Artificial Intelligence: A Revolution of the Investment Landscape (w/ Raoul Pal and Trevor Mottl)
RAOUL PAL: Trevor, great to have you with us on Real Vision. You were recommended by a mutual friend of both of ours, Mike Green, who said, listen, you've got to sit and talk to Trevor. And once I started looking into your background, I realized we share almost an identical background, slightly different timings. So talk to us a bit about your career, where you came from. TREVOR MOTTL: Sure, so thanks for having me on. This is wonderful. I started my career at Credit Suisse in the exotic derivatives team in London, and then I moved to New York and spent more time in exotic derivatives. And I made a stop at Goldman Sachs, where I structured equity derivatives through the financial crisis. And after Goldman Sachs, I went to a firm called Susquehanna International Group where I ran macro and derivatives strategy. And that was a great experience where I was exposed to really thoughtful professionals who were at the cutting edge of making markets in both options and ETFs. From Susquehanna, I moved over to the buy side and worked with Pierre Lagrange at GLG Partners, and it was also a wonderful experience. I met a lot of wonderful people. And it really changed the way we worked on risk management at the firm. RAOUL PAL: Were you in GLG London or New York? TREVOR MOTTL: I was in GLG London. RAOUL PAL: Were you in the same office as I was? Were you in 1 Curzon Street? TREVOR MOTTL: Absolutely. So the long rows of desks and, you know, cramped office space but a beautiful building and a beautiful location. RAOUL PAL: Yeah. So then after GLG, where did you go from there? TREVOR MOTTL: So I moved back to New York, and I ran long-short risk for Balyasny Asset Management. And that was a phenomenal group of people. I have enormous respect for Dimitry and what he's built. I'm really happy to see his performance. And that really demonstrated what a high-caliber, integrated, quantitative, and fundamental firm looked like. They did a really good job bringing the quantitative pieces together with the fundamental pieces, and they continue to do so. I think that's a large reason for their current success. So I was really happy to see that. RAOUL PAL: They retooled as a business, didn't they, a few years ago? I guess this was part of that restructuring where they changed into a much more modern firm instead of risk management. TREVOR MOTTL: I think it was really the integration of the data science. And there's a great guy over there, Carson Boneck, who's done a great job leading out the data effort, as well as bringing in and focusing the long-short investors. They did a really nice job, so I'm really happy to see their success. RAOUL PAL: So how, then, did you get into Lazard? And talk us through what you're doing at Lazard because this is super interesting. TREVOR MOTTL: Sure, so about three years ago, I saw the opportunity to really bring tools of machine learning, AI into the investment process. One of the things I realized is good investors tend to have very repeatable processes. And the tools for AI I had developed to a point where you could bring them in and start integrating them to do jobs that investors really found challenging. So I moved to Lazard about two years ago, and I run a group in Palo Alto called Lazard Labs. And our focus is to bring cutting edge AI and data science into the investment process and develop investment funds, so hedge funds, long-only products for our investors. RAOUL PAL: Let's dig into this AI because this is a massively misunderstood or just not understood at all thing. Took us through how you conceptually see it first then we can talk a little bit more detail about how it actually works in practice. TREVOR MOTTL: AI is a really broad catch-all term. And all AI really is is more advanced analytics and modeling. So the advantage of AI is we can model things we couldn't model before, and we can analyze data that we couldn't analyze before. So a great example is textual data. AI is made and analyzing large amounts of text data, let's say all of the 10-Ks for this year. That text can be analyzed if you ask it very specific questions. So one could load up all of the 10-Ks into a database, write a program to extract let's say the length of the risk disclosure section of the 10-K and then see how that's evolved for a given company over a period of time or compare across industry groups the length of a risk disclosure section. That's a very simplistic application, but that's one that a fundamental investor couldn't do until they had the tools to actually do large-scale textual analysis. That's an example. RAOUL PAL: What kind of information does that give you, then? I believe it's part of a much larger data set that you're looking for something. Are you spitting out trades based on that, or are you doing analysis and research based on these kind of things where you would then look in further? How does this all work in that particular example you go? TREVOR MOTTL: So that's a simple example. That's sort of a toy example, but it gives you some context as to how natural language processing could be helpful. What we think about is a broad array of these types of signals. So what we're doing is asking questions of a computer, building systems to go and answer those questions using AI tools. And I think this is the key takeaway, which is AI doesn't ask the questions for you. Investors still have to ask the questions. And I think good investors ask a lot of