Transcription:
Penny Crosman (00:03):
Like other fintechs, online lending software provider Upstart had a difficult time during the COVID-19 pandemic lockdown, but the company recently announced its second-quarter earnings and its stock price nearly doubled. Upstart's leaders have forecast positive earnings before interest, taxes, depreciation and amortization in the fourth quarter. The company's CEO Dave Girouard is here to share details of Upstart's expected turnaround. Welcome, Dave.
Dave Girouard (00:28):
Great to be here, Penny.
Penny Crosman (00:30):
Thanks for coming. So Upstart's chief financial officer said last month that the company has had a two-year nuclear winter. What are some of the things that have made it so hard for Upstart over the past couple of years?
Dave Girouard (00:46):
Yeah, sure. So I think it's sort of a confluence of factors that came about in sort of the post-COVID period. And all of these things are a little bit overlapping, but the first really being bank and lender liquidity challenges. I think that was not lost on anybody. That was something 2022 and the first half of 2023 in particular where banks were just seeing loss of deposits. There was a lot of fear and doubt about lending. So a lot of banks pulled back, a lot of credit unions as well. Secondly, of course, interest rates going up a lot. The zero interest rate environment wasn't just about the fed rates, though that was part of it. It was also about the flow of capital and how much it was either flowing or not. And then finally, I think, and almost as importantly, is the American consumer really became overextended as the stimulus that was provided en masse during COVID was withdrawn and it ended.
(01:52):
The spending habits Americans had developed didn't really change. And it was sort of like in effect there was this large massive layoff that happened across the landscape because the government suddenly stopped sending out checks to people. And that led to a consumer being, well, the unemployment hasn't gone up materially. The consumer really has been in the post-COVID era sort of over their skis on spending, et cetera. And that's really made it a riskier environment, put that all together. And that means the rates on our platform have gone up a lot and that translates to a lot lower volume in the business that we run. And that's really been the challenge we've been faced with for the last couple of years. And again, fortunately beginning to change and improve in recent months and quarters.
Penny Crosman (02:42):
Thank you for explaining that. And I mean, overall it seems like household debt, consumer household debt is at an all-time high, if I'm remembering correctly the last time I looked at it. Does that worry you?
Dave Girouard (02:58):
Yeah, for sure. It worries us. We're very cautious about the state of the consumer, their income, the stability of their income, their expenses, etc., including, of course, debt. The models that we have, like any responsible lender, can understand and make sense of the debt a consumer already has before extending more credit to them. So it's not as if the world doesn't have ways to deal with that in terms of lenders. And so that's not per se a problem for us. It is just a sort of macro event of overspending by consumers that is a really very unique occurrence due to this post-COVID time, and it's almost unprecedented where credit performance, if you look at credit card loss rates across the industry, they're as high as they've been since the great financial crisis, whatever, where that was 15 years ago. And why is that? Unemployment is still near record lows and it's really just because consumer spending outran consumer incomes. We're hopeful that is normalizing and we're certainly seeing improvements in credit in our platform in the recent months. So we're comforted by that. But of course macro is always difficult if not impossible to predict.
Penny Crosman (04:18):
So how did Upstart get through this challenging time?
Dave Girouard (04:23):
Well, some of the things that we've been doing the last couple of years are the same things any business would do in a challenging environment, which is reduce spending everywhere. We can take some of our bets off the table that we just feel like we're too long-term and didn't make sense given the way the world was shaping up in 2022 and since. So yeah, I mean we have fewer staff. We've had a few projects that we've decided to postpone or cancel. So those aren't uncommon things that any responsible company would do in these kind of times. But at the same time, one of the things we're really careful about is we wanted to continue to invest and improve the platform so that it would be much, much stronger and better on the other side of all this. And I think we've really done that.
