Credit Scores vs Bank Data: Why Lenders Are Switching

Nobody gets into Fentech. Kids want to be firefighters and astronauts. Not very often do kids say, "Hey, I want to get into Fentech." The challenge I always found was the way lending works today has not changed very much in past 30 years. The traditional credit system is not rewarding you for that. Our models have an impact on if humans get access to credit, which could be life-changing for them. We can fix this. It's not that we got to fix this, but we can fix this. A gallon of gas here in Las Vegas ranges anywhere from 529 to 629 at Gallant. I never thought I'd see that. For that being said, how are you using that kind of data? You cannot afford to make mistakes with lending decisions. You could put a ton of money out there on day one. It is going to be risky, but you will get signal back very, very fast. That's where we are, our big believers in AI. If you ever are running an actual individual lending decision, Rue a neural network, you are in a world of trouble. Welcome to Fentech Confidential, bringing you the people, tech, and companies that change how you pay and get paid. You're building a Fentech product. You want to offer digital assets, but wallets? That's the hard part. Security, compliance, key orchestration, blockchain integration? That's why Fentech's payment platforms and custodians choose defense. They provide wallets as a service that's API first, multi-chain by design, and secured within PC. No single point of failure. So you can launch across over 50 blockchains, automate policy controls, and stay audit-ready without managing private keys. Stripes-rich, powers crypto payments through defense, moon pay scales wallets securely with defense, sphere, grew without compromising on control, or compliance. Even major FIs like Fidelity, AB&MRO, as well as custodians like Zodia and Tunston, trust defense to power their on-chain infrastructure. Developer-ready, compliance-approved production grade from day one. If you're building in payments, exchanges, OTC desks, market makers, or DFI, defense wallets work the way you need them to. Request your demo at ventechconfidential.com/dfns. Defense, secure wallets, built direct. Welcome to ventechconfidential leaders one-on-one series, where we sit down with ventech leaders to understand what drives our passion for ventech and their leadership lessons on how to get through things. I'm your host Ted Huff, the CEO and founder of Volear, and today we have Jamie Twist, the CEO, and Casey Kaplan, and you got one long title. Chief Product and Commercial Officer. That's right. Maybe we can make it a bit longer at some point, or at a third thing on, but yeah, it's good for now. And both of these guys are from Carrington Labs. Guys, I really appreciate you coming in. You flew a long ways, but you're here in Vegas with us, so I'm super happy to have you here. It's a pleasure to be here. Thank you for having us. It's great to be in the studio. So one of the things that, as having you come into the studio, I mean, Jamie, your background is really around the data science side of the house, which I find interesting having a data scientist in the CEO role, and it makes me curious to how you look at things. And then we'll dive into that in a little bit. And then Casey, I mean, you've spent what, 15 years or so building and participating in fintechs of all different shapes and sizes. Absolutely. One of the things that we were talking about is how Carrington Labs has built this really cool product set that leans really, really heavily into the A.I. House, but doesn't forget about the human piece of it. Nobody gets into fintech. Like kids want to be firefighters and astronauts and doctors and firemen. I mean, we could probably list out a whole bunch of typical things, but not very often do kids say, hey, I want to get into fintech. What was it about the space that got you interested? And then we'll dive into what really drove Carrington Labs to become? Sure. I'll start because I probably have a fun story. And Jamie is probably more of a technical analytical answer, knowing him. That will usually be the case. I first got into fintech because in grad school, we used to do bar hops alone. And one of the things that I found really annoying was at the time four square was really popular. You could check in and become mayors of venues. And I thought there was too much friction to take out your phone and check in and realize that wherever you're normally going, you tend to make a purchase there. So I said, how can I link a credit card transaction to automatically checking in on four square? And that caused me to learn how payments and cards work. I then went on to found a prepaid debit card program at the time. And then from there, I became a program manager and that evolved into a fintech ecosystem and then got into lending and credit. So I kind of started with the payment side and then evolved from there just because I was trying to solve a problem that I had and found it fascinating. How complicated payments was at the time, I guess still is, but once you start to connect the dots, it's really interesting and you see all these opportunities. So Jamie, I'm guessing that yours didn't have anything to do with the pub and maybe a pint of beer. I'm guessing it didn't start there. We probably should have gone to the other order. It's mine is going to be a lot less exciting. Publicly I've been gazing. But so I've spent my career working at the intersection of data and technology and financial services. And I've always been attracted to what are the hardest most wicked problems in that space. And if you look across large banks in particular, the hardest problems they have sit in credit risk. How do we take a vast amount of data, which is often poorly structured and use it to make a very difficult decision about whether we should lend this person or business. And so I built credit risk models for years for versus as a consultant and then I went in house at a couple of banks and then I finished up my banking career as the chief data officer of a global top 50 bank. The challenge I always found was when you're working in that big end of town and banking, you're spending 95% of your time doing pick and shovel work around data quality and lineage and extracting information and only a tiny amount of your time to actually using it. And so it's been fantastically liberating to move into a startup environment where you can really focus on the problem itself. And how do you actually use data to make these really interesting difficult decisions around lending. Nobody comes out saying, hey, there's a standard has been in place for decades. It's not like you're sitting on the deck on a Friday night and you're like, you know what? I think that what we've been doing for the last three plus decades, I think it's time to change it. I think it's time to fix it. So that doesn't just happen like on a whim. But what was it that what was the gap, I guess, really that you found that said we can fix this. Not that we got to fix this, but we can fix this. Carrington Labs grew out of a lending business that was and is very much a mission driven business. The goal has always been to make loans to people who aren't well treated by the traditional financial services sector and do so in a way that's safe and affordable. So often a small amount of money for a short period of time. By design, that lending business ran at very thin margins. That was always the goal have something that isn't going to extract a lot of money from people, charge them $10 for a loan, that sort of thing. In order to make that work, it became very quickly apparent that simply pulling a credit score for those people wasn't going to be very useful. Or thin file or no file customers didn't have a credit score. And also credit scores, they turn out not to be very predictive outside of that sort of kind of middle, maybe 60% of the market people above and below that credit score doesn't really tell you very much. So in that business, we were looking for a way to figure out someone has come in front of us. We never met them before. How do we understand them and work out whether we should lend them and how much we should lend them. The way lending works today off the credit file and the credit score has not changed very much in the past 30 years. And there have been these tremendous advances in technology