Singularity or SkyNet: Exploring AI with Greg Bond of Man Numeric

Greg Dowling (00:05):

Welcome to the FEG Insight Bridge. This is Greg Dowling, Head of Research and CIO at FEG. This show spans global markets and institutional investments through conversations with some of the world's leading investments, economic and philanthropic minds. To provide insight on how institutional investors can survive and even thrive in the world of markets and finance alpha or apocalypse singularity or Skynet, AI has been a big focus for FEG this year. So we are honored to host Greg Bond, Chief Executive Officer of Man Numeric to speak to us about artificial intelligence, potential benefits and risks. The Man Group is an asset management firm dedicated to combining its investment management experience with cutting edge technology today. Today hear about the history of AI, we will define the buzzwords and address whether it'll take all of our jobs or unleash economic growth and productivity. Finally, we will end with how Man incorporates AI into their business practices. Listen now, do not get terminated. Greg, welcome to the FEG Insight Bridge.

Greg Bond (01:17):

Great, thanks for having me Greg.

Greg Dowling (01:19):

Would you mind introducing yourself and Man Numeric?

Greg Bond (01:21):

Sure. I am Greg Bond, I'm president and CEO of Man Numeric, which is the bottom up systematic equity and credit engine of the broader Man Group. We've been trading securities since the late eighties and really again with a systematic bent to it.

Greg Dowling (01:35):

We're excited to have you here and we're going talk a lot about AI, but maybe to level set here, could you define what AI is and maybe what it isn't?

Greg Bond (01:45):

Yeah, it's interesting. There's a lot of jargon that gets thrown around and we think about AI as being something new, but it's been around for a while. The way we think about it, sort of, it's concentric circles, right? The outside part is the artificial intelligence. That's the broader scope, and that's any technique that enables a machine to mimic human behavior. So any sort of technique that allows you to do that, you can think about maybe expert systems back in the seventies. People are writing a bunch of if then clauses in their code to try to answer questions that pretend to be a human. But I think we've moved in the modern world more into what people call machine learning, which is using AI techniques where statistical methods are the underpinning, right? Helping the machine improve with experience through statistical techniques. And that can be, you know, fancier things, but also more straightforward things like linear regression.

Greg Bond (02:29):

There's a lot of jargon, but at the end of the day, there's just trying to do something with statistics to help the machine understand and sort of fit the world. And then a little more recently, maybe in the last five to 10 years, there's been a subset of machine learning, which people call deep learning. And that's really trying to figure out a way to have the machine or mimic some of the operations of the human brain, you know, to a point, right? And this is where you hear about multi-level neural networks and all of that kind of jargon, but it's really trying to model the human brain. And then you could argue that a subset of deep learning has been the transformer model. So this is where, you know, you hear about Chat GTP, large language models, and here this is just another type of deep learning where we're trying to work with, it's particularly good with natural language processing and other things and more about language.

Greg Dowling (03:13):

A lot of jargon there. Neural networks, I mean that that's the brain for people that don't know that. And also just transformers. And that's been the big breakthrough from what I understand. Maybe you can explain in a little bit more detail how a transformer works and how this led to Chat GPT.

Greg Bond (03:29):

It's very basic level. It's an interesting way to predict what the next word should be. Sort of a logical step and a sentence or a question. And it's a fairly straightforward technology, but it requires a lot of fitting, right? A lot of powerful computing, a lot of, you know, chip power, etc., To have that go forward. And so that's been bringing what I think is somewhat of a very technical skillset and technical things into something that we can use in our day-to-day life. And I think that's been a bit of the breakthrough. That's why it's been adapted by so many people that it's sort of brought machine learning into the retail world. And I think that's where a lot of the excitement is. And I think the key is that if we think about AI and sort of the broader press, we just see these big cycles of interest in AI going back with the deep learning and even prior.

Greg Bond (04:12):

So there's always a little bit of this excitement, disappointment, excitement, disappointment with AI over the last several years and you know, arguably Chat GPT is that latest version of some of that and the excitement around it. But I think we just need to have some caution. And I think what we're really seeing where the uptake is, you know, particularly at Numeric or, and the broader Man Group is just a steady adoption of a lot of these technologies over time, sort of using it, finding places where it could help in an investment process in particular. But it's really trying not to react too much to the press and the interest from our investors, etc. But trying to figure out where it might fit and the excitement. But I think this time with the development of these models, it might be transformational versus some of the other hot buzzwords that we've seen over the last 10 years or so, cryptocurrency, etc., That I think that this is something that we'll find actually useful.

