Ep11: Leaders in Lending w/ Sam Sidhu at Customers Bank

One year later: AI underwriting and consumer lending performance during the pandemic, presented by Upstart. I'll now turn it over to our two speakers: Jeff Keltner, Senior Vice President at Upstart, and Sam Sidhu, Vice Chairman and CEO of Customers Bank. Welcome, guys.

Absolutely, thanks Jeff. Looking forward to it.

Yeah, well let me just start by saying congratulations on your recent appointment. I know you're Vice Chair now, but as you know, CEO-elect of Customers Bank. That's really exciting.

For those in the audience who might be a little less familiar, can you give me just a little bit of background on Customers Bank? Who you guys are, what you do, kind of the bank in general?

Yeah, and you guys have been growing pretty rapidly. You say $18 billion. In my mind, I go back to when we first started talking — I think it was like eight or nine, and you guys were under the $10 billion threshold. So congrats. What's been driving some of that growth? I know this wasn't on my questions list, but I'm curious because it is a remarkable trajectory of growth that's not common in the banking industry.

Yeah, it's true. I did want to remind you real quick, Sam, that you promised to ask me hard questions here, so I'm not just asking. I think for the audience, we want to keep this as a kind of interactive dialogue. And as Jenny said, if you throw in some questions, we'll try and get to them throughout the presentation, but we do have some time reserved at the end for that. One of the things that struck me about your strategy, Sam, has been the execution of fintech partnerships. I'm curious how you think about that kind of build versus buy and the partnerships that you have created, because you've been at the forefront, I would say, of the banks I've worked with in terms of really leveraging that strategically. So I'd love to understand the thought process behind how you think about that as a part of the strategic mix at the bank.

I'd love your sense on how you — I know you're somewhat newer in a full-time role at the bank — but one of the things that's unique to me is your ability to execute. I mean that both in terms of finding and onboarding a partner and then managing the relationship, because I see banks that really fail at both. They have the appetite for partnerships, and then they can't get it through all the various committees and it kind of gets mired in internal process and doesn't go anywhere. And similarly, I do see banks that are able to sign a partnership and then — I came from the software industry, we always called it shelfware — it's great, it's signed, but it doesn't actually produce results for the business. And you guys have been effective at both bringing new partnerships live and making them meaningful parts of the bank and the business. I'd be curious what you think it is that you do differently or better or uniquely well in that kind of process internally at the bank to enable you to do that. I think it's something many of the attendees probably are trying to do and struggling with the effectiveness that you guys have managed. So your thoughts on how you do that and what makes you different, I'd love to hear.

I like the phrase muscle memory too, because I do think there's a degree to which the first is the hardest, and these are often in banks untread paths. It gets easier, and I think you guys have done a good job of breaking that new ground and then saying, hey, now we know, we've learned, we got a little better at our process and our team gets a little more comfortable. So I think that's a good way to think about it. I did want to ask you — of course, we want to talk about consumer lending, so I apologize to the audience for not taking too long to get there — but I wanted to ask, before you guys ended up being very large in the PPP lending program throughout COVID, which given your size at the time, you were punching way above your weight class in terms of what you did through PPP. Can you tell me a little bit about how you did that? What was it that enabled you to be successful in that, and why was that important to you? Because I think a lot of PPP banks were serving their existing customers, just making it available to the businesses they already serve. I think you guys went quite a bit broader to try and support the program holistically. I'd love to know why and then how you executed so quickly on that.

I mean, for a bank that was $9 billion in assets just a handful of years ago, $11 billion in funds dispersed for PPP is quite an achievement. So congrats, well done. That's an impressive execution. I will switch now to consumer loans — that's the thing we do and the thing that we work on together. But Customers Bank made a real decision a number of years ago to enter consumer loans in a pretty substantial way, not just in partnership with Upstart but broadly. I'd love to get your thoughts on that, because it's not something many banks do. I see most banks think consumer is mortgage, and anything that's not a mortgage they don't really touch at any scale. And you guys chose a different path. So I'm curious why that was strategically something you wanted to invest more heavily in.

And you guys are in a number of categories of consumer loans at this point. How many different things? I know we do unsecured together, and we can talk a little about that, but can you tell me about what your broader consumer portfolio looks like at this point?

