Robert Flatley, Founder & CEO TS Imagine

[00:00:00] And look, tokenization is not a marginal upgrade. It is a structural reconstruction of the economics of intermediated finance. I mean, it's going to change all of it. And there's kind of three things, your favorite word coming up, there's three things to anchor that on. Atomic settlement collapses intermediation. Collateral becomes mobile. Infrastructure economics concentrate.

[00:00:23] Any one of those things would require people to kind of act urgently. But together, they kind of create this world where firms that build the rails will own everything that flows through them. And the firms that wait are going to pay rent for someone else's infrastructure for the next 20 years.

[00:00:39] Dean Curnutt Hello, this is Dean Curnutt and welcome to the Alpha Exchange, where we explore topics in financial markets associated with managing risk, generating return, and the deployment of capital in the alternative investment industry.

[00:00:59] Robert Ruffin I was excited to host this conversation with Rob Kaplan, Founder and CEO of TS Imagine on prediction markets, AI-driven workflows, and the structural changes reshaping financial market infrastructure. We begin with Rob's path from software engineering into capital markets, including leadership roles at B of A and Deutsche Bank, during the rise of electronic trading and through the GFC.

[00:01:22] That experience informs a broader perspective on how market infrastructure evolves during periods of stress and technological transition. The conversation then turns to artificial intelligence and the distinction between large language models and reinforcement learning systems. Rob explains why traditional, deterministic workflows in settlement and collateral management create different challenges than probabilistic systems such as risk management.

[00:01:49] He argues that the next phase of AI adoption will focus less on generating language and more on learning and automating complex workflows across financial systems. We also explore prediction markets, an area where Rob and his team have spent significant time building infrastructure and risk frameworks.

[00:02:08] He discusses how markets tied to elections, Fed policy, GDP, inflation, and geopolitical outcomes are beginning to move from retail experimentation towards institutional relevance. We also discuss tokenization and settlement infrastructure. Rob outlines how stablecoins, digital ledgers, and atomic settlement could reshape financing, custody, collateral mobility, and the economics of intermediated finance.

[00:02:35] We discuss the implications for prime brokerage, repo, clearinghouses, and 24-hour trading environments. I hope you enjoy this episode of the Alpha Exchange, my conversation with Rob Flatley. My guest today on the Alpha Exchange is Rob Kaplan. He is the founder and CEO of TS Imagine, a risk management service available to institutional investors around the world. And Rob is someone I've known for more than 20 years.

[00:03:05] Rob, it's great to have you on the podcast today. Dean, it's great to be on the preeminent podcast in Capital Markets. I appreciate the offer. Fantastic. I appreciate the kind gesture. And as I said, you and I go back to our days at B of A. And I want to say 2003 or 2004, you were running electronic trading. I was running derivative and convertible sales. And we collaborated on some of the earliest electronic option trading technologies there were.

[00:03:35] It was a precursor to so much development. You've done a lot since we first met. You've created and sold businesses. You're now running a large organization that you brought together through a merger. Briefly walk us through your career post Bank of America, the areas you focused on throughout the 2010s and up to now. Happy to do that.

[00:03:58] After working with you at Bank of America, I moved on to run electronic trading and market structure at Deutsche Bank. Pretty much through the credit crisis. Left there in 2010. And my original roots were software. Like everybody, I started out, I didn't know what I wanted to do. And eventually fell in love with software. Because I was implementing it and I was buying it. And it seemed like the guys who were selling it were having more fun than I was. I was in the software business long before I joined Bank of America.

[00:04:26] So I learned capital markets and financial services with a lens as a data architect and a software engineer and a salesman. Deutsche was a great place. But I think by 2010, you know, I saw the writing on the wall in terms of banks transforming post-credit crisis. And I had a good view of what I thought the right software angles were post-credit crisis. Selling software for financing. Selling software for regulatory reporting. Delta One was kind of on this big trajectory at the time.

[00:04:56] So I ended up building four companies inside of a company that I founded called Core One Technologies. And built that up into a mid-market company and sold that to IHS Market about five years later. And then I knocked around like you're supposed to do and tried crypto and some other things. Built some businesses along the way. And then in 2021, I hooked up with Francisco Partners, a great California-based private equity firm. We landed on a thesis that we should get back into the software market.

[00:05:26] People are frustrated with the software that they use in the investment process. Be that portfolio management, trading, risk management, financing, etc. So I bought two charter franchises. Trading Screen and Imagine Software. I was a customer of both products during my time at Bank of America and Deutsche Bank. And I had a pretty good idea of what those platforms could do. And I felt like those were the right assets to bring into the market and modernize.

[00:05:53] So that's kind of the resume version of what's been going on since Bank of America days. And I'd say, you know, what I actually took from all of it and what I think matters for what we're going to talk about is it kind of comes down to three things. Every industry, because I was in the newspaper industry, you know, in the beginning of my career and then moved into financial services. Every industry thinks infrastructure is permanent. And none of them are right. Newspapers had massive publishing infrastructure. That's all gone. Trading floors had armies of people.

[00:06:23] They don't have armies of people anymore. There's people, but not armies. And capital markets got comfortable with its rails and automation. And now AI and tokenization is set to flatten that too. So the pattern is the pattern. You know, every generation has its, we'll always need this job until it doesn't. That's what's so great about this industry. Well, we're going to talk about a couple of different topics here.

[00:06:46] And a lot of this is inspired by what I would just say is some excellent thought leadership that you and your team are doing. It's available on the TS Imagine website. There is a raft of very interesting, well-constructed content on there. And I've taken some time to go through it. You and I had breakfast recently. We talked about prediction markets. That's going to be a fun part of our conversation. We'll talk about some of your interesting views on where the moat is and isn't in SaaS.

[00:07:15] And then tokenization and what that means for carry, for repo. That'll also be a topic. As you talked about your career post B of A, your time at Deutsche Bank, and you mentioned the financial crisis. I can't help but ask, and I'll ask you through the lens of one of the statements I make often, which is you learn the most about markets during the periods when things go horribly wrong. And it's hard to argue that that's not the most significant period in modern finance.

