We continue this series with a new episode of This Week in Intelligent Investing, the podcast co-hosted by Elliot Turner and John Mihaljevic.
In this conversation, Shaun Heelan shares key insights forged in the high-stakes world of structured credit and translated into a rigorous, concentrated approach to equity investing. As the Chief Investment Officer at Munich-based MAAT Investment Group, Shaun brings a unique and deeply analytical perspective to modern value investing.
Shaun’s career path is anything but conventional. Starting out at Goldman Sachs during the European dot-com and telecom implosion, and later running highly profitable correlation books at another leading investment bank during the lead-up to the Great Financial Crisis, he learned early on that downside protection must always be an investor’s primary focus. Today, he applies those hard-won lessons in valuing European small and mid-cap equities. He approaches public market investing with the mindset of a private equity owner, mapping out assumptions, scrutinizing accounting realities, and actively hunting for asymmetric payoffs.
Beyond his differentiated portfolio construction methods, Shaun offers a grounded, compelling counter-narrative to the recent market euphoria surrounding artificial intelligence. Drawing on his academic background in high-performance computing, he expertly demystifies Large Language Models, describing them as “smart librarians, not scientists.” For investment professionals navigating an increasingly top-heavy market, Shaun provides a fascinating framework for understanding why AI will likely reinforce, rather than destroy, the intangible moats of established network-driven businesses.
To bridge theory and practice, the conversation dives deep into real-world applications. Shaun walks through his forensic diligence process by breaking down why his firm passed on Vistry Group (UK: VTY), a homebuilder frequently touted as the “NVR of the UK,” citing aggressive margin assumptions and hidden balance sheet liabilities. He then pivots to a compelling long thesis on Edenred (France: EDEN), demonstrating how regulatory fears and headline noise have created a deeply mispriced opportunity in an asset-light, moat-protected compounder.
What You’ll Learn
How crucial risk management lessons from the 2008 mortgage-backed securities crash apply to underwriting modern public equities.
Why Value-at-Risk (VAR) models create institutional blind spots as well as opportunities for long-term, structurally advantaged investors.
The mechanics of sizing highly concentrated positions and managing path risk in an era of surging single-stock volatility.
Why LLMs function as “smart librarians” and the fundamental danger of conflating computational pattern-matching with true business wisdom.
How AI is strategically poised to strengthen the barriers to entry for businesses possessing non-technological, intangible moats.
A forensic teardown of Vistry Group, highlighting the red flags of aggressive partnership accounting and overlooked land creditor debt.
The case for Edenred, exploring why the market is systematically mispricing its regulatory risks and underappreciating its network float.
Enjoy the conversation!
Transcript
The following transcript has been edited for space and clarity.
John Mihaljevic: I’m joined by my co-host, Elliot Turner of RGA Investment Advisors, and our special guest, Shaun Heelan, Chief Investment Officer at MAAT Investment Group, based in Munich.
We have a lot to cover today. We’re going to talk about Shaun’s investment philosophy, his and his firm’s views on the complacency in the market and some concentrated risks. We’ll also have the AI debate, where Shaun has some great insights. Shaun, let’s start with your background and the story behind your firm. Tell us about your path in investing, some lessons you picked up along the way, and how the core team came about at MAAT.
Shaun Heelan: My path in investing has been a strange one—a long and winding road, led by curiosity, much like many people in this field. I started 24 years ago at Goldman Sachs in London. I’m Irish, born and raised there. I was recruited after completing a master’s degree in high-performance computing, which gave me some expertise and ability to speak about the current AI debate.
I went to work at Goldman as an intern in the summer of 2002. As you may recall, that was right in the middle of the collapse of the dot-com boom. What’s lesser known, especially for American audiences, is that there was a huge dot-com boom in Europe as well, particularly in the telecom sector. There was an underdeveloped high-yield credit market trying to be developed similar to what you’d seen in the US in the late seventies through early nineties. That market imploded.
I was there just as it really started to unfold in 2002. I was very fortunate to get an offer to come back and start early in December because I was finishing my thesis. Goldman had two floors in London with a bridge between them at Peterborough Court and Fleet Street. Half the trading floor had been fired—it was like a neutron bomb. That really impacted me.
The mood had changed obviously, and I think it’s informed my career ever since. I came with a natural sense of conservatism, but it created a heightened sense of risk awareness of what can go wrong in the markets and how quickly moods can turn. It also taught me one crucial lesson: the downside is the thing you need to consider first and foremost in investing. You can have prolonged periods where things look very easy, where conservatism seems like a terrible mistake. But when the correction comes or when sanity prevails, it’s better to have been conservative all along—much like the GMO message that’s been sent around for 20 years.
I also remember Buffett talking at the Sun Valley conference in 1999 and a book coming out shortly afterwards predicting that equity returns for the next decade would be in the low single digits. That had a huge influence on me. I was very lucky that as the high-yield market was imploding, Goldman in London was a very active merchant bank—something different from what you have in the US—almost like having a proprietary desk, but one that makes more structured, long-term investments. They were very famous for this.
I was put into a group that ended up doing two large rescue financings—one for Vivendi, which was a much bigger company at the time, as well as Universal Music Group and studios, and then Rhodia, a French chemical company. Because we took merchant risk, that deal ended up being probably the most profitable thing in fixed income at Goldman that year. I didn’t particularly enjoy it—I’d joined a trading desk but ended up in a merchant banking capital-raising role.
I was going to leave and head back to Ireland to potentially work at Google, which was opening up there, and I had a computer science background. But I was persuaded by friends and my father to visit the States because I’d never been to New York. I went over for training and was very fortunate. My bosses in London were quite fond of me and thought I’d done a good job. I’m a very blunt, honest person—probably too honest, if you ask my wife. I told everyone before I left that I was probably going to quit as soon as I got back. After a few weeks, they came to me and asked if I’d like to interview with some desks in New York.
I loved New York from the moment I landed. I interviewed with a few desks and was fortunate to get offers from pretty much all of them. I joined the mortgage-backed securities trading desk at Goldman. At the time, Goldman in non-agency mortgage-backed securities wasn’t a major player—maybe 12th or 13th out of 16 or 17. They were very active in the agency space, with Ginnie Mae, Fannie Mae, Freddie Mac, and CMOs, but not in the other space.