I like that. So let's talk about COVID — that's the title of the webinar, right? Consumer lending performance through COVID. I think when I was talking to banks three years ago, there would be a concern that this is kind of the riskiest thing we could do. There's no house, no car, nothing to take back. This money is just — it's not even like a student loan that can survive bankruptcy. An unsecured consumer loan is the riskiest kind of loan in many ways that a bank can engage in. And there was nervousness about that. But then that nervousness was certainly enhanced by — we're at the tail end of a pretty good period of years from a credit point of view. Is this a moment in which to engage? Am I going to really at the end of the boom times invest in this risky category and leave myself at risk during what is inevitably to come in terms of a downturn? Tell me how you guys were thinking about that particular risk at the time, because I know we started working together a number of years ago in the consumer space. And then I'd love to understand the evolution of what you saw through the last couple of months or a year and a half of a very unusual economic circumstance, but certainly one that could probably be described as a macroeconomic disruption of some real proportion.

I know some of these answers, but were you actually increasing your target returns and pricing on the loans for that alpha? I will say for those who aren't as familiar with somebody who plays broadly in the industry, we saw a dramatic pullback from banks, from capital markets — just a high degree of nervousness. Which I think, to your point, was a massive opportunity if you believed in what you were doing. What is it Warren Buffett says? When everybody else is nervous, get greedy. And everybody was nervous, and it was a moment where you could go and take advantage of that. How did you think about — were you tightening credit boxes? Were you raising rates? How did you think about managing the fact that it was a nonzero risk that you would see deterioration in performance? And then we could talk about what we actually ended up seeing in the portfolio, but I'm curious how you thought about managing that increased level of risk even if you were trying to take advantage of the opportunity.

We're glad to be doing $100 million a month with you, because it's been great. I guess we should back up — I forgot to do this earlier — and talk about the nature of the relationship we have with Customers Bank. We've obviously been partnered together for a number of years. How do you think about how the partnership works? Maybe I'll give from Upstart's point of view how we think of how we work with banks broadly, but we'd love to start with your perspective on how this partnership works for you.

For those who are less familiar with Upstart, we are an AI lending platform. Unlike many fintechs, we are entirely partnered with banks and have been from our earliest days. We provide a digital origination experience, including probably the thing we're best known for: the application of AI to credit underwriting and deeper analytics to find those borrowers who are more creditworthy than their credit score might indicate. I think we can talk about that in a minute, Sam, because there's been a lot of question about: is that really going to work? How does it work? Is it going to survive a downturn? And we have some data on that now as we've come through the pandemic. The other area we apply machine learning and artificial intelligence to is the simplification of the borrower's experience — how do we get through KYC, ID verification, income verification without asking for a lot of documents? Those things really end up making a huge difference in the consumer experience, but importantly in the economics of the program and the cost that you bear to actually process those loans. I think one of the reasons so many banks have not been in unsecured lending is because the costs are too high to justify the revenue. When you can reduce those costs, you can make this really economical. We offer the ability not only to have the origination experience but also to drive demand to that experience for our partners, so we're actually helping you guys source that $100 million a month, and servicing on the back end for partners who want it. We've been doing all of that with Customers Bank. You guys were the first partner that was actually keeping loans on balance sheet with us way back in the day, as opposed to programs that were selling loans into the capital markets. So we've evolved a lot together and are now looking at the second product we've entered: auto refinance and eventually auto purchase financing, which I know we're talking about how we get into that game together. So we've been excited to do that. With that background, I did want to ask you really quickly: one of the things I've been talking about — AI and credit underwriting for 10 years now almost to banks — and one of the questions that was hardest to answer for any banker was, hey, you showed me some great data, and it's all from good times. And yeah, there's some better differentiation and the credit's better, but what happens when times are not so good? Is this going to deteriorate much more than the thing I've watched through a number of cycles, like credit score? So did you have those concerns, and then how would you categorize the performance of those models through the current macroeconomic stress environment brought on by COVID?