and data computing and storage speeds, the availability of kind of the digital exhaust from your financial life, machine learning and artificial intelligence. And those have almost entirely bypassed traditional lending very much still stuck in the late 1980s. And we realized by starting to stitch together other sources of data, particularly banking transaction data, you could paint a much richer picture of an individual and make a far more accurate assessment of whether you should be lending the money or not. and we realized that IP that we built for that was itself actually quite an important product, both in terms of being a financially viable driving the industry forward, but also being, I think, a much fair and more inclusive way of bringing people into the financial system. So, when you talk about the inclusion, he said it, it carries on to the answer. What does that actually mean when you start looking at the product itself? I think keeping in mind how we got there is really important. Also, so our Carrington Labs, for those of you who don't know, we specialize in cash flow underwriting and credit risk analytics. I think the term cash flow underwriting is still this elusive thing to people. People know the term if you go two, three levels deep, that meaning changes to people. It's more than just inflows and how it flows of your money. And a lot of what we're doing is creating really advanced behavioral features that go into machine learning models to understand credit worthiness. And when we productize or offering, that's something we keep in mind. We generally feel that every lender and credit provider has unique set of customers who's trying to do their own unique business objectives. And our products tailored to that. So instead of just having one generic score similar to what credit bureaus tend to offer, we have personalized scores. And we've spent years now building the capabilities to have an automated model built pipeline based on the unique data points that a lender or a credit provider might have, so that they're able to get really predictive analytics and probability of default percentages for each of the products that they might have in their portfolio. What that then comes to you as you think about it, inclusion is when you have data sets on customers that are maybe other than bureaus like the open banking data, transaction data, you can actually shape that your business objectives as well as helping customers in a compliant and regulated way. And that's how we productize it. So when you look at our products, we did is understanding credit risk. It's understanding how should you size loan limits and then it's understanding if people are going to be able to service that once the loan is out there. And combining all this together, you find that you're often able to approve people who would normally be declined through the traditional system while also being able to be more profitable at the same time. So it's great. You know, these win-win scenarios, but that is truly what we see. Well, I thought it was interesting. Lexus Nexus put out a report where they look at it and like over 50% of applicants that come in, board lending, aren't able to have a reliable score. And I think we kind of talked about this earlier as we were prepping for today is like, why is it not reliable? Is it because it was the point in time? Is it because we don't get to see all of the things that they go on in their financial life? Like all of these different pieces, the people I think that get hit the most, hit the hardest, I guess you would say, are people who are self-employed, are a gig worker, may have a side hustle. We'll call it that don't get treated as fairly. Help me understand how you guys are approaching it. Carrenson, how are you approaching these unique cases? I'll start with the shortcomings of the credit score approach to lending. Essentially, and everybody should look at their credit file and see what's in there. It's a useful exercise. But essentially, that talks about the way that you've engaged with credit in the past. In the country you're living in now, they don't cross borders very well. And so if you have had a credit card for a while and you may be taken out alone and paid it back, you will have a credit file and a credit score and it'll probably be pretty good. And you will be eligible for a loan using that traditional approach. But as you rightly note, that leaves out a lot of people. It leaves out people who are new to the workforce, new to the formal financial system, new to the country. People have come up in a different socioeconomic pattern that means they haven't had that kind of engagement with traditional banking and lending. And so a huge portion of America, a huge portion of people around the world are essentially by design cut out of the formal financial system that way. In a way that's unfair to them, but also isn't good for the lenders because they miss out on business and isn't good for the country because people who could responsibly use more capital to drive their lives forward aren't getting access to it. We will look at different sources of alternative data and your bank transaction data will be usually a very big part of that. And almost everybody has or can have a bank account and have transactions that are clearly listed in that bank account. And using that, we can draw a very detailed picture of how you live your financial life. Certainly there are some basic things like how much money do you make, how much money do you spend, do you have savings, understand your PNL and your balance sheet, so to speak. And that will give us a good sense of whether you can service a debt right now. We'll also look at things that are much more behavioral in nature. So in the past, if you've had a cash crunch, let's say there was a month when it was difficult to pay your bills, how far in advance of that cash crunch did you see that and adjust your spending? Did you pull down your nondiscretionary spend? Some people see those cash crunches 21 days in advance, some people at seven, some people at three days in advance. Well, we can see that in the transaction data and that is enormously predictive of your ability to repay a loan. We also see it around a lot of behavioral factors. So in the past, when you've had an obligation, have you done everything you can and deployed assets and moved money around to try to meet that obligation to pay that bill, to repay that loan, let say a direct debit hit an empty account. That's also enormously predictive of your attitudes towards money and your willingness to repay a loan. And we find with that approach, first of all, you can bring basically everybody into the lending system, as long as they have that transaction account and some history in it. But it's also a much more fair way of engaging with somebody because you're actually getting to what drives their creditworthiness, you're getting to their financial situation, you're getting to their character, and it's a much more reasonable assessment that is much less to do with the kind of family they grew up in or when they moved to a given country. If you look at the last 30 years, you have millennials and younger who tend to be taking out less credit, which means their credit files might not be as robust. And part of that is because they've seen their parents going to debt, right? And that's been traumatizing for many of them. So as you have less credit cards or less forms of credit because you might have a job, you might be responsible and you don't need credit as much, you're just paying it on a debit card through your salary, the traditional credit system is not rewarding you for that. So some of these newer approaches to cash flow underwriting and looking at that behavioral information that Jamie has touched on, it becomes really important and it turns out it's really predictive. One of the things that as we look at the way that Carrington Labs became its own piece, I mean, it wasn't like you had an idea and needed to test it and figure it out and try and find data from God who knows where to get it into action, building it inside of before pay, gave you millions and millions of data points to be driven from how do you feel that has benefited you and being able to address the market needs? So we built the Carrington Labs capability inside a lending business and that's a fantastic place to start when you're