Leaders in Lending
Leaders in Lending

Episode · 2 months ago

The Opportunity with AI Driven Lending: Paul Gu

ABOUT THIS EPISODE

There’s a lot of buzz about the “artificial intelligence revolution” and the impact it’s going to have on businesses as we know it. As AI and machine learning emerge in the consumer lending space, the technology has proven to provide dramatic wins for both lenders and consumers.  

This week we are running an older, listener favorite episode featuring Paul Gu, co-founder of Upstart.

You're listening to leaders and lending from upstart, a podcast dedicated to helping consumer lenders grow their programs and improve their product offerings. Each week here, decision makers in the finance industry offer insights into the future of the lending industry, Best Practices around digital transformation and more. Let's get into the show. Welcome to leaders in lending. I'm your host, Jeff Keltner. This week we're sharing with you a slightly older episode, my conversation with upstart CO founder and head of product in machine learning, Paul Goo and this is one of the first episodes we had. Paul Dives in great depth into the history of upstart, how we chose to start the business and unsecured lending, and quite a bit on the application of artificial intelligence and machine learning to the overall lending process. Where up start is doing that today, how we do it and where we see it going in the future. If you're interested in any of those things, this is a great conversation. So is out further ado these enjoyed this conversation with Paul Goole. Paul, thanks for joining me today. Of course, happy to be here. Yeah, so I want to start with kind of when you and I first met, or really when when you got into upstart, because you were an Undergrad at Yale studying, you know, interesting things and decided that it was worthy of your time to drop out of school and found a company in the lending industry, which is maybe not what you know every teenager in their late teens thinks at the time. So walk me through a little bit what it was about this space that was interesting to you and you felt like something you had to go and do. Yeah, so when I was in school I was studying computer science and Economics. Spent time sort of preparing myself for for sort of life in a sort of quant hedge fund world and spent some time there and essentially what I realized was there were a huge number of very smart people applying novel technologies to solve an incredibly narrow problem, the problem of slightly inefficient securities prices on a variety of sort of traded securities, and that was great, but it wasn't obvious that that was solving important problems for a lot of of real normal people, and I thought if you could simply take the same techniques that were used there and apply them to a problem that would affect real people. You could do something that was very important. So that was sort of the first half of the story. Then, as it sort of I got to thinking about where that application might come from, the sort of most obvious place was in the sort of closest adjacen C, instead of incorporate finance, moving over to personal finance, and the more I learned about it, the more I realized that huge numbers of people have very limited access to credit. Essentially, anyone who isn't born with money has a hard time getting access to money before they've built up a long, rich credit history. And of course that means for the vast majority of people there are at least some limits to their ability to get credit when they need it most. And if we simply applied the techniques of machine learning in ai that had really been demonstrated to be incredibly powerful in other domains and applied them here, we could do something very important for people. Yeah, I love the mission that you guys started with directly around helping the consumers. I am curious when I get the consumer lending space. You know there's lots...

...of kinds of loans and the personal loan, the unsecured loan what banks maybe you've called a signature long it's kind of a small space overall in the consumer lending ecosystem. Why was that the place when you said, Hey, we can use these techniques to to improve access? Why was that the kind of loans as we were starting upstart that she said, this is where we're going to focus our time, versus some of the larger categories like mortgages or or helocks or auto loans or something in that nature? Yeah, great question. Principally two reasons. But first is, if you want to demonstrate the efficacy of machine learning applied to something that's been done a very long time by many, many companies, many banks, you want to sort of apply in the place where it's going to matter the most and be the hardest, and that's an unsecured personal loans. Do you think about every other asset class? Car Loans, home loans, they're backed by something. Even credit cards are sort of backed by the further utility of being able to use your credit card in the future. When you give someone twenty dollars and say please pay me back, you're really backed by now thing, and what that means is you really have to be very good at making a decision about who you're going to lend to and who you're not going to lead to, or else you will very soon be out of business. And that meant that if we could do a better job of underwriting the risk here using Ai, we would be able to generate incredible economic gains for both sort of the lender and the consumer, and so that's what we were able to do. The second reason is that, from the consumer's perspective, the unsecured person alone, by its very nature, is the most flexible kind of loan. It can be used for any purpose. It's not something that you can only get sort of at the specific moments in your life when you're buying a car or House. It's something that you really can get at any time for any reason, and in that sense it sort of was the broadest products, sort of a natural starting place, and we always felt that if we did those first and second things, it would be much easier to go from personal loans to other types of loan products, which were now doing, than the other way around, where if you start with something really safe and really limited, it would be very hard to get out of that box and go to other spaces. Interesting. I don't. I don't think starting with the riskiest space is the way many lenders approach the space. But I take your point that it's kind of where the technologies make the biggest difference. What do you think the biggest shift is? I mean, I know many lenders have talked about ml or ai and we talk about ml or ai a lot in terms of the context of how we underwrite credit. But talk to a little bit about what you think the shift is form from a traditional approach to lending to one that's really ml or ai driven and how you approach the question of identifying risk in a different way. Yeah, so, you know, these words are often used as kind of like buzzwords and sometimes they cease to lose their meaning because of that. But when we say them, what we really mean as principally two things. The first is that we've taken an approach of using an immense quantity of data in the problem. So unlike in sort of traditional underwriting we're only a handful of data points are used, we use over a thousand variables. We use millions of rows of repayment data to actually train them all, and those data points that we're...