Greg Dowling (05:01):

Greg, is it fair to say in a simple kind of way, maybe when you're sending a text, and I'm sure all listeners have sent a text and you're in there on your iPhone and you're typing the sky is, and it will anticipate what the next word is, might be blue, that's probably the highest probability, but it could be green or red, or something like that. That's kind of what the generative side is on steroids, right?

Greg Bond (05:24):

Correct. There's a lot of details around that, but I think again, is Chat GPT or the large language models, is it actually creative necessarily, right? I think that's, that's a question, but it's more about an algorithm that's making a good guess as to what the next thing should be. But I think there's some interesting things around it. So one simple answer is Chat GPT may not be necessarily creative, but we are seeing some evidence that for certain areas that it can sort of do some initial hypothesis generation, it's done some work and AI in particular has figured out some new moves in some of the games like AlphaGo from a few years ago, again, around some of that deep learning. It's sort of how do, one wants to define creativity, but there might be some sort of opportunities here around meta creativity, creating meta researchers, things that are sort of in between and helping one become more productive and trying to generate hypothesis for example.

Greg Dowling (06:10):

That's pretty interesting. You kind of mentioned it, but like Chat GPT was I believe, the fastest app to get downloaded over a million times. Like it took like five days or something like that for it to happen. How did this happen? Like I'd never heard of it and then all of a sudden like everybody around me was talking about it.

Greg Bond (06:27):

Well, I think it's a couple of different things right now. You've got, first of all, it is something that goes into the retail audience, right? It's available to everyone. It's not a fancy package from Python or R, it's right there. It's something that that you could immediately activate. It's also something that you can, again, on phones, on computers, we have just the broader kind of network effects that we've seen in our society, right? People start using things like Facebook and then the more people that use Facebook, the more valuable it is. And so when we talk about these things that I think it's just the, the confluence of really interesting technology, people's comfort with technology and then just the availability and quick adaptation of it. It's just very easy to sort of download and use, particularly from a personal perspective. Now in businesses, people might find it harder, right? To download it at work for example. There's a lot of privacy questions and IP sharing questions around that. But I think for the typical user, it's just really interesting technology.

Greg Dowling (07:17):

That's absolutely true. And I love what you said earlier, you're talking a little bit about the history of this and I guess it probably traces its roots back to the fifties, right? So the code breakers from the UK and others where they were trying to do this and there was even money from the U.S. and the British government we had this AI winner, right? So this is not new, we wouldn't recognize the old version of it, but this has been talked about for a long time.

Greg Bond (07:42):

As I mentioned, techniques that people learn in their stats class in college, again, we forget that things like linear regression, just simple models, that's a form of machine learning. And I think we can get enamored with the whizbang cool latest sort of non-linear, really interesting models. But at the end of the day, I think basic research processes and things, we shouldn't change those because we have these new techniques. It's really going back to hypothesis testing, understanding when and where to use the artificial intelligence or the fancier algorithms versus the more simplistic algorithms. And I think that's a bit of the challenge that we're seeing that you can get enamored with that wave of interest and sort of maybe abandoning your basic scientific testing and things that we try to do as systematic investors. What the power is for us I think, is that in places where there is a lot of data, things that you're trying to understand in the market, those are the types of places that you want to use these higher level techniques. But you need to go in with a strong hypothesis and you need to prove to yourself that this is actually adding something above and beyond what you already know how to do. And those are the strong tasks that we require folks to go through when they're doing research projects because it's really important that ultimately you give up a little bit of the explainability of what you're doing when you go down this route of machine learning. And so sometimes simpler is actually better.

Greg Dowling (08:56):

You kind of mentioned it there, but what is it good at? What is it not good at?

Greg Bond (08:59):

So I think it's good at things where there is a lot of data. I think that's the general principle in how we think about it. That when we see somebody of trying to apply AI to very narrow sets of data, maybe trying to predict macroeconomic events that might be every 10 years, those are harder things to do. Versus thinking about trade data, stock data where you have thousands and thousands of data points and that allows you to, you know, fit a model and have what we call sort of in and out sample testing. So you have enough data to sort of study and then you have enough data to test that idea. That's important. The other parts, and generally finance has been a hard place to do AI. I think what we look for as a term of art is something where there's a good signal to noise ratio.