[00:07:43] What might be a couple of takeaways from the lens of, as you say, data architecture of systems that you experienced through your time at DB through the GFC? That's a great question.

[00:08:25] What might happen? But also use that as a weapon to go grab new business. That was the lesson that we learned. We actually did really well on the Autobahn platform in 08 because Deutsche Bank was in good shape through the crisis. Some of the other firms like Bear Stearns and Lehman, we competed directly against those firms for customers.

[00:08:45] So we were able to scale our technology, both from a volume perspective and also, I would say, ease of integration and onboarding to actually turn that into a positive result for our firm. Going back a long, long time here. But I'm remembering being a young up-and-comer at Lehman Brothers in 1998. I got my first big break. I was covering long-term capital. And of course, that's when they imploded.

[00:09:09] The risk systems were really insufficient in terms of really understanding where the exposures were across swaps, across options, across different geographies around the world. What is that like through the financial crisis in terms of really understanding having systems that are sufficiently flexible but also powerful to really understand where the true risks are?

[00:09:36] You've got two types of systems in the market. I think everybody knows intuitively they exist, but they don't put language around it. There's deterministic systems. That's a trading system, a settlement system. We trade it at $65. It has to settle at $65. We're going to book it at $65. All the instructions are deterministic. They're idempotent systems. If you put the same transaction in twice, you need to get the same answer.

[00:10:04] So those systems need pure linear scaling. We just have to make sure that the answers are always right and that the system is going to calculate things correctly all the time. And for the most part, that's just a pure, you get rewarded for good scaling architecture in a system like that. The other kind of system in this market is probabilistic. So these are systems like a risk system, value or risk, looking at probability distributions and outcomes.

[00:10:29] And those systems need to operate under stress that usually only happens when something goes terribly wrong. There's three keys to a system like that. One is it has to be multi-asset class. You've got to look at a picture of your risk across all of your customers. You've got to have all your customers' positions and asset types in order to effectively risk manage it.

[00:10:57] Secondly, it has to also be able to understand what your client looks like on a levered basis. And this concept comes back and bites Wall Street all the time. There will be a client who's a client of the prime brokerage business and the clearing business and the systems don't talk to each other. So to be able to look at the client exposure, not only market risk, but credit risk on a levered basis.

[00:11:18] And you've got to not only understand your own financing and margin with that customer, but you're also offering leverage on 40 or 50 derivatives exchanges globally. So that's a hard orchestration problem for a probabilistic system. But those are table stakes. We're table stakes then and are definitely table stakes now. And that remit has only gotten more clear and sharper over the last 15 years since the crisis because the windows for measuring that are shrinking.

[00:11:47] And I think we'll talk about that later when we get into tokenization and 24 by 7 trading. The margin period of risk completely evaporates, right? So you've got to understand where you are all the time. And you've got to do that on a multi-asset class basis. Well, you mentioned that you've brought together these two businesses, TS and Imagine, several years ago.

[00:12:07] This push towards AI and this LLM game has been years in the making, but it's really accelerated the last several years where they're accessible and they're being implemented.

[00:12:19] And so before we start really diving into some of your thought process on things like prediction markets, just big picture as someone running an organization that's so focused on data and systems, I'd just be curious if you could just try almost from a 30,000 feet to tell us what is new and different in terms of what you have your team focused on as a function of this just incredibly rapid development on the LLM side.

[00:12:49] How is today different than five years ago? We're in the middle of what I think is the largest invention of this century, which is the adoption of AI. We're not even at the middle, we're at the beginning, but it feels like the middle because it's gone so far so fast. Most capital market software companies really spent the last two years watching that first AI wave come in. And for the most part, Dean, they were left unaffected. The chatbots couldn't crack protocols.

[00:13:19] The LLMs couldn't navigate their way around multi-party workflows. But, you know, safe from the first wave is not the same thing as safe from the second wave. This is going to come in waves. And for the most part, this wave, I think, helped people prepare themselves for different development life cycles for software and how they deploy and features that they could add in.

[00:13:39] But this second wave, which will really run workflows for customers is a threat to not only software companies, but people that work in banks and IT. Because once the LLMs, which are excellent, by the way, do the work that they can do and max out, we're in for a second wave, which is reinforcement learning, which is really around displacing complex workflows with agentic coding.

[00:14:08] You talk a little bit in some of your pieces on just the way in which these advances, I think you just referenced it, are able to learn about processes over time. You're sort of training them to become integrated into a process that seems to so require human intervention that, as an outsider to the technology, it's almost incredible for me to even contemplate that a machine could do this stuff.

[00:14:36] Was there a switch that got turned that enabled this? Where was the step function? When did that occur? We started using LLMs pretty early. I think they were very capable even in their first iteration, but they weren't easy to use as a full stack. So Claude Sonnet 3.5, people didn't really care about it en masse across the industry. And when Claude Code came out, it changed everything.

[00:15:05] Because Claude Code actually just made the entire stack, in effect, transparent to the end user. So anybody could use Claude Code. You didn't have to wire different parts of the technology stack together to get an outcome. That was, I think, the big flashpoint. I mean, it was definitely used for research and reporting and anything that required language to be produced. There was just this steady adoption over time. But when Claude Code came out, I just felt like a switch turn.

[00:15:33] The velocity of change increased dramatically at that time. Well, I mentioned the thought leadership that you guys are making available on your website. And if I were to categorize them or choose an area of emphasis, it's on prediction markets. And I think you clearly have an interest there. I think what I'm seeing is you've got certainly an interest in terms of, as a data scientist and a technologist, in terms of integration. But also even just as a, almost an options guru.

[00:16:03] You're really writing about these prices in a way that is very native to the Alpha Exchange audience. So we can nerd out a little bit on some of this because it's clear that you're looking closely at these. What got you interested in prediction markets? And then I want to just talk about the raw probabilities implied by the prices and so forth. Tell us how you got to start focusing on this.