questions, some of which we can answer, some of which we can't. Doing a discounted cash flow model is a question we can answer. We know how to do that. But going and analyzing the sentiment of a earnings report is something that's very challenging for us because we'd have to go and outsource that and manually score every single earnings call. So AI gives us the ability to go and broadly analyze things like earnings call sentiment, news sentiment, changes to risk disclosure, linkages between companies. So when we're disclosing competitors in a 10-K or a Q, we can actually go and identify those competitors, build a graph, and build maps of companies to understand how they compete, as well as understand how they're interrelated, which right now is really important because we're talking about potential bankruptcies on the horizon. And one bankruptcy can lead to another bankruptcy. If a supplier goes bankrupt, that can impair a producer, and that producer could go bankrupt because they're not getting the supplies they need in time. RAOUL PAL: So when we talk about these data sets, I think people are shocked at the amount of data sets available. You and I talked last week or so about Orbital Insights and the kind of data sets they've got. Talk me through some of the data sets that are available or somewhere that are in private hands, because I know, for example, JP Morgan has real-time banking data anywhere in the whole United States, which is an extraordinarily powerful data set. Talk me through some of these data sets out there because people don't really understand the magnitude of this. TREVOR MOTTL: There is an enormous amount of data that's now available. Much of it's available via API. So integrating those data sets into a relatively simple Python script is very easy. So data that's available via API runs the gamut from CIC data, most of the economic data that you'd get out of the FRED database. We can scrape or pull down KNQ data, earnings call data. There's tickerized news data now that's available from both Bloomberg and Reuters. And while that doesn't sound like a big step forward, it's actually a really valuable tool and data set because Reuters or Bloomberg have effectively taken all the news related to a given stock and classified it for you in one specific feed. So if I want all the news for Microsoft, I can going to get that, and that's true for any other company. Obviously, I'm beholden to Reuters' or Bloomberg's classification, but doing that on my own is rather time consuming, and they've done that for us. So that's a wonderful addition from a textual perspective. Additionally, there's sentiment data available, and then there's an entire swath of alternative data. I call it the alternative just because it doesn't really fit into the normal data set that typical investors are-- RAOUL PAL: --Bloomberg Terminal, yeah. TREVOR MOTTL: Yeah, exactly. So Orbital Insight is a really interesting company that have produced high-quality geospatial data, and there's company-specific data. So they've mapped things like all of the sites of big box stores, all of the locations of new construction. They've mapped how high floating top oil storage containers are. So they've done some really interesting work. RAOUL PAL: They've even done corn by looking infrared imaging to find out how advanced the corn planting season is, and this is all done granular satellite data. It's truly extraordinary, right? TREVOR MOTTL: It is, and the company really found its roots in doing agricultural research, looking at patterns, color, infrared, the array of imaging on planting to see how far along plantings were, how many acres had been planted, et cetera. So there's an enormous amount of data you can extract from the sky. And they do a really good job of not only doing data extraction, but they process the data for you and make it available as an index or as a feed. So all of the compute-intensive work is outsourced to them. And there are many more data providers that are effectively outsourcing the compute and providing prepackaged signal data, which an investor can integrate into their models. Orbit will be one of them. RAOUL PAL: And how do you then go about it because there's so many data sets, right? So first, you're trying to get as much data then you need to answer the questions. And I'll come into all of this because I've got a ton of questions I want to ask you. But data sets, so you've got Google, social media, all the newsfeeds, all of the price feeds, all of the 10-K data, the alternative data, the credit card data. How do you go about even thinking about what are the data sets available and which ones you need for what you're trying to achieve? TREVOR MOTTL: I actually go the other way. So rather than going and hoovering up a whole bunch of data and looking for some random signal in it, the approach that we take and one that's less challenging to conceive of is we ask questions and we test hypotheses using data that's available. So we'll make a hypothesis about some feature in the market. We'll go and find the right data. We'll build a feed. We'll build an experiment. We'll test it using a robust set of tools. We'll look at the output of those tests, and then determine whether or not that's a worthwhile pursuit. And I think this focus on questions is something that really is differentiated. RAOUL PAL: Doesn't that then come down to the human element, right? The quality of the questioning mind, which is pretty much your team getting together and saying, OK, these are things that