#33 Jacob Ayres-Thomson & Hassan Salamony - How can AI leverage vast amounts of public data to empower investors.

The Nalu Finance Podcast

In this insightful interview with Jacob Ayres-Thomson and Hassan Salamony, we discussed how Artificial Intelligence can leverage vast amounts of public data to empower investors. 

Jacob and Hassan dive deep into the power of AI in financial markets, including:

• How AI extracts actionable insights from vast, noisy public data sets

• Why explainability matters—and how 3AI helps investors understand AI-driven predictions

• How their models aim to detect surprises before markets react

• Where AI is heading in the investment landscape

  
  

They also share their experiences from Goldman Sachs, PwC, and beyond, offering a rare look into the intersection of machine learning and institutional finance.

🎧 Listen Now On: Apple Podcasts | Spotify | Youtube | Podomatic

Hassan is the COO and co-founder of 3AI and Jacob is the CEO and co-founder of 3AI.

🎙️ Transcript:

Stefan Wagner: 00:42 Thank you very much, Jacob and Hassan, for taking the time to speak with me today. 3AI, you know, Jacob, you founded it. Hassan, you joined shortly after that. What motivated you actually to establish 3AI Investment Solutions?

Jacob Ayres: 01:00 The motivation really wasn't to establish a company. The motivation was to solve the problem. And the problem is to understand what drives stock total returns. So we're on a quest to look at everything, to see everything, to understand everything.

Stefan Wagner: 01:19 And who then can use your solution? Who is your, in a sense, what are the benefits of, and who is your ideal client to use what you can offer?

Hassan Salamony: 01:29 So I think once you've developed a solution which is better assessing stock returns or identifying what are the key drivers of stock returns, the application of that is actually pretty broad.

Stefan Wagner: 01:43 Who is, in a sense, your ideal customer?

Jacob Ayres: 01:47 Anybody who cares about the future performance of stocks. So at the moment, if you look at our website, the two core products that we're starting with, it's insights and indices. So indices, essentially, we're creating enhanced indices, so we can rapidly filter to any thematic, country, sectors, whatever people want to look at, and then create an enhanced version of that.

So the best enhanced version is global. The more stocks you look at or have an ability to choose from, the better your performance. And actually, I've written a paper we haven't published on that as well that shows that as mathematical truth through stochastic simulation. For a given skill level, the more things you look at, the better.

So it's a bit like if you go to one of those talent shows and they have 1,000 people or apply for 10, well, the one with 1,000 is probably going to find a better singer. And then the other is insights. has been an interesting area. We originally started speaking with the hedge fund community, sharing with them our predictions. The feedback was very strong. They were finding significant alpha.

Stefan Wagner: 02:52 Which actually, can I jump in here? How would you define alpha? In your mind?

Jacob Ayres: 03:02 I would always go back to the person asking me and say, how do you want me to define it? How do you generate alpha? I'll answer. How do you generate alpha? By having some level of accuracy and prediction on the future. That's the first point. And the second point then is how do you execute and all the rest. And alpha is that final thing. So, every hedge fund is supposed to be finding alpha, while each one of them is looking at different things. They've got different frequencies of investment, their strats, does it work with their strats? Basically, for them to find alpha, we need to be seeing information that gives predictive accuracy on the future that they're not seeing. They don't care if we have the same predictions as them, we're useless to them.

They go, guys, you're doing a good job, but it's useless for us because it's the same that we're doing. that the information that they're seeing, they're finding it good alpha, they're finding about half of the recent funds at about half of the alpha, which shocks us because that's a company with over a thousand employees, managing lots of capital, and we're small. But basically, they're finding orthogonal information in what we do. But what we found with funds' recent feedback was actually that these very advanced systematic funds, they want to develop their own or do develop their own systems.

So they've got many quants or buildings on the models. And they realize, as we do, that the real key is engineering information for machine learning to look at, more than it is on the machine learning. in terms of stock prediction. In stock prediction, you don't have as much data as say, for example, chat GPT, which is basically learn the entire web and you can test it. It pretty much knows everything. And it knows things it shouldn't know, by the way, but we'll leave that out of the chat.

So what we found with hedge funds is what they're really interested in is interesting new data sets. Now, we've created a data set that is absolutely perfect for hedge funds because not only is it perfectly mapped already to securities and specific to every security, But it's already essentially vetted by us through machine learning that it's useful into the forecasting problem. But also, there's, if you like, a not repeatable scientific method which uses the human brain, which is critical to virtually everything humans do, and yet we don't seem to work out how to create, automate, and share it, which is common sense.

And so thinking about really what is a stock, all of our factors, we've never removed a factor. We simply add more. But all of those factors have to pass a common sense test. We think about companies as colorless, odorless cash printing machines. From the perspective of a shareholder, we don't care what they do. We simply care, are they printing cash, the speed at which they're printing cash, the speed of making change at which they're printing cash, and does this printer, is it small, cost little, or large, cost a lot, because you've only got a limited amount of cash, your floor space. And does it have kind of smoke coming out of it, electrics coming out the wire? Is it still going to exist tomorrow? The probability default, which for an equity investor is twice as important typically as a debt investor, because an equity investor loses 100% debt investor only loses on average 50 or so.