I’d seen CDS contracts in London, and they were quite active there, but the mortgage side in New York hadn’t really seen anything like that. There was an active market in total return swaps. The memory of the 1998 LTCM crash was still very present. That crash was right at the epicenter, and the long-term trades that killed them were mostly basis trades on swaps against various mortgage products.
I was part of a group that developed that market further and made ourselves more competitive. We had a cold-start disadvantage: we were cautious with our balance sheet and had a high equity proportion, but if we were competing with Barclays, Deutsche Bank, or money center banks, they’d charge their trading desk a financing cost more like what they’d get in the money markets—LIBOR minus. We had about an 80 basis point disadvantage on anything we held in inventory.
We proposed developing a new product: pay-as-you-go CDS. Something similar was happening more natively at the time, but we ended up getting a big lead and dominating that market at Goldman. I was part of the group that developed what became the standard contract for subprime and CMBS pay-as-you-go CDS. The first trade was actually bespoke—the famous basis trades done about a year before anything else. These became the standard for the ABX and CMBX indices.
The Goldman model that I helped develop with the quant group became the market standard model for those CDS contracts. As a result, I was negotiating with the dealer group as to which model should win. We won. Everyone thought I was much more senior than I was—people thought I was an MD or partner because of my deep voice and Irish accent. I started getting offers from every bank. The best one was from Countrywide in California, which was tempting, but I could see there might be problems down the road. I didn’t take that step.
I went to Merrill Lynch, where I ran their correlation business. After six months of originally trading single-name credit default swaps and mortgage-backed securities, they handed me a business they’d had for two years in which they hadn’t done a single trade. We had very low budgets. In the first year, I was asked to make five million in PnL, which sounds small now but was meaningful then. We made 45 million, and the next year we made over a billion, and every year thereafter until I left, we made over a billion with a team of around three or four people.
I’d say I won the lottery multiple times over working at Goldman and Merrill with some of the best risk takers you can find. It was an amazing training ground. I learned everything from the ground up—working from 6:30 a.m. and going past midnight.
At Merrill, we had a great distribution force and brilliant people all around. It just wasn’t as well-organized or cohesive as Goldman at the time. Because we were so profitable, I got to see a lot. If they had problems in Japan or London, often very big problems, we could figure things out quickly. We had built good analytics that could be deployed across the firm. They’d parachute me or my team in to deal with that. I got to see a lot of what happened within Merrill leading up to the financial crisis and interacted with very senior management, right up to the CEO.
I learned a lot about how people behave under extreme stress, about managing risk. When dealing with mortgage-backed securities, those principles apply to equities. I’ve always been a fan of Warren Buffett. He had that quote that when you’re looking at an equity, it’s just a bond with undefined coupons that will come for an indeterminate amount of time.
The perfect training for me was doing that in structured products—MBS, CDOs, or anything else. The beauty is you go from a spectrum of ultra-safe securities with no uncertainty around timing at the front of the capital structure—AAA rated—all the way down to things that are incredibly leveraged at the bottom. You’re playing with a variety of assumptions and get to see what the results are every month. You can compare what you predicted versus what actually came out from the surveillance reports. That was a fabulous learning experience.
I learned the principles of what makes a very good trader—how to evaluate risk, how to think about risk-taking and liquidity. I also learned how to behave in a marketplace where people are doing silly things. It’s always difficult in any industry when competitors are doing silly things not to copy that, but we found ways to manage it.
That was my background in fixed income. It’s a weird pivot, but I then went to the buy side with Brevan Howard and BlueCrest, trading mortgage-backed securities and caught the upswing of that. I finally thought the mortgage trade was over and made a crazy career move—I took a 50-to-1 pay cut and moved to Munich to learn how to value companies and invest in equities.
John: MAAT has a highly concentrated portfolio and a value-based philosophy. How do you contrast that with the concentration we’re seeing in the tech-heavy indices and passive investing? In your Q4 letter, you also referred to some cracks you’re seeing in the credit market. Tell us about your macro view.
Shaun: When it comes to what we’re doing at MAAT and the principles we apply, it relates back to what I’ve done earlier. The securities you’re investing in or the tools you’re using in building a portfolio can be quite different, but the principles of how you use them should be a solid foundation. If it’s built on granite, it will last for an incredibly long time. If it’s built on sand, it will fade away very quickly.
Anytime you confuse a trade for an investment, I think you get into trouble. A trade can be built for the short term, and the narrative doesn’t really matter. An investment has to be very different. I saw first-hand people who were trading products built for duration succumb to short-term assumptions—thinking nothing would go wrong in the short term and they’d always be able to exit. That’s kind of what you’re seeing in private credit now. The premise is: I’m going to take five or ten-year loans and put them in a fund where you can redeem the equity within a month. It’s the classic asset-liability mismatch.
What I learned on the structured products side was to really think about what assumptions you’re making. You’re always making a ton of assumptions, and it’s very dangerous if you don’t know what they are. In structured products or MBS, you know exactly what you’re dealing with—CPR or CDR recoveries—and you can do that down to whatever level of specificity you want: at the property level, MSA level, zip code level, whatever. The same principles apply when you come to investing in corporate debt or equities. The variables change, but the principles apply all the time.
Step one: Do you know what assumptions you’re making? Step two: Do you understand them well enough to make a reasonable parameterization? And step three: Are you checking to see if you’re correct? I love a quote from George Box: all models are wrong, but some are useful.
We bring that to the research side. We try to build our approach in looking at companies in a very specific way, as if we were going to own the company for 10 or 15 years. I’ve taken companies private at my prior employer and understood these companies as well as the owner-founders would. It wasn’t uncommon for some of the CEOs to come and invest with us afterwards and stay for many years. Some of our best investors came that way.
If you’re going to use a model, use it for what it’s useful for. Know the assumptions you’re making, know where you’re strong. This goes back to Munger and Buffett’s comments on knowing your circle of competence. You can expand that circle all the time, but you’ve got to do it slowly and check if you’ve actually learned. The illusion of knowledge is a very dangerous thing.
We deploy that actively when looking at a company. We do very deep research and prepare a primary report. We also prepare a model, but the model isn’t built for me—it’s built for everyone else in the investment group. The assumptions are very clearly laid out at the top, the outputs at the bottom. We treat the model as a way of sensitizing things and reflecting how we think about the business. It’s a gross oversimplification. Anyone who’s run a business knows you can try to look at it through a spreadsheet, but it’s far from that in reality.
