AI at work: Inside the new era of hedge fund research and trading

Speaker A: Foreign.

Speaker B: Hello, everyone, and welcome to another episode of Hedge Fund Huddle with me, your loyal host, Jamie McDonald. And today we are talking about the tiny topic of AI. Um, uh, now, more specifically, we're talking about, uh, the practical uses of AI and how people are using it in, in terms of, uh, investing in trading today. And luckily, I'm joined with two experts to help me out. They are Andrew Delaney, who is president of the A Team Group, and Nishant Gurnani, who's a partner at Versa Investments. Gentlemen, welcome to the show.

Speaker A: Thanks for having us, Jamie.

Speaker C: Yep. Really great to be here.

Speaker B: Good. Well, guys, I like to start by just getting a bit of background, uh, from our guests about how they started, uh, their work careers and how they got to where they are today. So, Andrew, why don't we start with you?

Speaker A: Sure thing. Again, thanks for the opportunity to share everything here. 18 Group is an online publisher, and I'm a journalist by trade. And what 18 does is it focuses on the business of data and technology in global financial markets. We've got four main areas, one of which is trading, uh, tech. And we cover a lot of the use of AI in the trading and investment workflows in that segment. And for each segment we have, we offer sort of, uh, news analysis conferences, web webinars, um, and so on, so forth. So lots of free content, and I look after most of that. But as I say, I'm a journalist by trade, but I started my career, you know, with data very much at the forefront of what I was doing. Straight out of college, I got a job, luckily, as a news assistant at the Wall Street Journal. And as part of that job, you know, the deal was basically, we'll teach you how to be a journalist, but you've got to deal with data in our. The back of our book. So, uh, my job was basically to take, um, what was then known as a telerate terminal. It's a little video screen sat on my desk with a little keypad. And every evening, uh, just before midnight, which is when we put the newspaper to bed, I would punch out the government bond prices from counter Fitzgerald on a little printer. And then I'd turn around to a screen on my desk, another screen, and punch those numbers into the galley at the back of the book. And it was a fair, fairly sort of, you know, a menial task, and probably not as glamorous as I'd like it to be. But it taught me two really important things about data. And the first one was that, uh, the importance really of exclusive or difficult to Find information and analysis and insights. Basically, um, this was the European edition of the Wall Street Journal. The people who bought the newspaper, many of them bought the newspaper for those bond prices. Cause you couldn't get the anywhere else in Europe. This was pre Internet, pre just about anything else electronic. Unless you had a Teller 8 terminal, you couldn't get those bond prices. So exclusive data, very important. The second lesson I learned was the importance of integration. And of course, although the Wall street journal and teller 8 were both owned at the time by Dow Jones, I was the integration layer. There was no integration between the two, uh, systems that we were using. And so I just used to punch in 767-67-6777 and get those pages, print them out and get those bond prices, tap them back into the galleys for the paper. That was the level of integration we had at the time. And so that's how I got my start in this career, sort of moving forward a bit. I ended up in New York for 20 years. Initially I was the launch editor of publication called Inside Market Data, which became sort of the bible of market data. And then later in uh, 2001 we launched a team. And a team, as I mentioned, lots about data, market data, reference data, etc. But to bring us full circle to this podcast, last year we acquired the alternative data conference business of Eagle Alpha. We are now running the Eagle Alpha events. But as part of that, this sort of connection between alternative data and AI became a very important part of what we're doing. A lot of alternative data services are unstructured and we found that as AI became more accepted, AI models could be used to add structure to alternative data services, making them more important. So that's become a very important part of what we do. So a little bit of a roundabout way of getting here, but that's how I ended up on this podcast.

Speaker B: Oh, thank you, Andrew. And uh, a good little history lesson on how far we've come, I mean, in our lifetime, to be punching bond numbers into newspapers. Nishan, kind of same question to you. Before we talk specifically about Versa, perhaps you can give a little bit of a background about, uh, about your career and how you got there.