Stefan Wagner: 06:03 One side, how do you ensure, for example, that there's no bias in there? I mean, you know, the minute you build something, usually there's bias in it, the AI is built still by you need to find bias for me. I mean, I presume, you know, by selecting which check of which factor you for example, or you're leaving in or if not, that common sense, you know, you could have a bias towards certain things, you know, your common sense probably different. Definitely, I experienced this right now, my common sense is very different than my common sense of my 14 year old teenager.

Jacob Ayres: 06:36 So that common sense, if you like, is that we're talking about is it's not to have story based data in there, like we said, like some ESG data might not relate to long term welfare creation of a business, but will definitely show alpha over the last 10 years because of an influx towards ESG investing and requirement being coming down from loads of investment companies and other clients of theirs to have ESG in their portfolios, that creates an alpha whether or not those companies are outperforming. But if those companies don't outperform in the long term, as price reverts to value, they're going to underperform if the factors didn't actually relate to the long-term ability of the company to outperform in wealth generation for shareholders.

Stefan Wagner: 07:13 I think it goes a little bit towards what you said. ESG is a theme a little bit if it doesn't hold up.

Jacob Ayres: 07:17 Well, some of it and I think some not because, for example, employee satisfaction How happy are employees? Some ESG people are getting that in their data. But it's a pretty damn obvious thing that a happy human is more productive, more creative, thinks clearer. And companies are just a combination of legal contracts and humans and cash at bank or not. and ideas and IP. So, a business is fundamentally a human thing, but there's been quite a lot of talk in AI about bias, and I think it's probably one of the most misdirected conversations in AI right now, because the bias that has kind of, if you like, disturbed people about AI is usually just the AI becoming a mirror to us.

So if we look at bias with regards to things like, for example, looking at demographics of people and taking a different view on them, etc. That's because the data sets it's trained on have that in them. But AI doesn't really have bias. It's like Kasparov said, I think the reason why AI is so powerful is it does not contain the prejudice of the programmer, it simply learns the truth.

So the key is to give it good data. You see, and in our world, it's kind of trying to put on that Warren Buffett or Terry Smith cap, but also bring in the bit of the trader thinking as well around technicals and other information sets that they wouldn't really be able to process as a human as well. and enabling the AI to dig in and do the research and ask all the questions, but it doesn't contain bias.

The AI, in our sense, is actually over 100,000 different AI models. Each of them is correct, based on analysis of the past of what it would feed back. and averaged across the lot, obviously a much better result. It's called Ensembling in Machine Learning.

And so what we're really getting out of our AI is, at the click of a button, an inordinate amount, an insane amount of research effort, which is statistically validated compressed into single number and explainability around those numbers, the AI explainability that we've subsequently developed for it as well, that enable us to go factor by factor level, our 326 factors, we can see two key aspects for every factor. How much alpha is expected to contribute for a stock like this, Tesla or Microsoft or Google. But also how influential was it in deciphering the alpha? Because you could have a factor where it says, okay, we're not expecting any alpha out of that. But it was a very influential thing. I've seen lots of positives and negatives about that, because the way it looks at everything is all conditional.

So if you look at medical research today, and you begin you go into machine learning, You look at medical research, it looks like something from a century ago. Because they go, okay, we're looking at cancer patients, we want to look at five year survival rates, and we're going to test people taking this tablet or this injection or this surgery, and look at the five year survival rates, looking at a single point in time to the future of whether people died or lived. according to an influence of that cancer, doesn't look at any other aspects about that person at all, but would learn so much more if it did.

Whereas our machine learning is looking at, and typically is funded through some research at 200 people, 100 people, like a tiny data set. We're looking at equivalent of 160,000 people, 160,000 stocks. And the AI has enough data that it can ask lots of questions, it kind of dots on the questions, so it learns a lot deeper about the attributes. That's why our live performance is significantly higher on a different Richter scale to the requirement for FDA drug approval. So the standard requirement for FDA drug approval in the States is 95%, so about 5% p-value. So a 19 out of 20 chance that it's right. Whereas ours is beyond trillions that it's right. And that's in multiple live examinations. How do you deal with

Stefan Wagner: 11:24 Surprises. What kind of surprises? So for example, Silicon Valley Bank, for example, suddenly there was a literally a loss of trust inside the company. I mean, because this is a herd behavior is a bit irrational. If everybody would have left their money there, probably the bank would be fine.

Jacob Ayres: 11:42 So, to the extent, and only to the extent, that the information that we look at could have given insight into that happening.