Yeah, so we have something like 1600 variables that are contributors to our model today. It's really amazing that you find there is very little power — even I think you've got to get over 100 variables to see something like half of the explanatory power of our model. And then every little variable is not super important. You can take out any one variable, including a credit score, and it wouldn't really change the outcome, because of the 1600, many are saying similar things. They're related. But it's the ones that are related that say slightly different things, and understanding how that really reflects a difference in creditworthiness gives you uplifts. I think to take advantage of this, you really need a combination of three things, and not many have put all three together. That's the 1600 — I think of the data like a big spreadsheet. 1600 is how many columns do I have? 1600 columns is great. But then I need to make sense of that. I need a lot of rows, because I need a lot of people who look very similar in one column — have a similar credit score — and look different in others to understand how to find that slightly lower credit score borrower who truly is creditworthy. We know they're there. The beginning insight we saw was in a subprime pool: you have 20% defaults, which sounds awful until you realize it means 80% of people paid you back. Could you just find the 80%? That's really where we started. So you find that if you have those columns, you need a lot of rows. We've now, across all of our partners, originated over $10 billion in unsecured consumer loans. That's hundreds of thousands of rows of data, so to speak, in the spreadsheet. And then you can't use your old logistic regression that goes to a scorecard that prints out on a five-page PDF to take advantage of those things. You can't say if this variable goes above X — you need really sophisticated algorithms, and those algorithms require both the large number of columns and the large number of rows. So we found that it's really important that if you have those rows and those columns and can apply those techniques, then there's tremendous ability to actually find those relatively creditworthy yet not yet high credit score borrowers. And even to find those high credit score borrowers who actually represent a somewhat higher degree of risk. So it really does take all the variables. You could take one out, but every one of them adds a little something to the model, and it's the sum of all those little somethings that ends up making a really tremendous difference in your understanding of the creditworthiness of a specific consumer.

Yeah, I think it's a great question. Obviously, our models are all trained on the performance of loans that didn't have that relationship. And I think particularly as you get into larger, different kinds of loans — let's say mortgages — that becomes really crucial. And I actually love the model you guys are developing, because I do think that a hybrid of the two will be valuable in the future, particularly less from a credit point of view and more from a service point of view. But I do think you see — and we see this in pools of loans — that for customers with long-standing relationships with a particular bank, they perform better on a loan from that bank than they do on a generic loan. There are elements of that that we see, and we'll price with our bank partners that have a long-standing relationship. But I think when you can see it, it means you can also see the absence of it, and so you can properly price and manage that risk. So I think there will be real value for banks that can cross-sell their customers on multiple products and thereby take advantage of the lower risk that comes with a long-term relationship — understand how to price and quantify that lower risk, so it's not just a 'hey, you're a customer, we'll give you 50 bips off,' but you can really understand how much less risk is this customer who's paid off three loans with me than a new consumer. But of course, there's also real value for the bank in being able to bring a low friction, low cost experience that can bring new customers into the bank, because the great thing about that loan to a non-existing customer is that's a new customer, and you can sell them another loan where they won't be a new customer, and you can help them with other needs. I think being able to do both is really important. We've certainly seen that particularly in the small unsecured loans. People want fast, they want easy. For the longest time, my CEO would go into his bank — and I won't name it in the CBA webinar to make them feel bad — but if he said 'I'd like an unsecured loan,' they would pull up an 800 number to schedule an appointment in the branch. And he may have liked shaking somebody's hand, but in that moment, he wanted to know how much he qualified for, and they weren't giving him that information. So I think providing the best consumer experience means meeting them where they are, which in this case may mean in their pajamas on the couch, and they want an answer then. I think you can ultimately combine that with relationships and high touch service as it's needed, but not necessarily for everything, which is the default that many banks go down to.

Yeah, it's a good question. I see banks that go different ways, and I'm always recommending that banks think of a digital product as a way to build a national footprint. I will say there are still many banks that want to go digital but they want to know that the customers are in a footprint where they can walk into the branch and talk to somebody if they want. I think often they feel like their ability to cross-sell other types of products is related to everything else I've got is sold in the branch, so if I want to bring a new customer in, I need to have them be able to come to the branch to cross-sell. But we really push them to think about a way to broaden. We can obviously restrict when we're doing referral flows where we're helping you find new customers — we can restrict that to the geographic footprint of our banks. Let's say I'm in these five states, can you just get me loans from here? That's fine. But I find many of them are increasingly interested in using these digital experiences as a way to broaden their experience, to see what it's like to have a customer in a state they haven't been in, and to kind of dip their toe in the water of moving towards a national digital experience as well as their in-person experience. And then the other thing I'd say is we do work with our banks that have more substantial existing branch footprints to say, how do we bring this experience into the branch? How do we enable the branch employee to talk about this kind of product to the consumer and then originate it in the branch, maybe on the consumer's phone, but in consultation with that banker? We've had a lot of success with some of our partners driving adoption within their existing customer base through their existing branch network. It actually can work quite well. So we always want you to serve your current customers, but think of this as a way to expand both your customer base and your geographical footprint, because there's a lot of opportunity and demand out there for that.