[00:16:29] So I love retail products, consumer products that make their way up to the food chain and disrupt enterprise products. I love anything like that. Crypto, DeFi did the same thing. Prediction markets are the latest consumer product to make their way up the food chain into enterprise. And I just think it's fascinating how people sort of try to ignore it. And then they try to ban it. And then they want to regulate it. And then they finally give up and say, fine, let's just use it. That's a fun cycle to watch.

[00:16:58] Today, I would say institutionally, they're kind of a curiosity being tomorrow. I think it's an asset type or asset class. The instruments themselves aren't new. People have been betting on uncertain outcomes for centuries in one form or another. But this latest wave, I think, had a really interesting start. So what happened in October 2024, three weeks before the presidential election, Kalshi won its lawsuit against the CFTC. The court ruled that event contracts on elections were legitimate derivatives.

[00:17:28] They weren't gambling, which sounds a little like a technicality. But it was actually a very big structural shift because it brought the whole asset class or asset type, whatever we want to call these things, under the umbrella of regulated finance. And the timing was immaculate because it was at the same time of the presidential election. So all of a sudden, everybody could bet on the election.

[00:17:51] And people in this country are tired of mainstream media narratives kind of hacking reality. They have their lived experience. And they're like, well, there's no way this guy's going to win. That guy's going to win. And I feel like prediction markets give somebody an outlet. So rather than canceling your subscription to the New York Times or Wall Street Journal, you can actually take a position, which is really interesting because those you could do before is yell and roll your eyes.

[00:18:18] From that sort of experience as retail users, that market lighting up, I think it just flashed right into capital markets. Well, that's GDP contracts and CPI prints and all these weather and anything else that might impact financial markets. And it hasn't cooled off since. And I don't think it will. And if you compare that with crypto's enterprise adoption, which was pretty uneven because the

[00:18:47] use cases were less obvious in crypto. In prediction markets, you have a very sharp institutional use case from day one. Direct macro hedging on variables that there's no options market for. There's no options contract for GDP. And the Fed also already validated the forecasting accuracy. So the adoption curve here, I think, steepens. And it's going to happen a lot faster in terms of institutional adoption that crypto did,

[00:19:16] because I think the institutional value proposition is concrete from the start. Most of this is underpinned actually by the Fed. I'm not sure you're aware, but they wrote a working paper in February of 26 titled, Kalshi and the Rise of Macro Markets. The conclusion is, and the punchline is, Calci implied forecasts match or beat Bloomberg and the FRBNY survey across Fed funds, as well as CPI and unemployment. So on Fed's funds specifically, though, it's super clean.

[00:19:45] Calci's mode matched realized rate by the day of every FOMC meeting since 2022. They have a perfect record. Bloomberg's consensus doesn't have that. And Fed funds futures don't have that. So when the Fed publishes a paper saying a prediction market forecast the policy, the rate policy better than the policy market itself, I don't think you can call that a fringe anymore. And I don't think it's a hot take. It's just a reality. There's something great about these.

[00:20:13] First, obviously, the binary nature is for anyone. It's so easy to get your arms around. In options markets, the probabilities are a little downstream. You have to make certain assumptions. These are raw and very specific. Obviously, there's some hair on these markets for sure. And hopefully they don't get just contaminated by too much of that. Will Beyonce mention Jay-Z in her Emmys acceptance speech? Or they call them mention bets?

[00:20:42] I'm remembering also in 2024, I mean, this was fascinating when I don't think Calci was as obvious back then. But certainly Polly Market was who will be the Democratic nominee. Now, Biden had his horrendous debate. And what was so fascinating was the way in which Biden and Harris were obviously inverses of each other. And that went back and forth. But what you noticed was when Harris rallied that she would be the Dem nominee, the chances of Trump winning went down.

[00:21:12] And I think that's all the information Nancy Pelosi needed. That, hey, the market's telling us something here. Super interesting. So I mentioned that these are binaries. And options have got obviously a date, certain expiration, but we don't really know what the price is. It's either in the money or out of the money, but we don't really know by how much. You've talked about the way in which prediction markets could be or would need to be integrated into risk systems.

[00:21:39] And that as we do stress testing, that we'd have to think about that differently. I was hoping you could walk through some of your thinking and writing on that unique aspects of prediction markets and these binary options. Here's where I see this going. And here's what my customers are interested in. Number one, direct macro hedging. Every book, whether it's credit or equities or rates, has, for example, GDP sensitivity, right?

[00:22:08] People hedge with S&P puts and hope correlation shows up when they need it. Prediction markets let you hedge the outcome itself, whether it's GDP or CPI or unemployment directly. And the Fed was pretty explicit on this. For these variables, the prediction markets are really the only place to get market-based probability distributions for that specific event. As I said earlier, there's no options market for GDP. The second thing that people like is event risk quantification.

[00:22:37] Traditional risk models treat elections or FOMC decisions or regulatory moves as basically shocks, right? Stuff happens. And prediction markets price those probabilities in real time, tick by tick, as the markets take in more information. And look, the Fed paper again, I mean, it was really good. They showed prices move with statistical significance across the distribution, with mean, variance, skew, kurtosis.

[00:23:03] So if you're a risk manager and you're not using that signal, you're not being conservative, in my opinion. Just ignoring part of the risk is actually measurable. So I think a lot of people don't opt into things because they think, I'm going to be safe, but I kind of view it the opposite way. If it's there and you're not using it, I'm not sure that's the safe path. The third piece is something near and dear to your heart, Dean, uncorrelated alpha.