we're observing that may be a hypothesis, and then we need to test it. I mean, that requires a lot of intellectual power from the team that doesn't lie within the AI itself. It actually is the human element. TREVOR MOTTL: And I think that that's a key piece. So that good AI today requires those questions. We build models to answer questions. We don't build models specifically to ask questions. And while there might be some people who are looking at unsupervised AI tools that are producing really interesting insights, we're really focused on having the AI as something like an analyst or augmenting the analytical process so that we're building better portfolios that are potentially more robust, that are more reactive to changes in market environment, that potentially are more adaptable to regime shifts and also idiosyncratic shifts in individual companies. RAOUL PAL: I'll come into a lot of that in a bit. I want to get an understanding of the different players in the space. So what makes Renaissance Capital, Renaissance Capital, Two Sigma, Two Sigma versus what you guys are doing? How is the space defined, in terms of what types of players are involved? TREVOR MOTTL: Well, rather than commenting on some others, I see a wide array of applications in AI across finance. Many fundamental investors have now embraced scraping the web to extract specific data sets, and I think that is becoming very commonplace across the long-only and hedge fund world. So going to specific websites or data providers, scraping data, pulling that into a database, potentially writing an application where you can pull that into excel, I think that's pretty standard. And that just becomes another feature in your model or enables you to track something that you had done manually in an automated way, things like car sales by VIN, things like that. So as you move to some of the more quantitative shops, and again, I'm not privy to what they're working on, but using larger data sets, using potentially AI tools to screen signals for forecasting Sharpe from those signals, that's another application that can be used. We talked about alternative data sets a second ago. Really assessing the signal quality of those alternative data sets because many of these data sets are relatively costly. And before someone spends a lot of money to acquire them and then has the recurring annual cost of both purchasing data but also maintaining and integrating that data into a process using machine learning and AI tools to evaluate the signal quality is an important step. So many alternative data purchasers are using AI and machine learning to assess that signal quality. RAOUL PAL: Fascinating. And it sounds to me that one of the misunderstandings people have is there is no black box. It doesn't really happen that way because the elements of questions. And sure, some others may be more advanced in creating AI questioning. Sure, I'm sure people are working on it. But that doesn't sound that there is a black box that comes up with magic answers, and he who develops the best black box wins. It's a much more nuanced process than that. TREVOR MOTTL: I agree with that. And the concept of black boxes, it's an amorphous term that's been thrown around for models and systems that we really can't explain that well. And there is an element of that in AI. There's an element of that in machine learning. But really, in any investment process, we're asking a question, which is if we do something, do we deserve to get paid for it? So really understanding what we deserve to get paid for, what value we're bringing. RAOUL PAL: What does that mean, deserve to get paid for? TREVOR MOTTL: So think about back in the early days of block trading. There are some very well-known investors who made a lot of money setting up hedge funds that effectively did block trading as well as desks that did block trading on the sell side. Block traders are effectively providing liquidity at a time of need to institutional investors or large investors, and they deserve to get compensated for the risk they're taking of making a large purchase or sale at a given period of time. And effectively, they're creating somewhat of an insurance product for an investor who wants to get out of a position. And they should get a return for that service that they provide. They're taking some risk. They get some return for taking that risk. For a factor investor, let's say a value investor, there's a different service that a value investor provides to the market. A value investor is typically buying stocks that are low-multiple stocks. And if they're building a long-short portfolio, a market neutral value investor, they're selling sure stocks that have high multiples. So they're effectively doing two things. Number one, they're providing capital to companies that have low multiples that have likely seen their multiples decline. So in a way, they're writing a put on low-multiple companies. And for writing that put, they should get some decay, some carry on that. They're also selling short companies that have high multiples, betting that those high-multiple companies either remain the same or go down in value. And they're effectively writing a call option on that aspect of the portfolio. So they should get paid that decay in both wings, and they should make money if there is mean reversion. You see multiple expansions in the ones you bought, multiple contractions on the one you sold, you should make money in the middle for taking that risk in the belly. And every single strategy