Beyond LLMs: Modern Machine Learning is Changing Investing

David Wright: Very much so. So, starting point, why do we use it? Markets have become more complex. So you can do this, clearly what we are like, we do this statistically, but if you look at the number of different things that drive relative stock returns—over whether it’s a one‑month or probably out to a three‑month horizon—the number of things that explain relative performance has exploded in the last five, ten, twenty years.

In the short end, less than a month, it’s just a huge number of different anomalies driving returns. We as humans could never understand that. More traditional, relatively narrow factor models can’t explain enough. So machine learning allows us to really identify the patterns and relationships in what has become an incredibly complex market. LLMs are useless in that area, they’re not good for building prediction models, they don’t work particularly well with numerical data. But things based around decision trees, which is what we’re using, thousands and thousands of decision trees that we train in a very sophisticated way, they allow us to build what we need, a level of accuracy and a level of understanding of what has become a very complex market.

Dave Nadig: In a fairly short time window, too, right? I mean, you’re not looking at signals that are years long.

David Wright: No. So we take around 400 different characteristics, and let’s work down from least jargon to most jargon. Characteristics about a stock often might be called signals; we can maybe talk about why we don’t necessarily call them signals. Today, in the machine‑learning jargon world, we’re talking about features. They might cover lots and lots of different aspects of companies, but the secret sauce of what we’re able to develop here is not about how any one of these things works in isolation. It’s about how they work in different combinations and condition each other, how in some instances you might want to use it in a positive way or a negative way. Again, all of these complexities, machine learning is just such a powerful way of accessing that alpha.

New Developments

Dave Nadig: So you’re both processing all of this huge data coming in day after day after day, but you’re also revving your own model. You said you may now branch into doing some neural‑network stuff, you told me earlier you’ve done some testing with LLMs on text‑based analysis. So you’re also doing this whole research project on the firm’s process while you’re investing the money. What’s exciting in that expanding fan of things we find AI can do for us, what is new and interesting there?

David Wright: It almost seems like a point now, my team are like, anything that I say we won’t do, they’re going to go and investigate and try to mess with me in these situations. So, again, LLMs, they’re not going to build you a forecasting model, but quants have been playing around with these for a long while. Natural language processing was a sort of early idea, more advanced LLMs the way to analyze text. The big thing for us with LLMs is, could we do sentiment better, could we build trending?

Dave Nadig: Like looking at news and things like that?

David Wright: Exactly. So the news, I think, has been the most interesting one that I see. Give the example here of a traditional trending signal, like buying analyst upgrades, selling downgrades, that we assumed as an industry wasn’t going to work anymore, and the decay would be one way. In practice it’s kind of more episodic; sometimes it works and sometimes it doesn’t. We built trending models by using LLMs to analyze news articles, to pull sentiment from them, and it looked a little better than what we were getting from a traditional analyst signal.

However, once you regress that, once you analyze what you’re getting from that news article on things like earnings that are already known by the market, there’s almost nothing left. So that is the key thing: an initial correlation might suggest you’ve got something new, but if you feed and remove things that the market knows, there’s often very little left from these types of things. So, LLMs, again we thought there might be some power at signal building, but it seems more limited than really many people are suggesting.

Dave Nadig: Is there a sort of standardized tool‑versus‑customized‑tool problem here? Because one of my experiences just using AI in a day‑to‑day world is they’ve become almost more and more generic, right? They can do more and more things, but that also makes them sort of less specifically useful for some things. How does that relate to what you guys are doing in your retraining process? Because, am I right, you’re still retraining every three months?

David Wright: Yep, so we retrain every three months. We have essentially a training framework in terms of the length that we train over, the allowance we allow, how many different trees it trains to combine together in the model, the amount of data that we use to do that training. So we have a set framework, and about every year we slightly restructure that framework. For example, a recent research project has found that as we incorporate more features, so we are almost encouraging more complexity, training over a longer period produces more accurate forecasts. So the framework probably changes on about an annual basis, and then every three months we do the retraining on the most recent data period, so that we pick up on the changing anomalies within the market.

