How machines will transform private capital markets | Ep. 83

Speaker A: The investment outcomes of those algorithmic investment decisions were vastly superior to at least a majority of investment decisions that are observed in the market done by normal private equity teams.

Speaker B: Would you expect to see a kind of a one off improvement in someone's performance if they had, let's say, cracked the code? Is that what you're looking for?

Speaker A: You get quickly to a level of complexity, 99 factors, you cannot possibly understand anymore, the why of the causal linkages between them.

Speaker B: The secret really is how they interact then, because a lot of these things just sound like common sense.

Speaker A: And if you start trading in public markets, other people see what you're doing, they follow you and then the opportunity goes away. Well, in private equity we don't have that problem. If we, private equity are not going to be able to do this, you know, we're going to have the Googles and Metas of the world come in. They understand data, they understand the algorithm. You're just going to have a product that blows us out of the market. It would be fantastic to kind of build out the vanguard of private equity by bringing science to secondaries and designing products that have widespread appeal for investors in private equity, institutional and mass affluent at dramatically lower cost of the investment.

Speaker B: Professor Oliver Gottschalk, welcome to Fundshack. You're one of the best known academics in private markets globally and you've also had a commercially available product under Gottschalk analytics whereby you help GPS and LPs understand the return drivers of product private markets funds. And this can be quite a complex thing to ascertain because unlike operating companies which just make sales and keep their costs down, the return sources and drivers, the value drivers of private markets funds can be difficult to determine. And so you help people understand that. You've also been doing a lot of work in the machine learning space and I believe you gave a speech, uh, at Super Return entitled Machines at the Gates. And in that um, you make the case, and I think it's more than a theoretical case, it's empirically backed that, that human um, allocators can now be surpassed with machine learning algorithms. At least your one. Is that overstating the case or is that the case? Because it's quite the claim?

Speaker A: No, it's actually quite accurate and it's very fascinating, uh, with respect to the opportunity that that offers. So all of this is in the continuation of my academic work. We've been studying private equity empirically for over 25 years. By now passionate about trying to understand drivers of risk and return in general, particularly those that reside with the fund manager. And as you say, it's uh, non trivial to find those drivers of returns because basically, you know, from many other asset classes there's uh, the concept of the random walk. So yesterday's returns have nothing to do with tomorrow's returns in private equity. We're going to look at this a bit more nuanced. If you just look at the stability of returns as a fund that was, you know, fund manager was top quartile somehow in the past, or they're top quartile again, that's instable to say the least, depending on who you ask. Um, if you however look at past performance as so much as it's maybe indicative of certain underlying skills of the fund manager, you measure then components of performance or certain elements in the track record that are signals of skill, it seems intuitive that some of these skills likely persist over time and they may be valuable in the future in the sense that they may be the source for subsequent outperformance. A lot of my work is looking at those persistent indicators of skill and you can easily imagine that while it's been for the Gottschalk Alex business, very powerful to measure those skills, to paint a picture of the different chromosomes of the valley creation DNA of a fund manager and put this in front of human decision makers to assess what's the skill set. I see. Is the skill set likely to persist and is the skill set likely to be valuable going forward? It's now a natural evolution to say, well, let's take a related set of even broader and more complex indicators of skill, subject this to machine learning techniques and basically look lots of historic data in order to identify a uh, predictive ability for performance drivers of various parts of private equity. And that's indeed fascinating because that's a game changer with respect to how I believe private equity is going to go about allocating capital going forward.

Speaker B: And so you've done that and what's been the results?

Speaker A: Well, it's uh, very interesting. So I've been at this for eight years now. The first presentation on this topic was a super turn keyNote back in 2018, which kind of dates me and this research, um, because I've always been fascinated, uh, with how much these statistical techniques can do. And of course the methodologies get better, the data gets better, better, and we're now 2025 basically ready for primetime. We see very clearly in the backtesting that had an investor been using the models that we've been building and that actually were on the table in 2018 1920. The investment outcomes of those algorithmic investment decisions were vastly superior to at least a majority of investment decisions that I observed in the market done by, uh, normal private equity teams.

Speaker B: So were you beta testing or was anyone following your algorithms?