...using are a mix of both traditional and non traditional data points. We really are sort of seeking the broadest possible set of data that can be used to drive signal about who's going to pay back loan. And the second half that is really about the way that we learn from that data. So traditionally a person would go look at the data and try to draw insights from it, or they might use a classical technique like a regression technique. That essentially means fitting a straight line through the data. Now the problems that, while they're they're very easy to draw and understand, they don't fit the world all that well because most of what happens in the world is not straight lines. It depends on actual interactions between different variables, between the sort of different factors that that that drive whether person is going to pay back or not, when they're going to pay back, if they're going to prepay, when they're going to prepay, and so to actually make precise individualized predictions at the month level, at the borrower level, what we need to do is we need to actually use sort of state of the art machine learning algorithms. These are algorithms that actually are highly flexible, are able to sort of fied patterns in sort of tens or hundreds of relationships in our variables interacting together and the relationships between them, and when we sort of combine that kind of big data approach with the sort of modern algorithm approach, that's really what we call ai lending. Interesting, how when you think about the uplift you've got, and we can talk about the actual increases and accuracy and how you see that. But you talked about more data, both in terms of I always think of data as a spreadsheet, both in terms of more columns, more data points per applicant, and more rows, more historical data look at. I know there's also the use of alternative data points and models. How do you think about those different components and how important each of them is or isn't to the improvement and accuracy? Is there one of those that's really pulling the bulk of the weight and driving the outcome? Is there a split? I mean, how do you think about what's really driving the efficacy of the models? Up Starts Building. Yes, the two answers there. The first is, I would say these sort of different components work together in harmony, and that's not just sort of a generic talking point. It's actually really important because if you think about using a high powered machine learning algorithm, you tried to use it with only five variables, well, it's not going to do anything more for you than the straight line because you've hardly given it enough data to work with. It's not going to have enough to draw and to actually use the sort of powerful, complex nonlinear learning patterns that it's capable of. Similarly, if you were to take over a thousand variables and try to plug them into a straight line model, it wouldn't work. The model would also not be able to it would tell you that nine hundred of the thousand variables are useless, because it turns out that if you say I'm only going to use this variable in a straight line way, independent completely of the other nine variables, well, most of them actually aren't that useful. Most of the data is useful only in the context of certain other variables, and usually in a way that's highly nonlinear. So they really do go together. I would say, though, if you know, if we had to sort of pride apart, and we do for certain of our products, where we are sort of selling subsets of our model where someone can just access the algorithms on traditional the data without necessarily going through...