Greg Bond (09:41):

So think about when you're teaching a computer to read a picture, you know, it's very clear, right? It's very repeatable. I can identify the picture, but when you get into the financial data, things are often fuzzy and the cyclical noise ratio is not particularly good. So in the investment setting, you're trying to find places where at least it's a reasonable cyclical noise ratio that you can try to fit. And I think the other hard part for AI is this concept of stationarity, meaning correlations might change. The relationship between stock and bond returns may be one thing in one decade it might change in the next decade. And you're trying to have your model fit. And if things are changing that might present a problem for an AI algorithm that can only really look at the past to fit the future. And I think that that's been an area where you will caution and thinking about what kind of time period are you trying to fit your model because you want a lot of data, you might go back deeper in history to gather more data, but that would make you more susceptible to correlations changing and relationships changing, fitting a model in a low interest rate regime versus a high interest rate regime.

Greg Dowling (10:42):

That seems to be challenging, right? Like if you think about finance versus physics, right? Like the laws of gravity are constant, finance changes, right? Correlations change, it's a dynamic market. It seems less usable in finance the way we could maybe use it in science.

Greg Bond (10:59):

Correct. You have to be very careful how you fit the models. You want to caution what we say is sort of what we call overfitting a model, have too many degrees of freedom when you're looking at things. And so one of the ways I describe it, and let's say we are using a neural network, you know, neurons in the brain, maybe one or two for our processes, good, we don't need all of those. What we're trying to do is find relationships that might not be intuitive to us or upfront as we look at data in a more linear world and trying to use a little bit of this technology to help us find where there might be non-linear or interesting relationships that are in the data. So it's just a matter of degree. And I think that that's really where sort of the science gets left behind. It's more of the qualitative stuff. How do we approach this from a fitting perspective? How do we want to sort of take in our experience and thinking about how many degrees of freedom we actually want to let these models have? It's a cautionary tale.

Greg Dowling (11:49):

The whole creativity part of it. I mean there's been, you know, it was a Google data scientist that quit because he thought the AI was, you know, becoming self-aware or something like that. But you know, it does come across things that may not be intuitive to humans. And there's been examples of this in go and chess where it will make a move and the person playing will be like, well that's a dumb move, but it's actually a really smart move. How do you even interpret the data? It might be something that is correct or it could be a hallucination. How do you kind of test this?

Greg Bond (12:24):

In game situations like alpha go or chess and the rules are set and you can play the game over and over again and get a sense that that's actually going to work. I think going back into the financial markets and the real world, obviously the rules change. People are invested in the markets, their behaviors can change like we talked about. So I think a lot of it comes back to the hypothesis that you bring to as an investor with many, many years of experience. You don't want to abandon your experience. I think it's a nice interaction of human and machine to try to get the most value out of these technologies. It's also really, really important that what does come out of the machine, that you try to visualize it and explain it. Because ultimately, particularly in an investment setting, we don't want to get in a situation where an algorithm's underperforming or a strategies underperforming.

Greg Bond (13:07):

And I can't go to our investors and just say, hey, the machine told me to do it. So we spend almost as much time building visualization tools, tools to help you explain what the machine is doing. And hopefully that ties back to your hypothesis that the relationship that you're seeing from that actually makes sense. And sort of using that to reinforce your basic investment process. I think again, it's a compliment to what we do, but I think it's not the only thing that we do. And I think that's the important part is how do we bring these together in an interesting and reinforcing way.

Greg Dowling (13:36):

Now you've mentioned neural networks a few times and I said, hey, that's like how the brain works. But is it also fair to say that in certain computations you do one variable at a time, you sort of move one thing, but a neural network can do multiple things all at the same time. You can type into Chat GPT or Perplexity or any of these apps and it comes out fairly quickly.

Greg Bond (13:58):

There's a speed question of models that are already fit. And I think, you know, in the large language models set up, what you're seeing is a model that's been fit, which is why it can process very, very quickly. In other cases, like when you're in the financial markets themselves and you're trying to react quickly to the market, you need to be vast in terms of fitting the model. It may be even in real time. I think the key too, with the way we think about broad and algorithms and and technologies, there's multiple ways, you know, it's not just neural networks or other techniques. And I think the key when we look at the data is to have multiple ways of doing it. It's something called ensemble. So an ensemble of methods because every method has a pro and a con to it. And the idea is if you can do over different kinds of techniques, you'll get a better outcome. Ultimately you'll feel more comfortable with the results, whether it's a neural network or other technique. It's important that you look at all the techniques and sort of bring that together in a way that makes you feel comfortable with the result.