The other thing is that we’re hyper-aware that all businesses are really just people. A better group of people with a better sense of philosophy, shared values, and good communication will get much better returns from the same assets than another group with poor incentives. We’re always trying to look at the qualitative. I’m a huge fan of Nick Sleep at Nomad Partners, and if you listen to him toward the end of his career, he was constantly talking about the qualitative aspect of investing. If you listen to Buffett and Munger over the years, they’ve said it’s 99% qualitative. For people new to investing, that sounds silly, but to me it’s obvious through experience.
The same is true for mortgage securities during the financial crisis—it’s how you thought about what could happen in the world. There’s a great quote from Jeff Gundlach about people trading subprime in early 2007, getting very excited because spreads exploded. What they failed to realize was that spreads were exploding because interest rates were falling rapidly—government rates—and that’s because a crisis was unfolding. You have to have that nuance. Be thoughtful about what parameters are happening, be thoughtful about what can happen even if it hasn’t happened before.
We try to bring that sensibility to MAAT. When we founded the firm, we founded it based on a very strict set of criteria. We wanted to make sure it wasn’t set up as a trade. We wanted something we could build for 20, 30 years and be very proud of. I’d worked at lots of different funds—I’d seen some succeed, some fail. I went around and inverted the problem, thinking about what were the common factors that made those firms fail. Almost always it was because they had not set out the rules of the game ahead of time. They did not have shared values or principles, and ultimately people were tearing each other apart. After a run of success, maybe people started saying it was all them. Or you got a situation where people felt disrespected and disengaged.
One of my pet peeves in the investment industry is that you get incredible brain power—remarkable people with great educations and work ethic—and they come into a firm and suddenly no one listens to them. They don’t have a say. They feel like they’re just furniture, and in many cases they feel they’re just there for optics.
We tried to build a firm where everyone on the investment team has full engagement. The analogy I’d make is that you can have the engine of a Ferrari, but if you’ve got the transmission of a Fiat Panda, it’s not going to work well. When we have an investment committee meeting, nobody shows up ill-prepared. Everyone’s done the work on the primary report. If there’s a difference of opinion—and there always is—we really encourage that, but it has to be parameterized in some way. It doesn’t have to be the most accurate thing, but it means that people are thinking.
When it comes to making investments, everyone has a say and everyone has thought about it deeply. Their point of view has been expressed, and it’s active engagement, not passive engagement. That’s really important to get the best out of people.
On concentration: I think it’s one of the most misunderstood things in investing. If you look at the financial crisis, the big short was in effect a very concentrated bet, but it had incredibly asymmetric payoffs. In my book, I had a 45 million maximum loss. I had call options to get the capital back. But I stood to gain 6 billion over three years, which is close to what happened ultimately. That’s a crazy thing to turn down, especially if you think the odds are massively in your favor.
Likewise, in the Euro crisis in 2011 and 2012, we were buying heavily distressed mortgage-backed securities. There was a change in regulatory capital rules for both banks and insurance companies—Basel II and Solvency II. That created a rush to the exits in about six months in 2011. Because we were well prepared—we’d studied the impact—we realized that a lot of the people holding the securities didn’t realize what the impact would be on them.
We had 100% cash for that space. We were able to deploy it and buy securities at 3 to 5 cents that had been trading at 70 to 90 cents six months beforehand. We didn’t know it would get that bad or that the opportunity would look that good, but we knew there was a decent chance of it. The opportunity cost for us not participating was only about 7 or 8 percent per year, which we could live with.
In that case, we had 100% of my portfolio at Brevan Howard concentrated in those securities, and they made 100% a year for the next few years. I don’t think concentration is good or bad. I think it’s a tool—a form of leverage. It can be very bad if you don’t know what you’re doing, if you pay the wrong price, if you’re in with bad partners, which can often happen in public equities with an unstable shareholder base or a major holder. You just have to be mindful of that.
I wouldn’t say it’s something that should be used often. It’s for when you find a situation where the downside is incredibly limited and you can demonstrate that over two to three years, and the upside is hugely asymmetric—maybe a three, four, five, tenfold or more. When you see those opportunities, by definition they’re quite rare and usually quite small. Usually something very bad has happened. You’ve got to go for the jugular.
You’ll hear that from Stanley Druckenmiller. He says he learned it from George Soros. But Buffett and Munger did the same thing. Buffett had securities that were 40, 45% of his portfolio at different points in his career. Munger had something similar. Joel Greenblatt, the same. Chris Hohn, if you look at his performance over the last 20 years, it’s incredible, and he runs a concentrated book too. In the hands of someone who knows what they’re doing and is very disciplined and good at assessing risk, it’s a very powerful thing. It’s one of the few ways to outperform the market, married to a longer-term horizon.
People push back, and it’s understandable why—there have been so many famous blowups. But what I find most surprising at the moment is the amount of concentration in the marketplace right now. If you look at the S&P or the NASDAQ, just from what’s going on in AI, you’re talking about concentration exceeding what you saw in 1999-2000.
If you look at the MSCI World and how much US indices are in there—70% plus—and then apply the concentration of technology within the S&P, which passes through that, it’s really quite overwhelming. When I say it to people, they’re often shocked to hear how much of a passive investment or tracker they’re exposed to.
Historically, anytime one given sector became extremely concentrated—industrials or the go-go companies in the late 60s, consumer staples in the Nifty Fifty in the early 70s, energy companies around 1980, technology companies around 2000, and financials around 2006-2007—anytime they got up to 25% or 30% or higher, what happened in the following 5 to 10 years tended not to be pleasant.
We deploy concentration because we think it’s a competitive edge and we think we know what we’re doing. We think it should be embraced, but with the proviso that you really understand what you own and you’ve got good discipline about assessing the downside versus the upside. It can be very powerful in the right hands and disastrous in the wrong hands. We like to think we’re the right hands for that.
Elliot TURNER: Single-stock volatility has been meaningfully higher in recent years—trending upward and then exploding since COVID. How do you think about concentration in a world where the volatility of any given stock is far greater than it had been in the past, and what does that mean for underwriting positions, the required margin of safety, and how you build a portfolio?
Shaun: I think what you described is an undeniable fact: increasing single-stock volatility. In some ways it’s positive, depending on your structure; in some ways it’s negative. I’d take it as a net positive, thinking opportunistically.