Speaker C: Sure. So, uh, I'm Nishanth Vernani. I lead the Futures and FX strategies at FirstServe, which is a large quantitative systematic investor. My background was pretty traditional from purposes of, of being a quant. I was a geeky kid who loved math and science and computers growing up. And um, I studied math in college and statistics in graduate school and then in College I was very fortunate to be able to spend two summers at some very, very well known quant shops. I spent a summer at AQR and then I spent a summer at what was then called SAC Multi Quant but is now called Cubist in terms of my background on data and AI specifically. I took this class a long time ago as an undergrad that sort of changed my life. It was a year long sequence on AI and machine learning taught by two luminaries in the field. And it was very clear to me that this was going to be the cutting edge of what needed to be done. And the progress we've made since then has sort of been pretty, pretty incredible. Finance specific. One of the jokes that my friends used to crack in college was that normal people would list other careers that they might be like a doctor, a lawyer, and I would only list finance careers. So I've always, always been interested in financ. And before Verser I spent a little bit of time out in San Francisco working for a fintech again focused highly on alternative data sets, specifically with regards to underwriting.

Speaker B: Nishan, just giving that introduction. Did you ever just start trading stocks or indices yourself? Did you ever just think I could try uh, and do this myself?

Speaker C: I definitely did stocks as, as a teenager and uh, definitely not indices. I don't think I was that mature in my development just yet. But definitely stocks and definitely some, some other harebrained ideas of things that I potentially could have traded and, and uh, and done.

Speaker B: Um, okay, well let's get into today's topic. We're talking about the practical sides of AI rather than anything more theoretical really. What tools are being used today by hedge fund managers, by personal uh, traders to um, either filter ideas or monitor ideas or help them with risk profiles. And Andrew, perhaps we'll start with you. Going back five years, we were really talking I guess specifically about just generative AI, large language models helping people to cipher and filter better ideas. But what's happened over the past, let's say three to five years to where we are today in terms of how AI is being used?

Speaker A: Sure thing. Well we've been following AI for a bit longer than that. We saw things like machine learning, deep learning coming through. I'd say probably um, six, seven years ago. And then of course we had the launch of ChatGPT, the birth of generative AI. Um, I would say since then we've been following that pretty closely. We've conducted probably six or seven market surveys over the past couple of years looking at how people are using this data, we run a number of advisory boards where we take practitioners, um, in our marketplace to lunch and pick their brain over something nice to eat. And from all those activities we've really sort of whittled things down to three types of use cases that we are seeing in the marketplace. And these are efficiency, growth and control type use cases. In terms of efficiency use cases, these are things like summarizing meetings and actions, things like using AI to code more efficiently, test code more efficiently, and um, to extract data from unstructured documents like alternative data sources. In terms of the growth type use cases, we're seeing things, uh, people using AI for asset allocation, investment modeling. They're using AI tools to drive client retention and really looking to identify cross selling opportunities. And then finally on the control side, we're looking at things like risk modeling, stress testing, scenario planning, some credit and market risk assessment and regulation interpretation, and to little extent digitizing our contracts. So um, they're the kinds of things we're seeing, uh, in terms of um, models being used, obviously all the household names as they are now from Copilot, increasingly, um, Claude and so on and so forth. But I think the real action in the hedge fund area is around developing your own AI stacks, own large language models and even more specialized models to add that secret sauce. So that's sort of been the development that we've seen out of the past two years, I would say.

Speaker B: Oh great, Andrew, um, Anishant, over to you. Um, perhaps you could start by talking a little bit about Versa, which strikes me as a very AI driven boutique hedge fund. You can talk about which strategies you employ and then once you've spoken about Versa, perhaps you could just go into a little detail of how you're using AI specifically today.

