Trading Tomorrow - Navigating Trends in Capital Markets

Building Responsible Innovation in the Age of AI

• Numerix • Season 5 • Episode 48

Share episode

As data and AI reshape financial services, the question isn’t just what can be built but also, how it can be built responsibly.

In this episode, Melissa Koide, CEO of FinRegLab, joins host Jim Jockle to discuss how technology and regulation are evolving together. From credit inclusion and data privacy to explainable AI and model oversight, Melissa shares insights from years of research bridging public policy, innovation, and consumer protection. 

Jim Jockle:

Welcome to Trading Tomorrow, Navigating Trends in Capital Markets, the podcast where we deep dive into technologies reshaping the world of capital markets. I'm your host, Jim Jocko, a veteran of the finance industry with a passion for the complexities of financial technologies and market trends. In each episode, we'll explore the cutting-edge trends, tools, and strategies driving today's financial landscapes and paving the way for the future. With the finance industry at a pivotal point, influenced by groundbreaking innovations, it's more crucial than ever to understand how these technological advancements interact with market dynamics. As finance and technology continue to converge, one question we often hear is not just what can be built, but how it can be built responsibly. Across areas like credit, payments, underwriting, and identity, both regulators and innovators are exploring new models and learning from the results. To help us examine this evolving landscape, we're joined by Melissa Cody, CEO of FinReg Lab, a nonprofit research center that tests new technologies and data and facilitates dialogue to inform public policy and industry practices. Finreg Lab's work focuses on evaluating how technology and data can be used to increase financial inclusion and improve financial services for consumers, small businesses, and communities. Before founding FinRED Lab, Melissa spent nearly five years as the U.S. Treasury Department's Deputy Assistant Secretary for Consumer Policy, where she led research and policy development on how financial institutions were leveraging data and technology to expand access and improve consumer well-being. During her ten years, she also helped establish the NYRA savings program and regularly testified before Congress. Melissa currently serves on several advisory bodies, including the New York Feds Innovation Council, FINRA's FinTech Industry Committee, and the New York State Department of Financial Services Financial Innovation Advisory Board. Drawing on experience that bridges research policy and practice, she brings insight into how innovation and regulation are evolving together in today's data-driven financial system. Melissa, so wonderful to have you today.

Melissa Koide:

Yeah, happy to jump in.

Jim Jockle:

Arguably, it's probably some of the most important work to make sure everything is operating the way it should. So why don't we start there? From your vantage point, you know, how would you describe the current balance between innovation and inclusion in financial technology?

Melissa Koide:

You know, I mean, every day you think, goodness, could the technology sort of advance more? And, you know, data is instrumental to ultimately getting real technology uh analytics advances, but every day is a new day. And I do think we are at this really tremendous uh place where there's the opportunity to prudently and accurately expand market opportunities using data and technology that's got benefits for the industry, right? These are new market opportunities, but also importantly, opportunities for consumers and small businesses, many of whom millions have, you know, traditionally not been able to be really evaluated or assessed for access to the financial system. But getting to those advantages means that we have real legitimate uh matters that we have to get our heads around and tackle, especially as we're thinking about artificial intelligence and data use. You know, there are questions around explainability, trust in the models, confidence in the models, and you know, the consumer data access and its use and questions around consumer data privacy. So enormous opportunity, and we'll talk, I'll talk more concretely about what we see there. But that is in part some of these uncertainties is really what drove the creation, why I stood up VINREG Lab roughly about eight years ago. Um, I had been sitting on the U.S. government side, I was the head of a title that's um too many words, but nevertheless head of the Office of Consumer Policy in the U.S. Treasury Department. And you've got lots of points of view, right? Um, concerns and anxiety from the consumer advocacy community about how our data and technology are going to be used safely, lots of enthusiasm, especially from the non-bank and the fintech sector. And banks and others are trying to make their way through these, you know, opportunities and questions at the same time. So having some place where we can actually generate empirically derived, informed data and insights combined with the opportunity for dialogue among all those different stakeholders. And by the way, they're all going to have a voice when it comes to evolving policy, law, and regulation. That's the impetus, and that's what we um have set out and have been doing over the past eight years.

Jim Jockle:

So now you study how technology and data intersect with policy goals. You know, how do you decide which areas to test or engage in?