building an analytic capability for credit risk. There are a number of advantages that we got by starting that way. The first one is tremendous access to data. So when we built our first external models, we had literally billions of lines of transaction data on which to train those models and understand how to build features and work through all of the pieces there. The second advantage we had was the ability to get very rapid feedback. So as we build models, we deploy new features, we try new techniques, we can put those live in production in our sister lending business and very quickly get a sense of do they add value and how much value do they add. And the third one, which is tremendously powerful and this is very compelling for clients is alignment. So we are 100% aligned with our clients and ensuring that the Carrington Labs approach in those models are fit for purpose and do deliver a better lending commercial outcome. Our models have an impact on if humans get access to credit which can be life-changing for them. So this is a serious stuff that we really need to think about. We're not just a few people in a garage, we have a cool idea and we're just throwing some tech out there to see if people use it and hope for the best. Which unfortunately, for certain types of product, great way to rapidly prototype and see what's happening. But when you're getting to regulated products that have a true impact on humans, you need to get it right. You want to support the borrower, but you also want the institution who's giving that loan out to make sure that you're doing it right. that they're going to get the money back. So we think about that a lot. And we, we tests on our own lending business. We really think through the implications of how we're building these models. Part of the genesis stories is, you know, the lending business started. We thought there were better solutions than prebiro credits scores. We got into crash flow underwriting and built these sophisticated models before we knew it was called cash flow underwriting. And we've iterated from there. And I think part of the genesis story is, you know, I think Jamie was in Europe or the UK having meeting with banks. I was in the US doing similar things, just talking to some of the connections we had. And we were explaining what we were doing before we officially launched Carrington Labs. And people were saying, wow, that sounds great. Like, can you roll that out? So we realized there was market opportunity and demand from other lenders to use this type of solution. We realized we kind of cracked it on the lending business that we were involved with. There was this synergetic opportunity where I came together. Something I hear a lot working with Fintechs and financial institutions is that they can't afford AI tools to be a black box. They can't do it for compliance. They can't do it for payments. They can't do it for lending. They can't do it for customer acquisition. And frankly, I mean, we all know the regulators have stated very clearly, no AI whitewash or no AI washing. If you're using it, it's got to be explainable. All these different pieces come in. And I think it's been really interesting that you guys launched an MCP server-focused approach to this that was purpose built specifically to identify the credit risk models. We talked a little bit about how the perspective of the agentic workflows are changing over time. What would you tell leaders that are in any sector of regulated, financial, and financial technology? What would you say to them how to look at the AI architecture? And what is it changing for them? So I'd say a couple of things to financial sector leaders. The first one is as you look to deploy AI, be very thoughtful about the use case. And whether that use case is fault tolerant or fault intolerant. So there are some areas such as generating marketing copy where you don't want to make a mistake, but actually depending on what the nature of what you're putting out there, it might be okay if there's a quality issue or if it doesn't say quite what you thought it might say. Then there are many areas, probably more areas that are fault intolerant and clearly lending is one of those. You cannot afford to make mistakes with lending decisions and you cannot afford to have the laws appropriately very clear and very strict on that. And so if you're working in a fault intolerant area of your business, when you deploy AI, you always want to make sure you're embedding it in a workflow that puts a tight layer of control around that. And so the way that we think about that is we use a tremendous amount of AI in the upstream model creation process. But then as those models come together, we have what we call a control point. And that's where a human reviews the model prototype that's come out of AI and checks every feature that's in there and signs off that this is an appropriate way to be making credit decisions. And from that control point forwards, everything is deterministic, just repeatable statistical explainable, same input, same outputs. And we can tell you exactly why someone did or did not get approved. Casey, like one of the things that I find slightly hilarious is that a lot of companies are putting dot AI on the end of their their URL. They're saying, Hey, we have this new AI tool without really explaining it. And really in a lot of cases, it feels it correct me if I'm wrong, but it feels like it's just an automation layer that's sitting on top of it. Like the next level of robotic process automation, maybe using a little bit of extra data, but not a whole lot. You guys have approached this from more of a and correct me if I'm wrong, more of a like a deterministic type layer. Yeah. Why does the distinction matter when we start talking about deterministic versus inference? Yeah. It matters a lot as it turns out. So if you take lending, for example, a huge amount of Africa is into origination, right? Should you approve someone for alone? And there's some good companies out there that are automated workflows and they're partially using AI to do some of those things. I think there's the broader question of do you need AI to really automate some of those things or could you do these traditional workflows to do that? But that aside, with lending, you know, basis points matter, right? Like there's the regulatory piece where it has to be explainable, right? You need to make sure all the features you can understand the features in the model, how it's impacting that and you can explain the to regulators. That's also important for you to improve your credit criteria, right? If people are defaulting, you want to be able to update the criteria that you're approving people on. When it's more of a probabilistic solution where you're utilizing AI, I've more to make those decisions, you can't explain it. And we have helped lenders where they're quite literally calling the open AI API with bar or data and effectively asking chat GPT, should we lend to this person and their feedback was, oh, we get different answers at different times, right? We spend a lot of time thinking about where does it make sense to actually use AI because you get benefit from it versus using machine learning or other approaches which are deterministic. When we look at some of our AI forward solutions, so we do have an MCP which can communicate the results of our models. I think that's a good example, right? So if someone is using an agentic workflow and it is early days for that still, to get the contextualized information back is quite helpful. But also if you have a credit team or a support team who's trying to understand why was the MCB can be quite helpful there because they can effectively ask questions to the MCP server around the specific bar or without having to read through hundreds of lines of schema and seeing all the features. I think that is evolving very quickly and the use cases are evolving really quickly, but I think right now in the regulated environment, there is that distinction where when things can be probabilistic and when it is sometimes okay to make mistakes or when you need to get it right. And I'd also say it depends on country as well. So we support Lenders globally. In Australia for example, you have to explain why you approve someone for alone and in the United States you have to explain why you declined someone for alone with adverse action reasons, right? So