Speaker A: Well, there are certainly brilliant investors out there and certain are without a doubt better and always will be better than the algorithm. But if I do certain simulations where I look at um, let's say a universe of US public pensions, which, uh, nothing, uh, to say against US public pension schemes, the only reason I like them is a fruit fly, is that they conveniently are legally obligated to put everything they do, including performance outcomes, in the public domain. So I designed a little experiment. I looked at what US public pension have done in private equity in the buyout space. To be particular, look at the large programs, those who invested over a certain time period more than 10 billion into private equity, and assess then the degree to which there would have been improvement had they followed the prediction of an algorithm that was available at the time. So if I look at decisions in 2019, I honestly trained the algorithm and exclusively on data up until 2018, and then see what decisions would have been, uh, predicted and supported in 2019. If we do this systematically. I designed a very, um, prudent test without any unrealistic assumptions, looking at a given pension plan at the time, looking at what they actually committed to, and then just said, for instance, okay, you subscribed in this vintage year to seven private equity funds. Let's not look at your actual commitments and the scaling, the calibration across those seven funds as it happened in reality, but let's listen to the algorithm and let's overweight those funds that were predicted to be outperforming by the algorithm and vice versa. So not assuming access to additional managers, I'm not assuming shifting to esoteric high risk strategies, I'm just assuming a tiny bit of recalibration, reshifting within the given portfolio of each of the 17 public pensions that I looked at.

Speaker B: So that deals with the access problem as well. You're just rebalancing an existing.

Speaker A: I'm just rebalancing. Exactly. I want to see if this is a real honest backtest that's possible doing this over a number of years. For my 17 large US pensions, they invested 96 billion, they committed 96 billion to private equity funds at that time period. I observe as of today, collectively these pension schemes would have been $6 billion richer in terms of TVPI and 4.5 billion of those would already be Cash and cash gains. So as the amounts deployed in private equity are large, 5, 6, 7 cents on the dollar turn out to be meaningful amounts of capital. And there was a first, very carefully M designed m modest conservative backtest indicating that there's actually, uh, wealth creation possible. If one was to add the algorithmic prediction in the decision making process, if I may I then it was of course also curious to see, well, if we get a bit more aggressive on the prediction, on the assumptions, what would have happened had these pension funds actually all listened to the algorithm. Not in the sense that they just recalibrated rate, but that they shifted entirely to whatever the algorithm predicted as the 10 likely best performing funds in that particular vintage year. The numbers change, it still works out, but you don't gain 6 billion, you gain 15 billion. So you have 15 cents on every dollar invested of outperformance. If you assumed for a second that you could follow exactly the prediction of the algorithm. Now from an indicative standpoint that points to the superiority of the algorithm in particular, as across the 17 different plans, not a single one would have been worse off. Some would have gained more, some would have gained less.6 cents on the dollar on average in the conservative backtest. But that speaks to the ability of the algorithm to be pretty robust and reliable in allowing investors to get pointed to likely outperformers.

Speaker B: Private equity is notorious for its lack, uh, of data or opaque data. How do you deal with that problem when you're training the uh, algorithm?

Speaker A: Well, sometimes I look back and having chosen to doing empirical work in an asset class that calls itself, uh, um, private is a bit challenging on the get go. But, um, I became very pragmatic with respect to data. Many people heard me say this before. Whenever it comes to private equity data, I refer to the saying that, um, in the land of the blind people, the person with one eye is the king or queen. Um, and I follow this very pragmatic approach in everything I do empirically in the asset class. So I don't ask the question whether something is perfect, but I ask the question do I have evidence that looking at certain subsets of data, however imperfect they may be, with the methods I can create, do I get results that directionally point me to outperformers and that of course, uh, underlies every type of analytics. We uh, live in a world of imperfect data. I see lots of very accurate data in my advisory role to large LPs and GPs, in my research, uh, role at the university. But I cannot use that data commercially. I can use it Only to kind of sharpen my pencils or build my toolkit. And if then I see data work really well, methods work really well in that data, I can kind of take it in the world of commercial data with a few more question marks around it, but I can compare if it still does the trick, if it still points me historically to outperformance. I use that methodology. And in building the machine learning algorithm, I applied the same tool. It's certainly not the only possible algorithms, not the best possible algorithm. But in literally thousands of back tests my team and I have identified that the way we leverage the information in this inherently imperfect data consistently develops predictive models that point to a relative outperformance. And that's frankly all we can expect the algorithm to do.

Speaker B: So can you give me an example of something that's perhaps surprising or not? That the algorithm is identified in terms of a value driver that has predictive potential, predictive, uh, power.