(05:10):
That comes down to both the technology and the business model. And on the technology side, a lot of our energy in the past has been focused on what we call separation, which is understanding the relative risk of two different applicants. But the other side of it is calibration, which really means being calibrated against the greater macro economy. And that was something we just sort of took for granted. And we've done a lot of work so that our models are much more accurately calibrated. We continue to create more separation. So from a technical perspective, we can only wish a couple years ago that we had the tools that we have today and we feel more, much, much better prepared. You can't predict the macro, but you can certainly handle it with much more fine precision. And that's what we hope to do with our lending partners is help them manage their lending responsibly through any kind of macro climate.
(06:08):
On the business model side, one of the things that we really needed to look at is the marketplace structure we operate in. We have lenders, banks and credit unions as well as credit investors on one side of this market and consumers on the other. But it's all been effectively, it was an at-will platform, meaning a lender or a credit investor could come and go as they please in this marketplace. And what we sorted out is we really need to have long-term partners who have capital committed to the platform to reduce the volatility when the macro economy comes and goes. So that's been a big part of what we've done in the last couple of years quite successfully. In fact, over half of the capital in our platform today is in the form of longer-term commitments. And that's a big leap forward for us. And that's between the technology and the business model, lots of other things as well. We've really built a much stronger company in the last couple of years.
Penny Crosman (07:05):
I want to go back to what you were saying about adjusting your models to economic factors. I thought that was interesting and I just wonder, because I write about AI-based lending a lot and I do get pushback sometimes from people who feel like it's still kind of unproven or there might bias or there could be a variety of things that can come up. And obviously there's a counter side to all of that. But would you say that in general in your experience, it's easier to recalibrate an AI-based model than a more traditional linear regression or some other kind of model that doesn't really use artificial intelligence?
Dave Girouard (07:53):
Well for sure. I think if you think of artificial intelligence as just a lot more data and more sophisticated math to understand, interpret that data, then invariably it's going to be an enormous advantage, particularly in volatile times. And I think that the one truth in what you hear from skeptics is that it is hard to prove this overnight. Building an AI model and proving its efficacy is something that has to be done over a long period of time because there's certainly in good times plenty of lenders, their books will perform, they'll be happy.
(08:33):
And so what's wrong with that? But AI in lending is really focused on routing out the inefficiency and the inaccuracy, which is harmful both to the lender as well as the borrower. And so there's just kind of locked up economics that are in there and having really accurate data as every single payment is made every day, what does that mean? What does it imply for changes in the economy? How do you adjust your models based on that? The sort of accuracy and precision with which you can make decisions based on what's going on day by day, loan by loan is really brought to the front by AI. And I think that's going to be a powerful set of tools for any institution that is lending at scale.
Penny Crosman (09:20):
And when you say accuracy, do you mean you're able to take in more current information and a wider swath of information and therefore your credit risk models are better able to predict with accuracy?
Dave Girouard (09:44):
Well, think about the macroeconomy. You're thinking about something that could be stable, it could be shifting. And the first challenge is just to actually know where we stand today. How is the economy affecting the performance of credit? And just being able to essentially unbundle the risk associated with an individual from an applicant with the risk associated with the macroeconomy and how that might be affecting risk. That itself is a big challenge, but the nature of what we do is that we can do that really well and kind of segregate and identify macro-specific risk. Now that doesn't mean we can predict where the macro will go in the future, but it provides a really good starting point of knowing where it stands today, ensuring that originations that are happening today have reasonable conservatism built into them with the likely direction of the economy or the potential direction and able to sort of, as soon as there are more signals from what's going on out there, you can react and make decisions that are more accurate and make them much sooner. And that's ultimately what you need to, any lender's going to look at their book and say, wow, looks like we're underperforming in this segment or that segment. We're going to make some tweaks and some adjustments. This is just doing it much, much faster with much, much more precision. And ultimately we believe that will lead to long-term much, much better performance.
Penny Crosman (11:15):
So you have predicted or you've provided a strong outlook for the third quarter and you've predicted an even stronger fourth quarter with net profitability. Can you say a little bit about that and how you're getting to that point?