Melissa Koide:

Yeah, great question. So we do tend to, we're a relatively small team. Um, a number of us are former government and public policy officials, and we do really dedicate the work that we're doing to help inform policy makers. And so we spend a fair bit of time. We're located here in DC, really talking to regulators, but also just as importantly, talking to the financial ecosystem and increasingly the technology ecosystem. We're really motivated where there are possibilities where a data use or an analytical approach can have high impact. So where there's significant inclusion promise, but also where there are thorny questions around the responsible implement implementation or the use of the data or the technology. And then we look for areas where empirical evidence or the lack thereof it is creating barriers to the adoption of some of these beneficial technologies.

Jim Jockle:

So now, where have you seen technology meaningfully expand access to financial services? And where have you seen trade-offs emerge?

Melissa Koide:

So when we first got going, this was back in 2019, um, we wanted to start because from an inclusion standpoint and from a new market opportunity standpoint, we knew that the data element was instrumental. We weren't even talking about generative AI, language models. Banks for a long time had been using machine learning models, but instrumental in sort of the advantages, the nuance that those models can bring, but also the inclusion and the consumer expansion opportunities is by starting and really trying to understand the data insights. Here in the US, we have 20% of US adults who lack sufficient credit history using credit history data to be scored under some of the most widely used models. And we have another 30% who may not be able to access things like affordable credit because they have thin file. They're there, or they may be non-prime. So insufficient data or no data at all. But yet at the same time, we have millions of people in US adults and households across the country who engage in transactions, right? Your bank account, or think about a prepaid card. And so we really wanted to interrogate and evaluate the extent to which that consumer transaction data, what we shorthand is cash flow data, where it may actually offer not just the inclusion benefit or the new market opportunity benefit, but also the benefit of improving accuracy in a credit risk decision. And so that's some of the work that we've been doing over the years. Um and this really then does uh engender questions around access to that consumer data, data privacy, and important needs around the data infrastructure. How do the data get accessed and transferred? Um, these are all sort of elemental, I think, starter pieces to ultimately understanding when we're looking to things like generative AI or even just supervised machine learning, how are you training those models and really deploying them to their full benefit? A lot of that is contingent upon having rich but also representative data.

Jim Jockle:

So perhaps you can share some insight into your process. You you mentioned you have a relatively small team. Um technology, especially within it with AI, is advancing so rapidly. You know, what what is how are you integrating um interrogating the data?

Melissa Koide:

Yeah. Unfortunately, uh these projects are often quite bespoke. Um so we we really do it it will vary. Um in some cases, we've actually, well, we've begun to build the team internally that have industry expertise, and I'll mention this later, but I think, you know, the ability to build industry grade models or production grade models are actually pretty important for building shared confidence in the kind of research that we produce. But for instance, you know, we have done work on machine learning models, and there has been in the past a lot of regulatory concern about this black box nature of those types of models. So there we partnered with a team at Stanford in the Graduate School of Business to really interrogate and evaluate some of both the proprietary techniques for generating explanations that are required legally, uh, but also techniques that are even open source techniques for generating some of those explanations. And also in this regulatory need of evaluating for finding less discriminatory model alternatives. So it will vary to answer your question. Um we've also hired uh outside experts to help us do some of the work that we've done. And as we think about work into the future, we will definitely be um partnering with others. That's an important component of how we go about our work.

Jim Jockle:

Well, regulators often raise concerns about fairness and black box models. Uh what you know, what kind of transparency or safeguards seem most effective at this stage?

Melissa Koide:

Yeah, well, that is in part what we were evaluating with this particular explainability work around machine learning models. And part of the question is do what extent do, and this is in supervised machine learning territory. So it's it is to acknowledge it, it's not the same thing as non-deterministic generative AI and language models, but but this is very much an approach that especially large institutions have been able and have been deploying, but yet they are more opaque and they are more complicated to interrogate for explainability. And so what we found in short is these models can be deployed, but it really does for confidence, frankly, for the institutions themselves, let alone the regulators in the broader ecosystem, it does require human oversight and expertise to understand what are the data that you're putting in, what are the potential correlations in the data, so that when you are looking at techniques that are essentially evaluations of the models after the models have derived their outcomes, there is real human insight and understanding of how those models were being used and deployed and the data that were going into them.

Jim Jockle:

So measuring impact obviously can be really challenging. Um, I guess first, how do you define your uh agenda in terms of which data or policies to interrogate? And we through that process, um, how do you evaluate whether a new policy or technology is really achieving its goal?