Greg Dowling (14:48):

Should we fear that AI will take all of our jobs?

Greg Bond (14:52):

Well, I think that gets a lot of the headlines. I think the other side of that, the risk of not doing this in terms of some of the productivity gains that we're looking to see, it's interesting. You look at the broader U.S. economy, global economy, productivity is dropping, whether you look at it at the macro level or at the micro level. There's an interesting paper that came out in 2020, lead author was Nicholas Bloom from Stanford basically trying to answer the question, are ideas getting harder to find and they look across multiple disciplines, whether it's semiconductor research, cancer research, agricultural research, and generally you see research productivity is dropping across all of those fields. And from that paper we were able to look at individual stocks for example, and taking a look at the broader kind of technology names that have been out there.

Greg Bond (15:35):

And you can look at like Apple computer, which is now just Apple Inc. But back way back in the when it was launched Apple computer, great innovation there, great company. But looking at some of the data, you could argue that the research productivity sort of thinking about what they're getting in terms of sales growth per unit of R&D investment or other ways of measuring the investment in R&D in terms of people etc., Has been dropping about 35% a year more recently, right? So that's an issue that affects a lot of technology stocks. Yes, they're managing their sales growth, they're growing, but it's taking a larger investment to get that same level of sales growth, for example. As a result, we need something in this research productivity domain to help us out. One way to offset is to add more people or perhaps using algorithms and using digital researchers to help offset what you can't do organically. And I think there's a broader question on demographics and the shrinking of populations that you might need some of this backup, if you will, from the R&D perspective, that Chat GPT could help fill the gap or the other technologies that we've talked about.

Greg Dowling (16:33):

Yeah, it's funny, if I'm quoting this stat correctly, I believe people were concerned about ATM machines putting bank tellers out of work. And there's actually more people in finance now than were back then.

Greg Bond (16:45):

There's some interesting work by David Autor at MIT, another economist here locally, I'm based in Boston at B.U., where they did look at that. One of the reflexes is that the concern about jobs, it's going to take my job, but on the other hand, it can make you be more productive in that job. So now we can hire more people or you could do different things in the office that are higher added value for your clients or your investors. And I think part of it in the ATM world is that now you can actually have more bank branches because you'd be more productive. That one teller can now do many more things in an office. So there's a lot of fear on the productivity. I get it. And ironically in this case, the technology is coming for some of the jobs that historically have been immune from some of the modernization.

Greg Bond (17:27):

It's software developers, it's other people that are using data that have gone through many years of schooling to try to build up those skillsets. But on the other hand, I think even if you built that skillset now with these sort of co-pilot options, when you're trying to code, it's gonna make you more productive. And I think we can focus a bit more on the productivity rather than necessarily the negatives of potential job loss. But it is an interesting time around that the technology is coming after some of these other jobs where historically had proven immune.

Greg Dowling (17:54):

Yeah, it's not like horse and buggy manufacturers. These are lawyers, these are healthcare professionals, these are even people in jobs like myself, right? We have a lot of data. You can definitely see that it will be interesting. So yes, maybe it can help economically, but should we all fear for our lives? Is this an existential threat? People like Elon Musk talk about it. This is the end of civilization as we know it.

Greg Bond (18:17):

There's that threat that's out there. I think I would focus on other threats or other issues that are potentially more likely, at least in the near term. And I think part of that is on sort of bias and errors that can come through these AI techniques and the data that you fit. And AI does discover the biases that we already have as humans. And that might get reinforced in screening resumes for example, right? We need to be careful there. I think there's a lot of questions around the misuse of this technology, particularly around disinformation. And society, we typically adapt to these things, right? We've started to adapt to phishing calls on your cell phone and your email. We we're more aware of it. But what's a little bit different this time is that you're not sure if you're interacting with a person on the other side.

