Invest with AI
00:05:05 · Doug Clinton
[00:05:05] S0: Yeah, I'd say very early on, one of the things we sort of realized was in '23, the paradigm for using LLMs effectively to invest was prompt engineering. I mean, that was all the rage then and really understanding what what is the context that you're providing to the models in terms of the task you want it to output. And so like you just said, Brett, if you just say, give me a stock of 10 portfolios, you're probably gonna get, I would say, eight of the 10 largest companies in the S and P 500 would be my my bet, which is fine. That might be an okay portfolio. But I think when we think about really using AI to make intelligent investments, that's not what we want. And so giving it a different framework where, you know, maybe you narrow the field to small caps, and maybe you define that within certain parameters for market cap. Maybe you pull in some external data from sources where maybe it's fundamental data, maybe it's your own research about a handful of companies. I think that's where it starts to get much more interesting using these language models and seeing how they sort of interpret what is all the data that's out there that humans are are looking at. Right? And and language models can process way more qualitative data and quantitative data than any human analyst can. So trying to figure out how can you get that right data and the right selection set into the models is kind of step number one. And then I think as I'm sure you guys have seen too, right, the frontier has evolved very quickly where it was all about prompt engineering a couple years ago. I think now it's about agentic workflows, you know, orchestration. And I think even beyond where we're at right now, the thing that we're getting really into is how do you sort of manage organizational context? You know, instead of just narrow task context, you know, you're trying to pick a large cap portfolio. That's one specific vertical that you might be working on. How do you look at the entire corpus of every portfolio you've ever generated? Or maybe if you have multiple team members generating different portfolios with their agents, how do you sort of understand the entire broad, knowledge base that's in your company and make it useful to the models when you're doing investment work.
00:07:15 · Khe Hy
[00:07:15] S1: Doug, on that point, did you mention kind of the evolution from prompt and it seems like we went from prompt engineering to context engineering to agentic workflows. And now with kind of the Fable like models kind of intent engineering or, you know, output based prompting, I'm curious how that last bit... I mean, we're still early in the Fable five days, but, has that last bit changed your approach?
00:07:46 · Doug Clinton
[00:07:46] S0: I'd say our approach is always changing slightly, and I think it has to, actually. And and I think about these models in some ways as a reflection of of how alpha changes in markets. You know, what what might generate alpha today probably won't generate alpha tomorrow. And whatever you're doing to generate alpha, and and that's a very broad word that probably needs a real definition, but let's just say outperformance relative to some benchmark right now. Whatever you're doing to generate that outperformance today with the models probably won't work tomorrow because there are other people that are experimenting with these models or finding out creative ways, I think, to make investment decisions. It might be much more of a human oriented process than what we do at Intelligent Alpha where we rely very much on the models to kinda work end to end. But, you know, I think we're always trying to change our approach and see what has sort of maybe stopped working based on what we were doing in the past and all the data that we have over the past three years. And then our intuition, and I think this is where the human still comes in, sort of our intuition of where the market's going, what might work in the future. And in some ways, that might be a little bit like how some quants operate, where a lot of times thinking of the next factor or thinking of the next sort of dataset that you might believe has alpha is much more of kind of an intuition than than just math.
00:09:07 · Brett Caughran
[00:09:07] S2: Yeah. This is this is a huge structural difference versus building code, for example. One of the stats I've been... I sort of pointed to all my students is only 17% of stocks outperform the S and P 500 over a rolling ten year period. So probabilistically, that stock you pick at random is going to be, a stock that underperforms the S and P 500. Sort of broadly discusses 70 of long only managers underperform the S and P 500. And so public markets alpha is truly a power law game. How do you how do you think about capturing? How do you think about capturing that power loss or right tail alpha essence in the agentic structure? Because if we're sort of like taking our investment process and turning that into an agent structure, if I'm a median investor, I'm actually augmenting a median process, which structurally does not have the essence of that of that alpha capture ability.
00:10:09 · Doug Clinton
[00:10:09] S0: I think what's interesting to think in that context is is how much of the intuition coming from the LLM is valuable versus the intuition coming from the human. And as I said, with Intelligent Alpha, the bet we are making is this long term bet that over time, the models do continue to get better. We've seen that in the benchmark data. And that ultimately, probably leaving the models to their own devices with very little sort of nudges or interpretations or or influences from humans is the best process. But today, I actually think that's that's a little bit different where I still think the creativity of a human to say, you know, I wanna be 30% weighted in some stock because I have really high conviction that it's going to work for reason x. Those types of intuitions, and maybe there's some deeper research that goes on behind the scenes, things that an AI can't get access to, you you know, those things, I think, are still the core blocker from just turning the whole thing over to AI. And so what I would say we've seen work really well right now is almost this sort of, like, hybrid model between what would look like maybe a traditional quant portfolio a traditional fundamental portfolio. And so, generally, our portfolios, they don't have a thousand positions like you might have in in a quant portfolio where it's literally... Right? You're just looking at numbers, a lot of small bets, and you're trying to get that 52, 54% batting average. You know, we generally have something on the order of a couple 100, you know, maybe up to 500, depending on what strategy we're looking at. But there are, in our portfolios, some conviction bets that we allow the models to make that may look like, you know, several percentage points in terms of concentration in the portfolio. And AI, I think, has been pretty good in what we've seen so far in our portfolios at knowing where to be convicted and and maybe where to dial it back a little bit where it doesn't have great conviction, and it's gotten better over time to the point of our thesis. And so, you know, I think that humans can still play a really valuable role, which is bringing in that data that the models just can't get access to, pushing the models to take really big bets.
00:12:18 · Doug Clinton, Khe Hy
But if you're not gonna have that human influence, sitting somewhere between probably a traditional fundamental process and a traditional quantitative process works really well with the LLMs. [00:12:30] S1: How important is that data and what role does it play? Is it, you know... Yeah. Could you talk a little bit about the the data, the types of data that you bring into it?