Melissa Koide:

Yeah, no, that's a great question. Um, definitely setting out expectations and policy-relevant insights from the beginning so that the work that we're doing is responsive to some of these legal and market questions. You know, I mentioned the cash flow research we did there. Part of how we do the work and how we evaluate if it's useful and if it's having market impact is to evaluate the extent to which the regulators in the broader financial ecosystem are engaging in the research itself and importantly in the dialogue work that we do on the other side of the research. So, with the cash flow research, for instance, we engage with six fintechs who were using this kind of transaction data to make credit decisions. Um, we had an advisory group that helped to inform how we design the research. We've also done that where we've had advisors even from the regulatory community helping to make sure that we are structuring the research in ways that are responsive to the questions that they're grappling with. In the case of the cash flow work uh after the research, which I'd love to say took just a couple months and it was a bit longer than that. We then convene many of those different stakeholders along with the consumer advocates and also those regulators from the Consumer Protection Bureau all the way to the big bank regulator. And they got to be a part of the conversation in where we led a discussion about what does this mean for fair lending laws. We led a discussion about what does this mean in terms of adverse action notice requirements, right? Laws that are explicit around if you're making credit decisions, you've got to be able to generate these explanations and you have to have confidence from a model standpoint. So that work went on, I want to say it was probably about six months. The regulators, on their own accord, but but very much had been a part of our process, decided to jointly issue a statement to the financial services sector, noting that this type of data can have inclusion benefits, but it also, you know, is important to make sure that you are using it prudently and that you are conducting your fair lending assessments. But that joint statement, we think, really helped to compel the industry at large to lean in um much more aggressively in using this data. So banks are using cash flow or transaction data, both obviously on their own customers, but they're sharing it through intermediaries like EWS. Um, we've seen fintechs continue to use this data. And importantly, we've had an explosion of an infrastructure, right? The data aggregators, a number of them are in existence and are facilitating access to this transaction data. And that's an important piece, right? Plaid, this goes beyond just credit underwriting, but PLD reports that is now powering half a billion account connections daily. That's tremendous. Um, you know, over 10,000 financial institutions being linked. We also have seen the proliferation of um what are essentially score creators in the cash flow data space, right? Nova Credit has been active in the space. And that's really important just to make note of this, because we also have thousands of smaller financial institutions who may not have the capacity or the skill set to bring in the data and to build more complex models. And so being able to leverage third-party vendors who are building things like scores, which make it a lot easier to just lift and use, has real market and ultimately consumer benefit. And I would say small business uh credit access benefit as well.

Jim Jockle:

And many firms experiment through pilots or sandboxes. You know, what distinguishes programs that scale from those that don't?

Melissa Koide:

Yeah, great question. And I feel like we've actually uh through trial and error learned a number of important lessons, um, which I think would be helpful because there are right now um proposals on the hill to actually enable sandboxes to be used, uh, including by the regulators. So I'll just tick off a few. Uh, and you you kicked it off. Um, one, having clearly defined metrics and thinking about what you're trying to ultimately answer and and what questions you're trying to inform up front, I think is an important piece of this. Very much uh generating evidence and insights that are not just informing one particular firm, but really helping to inform the broader ecosystem and the policy dialogue that's happening. Um, and that I think means bringing in uh attention to legal and regulatory questions. You know, I've mentioned fairness a number of times, um, you know, needs for explanations, need for model risk management, successful in doing that is an important piece of this, but building that in from the start, bringing in all those different stakeholders is also critical, right? You don't want to get so far out, but yet you haven't really educated all the different voices that are ultimately going to inform whether or not the policy community is comfortable with something. So make sure the advocates are also a part of your effort. And then, you know, helping to think about like down the road, what are the potential policy and regulatory framework changes that may be needed for broader adoption? I think those are important pieces along the along the route.

Jim Jockle:

Obviously, emerging areas of data, whether it's unstructured data, identity data, social data, um, uh alternative data, et cetera. You know, how do you see the tension between predictive power and fairness of all that?

Melissa Koide:

Yeah, that's a great question. And it's, you know, this notion of sort of prudent financial services, yet consumer protection, yet inclusion, are they separate, separate objectives from each other? And this is where I think building the research and the effort, we're acknowledging it's not just choosing one over the other. And this is where we are especially excited about, you know, more representative, more nuanced data insights combined with more sophisticated model analytics, because you can actually get after both accuracy questions and also inclusion questions and even fairness questions. But you are building those into how you're designing your data selection, your model choice. And then from an evaluation standpoint, making sure that you're really interrogating across those. I think getting after both of them is possible, especially because of the data and the analytics that are in front of us.