Greg Bond (18:56):

Is it a machine or a person? These things are so realistic. That's different. I think there might be a business opportunity if you could actually have a proof statement that you're actually dealing with a human on the other side of things. I think that would make you feel more comfortable in some of those interactions. So I'm sure somebody will come up with that technology. So I think we're in an interesting period right now where disinformation is a risk, but we are slowly as a society getting used to disinformation. There's a risk there. We are adapting, but I think there's a lot more that can be done there to help clarify. And then we get all kinds of issues of cybersecurity now with development of these algorithms, protecting your data, protecting your algorithms that you're using. It's very, very important. So that's just going to ramp up as well. So yes, there's definitely the risk of existential crisis in in the world, but I think in the short run we need to think more about the bias, the misuse and sort of the cybersecurity issues first.

Greg Dowling (19:42):

Yeah. And we're coming up to a presidential election. Probably plenty of opportunity for disinformation. So um, Skynet, it's not going to happen? No Terminator?

Greg Bond (19:53):

I won't opine on that, but I will, we'll figure, ask me in a hundred years. I don't think there's any evidence that we've sort of reached this economic singularity or anything like that where you know, humans become superfluous to the society but we should keep an eye on it and it's an opportunity as well as a risk, right? I think we need to balance both of those things.

Greg Dowling (20:11):

So you're not signing up for neural link? Not yet?

Greg Bond (20:14):

Uh, not yet.

Greg Dowling (20:15):

Not yet. Okay. All right. Good to know. Please let me know. And by the way, sometimes if you've ever called you know, airline customer service or your credit card or your bank and you're in some sort of giant call center, I think I may actually rather talk to a robot.

Greg Bond (20:31):

Again, there's pros and cons to all of that. I hear you. If you were dealing with a machine, maybe you could call right through and get right through and get your question answered. But sometimes maybe if I'm talking to the bank, I want to make sure it's a person.

Greg Dowling (20:45):

Yeah. We've been talking about this in generalities. Let's talk a little bit more specifically about Man Numeric and how you are using AI. I'd like to kind of break it up a little bit.

Greg Bond (20:57):

Sure.

Greg Dowling (20:57):

One on sort of non-investment, just productivity on the other side investments, how are you using on the investment side? So maybe just on the non-investment side first.

Greg Bond (21:06):

What we're seeing in Man Group in general has found in exploring the use of Chat GPT, this concept of co-pilots, we have our own sort of Man GPT set up, you know, we're worried about, you know, data leakage, IP theft, all of that. But we've been able to, I think to work through that in a very disciplined way, which I think it's hard for organizations, right? These are new technologies and I think having a specific way you want to use these new technologies is really good to go through your compliance department and your legal department to make it a very clear what your goals are there. So I think we've got sort of that system that we've been investing in. The interesting part is this concept of what I keep referring to as a co-pilot, which imagine you're sitting down and trying to program something and often sitting at a blank screen is always the hardest place to start.

Greg Bond (21:49):

Just like when you're trying to write a letter to somebody, that blank piece of paper, maybe Chat GPT can help you start that language. So the concept of a co-pilot is very similar on the programming side. Have it ask questions and use it to help get that code started so it makes you more productive from that I call the blank screen problem. Also, we have tools that sometimes can be difficult for people to use, maybe data visualization. So it might be instead of you trying to figure out what button to click to look at a performance number or cut your data a certain way, you could just say, Hey, can you just show me how to do this by country or by sector or by stock, rather than having to click all the buttons and it can sort of figure that out for you. So that's another area where it just allows you to interact more fluidly with your technology. And I think that's exciting.

Greg Bond (22:30):

That's what I tell people you're asking me, well how should I handle the world of AI? I'm trying to get into finance and into a career in any sort of field these days. And I think data science is important for many, many fields, right? And I think just having a little bit of programming experience and then using these co-pilots to help even become a better programmer will bring sort of the world of data science to a broader community. That's why we're pretty excited about it from both the day to day trying to to run our operation, but also maybe on the recruiting side maybe we can open up our recruiting to a wider audience. Where today you're sort of stuck with finding people that are both technically talented and creative. Maybe we can just focus more on the creative ultimately because the technical piece will come a little easier given where we are with these technologies today.

Greg Dowling (23:10):

Yeah, we've been using a lot too. Not so much on the investment side but more on the blank page. So it's great for content if you're like, hey I got to write a market summary, can you create an outline for me and get you started? It's also great for maybe we're doing a co-investment and it's on industrial garage doors and you're like, I don't know anything about industrial garage doors. So like where do I start? Like oh I can just type it in there. But you're right, this data leakage thing is real, especially in the financial services area. You can't have IP go out, you can't have clients wiring instructions go out. So you really have to lock it down. It's wonderful, but there are a lot of guardrails you have to put around it.