Dave Nadig: I want to pick right up there, the anomalies that you guys are tracking. Some of them, I think, are probably pretty obvious, and some of them you may actually not have a clue about. What’s the velocity of change in those signals? When you’ve got something that you’re like, “this is really working, we’re going to weight this pretty heavily, this decision tree is paying off,” is that a thing that’s going to last for a whole quarter or a whole year?

David Wright: Yeah, if not quite a lot longer. And I like the language we end up using, when we’re referring to the algorithm‑type stuff, like it has this sort of human behavior. So it’s retrained over like 15 years. There are still plenty of the anomalies that you go right back to the earliest part of that 15‑year period, and if you just trained on, say, the first three years of the training period, you would also be able to identify the same anomalies in the latest three‑year period. So plenty of them are persistent, but then some of them are more transitory as well.

Just to put a number on this, using 400 features, there are a billion combinations of four features. So maybe a proportion of them are going to be identifiable through the entire training period, and then some of them are more identifiable and more transient, more recent things that the algorithm is able to capture.

Dave Nadig How much of your ability to identify these signals, sorry, features, and exploit them, depends on your own compute bill? I mean, you must be rather large users of other people’s AI compute, I’m guessing . But that is also coming down at the same time there’s more competition for it from the rest of the world that’s discovering AI. How do you think about that trade‑off? Because presumably you could always choose to just spend more money on compute.

David Wright: Yeah. When I look at budgeting for a quant team, and you go back a few years, it was predominantly about data. Now, you’re right, it’s what we’re spending on cloud, it’s what we’re spending on new chips. On the hardware, we tend to do most of the training internally. We buy new chips every couple of years to make sure that we’re getting the efficiency gains. We put out bigger, more moonshot projects on the cloud, where maybe it can be a bit more efficient.

But in general, with the speed that things are improving, we’re spending less than we expected to spend, even though we’re doing far more complex training. Again, if you change from, like, a 15‑year training to training over 20 years, it’s almost exponential how much more data you’ve got to use over that period, but we can now train that way quicker than we did shorter‑term models. So we see huge evidence that we get better, we understand the way we can do things more effectively, and then the hardware we’re using is getting more efficient and more effective.

Dave Nadig: I wasn’t clear on this. So you guys have your own hardware, or are we somewhere in some data center somewhere?

David Wright: So we have it in‑house in Geneva. There’s a 200‑year‑old bank; we’re pretty good at housing stuff that we want to be careful with and look after for many, many years. So we have our own server banks at the bank. We have a dedicated server bank with Nvidia chips, with CPUs, dedicated just for my team. And, again, every couple of years we’re having to pretty much pull that out and start again and buy the latest pieces.

Dave Nadig: That’s wild. So, a bit of a topic shift here, that puts you also on the other side of the AI conversation, which is you’re out there trying to get access to Nvidia chips and all that other stuff. How do you think about that trade‑off? Because there’s real scarcity in some pieces of this, but you’re a pretty high‑value user of that scarcity. How do you think about that math? It seems like a nearly impossible trade‑off, you just have to have belief that this much more compute is worth the million dollars to get this much more compute. Is it just a bet?

David Wright: So, firstly, I think that argument becomes much clearer when you’re a true high‑frequency quant shop. Like when you’re a pod at one of the multi‑strats, you’re almost like a market maker, and we know the kind of names here, there’s plenty of them.

Dave Nadig: You guys don’t do that.

David Wright: We don’t do that. So that’s not us. There, it’s more clear‑cut that the spend can be really beneficial. I would still come back to us: the real edge for us is, are we building the right features that we can train the model with? Are we improving the quality of the training that we do? Are we coming up with new ideas to implement the strategy in more effective ways? The compute becomes a little bit like a base case that you’ve got to have, a good‑enough stack and setup, and that isn’t available to most people. But we’re not at the stage where that’s going to be the edge. It still really is the research quality that we do, improving and refining every different aspect of the process, end to end.