Speaker C: Sure. So Versa is a systematic quantity investment firm based in New York. And our focus is explicitly on absolute return strategies. We are purposely designed as a systematic, research driven boutique. And alternative data and AI have sort of been our pillars from very early on. So we started working with alternative data very early on. And one of the things that we're going to get into is that there is no extracting insights from alternative data without having AI techniques there. You can't construct sentiment scores from text unless you use natural language processing methodology in that. So this has sort of been a core part of what we're focused on. We have three main strategies that we run. We run a systematic equity strategy, a event driven strategy that focuses on merger arbitrage, and then a managed feature strategies which I'm personally in charge of and when it comes to our philosophy on alternative data and AI, this sort of cascades firm wide. So I want to talk about some of the sort of specific examples that we use on a day to day basis, some which Andrew alluded to, but I want to sort of use some specific examples that we think of. So our philosophy is fundamentally we are systematic investors and our job is to speed up the velocity with which we get and are able to make good investment decisions. That sort of framework is generalized so that it is not specific to quantum or discretionary. Our goal is to get good actionable investment ideas. And so AI is used on a day to day basis. In helping us do that, first and foremost on speeding up research, asking more detailed questions, helping us organize our day to day management, we use a tool called motion AI that sort of dynamically adjusts tasks and projects based on priority. So all we see is that throughout our investment process, these efficiency gains, even though they may not be super glamorous on an individual basis, they compound so that we're able to do things that you're not able to do before. In terms of the research specifically, there are really two big ways that we see that AI is impacting us. One is the speed at which we're able to evaluate research ideas has increased significantly. So if I have a question like I want to know the number of dissenting votes in the FOMC meetings going back 20 years, that's a question I could have answered before AI, but the speed at which I can answer that right now using either off the shelf large language models or fine tune models that we fine tuned ourselves has gone up significantly. The second thing that we're able to do is tackle a uh, complexity of research ideas that we were unable to do previously. So as a very concrete example, consider an idea that you have and um, the idea is currently there are a lot of high quality podcasts that upload on a daily basis where investors come on and give a lot of interesting color on market sentiment and their views. And so maybe your investment thesis is I want to get some sort of consensus understanding of what people are saying. 15 years ago this was very difficult to not possible. 2012 we saw the big Vision paper come out of Alexnet. 2018 is when we see Google launch the BERT paper. 2017 I was at a conference where actually the attention is all you need. Transformers paper was announced and even this idea of summarizing synthesizing a vast amount of market sentiment from podcasts would have been impossible then. And today using Claude, using the Latest large language models. I could do it in a weekend and not even a weekend in few hours. I could make a large, large amount of progress on there. So I think now the question has really become from the investment process, not about are you using AI? I think all knowledge work in general requires using AI, meaning fleet, in order to speed up efficiency. But the question is really around how are you using it, where is it providing the most value and how is it fitting into your investment process in general?

Speaker B: Nishan, I have two quick follow up questions on that. Firstly, one thing that struck me as you were talking was when again this is going back to when I was running a book myself. Conviction was kind of everything when it came to an investment idea. I needed to personally have conviction in an idea to know when to add or when to, you know, take positions off. Uh, and um, part of getting the conviction is the friction that you feel, you know, putting work into an idea. It was reading the 10Ks, it was meeting with management and it was, and it could only really come from within. Now if you're relying on AI, obviously your conviction can only be as high as the conviction you have in the, in the AI platform. So question number one is if conviction is still a big part of the trading, how do you get conviction in AI? And part two of that is, as Andrew was alluding to, there's a lot of platforms out there. You mentioned, you know, Claude and Motion AI and Gemini. Do you rely on third party AI platforms or have you started to kind of build your own? And which do you see as, as more sort of useful to you?