Greg Bond (23:49):

And we're learning about what guardrails and how to do that. I think that's like any new technology, there's always that learning curve upfront both how to use it but also how to bring it into organization in a way that works. And from the IP perspective, all of the things that we talked about. I think it's also from this sort of summarization, like you said, a lot of information content. Imagine you're a discretionary portfolio manager and you're looking through hundreds of analyst reports. It might be nice just to summarize it down into a few bullet points. Again, making you more productive at your job. Going over the question on how Numeric uses it. So again, as a systematic manager, the actual GPT stuff is useful more from the productivity and the coding side, but from the investment side and from the broader AI perspective, we found it very interesting and very useful.

Greg Bond (24:33):

Going back to the, coming out of the global financial crisis, finding ways to have AI help us learn about the risks in the market, right? We can sit there and we feel the risks. We can measure risks with more traditional models, you know, sector models, factor models. There can be interesting connections that occur in the data or depending on the market environment that you're in. Going back to even going through Covid for example, who thought that Clorox and Peloton would be related in any way that's not cut defined by sectors or other things, it's just the way in that particular environment. And so having machines help you figure that out, helping to understand some of those risks is useful. So that was really some of the first things that we did coming out of the global financial crisis, understanding risk. And then over time it's sort of hit multiple parts for understanding stock selection, bond selection.

Greg Bond (25:23):

There's some really interesting things you can do there with all the data that's available. Taking stock features and trying to forecast returns and doing that in a way that hopefully is additive to what we already know how to do in our long legacy of investing in these securities and trying to prove that it does add value. I think part of it, when you hire researchers, everybody's very excited about AI and wants, that's what they want to do. And we have to sort of bring them back a little bit and say, okay, let's try the simple first and then let's see if the more complex actually adds value.

Greg Dowling (25:53):

A lot of people talk about quandamental, you know, grounded in fundamentals, but you're using quantitative analysis but then there's certainly people out there that just trade based on sort of non-obvious factors because machines can pick up patterns that sometimes we can't, but they're also kind of unexplainable, but they could be, you know, spurious like stock returns and snowfall in January. Go between those, how do you sort of two areas, like you said Clorox and Peloton, that's pretty interesting, right? But then what about other things that tells you you're like, I don't know if I get it.

Greg Bond (26:27):

So the key is that this is where some of the art comes in with the science and having the strong intuition going into it, coming up with ways that hopefully you can not let the machine, as I talked about earlier, sort of overfit the data, right? Kind of find these crazy relationships that might not hold going forward. Looking at the data and trying to interpret the data, you know, having the tools like sort of help you visualize what the machine's actually doing. But it also comes back to your investment philosophy. So if your philosophy like ours is rooted in some fundamental notion, right? We like cheap stocks and they've got good quality and they've got good earnings trajectories and you're sort of rooted in that. If you have that as your base investment philosophy, is there something that you can garner from the AI work that helps reinforce that?

Greg Dowling (27:13):

So refining the models versus finding some completely random thing to throw in.

Greg Bond (27:18):

Correct. And I think the other thing that's important, like any good research from a fundamental perspective is if your hypothesis is hey this should be a better fundamental predictor of same store sales or something that people focus on, let's not just focus on does it predict stock returns better? Does it actually predict underlying fundamentals better? Right? And I think that that's what we call a secondary test, which is yes it's great at picking returns, at least in our back test, but hey, it actually is a good predictor of this company's same store sales next quarter. That makes you feel better that you're actually picking something up. The other thing which is kind of interesting is went through the work. We do all kinds of adjustments to our models historically, right? Certain models have a bias in them. Maybe they always like large cap stocks or small cap stocks and those things typically if they're very persistent and static, they don't really help you predict returns very much. And what we were finding is that the AI technologies actually discover that as well. They actually sort of come up with ways to interact this model with the size of the company, things that took us many, many years to learn and do and implement it actually understood and and developed those relationships by fitting the data. So that also gave us some comfort that it's rediscovering things that as again, organic machines, organic researchers that the digital machines have actually figured that out as well. So that's been helpful to make us feel more confident.