[00:23:28] If GDP market is at three spot two and the rates curve is applying two spot eight, that's not noise, that's disagreement. And obviously, disagreement is where trades come from. That's what my clients are looking for. However, the challenge we have in operationalizing this stuff is pretty significant. There's a stack of things that need to happen. So at the base level, and this is the boring part, which is where I excel, of course, is

[00:23:54] the boring plumbing, is putting together a data ontology that links the event contract, the contract ID, to a normal security master. What asset class is it? What geography is it? What index is it in? Or indexes is it? What ETFs is it in? What's the sector? When you go to Calci and Polymarket and you pull it up, it looks like eBay, right? It's just a really long list of contracts.

[00:24:20] So how do I give it meaning in context in the shape of a risk management system or a portfolio management system? So you've got to build that data ontology to build the linkages between the contracts, the securities. And then from there, you can start to look at risk factors and probabilities, and then ultimately link all that to a yes or no outcome and decide how much weight you want to put on that contract. So the security master mapping is the first piece.

[00:24:49] We've done that, by the way. And if I could just jump in, is that done specific to your company or is that a utility for the industry? So that's specific to our company. We had to go out and meet the people at the prediction markets, get the data feeds, get some APIs in place to pull the data on a frequent basis, try to get the historical data, et cetera. So you pull all that together.

[00:25:14] And then problem two is the math part, fitting a distribution to a price or a band of prices tick by tick across hundreds or maybe thousands of contracts. So we've solved that part of the framework. When you get into running it, I think it's going to get a little bit interesting, right? So doing it in production with sensible air handling for moments when prices go non-monotonic or if the fit falls apart, that's a different sport. The industry is not there yet.

[00:25:41] So we've got the data problem and the UI intuitive problem solved. And now we're into the, OK, so how do we run this? And that's what we need to ship, right? We need to ship something that not only pulls the data together, but can manage some of these outlying events and exceptions handling. And look, we're doing this while most of the industry is still deciding whether to take these markets seriously. So we feel like we're a little bit ahead of this, but that's where it is today.

[00:26:09] There's also a lot of really positive signals, I must say, from the regulators. So I was at the BOCA conference, the FIA conference in early March, and CFTC chairman Selig spoke about prediction markets. Prediction markets actually abducted that whole conference, by the way. But he spoke about them and he called them truth machines. And I thought that was a really cool way to articulate it. And I thought at the time, OK, I'll take that. You know, if these are indeed truth machines, what I need to build is a lie detector.

[00:26:38] Sometimes these markets are wrong or they're noisy or they're too thin. Our users need to understand, if I want to operationalize this stuff, how do I put rigor around it to separate the noise from the signals? Are clients of yours, institutional investors who traffic in derivatives, are they mapping a Fed easing yes or no contract versus what's implied through the three-month, two-year yield curve, you know, that sort of thing?

[00:27:08] Is that happening at this point? I would say we're in the period of experimentation there. They're calibrating how much to weight those events and whether those events are indeed leading indicators, which in a lot of cases they are, or whether a lot of that price is already in their portfolio. It's already priced. And that's what we're working with them on. And that's really a good sidebar conversation around AI.

[00:27:35] So we're doing all this with this great once-in-a-century invention that's come along. And as I talked earlier about deterministic systems and probabilistic systems, on the probability side, we are using AI agents to calibrate the relevance of the contracts, the contracts that we pull into the platform. We use AI to assemble the correlation matrices so people can focus on the contracts that are most relevant to their portfolio.

[00:28:03] And then if you take something like the straight-of-hormuz contract and you looked at the benchmark crude price, maybe it had a negative weighting in the first month or second month, 20% versus the benchmark, and maybe 10% six months out or a year out. You might not like the way we put backwardation together in that risk factor. Our clients can then go in and train the agent and say, hey, we love the front month, but can

[00:28:33] you lighten up a little bit further out on duration? AI is being used to actually help the clients train our system on how to calibrate those factors. And then each client can manage those factors the way they want to. So it's kind of fascinating to take two innovations and put them together and have this great experiment. But the productivity is wild. On a 90-day sprint to get a production product out, we're about day 45 or day 50, and we're going to get a product out in 90 days. Nice.

[00:29:02] Well, it's interesting. We have yield curves. We've got vol curves. They're upward sloping, downward sloping. And as you were referencing with the Strait of Hormuz, you got a time curve. That's unique. I remember back the last time the Democrats and Republicans had a shutdown, and I think this was maybe November or so, and you had a Kalshi-developed time curve on how long the shutdown would last. And if I'm a politician, that's just really interesting information for me.

[00:29:31] And that's a unique thing, that options markets really can't do that specifically. Not like that. I think that was interesting. So you mentioned being at that conference. You talked about retail sometimes being the source of this, the starting point for crypto. If we go to Bitcoin, we saw that the IBIT ETF really became a thing. And it's just easier in a lot of ways for the non-native crypto folks to just say, look,

[00:29:59] I want exposure to this widget called Bitcoin, but I don't need cold storage. I want this thing to show up in my Fidelity account. I want to do some options on it. It's just easier. And with prediction markets, it was supposed to come out this week, but Roundhill has got a yes or no contract on, I think it's on the election, the 2028 election. And I'm just wondering if you think, one, is that just a sign that these things are absolutely

[00:30:25] coming to markets that the SEC is blessing the launch of an ETF? I do. First of all, if you just look at the structural differences between something like Bitcoin or a synthetic product on Bitcoin and prediction markets, these products are futures. They're regulated by the CFTC. And they're still fighting that out with the states a little bit, but the path is to be regulated by the CFTC. There are exchanges that are licensed like Calci and others, or Rothera, which I think starts up in Q2. Licensed to do business.

[00:30:55] They have FCM members. They are DCMs and DCOs. They're as regulated as anybody, I think. So you don't have any of that regulatory blockage that you had in crypto. They're derivatives contracts. So people, for the most part, know how to gear them in their systems. There's going to be some work around settlement and margin and things like that. But those are problems that can all be solved over a period of time. So again, you don't have those types of blockers. And you don't have a custody issue, which was also a big blocker. I think this just happens much faster than crypto.