Speaker A: I'll give you an example from my early research, but then I have to um, uh, caveat this a little bit with respect to machine learning. Um, a couple of years ago a co author of mine came to me and he stumbled, ah, across a great data set about religiosity. He had data county by county in the United States about how many people go to church regularly. And we discussed this and uh, I wondered why on earth could there be any link with private equity? Well, religiosity, according to social sciences, is a proxy for risk aversion. Apparently you're hedging yourself for what may happen to you in the aftermath by going to, uh, church. Not my opinion, but research out there. So we designed a study where we linked private equity both with respect to the volume of private equity activity and the outcome of particular bankruptcies, to religiosity, linking the religiosity in the place where the investee company is headquartered to those investment outcomes. And we controlled, as academics do, for a gazillion other things, the industrial base, the population, county by county across the US we found indeed that religiosity matters in the sense that if a, uh, business, all else equal, is located in a highly religious county, it is less likely to undergo a, uh, buyout. So you have lower buyout activity. At the same time those buyouts that happen have a substantially lower bankruptcy rate. So how do we explain this? Well, if you have business founder, you're presented with a private equity option and you assess its risk, you're going to be very prudent and you're not going to go for A deal that allows you to cash out, um, and possibly makes the company suffer and go into bankruptcy. Because next year in the church bench, you're just sitting next to the person that got fired after you cashed out. So this will drive a higher bar with respect to hesitation to do the buyout. And then at the same time, it's going to be sure that those deals that do happen are much less likely to fail. So you find, uh, sometimes with empirical work, a few unexpected links. Now, with respect to machine, can I just ask. Of course.

Speaker B: But you need the hypothesis in the first place, don't you? How did you come up with that, for example? Because that's really out there. I would never have thought of that in a million years.

Speaker A: I mean, discussion, brainstorming with a co author. Academics combine data sets, have some intuition tested on the data. Sometimes you find something intriguing. Now, in the machine learning world, it's a little different. First of all, there are these broad two areas of machine learning, what's called supervised and unsupervised. Unsupervised is basically you throw the data at the machine and you expect it to figure things out, um, without any hypothesis, without any design of the particular measure that you go in with. For the most part, that's not what I'm using. I'm still in the world of what's called supervised machine learning, where a human actor has identified certain measures. Come back to what I said earlier, measures that are likely reflective of distinct skills, and then you create a whole range of those and you let the machine figure out not only which one matters, but how they matter in their interactions. And that's important. And that brings me to the point where I cannot really say, what did the machine learning pick up that you didn't expect, Oliver? Because, um, if I, the current model have 99 different. Those variables, they're called features in the language of machine learning. Now, 99 features, I have 99 factors, all of which could have some impact on the performance outcome, the return of the next fund. And they don't only influence that performance in isolation, but in interaction effect. And so these interaction effect may be nonlinear, so you get quickly to a level of complexity. 99 factors. You cannot possibly understand anymore, the why of the causal linkages between them. Of course you get a feel. Of course performance matters. Of course change in scale matters. I can measure this one variable at a time. But it's important to look at the, uh, final machine learning model. It's not one model. I have, like for every prediction, I have 60 different machine learning agents, each of which predict the world from a different angle and then there's some method to combine those. This has inevitably complexity that I cannot possibly understand nor explain to someone. So at some point you have to trust the machine and you only develop that trust with this very systematic and honest back testing that I referred to earlier.

Speaker B: It's this black box component of AI, isn't it? Whereby there seems to be an inherent trade off between how transparent something is and how powerfully predictive it is. You can't have it both ways, inevitably.

Speaker A: So. And the way, you know, for a possible real world application of this we got to kind of build trust is to say, well, let's start with a simple model with a few indicators where we would agree that they probably matter. See that the model does the trick while we can still unpack it well, then add more indicators, add additional complexity and, and at the end of the day, if you see step by step that, well, the model still works, it still works. At some point you got to let go and say, I cannot understand the details anymore. But well, I understood it for the first six steps. So if I get the same robust results on the backtest, it probably will continue to do meaningful things. But you got to be very careful in designing the backtest, right, and avoiding the whole range of traps that are out there in order to basically create false generalization. As I always warn people, the models are indeed so powerful. If you don't pay careful attention how to design the training and validation, artificial intelligence turns into very powerful artificial stupidity. And there's lots of AI washing, machine learning washing out there by people who will try to blind you with methods where if you know what's going on, um, they've been probably more cheating than uh, others.