Dave Girouard (11:34):
Sure. I would say first of all, it's not really due to any kind of big change in the economy. Interest rates haven't really come down much at all. We're hopeful they will, but they haven't. The risk that we're seeing out there we believe has peaked and tends to be coming down a bit, but that's not really driving the forecast that we're providing or the optimism we have. Generally speaking, our models have taken some pretty nice leaps forward, and that's generally how our business has succeeded and grown in the past is when the model's more accurate in separating risk, it ends up converting more and improving more people. And that's how we generally grow. And we've seen in recent times a nice leak forward. So it's those improvements in accuracy, improvements in automation that generally lead to a brighter forecast for us.
Penny Crosman (12:28):
And are you taking on more bank partners as well? Is that part of it?
Dave Girouard (12:34):
For sure. We continue to add bank partners, that's very fundamental to our business. There's more than a hundred on our platform today. It is central to our business model that we aren't here to compete with banks or credit unions, but really to be a technology partner to them and not just in where we started in personal loans, but across every form of credit that really matters eventually. So yeah, we are happy that we continue to build out our roster bank partners and it is important for us to have diversity and it's good for the consumer as well where there's a lot of banks have different takes on what they want to offer and what they don't want to offer, etc. And this sort of marketplace structure really brings the best possible offers of credit to the consumer.
Penny Crosman (13:20):
I think in your second-quarter earnings call, you said that Upstart recently launched an improvement to its core credit pricing model. Can you share anything about that?
Dave Girouard (13:32):
Yeah, we've pretty recently launched what we called internally Model 18, and we've kind of shared that nomenclature outside, though we haven't always put numbers to models, but Model 18 was the first that used what we describe as "APR as a feature." What that means is APR of course is you can think of as normally the output of a model. It says what APR should I charge this customer based on all these factors we know. But the interesting thing of course is that the APR itself affects the performance of the loan, and it really does this in two different ways. One, which is really the smaller effect, is if you charge someone a higher APR, their monthly payment is going to go up and if their monthly payment goes up, their likelihood of default goes up at least some amount. But the larger effect really would be termed adverse selection, which is someone likely to accept a 15% loan on average is less credit worthy than someone who's willing to accept an 8% loan.
(14:37):
And that's just a market dynamic, right? The best borrowers have lots of choices, and if you have much higher APR for any particular borrower, you're likely to be adversely selected and that results in underperformance. So model 18 for us was the first time where we actually included APR in the model, and it's a bit of a science and math nerd thing to say, how can it be both an input of the model and an output of the model, and that's required some nice sort of technology leaps forward, which we saw. And the bottom line of it all is it's resulted in a much more accurate model, and that's been a big part of what's been boosting our performance recently.
Penny Crosman (15:17):
That's interesting. And just for my understanding, how and when do you gauge accuracy? Is it like six months after a loan is a made a year after or two years after?
Dave Girouard (15:32):
Well, accuracy, there's a lot of different ways to define accuracy, but generally we use very classic statistical methods of measuring accuracy of a new model, sometimes area under the curve. There's several different ways to do this, and that is really particularly focused on separation. How is it separating risk? And every time we launch a new model, we have a very good sense of improvement in accuracy, but there's a very different thing, which is how is this model performing in the real world over time? And that's something like anyone else we have to look at and see the monthly loss rates, are they above or below what you expected them to be? We've come up with some really nice tools to accelerate that as well. So if it's a 60-month loan, you don't necessarily have to have a model out there running for 60 months before you have really good feedback.
(16:28):
So we've come up with some very novel ways to do this, but that's again, a function of both separation and calibration and all these things make the data science behind lending fairly complicated, but we're so optimistic because the opportunities to do things better than they've been done in the past is so great. And the win is there, it's shared by everybody. It's better for the borrowers, it's better for the lenders, just better technology. And it's no small effort to do it. We've certainly learned that the problem is probably harder than we thought it was a few years back, but we're really getting to a place where we feel like if you're not using AI to do lending, you're going to be left behind at some point because it's just invariably more efficient, more accurate, better for all.