in our solutions, we have a in our credit response solution, we have a solution where in our response we will give you score rationale where we have promoters and detractors, right? So kind of you're covered on on both ends and both jurisdictions. And those map to the features that are in the model that the Lenders approve themselves before it goes into production. So we think that transparency with Lenders and what's going to the risk model is really important, but we'll also keep the explainability. You know, it's interesting as you talk about that. It makes me curious because there are many layers of what we call AI, I mean, image and machine learning. Some people will call that AI. Yeah. Some people will will look at the inference of a particular information as AI based off of data. Then you've got the large language models and you've got all these different pieces into it. Help me help me and everybody understand like how how it's different between just you'd mentioned throwing it into an open AI API. Maybe with a little bit of context, maybe with a series of prompts, maybe, maybe we could even now with Claude Finance being launched here in the last six months or so. How are we making sure that we can maintain the explainability, maintain the fairness and maintain the, I guess I wouldn't say maintain, but ensure that the bias doesn't creep in because every time we look at something, when I'm using it, it wants to tell me the answer I want to hear, not the answer that I need to hear. How are you at Carrington Labs, like balancing between, I want to hear this versus you need to hear. We could probably tag team this one a little bit here from different perspectives. I think that's where we have the concept of a control point. So, so downstream of the control point, solution, we believe everything has to be explainable. And that's where we rely on machine learning and deterministic solutions. Upstream of the control point, that's where we, you know, big believers in AI. I'd say specifically Gen AI to get the creativity. When you're using Gen AI to try to get more deterministic like features, still not guaranteed though. There are a number of strategies that are emerging and the solutions are getting better and better and better every day. There's the concept of guard rails that you can put in. You can play with like the temperature of what the LLMs are doing to try to get more consistent outputs, but it's still not guaranteed. So there's risk from that regulatory perspective. When we look downstream past the control point where it has to be explained, well, that's where we think, you know, machine learning ultimately produces what regulators want right now and that that's been tested and it's commonplace. So that combination right now is really effective. So if you are using AI internally to build your tooling, that is one thing right. And there's a number of strategies and there's a concept of skills are emerging and utilizing various markdown files to provide extra context to the agents as they are building is commonplace at at time of this recording. Maybe by the time it rolls out to a world will change. But I think when you are in production in a regulated environment and you are not able to make mistakes. So there's massive regulatory risk for I guess making a decision that would adversely impact someone you can explain. That's what you need to be mindful of when either building yourself or working with their party providers. As Casey was talking about that, all that comes in my mind is like you mentioned things are changing so fast. And you know, it feels like every day that you know open AI's models ahead on this and then the anthropic models ahead on this and then the Google Gemini's head on this and you know, like all these models are like jockeying for being the best on that. Two part question one. How in the heck do you keep up with it and two, how does that benefit us to remove the bias that we talked about to take a step back. There are two broad categories of what you might call AI and they're very, very different. And I think a lot of people blur them together in a way that makes it hard to resolve these questions. When most people think of AI, they think of generative AI and related techniques that are based on what we call neural networks which are essentially mathematical recreations of the brain enormously powerful, enormously flexible, absolutely black boxes. No one has a good way of understanding what's happening inside. Random this is a core element of how they work. They don't work without some random elements to them. And so you will get different things every time. And so most people when they think of AI, I think of those neural network based technologies. Then separately, we have a whole bunch of things like machine learning and other statistical techniques which are just more statistical and mathematical and those are generally explainable and they are reproducible. You put the same things in and generally get the same answers out. When we think about lending and credit, there's tremendous power in being able to use generative AI and other neural net based technologies and approaches. And using that to give yourself insight into data in order to set up ways of building other kinds of models and all sorts of explanatory power. But if you ever are running an actual individual lending decision through a neural network, you are in a world of trouble because you will not know what answer it comes to and you may get a different answer every time. So those are tremendously powerful tools. But when it comes to actually making a decision about an individual loan, you should always be in that more deterministic and that explainable space machine learning and related techniques. And so to answer your question, when we think about the relentless march and progress around these models, we always think about how can we make our explainable, predictable deterministic machine better using whether it's Claude or Gemini or the latest Open AM models and so on. But not hand them the keys and let them make decisions on their own. And I think despite the rapid rate of progress, I think the world is very far away from a large language model that actually has that level of transparency and predictability. In watching, guys, we started talking late last year about what Carrington Labs was doing and the approach you were taking. I mean, you guys have been crazy busy with partnerships and announcements and all these different things. And I don't want to get a little bit. Let's get into some of the specific pieces, right? Because these are real products. You do have real partners and you are delivering real results. Walk me through like a real-life scenario where let's just say a Fintech company is declining 40 plus percent of their applicants. And they're starting to see their margins just get just demolished by all these declines. They've got their their cost of acquisition has gone through the roof because they're having all these declines. But what does what does it look like for someone that comes to Carrington Labs and says, this is my problem. How do how do I fix this? We'd normally start by understanding what they're trying to achieve, right? Because they might have a problem, but it ultimately comes down to how does that compare against their objectives. The correct economic answer is how do we maximize our margin? Is usually the right thing they should try to be doing, but not everyone has that mindset. So we sometimes see startups who are let's go with that though. Let's let's go down that path. We're trying to maximize our. All right. That was the fun one. We're trying to maximize margins. So commission margin and absolute dollars. Yeah. Yeah. Right. How do we get the most dollars? Right? At the margin level. And there's a few pieces to that when it gets a lending and credit. So one is or who are we approving? The second is of the people we are approving. How much money are we giving them and how is that priced in order to the duration? And then once it is originated, how is that loan performing or the type of credit performing? And those are the three areas that we tend to look at and you can use that same concept for pre-qualifications as well. Well, you'd mentioned earlier today that motivation. Yeah. How do you determine like whether or not your lending process has too much friction, whether or not