Greg Dowling (28:39):

Another thing I would think it would do well would be on the trading side, like best execution.

Greg Bond (28:43):

Yeah. So that was the other part is I went through sort of the risk and and picking stocks and bonds, but also on the trading side. And there's some really interesting technologies there around trying to identify which broker, hey, which broker do I want to trade with? And you know that, you know, broker A, B and C, they've been doing a really good job, but I've got C, D and E that are also trading. And so what you want to do is sort of this learning process. You want to typically put your flow to the brokers that are giving you the best execution prices, but you also want to monitor what the other people are doing in the market, right? And and maybe giving them some flow to help you learn and understand how they're adding value. And then maybe it changes over time. Maybe you have one getting better, one broker's getting worse.

Greg Bond (29:19):

And one of the technologies that we use is adaptive intelligence routing called A.I.R., which is a way for us to send our flow to the brokers that are giving us the best prices but also sample the other brokers that are out there using that information. And it's this balance of getting the best price but also exploring the other options that are out there. And this is something that's been around and developed even coming out in World War II and I think it's something that's actually helped us sort of get best execution ultimately in the long run for our clients.

Greg Dowling (29:46):

Well I wanted to ask you a couple final questions on AI and more kind of a forecast. So first one will just be AI in general, then we'll maybe ask about what you're working on within Man Group. But what is next for AI?

Greg Bond (30:00):

Well I think we're in that phase on the other side now. You talk about the sine wave where people are now, the hype is there, people get excited about large language models and now it's where does the rubber actually meet the road? Do people actually find productivity gains? How do I take this technology and get it into my organization? And so I think that's really the next step is how do we actually adopt these technologies and make them useful and ultimately see it in the productivity data from the U.S. productivity, global productivity at a company like ours. Are we actually seeing coders being more productive with these types of technologies? And so that's sort of the hype. And then on the other side, the actual adoption of it, I think it's really interesting technology. And I think that that's something that could be, as I said, transformative, but we need to see it manifest itself as we work through our daily routines and seeing the actual impact in our data and our output productivity.

Greg Bond (30:50):

I think that going forward from our side and continual examination of the new technologies that are coming up, can we actually use large language models on the systematic side, not just on the discretionary side. I think there's a lot of interesting questions there. I think there's some interesting questions on when we go to work and simulate and look in the back tests for testing models, some of these technologies may allow us to build different types of environments to test the models, maybe the environments that the market hasn't seen before. And that might allow us to be even more robust in the way we think about the world rather than just relying on actual historical data. Could we use some of these technologies to generate kind of alternative scenarios that will hopefully going forward allow us to minimize risk and draw downs in the portfolios.

Greg Dowling (31:32):

So we talked about gaming earlier and machines beat humans, but humans and machines together beat the machine.

Greg Bond (31:41):

And I think that that's an important point that this is not one or the other. I think together in sort of in a reinforcing way, these things are additive. It really depends on your perspective. I think there are organizations, and I think you mentioned quantamental earlier, I think there's added value to that. The way I think about it is it's hard to do that 50/50. It might be that if you're more of a discretionary manager using sort of bottom up analysis that maybe 20% systematic helps that, right? You're 80% discretionary, 20% systematic. Or maybe if you're a systematic manager, maybe you go 80% systematic, 20% discretionary, right? I think there's a blend there. Culturally it's hard to have sort of forced two organizations together, two investment teams together. One's quantamental, one's discretionary for example, but very systematic versus discretionary. But having the ability, I think there's information in both sides that help reinforce what you're trying to do. And I think that's the fun part about this business is trying to figure out best ways to do new things. And there's not always just one answer. There's multiple ways to compete in the market and I think a good blend of that is really interesting. At the same time you have to approach the world with a very strong investment philosophy and don't change your philosophy because of the technology. Figure out how the technology fits your philosophy. And I think that that's really important to kind of stay rooted in that.

Greg Dowling (32:58):

Very wise words. If listeners wanted to learn more about AI, are there any books or papers they can read?

Greg Bond (33:05):

I'll be very self-serving here and point listeners to a couple of different things that Man Group's put out there recently. Barton Luck, one of my colleagues over at Man AHL, put an exhaustive study together overview of all of the generative AI literature. It's on the SSRN Social Science Research Network. It's very exhaustive, which is really helpful if you want to get a very quick read or a very long read depending on what level you want to go into that paper of what's going on in generative AI. Also on the Man Institute website, there is a summary into transcript of an academic advisory board that we did a few months ago around generative AI, where we brought academics and people from Man Group together to discuss some of these issues that you've brought up today.

