Introduction

Thank you, Thomas. I promise this will be the least technical talk of the evening.

Thank you. If not this evening, maybe this week that you're going to hear.

So I run an investment firm called Goose Hollow Capital. We're a macro investment firm.

We run some managed accounts. We also run some public ETFs.

Overview of the Talk

And I'm going to share about how we've been using AI in our process. So I think this is like a practitioner talk, so I'm going to talk about more practical things in this one. And so I think there's three things I wanted to talk to you about today. So the first one is, how has the investment landscape changed? What has changed in the last 25 years that I've been involved in markets and in finance? And that's something I want to talk about.

The second thing is what is that we do as investors and what is our process? And the third one is how do we use AI in what we do, right?

So this is something my compliance guy told me to put it in there, and it basically says that some of the stuff I'm going to talk about is probably not true. So...

Changes in the Investment Landscape

So if you think about like 2001 or 24 years ago when I was in grad school, I studied machine learning. It used to be called statistical learning in those days.

And I came to Wall Street and I said, hey, listen, I studied this thing called statistical learning. Can we do something with it?

And there was no data. First of all, there was no data and the data that was there was extremely noisy. And then the other problem we had was we didn't have any compute.

You know, remember back in the day Pentium 4 was one of the best computers you could have, right? So the two big things I think that has changed for finance and I guess for all of what we're doing now is these two things, right?

We have an exponential growth in data, and I'm not just talking of market data, but every other possible kinds of data, right? So for instance, like, How many of you know that the US government has sensors in every ocean on Earth, and you can actually read those sensors in real time? Not quite real time, but quasi real time.

So if you think that there's a big storm coming, you could know weeks ahead of time if you were reading those sensors, because obviously the water temperature in the Caribbean and elsewhere has been going up for a long time. So there is an enormous amount of this data that's become available, right?

And that data is possible for us to use because we can store all of this data and we have the compute to actually do the compute with it. And this is just the chart on the top left is how much of the GPU and CPU speeds improved. But on the bottom left, you can kind of see that's in zettabytes. Zettabytes is 10 to the power of 21 bytes. So there's 21 zeros after one in terms of how much that is, in terms of how quickly that data is growing.

Investment Process

So what do we do in our daily lives? So if you think about a typical investment firm, this is kind of the process of what we do. We start with, we try to analyze the market.

What does that mean? We get a bunch of data, economic data, market data. and other information, news that gets released and other announcements that come out. So we try to understand what's happening that and analyze that better. And that's sort of what I spend a lot of my time on, understanding what's going on.

What is the story? Then you have this second step, which is like, OK, now that I know what the facts are, what the ground truths are, I need to identify what is the investment. What does this mean?

OK, Iran threw a bunch of bombs at Israel yesterday. Israel might retaliate. But what does that really mean? What's going to happen because of that?

So this is sort of what we do is we spend a lot of time between the first three steps, which is analyzing what's happening in the market real time, identifying trades, and what I call investment research, which is more like given that we know something, it's kind of independently verifying facts. So given that you've heard that this is true, can you actually go back to the actual source of the data and verify that's in fact the case?

And then we also spend a lot of time model building, right? So, you know, these are mental models. So if I tell you there is some great battery technology out there, you know, the sodium batteries are going to be a thing, you need a mental model of what that's going to do, right? And you need to build that model. Now, the model doesn't have to be right. It just needs to kind of help you think about the problem, right?

Okay, what does sodium batteries mean? Sodium batteries are going to be cheaper than lithium batteries. they will make storage way less expensive, and you might end up not needing as much natural gas and coal and other things we have, because we would probably put a lot more panels, store all of those energy that the panels generate in the batteries, and sort of solve our energy problem.

So, this sort of model building is what we spend a lot of time on, and it's a lot of work. It's a bit tedious because sometimes we're not experts on things, right? We don't know anything about sodium batteries. I'm not a sodium battery expert, but I need to know because it's going to have implications for everything I do.

Scenario Analysis and Counterfactual Reasoning

Similarly, the other two things that we spend a lot of time on is doing scenario analysis and what I call counterfactual reasoning. So most of the time when you come up with an investment idea, it's quite easy to kind of find all the reasons why you're going to be right. Because the human brain works in a way where we're always looking for things that are going to confirm what we think is going to be true and what we think is true.

So it's very hard for us to find counterfactual examples or reasons against ourselves. And this is something we always struggle with. And I'll show you, the LLMs are pretty good at doing that sort of thing where they can actually reason against some of the stuff you're saying and poke holes in that.

Financial Reporting

Then the last one, the most boring one, but I think where most of this is going to get used, is in financial reporting. We have a monthly report we have to send to our investors.

It's completely automated. The LLM goes to our database, picks up all the data, and writes the report. Of course, we have to review it, edit it, and so on.

But we don't actually do much work now because all of that is automated. Same thing with, like, you know, we have weekly updates. You know, we send out weekly updates to some of our investors.

We say, oh, this is what we think is going on. All of that is automated, right? 1We tell the LLM the things that it needs to talk about, we give it what we think is going on, and then it generates the video.

So there's a lot of this sort of stuff that is like the stuff that you've already probably seen already that people are doing.

So in terms of like, you know, thinking about all this stuff, one way to think about it is to say, you know, what are the things where there's a high level of automation possible and where there's gonna be actual strategic impact, right? Like where, you know, doing this is gonna save you a lot of dollars and also probably make you a lot of dollars.

And that's the top right, right? The market analysis part of it.

We've spent a lot of time on the bottom left because that was the easiest thing to do. Automating all the reports we sent investors or reviewing certain transactions like files we get from our counterparties, all of that stuff, automating that stuff was the easiest thing to do. But it's not really going to necessarily change how we do our stuff.