the motivation is there to the customer like how is Carrington labs helping identify maybe some excessive, I'll call it excessive friction in the process. So we love looking at lending funnels. This is like it's probably a passion area of mine and breaking down the data, but it's on a fun. There are different types of last-discities. So pricing last-discities is clearly one of the things you need to consider if you're pricing too high. People might create an account, but they're not going to necessarily take the loan even if they're approved, which means the marketing dollars you're spending are not going to translate into revenue. Then there are other things in the actual funnel itself. So if you are doing in some cases there are borrowers who are really sensitive to doing a pull from a bureau because they're worried that that's going to impact their credit score. There could be other bars who are maybe a bit hesitant to use scraping and open banking to pull on the transaction data. It really just depends. So depending on how motivated a bar is to take out the loan has a big impact on throughput. If people want the loan, they're much more likely to go through the process. Now as a lender, what I would encourage everyone to be doing is have your analytics in place. Look at every single step in your process, see why there is drop-off, and do research, right? There's quantitative and qualitative things you can be doing to understand that at each step of the way. Introduce a financial health tool, right? That might be one way to get people to aggregate more of their bank account data. Let customers know that maybe if they've been approved for a smaller amount using some of the traditional approval processes and they're interested in more, think more of their account and use more of a cash flow underwriting approach to better understand their serviceability and maximize some of the limits they have they can have. But breaking a down at each step of the way I think is really key, but it does come back to the lender's objectives, right? It's not just about saying yes to more people. If those people are high risk and not going to translate to margin and dollars in the P&O. Support provided by Skyflow. What if you could build fast but not break privacy? What if you could ensure data privacy, governance, and compliance with just a few API calls? What if you could worry less about PCI requirements while actually improving privacy and security? How much more time would your team have to truly innovate? How much faster could you build and ship new features? How much more powerful could your app be? Skyflow is a zero-trust data privacy vault delivered as an API. Skyflow's radically simple design lets you collect, secure, and tokenize personal information like car data and payment details. And with built-in features like encrypted data analysis and sharing an automation and advanced governance your days of choosing between data security and data usability are over. Whether you're just concerned with PCI compliance or need to go further to include CCPA, GDPR, SOC 2 and beyond. Skyflow has you covered. With Skyflow, you can visit skyflowsecure.com today to learn how. You guys have deployed wage advances, SMB lending, even auto leasing through Flexcar, I believe. This is a completely different risk profile than somebody who's looking to get alone to buy a new TV or to get braces for their kids or something more lifestyle oriented. How do your models adjust for these types of variances? The standard loan seems relatively straightforward and I'm probably oversimplifying but it feels like it's pretty straightforward. You got these things, you pull this data, you look over here, you figure out what your risk tolerance is but when you have that many things in a closed ecosystem, I can only imagine the way as a data scientist look at it changes the perspective. That's absolutely correct and we find that the same person can have or same business can have a very different risk profile for different products and in different circumstances. I think one of the shortcomings of the traditional ways of lending is that people tend to be treated as constants even in these different contexts. The way that we think about that is that if we're working with a lender who has an existing business and has an existing set of data, we will retrain a custom model just for them on that data. The risk factors to get selected for inclusion in the model, the way that flows through to the elasticity matrices, all of those reflect the actual experience of that particular product for that particular lender, for that particular customer's segment. Where lenders don't necessarily have that data, we work with them to start them down a path which will enable them to get that capability fairly quickly and dial up the portion of their lending decisions that are based on cash flow underwriting. You can very quickly see that performance improvement even in a situation where you're starting with no data from the beginning. In that scenario, we will have off-the-shelf models that someone can utilize from day one if they have no data. Once they put loans out there and there's a couple of strategy they could use for how aggressively they want to grow, we'll get that signal back. And once we have that signal on performance, we very quickly start putting them on a custom model which might have some of the standard off-the-shelf features and weights along with the custom signal we're seeing from their specific portfolio and over time that will go more and more and more towards their specific portfolio. Within that, you could put a ton of money out there on day one. I think the more conservative and safe thing to be doing is ramp up over time. So we'll only put out a small number of loans, get the signal back, refine, improve and iterate really quickly. It's kind of like your traditional product mindset on how would you go? Unless you're VC backed and blitz scaling and you want to take all the risks, take that first approach, but otherwise go with the letter. One of the things that you've mentioned a couple times is data learning, adjust the model, data, learn it, adjust the model. We've mentioned that a few times. What a lot of people that I work with are concerned about is, that means that I have to understand how to adjust the models. I have to understand how this decision impacts that decision. How are you guys at Carrington Labs looking at hate to call it this, but the model is learning from itself and learning from the behaviors that happen based on the decisions that have been made, based on the factors that come into play. How much real self-learning actually can happen? There's quite a lot of learning that happens as the model gets more data over time. If a lender is already lending with some scale using traditional mechanisms, we can very quickly use that data to put a model in place. That model will almost always be immediately additive to what they're doing. But then as more data flows through, as we get a wider range of limits, this often leads to a much wider range of limits and higher average balances, we will retrain that model on a regular basis and it will get sharper and sharper. Generally, what we, the way we work with clients is we will host the model for them. We see their data coming through it because we're doing the scoring and we will just retrain that model in the background and they will just see that performance improvements. We do ask them to be across and understand the individual factors just so they feel comfortable with what's going into those decisions. But it's not something where they themselves need to be deep data scientists necessarily. I'd also just say one thing. You never cross contaminate data, so a client's data is a client's data and we don't share that with someone else. The models are purely for the specific client. When you look at lending, it's really interesting because it hasn't, that's a innovated at the pace of other industries. I always like to point to e-commerce because everyone gets e-commerce. So e-commerce, you have the trend of omnichannel and personalization. We don't really have that. When you're looking at risk, traditionally what you're doing is looking at a bureau score and then you're using these generalized risk bands and that's how you're evaluating if someone gets credit. I think that's fairly antiquated. The new