There’s a symbol in Chinese and Japanese that says “crisis” and “opportunity” together. If I go back to when I was at firms like BlueCrest, DW, and Brevan Howard—much like you’d see with multi-manager shops today—they have very strict risk limits. If you lost 6%, you were pretty much out of a job. I used to equate that to driving a Formula One car with a spike on the steering wheel. You’ll get around, but you’re going to be very careful. Those firms are amazing, but it is very limiting and frustrating, and they’re selling to an institutional audience.
Why increasing volatility is both a risk and an opportunity: if you’re in one of those seats where your risk limits are strictly defined by value at risk or stress limits, as volatility picks up—there’s typically a 12 or 24-month lookback period—it inhibits what those people can do. They have to exit the market. You’ll see lookback funds as well where if vol goes above a certain level, they have to get out.
If, on the other hand, you run a structure with investors who are like-minded and think more like business owners—who are entrepreneurial and realize that when competitors get knocked out, you can do quite well—you have a different timeline and you will not be constrained by VAR. We have deliberately gone for a niche and targeted people who understand that. We go to high-net-worth individuals or family offices who have built a fortune themselves, understand that business is unpredictable, and embrace it.
How do we get them comfortable? We try to be as transparent and honest as possible. We explain our thinking and go through it. We invite them to critique us. In our letters, when we do case studies, if you disagree with an assumption or think we missed one, or think we’re asking the wrong questions, we want to know. That tends to build comfort. They know that we are thoughtful, that we’ve done our work, and that we’re thinking in a way they would think.
The more volatility you have, the more you restrict a lot of players from competing with you. The more volatile it is, the more opportunity you’re going to see. If I go back to 2011 and look at those bonds I mentioned—bonds trading at 80 or 90 cents that traded to 5 cents. If you’re limited by a VAR model, the VAR is going to look incredibly high, and most people cannot participate. Indeed, that’s what I saw with my competitive set. They just could not get engaged. The VAR was off the charts. VAR is a very useful measure for aggregating risk across a lot of securities in a big portfolio if you know very little about the things in them. That’s what it was designed for. It originated at JP Morgan in the ‘90s, specifically for board members so they could aggregate all risk into a number.
The irony is it was very heavily criticized post-LTCM and the ‘98 blowup because it obviously didn’t capture everything. My ultimate takeaway is that for our approach, the VAR will look infinite and tell you not to invest as the price goes to zero, even if there’s value in the asset. It just doesn’t consider the value underlying it at all. The more volatile it gets, the better it is if you’re disciplined about how you enter and mindful about the upside-downside risk.
For our space—we focus on European small and midcaps—there’s a risk factor that relates to your question about volatility: path risk. As a former derivatives trader, you have to think about this all the time. The old joke goes that you can drown in a river that has an average depth of two feet but has a spot where it’s 9 or 10 feet deep.
For our companies, we tend to get the valuation right, but that’s not enough. If I look at a company today trading at 100 million euros and I’m convinced it’s worth 200 million and will get there within three to four years, but it’s got some near-term problems we’re willing to endure—we’re willing to participate in a volatility path that others won’t, and that’s our advantage with that longer-term horizon. If it first trades down to 20 million—an 80% decline—for any reason, we may still be correct, and now it’s worth 10x. But we will not get the chance to stick with it because a private equity firm, or more likely today, an industry buyer or family office will step in and buy it. If it dropped 80% and they bid a 100% premium, maybe I was right over five years, but I’ve lost 60%.
When it comes to mitigating this, that’s where correlation comes in—something I used to do as a correlation trader. When we’re building a portfolio, we are incredibly mindful of where the value will be in three to five years. We try to understand that really well and how the mechanics of the business work, the competitive dynamics. We’re really underwriting it as if we were taking it private. There are limits because it’s a public company and you can only do so much due diligence. But in addition to that, for every individual company, we’re looking at the path risk: What was it in the past? What does it look like now? What events could come that are not related to operations—it could be industry, could be another shareholder, could be anything that might cause a sudden gap down in price. If that’s likely, we have to reduce the position size.
Then there’s the correlation. Just like multi-managers, that’s how they monitor it. They overallocate capital on a nominal basis but look to lower the correlation. We do the same thing. If you look at our letters, you will see very disparate names, and the things that are likely to drive long-term performance are quite different for each. We hope that all of them work out, but it’s very unlikely that any one factor would drive them lower in unison. The exception would be something like the financial crisis where all assets fall in unison because capital is withdrawn, or a war. But we’ve even seen it in our portfolio this year—different names are performing on different bases.
Those are the elements of building a successful portfolio. It’s difficult to do well because historical correlations are not indicative of the future, especially if you’re looking at securities that have already had a big decline. We have to make a guess about that, but then we have a very different approach to portfolio management and capital allocation. That’s maybe our secret sauce. This goes back to my background as a correlation trader. One of my partners has a degree in advanced mathematics from Oxford. Another is an engineer. We’ve built our own software and analytics to think about how we parameterize these things.
We link all of our research from inputs on the operating side down to a return vector. We get expected returns across scenarios. We use a simplified version in the letter and then run that through a modified version of the Kelly criterion—a bet-sizing algorithm. Mohnish Pabrai has written about it. There’s a great book called Fortune’s Formula about that. It’s a modified version. When you’re doing it for an array of assets in a portfolio, you have to think about the covariance and correlation. If you get that right, it should give you optimal bet size. We don’t use this in a crude manner. There’s a lot of noise in what we’re doing. What we do use it for is to force-rank our choices. From first to 10 or 12 names, we feed in our return vectors and expected correlations, and it shows us what the right order of priority should be from the implied bet size. That forces us to stay honest.
One last thing: in many investment firms, you get an incredible performance from a given asset, and it becomes a marketing tool, justifiably so. But the price might have gone up three, four, or fivefold. If it doubles in price over the course of a week, unless something fundamentally has changed for the business and the cash flows it’s going to produce, the return has dropped by half. If it was 20% before, it’s 10% now. But all the behavioral biases would lead that firm to boast about that position and leave it as a big size. For us, we’re forced to address it when we run it through these criteria. It’s going to drop—if it was first, down to maybe fourth or fifth. That forces us to stay honest. We’ve built the whole investment process to keep an eye on what’s going on, making sure we’re not succumbing to biases. It’s not immunity, but it sets off different alarms.