Penny Crosman (17:17):
I did want to ask you about in the bigger picture, what is your philosophy here? As I said, I've been covering AI lending for a long time, and I think of it as primarily or partially a way to get loans in the hands of people who don't have a high FCO score or don't have a FICO score because an AI-based lending algorithm can incorporate lots of other kinds of data and doesn't have to be so dependent on a FICO score. And I just wonder how do you look at it in the bigger picture? Why is there such a need for AI-based lending?
Dave Girouard (18:00):
Well, it really comes down to irrespective of your FICO score, the lending system as it exists, the legacy is enormously inefficient and inaccurate. So it approves tons of people for loans who aren't able to pay them back. That's not good for anybody, either for the lender or the borrower. There's a ton of people that apply for loans and are not approved for them, and that's not good for either person. That person could have paid them back. And then everybody, even someone with a super-high FICO score in our view, is probably paying more for credit than they should because they need to subsidize the whole system. And so if you can build towards models that ultimately deliver the right price to the right borrower at the right time, that's a far more efficient system. And everybody benefits from that. And that's kind of the world we're headed toward.
(18:54):
You'll never build something omniscient. I mean, the perfect model would know exactly who's going to default and who's not. It would only loan to those who aren't going to default, and they would probably be lent at almost a risk-free rate. Now that of course is impossible, but we can take giant steps toward that, and that's really what we're trying to do. My co-founder, Paul always had this saying that I found to be pretty striking. He basically said, any loan that defaults should never have been made, and any loan that is paid back, the borrower probably paid too much. And that sort of basic truth shows how far we want to take this, how much better we think the system can become. And again, credit is so fundamentally important to people and to businesses, to the American economy that making significant improvements, meaning reducing the price of borrowing is just a very, very fundamentally good thing for Americans and for our economy.
Penny Crosman (19:59):
So you do get regulators from time to time coming out with statements about the use of AI and lending. It can't be a black box and you've got to comply with all existing banking law, including fair lending law. It seems like every six months or so, they want the CF PP or one of the agencies comes out with a message like that. What's of your general answer to those concerns?
Dave Girouard (20:30):
I think the expectations of AI by lending regulators, whether that's CFPB or O-C-C-F-D-I-C, et cetera, they're all quite reasonable. The responsibility that a lender has a bank or a credit union does not change. If they use sophisticated tools like ai. They still need to be responsible for the performance of their models to be compliant with consumer protection laws, et cetera. And we've been working really since our earliest days to make sure that AI, as we develop it, can abide by all those laws. And in many ways, I think exceed what the legacy systems have done in terms of ensuring fairness and model governance, et cetera. So they're all sort of good things to be concerned about, but I think we've really addressed over the years in ways that we can clearly say that these systems are better for the lenders in terms of having better knowledge, better insight to what you're doing and a lot of control over what you're doing like a lender should.
(21:36):
But at the same time, much, much better for the consumer as well and everything including the adverse action notice if you are declined for a loan, there's a responsibility for the lender to tell you why. And actually it is quite possible. And in fact, we believe the notice is generated by an AI system can be even better than you would get as somebody just says, well, your credit score is not high enough. That's not particularly actionable or obvious. But I think that was a long answer to a simple question Penny. But I do think that AI can and is compliant and within the law and in fact can improve on a lot of the ways that the systems of the past have worked.
Penny Crosman (22:21):
Alright, well Dave Girouard, thank you for joining us today and all of you. Thank you for listening to the American Banker Podcast. I produced this episode with audio production by Adnan Khan. Special thanks this week to Dave Girouard at Upstart Rate us. Review us and subscribe to our content at www.americanbanker.com/subscribe. For American Banker, I'm Penny Crosman and thanks for listening.