likely has a rationale for being paid, and understanding that's very important. So for an investor, like these black-box strategies, there are aspects that you know. But there are also some features of the strategy that effectively you should be paid for. You're taking some risk that you should be rewarded for. RAOUL PAL: So you're using AI to help with portfolio construction as well in understanding how to construct the best portfolios? Because now you're talking about embedded within the bets that you take, there is an implicit risk-reward, let's say. Do you do that at the portfolio level as well and allow hypothesis testing in that basis, away from the standard VaR risk management models, that kind of stuff? TREVOR MOTTL: When I think about an investment process, I really break it down into four levels, and this is true really for any investor in any asset class. The first piece is universe selection. So what assets are you going to include in your universe? What do you pick from as potential investments? And we use machine learning tools to identify universes, and we use clustering mechanisms to identify assets that are similar. RAOUL PAL: What's a clustering mechanism? TREVOR MOTTL: So at the simplest level, GICS classification is, you could argue, a clustering mechanism. We're labeling companies by what they do in sector, industry, sub-industry groups. Well, there are some sectors, let's just say automotive, where there are some real outliers in that sector. And the properties of the company may be dissimilar. So what we do is we use clustering to find companies that have similar properties, similar revenue growth, similar multiples, similar volatilities, similar betas. We can go on and on. But there are an array of machine learning tools available now that enable clustering that potentially is more precise than the GICS classification method and adequately deals with outliers in industry groups and potentially reclassifies them with other stocks that have similar properties. So in universe selection, we're using machine learning. In stock selection, which is the second phase of the investment process, we use an array of tools to select stocks that we think will go up or sell short stocks that we think will go down. In the portfolio construction process, we're extending some of the algorithms that we're using for individual stock selection to understand, at a portfolio level, what's the likelihood of this portfolio achieving a given objective? Let's say we want a portfolio that produces absolute returns with a specified vol. We are using an array of tools to help to construct those portfolios. And then the last phase of the investment process, I call it drag, which is really how much does it cost to run a strategy. And that includes transaction costs, setup costs, a lot of the variable costs like hedging, et cetera, as well as headcount and infrastructure costs, very focused on the cost of buying and selling, so that transaction cost element. We're using machine learning to better understand transaction costs and model them. So at every level, we've got an element of either machine learning or AI. RAOUL PAL: And then within that, I know you're running different portfolios, so different portfolios have different strategies, but are you looking for mean reversal strategies, long-short strategies, momentum strategies, macro strategies? What is that universe of things that you do, and how do you decide what goes into a portfolio? Do you have a multi-product portfolio, or are they all broken up by different strategies? How does that work? TREVOR MOTTL: Sure, so we're primarily focused on equities right now. And we have long only strategies as well as long-short market-neutral strategies. And that's not exhaustive, but that's where we are right now. RAOUL PAL: Within that, also if you can talk a bit about time horizons as well. Because again, everybody thinks you're operating in the micro millisecond, but that's high frequency trading, which is a whole different business. So if you could talk a bit about that as well, it'd be useful. TREVOR MOTTL: We're really focused on longer holding periods versus high-frequency trading. It's a different business that they do. There's an enormous amount of infrastructure required to do that business, and there are some unbelievable players in that business. Where we really focus our attention on is the one month down to potentially one week holding period. But we're not looking for very short-term trades. We're really looking to pick up several percent on a trade basis. And obviously, we're diversifying our portfolio. So if we're looking to make 2% or 3% on a trade, we have a wide array of trades that we have on. And the turnover of the portfolios is dependent on market conditions. With more volatility, we typically have more turnover, lower volatility, lower turnover. But our holding period is very much aligned with typical hedge fund holding periods in the long-short side and typical long-only mutual fund holding periods on our long-only portfolios. RAOUL PAL: So when you're looking at, let's say, capturing the 3%, I presume you're then also kind of risk adjusting that 3% because 3% in Bitcoin is nothing, versus 3% in a 10-year treasury or 2-year treasury note, those are the extreme examples. TREVOR MOTTL: Of course. RAOUL PAL: So I guess I'm looking at it in those terms as well, the quality of the returns. TREVOR MOTTL: Absolutely. So like many long-short strategies, we're focused on producing good, solid returns with