closely is what we call model calibration, and that’s about how the question of credit performance. And as has been said several times, credit performance really has been exceptionally strong for us in this time period. And so what was weaker in this period was the model’s ability to approve as many people or convert as many people. And that certainly was a direct result of the increased conservatism that resulted from the model observing a couple of months of elevated risk signals in various pockets of borrowers.

And so that you can think of more of what we call a macro change that the model was responding to. I think with the benefit of hindsight, you could call that a bit of a false negative, I suppose. But of course, I think in the moment, there is a correctness to reacting to the signals that you’re seeing. And I think we directionally think that is the right thing to do. That’s what we’d like our model to continue doing. I did say that some of that reaction, we think, was due to a certain natural noise in what I call sampling or measurement error. And we did come up with some really good ways to reduce that. And so that noise going forward will be a whole lot less, which is a really, really great technical win for us. But ultimately, there is some level of directional responsiveness that we always want the model to have to the latest changes in what’s going on in the world.

And if that means that for a month or 2, the model gets more conservative, we think that’s just the right thing to do.

James Faucette: And then, as you look forward to the December quarter and as you’re forecasting, how are you thinking about exit rates? You made it pretty clear that you think that there will be a little bit of lagging or continuing effect as we go into the fourth quarter. But are you expecting that by the time we get to the end of the quarter, you’ll be back? And how are you feeling about the right way that we should be thinking about the run rates as we go into 2026?

David Girouard: I think we’re quite optimistic about the quarter. I mean, I think we have good growth rates. We are taking an appropriate level of conservatism. We have a very, very good pipeline of model improvements that very typically will drive conversion rates up. So in our view, actually, a lot of things are working really well. And it’s really important from our perspective to say that the model taking a bit of a conservative breather is a feature, not a bug. And if others aren’t doing the same, maybe we’ll figure out why over time. But it’s the strength of the model, not a weakness, that it’s making different decisions or taking a different take on the market. But in the grander scheme of things, we think the consumer’s health is good. We think our models are getting better. The new products are breaking out. So we think we’re in for a very strong 2026 and feel very good about the fourth quarter as well.

Operator: We’ll move next to Rob Wildhack with Autonomous Research.

Robert Wildhack: One more question on this subprime, superprime point. I mean, Sanjay, I think you mentioned that the UMI is lower for subprime, higher for some of the higher FICOs. If we zoom way out, we all see and hear a lot of headlines around this K-shaped economy, where super prime is doing quite well and subprime is struggling. So why do you think there’s that difference between what the UMI suggests and what we’re seeing and hearing more broadly?

Sanjay Datta: Rob, it’s a good question. I mean, we see directly, obviously, the data we have at our disposal. I think maybe it’s important to be precise with labels. So just to be very precise, if you think about the sub-660 population as measured through the traditional credit score lens, that is a population that, in our estimation, is in reasonably, I would say, actually quite good shape with respect to what their same default trends were pre-COVID. And so consequently, the UMI is relatively modest. If you go into the primary end of unsecured lending, so now let’s talk about the 720 to 750 segment. Those default rates are quite elevated compared to those same default rates pre-COVID, and their UMIs are consequently quite a bit higher.

And of course, we would talk about that segment as being a prime segment in the context of unsecured lending. Now, if you go to an even higher FI segment than that, let’s talk about the 800-plus segment. That is a population that I think is actually doing very well. They probably don’t do a lot of unsecured borrowing, though. So they’re not maybe in our label set or in our data set. And so I think you have this U-shaped thing in the economy where at the very low end or maybe the low end of the unsecured lending spectrum, let’s call it, 600 to mid-600s, things are very good. And at the very high end, maybe even beyond the unsecured borrowing population, things are quite good, and then there’s like a peak in the middle. And so I think we all use different labels to refer to different parts of that spectrum.

whether one part is prime or subprime or super prime or even not even in your data set. But I mean, very specifically, I think that’s what we see.

Robert Wildhack: And then just quickly, a couple of the OpEx lines caught our attention. Engineering and G&A were both lower sequentially, better than what we were all expecting again, better than what was implied by the guidance. Can you give some colors on the drivers there?