Yeah, I think that service and running the deferral program on the tail end was really important. For context, the shape is remarkably similar to what we saw for the industry and across all of our lending partners. We had that kind of increase, but a much more muted increase in overall delinquencies. This payment impairment for the audience is a combination of both people who are in a deferral program or people who are any number of days late on a loan, so it's kind of the total impairment whether they're just not paying or whether they've entered deferral. What's kind of remarkable is it's come back to normal. The industry is still slightly above normal as of the end of last year, but the Upstart portfolio is really back to where it was before, which is to say we saw a very small increase relative to industry and almost a complete return to normal. And then as we talked about the AI models, this is one of my favorite charts, but it's really hard to read. I think it explains why if you looked at the gross numbers, the Customers Bank impairments were much lower than the Upstart overall platform. That's because as we work together, Sam, you guys don't take — you have a limited risk appetite. We have other lenders who sell into the capital markets and don't retain, and they're willing to originate loans that are riskier as we perceive them than you are. This is a chart of how risky Upstart perceived the loan in terms of tiers and the credit score. What you can really see is that our banks that were limiting risk by the Upstart risk prediction saw not only better credit performance in good times, but this is impairment rates. So when you were in tiers one through four, the less risky category of loans, your impairments were much lower than if you were over here in what Upstart declared as the more risky loans, in a way that just wasn't true if you were using only credit score to do that. If you tried to limit your credit risk by just looking at credit score, you didn't have the same kind of relationship where you were really limiting risk of impairment here versus what happened in this. This is why, because you guys are looking in the lower risk tiers, your portfolio had less impairment, less losses in good times, and less impairment during the current crisis. So I think this data has been really quite stunning in some ways to see how much more predictive of risk impairment to the portfolio the Upstart model was. You can just look at the average column for credit score and the average row at the bottom for Upstart tier and go, 'Wow, if I had to pick one of those to limit my risk, it's not hard to figure out if I want to use the columns or the rows.' The question I have for you coming out of that — and I'll stop showing these — do you think this was a big concern? 'Okay, I've seen the models and I've seen the accuracy metrics in good times, but what's going to happen in bad times?' Do you think coming through this experience will be a point when more banks become open to leveraging these technologies, because you've kind of seen them through a cycle and you now have some evidence to look at, not just assumptions about how they'll perform? And I think so far the data is really compelling.

That's right. We certainly spent a lot of time improving our models for how to adjust in a way that is kind of unprecedented — the rapidity of change of the economic situation in this instance was so much greater. Normally you've got months to have this happen, and here it was a matter of days. But I think we learned a lot, and I think we'll be even better positioned the next time we go through a cycle. But certainly the data to me seems to indicate that if this was your concern as a bank — what's going to happen during a stress period — there's pretty strong compelling data now that it's really not a major concern. So I'd like to turn over to audience questions in a minute, but before I do, I just wanted to end by asking you: you're kind of taking over the helm of the bank, you guys have been really on the cutting edge of interesting stuff in fintech partnerships and consumer lending. What's on the horizon? What are the next two or three years for Customers Bank under your direction look like?

That's interesting. I like that you talked about leading with the assets. It's one of the things I'm starting to hear more from banks I talk to: this question of can we lead with lending? The traditional model has always been you sponsor a little league team, you bring in deposit accounts with relationships, and then you offer them loans as their customer. I think this method of actually bringing people into the bank through a revenue-generating product — frankly, it's kind of better to start off with a loan, it is where the revenue comes from. But I think it's really interesting because most banks I see are still hesitant to lean that way, but I think the world is shifting that way. If you can do it, why not? It's great to have a partner who's pushing at the front of that.

Perfect. Well, thanks for joining us, Sam. I've got a bunch of questions coming in from the audience. Jenny, I don't know if you want to ask, but I can pop on here. I've got one that came in that I'm actually kind of interested in in the context of this relationship question. You mentioned that you started out with some whole loan purchase programs, but you're now kind of in an origination model, at least in terms of your Upstart relationship. How did you think about the importance of originations and kind of originating in your name versus buying loans on the back end as more of an asset allocation thing? Why do you think that's important to you?

I like it. All right, let's see what we got here. In the quick, I got my questions pulled up. Out of the $10 billion originations, what is the average loan amount? It's a great question. Sam, do you know the average loan amount in your portfolio? I'm sure Ed's going to ping it to me in Slack if you don't. I think it's between $12,000 and $15,000. That's my guess. Ed probably knows best.