Greg Dowling (33:42):

Very good. Well I appreciate that. When you're not training robots and doing artificial intelligence, any hobbies that you spend time on, Greg?

Greg Bond (33:50):

I've been spending some time on sports analytics. That's always another area of fun that fits nicely with the systematic world on equities and other things. But sports analytics has always been a hobby of mine and continue on the side to do some of that fun stuff.

Greg Dowling (34:04):

Does it work?

Greg Bond (34:05):

I found some interesting things. I think Moneyball and you can take that concept of Moneyball and apply it to almost all fields now. You can watch the movie and they did it in baseball, but you're seeing it all over the place now. Politics, wherever there's a lot of data, there's information that you can use to help you make better decisions. And I think that's exciting. And I look at my son in college, every major that's out there, whether it's humanities or science or whatever it is, there's some kind of data science path that you can take. Communications, English, Language, Data is becoming more and more important and I think it's important for people to start to interact with that data. And I think with generative AI and these technologies, the barriers to that have really dropped significantly. So I would argue people should just go out and start playing with some data, whether it's Python or R or whatever the technology, just get out there and start looking at data.

Greg Dowling (34:51):

Also good advice. Hey, on the sports analytics, because I think it's interesting. So you're talking Moneyball you're talking about the Michael Lewis book that he wrote about Billy Beane and the Oakland A's and very low budget team. They had a find an advantage, they decided to use quant and stats. Nobody was doing that. Now everybody is doing that and maybe taking that to markets and sports, how do you figure out like, okay, this works because nobody's doing it versus oh now everybody is doing it.

Greg Bond (35:25):

The advantage that they had has largely been exploited and now everyone, as you say, is doing technology and data science. And there's some very interesting stats on, I think there's something like 750 people in major league baseball, either in total across all of the teams doing data science or software engineering. So that gives you a sense, at least from the movie, it's like two people sitting there with one computer. And now I think on that spectrum, the A's are actually at the bottom and the Rays are now at the top of that and sort of how the market's evolved. And so I think part of it is you have to do that to stay ahead and a continual investment and going faster. But there's also ways that you can try to think differently and maybe you're moving from pure data analysis into more how do I take scouts information?

Greg Bond (36:07):

So take that discretionary thing from that side and blend it with the systematic, like we talked about. So a scout observing a player, can I use that in my otherwise systematic process? And so I think there's a lot of interesting quantumental work that people can do, on the sports analytics side. And I think you're seeing a lot of that and the more sophisticated teams are trying to figure out how do I digitize the swing and use that to help predict whether that player is going to be good or not, right? I think that's really fun and really interesting, but like everything in life, you have to continue to invest in it to stay ahead.

Greg Dowling (36:36):

Yeah, no, I mean, and you're seeing it live, right? Like Stevie Cohen bought the Mets and he brought his, you know, quant team from point 72 to help with the Mets, which didn't actually help too much last year, but maybe it was too small of a sample size.

Greg Bond (36:49):

Yes, we have to see it. Yes. I think that's over time we've seen these techniques work and it still helps to have a big budget too. So I think blending the analytics and the dollars is ultimately a really good way to go about it.

Greg Dowling (37:01):

Pulling on that thread just a little bit, because I actually think Moneyball, although it's a sports book, is a great investment book. And so if you have any interest in sports and investing, it's a perfect one for you. Hey, thank you so much. I've learned a lot. I hope our listeners have too. So thank you Greg. Thank you very much for joining us.

Greg Bond (37:20):

Thanks for having me.

Greg Dowling (37:21):

If you are interested in more information on FEG, check out our website @www.feg.com. And don't forget to subscribe to our communications. You don't miss the next episode. Please keep in mind that this information is intended to be general education that needs to be framed with unique risk and return objectives of each client. Therefore, nobody should consider these to be FEG recommendations. This podcast was prepared by FEG. Neither the information nor any opinion expressed in this podcast constitutes an offer or an invitation to make an offer to buy or sell any securities. The views and opinions expressed by guest speakers are solely their own and do not necessarily represent the views or opinions of their firm or of FEG.