way of building things and one of the things we offer at Carrington Labs is you can get the probability of default percentage. It actually sends the unique risk of each borrower which allows you to be super sharp. If you understand that, you can then map that to the underlying economics that each borrower will deliver in the lifetime value. You have a really good sense of the overall piano performance. That's just, I think people get that concept at a high level but it's not there in practice yet. As the data is available, as we can retrain models really quick, as the model is going to be personalized, each product that a lender has, you start to see this and it's a really exciting time. So, Gene, my brain immediately goes over the data side of it. The data side of it is like, isn't there benefit in having consortium data versus making the models siloed from each other? Doesn't the benefit outweigh the risk when you do siloed them or help me understand why it's siloed and not consortium level? So, the reason that we keep each lender's data fully separate is for the interests and benefit of that lender. A lot of the banks and non-bank lenders we work with, they really want their data to be fully isolated and no one else gets the benefit of that and we completely understand that. Now, we do offer if someone wants to be part of what we call the signal sharing consortium, they can do that. We'll still never share the actual data but we will potentially share things like signal weights across that. We generally find lenders really prefer to keep their own data to themselves and we also find that most models actually work very well with just that one lender's data. It's more tailored and specific to them and if that lender has even a little bit of scale that's usually enough to give them a pretty sharp fine-tune model. I think that's interesting because if I were lending money, I would want to know whether or not Casey Payback is loan or not before I decided to do it versus looking at a credit report that shows that it paid it back. But did he pay on the due date? Did he pay like the date that he got the statement invoice? You would think that they would want to share that kind of data to understand the propensity of repayment. Is that not important? Well, this is one of the really remarkable things about using that cash flow data, that bank transaction data, is that if Casey is coming and applying for a loan-- Sorry, Casey, we're not mean to keep you sitting like that. I'm very creditworthy, everyone should know that. If a customer comes and applies for a loan and consents to having their data shared, we can see their financial activity at a high level of detail. So we can see things like, we see a paycheck coming in every month, but then this month, it disappeared and then it came back the next month. So what was happening there? Or we can see a regular credit card payment going out. But then we see that coming down, we see interest charges starting to creep in. Or we can see payments that used to be completely on time starting to drag later and later, whether that's a loan payment or a utility bill or something else. You can even with the right, with the right logic, look at informal lending between friends, if your cousin sends you money, is that a loan? or a gift, and did you send it back, and did you send it back after you'd already spent a bunch of your next paycheck, or did you send it back first thing? All of these things are tremendously powerful and predictive, and with the right logic layer and the right model features, you can extract these tremendous insights into somebody's financial behaviors into their character, their attitudes toward debt from that transaction data. Not an optical off-tab, but I just want to drill into that one because it's really important, so I think the common assumption is if you have more data, you can build a regression model and see what's significant across a large number of people. That's in a sense of generalizing things, and when you get to future generation, it is so important the quality of the features that you can generate. We see it all the time, people brag about having thousands of features in their model, and those features are literally, it's a grocery spend in seven days, 14 days, 30 days, 45 days, 60 days, and it's just like a bunch of the same thing over and over and over again, and they're using this brute force approach. Doing this is not wrong, right? Having one of those features is good, and figuring out if the seven day grocery spend is more predictive than the 90 day grocery spend is probably worth doing. But having that logic layer, and one of the things that we use quite a bit of AI to do fairly advanced future generation, the predictive quality of the feature is so important. That's where we get to these behavioral attributes that power these features. We see quite a bit of model uplift from even a handful of these advanced features can have a significant impact on the genie of a model. Yeah. I had David Glazer, the CEO of Dwall on, and we were talking about a number of different things and the precursors to the economy and the types of things that they're able to see as a payments provider that is mostly intra bank payments providers and seeing how just before the president got a nog, the inauguration, how basically all spending stopped. And then it started back up and didn't take fully off. And then seeing how the differences and the inflation and the costs and all that at Carrington Labs, you guys are able to look at things like the cost of fuel. I mean, as we're recording this right now, the cost of a gallon of gas here in Las Vegas ranges anywhere from 529 to 629 at gallon. I never thought I'd see that, but we have that going on. And I'm imagining that that source to impact a lot of the other pieces that you all are seeing as well. I bring that up because I think one of the big pieces that we don't, the traditional Fintech side of the house doesn't really think about the economic impacts to these types of models. And that being said, how are you using that kind of data? And it's so how? Yeah. So we use data. We do get a lot of insight overall into what's happening in the economy and how are people performing. From the point of view of making lending decisions, what's most interesting to us are two things. First of all, to what extent do changes in the broader economy affect this individual person? So we'll look at that by looking at their own financial history and you see some people where their work really ebbs and flows the economy. You know, a lot of construction workers, for example, will get very, very busy and then they'll be less busy, I'd quieter times. And you can see over a period of time, do I have to say does this person have individual economic exposure to the economy? Then the second big thing that we look at there is how do people respond to those moments of abundance or scarcity? So you talked about people filling up their cars with gas. One of the things we look at is, does somebody shift down to cheaper providers, whether it's of gas or groceries or anything else, when times get tight? Some people, you will see them stop going to Safeway and start going to Costco if their balance is running low and some people want. That kind of behavioral response to your own financial situation is an enormously important part of somebody's creditworthiness, but also actually gives you a lot of insight into broader economics as well. As you're talking about that, in my mind I was thinking whole foods to your grocery outlet, but that's a big jump. Not a small one, like you were talking about. But it also gets me thinking about a lot of times when not in lending, but in a lot of places in Fintech, the underwriting process tends to be a point in time. Have you started to see a shift of point in time underwriting to more of an ongoing recursive underwriting process to decide I need to tighten my internal policy, but for this individual based on the model, and I'm going on an tangent here, but in my mind is like, all right, so we see that this person owns a vehicle that has a very large engine that gets extremely low gas mileage. We're seeing them continue to use these purchases at these pumps, but the income hasn't increased to go along with it. Maybe we make an adjustment. Are you seeing that shift? Is it going that detailed? It's a really interesting