John: You’ve described LLMs as “smart librarians, not scientists.” You distinguish between information and knowledge, and you believe AI will reinforce traditional intangible moats rather than destroy them. Tell us more about that.
Shaun: First of all, I think AI and LLMs are absolutely revolutionary. They blow my mind on a daily basis. In no way am I saying this is not a revolutionary technology or that it’s not going to change the world. I think it’s going to make the incredible compute power we already have far more accessible and democratize it. That’s wonderful, in the same way the graphical user interface did before you had a command line.
But where I get concerned, particularly as it pertains to LLMs, is the idea that these things have understanding, intuition, feelings, or empathy. That’s just not the case. This is not a new problem. Joseph Weizenbaum was a computer scientist, a professor at MIT in the 1960s. He did an experiment to show people that machines can deceive. He came up with a program called ELIZA in the mid-60s. What it did was very similar to what you see with chatbots today or the first versions of ChatGPT. He created a little algorithm that was pattern matching. It would have you give input—”Tell me about yourself” or “How do you feel?”—and it would look for keywords.
If you said “I feel sad,” its response would be “Why do you feel sad?” There was no logic—just a lookup table. If you said “dream,” it would say “What did you dream about?” He thought people would see through this very quickly. They didn’t. It became a huge sensation. In the late ‘60s and early ‘70s, people were legitimately talking about creating virtual therapists like ELIZA. Weizenbaum was horrified. He thought this was insane. A decade later, he wrote a book called Computer Power and Reason. He was saying that what you’re seeing from the computer objectively is just calculation. It presents in a manner intelligible to us, but we’re putting a human face to it. It’s just pattern matching and a scripted response.
His conclusions were roughly this: The computer can calculate, but it cannot understand. Just because the computer can do something doesn’t mean it should. He thought it was immoral to push this as a therapist when it does not understand emotion and has never experienced loss. His final point: don’t confuse cleverness with wisdom. Like Buffett’s analogy about hiring people—he always talked about the three I’s: intelligence, intellect, and integrity. If they don’t have the last one, you want to make sure they’ve got neither of the first two either, because they’ll destroy you.
The risk when you deploy these things is that you’re attributing great intelligence. People often conflate intelligence with wisdom. They’re not the same thing. Then they hand over decision-making to it. That was Weizenbaum’s big fear—that people would outsource their thinking.
By way of analogy, if you look at the financial crisis, a very long-term cause was ratings-based capital calculations. That goes back to a decision in the early ‘70s to create the NRSROs—Moody’s, S&P, and later Fitch. The idea was these agencies can assess the risk. Everyone at the time would not rely on ratings. They would do their own assessment. But by the time you got to CDOs, people just trusted the model blindly, not realizing that the incentives had been perverted.
When you look at LLMs today, on a simplified basis, they’re almost an elaborate version of predictive text in your Motorola phone from 20 years ago. They can do it to far greater depth, width, and speed. But they fail if you put them through certain tests. There’s a lot of evidence now that if you give them too long a context window or keep questioning in the same context window, the results get very bad very quickly.
There’s a great paper released late last year called “The Artificial Hive Mind,” which tested every major LLM asking open-ended questions like “What is life about?” Almost all of them gave close to verbatim responses because of the way they’re trained recursively. They’re trained to give you back the average.
When you go to the analogy of librarian versus scientist, the LLM is an incredible model for accessing and retrieving information that we already know and putting it in terms you might understand. It’s like a great tutor. But that’s based on the idea that all knowledge is known at this point and the LLM is just a way of retrieving it. That’s not how human knowledge develops over time. That’s the distinction between information and knowledge.
When we say librarian and scientist, we crib that from Demis Hassabis, the founder of DeepMind. He’s gone on record talking about this multiple times. He says it’s more like a repository of data, and it’s a brilliant tool. In the same way a Kindle can hold a thousand books and you don’t have to have physical copies in your room—that’s revolutionary. But it’s limited. Let’s not confuse it for what it’s not.
He gives what he calls the “Einstein test.” If you went back to 1904 and trained an LLM on all human knowledge up to that point, would it be able to come up with the theories of relativity? His answer is absolutely not. It’s regressing back to the norm. What it would do is give you an answer in Newtonian physics. Einstein had to do a lot of lateral things and question key assumptions. How can an LLM ever really create? As they’re structured currently, it can’t.
If you look at Pluto, for about 80 years that was a planet. Depending on when you train the LLM, Pluto either doesn’t exist, is a planet, or is no longer a planet. Which is true? I’d go back to Karl Popper—all the facts we currently have are really just things we’ve yet to falsify. With that mindset, you’re always questioning things. An LLM doesn’t think that way. It thinks it’s been trained on everything in the world and that is the truth. That’s the limit of it. It has the illusion of great wisdom and knowledge, but it’s limited to what it knows.
Humans have been fooled by things like this before. There was ELIZA back in the ‘60s. My son is an avid chess player—there’s the Mechanical Turk that fooled people for nearly a century. It’s still an incredibly powerful tool, but you have to be aware that this is probably purely a probabilistic extrapolation machine trained to provide certain answers, largely uniform. Like any tool, it’s really powerful, but you’ve got to know its strengths and weaknesses.
Elliot: In his book on AI, Ethan Pollock makes the point that the very fact that LLMs struggle with hallucinations is one of their greatest strengths and greatest weaknesses because it illustrates their power to be creative. LLMs are in fact capable of being creative, but it’s bounded—it takes a prompt and some orchestration from the human. How do you think about that, and how does AI affect the barriers to entry for companies that sell intangibles?
Shaun: On hallucinations and creativity, that’s absolutely true. The paper “The Artificial Hive Mind” made exactly this point: when you ask general open-ended questions, you get a uniform response—whether it’s Gemini, ChatGPT, Anthropic’s Claude, whatever. DeepSeek too. How do you get around this? You have to ask questions that force it to deviate from the most expected, most accepted answer, which is going to be average.
It’s a bit like when Matt Damon was on Joe Rogan talking about this for people developing scripts. I think LLMs can help with creativity by forcing you to make a lateral leap. I remember as a boy growing up, there were books by Edward de Bono about lateral thinking—trying to get away from conventional norms and think in a different manner.