Jim Jockle:

And looking ahead, which parts of the regulatory landscape do you expect to evolve the most quickly because of improved data, access to data, or even AI?

Melissa Koide:

So I see several areas of AI in machine learning and the data helping to improve financial services at large. And I think, frankly, especially over the next three years, the regulatory landscape, the regulators, to be more concrete, are going to be watching and learning. And given what the language models that are now in front of us may offer, I think it's an exciting time for industry to be testing and learning and really bringing that to the regulatory community and all the other stakeholders too, because there is the space for this kind of test and learn opportunity with the Trump administration. And I think how they're really trying to allow for innovation. But where are things now? I mean, already we know that AI is being used in things like back office automation, um, fraud detection, anti-money laundering models and techniques are already happening now. And I think credit underwriting and things like identity proofing, are you who you say you are, as well as identity verification, are actually pretty close behind, especially with some of the larger institutions. And that I think, you know, and then if we overlay sort of these opportunities around models training other models, this agentic AI development that's really sort of come on the scene over the past six months, as well as e-commerce applications, like there's a lot ahead of us that I suspect will be happening pretty quickly over the next um two to three years.

Jim Jockle:

And finally, what's one way innovators and regulators could better collaborate to align on innovation with the public trust?

Melissa Koide:

So again, I I, you know, I'm I'm a little bit sort of beating our drum here, but I do think getting some of this independent empirical research out uh done, and that does take time, and it does take engagement from industry partners to make this kind of thing happen. Um, but just to acknowledge, when I was sitting in the Treasury Department and, you know, having meetings upon meetings with industry stakeholders and innovators, our techniques, they're great. You should just trust them. The next meeting would be with consumer advocates who are making the case, you know, this is the end of consumer data privacy, stop it all. You know, those voices are going to continue to be there. Um what I know that we need are more platforms where everybody gets to come together and we're engaging in this evidence-based research and insight. And then we're getting beyond just sort of trading the competing claims, but actually basing our market product evolution as well as the regulations on, you know, fact-based insight.

Jim Jockle:

I think there's a lot of areas we could we could have that model work, not just in uh fintech and regulation.

Melissa Koide:

That is true. Fortunately, my day job is to focus in this space.

Jim Jockle:

Okay, well, you could solve some more problems. I mean, you know, you've you've got a framework. So unfortunately, we've added to the final question of the podcast, and we call it the trend drop. It's it's like a desert island question. And if if you could only watch or track one trend in fintech, what would it be and why?

Melissa Koide:

Yeah, so we actually just put out a paper on this because we are definitely tracking it and thinking hard about it. Um, this is the area, the space of agentic AI and e-commerce and the possibilities and questions around stable coins. Um, you know, I think this has the potential to fundamentally affect how financial transactions and e-commerce are going to be happening. I'm not sure how much you all have been tracking this, but I mean, we could potentially be moving to a world where AI agents are autonomously executing financial decisions on behalf of me, on behalf of other consumers, small businesses, and then the the sort of supports and you know, the Genius Act and real encouragement from US government around stable coins, you know, is that going to create a new set of payment rails that are enabling those kinds of transactions to happen? But yet on the other hand, we still really must ensure that we're building the right tick guards and doing that from the ground up. So, how are we? The questions that are on our minds are how do we ensure that these kinds of systems are actually designed and working in the consumer's best interest? How do we maintain the transparency and the accountability when the decisions are being made? What kind of consumer protections do we need in place? Where does this is the question that's really on my mind right now? How do we understand what the objectives are with consumer protections? Like who bears responsibility? Does a consumer get made whole? Yeah, does a transaction happen? Is there a way that we can answer some of those really confidence-building needs, but do it from a technology standpoint, as opposed to rushing in immediately and naming it from a policy, legal, or regulatory standpoint? I think those are some of the interesting opportunities that technology offers ahead of us. But, you know, if we don't get this right, it could be quite scary. If we do get this right, it could be really powerful. Uh I'm excited about it. And I'm I'm on the old side, but I'm still pretty excited about it. But that's that's an area that I'm watching closely that we are definitely wanting to understand. So yeah.

Jim Jockle:

Your excitement is clear. Melissa, I want to thank you so much.

Melissa Koide:

Thank you. I appreciate it. This was fun.