[00:31:25] And I think if your clock is sort of set to where crypto took 10 years, I think you're going to get blown away because this will absorb faster into the market because you actually don't even need to trade it. You could just use the data. The data is valuable on its own. Operationalizing that data is not as hard as owning the contract and trying to figure out how to clear it and settle it and explain it. You just use the data. So there's two things. And you talk about this in some of your writing, which is price discovery is valuable.

[00:31:54] You want to know where things live, where things clear in terms of prices. But trading them, that takes extra. The market's got to get much deeper for that. And obviously, in some ways, you're presupposing that that happens. I'm wondering where you think we are. There's always this great chicken and egg with folks that are willing to take risk and people that want to take risk. Hedging and speculation, that two-sided price discovery that makes a market deep. We're obviously not there yet.

[00:32:23] What do you think needs to happen? And over what timeline would you be hopeful that it can occur? To the beginning of your point, people love price discovery and price formation. I mean, they can decide how they want to use it or discount it or not. But the fact that it's there is amazing. So I think what happens is we'll end up with a bimodal market where you have this retail market that's feeding prices in. And then you have this institutional block market because you can't go in and buy a $10

[00:32:50] million yes bet on something or a $20 million yes bet. But you can be informed on that price. And that could be through just an RFQ mechanism or a structured product. And again, we have electronified RFQ mechanisms in this market. There's all kinds of markets for that. Everyone has the tech. That can happen really quickly. And I think that bimodal market, similar to the muni bond market where you have the muni center with dentists from Connecticut buying municipal bonds, and then you have institutions buying blocks.

[00:33:19] So I think you'll have a nice informed market there. I think the real blocker, once you get past what we talked about earlier, which is data normalization, risk fitting, getting the data into the systems, the biggest problem, this is the rate limiting step for the whole asset class. Today's contracts are fully collateralized, which is safe, but it's economically kind of dead for a market maker. So a market maker who'd earn 24% on capital under normal margin earns 2% on full collateral.

[00:33:47] If you multiply that across institutional flow, you sort of understand why the money is still sitting out. The math just doesn't pencil for them. I think the fix is really risk-based margin, same as every other derivative. Look, it sounds simple, but it's not. The margin engines at prime brokers and CCP's were built on 50 years of assumptions that really don't hold for a lot of these contracts. Long and short aren't mirror images, bell curves, kind of make sure of that.

[00:34:17] Margin has to handle asymmetric risk, nested events. Contracts don't drift to settlement, they jump. A contract could settle at any time, two in the morning on a Sunday. It's a multi-year project for the industry. There's a lot of people focused on it. Until it's done, I think institutional sizing stays pretty small because nobody puts size into a market where the capital math is structurally insufficient. There's a path to get there. The exchanges obviously want to get there. The market makers want to get there.

[00:34:45] And I think they'll get it right over the next 24 months in stages. So things will progress incrementally. Until it does, though, prediction markets from an activity will stay retail with institutions kind of waiting for that leverage capability. And do you think of that retail presence as transacting on the CalShe website or the Polymarket site or app? Or do you see it becoming a drop down on my Fidelity account as well?

[00:35:15] It will be absorbed into retail brokerage. I mean, if you look at Robinhood, I mean, I think they're the central feeder into markets today. 25 million-ish retail users that love those products. They want to bet tech stocks because they feel like they know something about it. They don't just want to bet the stock price, they want to bet earnings, and they want to bet market cap, and they want to bet on the stocks that aren't public yet because they really like those stocks. So Robinhood is the feeder for that, for the industry.

[00:35:44] And obviously, everyone else will have to do it even if they don't want to as a competitive necessity. Obviously, there's some hair on them. We know about the Maduro raid and the, quote, insider knowledge trading that occurred in advance of that. And there have been other issues. Some of them are in sports, but others in some of these areas of prediction markets where privileged information is clearly incredibly valuable. How do you think the industry can move forward as someone who wants it to thrive?

[00:36:14] What would you like to see from a regulatory standpoint? How do we solve for that risk and allow these markets to grow? We'll have to do something. I think the first things are pretty easy and obvious. These mentioned contracts just have to go completely. We'll have to come up with different sectors and, I think, different contract types for those sectors, whether it's elections, whether it's other geopolitical events, economic events, et cetera.

[00:36:41] So we get some normalized way of looking at things, and things can be bucketed appropriately. In terms of the insider trading, yeah, we'll need rules and regulations and some surveillance around these markets to make sure that they aren't gamed. I'm fine with good information flow and people with good information being in the market. I think that's one of the foundational blocks of the commodities market.

[00:37:06] That was their purpose in life was to help farmers who understood wheat yields and corn harvest had information. I think that's okay. I just think some of these more manipulative episodes will require some surveillance. But look, we've solved surveillance issues, I think, pretty well in the equity markets. And I feel like we can also get there with the same types of tools and technologies in these markets as well. Well, one of the areas of just real interesting price action in markets has been in software.

[00:37:35] And just the meltdown in some of these SaaS stocks, much of it do a memo, a Citrini memo that purports to be written in 2028, looking back on 2027, but sure got people thinking about a world in which these companies were really vulnerable. The moat just didn't exist. And you've done some writing on this as well. You kind of distinguish it along a spectrum of vulnerability.

[00:38:04] Perhaps the way I'll start is you talk about the AI solving for IQ in some ways, and that being kind of the first layer of just substitution for what humans would do on reporting. Walk us through your thinking there and then where the moat can persist, where you can be safe in this world. I'm going to get a little bit technical on the structure of these companies because I think you have to get into the details to kind of appreciate the subtleties here.

[00:38:32] LLMs are probabilistic systems, and they produce language, which is sort of inherently uncertain. As I mentioned earlier, it's hard to take a probabilistic system and put it into a deterministic execution because you might ask it twice and get two different answers to the same question. If you look at the software industry in SaaS, there are horizontal SaaS companies and vertical SaaS companies. Horizontal SaaS companies in legal tech and HR tech and health tech.