...all the alternative data. We sell that product because if you had to pick a single thing that was most important, you would go with the more powerful enhanced learning algorithms run on traditional credit data. That sort of first step gives you a sort of very significant boost to accuracy, but of course layering in that additional data takes you that much further. Let me just dig in that for a little bit because I know you said traditional credit data, but many people think of traditional credit data as ten, fifteen points of data off a credit file. What do you really mean when you say traditional credit data, because my understanding is it's a little more sophisticated than just a credit score or a credit score and a couple of extra data points. But you're digging a little, a little deeper than typical. Yeah, so when a lender polls credit, they're really getting two different kinds of things, really three different kinds of things. The first thing, the most basic package, is they're getting a record of every single transaction that a consumer has had that has to do with credit, every single payment, mispayment, delinquency, application for credit. All of that is a coorded on a person's credit history. The second thing they get is they get the credit bureaus interpretation of that data. The Credit Bureau will say, well, let's summarize that, this sort of vast history into a handful of data points. Most commonly they'll tell you, for example, this person has had three credit inquiries in the last six months, or this person has been delinquent twice in the last two years. And now have a handful of sort of common summary Statistics about a person's credit report. And then the third is actually past that same information into a model that is built by sort of third party scoring agency, most commonly fico, and the output scores that say this score was based on some of these summary Statistics, things like the number of inquiries in the last six months. And so you get these sort of three things, the sort of raw, complete history of a person's credit life, the summarized statistics from the Credit Bureau and then finally the sort of super summarized scores. And in sort of traditional credit scoring a lot of the approach is to lie on the latter two categories, either the summarized statistics or even the sort of completely compressed credit scores. And when we talk about credit data. At upstart we are talking about the full, rich raw history, which is often highly unstructured, can be permuted in almost infinitely many ways, and so there's actually a lot of really interesting things that can be done using machine learning algorithms to figure out how to generate the most signal from this history in a way that doesn't just sort of completely overfit the data or completely cause your sort of servers to melt down because you're sort of trying to run something that that is sort of too big of a computation. And that's where I think a lot of the interesting work we do with traditional credit data actually happens. And so you talk about servers melting down and thought of that. We have servers in the upsite that would melt down. I think they're all somewhere in the cloud, but not our servers. But you want to Melt Down Amazon servers and burn down a database. But how do you think about the trends and technology that have enabled this? I mean, is there is there a reason, and from a technology point of view, that these...

...kinds of capabilities are coming to the four now versus five or ten years ago? There's something that's kind of just kind of shifted. That's opened this up to be, you know, more accessible or possible today. Yeah, absolutely, I'd say the first big thing is is compute availability. You know, when I say and when I sort of jokingly said servers melting down, I of course did actually mean them melting down. What I really meant was just sort of not having enough compute to do the work you want to do in the sort of space of time that you want to do it. And even today, in a world where you can, in theory, arbitrarily scale up the number of e c two machines that you've got on the Amazon cloud, we still are actually constrained by the amount of compute we can we can access, and the way that manifests itself is when we run into compute problems, we start seeing run times that take longer and longer. So for our full model training process, we've often in our history got into places where the complexity of the of the learning Algorithms interacted with the sort of amount of compute and the efficiency of the sort of compute we had available to us that we faced run processes spanning seventy two hours for a single run of the model training. Now you imagine that you do that once, it doesn't sound so bad, but actually you don't want to just do it once. What you really want to do is you want to find the optimal model, the sort of most predictive model among the university possible models, and that means you want to search. And when you want to do search you need to sort of go through many iterations of different models. If each run is taking you forty eight hours and you've got a thousand different runs to do, you're looking at a multi year project to find a model, and so of course that's that's not going to work very well. Now that that's sort of the state of things today, and a lot of the investments we make are in improving the efficiency of those learning algorithms and figuring out how can we shortcut the search so that instead of taking forty eight hours, it only takes twenty four hours and instead of doing a thousand sort of searches, we only do I have to do twenty searches, and suddenly you've got a problem that's much more tractable. Now, if you rewind back ten years, you're looking at a compute that's a frat sort of almost an order of magnitude less than what was available today at the same cost. So it just would have been much, much harder. Of course, in parallel with the sort of improvement in the sort of underlying sort of infrastructure, there has been many advances in the actual algorithmic technology, advances made sort of both from the kind of theoretical math and statistics side and also in the sort of implemented computer science side, and of course we benefit from a lot of those advances and we're applying them to the sort of unique problems of lending. Yeah, it's fascinating how those two things have interplayed, the availability of computing and the sophistication of algorithms available because it's a compute available to use them, and they kind of kind of feed each other a little bit. It's interesting. Now the other thing I know I'm started kind of focuses on is not just the credit underwriting but actually the process simplification. Right the in the mission statement it's effortless credit, right enable effortless credit based on true risk. How do you think about the application of these kind of technologies to effortless? What does that mean, and how do you go about reducing the effort of the friction and the lending process. Yeah, you know, I...