good, sharp ratios and on a risk-adjusted basis, putting up good numbers. And our models take that into account. So our models are aware of the beta of a stock. Our models are aware of the realized sharp ratio of a given security. They're aware of the realized sharp ratio of a given portfolio that's been constructed. So risk-adjusted returns are critical to us, like they are to, I'd hope, almost every other investor. RAOUL PAL: And do you test what is a risk-adjusted return? Because there's a whole other nuance within that because, you know, maybe Sharpe is not the best way of looking at potential future expected returns or whatever it is. Do you look at trying to understand that and dig into that side of it? Because many of us make assumptions about things that maybe don't hold up when you test them. TREVOR MOTTL: I use Sharpe as kind of a relatively straightforward example that most people discuss when they're speaking about risk-adjusted returns. But when we're thinking about risk on the downside, we have the typical array of features that you use to assess downside risk. But we also incorporate regime-type models, which we view do a better job at understanding the tail properties of individual names. And if I'll go back to some of the points I made on clustering, it's pretty interesting when you look at how stocks cluster through time. And in some periods that we view as very favorable to picking stocks, you see relatively numerous distinct clusters with stocks behaving in a relatively idiosyncratic pattern. But through periods of stress and turmoil, what you frequently see is aggregation of clusters, and you start to see pictorially see that risk-on, risk-off feature in those clusters. And in periods where you see clear winners and clear losers, you also see that. So if you go back to the Trump election, there were some clear clustering occurrences that you'd see between stocks that had paid high taxes versus stocks that paid low taxes, small caps versus large caps. So the features of those clusters became very distinct, and you really saw a bifurcation of the market in that period of time. RAOUL PAL: So when you see something like that, that sets up an opportunity to say, can we capture some alpha out of noticing these clusters or the changing clusters that maybe the market hasn't picked up yet? Well obviously, the market has in price but not cognitively. TREVOR MOTTL: So we use those cluster definitions as features in our model. So rather than responding in a human sense to it, we build that feature into the model so that the model understands, during period of where clustering looks like this, X, this is what the risk profile is. It could be. This is what the risk profile has been in the past. And when clustering is in a Y state, then the expected risk profile is likely different. So it's a feature that we use to train the operative. RAOUL PAL: What's the difference between AI and machine learning from your perspective? TREVOR MOTTL: It's kind of a relatively blurry definition. RAOUL PAL: Yeah, because some people use them interchangeably, and other people get angry at you if you use them interchangeably. So I'm just fascinated by what your view on it is. TREVOR MOTTL: So I think of machine learning as any model that typically goes beyond the linear, multi-factor models. They're an array of models that are deemed a machine-learning model, K-means, Lasso, they have names like that. And they're basically changing the way that you group data. You're basically treating your features differently than you would in a multi-factor model. The transition to AI I really see as when you move into deep learning processes-- so using convolutional neural networks, recurrent neural networks, LSCM, long short-term memory-unitbased networks, I think that's really where the term AI comes into play. And those tools are typically valuable in things like natural language processing. They're useful, in some cases, in doing prediction. But it's very blurry, and there's an edge in machine learning that you could argue is AI. And there's early stage AI that you might say is just machine learning. So I think of them as tools that go beyond the traditional toolbox. RAOUL PAL: Is this not a very expensive and difficult process to enter in, to start a business like you've started within Lazard? Because A, you need to select the data sets, decide what data sets, pay for data sets. You need lots of people to program, a lot of people to ask questions. It's not a one man in his basement hedge fund. I mean, you really have to have a hell of an infrastructure to get going. TREVOR MOTTL: That used to be the case. So I think this is why it's an interesting time to have this conversation. Over the past three to five years, the tools to implement AI have become A, open source. So Google, Facebook, Uber have open sourced some of their core tools that you and I can go in and use them for effectively free right now. In addition, Compute, which used to be really, really expensive, is now available for rent in the Cloud. So you can go and rent by the minute a $100,000 computer in Google's Cloud, and we could do that together in 20 minutes. So that's a big step. So you can kind of pay a la carte for Compute. RAOUL PAL: So you don't need a huge room full of servers and that kind of stuff. TREVOR MOTTL: Exactly. More and more people are not only well-versed in Python, but the libraries in Python made it much, much easier to go and implement some of these strategies. So you're right about the data