[00:38:59] A lot of those companies are really producing language. They can produce the language and the outcomes using an LLM. And those companies, I think, have to do things like right now. I mean, I think most of them were doing things a couple of years ago. It's not like they were just sitting around waiting for somebody to write a paper and complain about it. They've been working on it for a while. They're in the first wave of vulnerability here because of the nature of their products and their output. Their UI is their product in a lot of cases.

[00:39:28] Those are the guys that are in trouble in this first wave. In our world, in capital market software, most of the products are vertical vended products, meaning they do everything in an asset class or a group of asset classes. And they're very workflow heavy. There's not a really great way for an LLM to learn workflow because workflow is layers of compliance

[00:39:54] and exceptions and risk and all kinds of idiosyncrasies that would make it very difficult for an LLM to reproduce it. Plus, LLMs aren't stateful. They try something, it doesn't work. They try it again and it just stops. They can't do anything. There's another kind of model called reinforcement learning. And reinforcement learning is a deterministic system predator.

[00:40:22] It doesn't try to describe workflow. It tries to learn to run it. So you deploy an agent into an environment, it'll fail a thousand times. It gets a signal, it adjusts, it eventually learns what the system will accept. Not because it read the manual, not because it's good at telling you what the system does, because it tried and failed and tried again until it worked. It's not outputting language at all. It's not working on a deterministic pipe. It's becoming the pipe.

[00:40:51] Its job is to become the workflow. That architectural reason is why this second wave that's coming probably in the next 18 to 24 months is more of a threat to the way most of the software works in our space today than LLMs were. LLMs hit our space and bounced off. These will be much more formidable. So it's either figure out how to do it within your own framework or someone else will do it for you. You mentioned the exceptions report.

[00:41:19] Something just along the chain of linkages that needed to happen to make a trade clear, something was out of sorts. And I think what you're saying is in the previous versions, that's it. You're stuck. And what you're saying, and I'd love for you just to describe it a little bit more by way of example, so I think the listener can get his or her arms around it a little bit better.

[00:41:43] Give us an example of the way in which these new LLMs with reinforcement learning can solve for the exception. Let's just lay out what a GPT class LLM is and then talk about reinforcement learning and the difference. A GPT class LLM is trained on language, billions of documents, conversations, code, research papers. The training signal is really does the token match what the human would write, right? That's their positive signal.

[00:42:10] The model gets very good at producing very plausible, very coherent, contextually appropriate text. And that's really useful, and those things are amazing, and they've gotten amazing very quickly. But the feedback loop it was trained on is linguistic. Did the sentence make sense? Did the answer sound right? An RL model is trained on outcomes in an environment. There's no text, Dean. There's no language signal at all. There's just a task and a state space.

[00:42:37] There's an action the agent takes, and there's a reward. Did the task get closer to complete or further away? The model has no idea what anything means. It only knows what works. If you press button A and state X produces reward and pressing button B produces penalty, the model learns to press A. And it just keeps going in all these circuits until it completes all the tasks that the system will expect. And it'll go tens or hundreds of thousands of times until it's reliable.

[00:43:07] And that's it. So the reason this matters in capital markets is that workflows kind of don't care about meaning, right? The fixed protocol doesn't care whether the agent understands what a fill commitment is. It just cares that the message was formed correctly. The CCP margin system doesn't care whether the agent grasps the theory of collateral substitution. It just cares whether the right fields are populated in the right sequence in an authenticated session. That's an outcome problem, Dean.

[00:43:35] It's not a language problem. A chatbot trying to navigate that workflow. It's kind of like hiring someone to read every book ever written about driving but has never sat in a car. They can describe every maneuver perfectly. They just can't drive. A reinforcement learning agent is the opposite. It has never read anything. It has no idea what a car is conceptually.

[00:44:01] But it has sat at a simulator for six months, crashed tens of thousands of times, and now it doesn't crash anymore. In capital markets execution, not crashing is the job. So these worked examples, I would say, a good one since we talked about collateral a little bit. You have a buy-side firm needs to move collateral in a clearinghouse account. And an LLM can reason through that correctly, like what needs to happen. And it knows what has to happen next.

[00:44:29] But when it hits the actual system, let's say there's a message format that's incorrect. It gets rejected. It tries again, but it has to hold live connections across four different systems to do that. It can't do that. LLMs are stateless by design. They just stop. The reinforcement learning agent, it'll fail the same way on day one, but it'll keep going thousands of times, each one a little bit better than the last. And it doesn't learn from reading, it learns from failing.

[00:44:57] So eventually, it's running the full workflow in four minutes, and it used to take a person 30 minutes. That's the difference between those two approaches. And that's why reinforcement learning is so interesting to me as an ingredient into our space. It's going to really change, I think, the dynamics of software in all verticals, including ours. So with respect to your business specifically and the value prop that you're putting forth out to clients,

[00:45:23] talk a little bit more about how reinforcement learning is enhancing what you're able to do, maybe be specific, and then the mode. How does it reinforce your capacity to keep those clients? The way I think about the mode and sort of our position and other companies' position is there's four types of capital markets fintech companies. There's these really long, decade-long roll-ups where somebody's acquired 150 or 200 systems.

[00:45:51] And they're all workflow systems, and they're all vertical. They do something really good, but they have a bunch of them. They can't defend everything. RL doesn't even need to attack a complicated platform. It's just going to attack one module at a time. And if you have 150 systems, the question isn't which one do you invest in? It's which ones do you abandon? Because you can't fight a war on 150 fronts, right? You'll just lose everything. So I think those firms have some difficult decisions to make moving forward.

[00:46:18] You have the exchange diversifiers like stock exchanges that bought tons of fintech over the last 10 or 15 years, billions and billions of dollars. While this is happening to some of that fintech that they have, their data businesses are also being commoditized through other AI and energetic AI products. These are fundamentally financial engineering cultures and finance cultures at those firms, not software engineering cultures.