Now the thing where it's very difficult to do is the one on the top left, which is the economic and scenario forecasting. Because LLMs are not really, they're language models, right? They're not econometric models.

Teaching them how to call econometric functions and building models, that's like a whole other game and we've been doing some work on that. But it's highly impactful because obviously the LLMs can help us think through a certain problem well and build a mental model for it.

That'll help us invest and invest better and come to decisions quicker. So this is the way I think about this.

Use of AI in Investment

I'm going to show you a few examples of some of the things we're doing. I really like this quote from Herbert Simon, which some of you might have seen. But basically, he talks about how we are overwhelmed with information.

The amount of data, the amount of news, the amount of social media content and real-time information that we get is just ginormous. And there is no way for a human being to actually process this information in any sort of fashion, right? So that's why people use machines to do it.

So if you're a small firm like us, what we end up doing is we use LLMs to kind of help us summarize this information and process this information, right? So in this example, every email, every research we receive, we essentially read it machine read it, the LLM takes the key points, highlights what the trades are, and sends me a Google chat.

It says, hey, listen, Morgan Stanley is telling you to look at Mexico now. And if I actually want to read the actual piece, I can go and open the piece and research and read it. So this is now, to me, saves an enormous amount of time because the amount of

things we have to track, substacks and blogs and websites and actual news that comes out. It's just ginormous, right? It's not possible for us to do it.

And this is something where we've gotten a lot out of this because we have it reading every piece of research we receive. The other aspect of this that I was saying to you was the research and exploration part of it, right? So one of the problems we have is that sometimes we don't know anything about a problem, right?

And you're very... I'm kind of hesitant to say that to someone, to say to someone, I have no idea what to do about this. But if it's in an LLM, I'm much more open to going to the LLM and saying, listen, I have no idea how to think about this.

Can you give me a framework for this? So this is an example where what we do here is that we want to understand why South Korea doesn't have enough children. Why are people in South Korea not having enough children?

So what we do here is we go to the LLM and say, build us a system dynamic model, right? So system dynamic model, some of you might be familiar with it. It's a dynamical system model.

We have stock flow, stock and flow variables and relationships between them. And we can actually elicit the LLM to give us all the variables that are important for this particular problem.

And so what you will see here is that it comes up with this graph, which is the population in South Korea, which is what we are interested in, is a function of the urbanization rate, which is actually a function of housing affordability. Obviously, if you have you know, house prices go through the roof, people are not gonna have children, right?

Because then you can't really, so what is the reason, primary reason that, you know, South Korea's population is declining? Well, these are all the variables and some of these variables might be changing. So the government's actually now trying to improve housing affordability, right?

So that'll have an impact on the population six months from now, a year from now. But that understanding how long that's gonna take for them to implement the policy and then the actual impact in terms of the population growth, might be three years from now, right?

So that's again something where the LLMs can help us because they have all the variables and they have all the relationships as well. Now, the problem sometimes we see when we do this sort of exercise is some of the variables and the relationships are often wrong.

Like it'll get the sign of the equation wrong, or it'll have the magnitude wrong. But that's something that you can always deal with, because you have an econometric model, you have a way of estimating that, so you can actually deal with that. But many times, we don't know what this graph should look like.

And that's where they can be very helpful.

Practical Examples of AI Usage

The other aspect of this process, and I wanted to show you a little, this being a talk and showing a little demo might be good, is also like we started to work with the LLMs with some of our internal data. Obviously, we have a lot of little tools we've built and I'm going to show you one of them. This one's just using ETF data.

What we're trying to do is, We want to understand the impact of various events and things on the performance of certain ETFs, right? And so how we did that.

So in this case, we're interested in these five kinds of analysis, you could say. One is, of course, we want to know, OK, how did a particular ETF do when a certain event happened?

The other thing is what happened when interest rates went up or interest rates went down, what was the consequence on a certain ETF after that, and then geopolitical events like, okay, Russia invades Ukraine, what happens to certain markets and so on. So let me see if this thing will pop up here.

Perfect. So this is our ETF app. Now, of course, this is like

anybody can pull it up and the UI isn't that great, but this is our public-facing UI, which looks not so good, but it is functional. It does what it's supposed to do. So in this case, just to give you a flavor of what it can do.

So for instance, if I want to know, because NVIDIA has been in everybody's mind, I want to know every, Every ETF that has more than 5% in Nvidia, it generates a list of tickers and gives you a little summary of all the ETFs that have got more than 5% in Nvidia. So in the background, all we're doing is it's a text to SQL application, right?

So we have a bunch of data in our database where we have calculated all of these sort of things. Some of them are done on the fly. So if I ask, okay, what's the six-month correlation of the TLT and S&P, which is the Treasury ETF and US benchmark, it's going to come up with like,

the correlation and it's going to calculate it on the fly. So you can do a lot of these calculations on the fly that we're interested in. Of course, one of the issues we have with using something like this is that, particularly if you expose it to the public, is that the data has to be grounded.

We can't have it make up stuff. So we have to give it data and say, you can only use this data and if you don't have the answer, just say, I don't know. Because obviously, there's all these complications that come when the LLM just makes up stuff that's not really true.

Being in financial markets regulated space, we'll be very careful about the whole hallucination problem. The idea of this tool is not to answer every possible question, but it's a very open-ended tool.

You could ask questions on, give me the ETFs that have done well when Trump won the election. You know, I'm just trying some canned queries here, but we can ask any other questions as well. And it kind of comes back with, okay, when Trump won the election last time around, you know, these were the ETFs that did well.

Happens to be the short volatility ETF because volatility went down. And then there's a few other tech-related ETFs that did well as well.

So...

Conclusion

Yeah, so that's really the talk.