...started to really think of these not as two distinct things but as part of one thing, which is, the one hand, when you think about the what we mean about the term underwriting, really it is just everything you are doing to ascertain the risk of this loan. It doesn't matter whether you're doing it at the front of your process or the back of your process. It's all for the sake of one thing, which is you want to make sure that the person you're giving a loan to the high chance of paying it back, and whether you're doing that by pulling their credit report or by verifying their income, it's sort of all for the same purpose. It's all underwriting. Now, the flip side of that is there are, and always have been, ways to gather more information about people that create a lot of cost and friction, and in some sense there's almost a perfect trade off between these things. You could decide to follow a person around every day for their whole life and you probably have a pretty high chance of figuring out the sort of person who's going to pay you back, even without any fancy ai right. But of course that's going to be incredibly costly and probably not a lot of I want to sign up for that service right. So, if you're going to kind of scare everybody away, and and essentially you know that's that's what we've seen. There's an incredible trade off curve where the more work you ask a user to do, the more they go away, and oftentimes those conversion falls are very substantial. You're talking about on the order. If you ask someone for a document, you can expect maybe fewer people to make it through your process. And if you're sort of, you know, a traditional lender that sometimes asked for several documents, you can just sort of do the math on that and look at what kind of conversion loss you're you're looking at. And so for us we think of it as there is this efficient frontier of trade offs between the information quality you gather and the amount of work that you put in front of your user, and the sort of real magic of our AI system is that we can really achieve much higher levels of this frontier, meaning we can choose between much higher combinations of information quality and ease of process for the consumer, where, for the same level of work we can get much more information from the consumer or for the same level of information. We can do it with much less work to the consumer. That's manifested itself today in our combination of extremely high underwriting accuracy, where we often see multiples of the accuracy that traditional sort of credit models can achieve. But at the same time we're doing that while over seventy of people that get loans with us are doing that, are getting those loans in a fully automated instant way, meaning there's no documentation that they had to upload, no sort of human touch on the upstart side, no waiting around, no leaving the session, just sort of all one smooth, instant digital experience. Yeah, I want to talk about that instant for a second because I've talked to a lot of lenders and particularly in the context of online loans and particularly in the context of, you know, walk up or net new traffic, which my understanding is the vast majority of upstart loans are going to first time borrowers, people without an established relationship, that the movement to online with walk up results in very high levels of frauds. So, as you talk about seventy no documentation, no kind of manual review, how do you think about preventing high rates of fraud from entering the systems. That's happening. That's kind of, I think, an astounding...

...stat for a walk up traffic business in an online context versus what most lenders have typically seen. Yeah, I mean the vast majority of, well over of business that we do is for for sort of first time customers in the whole upstart ecosystem and for any particular bank partner of ours. And of course there are there are a lot of fraud attempts in that in that system Um and so I think one of the challenges with moving from kind of traditional retail lending over to online is you just have this sort of huge exposure to fraud risk and we've seen many different kinds of fraud risk. There's, of course, sort of first party fraud, or someone pretends to be you know, our third party fraud, where someone pretends to be someone else. There's first party where someone just has no intention of repaying. There's sort of Internet specific behaviors. One of the most common ones is called loan stacking, which didn't really make sense in the kind of analog world where you know it would be a lot of work to go to a whole bunch of different banks and try to get loans. But Online, it turns out you can open five tabs and be on five different lenders websites around the same time and you can get loans with all of them before any of them find out from each other or from the credit bureaus that anything has happened. And so you really have to sort of prevent against all these different kinds of risks, and that's sort of a lot of the special sauce of the system that we've built up is we've built dedicated ml models targeting each of these kinds of risk. We've built them into the right points in the flow where what we really have is sort of you can think of it as kind of a sort of decision tree of at the right sort of moments we're asking the models to to sort of pass judgment on the probability of certain risks and depending on that risk level, we can then take different actions to mitigate that risk. If we successfully do it, great, if we don't, then there's sort of this question again of whether it would be worth it to sort of take another action, and so at each step you're sort of asking a series of kind of mathematical questions and if the answer is the risk is high enough, then you need to sort of take the next action in a series of actions and kind of building that system over the years we've really developed the expertise to keep broad to really minimal rates. We're talking tens of basis points of sort of fraud across the platform, which, you know, again in the context of almost completely kind of first time unknown online, fully automated sort of loans, is is, we think, a really powerful statistic. Interesting. How do you think about what happened during I mean the last twelve months have been a kind of unprecedented time, both from a macroeconomic environment point of view and from an unemployment potentially fraud driver's point of view? How have you thought about your model's ability to perform through this kind of unprecedented economic experience, kind of disruption to daily life, and what are the results you've seen in there? You know, anything you've learned that you think would be interesting to listeners about what you've seen in your experience over the last twelve, thirteen months through this kind of very strange environment? Yeah, extremely strange environment. You know, lending. Of course we always talk a lot about the economic cycle. You know, when times are good, everybody seems to be lending. Times are bad, no one wants to be lending and sort of when times are good, everyone just sort of worried about when times are bad. And one of the things we've always believed is...