Speaker C: Sure. So with regards to conviction, the way we think about it is the AI agents and we've been spending a lot of time on building out our agent capabilities should be thought of as junior researchers and their goal is to help you get to decisions faster. But the conviction still has to come from you. So let's sort of give some, some concrete examples. In the past, if you're a junior researcher who joins our signal research team, one of the projects that you might be asked to do is I as the lead have a specific academic article that I've read maybe in the Journal of Financial Economics. I think the idea is interesting and your job would be to read the article, implement the idea, test it using our internal evaluation framework and make an argument and a research case for why that signal is predictive and should be added to the strategy. What we can do with AI tools is we can systematize this process. So not only do we have researchers doing this Work, work. But we can have AI agents automatically read papers, suggest ideas, but then they go through the natural research process where the PM or uh, strategy lead still has to evaluate it on a rigorous basis. The thing with using AI tools in general is evals are really important. There are constantly new models coming out. So two weeks ago we saw Opus 4.6 on Claude OpenAI launched, uh, Codex 5.3. So we're constantly seeing these new things come along. And so the conviction and confidence of these models is a function of structured evals process. So leading it back to what you said, Jamie, in your context, what you would do is you tracked insurance stocks, if I remember correctly, and you would have a historical data set that you've worked with, where you sat and you did the work and you did the effort. And every time a new model would come out, you would get to score it on that. And depending on the quality of the score, that's the conviction level you have in that particular agent that is using the tool. So a really fun example for everybody who's listening to the podcast is go into your own favorite LLM that you like and try this very simple evaluation. So just go M. So my car needs to be cleaned. The car wash is 50 meters away. Do I walk or drive? And if you just try something really simple as this, uh, like very simple reasoning eval, you're going to be shocked by some of the answers that you get. And if you go back models as you can click through, you'll see how it's getting better. So conviction is still very human driven, but getting the idea to a point where the human can start working with it and think of an actionable idea. That's the velocity I'm sort of talking about where I haven't, I have a random idea. I read this paper. I don't necessarily have the time to spend a week looking super deep into the paper because maybe my initial conviction on the idea is low. I can have Claude in the background running, and I explicitly do. Right now I have Claude running on a couple of different problems that I'm looking at and it'll give me back enough structured output that as a, as a human then I can say, actually this is really interesting. Let me go and pursue this further. This has legs to it, this idea, actually. Thanks, Claude. You did the work. I've convinced myself this idea I need to throw away. So that's on the conviction side. On your second question about internal versus external, which Andrew also alluded to, I think you're going to see an Evolution of both use cases. So if we take a step back, when we talk about AI, we are really talking about large language models specifically because that is one of the key tools that is helpful in the investment process. In order to train a large language model, there are two steps. There is a pre training step where you run the model on a large corpus of text across a large amount of compute. And that is sort of learn some generalized knowledge. And then there's the second step which is the tuning step where you suddenly decide this is my specific use case and then you feed it examples to help understand. So the way that I suspect investment firms are going to do and the way that we started thinking about this problem is you have to design your systems so that you use whatever model is best. And that may be something that's off the shelf. It may be something that's off the shelf that you then fine tune for your specific use case. It may be an open source model that you downloaded off hugging face like Deep Seq and then decided to have a bunch of internal evals that help you tune it specifically to say FOMC statement analysis or unstructured data parsing, whatever be that use case. And then the third, which I think is actually the toughest and we won't see firms do this. And I have my reasons why is training very large end to end models entirely from scratch themselves. I think the whole purpose of LLMs and fine tuning has been that you can take something that another person has trained on a generalized corpus and then make it smarter for your use case. And that's the real, real moat. Because the truth is most investment firms, including the largest firms, do not have the level of compute that the Googles and the Microsoft's and the OpenAI's anthropics have in order to do it. And frankly they're solving different problems. One is like the Gemini model is supposed to be generalized to generate text, video, transcripts, whereas investment folks are uh, really focused on the investment problem. And so the highest leverage thing is take something that exists, adapt it to your specific use case and improve it. And that's where we're going to see this proprietary differences where firms that have spent a lot of systematic time on how to improve their internal models will diverge in the skills with which they're able to deploy them.

Speaker B: Perhaps Andrew, you can comment a little bit on what Nishant just said, what else you're seeing elsewhere in the market in terms of using your own platforms versus external. And then also Andrew, um, Nishant said Something there which was think um, of these platforms as a bit of a junior analyst. So there'll be people listening wondering how many jobs will be open, ah, at the junior level in years to come. And I wonder Andrew if you could maybe just comment on that a bit.

Speaker A: Sure thing. So um, I totally agree that the uh, world has been looking at how to onboard AI if you like. Um, we've got this sort of uh, idea that um, the firms are building their AI stacks. They're putting in place governance rules, policies and so on, so forth to sort of really put it, I guess, uh, really put the best foot forward, have the right tool for the right job. And I think that is a process that's ongoing. I don't think anyone's got it down um, as just yet. We're seeing a lot of appointments of chief data and AI officers now. People who are incorporating that kind of discipline into their adoption of AI, making sure that the people within the organization know uh, which tools should be used for which, which processes and which tasks. So I think it will be a mix of internal, external. I don't think, as Nishan said, I don't think it'll be a massive build from scratch. But the nuance, the uniqueness will come from the mix, uh, the mix of what you've got internally and of course the data you've got access to which we can talk about, I know we're going to talk about in a little bit. So I think that is the path we're on in terms of using agents to do various tasks. I mean we are seeing that in real life part of uh, we've just completed a survey a bit wider perhaps but certainly within the investment bank and investment management side of the world and we're seeing large organizations put in place teams of AI agents to perform tasks. We're seeing that, we're seeing valuation of these agents as if they are employees. They get ranked, they get um, evaluated, they get trained, they get told off and told to go and perform better if they don't meet certain requirements, um, and ultimately they get terminated if they don't work. So you're seeing sort of a whole sort of corporate structure of uh, these models starting to, or these agents, I should say starting to emerge I think in terms of the light of the end of the tunnel if you like, or the uh, the silver lining for perhaps junior uh, staff is that collecting the data that needs to be used to train models and indeed to pull into these models is still something that finding unique data sets is something that's very much a mix of manual and automated. We see a lot of uh, human in the loop for this kind of stuff. And it's getting back to trust in the data and making sure that people um, do feel that they do have, that they're getting the right data. That's, that's, that's being used to train these models. I think that's an imperative that that will continue.