Speaker B: So it's going to require for a long time, uh, human intervention, intelligent human intervention in order to train these things and manage them and update them in

Speaker A: the design of the model calibration phase, without a doubt. Um, and that's why I'm passionate about my own involvement in this area because I do believe that the deep understanding of private equity data, exactly what I've been working on for over two decades, um, is very helpful in order to maintain a kind um, of intuitive feel for whether the results are plausible or the model is just hallucination. I'll give you one example if I may. We uh, spoke about the application of this to the primary fund selection with my US Pension fund example. There's another application that's actually even more interesting with respect to the Backtesting results, and that's in the secondary market, specifically on LP Diversified, where the challenge is basically that you're presented with an opportunity, uh, to buy a primary buyout, uh, fund in the secondhand market. And you ask, okay, the fund has a net asset value of 100 million. What do you want to pay for it? And the incumbent existing approach is to say, well, Great, I got $100 million worth of NAV. Let's look at the underlying businesses one by one, focusing on the large ones. Do a value prediction, discounted cash flow, and figure out a value that in a sense is second guessing the GP who's running the fund and owning those assets. Um, I come at it from a different angle, not least because my research has confirmed what academic colleagues of mine have written about in the Journal of Finance, that the market for those, uh, secondary stakes in private equity is inefficient in the sense that the price paid doesn't, uh, correlate in a particularly strong way with the ultimate go forward value of these fund stakes. So the approach to do the bottom up valuation is imperfect to say the least.

Speaker B: To the up and the downside, it just doesn't correlate.

Speaker A: Absolutely. It doesn't correlate. It's all over the place, literally. There are mispriced opportunities, lots of mispriced opportunities in both directions. I have a different approach to those, uh, valuations because I rather ask the question, okay, here's 100 million of nav in the hands of a certain fund manager. First of all, what do I know about this fund manager that makes me understand whether the 100 million NAV is rather optimistic or conservative in terms of valuation policy? And then what can I understand about the likely ability of that fund manager to add value to those assets going forward for the next couple of years? Let me give you an example for the ladder. The fund manager has been raising this particular fund that I'm considering to acquire six, seven years ago, they raised a subsequent fund five times larger. They're probably very busy deploying that larger fund. So relatively little capacity left for value creation of the earlier fund. Or one fund manager not to pick on anybody. But it's well known out there, a firm called Crestview Partners, they used to do oil and gas. The older funds have a good share of oil and gas exposure. Let's assume that secondary stake that I'm looking at has 50% oil and gas assets in there. Crestview no longer does oil and gas. So probably the people who know oil and gas are either retired or demotivated. So we would agree that all else equal, that Crestview Fund with its oil and gas components is going to be evolving less favorably in terms of value appreciation going forward than the same asset in the hand of a GP who still does oil and gas. And so you're nodding and uh, I would hope that if I explained one by one each of the 99 different features of my models, you would go like that probably makes sense. How powerful is the indicator? We don't know. How do they interact? We can't explain. But the starting point in the supervised learning is always based on the understanding of how private equity works, how it adds value. And how do we measure proxies. How do you derive proxies of those drivers of value creation? From the data we have.

Speaker B: Uh, right. So the secret really is how they interact then, because a lot of these things just sound like common sense. Like I would have looked for this. When you're doing a deal with someone, yes, you do a spreadsheet analysis, but you also look at where they are psychologically, I guess, and culturally and all of those things. But it's knowing how to weight that and knowing how that it interacts with all of the other factors. That's what the machine learning.

Speaker A: Exactly. My role is the human, uh, scientist who's designing the machine learning setup is in the identification of those features. And they're not only performance based, not only strategy based. I pick up things like team dynamics from another research, uh, project. I've been tracking systematically every single investment professional working for a major private equity firm since 2003. 20,000 people I know how they move from one firm to the other, how they got promoted, how the teams are diverse over time, how people leave, where they go to. So I can basically build the profile of human capital of a private equity firm over time and correlate this to, uh, growth in aum, um, for example, we've also been bringing in more recently fascinating data on social media public sentiment, uh, because we know both the GP and or the portfolio companies in that fund may be in the news for something. Some private equity fund managers are on the news for whether or not they pay taxes or who they party with or who they got divorced from and how they celebrate their marriages, their wedding. And all of this may, according to what I hear from my friends who in the secondary business that may trigger sales activity in the secondary market. Now are these good or bad traits? I don't know yet. But, uh, we're measuring social media public sentiment over the last decade link this into our model and see does the model get better if we train on those additional indicators?