costs. They are higher than some other businesses. But a lot of the data that you're buying is data that fundamental guys have already started to buy. So there's not really special AI data. The Compute is available, and many of the models and tools are free now. So the hardest part is building the team. And one of the reasons I moved out to Palo Alto was there are a lot of AI scientists out here. I kind of joke and say, if you want to make a movie, you go to Hollywood. If you want to build an AI business, you come to the Bay Area. RAOUL PAL: Those things have become more available, and it's not like high-frequency trading, which has this massive cost because it's an arms race. How crowded do these positions get? Because if everybody has roughly the same data sets and everybody has access to roughly the same kind of models and AI tools, how difficult is it to ask different questions to everybody else to avoid it crowding? And I also want to understand, what does that crowding in this space mean, because there will be some crowding, obviously, in times when everything changes, when everything goes to a correlation of one, which we've just seen? How does that play out in model-based stuff. TREVOR MOTTL: We're relatively early in AI, machine learning adoption into investment processes. So the risk of crowding I think is some time away. RAOUL PAL: So it's not like the CTA space or something like that where everybody acts at the same time? TREVOR MOTTL: Exactly. And the nature of the models and the nature of the tools actually increases the array of questions that one can ask. So if I'm building features out of text data from earnings reports, there might be an array of people doing that, but everyone might not have that as a core feature in their model. So until there's a large proportion of investors that have effectively increased the price of that risk and also increased the downside from that risk, then we're unlikely to see the effects of crowd. And I think if you go back in time, it took quite some time for quantitative strategies, the more traditional stat arb strategies, to get crowded. There's an amount of capital that needs to be allocated to these strategies before they really feel the effects of crowding. Now, I think from a crowding perspective, a more concerning feature is if our algorithms pick trades that are crowded. So we're getting into positions that are similar to quantitative traders, or we're getting into positions that are similar to fundamental long-short managers. If AI converges to those styles and those selections, then crowding becomes a risk. And for that, my point on drag, measuring the impact cost of trades, that's really where understanding the transaction cost dynamics comes into play. So if you're seeing increased transaction costs in giving names, increased transaction costs in given strategies that are well-trafficked, decreased returns of those strategies, that's a sign of crowding. And we can incorporate those changes into our model as a risk reduction signal. RAOUL PAL: So you don't think yet the space has an implicit short-vol position, which many of these do in the end? So once you get into situations where volatility explodes, many strategies-- equity long-shorts is a classic example, credit is a classic example-- suddenly, basically, they're short vol, and everyone's returns all fall apart. How has this space dealt with those kind of implicit short-vol bets that many of these portfolios end up constructing? I presume most of them have some sort of distribution of return profile within it. TREVOR MOTTL: So the strategies that we look at have historically had a typically longer vol profile and typically have done well through periods of vol expansion. So in the recent period of vol expansion, the strategies have done well. And in periods where you've really seen regime changes, they tend to be slightly more dynamic than some other portfolio constructions. So it's sort of a positive attribute of the strategies that we're building. There's nothing to say that you couldn't use these tools to build short-vol strategies, but we have a preference to monetize. RAOUL PAL: You're an option trader by nature. Well, some people are short-vol guys, and something some people are long-vol, and that's just how it fits down, right? TREVOR MOTTL: Exactly. RAOUL PAL: And so when you're looking at that regime change, what do you use, things like voli vol? Do you use price action, or are you picking it up from other data? What gives you the kind of signals that help you get that understanding? TREVOR MOTTL: There are layers within the model but at a very high level and a differentiated level, things like sentiment-- so broad news sentiment, sentiment for key names in the portfolio, key names that make up indices, and sentiment meaning positive news, negative news, positive social media, and negative social media, that type of thing. So that enables slightly more dynamic properties in the portfolio. RAOUL PAL: So Trevor, within this, do you think that AI portfolios, you only know your own, but in general, do you think they tend to be more diversified in structure than the typical hedge fund portfolio, just because of the number of inputs you take are probably broader than many of the hedge funds? Obviously, some people are exceptionally good at doing that themselves. But do you think that you end up with a more diversified approach? TREVOR MOTTL: I don't think there's a real rule for that. So typically, our portfolios look similar to long-short