If you look at what Buffett described as a great investor or great underwriter on the insurance side, it’s someone who can imagine things that have never happened and really picture them vividly. I think that’s where an LLM is incredibly powerful for investing. If you know how to direct it, if you know how to force it away from the average response, you will get very good outcomes.
One of the things we do is have the LLM question all of our assumptions, find weaknesses in the logic, try to find facts that contradict what we’ve assumed. To play the 11th man. I think it can be really helpful there. The weakness, as you said, is more on the precision side.
There was a company I had spoken to several years ago that was trying to develop a model to deal with legal contracts. In a legal contract, every clause, every “i” that’s dotted and “t” that’s crossed has to be accurate. You’re in big trouble if you get an error rate of even 50 basis points in a key clause. As a former bond trader, it’s similar. You cannot have a situation where it’s right 9 times out of 10. So anything where the error rate or severity would be really high, an LLM requiring precision is going to be a poor choice.
As it pertains to intangible moats, it’s really to do with businesses where technology might appear to be at the front, but the strength or the moat of the business is something different. An example would be Bloomberg. I’ve seen a lot of talk on X where people say “I built a new Bloomberg—I can pull this data in, I can do an earnings analysis.” That’s a very small aspect of what Bloomberg does. It’s really used for messaging, portfolio management, bond trading. You cannot live without it.
It’s the network they have. It’s their ability to take a lot of disparate data and put it in one repository that many people use, especially larger asset managers. I think they’ll be aided by AI. One of the complaints about Bloomberg in the past was it’s very difficult to figure out all the commands. It looks like something out of the ‘80s. If anything, the LLM interface will democratize it. They probably can reduce the sales force intensity. They’ve got thousands of well-trained engineers who can become a lot more efficient. I just don’t see a natural competitor coming around. They’ve got an installed base. A lot of people have built their processes and systems on it. That is the moat of that business.
Another one is Edenred. An LLM is very useful for them. They can reduce the number of programmers or make them more efficient. On the positive side, it makes a universe of customers that weren’t historically available to them—small and medium businesses. It was not cost-efficient for them before. They had an enterprise-type sales force. AI makes those customers a lot more accessible. They can automate the initial part of the funnel, get those customers in, and they benefit from the network. The marginal cost of onboarding beyond that is very low.
One more example: the UK used car market, dominated by AutoTrader. It has perhaps 90% of the share of classified ads in the UK. That’s a moat. AutoTrader is a notable case—it’s gone from its all-time high down 50%. The idea is that you can create a classified site very quickly and easily for free. The engineering barrier is no longer there. I don’t think that reflects where the moat for that business lies.
The moat is that they are a verb in that marketplace. Everyone knows AutoTrader the way you and I know Google. That’s where people go to search for vehicles. They have most of the high-value dealers on board. They’re the number one lead source. What they add in value is inventory turns. Used car sales in the UK—the actual selling of the vehicles, servicing, or financing—is a very low margin business: 7, 8, 9%, depending on the dealer.
What’s critical is that they turn inventory very quickly. If you turn it only twice a year, you’re not going to do well because you’re using the same parking spot. If you turn it 12 times, you’re fantastically profitable. What’s critical to that is the quality of the lead you get. They have 90% of the traffic and almost all of it is organic—96%. There’s 4% paid search.
People are saying they’ll create an LLM-based classified business and all the dealers will come. For sure, dealers will give their inventory to somebody who wants to come in—it’s free for them—but they’re not going to pay a premium for that. Google pretty much owned all of the search space, but they weren’t able to displace AutoTrader over 20 years.
Why? Because of tools AutoTrader has built for dealers that help inventory turns: inventory management, pricing for exit to maximize turns, pricing stock they’re going to buy at auction, financing partners. In our opinion, the barrier to entry just got bigger for someone trying to create a competitor.
Look at Heycar, a Volkswagen-sponsored business selling new and certified pre-owned vehicles. Huge support, 250 million invested in the brand—went nowhere. Never sold more than a thousand vehicles a month. Cinch and Cazoo in the UK, equivalents to Carvana in the US—collectively, these three companies spent a billion and a half plus trying to build a brand and could not compete.
The reason is: how do they get people onto their network? They’ve got a chicken-and-egg problem. They need people on the network to make it valuable to dealers so they can charge money. To get people onto the network, they have to pay for search. That’s cost per lead. What you’re seeing now with Google is the AI answer at the top, which stops people from going to rivals to AutoTrader like CarGurus. It makes the customer acquisition cost a lot higher than it was 5 or 6 years ago.
AutoTrader is so unmatched in the UK, which is a very niche market—left-hand drive, an island. Who’s going to come in to compete? Where would they get the data? There are rules in Europe that prohibit or limit the ability to scrape data. When you’re dealing with a product that turns over 10 to 12 times a year and it’s a perishable good, the data you require to price it accurately is very discrete. Who has the best data? AutoTrader.
The only way a business like that could be negatively impacted is if one LLM came to dominate—which doesn’t look to be the case—and started charging a toll for directing traffic. If Google couldn’t do it over 20 years, why would an LLM? Even if an LLM wanted to, does it make sense for them to get into a battle with the dominant data provider who has enormous market share? The cost to build a competitor just went up a lot. The advantage for that business was never technological. It wasn’t in the engineers or the user interface. It’s in everything else and the strategy of the business.
It’s similar for Edenred. It’s in their network. An LLM is not going to help you acquire merchants—go out to restaurants and sign them up. If you’ve got an existing network, AI might make sales easier or help you go to a wider network. It might make your engineering cost drop over time. For businesses where the moat is not the code but all the other aspects, we think AI probably increases the barriers. For things where it’s a simple automation tool, it’s a different story, and there I think people can get in trouble.
Elliot: I just want to lean in on some of the points you made, because I think they were fantastic and well-stated. I thought it was ironic that somebody used DoorDash as an example of where AI agents will inevitably kill it. When I followed the history of the company from when it was private, there were dozens of companies with very similar code, where the UX looked similar. They even in some cases had similar numbers of restaurants and good overlap in supply. Aggregating supply was a challenge, but others did it too. Execution matters, and the way the business runs matters. It’s literally running a service, not just software.