Sanjay Datta: Sure. Yes. I mean, some of it is just like our ongoing fixed expense discipline, which we’ve been focused on for some time. Some of that is, frankly, mechanical. As we reduce our outlook as a business for this year, we will reduce our expectation for things like bonus payouts and other comp accruals. And so there’s a bit of a mechanical adjustment to a lower outlook that sort of reduces the fixed cost base as well, which is working as designed.

Operator: And we’ll go next to John Hecht with Jefferies.

John Hecht: First question is, you talked about the use case for the broader unsecured loans. It looks like your HELOC loans are $55,000 to $60,000 on average. Can you give us the use case there?

David Girouard: Home equity loans are general-purpose loans. So people tap them for lots of reasons. We don’t have a breakout today of what the use case is for ours in particular. But of course, people know the most obvious thing is oftentimes used for home improvement, but quite often can also be used for other types of debt retirement or anything. So we think of HELOCs and personal loans as having, in some sense, being trade-offs from each other with respect to a general purpose set of funds, a little bit different rates, different process. But in some sense, they are substitutes for each other.

Robert Wildhack: And then I know this might sound a little bit like beating a dead horse. But just on this concept of the UMI and the tightening or conservatism and the dichotomy, just from what we’ve seen, some auto finance companies, some unsecured lenders, subprime, prime. And virtually everybody we’ve covered this quarter has experienced good volumes, but not only good volumes, but really positive credit trends. You guys talk about this concept of calibration over and over. I guess maybe what I’m seeking is, does your engine not disclose to you what it’s seeing that’s causing the difference between it and the market? Or are you able to see why it’s doing things differently? You mentioned pockets of weakness in certain populations. Again, what population or what demographic was that, or geography or something? I mean, is there something you can point to so we get an understanding of what this black box is doing to some degree?

Paul Gu: Yes. I think the principal way that you should think about this is that we’ve intentionally built our system so that it can respond faster than traditional credit metrics would. So in our experience, when other players talk about their credit performance, it’s a very backwards-looking metric in the sense that you’re typically looking at a somewhat mature cohort of loans and you’re measuring something like the actual charge-off rates. If you think about charge-offs in a lot of something like auto, you’re often talking about something that could go 180 days since it was first delinquent. And then there are mixed effects, and then there are sort of effects from new populations getting originated and mixed in there. And the confounding variables that come with all of those things generally create a pretty substantial obscuring effect to being able to tell what’s really going on in credit performance in real time.

And so we’ve built a system that is much better at precisely being able to, in real time, tell you what actually is going on when you control for all of those variables. So think of it as a system where holding constant all of the changes in your borrower population across, in our case, the thousands of variables that we use to actually underwrite and understand the risk of loans. When you control for all of those things, you control for the timing, the cohorts, the vintages, then what are you actually seeing? And how does that interact with any of these thousands of variables so that you can actually see the sort of underlying patterns? And that, I would say, is one possibility that you could see something that’s very segment-specific. I don’t think that’s the story we have in this particular period.

The other thing that it very simply lets you see is if there is an across-the-board move that would have been either detected 3 or 6 months later by traditional credit metrics or wouldn’t have been detected at all because it would have gotten obscured by the sort of changing mixes or new originations getting blended in. And in our case, we’re able to see that. Now again, as I said earlier, I think it’s possible to be overreactive to sort of that precise, fast-moving signal. And I think we optimized the balance a little bit better through some of our work this particular quarter. But ultimately, our goal is to be faster and more precise than anybody else in the market can be. And so we don’t find it necessarily surprising that there are periods of time where others are saying one thing, and we’re saying totally the opposite.

Operator: It appears there are no further questions at this time. I’d like to turn the conference back to Dave Gerardo for any additional or closing remarks.

David Girouard: All right. Thanks, everybody, for joining us today. We’re excited to finish the year with a flurry of activity and progress when setting ourselves up for an amazing 2026 for Upstart and our shareholders. Thanks for joining us today.

Operator: And this concludes today’s call. Thank you for your participation. You may now disconnect.