Speaker B: Andrew, just then you mentioned, you know, the constant strive for unique data sets. And um, you know, even when I was, you know, running a book 10 years ago, I was always so worried about crowded trades. But I, I can't help feeling that these platforms are going to continue to create these crowded trades. So perhaps, Nishant, you can, you can talk a little about that. I mean, how do you make sure that the prompts that you are using for, for idea generation and are not similar to the Citadels and the other big companies out there? And again, maybe going a stage further, let's take a black swan type of event like tariffs last year. How does Aversa, uh, perform in that kind of environment? And how do you protect yourself?

Speaker C: Sure, look, that's entirely a fair question and I think it goes back to investment edge. So as AI becomes more accessible, it is accessible, right? Like, I don't want to make it seem like that's a future statement. It is accessible right now. It is pretty easy to get started using it on a day to day basis. The edge doesn't come from having more models or compute or more data. It really comes about how are you using those tools in a meaningful way. So when we think about how we're using these things, our advantage really comes from our investment process, which we believe is differentiated. It thinks about markets in a very specific way. Our usage of alternative data for each one of the strategies is defined in a very specific, unique way that we don't believe people are doing. So let me talk specifically about our flagship managed future strategy which I work on. One of the things that we do there is, we have this view that you need to look at equity index futures from two perspectives. A top down macro perspective, but also a bottom up stock level perspective. And so we have alternative data that we collect on 24 markets, equity markets globally, that's 10,000 stocks. We aggregate those stocks individually and at the country level and we construct signals doing that. And we believe that's not a common approach. There's a lot of skill and nuance that goes into applying those things and thinking about that problem in general. And there is a Research focused idea that results in that. If, if I give a quick example just from the general AI world, one of the things that that has come up previously but maybe not been discussed in detail is so if we, if we roll back the history a little bit and just look at the timeline of the development of Transformers in 2017, this really important paper comes out called Attention Is all youl Need. That introduces the attention mechanism which is the heart of the Transformers. It comes out of Google and a year later Google actually releases uh, the first Transformer model which is called Bert on um, Sesame Bert. The T is the Transformer, yet It was the OpenAI GPT series that ended up winning. Why is that? The reason that happened is one, their focus was very different. The Google models were very focused on the Google problem of search and understanding. And so what BERT was really good of was understanding text. It was a really good reader of text and understanding. OpenAI took a very different approach where they really focused on this generative piece, thinking about how it looks and feels to generate text and thinking about what is the likely next thing that somebody is trying to say and generate stuff that goes according to that. And it turns out that that approach was actually the approach that ended up scaling better and led to the GPT advances. So there are two teams. Google is significantly better resourced as significantly more researchers. OpenAI in 2018 I used to go to offices because they're right pretty close to where I used to work at Brex. There are about 80, 80 odd people. They're working on the early GPT versions and it just turned out that that approach was the right one. So when we think about the investment process and commoditization of AI and this comes down to what Andrew was saying about the uniqueness of alternative data sets. Also it's about thoughtfully thinking about this is what I'm doing. And here's how this is going to lead to differentiated alpha. Ultimately rather than necessarily being concerned that oh, everybody's putting in the same inputs because look, frankly speaking, markets are competitive. If you just do the same thing as everybody else, you're not going to make money. And so a lot of the folks focus is certainly on that. And so when, when you reference time periods like Black Swan events, there are sort of two specific things we think about in that sense. One is just experience. We have structured our strategies across the board to have risk as a core part of their philosophy. The founding partners have navigated multiple cycles, dot com bubble, great financial crisis, Brexit and so having a good risk framework that takes a realistic notion that liquidity is going to dry up, a lot of stress scenarios can happen. And designing strategies that are sort of going to survive those, those uh, those periods is important. In my particular case, for the flagship managed features strategy, there's a focus on something that is referenced that we call convexity, which is the ability to do well in up and down market markets. And so it's part and parcel of the design of the strategy itself. And we've been trading on this philosophy where we look at cross sectional differences between equity markets worldwide. So regardless of whether they're all falling or going up, we should be able to make money. And this has worked well for us. Over the past eight years, the strategy has been live, not just during sort of COVID but SVB and so on, so forth. So the goal isn't really to be immune to market shock shocks. That's unrealistic. The goal is to take your specialized investment process and design it so that it is resilient to different market environments. And where AI and alternative data fit in is helping you design that process well and in a robust way and in a unique way so that you're not competing with others and you're actually able to make money in deferring markets.