...that there's sort of this itself, this kind of fascination with the economic cycle and the sort of focus of when the bad part of the cycle is coming, is itself a symptom of sort of an approach to lending that is not particularly robust or predictive. It means that lenders aren't really able to predict very well the sort of individual consumers level of risk. What they're really just doing is saying, well, overall, times are good now and you know, we assume right now we're gonna get a five percent default rate, but when times are bad that five is going to change a ten and it's it's you know, it's sort of just a kind of like almost like throwing paint against the wall approach and sort of when the seasons change, like everything goes from good to bad. Our belief has always been that really the variation between individual borrowers far outweighs the variation between macro economics. The macro economic periods Um and it's very clear that's the case when you actually zoom into the data and you see that even in the toughest economic climates, the vast majority of people in sort of low risk buckets still end up repaying their loans. The most interesting question is actually which individuals are actually in those low risk buckets and not sort of which period are you in? So now going back to this past year, of course, was sort of a very interesting time. Not only was it a turn in the economic cycle, but a very unusual one, and essentially, we what we saw was that we saw certainly a lot of changes to consumer behavior, but an aggregate we actually saw that the sort of loan performance of the sort of entire ecosystem across all our bank partners, all our sort of capital markets partners, the returns that they were expecting to earn, they pretty much across the board, consistently achieved or beat over the course of the last year, in spite of the pandemic. Now, having said that, I think a lot of that does have to do with some of the sort of unique circumstances that happened here, which is another piece of what we've sort of really built into our model, and that's about being able to respond to the sort of macroeconomic changes on on really a dime. I mean, when unemployment rates started going up in q two of twenty twenty, that was a time where it was really important to be able to respond very quickly and precisely to those those sort of changes in unemployment rates, and you wanted to be able to know like in which sectors, occupations, etcetera, there was extra unemployment risk that you needed to be modeling in and how to model it relative to traditional norms, and you wanted your model to handle that, because the alternative was sort of getting a group of people to sit around a table and kind of like use a backup onload map and then just sort of have to like take a blunt cut to your credit box, and that sort of really just means that you are either going to still take on too much risk or you're gonna essentially stop running your business during this challenging time. What we founders who were able to continue approving a significant number of people for loans and have those people pay back because they were the right people to lend to. On the sort of reverse side, as things started normalizing again. You want that system that can respond very dynamically to updating macro economic data. That's sort of a system that that we built and it sort of worked incredibly well over the last year. Is that system predicting macro economic conditions are? How do...