[00:46:44] So I think they're going to have to change culturally how they approach these platforms. They bought them and are good stewards of them. I'm not sure they're good rebuilders of them. I think that's an interesting market to look at. And then there's the mid-market guys like us. We've got integrated platforms across pretty complicated multi-asset class workflows. We own the training environment. Years of sort of cross-asset workflows, risk override decisions, exception patterns, you know, across different systems that are unified.

[00:47:13] Generally really hard stuff to replicate from the outside. That is probably the safest place to be because the RL models are going to have to learn on systems like ours in order to produce. So obviously we would be consumers of those agentic products and would make our products more efficient, cheaper to run, cheaper to own, and cheaper to deliver and faster to deliver by using those agents where we have people doing things today. We'll have more agentic AI doing things in the future. Most of us are PE backed.

[00:47:44] PE can go either way, Dean. They could fund the infrastructure investment and compound the advantage, or they could harvest the margin before exit. So you have the same ownership type, but two radically different outcomes, right? It just depends on what the thesis is and the holding power of that asset. And then you have point solution vendors. There's a lot of these, hundreds of these in the market, five to $50 million companies, collateral optimization tools,

[00:48:08] reconciliation modules, reporting engines, bounded, well-defined, measurable outcomes. And that's an ideal training environment for reinforcement learning. These companies built their business doing one thing extremely well, and RL was built to learn exactly that, right? So I think those guys are probably the most vulnerable. So that's kind of how I break down the sectors in our space. Everybody has a challenge here. Some of them are offensive, some of them are defensive, and some of them are both.

[00:48:34] Well, we've been talking about these incredible advances in technology, what they mean for business. Another part of the conversation is around just the way in which trades are settled, the speed with which they're settled. Folks like you use terms like atomic settlement, which I love that term. I think tokenization becomes a big part of that. Some people, when they think about tokenization, and maybe I'm one of them, think about non-fungible tokens. That's not what you mean, obviously.

[00:49:03] Just give us the big picture of what exactly is this new thing called tokenization. And then we'll just dive into some of the implications for things like overnight repo and so forth. Digital ledgers have been marinating in traditional finance for well over a decade. This all came out with the advent of Bitcoin and blockchains. And those ledgers have absorbed key lifecycle events of traditional security.

[00:49:30] So when I think about securities, I think about six things. How are things issued? How are they traded? How are they settled? How do you custody it? How do you finance it? And what happens at maturity? Entrepreneurs in digital ledgers and tokenization have figured out how to create all of that in a digital ledger. We spent decades building this big complex market.

[00:49:55] Each asset class had its own way of running things and doing things and messages and workflow. There's just been billions of dollars spent. And they've learned how to do all of that. And then it was just a matter of how do we actually now bring that into traditional finance? How do we take this great immutable ledger and bring it into a normal finance structure? Like we talked about earlier, normal finance sort of hated it. Then they tried to ban it and they tried to regulate it.

[00:50:23] And now traditional finance, some of the individual people who were the biggest critics of it are now embracing it. This is going to happen. The firms that built the infrastructure for tokenization, first of all, they're all super smart. They're run by smart people. Everybody in those companies is really smart. They're well financed. From a regulatory perspective, they're now sound. They're on very firm footing with the Genius Act. And they have a decade of experiential learning. And they're really primed to disintermediate.

[00:50:51] All of this is hitting and this embrace from traditional finance is hitting as firms are adopting agentic workflows. I don't believe this industry is going to look anything like it does today in five years. Between tokenization and AI, I mean, it's just a completely different world, which I'm really excited about. What sort of annoys me about the tokenization conversation is a lot of people, they limit it down to a conversation about efficiency.

[00:51:20] Faster settlement, cheaper plumbing, knock some basis points of operations, which I think is conveniently comfortable framing. And I think that's the problem with it. Worst cases, they call it a feature. And look, tokenization is not a marginal upgrade. It is a structural reconstruction of the economics of intermediated finance. I mean, it's going to change all of it. And there's kind of three things, your favorite word coming up. There's three things to anchor that on.

[00:51:49] Atomic settlement collapses intermediation. Collateral becomes mobile. Infrastructure economics concentrate. Any one of those things would require people to kind of act urgently. But together, they kind of create this world where firms that build the rails will own everything that flows through them. And the firms that wait are going to pay rent for someone else's infrastructure for the next 20 years. It's a really interesting time between now and call it 2028.

[00:52:17] I think the trap with tokenization also is it doesn't arrive in a dramatic moment. It comes in quietly. Settlement rails, collateral systems, money market products. I mean, people kind of read it, but they don't really focus on it. And it comes in one segment at a time. And then by the time the shift is obvious, the economics of the old system are already broken. It's super clever because there's links in this chain of events that have to happen. I broke it down into five key links. And we're already on link three.

[00:52:47] Acting urgently or at least getting educated about it is really important right now. The first link in that chain is tokenized money substitutes will kind of drain traditional deposits. So you have all kinds of examples. JP Morgan's on-chain fund, Wisdom Tree's tokenized T-bill, BlackRock's money market fund called BUIDL,

[00:53:11] same credit quality as bank deposits, better yield, programmable, 24-7 liquidity. Wisdom Tree went from 30 million to short of a billion in well under 12 months. That's not a pilot, right? That's a migration in process. So I think the threats to banks isn't these central bank digital currencies. It's actually people in the competitive orbit who are moving first. They're the ones that are going to capture the wallet.

[00:53:40] The second link in the chain is stablecoin rails bypass kind of the traditional correspondent banking structure we have. And this isn't crypto, right? These are programmable dollar rails. Again, instant, global, 24-7, almost zero marginal cost. And the Genius Act gave us a regulatory framework for all of that. And you even have insurance companies like Aon settling insurance premiums on stablecoin rails with Paxos and Coinbase. That was over a year ago.