Speaker B: What pushback do you get from people when you're explaining the predictive power of your model?

Speaker A: There are a couple of very serious obstacles to taking this to the next step, because as I mentioned, I've been kind of in stealth mode for this for six, seven years building out the algorithm, and for the last two years I see it actually works. The cat catches the mouse. It may not be the only cat out there, but reliably it catches the mouse, it points to outperformers. Um, in talking to industry practitioners from that area, there are two very natural, understandable hesitations that come up in addition to the general inertia that people don't like to change. And the one element is of course, coming from lots of the incumbents in the private equity market who would say yes, but, um, the human intuition would pick up something very similar. We look at all these factors already. We just don't have the algorithm. But based on our experience, we can basically factor them in as well. And then I refer to the aforementioned, uh, study in the Journal of Finance, and my own work in the area was as well. If you, if this is what the industry does, why is pricing so inefficient? Because you would expect that then at some point somebody would buy assets only at the price at which they should be traded in the secondary market.

Speaker B: Because some people have good intuition and some people don't.

Speaker A: At the very high level, you can also look at the data on the returns of the real existing scale players in the secondary market. Um, it's no secret. The performance of those large private equity shops over the last, uh, uh, 10, 15 years with the mature vintages is all very similar. The performance outcomes all are very narrow band around 1.6x, 1.65x return, very similar to one another. So that tells me it's unlikely that any of those players has cracked the code and generates dramatically different performance from others. The other major obstacle if I take this forward and says, okay, how do we actually take this out of the laboratory and build an investment program around that is because of the very unproven nature. Many people are influenced by what you know and what you learn correctly from public markets about all these backtests, where people come to the market setting up a new trading program, a hedge fund program, the fantastic backtest. The problem is as soon as you deploy, uh, doesn't work anymore. Why does it not work? Well, the opportunity is so small, it's quickly arbitraged away because if you start trading in public market, you drive prices. And if you start trading in public markets, other people see what you're doing, they follow you and then the opportunity goes away. Well, in private equity we don't have that problem. These are ah, discrete traits. Think about secondary traits. Uh, these are discrete opportunities. Uh, first of all, people don't immediately see what we're doing and um, what we do does not necessarily drive up the prices because the market dynamics are very different. So I see a much stronger ability to translate the back testing results into what is likely going to happen if and when we start to deploy capital on that algorithm.

Speaker B: So going back to the first point, would you expect to see a kind of a one off improvement in someone's performance if they had, let's say, cracked the code? Is that what you're looking for?

Speaker A: Yeah, you would basically have the observation that one of the large players is better at pricing, which means they don't buy things that are too expensive and they buy lots of things that are relatively inexpensive in the market. And thereby you would see that cannot they crack the code. And don't get me wrong, there's great dispersion in returns, uh, on secondary funds. My ranking of the best performing secondary funds in private equity just came out a few days back. And of course there's huge performance dispersion. But the strong outperformers, as I've uh, shown in that study, are relatively smaller players with a very distinct niche focus. So if you look at the scale players, those managing 5, 6, 7 billion plus per fund, those are the returns that are very narrow.

Speaker B: Right. And finding the smaller players is harder because one strategy is just to invest with brand name players. But if you're saying that the smaller players are doing best, then you're actually missing out. You're not going to get alpha. And so it's easier to get it if you're systematizing this.

Speaker A: Exactly. And if you want to create what's the holy grail, a scale player at substantial outperformance, you have to leverage the ability to algorithmically identify those mispriced M opportunities that you want to acquire at substantially lower cost. Because the other element of applying machine learning is of course that it kind of is a perversion of the economic principle. It used to say you cannot do more for more. Like uh, better performance at lower prices. That's not possible. That's exactly what machine learning allows us to do. We get better decision outcomes and we can take out 80% of the investment cost in order to identify those opportunities.

Speaker B: The thing about cracking the code to a complex system is often that it works and it works and it works and it works and then it doesn't. Mhm. Because these systems are non stationary and as you said with the public markets, they kind of react to observations about them. What's your degree of confidence that you have cracked it? I mean, how long is your kind of time series?