...you think about what you're doing to bring macro economics into your risk you know your risk profile. I get that the individuals greater, but you're still talking about ingesting some sort of macro economic status into your decisions. How does that work? Yeah, so we were not in the business of predicting macro that is sort of a whole business in of itself. We're very happy to industry hold different we're we're super happy to be market neutral with respect to macro economics. So what we want to do is want not to take market consensus, and so there are a variety of sources for market consensus statistics, both things that have already happened and things that are expected to happen in the near future, and we take those and we simply plug them into a system that's been fit against historical metrics, meaning we've taken things like historical charge off rates on consumer credit across different asset classes, historical levels of unemployment across different industries and occupations, and we've built models that connect these things to the way that they would impact our particular categories of loans. And so we've essentially built that into our model so that as soon as the latest unemployment numbers are released on a weekly basis, on a monthly basis, or the latest sort of monthly projections are released from various consensus groups of economists, those numbers essentially just get fed into the model update the macroeconomic assumptions and then you're properly calibrated to the sort of market consensus on macro economic numbers. That's a really interesting approach. I like that. So I want to ask one of you've used the phrase bank partners a couple of times. I just wanted to dig into the way you think about working with banks versus, you know, a lot of the fin techs that are out there, I know, pursuing charters, buying banks, becoming banks. You know what? What is the business model for upstart in terms of working with banks or you know the bank partners that you talked about? Yeah, we're we're technology company through and through. We've been doing this a long time and I think every time we've had an opportunity to sort of thing about which path in the road we want to go down, it's been very clear for us that we want to be more and more of a technology company and that's what we're good at. We're not good at being a bank, we're not good at any of the things that a bank is good at, but we are very good at building technology and we believe Fark strongly that the way that technology is moving, especially in the field of AI, it's going to be a sort of technology field where you sort of need aggregation and returns the scale. That's true across data scale, it's true across the R and D investment scale. It just doesn't really make sense for every single bank or every single business in a certain industry to really be developing their their own AI engines and tools and data sets. It's just not going to be scaled enough to be as good as one that is essentially shared across many in that in that industry, and so that's what we aim to be. We am to be the common technology provider and builder and we think actually banks are good at a lot of things. Those things sometimes don't include building technology, but they do include having great customer relationships spanning many products, having the sort of customer trust and loyalty and sort of brand understanding that comes with the sort of retail presence comes again with having those different products. Has that sort of relationship with with checking accounts, banking accounts, all of those things are things that that we don't do, we don't want to get into doing, but we...

...do want to support and enable our bank partners to be very, very good at lending and we think that piece of it needs a strong component of ai and that's what we bring to the table. Got It. I appreciate your thoughts. Has Been An interesting conversation, Paul, I think hopefully interesting for the listeners in terms of what I've starts doing that's unique in the space, and I have three questions that I usually ask guests at the end of the podcast, but I'm going to add a fourth for you that you don't know about. So here you go. You know, so you you came out of college, will, literally out of college because you hadn't graduated when you started at upstart, and you've been in the basically the consumer lending space for a decade now. More or less. What's been the number one thing that surprised you about the way the space has operated since you when you kind of came in from an outsider, I'm going to guess, Pretty Limited preconceived notions about the industry? was what's been the biggest surprise? Yeah, for sure. I think the biggest surprise has been I mean, I expected there to be a lot more people trying to build what what we're building. I mean it just seemed to me like such an incredible opportunity to create value for lenders and an incredible opportunity to create all for consumers. It almost is like, like, Bang your head on the table, obvious that this is probably the best application for ai that that you could have period. I mean best in terms of like it's sort of this combination of all the kind of pieces are there, like the sort of high volumes of data, the high economic opportunity, the sort of real time decision making, and at the same time, you know, it's sort of like this enormous, enormous market, one of the sort of oldest industry's most profitable industry is sort of the source of almost all profits of the consumer financial system. And yet you know here we are, ten years almost since I dropped out of college, and I guess it has been actually more than ten years and and almost we see so little, so little sort of happening in this direction outside of what we're building at upstart that I've just I've just been a surprise by, I think, the level of of sort of inertia and maybe some of the kind of institutional sort of barriers to innovating in this direction. But we hope that's starting to change now with some of the success we've had. I'm certainly seeing a lot more is from the banks I talked to. So fingers cross you're going to see more lenders coming this way and technology providers trying to tackle this problem in different ways. So now my standard three questions. What's The you know this is interally one for you because you've really, you know, most of your career one company. But what's the best piece of career advice you've ever gotten? The best piece of career advice probably, I get the thing I'll go with when in doubt, air on the side of being more technical. I think that's something that someone told me back when I was in high school and really it was sort of like. I think the idea was that in almost anything you're doing, whether whether the sort of nature of your work is usually done in a technical way or not, just because of sort of the power of technology to to increase efficiency, to sort of leverage things and to really train your mind to think in a rigorous way, it's almost always helpful to sort of aim to be on the more technical end of people in the sort of space of something that that you're doing. I think if nothing else, it forces you to think in the way that a mathematician or a computer scientist thinks,...