space. When we started, our first product was a credit risk model. It was looking at scoring at time of origination. The data is temporal in nature like you pointed out. As we've evolved, we really now attract the full borrower life cycle. We now also think a lot about the servicing element of that, because once you write alone the money's out there, so then it's about what can you do to make sure that you can recover those funds, detect early, distress, and a portfolio. We have a cash flow servicing capability that looks at that. If the open banking data is being persisted, if the connection is persisted, we can see like, are these changes happening? If fuel goes up, is there less discretionary income? Does that mean that someone may not be able to pay the loan or whatever that might look like? We do think that lenders should be thinking full life cycle, not just origination. By having the capability to rapidly retrain models based on new signals that's coming in, whether that is at origination or on servicing, and as well offers a huge uplift. It's early days for that. There's not many people who are doing advanced things on that servicing piece right now. What tends to happen is you get a monthly report on how the portfolio is performing, and that's what most people do. I think being proactive based on what the data is telling you in your real time offers a lot of opportunity. I think just to jump in on that, there are a number of products where you can actually adjust your exposure after originations of credit cards being an obvious one, where somebody may have a certain limit, they may not be using all of it. We can go into a lender with an existing book of business, look at the, say, 100,000 credit card holders, and say, okay, of these 100,000, these 5,000 are, we see some stress emerging. If you're in a position to pull back some unused limit, that's probably worth considering. Conversely, what we much more often see is, well, these 30,000, you've had a very conservative setting on them. They are doing quite well, and they can actually, if it's appropriate and lines up with regulatory objectives and your funding position, and so on, you could actually consider meaningful increases to them as well. Yes. So, one thing we keep in mind is, how do we drive more dollars to your margin? I don't know if that's true for all the technology providers out there, but ultimately that's what a lender is trying to do. Like James said, we see so much uplift by doing the limit in line management, and it's something that offers a lot of value. People know they should be doing it, but they just don't necessarily have the solutions in place right now to be doing it. Well, hopefully you guys don't mind, but I want to do a little bit of a side quest, because you mentioned something that is near and dear to my heart, and it's something that I do a lot of work around, and I speak about a lot. But you mentioned open banking, and in the US, open banking has started to feel not so open with financial institutions like JPMorgan, and the most prominently known, deciding they're going to start to charge to get access to this data. How is that impacting the way that your customers are leveraging data, the way your customers are accessing the data? I might let Jamie start with that one. So Jamie, I'd say is our expert on this, helping write some of the Australia standards. So what do you think, Jamie? Yeah, so I'd say a couple of things. First of all, I think any bank that wants to put up barriers to data sharing from his customers I think is showing a lack of confidence. because what they're doing is we're saying, we think we need an asymmetric playing field in order to compete with attackers and other lenders. And so I'm surprised that JP Morgan, which I think is a very capable bank, feels that it can't fight on a level playing surface for a customer's lending business. Now you said that. I also think that ultimately, this is data that the customers have created that they own. And I think the bank, just from a sense of fairness, should enable the customer to take that and share that with other providers as they think best. Our experience has been when barriers to data sharing go up, it's not that data sharing goes down, it's that data sharing moves into the back alleys and it gets done through people scanning and emailing bank statements to somebody. So it's done in a much less secure and much less coordinated way. So I think it's a step backwards for the financial sector as a whole if we raised barriers to sharing data. - It's so interesting that you mentioned the email piece of it because a previous guest that come to called Lunoz, they're helping accounts payables, accounts receivables, all these other things. And we were jokingly saying that, you know, APIs are really cool and everything, but a lot of people live in email and business is done mostly in email. So is email like a regression away from APIs to get the data in a more free and unrestricted way. So that's really interesting that you bring it up from that perspective. Casey, you're chomping it a bit to say something. - There's different sizes, right? So I think it's very easy to say, let's charge large financial institutions who are very profitable, or we should make a list, should not charge users to access the data, right? But when you look at open finance more broadly, and if you say smaller organizations, also have to share the data, which they should, but the regulation puts so many restrictions on how that data could be shared and the standards. There's a huge cost for those smaller organizations. So I think you just need to find the balance, right? It might have a significant P&L impact to build out the infrastructure based on what the regulation says to allow customers to access the data. I absolutely think customers should be able to access the data, but we just need to be mindful of what happens. And I think some of the open banking standards that have been proposed in the US are a bit more open than what's happening and say Australia, for example, but no, Australia, there's a lot of restrictions even around derived data. So if you are using the open banking data and generate insights off that you're quite restricted on where that data can go and how you can use it as well. So I think we just need to think through end-to-end both sides and get to the right outcome. There's been really interesting watching open banking in Europe and the discussion that I've had a lot around that is more of that was meant to make it easier for a customer to move their data to simplify the process to be better served. And I think in the US, it has gone to the perspective of how can we monetize on the data itself, not on the services we deliver. And so that's the piece that gets a little bit disconcerting for me is it as a business and as a consumer. I want to freely be able to have my data and be able to send my data where I want to do, where I want to send it. And Jamie, like you said, if I'm stuck in a scenario where I have to email it so that I don't to pay a fee for it, I'm probably going to do that. So I'm getting into my favorite section of every show. And so one of the things that we look at, lending and AI-supported lending, you know, it's estimated by 2037 that it's going to be a $20 billion space for loan origination using AI tools. And you look at that, what I want you guys to do, and yes, I bring out the good old crystal ball. - Ooh, an actual crystal ball. - An actual crystal ball. So what I want you to do is I want you to look deep, deep inside of this crystal ball. And Jamie, I'm going to have you go first. What I want you to do is I want you to look in, go out three years, see what three years is, and say, yeah, not good enough, go out five years, and then come back and tell us, what is it that you see happening using artificial intelligence and lending? - So I think looking at five years, there'll be two big shifts in lending driven by artificial intelligence. The first one is the actual process of applying for credit, getting credit, deploying that money will be embedded much more in a much more native way into our financial lives. Our agents will be out there seeking credit for us, pulling it back and deploying it in a way that's most efficient for us, without us having to go through applications that will seem very archaic. Then the second thing that will happen will be