Okay, we've got other crickets. Do AI models replace human underwriting 100%? What do you think, Sam?

Yeah, and I'd say from the Upstart platform point of view, all the credit decisions are fully automated, so there is a complete replacement from the risk analytics point of view. We're not saying, 'Hey, come in and tell me you think this person's a better risk.' There are still instances where for fraud prevention, for verification, for other things, there will be human interaction. Again, we try our best to limit those because we find if we can go to a no human interaction model, we convert roughly twice as many interested parties into loans. That's good for everybody. But we also don't want to do that at the detriment of preventing fraud. So we've kept fraud below 30 bps across our history across the platform. Balancing those two is really important. That's not to say you can't call if you have a question or want to talk to somebody — we're absolutely there to make that happen. But at the same time, as I said, a lot of people don't. They just want to get through it. The more we can not ask you to upload a pay stub or a W-2 or a bank statement to complete the process, the better off it is for the consumer, the lower cost it is for us and for the partner, and generally the better performing the loans will be. So we do find that it works that way.

Let's see, I'm trying to look through my questions. Oh, here we go. During the pandemic, how did you monitor and assess the performance of your personal loan portfolio, and how did you get confident going back and expanding origination? So kind of what was the cadence or the process by which you were looking at how this thing was performing and getting comfortable, as you said, increasing the originations as you got a little bit farther in and said, 'Hey, things are going well, let's start growing again'? What did that look like?

Yeah, that's great. I think we enjoyed the ongoing dialogue. It was nice to have a good partner who was looking at the data and making rational decisions, where I think some were making fear-based decisions. I think that process worked well, and kudos to your team for being on top of it and well-respecting of the data and understanding of it. I've got two questions here related to fair lending. This is probably — if 'what's going to happen when the economy turns south' is the number one question I get asked about AI models, fair lending is probably number two. I'm going to assume you'd rather I answer that question. I think whenever you look at the application of AI/ML to any area, the question of fairness is central. It was central for us as we began developing models many years ago. So we actually went and had a conversation with the Consumer Financial Protection Bureau before we started lending with any of our bank partners on how should you think about fairness in the context of the models we're building and the data we're using. I will say we were naive enough that we walked into the CFPB's enforcement office as opposed to their office of innovation, because we just weren't sophisticated regulated entities yet. But what we came up with with them was a process to really think through how should you evaluate fairness, understanding that the world as it exists is not totally fair. If you look at the distribution of credit scores among different classes of people — African Americans versus Caucasian Americans — they're not the same. So if you're using even something as standard as credit score, you're going to have a higher approval rate for white Americans than for African Americans or Hispanic Americans, because they just don't have the same credit score distributions. So we said, okay, that's true, so we can't just compare to like 'is everybody being treated exactly the same.' So we came up with a waterfall test mechanism that we built in conjunction with the bureau and perform quarterly on the portfolio and provide results to the bureau. The result of that in the end was the issuance by the bureau of what they call a no-action letter, which is kind of a 'hey, we've looked at this, it makes sense, and we don't see any cause for concern' on specifically the topic of fairness. Frankly, what we found is that in a world where inaccuracy is high — where we're turning down everybody below 680, 85% to 90% of whom might be good borrowers — when you can have a more accurate model, everybody can win. That really means two things. One, when the CFPB compared our results to what a traditional model might have done with similar risk levels on the same portfolio, they said they found that for every demographic, we could increase approval rates and lower interest rates. So we're charging people less, we're approving more people than a traditional model across all classes. And then the question becomes, is that impact being disparately felt by a protected class or not? What you find is that this data can often offset traditional biases — things that are not evenly distributed today. So we've not found any cause for concern in our work with the bureau, and that's something we will continue to do. But I think it's an important question, it's not an easy question to answer, which is why we went right to the front of it. But we really believe very strongly that the use of these techniques and alternative data points can help improve the fairness of the overall credit system because of how much unfairness and inequality exists in the system today. We think we're seeing that in the results of our testing and our work with the bureau. So I don't think of that as a back foot question; I think of that as something we really lean into — it's frankly why we got into the business in the first place.