market-neutral portfolios when we're looking at absolute returns. And they look pretty similar to portfolios that are long-only portfolios for our long-only products. So we don't have double, triple the number of names. Our bet sizes tend to be less concentrated at the top and more uniform through the middle. Many long-short portfolios, many long-only portfolios have concentrated bets and then the long tail of bets. Typically, our portfolios are flatter in their distribution, from a bet-sizing perspective. RAOUL PAL: Is that because there's no human emotion in it? Because we all want to size what we think is our best bets largest. And this way, it's saying, no, we just want the return profile of each individual bet. TREVOR MOTTL: I think that's a fair comment. RAOUL PAL: Yeah, and some people are brilliant. You know Stan Druckenmiller? He's a genius at sizing active bets in whatever, but it's actually one of the things that most people stumble on is taking too much risk when they think the idea is the best idea, and it may not be. So that brings me on to the return profile versus other strategies. I mean, it's interesting because in-house, you've got a number of different strategies within Lazard Asset Management. But also in the world generally, how do you notice the return profile of these kind of more systematic AIbased strategies versus others? Do they have anything concrete, whether it's in alpha or whether it's in vol-adjusted returns in some way? What do you think? TREVOR MOTTL: Well, I think that the difference can live in the downside properties so really decreasing that tail property that you see in either in an overweight, underweight long-only book or a long-short portfolio. Because you're bringing in additional features, you're likely running a less crowded portfolio. And therefore, you're not exposed to the emotional unwinds that can be negative for some of the other strategies. I think we're really early on to make any concrete determination. And I also think the world of AI is diverse enough where this is a toolbox. We can use these tools to enhance strategies. RAOUL PAL: It's not one answer. TREVOR MOTTL: It's not one answer. RAOUL PAL: And so typically the kind of strategies that you're looking at, again, you're running a range of strategies, what kind of vol are they, and have you experimented with high-vol strategies using AI, which tend to be longer-term time horizon strategies? Because one of my beliefs is that it's much harder to compete with humans on a longer time horizon where you've got pattern recognition and that perceived living in the future and too many expected paths to get there or machines to deal with it yet. So have you looked at A, what kind of vols you have now, and have you ever looked at kind of longer time horizon and higher-vol strategies to see what kind of return profiles you get out of that? TREVOR MOTTL: We have looked at higher-vol strategies. Our strategies tend to be lower-vol strategies, both in the long only and in the long-short world. RAOUL PAL: Why is that? Is that the kind of five to eight vol or whereabouts in the vol? TREVOR MOTTL: So in the long-only world, somewhere between 75% to 90% of the S&P type vol. And in long-short, typically in the four to eight vol range. But obviously, that's regime dependent. The longer time horizon is actually pretty tricky. As you extend the time horizon, the signal degrades. The noise tends to take over. And we haven't built strategies that do forecasting on very long-term basis using AI. We're far more focused on that one week to one month period where there's the potential for the algorithm to have a variant perception, versus some of the other investors out there. RAOUL PAL: And that's one of the points I've raised to people who've been concerned that how do we compete with guys like you who have A, huge experience, but are asking all the questions, have the tools, have the deep understanding of all of the makeup of the markets from a very granular level. It's very difficult for people to feel that they can compete in a marketplace. And one of my answers to people like that is be in the longer-term time horizon because then the only competition is humans. So you have edge there, apart from people's brains against brains and luck. TREVOR MOTTL: I think that's good advice, and I'll elaborate a little bit on that answer. When you think of investing, I really break investing into three axes. There's a pricing axis. You can be better at pricing an asset better than other people. Can be better at pricing optionality better. You can be better at pricing yield better. It runs the whole gamut. You can be better at picking people for a venture capitalist, picking founders. The second axis I focus on is holding period, which echoes your point. And that goes from the very, very short term, the milliseconds in the high-frequency world, to the multi-year horizon investors in the venture capital and private equity world to the Warren Buffetts of the world who are sitting there holding things forever. And then the third axis I think about is liquidity. And it's distinct from holding period because there are some assets that are relatively cheap to get into and get out of. There are other assets that have enormous transaction costs, as well as you might not be able to find a buyer. It could take years to find a buyer for very bespoke asset classes. So if you look across those three axes, there are plenty of places to find an edge. There are amazing investors in every corner of that