Speaker B: Um, so Nishant, uh, sticking with you, that's really interesting. So we've spoken a bit about research and investment idea generation being automated versus human LED and the relationship there. What about execution? And um, again, this touches on risk. To what extent do you have AI programs in place that will, without a human being evolved, change the percentage makeup of a portfolio, I. E Will trade without a human being involved? Because that, that seems to me like a, a bigger step. I, I was thinking earlier it's a bit like booking an Uber, but there's a, a human driving it versus Actually getting into a Waymo where there's now no human driving it. Like are we at that stage yet?

Speaker C: So I think it's a spectrum. Look, jb, even without AI, there are a lot of high frequency trading algos in the market right now that are trading autonomously without any human intervention. That's just the truth. With where we are when it comes to agentic systems in general, I don't think we're entirely there yet. I think what the advantage of the agentic approach is you've imbibed a little bit of intelligence in all the various components. So if we talk about execution in general.

Speaker B: I was being specific about discretionary trading there.

Speaker C: Sorry. Yes. So, uh, in discretionary trading, I don't think we are necessarily there because again, I don't think there is is a level of trust in the LLM output. But we are so close that these are the things that could happen. So let's be specific on an example. So you're watching some stock, you have a large position in Apple for whatever reason and you've built a bunch of agents out there that are looking out for black swan events, right? So you are reading the news feeds, you are looking on Twitter. As soon as somebody says something about Apple that you've designed and think is going to be super negative. If you have an alternative data source that is looking at payment volume coming in on number of iPhones sold. Right now we're at a place where you might have alerts and the alert goes off and then Jamie gets called and then, you know, you do something. With Agentic systems, I think we're a step further. We've given them all a little bit of intelligence. So not only will they call and say, hey, there's something wrong with this Apple position, they might have a recommendation that says, actually you need to cut your position by half. Half. I don't think we have reached the stage just yet where we are fully comfortable with them doing that execution because again, there is this human component that is still driving the investment decisions on the discretionary side. That being said, I think we're probably less far away from that than we think. It's a level of comfort. So if you've been following the news, Meta recently bought this open source, this agentic platform called cloudbot. And what cloudbot is, is a personal assistant that's in your emails, it's sending emails on your behalf, it's scheduling meetings and doing stuff. And if you start using that, there's a deep discomfort with doing that potentially. But once confidence grows, I. E. You systematically evaluate how its recommendations have behaved over time, I think you'll see people converge to a place where they're more comfortable letting it trade on their behalf. Just like you've seen people get much more comfortable with using Waymo. Like if you've been in a Waymo the first time you go, it's very scary, but then it quickly gets very boring because you're so used to the consistency with which it's able to do the thing that you want it to do. And so I think we're on that journey. I don't think we've necessarily got there on the discretionary side.

Speaker B: Yeah. Andrew, moving across to you, I wanted to ask a little bit about regulation and how that may play a part in all of this. We've obviously talked about AI as a good and helpful agent but I'm sure it can be used out there for um, you know, if you're long Delta and short American Airlines, you can presumably program something to click on Delta Airlines website as many times as possible to try and like give the impression that everyone's starting to fly Delta or whatever. But to what extent are regulators sort of trying to get involved and what do you see any effect that might happen?

Speaker A: We've had um, regulators come and speak um, at our events. They tend to come and um, try to assure everybody they're not going to be too heavy handed around this. Given our audience, they want to seem to have a light touch. I think uh, they're at the stage, very much learning stage at this point. Uh, some of them are running all manner of sandboxes and trials and so on and so forth where they get people to play with things that they think could help in things like reporting and so on and so forth. But in terms of actually getting more proactive um, around the use of AI in that investment process, I think we're still uh, waiting to see. There's been a lot of talk of the uh, sort of the EU AI act and so on so forth which are uh, more generic in flavor if you will. I uh, don't think that's trickled down yet into our world. And when I talk to our regulators, that is the financial regulators, we haven't seen them apply that as yet to our activities if you will. So I think the jury's a bit open really as we are now now not quite there yet.