Speaker A: So we've been training the model on over two decades of private equity data because private equity is super challenging as you need that long time lag between an investment decision and observing performance outcomes can have at least three years, typically five. Now think about the model building. We have data, uh, through late 2024. So I can only judge decisions that took place in 2019, 20. Now to train a model that I want to apply in 2020, again, I built the model in 2019, unobservable performance data in 2019. So I got to go back with decision until 2014. Mhm. So I got at least a decade of data to just observe performance and then to look at 2014 decisions. Any historic data on those opportunities from 2014. So you need very, very long time cycles, which limits the ability, uh, of actors that could potentially build something like that to those real world investors who happen to have lots of historic data or folks like myself who gathered this for a long time for their academic activity. Um, at the same time, um, you then need to be very careful. At what point does data decay? Um, the deals by CDNR and KKR from the 70s are fantastic for the history books, but the success drivers there are less likely to be relevant in today's market environment for obvious reasons. So in designing these models over time, my team and I have been working a lot on what's kind of the decay function, on how long data series do you train and at some point do you let the old data decay as it becomes naturally less and less relevant? That's all of what kind of led us in a very iterative process over many years to zero in on. Okay, here's the design of the algorithm that predicts performance in a robust fashion. And it's actually very interesting. In all my backtesting, based on the years I just mentioned, what do I do? I train on private equity data roughly between 2008, 9, 10 and 2014. 15, 16, 17. I apply it in the late teens and I observe performance until today. So in the middle of my backtesting I have arguably the biggest secular shift of private equity since the gfc because I'm Training on data pre product Covid Ukraine, however you want to call that shift, and I apply it to very different data M. And this cat still catches the mouse, which gives me additional confidence that, well, if the algorithm and the predictive power kind of survives that shift, can I imagine a shift happening in 26 that makes the model entirely non applicable? Of course I can. Would this shift probably render any other level of human experience and expertise in private equity investing equally, uh, irrelevant? Probably so. And the algorithm has the advantage that of course it learns in a systematic fashion, so it picks up very quickly changes like you have. We have now of course, especially in the secondary market. We got continuation vehicles, we got nav landing, we got different sources of liquidity and slowdown of other liquidity events. The model being retrained every quarter picks up these things and neutrally and objectively without any human bias integrates them the prediction going forward. Whereas it's much tougher for the human actor to throw out something that they may have learned a decade ago, because the world changed in recent years.

Speaker B: So the decay rate can actually sometimes be very rapid. You can switch from a constantly declining interest rate environment to suddenly a high interest rate environment, but the algorithm will adapt to that relatively quickly.

Speaker A: It will adapt to this more quickly and more accurately with less biases than the human actor. But it will never be perfect. So I have to rely on the model to pick up longer term trends of drivers of performance. Will experience, will skill, will the ability to have your own deal flow? Will the ability to buy at low prices be applicable in a relatively more or less high interest environment? Probably yes. And it's those type of factors, rather than very short lived macroeconomic drivers of return that are tried to put in the model and that inevitably training over long cycles the model will put in. But let me emphasize, there's a need of course not only to have kind of human intelligence go in the process of designing the model, also in the application of the model, my metaphor is not the kind of Google or WeMo self driving vehicle where we'd all sit in the backseat reading newspaper and let the car do things, but it's more that of your Mercedes, uh, Tesla, Porsche drive assist system. There's still going to be a person in charge with a hand on the steering wheel, but with the role importantly to avoid the accidents. Because the algorithm at some point in time may have wrongly recommended investing in Madoff or Abrash or whoever you have. Because there were some things going on that the algorithm could not possibly have known. And it's important if I discuss with um, possible investors anchors partners to implement this algorithm, I always remind them that we would want to ideally design a system where we integrate human experience with the prediction of the algorithm. But importantly in the sense that the human intervention can only adjust expectation downward. The algorithm says this is going to be a 22% performer. The human can say I still not going to buy it because I know the co founder adjusted a health issue or the key to people from the team just left something the algorithm couldn't know yet. We adjust this downward to 15. The algorithm should not be. The human should not be able to say oh, the algorithm says 12. But I know this is Joe's firm, it's a fantastic firm. We still do it expecting tranny because I know better. And that protects us through the use of the algorithm for the whole range of cognitive biases that often lead to bad decision outcomes for the human.

Speaker B: Yeah, that makes perfect sense. Another analogy would be commercial um, airliners which pretty much fly themselves. But like that film Sully where he lands on the, on the Hudson. Yep. You want someone just cross checking it all but you're managing the downside, you're not trying to optimize it as a human.

Speaker A: Exactly right. Exactly right.