Yeah, that's a great question. I think what you have to do is move the level that you're thinking about the model up a level. That is to say, I'm not worried about the specific credit score box as much as how do I determine what that box should be? What's the process that trains the model? Where is the data coming from? How are we assessing fairness? How are we assessing accuracy? And how are we overseeing things that might go to production? If those things are staying the same and I'm overseeing that process, then the output of that process should be good. In fact, as the model is trained, one of the benefits of having defined a test for fair lending is that we can run it on every version of the model and go, 'Hey, we tested it under the old model, we've tested it under the new model, and we still have good results, so we're comfortable going live.' I don't know that frankly there are any other platforms doing that. But thinking about evaluating and overseeing the process level — the oversight, the criteria to launch, the testing, the training process — that's the core thing, I think, versus saying, 'Hey, I want to approve every change in FICO score.' And of course, you guys do approve and dictate every change in credit score box. But I think that's where the thinking evolved at the regulatory side, where the banks have to get comfortable too, is saying, 'I don't need to look at the code for every model change. I need to understand how it's being evaluated. If that evaluation process is changing, now that's a different issue — I want to see very clearly what's happening. But if you're using the same training and testing and evaluation method, and the model gets a little bit smarter, I'm comfortable that that's still effectively the same model, just a little bit smarter version of it.' That's the shift in thinking that has to happen to effectively take advantage of these things. So we got a question: does Upstart provide the digital application that a customer would fill out, or only provide the underwriting engine? My answer is up to you. In our partnership with Customers Bank, we provide the digital experience, we provide the flow of loans, and the servicing. We have other partners that just want to consume the underwriting engine and do that through a programmatic interface. I think that's not as common in unsecured loans because frankly not a lot of people have a great experience for unsecured loans today, but it is available if that's how you'd want to consume it.

Yeah, it's a great question. We measure NPS very carefully, typically on originated borrowers soon after origination. Our NPS hovers right around 80. I don't know what our latest number is across all of our partners. Frankly, our bank partners that are retaining — Customers Bank being an example — typically have slightly higher NPS than our programs that are selling to the capital markets, because the number one thing that influences people's happiness is their rate. Because you guys are able to offer slightly better rates than those lenders that are selling their loans to hedge funds and others, that typically results in higher customer satisfaction. But Eddie, I'm being told it's 81 right now. So 81 is a pretty high NPS. We monitor that weekly, daily, following the trends, because it's really important to us as a metric of whether the experience we're providing is meeting the customer's needs. That's why we got into this business — the consumer is our true north, and this is one of the ways we make sure we are staying true to that true north. Just as a point of reference, I'm sure many folks in the audience know their own NPS. Where would you say top-tier banks are hovering? I don't want to make anybody feel bad, Sam, but when JD Power looks at the NPS for the industry as a whole, I want to say it's in the 20s or 30s. I know there are some on the lower end that actually — NPS for those who aren't as familiar: Net Promoter Score, on a 1 to 10, 'Would you recommend this product or service to a friend or colleague?' 8 to 10 are considered promoters, 7 to 5 are neutral, 4 and below are detractors. You take your promoters minus your detractors, so a negative 100 is a possible score, not just a zero. We do see some in the financial industry who go into the negatives, where there are more people unhappy than happy. 8 to 10 means people have to be very happy. I think they're telling me Customers Bank's is right at 81. That's well above where I think anybody — the only people in the industry that have an 80 are probably USAA, who typically has a really high customer satisfaction score. Most of the rest are in the teens to 30s. I'd say top tier is considered 50 to 60, depending on how you measure it and what service. But 80 is quite high and looks a lot more like typical tech products in the broad consumer industry than financial services.

Well, so I guess people are asking if approval rates have gone up and if consumer demand has gone up in 2021. We saw certainly a dip in consumer demand going through the pandemic. In our minds, it's mostly recovered, I think buoyed a little bit by stimulus, but we've not seen a substantial slackening in demand from consumers. And frankly, the average approval rate has gone up. Approval rates depend a lot on the marketing mix — where people are coming from. But I will say on the same set of borrowers, every couple of months the model gets noticeably smarter. Whenever the model gets noticeably smarter, we can approve a few more of any given population of borrowers and lower the rates for a good population more, because there are still many good borrowers that we don't recognize as such. Every time you pull one or two losses that you can predict and not lend to, it gives you a handful of borrowers that you thought were too risky that you could pull in and actually better understand as creditworthy. That happens for us every month now. We may choose to mail a slightly riskier group than we did last month and offset that from a gross approval rate point of view, but every month the model is getting smarter. I think Sam would attest it's maintaining its performance from credit, and we're not seeing as it approves more people a deterioration of the credit performance. It's just us really finding those properly creditworthy people more accurately than we did a couple months ago.