[00:54:09] So NASDAQ partnered with Kraken and Back for tokenized equity settlement. Correspondent banking network that we have today, that kind of hidden backbone that nobody really sees the global dollar flow over. It's getting disintermediated not by a competitor, but by a cheaper layer of infrastructure. The third thing to think about is credit origination and that moving on-chain. So on-chain mortgage origination is already live and it's at scale. So there's a company called Figure, for example.

[00:54:38] I think they've got close to $20 billion in HELOC. They've gone into that market and then mortgages will be tomorrow. So the origination to securitization chain, historically an extremely profitable part of credit, it's going to collapse into a single atomic workflow run by a smart contract. So I think we're probably 18 months to 24 months out from leveraged loans moving in the same direction.

[00:55:02] Those three links are already operational and trucking billions of dollars across those rails. The things that are still kind of in the works and happening, but happening really quickly, kind of the fourth link in the chain, which is continuous settlement removing the financing windows. So when WisdomTree got SEC approval for 24-7 trading with instant settlement, that really wasn't a new product.

[00:55:28] It was a demonstration of what you could do with settlement infrastructure and kind of old versus new. So a prime broker whose revenue depends on T plus one or partially depends on T plus one and T plus two settlement, they aren't really facing a competitive displacement. It's a structural displacement. The overnight financing window, repo, stock loan, SEC financing, it doesn't compress by T zero, it actually just gets eliminated. So there are things here that just go away when you get to that type of settlement.

[00:55:57] And then the fifth link in the chain is really atomic settlement removing the CCP risk premium. So CCPs exist to mitigate counterparty risk between trade and settlement day. So atomic delivery versus payment eliminates the risk the CCP has paid to absorb. So their mandate doesn't disappear overnight, but the economic justification sort of erodes with every basis point that migrates to the atomic rail. So these are structural changes that are happening right now.

[00:56:24] Very well capitalized companies join now by TradFi to drive more asset classes through them. I'm an individual with a simple portfolio. I've got 60% stocks, 35% bonds and 5% cash. In two to three years, what's the difference to the end user? Is this just, they don't see any of this? They're now in this tokenized world, but they just don't really even notice it.

[00:56:52] What's the transition from that person who's got a Fidelity account, traditional stuff to move into this world? So there's two options here that are coming because you have the cash markets also changing. The NSCC is going live this quarter on a 24 by 5 equity settlement. So the margin period of risk is no longer 4 p.m. to 9.30 in the morning. It's Sunday night to Friday night.

[00:57:20] So as a competitive response, they're like, well, we'll just change our infrastructure. You don't have to worry about these instruments, all these crazy new things. You can just trade your equities Monday to Friday. 24-7 is also happening at a bunch of traditional markets as well. There's an infrastructure move, mostly in the U.S. And then there's an instrument move, which is, well, why would you want to buy that T plus one cash product when you can have this great digital product that's the same thing and it's instantaneous settlement?

[00:57:49] So you have options. You'll have this dual running option of do you want the instantaneous settlement or do you want the slow one? I think most people are like, oh, I'll take the fast one, Dean. And that's just kind of an example on the equity market. You have more flexible and faster moving collateral on money markets, easier to exchange and trade things. There's less friction. Custody goes to zero.

[00:58:13] So I think from a retail perspective, the price that they pay for the friction in financial markets gets reduced almost to nothing. Well, you talked about the central counterparty clearing mechanisms and the sort of risks that they're taking, that time slice, so to speak. And if you conceive of this world in which everything's always instantaneously clearing,

[00:58:36] in some ways that almost invites a 24-5, maybe 24-7 ultimately mark-to-market environment as well. And this is where I wanted to get you thinking. You know, I go back to 2024. There was this VIX event in August. The VIX just went bananas. There was an overnight sell-off in JGBs, I think. The yen rallied a lot. The Nikkei went down 10%. The VIX went from 25 to 50. It was kind of a flash crash type of thing.

[00:59:06] It reverted after a couple of days. But the market kind of broke. And it started because of a very illiquid overnight session halfway around the world. And so my question is, is there some risk to marking-to-market continuously with the knowledge that the reality is liquidity is probably going to dominate from 9.30 to 4 still? Are we vulnerable to these overnight sessions and asking too much of them, I guess is my question. I think so.

[00:59:37] A lot of people talk about 24-7 in a crowd and they're super interested in it, but nobody wants to come to work on Sunday morning. We haven't figured out the economics of like, why don't we have two more days in the week, right? So that's not great. You also have, as you described, these twilight zone periods where there's no liquidity and you have these price shocks. That happens all the time in FX and other markets. I think that is a problem. You know, there's going to be a night tape and some other things, mechanisms that people have put in place to kind of buffer against that.

[01:00:03] But for the short term, I think you end up with a market where people are going to want to trade in the most liquid moments and they're going to want to do things in normal business hours. But eventually that erodes because I think people want to trade these event contracts and they want to trade stuff in Hong Kong that they hear about. So I think over time, you'll end up with gradually more and more order flow. Again, probably led by retail that decides when they want to do something and what they want to do. It's gradual.

[01:00:31] It's not going to be a big bang and all of a sudden we're up at two in the morning trying to figure out what we do on the trading desk. It's a gradual move in that direction. I was at a conference a year ago and they were trying to shrink trading hours. So you have a lot of people who think, oh, we should only be open four hours a day. I see both things happening. In Europe, most of the talk is really not about infrastructure. It's about instruments. So there's a Dutch business company that just got approval called One Trading. They're trading perpetual equities on a 24-7 basis.

[01:00:59] So there's a lot to be worked out here, definitely. Well, Rob, we covered a lot of ground. Found the conversation fascinating. Really glad that we had an opportunity to do this. Thank you for being a guest. Appreciate it. You've been listening to the Alpha Exchange. If you've enjoyed the show, please do tell a friend. And before we leave, I wanted to invite you to drop us some feedback. As we aim to utilize these conversations to contribute to the investment community's understanding of risk,

[01:01:28] your input is valuable and provides direction on where we should focus. Please email us at feedback at alphaexchangepodcast.com. Thanks again and catch you next time.