three-dimensional space. And I think investors need to pick their area in that three-dimensional space, understand their core strengths, and look to exploit those core strengths. And I know that's a very generic description, but the great investors that are featured all over the world have typically done that. They're focused on a specific area of that space and done that well. RAOUL PAL: Yeah, it's like Stan Druckenmiller did an hour and a half interview on Real Vision a while back. And one of the things he talked about was when he's taking a big position, he will only use basically bond futures, bonds, or currencies. Because in the end, he said I can get out in milliseconds 24 hours a day, and nothing else gives you that ability to manage risk. And his expertise is knowing when to apply massive leverage and into what is normally a lowball asset. It's super interesting. So just to finish off, I'd love to get your thoughts, because I'm a macro guy and that means I live in the future, where is this all going in five years? How do you see this space evolving? Because it's a very interesting, rapidly-developing space. And it'll obviously splinter and morph as it goes, but where do you think it's all going? TREVOR MOTTL: I think we're going to see a convergence of the quantitative and the fundamental. And there's going to be shades of gray throughout that. There'll be firms that are purely quantitative. They'll always be firms that are truly fundamental. And in the middle, I expect to see the adoption of more quantitative tools that influence the fundamental discretionary investment process where you see the investment process potentially becoming more repeatable, less exposed to negative emotional biases. You're able to integrate more granular data. I think our understanding of investing in markets will increase over the next five years because we are able to use these models, use this data, use this compute to answer questions. We never were able to answer before. RAOUL PAL: So that brings me up to one thing is right now, there's the world of human behavioral investing, which is what basically you're exploiting, the alpha from understanding that, how people react around earnings numbers, whatever it may be, right? Now, as AI comes up and becomes a more dominant force, where does the signal come from? Because you're then competing against each other. TREVOR MOTTL: I think that's back to the crowding point. RAOUL PAL: The Silicon Valley thing about AI is you get to a point where it takes over from human emotion. Not we're saying that AI runs the portfolios, but they're observing each other now and not observing the humans because there's less of that around this activity, which we've already seen with index funds, for example, have been very prominent in the market, versus the old value investors. TREVOR MOTTL: I think that crowding point is true for many of the innovations we've seen in finance over the years. If we went back to the '20s, we could go in, document all of the strategies that got crowded. if we started in the early '80s, block trading, at some point, got commoditized and got crowded. Stat arb got crowded then you saw quantitative strategies get crowded. Longshort market neutral had periods where it was crowded and became very aware of crowding to evolve. So there's not a right answer. There's so many degrees of freedom in investing that as an area gets crowded, it potentially creates other opportunities in other areas. RAOUL PAL: But do machines just trade against machines, I guess is the point? As we've seen retail volumes and other volumes falling, do we get to a point where it's betting against somebody else's machine, just philosophically? TREVOR MOTTL: Sure. In the high-frequency world, I think you see that today. And in order execution in trading volume, effectively, it's algorithm versus algorithm to effect trade. For a longer time horizon, I think that becomes a harder limit to expect because there will be variant perception amongst investors for many companies. If we all sat around the table to evaluate one company with all of the computers in the world and all of the data, we still could come up with two different answers or three different answers or four different answers. And I don't expect that to change. I always expect markets to evolve. And the emergence of AI as a tool will foster the evolution of markets. RAOUL PAL: Fascinating. All of this is fascinating to me, hence I drilled you with a million questions because there's so much learning for everybody on what it is. Because not only is it a whole different development going on, but what it also does is ask questions of us as investors, how we run our own processes. Are we asking the right questions? Are we doing the same thing? Are we observing the risk-adjusted returns in the right way within our portfolios? Because these are real questions. Just because it's a machine doesn't mean you shouldn't be asking of yourselves if you're an individual investor, too. I just think the whole thing is super fascinating, and I really appreciate you coming spending some time to explain it all. And I wish you the best of luck with it because it's an amazing space to be in. And it'd just be very interesting to see how it develops over time. TREVOR MOTTL: Thank you very much. This has been a lot of fun, Raoul. JUSTINE: If you're ready to go beyond the interview, make sure you visit realvision.com where you can try real vision plus for 30 days for just $1. We'll see you next time right here on real vision.