lenders themselves will be making decisions based on a much richer and fuller data set that gives them a much deeper insight into a customer, whether it's an individual or a business. Their creditworthiness, how they will respond to different financial situations, and they'll be able to deploy credit in a much more precise way, and overall, the availability of credit to the right people and businesses will be much, much higher, and they'll be a structural and permanent boost to the economy off the back of that. Casey, what do you see? - I think when I look five years out, I see what tends to happen is regulation, legs, innovation. So I think there's gonna be quite a bit of policy update that is more supportive of AI and AI driven processes as the AI solution's also evolve. The other more general thing I see happening is friction is gonna be removed from the process, which is not too dissimilar from what Janu was saying. And we see it across every industry and every technology that's been successful when you look at phones, we had telegraphs and landlines and car phones and feature phones and now smartphones. I think lending is gonna go that direction as well, right? So whether it's new data that's being used, new ways of modeling, new approaches to understanding and risk in maximizing margin, new types of accessing capital, I see all of that coming to fruition. I don't think it's a matter of, if it's just a matter of when. - Do you think we're gonna see the number of companies offering lending, expanding, or is it just the same folks maybe with a different logo? - So we see, especially in the US, a number of the financial institutions consolidating and some of that is because of legacy issues of only being able to have like a state license in one state at a time from a very long time ago. So I think we will see a lot of consolidation because smaller financial institutions who are lending right now just don't have the capital to innovate. I think once that starts to plateau, I think we can then see more innovation happening and it'll be more destruction. So a little bit of both, but we'll have to see how it plays out. - I was having this discussion with a scaled car retailer that was trying to work out whether they should be involved in lending directly or not. And we don't know the answer, but there is a lot of value in having the organization closest to the customer and the flow of data coming off those interactions and transactions, having that company be the lender because they have more data, better access to the customer than anyone else. So if we can build out the tooling in the systems that enable even non-financial businesses, perhaps smaller businesses that access to cutting edge decision technology, limit setting, origination system, servicing capabilities, I think you could see lending actually spread much more widely across the economy, potentially in a very positive way. - So if that happens then, as some of the things that are specialized now have become commoditized, and if you're a product person as well, that's where your product market fit, like really comes in from that end consumer that you're interacting with, right? So it's no longer about, do you have the best technology if everyone has access to the same technology? It's what's the proposition to the end customer and can you position that in a unique way and your specialization that will ultimately make you win? And that's really exciting. - All right, well, last question of today is, if I wanna know your view of your perspective, if you were to tell somebody's over here that hey, we've got one piece of advice for you that if you follow it, it'll change the game for you. What would it be? And why? Jamie, we'll start with you. - So my advice to any lender is bring together the process by which you approve or decline alone with the process by which you set the terms of that loan, the limit, the duration, the pricing. Too often those things are separate and as a result, lenders miss out on the opportunity to optimize the value of the customer. that's there before them because they don't think deeply enough about what's the right limit, what's the right duration, what are the right terms to these lines? It's a case, I would love if you could take it from the approach of the product market fit commercial side of the house. What advice would you give somebody who's looking at do I build a technology, do I market a technology, do I become a lender, do I expand my lending, like look at it from that perspective, what advice would you give them if it was the only thing that they could use this year? Let's say like the only thing that limits you in life is yourself and you need to think about what that means. I think if you are a lender and you specialize in the relationship you have with your customer, I would not be trying to build my own technology in house, right? I think there is great solutions out there from the origination, underwriting service inside and I would buy the best of breed and bring those together in a way that Jamie just described and I think the relationship you have with your customer, those interactions is the most important. I think if you have a generic product, you're not differentiated, so see how can you differentiate and really understand what are the pain points of your customers? That will require some testing, chances are you will create some products that customers might not use or don't work out the way that you want and I wouldn't take that as a defeatist approach that you shouldn't try anything else. I think it just means you need to take the learnings and try again. With that said, I'd be mindful of spreading yourself too thin, too fast, right? Like find your niche, get that flywheel going and get that working and then think about iterating over and over and over again. I think a lot of fintechs fail because they do too much at once and then you probably have the opposite out of the spectrum where there's a bit of an innovator's dilemma for more of the established businesses. They're not willing to try new things and they ultimately get outpaced by some of the emerging companies out there. Well, guys, I really appreciate you taking the time out today to sit down with me. There are three things really that I took away from today. One is that it's not a simple yes-no scenario. The second thing that I found interesting is that just because you're using the data you have access to from within your silo doesn't mean that you're restricted, that you can leverage the data that you have to still make great decisions. And then last but not least is that we are still in the early stages of this technology and what it can and can't do. Well, sorry. Thank you so much for joining us today. It's been really good to have you. If you got value from this conversation today, be sure to head over to YouTube, Spotify, Apple iTunes. That's where you can find more great conversations like this. But also go ahead and head over to fintechconfidential.com and subscribe. That's where we do a lot of deep thives. That's also where we host all of our newsletters. Find out what's going on in all the areas of fintech. And as always, keep moving forward. As we wrap up today's episode, I've got one last thing for you. If you're in the trenches fighting fraud and financial crime, you know it's a complex battlefield. That's where hawks, AI tools for real time payment screening, AML, transaction monitoring and dynamic customer risk rating come into play. These aren't just buzzwords. They're game changers designed to make your compliance more effective and less of a headache. Imagine slashing through false positives with precision and giving your compliance strategy the edge it needs. Head on over to get hawk AI.com to sign up for a demo and discover how their platform can revolutionize how you fight fraud and financial crime. This has been a production of DD3 media with all rights reserved. This is provided for informational purposes only. It is not offered or intended to be used as legal tax, investment, financial or other advice. We strive to provide accurate and up-to-date information but will not be responsible for any missing facts or inaccurate information. You comply and understand that you should use any of this information at your own risk. Cryptocurrencies are highly volatile financial assets, so research and make your own financial decisions.