Demystifying the Quant Black Box

Dave Nadig: Let’s shift to the advisor seat for a minute, because we’ve obviously had lots of people up here talking about different ways of doing stock selection, different ways of building portfolios. We’re all fishing from the same pool, right? We’re all playing with the same 3 or 4,000 names globally that we’re going to really think about investing in. One of the criticisms of AI and quant in general is that it’s a black box. How do you get an advisor over that hurdle? Because it’s not wrong in some senses, but it’s also kind of insufficient to say it’s a black box.

David Wright: So the first thing is, we’re pretty transparent. The advisors that are working with us, we’re going to give them a lot of information about what we’re doing, the data that we use, the features that we train with, a lot of the parameters that we choose in the way we train the model. I’ve given advisor teams access to some of my researchers and my portfolio managers to talk through these types of things. So the first thing is, we give a good level of information through that discussion.

And then, some of the choices that we’ve made to this stage. I’ve mentioned the idea that we train with decision trees, thousands and thousands of decision trees; we’ve been more reluctant to use deep‑learning neural networks. One of the reasons we’ve favored one over the other at this stage, amongst other reasons, is that there is an interpretability to it, there is a transparency. When we’re on a video call with an advisor, we can show the tools that can indicate the reasoning for every single position in the portfolio. We can attribute the end performance to show how much of it is a direct exposure to an individual feature, how much of it is one of these complex relationships that the model has learned. So there is transparency before the portfolio build and after the performance, and we’re willing to share this level of information.

Dave Nadig Can I bother you to get detailed in one of these? Can you give us an example of a position you might talk about, and a set of features, that you’d be able to talk to an advisor and say, hey, this ticker is in the portfolio because these five, six, seven features are very strong, and this complex series of features is something we’ve leaned into? How do you get somebody over that gap?

David Wright: Yeah. So, we’ve talked about this before and this is a struggle, because of the anomalies that we find, these relationships between the different features, whether it’s combinations of two, three, four, or ten that condition each other, that sometimes flip the sign on a signal…

Dave Nadig: Can you give the example of the data point we might be talking about?

David Wright: Okay, exactly, that’s where I’m going, but I get your point, let me get there. So some features have much more benefit in the direct exposure to them. Some of the short‑term, price‑based features, short‑term trending or short‑term reversal, linear direct exposure to that is most powerful. Some of the more fundamental and calendar features, they don’t have individual predictive power, so I wouldn’t call them signals, but when they condition other features in the feature set, they become incredibly helpful. So analysts, the relationship between a trending analyst signal, a reversal signal, and then the earnings calendar of a company, they all interact in very interesting ways.

Dave Nadig: Got it. So, for instance, the calendar is an interesting one to think about, because you might have a couple of data points that are always sort of looking like this, and that doesn’t make them a buy signal necessarily, but if it’s two weeks to earnings, everything’s off.

David Wright: It’s exactly that. But the challenge that we have, Dave, is that probably a third of them have that story base, and I want to give you that type of story. But the power in the alpha is that probably 70% of them don’t have that obvious story, so they’re hard to find. And that’s what makes them repeatable anomalies, because they don’t have that kind of story to latch onto, and I think for that reason they’ll last longer. We can see them in lots of different markets over different time periods, and I think the reason they don’t get arbitraged away is because there isn’t necessarily that obvious story to them.

Dave Nadig: Is some of that also that it’s proprietary, because you guys have found that signal and other people may not have, and therefore you avoid crowding?

David Wright: Yeah, I definitely think there’s that aspect. The type of thing that we’re using, this is not… our model is proprietary, our training is proprietary. But the data that we use, some of the techniques that we use, we know other people are doing it, but they do tend to be, as I’ve already said, in the higher‑frequency end of the hedge‑fund world. So they’re tending to chase the next second, the next minute, the next hour, maybe the next day. The people that are going after the next month don’t tend to use our tools. So I think those are anomalies, it is a less crowded space, and we’re the only people, as it stands at the moment, really trying to bring this type of idea to a much wider audience.