I tend to agree with your points. I know John’s written some really good stuff on how there are some areas where barriers to entry drop meaningfully and the pricing landscape might change. Where even just the threat of being able to code something might tilt how you approach it, or you might have lower headcounts. If a company can’t evolve from seat-based pricing to value-based or consumption-based pricing, there could be challenges.
Shaun: I think your DoorDash example was wonderful—I wish I’d used that rather than my own, Elliot. I’ve read some of John’s stuff, and I think it’s wonderful as well. I think you guys probably understand this better than I do, but I think you’re right on the money there.
John: Let’s move on to the case studies. You alluded to Edenred—that’s one of your newest portfolio additions. But let’s start with a passed opportunity: Vistry Group in the UK. It’s a good case study of your private equity-style due diligence. What made you reject the “NVR of the UK” thesis, and what were your forensic accounting concerns?
Shaun: It’s timely because they just came out with results last week, and an announcement about the departure of the chairman and CEO—Greg Fitzgerald, who’s had an amazing career at the top of that company.
Briefly, for the audience, Vistry is arguably the largest homebuilder in the UK—first or second. It’s the aggregation of three, maybe four, different homebuilders. It’s a fair analogy to NVR. NVR had an asset-light approach to building homes, which revolutionized the model in the US. It’s probably most famous because of Mohnish Pabrai, who wrote about it in the 2000s. When someone thinks of a homebuilder, you don’t think of a business that’s going to make returns like Microsoft. But if you go back to ‘93 or ‘94 when NVR emerged from bankruptcy, you made a similar return to Microsoft—about 18% per year. Quite remarkable.
We had prospective investors—about eight of them—almost in unison telling us to look at Vistry. I went into it excited and wanting to own it. As part of my research, I listened to some podcasts and heard some stuff that struck me as strange. Someone said it was a cost-plus model. When I looked at the results historically, I knew that wasn’t the case. That’s not how it worked.
When I looked at how Fitzgerald had aggregated the company—three homebuilders with a huge geographic footprint—I thought integrating these in four years was going to be a real challenge. Culturally in the UK, the difference between the North and the South is enormous.
When I went through their CMD and saw the claims about unit volumes—getting up to 30,000 homes, where 25,000 would be disappointing—and what they were claiming for margins, it seemed strange. But we investigated.
What we found was they were going to struggle on all the key variables. How we summarized it: this is a function of unit volume, margin, and average sales price. Can you parameterize them accurately?
We started with unit volumes. We looked at all the homebuilders in the UK going back 40, 50 years: What have they done? How easy is it to get to those targets? When you look at the history, it’s not easy at all. Very difficult. Nobody has done it, especially not while integrating three different businesses.
Then we asked: is there going to be huge government stimulus giving people more propensity to buy, or relaxation of very restrictive planning rules? They’d announced a roughly 35 to 39 billion program. People were very excited. But when you examined it, that program was over 10 to 12 years and it wasn’t sure when it would start. Breaking it down, you’re looking at adding maybe two or three thousand homes a year to the overall UK market. It’s not enough to move the needle for one company that needs to go from 17,000 homes per year to at least 25,000.
There’s been talk for 20 years about liberalizing planning standards in the UK. It is so difficult to do because it’s largely a municipal problem and very intensely political. People have promised it for a long time. It never gets delivered.
If you look at affordability, disposable incomes were really being squeezed. The commonest argument in favor of homebuilders—the housing shortage—does not mean people can afford to buy a house. They’re too expensive.
The final aspect particular to Vistry was their partnership housing model. They had moved to an asset-light model where they don’t have to pay for the land. They partner with local municipalities who need to build council houses or rental houses. Lower margins, maybe by four or five points, but much more rapid turnover of capital. So the return on capital would be enormous. That was the argument.
In practice, local municipalities in the UK are in dire financial straits. Many of them are doing very poorly. It was going to be difficult for them to fund this. And despite their best intentions, NIMBY tends to win out.
On the accounting side: this partnership model is one where they win a long project, anywhere from five to ten years. They don’t need the capital upfront, but they commit to a building plan and make assumptions—how much for labor, materials, timeframe. When they locked in these partnerships, yes, they had guaranteed sales, but they’d also locked in a guaranteed price. It was not cost-plus at all.
If they made a mistake on what house price inflation was going to be relative to build cost inflation, they had a problem with margins. Over three to four years, if the cost of inputs was much higher than house price inflation, margins would be much lower than expected. The margins they expected on these partnership homes were around 18% on a gross margin basis. That was the best you could do assuming your initial assumptions were correct.
All of the UK homebuilders had experienced great margins throughout the mid-2010s because home price inflation dramatically outstripped build cost inflation. That reversed with Brexit and hasn’t gotten much better since. Vistry was aiming for 75% of business to come from the partnership model. That means 75% of gross margins capped at 18%. If OpEx is 7, 8, 9%, and people were talking about 12 to 14% EBIT margins, how do you get there?
When they had the profit warning—they had three in a row over four weeks—the way the information was released was concerning. They didn’t do it upfront, which indicated they didn’t know the extent of the problem. It also indicated something rotten in the culture underneath. I don’t mean to suggest the executives knew this, and Greg Fitzgerald was an exceptional CEO. But it means integration of these mergers was not going well and the systems weren’t working properly.
Cultural problems in large, geographically dispersed organizations are very hard to reverse. We looked at things and said: if you correct for the accounting catch-up, the average margin on a net basis was in the single digits—5 to 6%. Our fear was simple: when did they win most of the new partnership business that’s going to be built in 2027, 2028, 2029? They won that in 2022, 2023. There’s a long lead time. Isn’t it very likely they bid for those projects with rose-tinted glasses?
It wasn’t that they were going to lose a ton of money or that there was fraud. It was just that the margins they were claiming seemed incredibly unlikely to us. If you mix in unit volumes—also very difficult to hit—and the margin challenges, where do you go with that?
I was open-minded to the idea that things could change. Maybe you do get revolutionary change in planning permission, but I want to see it first, not the promise of it. When we ran it through that lens, it looked pretty bad.
One last thing that applies to all the homebuilders in the UK: there are liabilities not immediately obvious. If you look on Bloomberg or CapitalIQ or FactSet, there are real liabilities you should add to the enterprise value that don’t show up on the standard form.
The two are provisions and land creditor debt. Land creditor debt is where you agree to buy the land and they give you vendor financing—you don’t pay until you start building. Provisions are exactly what you’d think. In the UK, this became a major issue after the Grenfell disaster. They found that a lot of council housing had not used the best insulation materials. Other problems were discovered.