Speaker B: Okay, and Nishan, just I guess tangentially on that. Are people like versa having conversations with regulators? I mean I'm sure that they're out there talking to, to people inside the market and then maybe it's a second follow up. Maybe this is more like an investment question about AI. But we're obviously in some kind of bubble maybe when it comes to tech and a new piece of um, transformative technology like uh, AI, um, what are the signs that a bubble might be bursting? What do we need to look out for? Do you even believe we are in a bubble?

Speaker C: Certainly. So answering your first question about regulators, we as a, as a SEC and CFTC regulated firm are properly regulated by the rights of the horticulture. As to whether they're coming and speaking to us about regulations, not to the best of my knowledge, but certainly, certainly somebody at the firm can correct me if I'm wrong? I don't believe so. One of my thoughts on the regulation piece is I do think there are existing rules in place that govern how algorithms are deployed in financial situations that I think work really well. So the person who deploys the particular model is the one who is ultimately to blame. So during my summer on the SEC multi quant debt ask the Knight Capital algo situation happened and there it was very clear that the blame did not rely on the algorithm itself. The blame ultimately goes to the people who are deploying it. So I think there is that framework that exists even for AI development. So issues with Waymo cars are immediately related to Google and like they are the ones that should be held responsible. So that's certainly my thoughts on the regulatory piece piece with regards to bubble, not bubble. To be entirely fair, I am not the right person for that. I am a systematic quant investor. So for me this is a natural evolution of the process. I really just think of these things as tools that are helping me be a better quant investor. And so there's a little bit of bias in my thinking because more data, more compute are my lingua franca on a daily basis. So I certainly don't see this necessarily being super bubbly. But I am not the right person. I am not looking at the capex spending, I am not looking at the mismatch between what compute is required and the power that is required to generate that compute. I think there is some mismatch between those things. I think the amount of compute that we are trying to build is not supported by the amount of power that we are able to generate. I think that's a key bottleneck that has been pointed out.

Speaker B: Um, and a question really for both of you. I guess there'll be quite a few people listening who are perhaps trading their own portfolios and they want to get better at using AI tools to help either come up with ideas or help them monitor them. What sort of advice can you give them of where to be looking to try and find the right platform for them? And I guess the second question is I never really thought about it until just now, but how careful do people need to be with their prompt writing? Should they actually spend some time actually trying to work out, you know, what they write in those prompts is obviously quite important. So maybe a few words on that,

Speaker A: uh, generically speaking on the prompts, uh, side, I mean I hear of people keeping prompt libraries and again as part of the governance, if you will, of AI deployment is, uh, this is the way we approach this kind of a prompt, this is the way we approach that kind of a prompt, to get the best results or to safeguard against perhaps generating something questionable that won't be defensible ultimately. So I think there's some governance to be done around that and I think people are starting to do that. And then my other bit really is about, uh, data quality. There is this constant search for new data sets. I can see it. And I think to some degree AI is something that encourages that and makes it feasible. But it is about ensuring you've got the right processes in place to make sure that what you get in the end, that secret source, the nuance, the unique approach, is really optimized by making sure that your data quality is good. Ultimately, that would be my two cents, as it were.

Speaker C: Yeah, I think Andrew is 100% right. Having a prompt library is something that we highly recommend from a quant perspective because we treat these as models and we want as much of a deterministic outcome output. We do need to store the type of prompts that we write. And over time you sort of learn how certain prompts help the model make certain decisions. The worst thing that can happen as a retail investor trying to apply these things is one day you put in one prompt and you get recommendation A. On the other day you put another prompt, you get recommendation B. And so there are simple things that you can do to make this process more robust. One is any large prompt that you're writing that's an investment thesis, you should record it in your trading journal and it should go there as a specific input in your investment process. The second thing that I highly recommend that people do, and this is something that we spent a lot of time building internally, is if you use a particular tool to get some sort of output that is used in the investment process, have another LLM score it. So if I just think of discretionary investors, if you are parsing FOMC statements, say you say you're a retail investor and you're trying to parse an FOMC statement, you can do that pretty easily off the shelf, Maybe say using ChatGPT, have Gemini score it for you and have that consistent framework so that you're having different models, different high caliber models sort of evaluate each other because it keeps the output a lot more honest and it gives you a better sense. One of the agents that we're building internally that I'm very excited about is what I call the always critic agent. So we were talking about conviction and I really am looking for an Agent that is just every idea that gets proposed to it. It tries to point out issues with the idea. And that's very helpful because. Because it's like having somebody out there who's carefully reviewing it. So there's actually a really great open source tool called RoboRev that does this continuously for code, but you can also do this for your research process. Have another agent sit and score the output of a specific agent, poke up holes, use that output as an input loop to improve the structured response. So you can have this back and forth going so that you get a couple of rounds of here's an idea. Yep, here's where it sucks. Here's an idea, here's how it's better before it even gets to the human to look at and say, hmm, interesting. Maybe I can action that. Or not.