The Case for AI Looking Ahead

Dave Nadig: And we only have like 2 or 3 minutes left. I feel like I’d be remiss not talking a little bit about what’s going on in the AI conversation in the country right now. It seems obvious to me, given the way you’re using AI, I’m not particularly worried about the racks you have in Geneva making a million paperclips and killing us all, that’s not what they’re designed for. But you’re one of the people I know with the most fingers in the process in this whole industry. Do you have any concerns in terms of risk management yourself, whether it’s the kinds of models that you’re using not performing the way you would expect them to, or LLMs by competitors disrupting markets in ways that are unpredictable? Where are you on the p-doom versus acceleration model?

David Wright: On a broad level, I’m still pretty bullish. Some of the stuff that I see and read about what it’s doing around health and things like that overcomes my concern about it wanting to take over the world tomorrow. On a purely like, my investment‑world business, again, you kind of hit on it: it’s so targeted and focused in what we do. I do think it’s an interesting point, do we have the potential, the more people that do this sort of thing, to influence the market? I think there is an aspect that is right back to the start of what we discussed, one of the reasons markets have become more complex is because there is more digital trading, there is more machine learning out in the market. So I think markets will continue to get more and more complex, and it will require more powerful tools to understand that complexity. So there is potentially a reinforcing element to this.

Dave Nadig: Interesting. Well, David, this has been a fascinating conversation, I could go another hour here. Thanks for joining us again.

David Wright: Thank you.

Disclosure

*Alpha: The percentage performance of a financial instrument above or below that explained by its exposure to the broader market.

Regulatory Statement: Before investing, carefully consider the fund’s investment objectives, risks, charges, and expenses. This and other information can be found in the fund’s prospectus or, if available, the summary prospectus, which may be obtained by calling (855) 994-4778 or visiting www.pictet.com/etf. Read it carefully before investing.

Investment Risks: Investing in Exchange Traded Funds (ETFs) involves risk, including possible loss of principal. ETF shares are bought and sold at market price, not net asset value (NAV), and are not individually redeemed from the fund. Market price returns may be calculated using the midpoint between the bid and ask prices. For additional information regarding fund specific risks, please visit www.pictet.com/etf.

Performance Disclosure: Past performance does not guarantee future results. The performance data quoted represents past performance and current returns may be lower or higher. Investment return and principal value will fluctuate so that an investor’s shares, when redeemed, may be worth more or less than the original cost. Current performance data may be obtained by visiting www.pictet.com/etf.

Tax Considerations: ETF distributions may be taxable as ordinary income or capital gains, unless you are investing through a tax-deferred arrangement, such as a 401(k) plan or an individual retirement account.

Market Volatility: ETF shares may trade at a premium or discount to their NAV in the secondary market. Brokerage commissions will reduce returns.

Non-FDIC Insured: ETF investments are not insured or guaranteed by the Federal Deposit Insurance Corporation (FDIC) or any other government agency. They may lose value.

Professional Advice: The information provided on this website is for general informational purposes only and should not be considered as investment advice. Consult with a financial advisor or professional before making any investment decisions.

Performance: Market price returns are determined using the official closing price of the fund’s shares and do not represent the returns you would receive if you traded shares at other times.

Pictet Asset Management exchange-traded funds (ETFs) are actively managed and do not seek to replicate a specific index. ETF shares are bought and sold through an exchange at the then current market price, not net asset value (NAV), and are not individually redeemed from the fund. Shares may trade at a premium or discount to their NAV when traded on an exchange. Brokerage commissions will reduce returns. There can be no guarantee that an active market for ETFs will develop or be maintained, or that the ETF’s listing will continue or remain unchanged.

Foreside Fund Services, LLC, distributor.

Permalink | © Copyright 2026 etf.com. All rights reserved