Speaker B: Yeah, that's. So the implications of that are that um, asset allocators as a type will be looking at people who are perhaps a little bit more systematically minded, a little bit more conservative in their temperament. We're not going to be looking for people who are seeking alpha, but they're more like caretakers. Is that fair?

Speaker A: Um, potentially. Um, but I think uh, the good news for the, for those people in the asset allocation business today is that they just got a really, really powerful tool to look at. And it's just like, you know, um, tumor detection. You don't have the AI only you don't have the doctor only, you have the doctor who looks at the picture and the AI pointing to certain elements that may look wrong. So I think the skill set of the human is going to be highly applicable. They can just take better decisions. Going back to my analogy with the pension funds, they're brilliant people, certainly working at those 17 pension funds. But uh, um, if instead of, or in addition to whatever advisor gatekeeper they currently use, they have the algorithm, they and their teams can drive substantially better outcomes for the people in their pension system. And these are important areas of improvement for private equity.

Speaker B: What does the uptake of this type of tool look like?

Speaker A: I mean that's the big challenge that I meant previously. It's inertia and it's the frankly, um, willingness of the human to admit that the traditional way of investing has, uh, so much room for improvement. I had several conversations and some of these are ongoing with people who do secondary fund investments, for instance, for a living. And then I show the backtest and yeah, okay, Oliver, but now tell me, what's your view on the pricing of these 10 funds today in the market? I give them the pricing of the algorithm and said, oh, no, no, no, no. This is. This, this, this, this, this, it's different. Okay, what do we do? We now set a date and we're going to meet in 2029 and we're going to bet a, uh, bottle of wine in order to. Who's going to. Who was right or not. They um. As I cannot open up the algorithm and explain exactly why, ah, there's a certain recommendation, there's a huge hesitation to trust the backtesting. In particular, if you've been in the business for the last 10 years of taking these decisions yourself. And I understand these limitations. But there's also ample literature on the fact that, yeah, it's the inventor's dilemma, right? The incumbent is very rarely able to, uh, be ready to disrupt themselves. It's often a third party new entrant that brings disruption. And then everybody else tries to copy and rush. And uh, I foresee a little bit of this for private equity. I discussed this years ago already, early days in this initiative with somebody, uh, very high up, one of the larger, um, asset allocators, um, in private equity in the world. And this gentleman basically said, yeah, Oliver, if we, private equity are not going to be able to do this, we're going to have the Googles and the metas of the world come in. They understand data, they understand the algorithm. You're just going to have a product that blows us out of the market, including the entire fee basis. Because saying it's been out there for a long time and it's kind of the working title for me for this project, it would be fantastic to build out the vanguard of private equity by bringing science to secondaries and designing products that have widespread appeal for investors in private equity, institutional and mass affluent, at dramatically lower cost of the investment. And that's out there. Whether it's going to come from a total outsider like a Google, um, from a startup initiative like the one I'm pushing, um, I do not know. It's unlikely to come from all the incumbents, but I know there are also several there who do great work in this area. But I have no doubt that if we sit down here in five years lots of this will have happened. Who does it, how fast does it come? Different question.

Speaker B: The low cost aspect comes just from the fact that the whole investment process becomes more predictable, uh, and less manually intensive.

Speaker A: Well, if you ask me today, what's the price on these 865 different fund stakes that could trade tomorrow in the market? It takes me to open up the computer and show you one spreadsheet and the cost to kind of generate this. Based on the models that we calibrated they would have tripled or calibrate but that runs on your standard PC and that crunches for a few hours every quarter. So that's the replacing the cost that you have. If you want to price a portfolio of say 80, 90 different fund stakes, you're going to value 2, 300 different companies from the bottom up. And that takes a lot of time from pretty expensive, highly skilled resources that you can better use for other purposes.

Speaker B: And the guys like you who are currently in the market. Isn't data your real moat? It's harder for uh, big tech to come in in a couple of years. I guess there are techniques where they can fill in the gaps but.