...which is in a sort of highly logical, rigorous way. And if you approach problems that are usually not approached as rigorously in a more rigorous way, it turns out you can often sort of find insights or optimizations about them that are missed when they're sort of thought about in generalities or or thought about without the sort of precision that's common to those fields. And certainly, having known you for ten years, my level of rigor and thinking and analysis has gone up substantially. It's kind of hard to have a discussion with you without without doing that. My second question is, what's the best advice you've gotten about the consumer lending space in general and this one. I think it'll be quite interesting given your history. Yeah, the best piece of advice we've gotten about the consumer lending Ingra have you ignored all the advice that you've got and your own path? No, no, certainly, and in our years we have spent a lot of time reinventing the wheel on many things. I mean, for example, we we didn't know about direct mail when we started in for many years we just we thought, we just thought there's no way that direct mail can be a thing that is done in the twenty one century as as sort of a successful mode of advertising, but it very obviously is a sort of successful way that advertising and consumer credit is done. I think the sort of best advice has been something along along the lines of don't be clever and innovative in stuff that you don't need to be clever and innovative about. When we started we really tried to sort of make everything about what we did different, and that meant all the sort of different terms of the loan, the way we structured servicing, fees and sort of every last part of it. We sort of tried to like rethink and most of those were like minor, minor optimizations and of course sort of the core thing we were doing was like a really big optimization. But because of all the sort of small, different things we did, we found ourselves constantly tangled up in kind of the greater like ecosystems sort of has has a set of norms of how they operate. You know, when you think about rating agencies and sort of regulators and sort of all the sort of different things that they're used to and and you kind of want to isolate your sort of innovation down to the things you want people to focus on and not get sort of spend all your time hung up on the sort of innovations that I don't really move the needle but do sort of create a lot of confusion and sort of headaches for for people, and I think we did a lot more of that early on a lot less now. That sounds like finding that efficient curve you were talking about between effort and accuracy and collecting data. Maybe you didn't quite have that right in the own in in the product organization in the early days. And then my last question is always what's your bold prediction for the future? Give me can be lending fintech life in general. What's what's your bold prediction? You seem like a guy that makes bold predictions. So give me a give me a nice bold prediction. I can bring you back and hold you too. M Hmm, a bold prediction. Usually see you lacking for bold Paul, for predictions. Frankly, you don't have to give me odds if you don't want, but I'll take them. Prediction without odds is not much of a prediction. That goes that rigorous thinking point you made earlier. I didn't think this would be the question to stump you. M I don't think I have a good answer that...

I want to I want to commit to no answer. No bold predictions. All right, you need a mathematical model to make our predictions and we don't have one on the fly. I'll take it. Well, Paul, I appreciate your joining us. I think is a very interesting discussion on application of ai across the lending life cycle, concepts around underwriting, and I particularly like to your point on macroeconomic where the difference in individual risk is much greater than the variance among individuals during economic stress periods. That was a really interesting point. So thanks for joining us. I appreciate your making the time okay, awesome. Thanks so much. Up Start Partners with banks and credit unions to help grow their consumer loan portfolios and deliver a modern, all digital lending experience. As the average consumer becomes more digitally savvy, it only makes sense that their bank does too. Up Starts AI lending platform uses sophisticated machine learning models to more accurately identify risk and approve more applicants than traditional credit models. With fraud rates near zero, up starts all digital experience reduces manual processing for banks and offers a simple and convenient experience for consumers. Whether you're looking to grow and enhance your existing personal and auto lending programs or you're just getting started, upstart can help. Upstart offers an end to end solution that can help you find more credit worthy borrowers within your risk profile, with all digital underwriting, onboarding, loan closing and servicing. It's all possible with upstart in your corner. Learn more about finding new borrowers, enhancing your credit decisioning process and growing your business by visiting upstart dot com slash four dash banks. That's upstart dot com slash foward dash banks. You've been listening to leaders and lending from upstart. Make sure you never miss an episode. Subscribe to leaders and lending in your favorite podcast player using apple podcasts. Leave us a quick rating by tapping the number of stars you think the show deserves. Thanks for listening. Until next time,.

In-Stream Audio Search

NEW

Search across all episodes within this podcast

Episodes (87)