At the time we wrote about it, in 2024, they had reported debt on the balance sheet of 541 million and cash of 320. So 180 million of net debt doesn’t seem earthshaking. But add land creditor debt—another 740 million—and provisions of 352 million. That was for a business doing adjusted EBITDA of 330 to 350.
The combination was nearly 1.1 billion, or three turns of EBITDA. All the bullish arguments said it was just a very cheap multiple, but that’s excluding things that are real liabilities. The provisions are real. Land creditor debt is real. You’re going to pay them the same as you would other debt.
Over time, provisions in 2021 were only 100 million and grew to 350. Land creditor debt went from 540 million to 740. We just never saw much cash really coming out relative to what they promised. They did some buybacks, but nothing close to what they wanted.
We were quite disappointed at the end. We’d done a lot of work. But this was a great case study for showing our investors how we think.
John: Let’s hit the highlights on one more idea: Edenred. You also presented Mayr-Melnhof Karton at MOI Global’s Best Ideas 2026 conference, which is available online on the MOI Global website and the Latticework Substack. We won’t get into that one. But give us your thesis on Edenred—a three to five-minute snapshot.
Shaun: Edenred, for those who don’t know it, is a food card and fuel card business. If you’re in the US, you may not know it. You’d be very familiar with it if you were from Brazil, France, or Germany.
These are employee perks—a way of giving extra benefits without being so inflationary. It used to be vouchers; now it’s a card. You can go to a restaurant, order food online through DoorDash, or buy groceries, and use this credit instead of cash. It’s issued at a discount. Later on, they went into the fuel card business. If you’re in Europe, you’ve seen these. It’s for truck drivers or for us driving our car—you can expense it properly.
Edenred is the biggest player, probably the most innovative. It’s a French company. The biggest companies in the space all came out of France because they originated the idea. This was done during very inflationary times as a way of not stoking inflation as much, because it’s not quite like cash compensation. A new CEO about ten years ago realized they were very dependent on regulatory markets. Why do these food cards exist? Because there was a tax deduction. If you gave this to the employee, it was not taxed as income. So you give them maybe a 10% raise without the inflationary impact.
Once these businesses got big enough, they could scale and get discounts from merchants the same way DoorDash or UberEats or Deliveroo get discounts from restaurants. So this is a business that’s done very well—very asset light, network-effect type business, fantastic growth over 10 years, very good acquisitions. People loved it in ‘21, ‘22, ‘23. It got to very high valuations—up to 18 billion. Today, it’s a market cap of 8.5 billion and an EV of 6 billion. Why?
It all goes back to what they were afraid of: these are regulated markets. In some major markets—Brazil, Italy, France—there was terrible abuse of smaller merchants. If a small operator was using Edenred’s service, they might be charged 15% on every transaction, but a large retailer might pay 1%. Think of it as a take rate. So regulators came in and capped what could be charged—4.5 to 5% in Italy, closer to 2.5% in Brazil.
If you cap that long tail of smaller merchants that pays much higher premiums and make the average 2.5%, it’s pretty catastrophic for margins. They had this horrible 18 to 24 months where Italy regulated, then Brazil did something far worse, and the French are debating it. Those are the three largest markets.
People said this company was earning ridiculous economic rent unfairly, all because of regulatory protection that’s now disappearing. When we looked at it, we inverted it and said: actually, they’ve been increasingly diversifying away from these regulated businesses. The regulatory impact in all of those countries has already happened. It’s already priced in. The horse has already bolted.
If you just value the non-regulatory-impacted businesses, it’s probably worth more than the EV today. In addition, a lot of the regulated markets are actually growing in their favor. The secret sauce that people miss is their enormous network—millions of merchants, millions of employees as end customers, and six-year contracts on average.
They have the opportunity to deploy many more services. Where 10 or 12 years ago they might have had one or two products—the fuel card or the food card—now they’ve got an average of 2.2. In some countries it’s up to 5. They can take advantage of this network and add more services at very little marginal cost.
We think next year the business is going to make somewhere between 1.35 and 1.4 billion in EBITDA. The company today is worth an EV of about 6 billion. It also generates an enormous float—about a billion and a half. There’s a lot of value there that people don’t really pay attention to.
If you take the company at face value—they think they can grow at 10% per year after this trough transition year through 2030—then you’re looking at a business with around 1.8 billion of EBITDA, very high margins. They’re actually pretty good capital allocators. It’s just hard for us to imagine a business that’s grown on average at 7% a year for five years trading at around 3 to 3.2 times EBITDA.
In a nutshell, that’s the argument. I think it’s just misunderstood, and people are focusing more on a sequence of very bad news events and really on the wrong items.
About the guest:
Shaun Heelan serves as Chief Investment Officer of MAAT Investment Group, based in Munich. Shaun brings over 20 years of investment experience across a broad range of asset classes. His career spans both the sell-side and buy-side, covering everything from highly liquid instruments to complex, illiquid assets. Known for his disciplined investment process and strong risk management, Shaun’s expertise has shaped MAAT’s philosophy and long-term approach.
The primary purpose of this podcast is to educate and inform. The views, information, or opinions expressed by hosts or guests are their own. Neither this show, nor any of its content should be construed as investment advice or as a recommendation to buy or sell any particular security. Security specific information shared on this podcast should not be relied upon as a basis for your own investment decisions -- be sure to do your own research. The podcast hosts and participants may have a position in the securities mentioned, personally, through sub accounts and/or through separate funds and may change their holdings at any time.
Featured Events
Best Ideas Omaha 2026, Omaha, Nebraska (May 1, 2026)
The Zurich Project 2026 (FULLY BOOKED), Zurich (Jun. 2-4, 2026)
Latticework 2026, Chicago, Illinois (Nov. 10-11, 2026)
Ideaweek 2027 (FULLY BOOKED), St. Moritz (Feb. 1-4, 2027)
Enjoying Latticework? Help us make it even more special.
Share Latticework (simply click the above button!)
Introduce us to a thoughtful speaker or podcast guest
Be considered for an interview or idea presentation
Volunteer to host a small group dinner in your city
Become a sponsor of Latticework / MOI Global
Volunteer by reaching out directly to John (john@moiglobal.com).