Speaker A: Well, I mean data is part of the uh, element that you need in order to play here I would add is the understanding of the data. Because data and private equity at some point will become much more of a commodity. I observe this space of course as a user of data. But you got big incumbents who have lots of data. You got new entrants who make great strides into getting visible with private equity data that they just generate based on scraping the Internet with very powerful large language models. And I believe this trend is going to continue. So data as the commodity as a barrier to entry in private equity is going to become less relevant. Understanding the data, but I'm biased there. Understanding the data probably remains a great limitation, um, but I wouldn't feel safe if I was running an existing private equity firm, um, to kind of ensure that I can still command the type of fees I'm commanding for the next decades purely for the allocation of capital. It's a different story for a primary biage fund. But if you run fund, a fund, secondary fund, you're running a long only investment platform, um, and there's room for disruption also in the pricing level with by the way an interesting angle to it. And I had this as feedback from a large pension scheme where anecdotally I kind of discussed that research about a Year ago, and the feedback was all over. If you're actually putting that fund together, we should anchor you. Not only because that may be a really interesting, um, investment profile, uh, in terms of return perspectives you provide, but we should anchor you because if you're successful, you're going to bring down the cost of liquidity in the secondary market. Right? Because you allow us. You basically do arbitrage on the pricing, you deploy more and more capital, you bring efficiency to the market. At the same time, you can do active trading. You will sell what is mispriced the other way around. So eventually you can bring down the cost of liquidity to private equity from what, 10, 12 cents on the dollar to 3, 4, 5. And we all know what this will do to the volume of secondary activity. And this particular pension fund says, well, forget about the fund returns. The driver for us as an LP is going to be the reduction in cost to an active portfolio management, which we would love to do, but we cannot because it's oftentimes prohibitively expensive.

Speaker B: Yeah, see, it's very easy to get worried about AI and it's taking all our jobs. But one antidote to that is looking at how it takes away the jobs that make things worse. And I would say if you're a fan of private markets, what you want is to see the fees come down that will benefit everyone, really. So, uh, that's a great positive.

Speaker A: I very much believe so. End. I mean, there are lots of analogies here at conferences like neither the integration of the Abacus, nor the calculator, nor Lotus123 and Excel have taken away investment banking jobs. Right. There are many more people today working in those industries as they were decades ago before these inventions. Just because more powerful tools lead to more investment, better investment outcomes, and then eventually better product offerings in that space. And on that particular point, um, I mentioned the algorithm and the training to basically have a forecasting model for those different investment outcomes. I can train different models to deliver different outcomes. Now, uh, you're an investor either in primary funds or the secondary market. What do you want? Do you want annualized return? Do you want absolute returns? Do you want alpha outperformance of the stock market? Do you want loss avoidance? Do you want a quick payback period? Now, uh, you pick any of the five or any combination thereof, and I can train an algorithm that delivers specifically based on the back testing either one or any combination of those five factors. So you can think about a product design for investment product and private equity. Here's something for the Family office. Here's something for the insurance company, here's something for the pension fund, here's something for your 4401k, which has amazing opportunities because certainly today there may be products out there that deliver one better than the other. But nobody's able to show in back testing that, hey, I trained this to deliver this outcome. And I, uh, put exactly those type of traits in the product that deliver those outcomes with a high likelihood. And just a starting point to see how much power we gain in terms of product differentiation, diversification in private equity, making the asset class more easier to navigate and better manageable, hence appealing for so many additional asset owners.

Speaker B: That's funny because, um, my final question, I was wondering whether to ask you it or not, because I thought it was a little bit out there. I was going to say, you know, down the road, do you foresee a situation where these tools actually allow for portfolio construction and looking at things like diversification, but you're saying, not down the road. You can do it right now.

Speaker A: I can do it right now. Portfolio construction and active portfolio management. Right. Because you would think about a private equity portfolio. You can basically say, I do, for the most part, trades in the secondary market. I buy whatever I find mispriced, I sell whatever I find mispriced in the other direction or I find correctly priced to rebalance the portfolio, I'm buying, buying, buying. I'm getting too much industrialist exposure, I got to sell something, leaving less money on the table to rebalance the portfolio and run through this periodically. Also, if there changes, for instance, in the needs and requirements of the plan, you have an endowment, um, change in regulatory context in the US Comes along, you know, Exactly. Not only that you have to sell a couple of hundred million, but how to reshuffle your private capital or private equity exposure in order to best manage your commitments going forward. And all of this can be done in a much more efficient and scientific and robust fashion thanks to those tools.

Speaker B: Great. Well, it sounds like we need to get you back on in something like late 2028 to see how things are going.

Speaker A: I'll m mark my calendar with pleasure.

Speaker B: Oliver, thanks very much.

Speaker A: My pleasure.

Speaker B: Thank you. You've been listening to the Fun Shack podcast. Wherever you're listening or watching, please click follow or subscribe. That's all I ask.

Speaker A: Thanks for listening.