Speaker B: That's some really excellent advice. Thank you both for that. In fact, if people are listening, they should go back and listen to those answers again. Some, some excellent practical advice in there. And so, um, gentlemen, we're, we're, we're slightly running out of time here, but I wanted to, considering we're talking about practical uses, finish on one question for both you, if it's not too personal. How are you both using AI in your, in your own lives, just to make life outside of work a bit easier? With me and my wife, it typically seems to be about what to do with the children, essentially. And then an argument ensues which we then ask AI to try and solve. But m. That aside, what kind of things do you, um, find, uh, useful outside of work?

Speaker A: You're saying? Yeah, I sit there and answer my daughter's homework, really, uh, with my perplexity on my phone. Daddy, was this mean. I'll be off quietly and um, you know, run a couple of questions on that. So it's become the new Google for me. Just on my phone for sure. And then, you know, I had a. My second project, which I haven't executed on yet, is our, uh, back garden here in London, which is a bit messy. So, um, my plan is to take a few pictures and then run, run these pictures through something or other and find out, get a decent design for the back garden. So I think there's all sorts of things you can do with this.

Speaker B: I like it, Nishan.

Speaker C: So I have a nerdier answer because I'm inherently a nerd. And so a lot of my AI usage is actually building lots of silly projects that I can use around the house. So I have this dashboard of the real time number of Bikes available downstairs by my apartment because I bike to the boxing gym every morning around 5:30 and it's really an unpleasant ride if uh, I don't have one of those electric bikes to take me through. So I have all these like little cute things about automating stuff. And I'm actually driving my wife crazy because I'm trying to build her a little app with our own schedules and like little things. So I am, I'm very much a kid in a, in a candy store who's just like any ridiculous idea I can think of. I'm using Claude to like spool up an app and start using it.

Speaker B: While you're preaching to the converted, I get obsessed about, about the time it takes to get from one place in New York City to, to another. And I know there's, there's a lot of different platforms that give you different answers and you see which one's best. Anyway, gentlemen, I have taken up uh, too much of your time, but you've both been amazing guests and I want to thank you for uh, for today's podcast. You know, if people want to get in touch with Andrew, the A Team group have their own website to, to see what they're up to. And of course Versa Investments, uh, have their own site too if, if you want to get any more answers. But um, gentlemen, if you want to give a few parting words, just a

Speaker A: thank you very much and great, great conversation, fascinating topic. I'm sure we'll be back next year to see where we're at.

Speaker C: Yeah, likewise. I've been a long time listener so it's been really exciting to be finally on this side of the mic. I think we're in a really, really exciting moment with AI and anybody who's listening to the podcast, just go out there, don't be afraid, just use the tools, experiment. There's a lot of fun to be had and a lot of efficiency to be.

Speaker B: Well, I think a year is going to be too long, the pace at which this uh, this world is changing. So Andrew, Nishan, thank you very much indeed and uh, thank you all for listening. Thanks once again for listening everyone. And please, as usual, give us a follow like or subscribe wherever you get your podcasts. The information contained in this podcast does not constitute a recommendation from any LSEG entity to the listener. The views expressed in this podcast are not necessarily those of lseg. And LSEG is not providing any investment, financial, economic, legal, accounting or tax advice or recommendations in this podcast. Neither LSEG nor any of its affiliates make any representation or warranty as to the accuracy or completeness of the statements or any information contained in this podcast, and any and all liability, therefore, whether direct or indirect, is expressly disclaimed. For further, further information, visit the show Notes of this podcast or LSEG combination.