Hassan Salamony: 11:51 So, I've just pulled up First Republic, so First Republic is still around, unlike a couple of the others. But this is probably not dissimilar to the detection that was occurring in these banks as they're running into some of these issues. So the ranking is very poor, it's 6.3%, so it's only better than 6.3% of our 20,000 plus universe. Very, very poor alpha forecast, you know, negative 10%. And then if you look at the factors that are really driving towards this, There's concerns around the business model. The risk models, the credit risk models are starting to flare up. There's changes on the balance sheet which are causing alarm.

Shareholder treatment is registering as a concern. Technical indicators, so what is happening with the stock price and performance is coming through. Valuations is registering positive, but that's because it's trading cheap. but it's not enough to offset all of the other features. So that there is your warning signal, that's your predictive value, and what we know is that actually when these things do start to flash up like this, there are problems ahead, you know, and often, sadly, we're actually seeing those problems unfold.

Jacob Ayres: 13:11 We also look at the amount of influence in how the AI thinks from factors, so not only the force of positive or negative alpha, but the influence, how much weighting is it putting into it as well. What we've discovered is that the machine learning system essentially thinks differently about different companies, much in the same way that, say like, Amazon might think a person like you that's bought this might want to buy that. This is doing that effectively through historic data. A company that looks like you, I would put less weight, so in the example of Carillion, onto prior historic financials, which had dropped by about 86%.

So basically didn't trust them anymore. So I've seen a price move like this. Your historic financials now make you look really cheap. I'm not actually going to trust them. I'm downgrading the influence of this cheap valuation by about an 86%. Another interesting one, if you compare, that we looked at was like, say, for example, Tesla has had a very high weighting towards analyst forecasts, earnings beats, but particularly things that are coming out of the minds of humans about the future. Because, essentially, it's still an option on all the wonderful things that could be achieved, such as like putting an AI robot in everyone's house and then basically turning the Tesla fleet into an Uber business and taking the whole thing out, you know, monopolizing cars, essentially.

Now, when you look at Berkshire Hathaway, which has quite predictable financials and a very stable business, you know, Sears, Candy, and all the kind of things that he does. Coca-Cola, exactly, yeah. We see that it consistently has a high weighting to the historic financials because it's recognised as a very predictable business. And so when one of the things that the Goldman's systematic investing conference that has a lot to do, I saw the people looking very much at market timing on factors, what we have essentially, the machine learning has discovered is the unique DNA, the unique factor DNA of every single company.

So every single company is essentially having different weighting to factors is learning what matters on a company by company basis. And it can change over time as the company changes. Yeah, well, it also has the company changes. Yeah. So the learning in our system is essentially Deep but not changing much anymore because each week we have another years of that of data So a year before to the following week, but we have that going back since the mid 70s So the percentage change week to week is is tiny now the space that we like to operate in or kind of occupies is enhanced indexing and

Hassan Salamony: 15:34 And effectively by that, what we mean is we can create similar attributes to an index, but sitting behind it is you've got the brain and the power of the AI doing the stock selection. And that can be very valuable in terms of future returns. And when we run our analysis versus different types of factor metrics and comparisons, for example, Fama French, what we typically find is that our alpha, our performance is all coming from a stock level.

So actually, you're getting the best of both worlds there. You're getting a product or an investment that is behaving and operating like a more traditional index in terms of volatility and risk controls and sectoral weights. But actually, you're getting the alpha that be coming from a much more sort of hands-on, bottom-up, active stock selection approach. And the combination of the two is very appealing and is actually very scalable as well.

Stefan Wagner: 16:29 But I assume you also need to make sure because I've seen this quite often when looking at funds. There are some very, very good stock picking people who work together in a fund, but they have no real view on the overall combination of all the stocks that they picked. How do you deal with that one in here?

Jacob Ayres: 16:47 Well, because the first thing is, and it's one of the great strengths of artificial intelligence over biological intelligence, is that it knows everything that it thinks elsewhere. It's a connected intelligence. We actually, interesting to that point, when we look at top-rated stocks, it's a mismatch. some look like growth stocks, some look like value players that aren't having much growth, but are very, very good yield and everything's rosy. There's certainly no signs they've got some kind of like utility type monopoly or something, but some do look like growth stocks.

So, it's learning to slice and dice different ways to find expected outperformance over that given time period. Time period is a very important question. I think that plugs in and those investors may or may not be aware of that. But the time period of forecasting does affect which strategies are better, clearly. If we put ours to shorter term, it becomes more technical, more sentiment. If we put it to longer term, 10 years, it starts throwing away technical, nearly all – like very little information. It's all about the long-term financial lookbacks. Even analysts become a bit less useful at a longer term.

Now, if you're forecasting across different terms, you can both be right and disagree with each other. So, like a very short-term trader can go, I'm going to go long on, let's say, oil, and they're completely right over the next 30 minutes. It's a brilliant trade. The investor goes, I'm going to short oil. And I'm thinking over two years. And they're also right, because over two years, it drops 30% in price, and they do well as well. So they're both right, but they're both going in different directions. But if they have the same term of thinking, so if you get those humans to talk to each other and say, well, over what term exactly do you think this is a good investment? Then they can actually combine their information better.

But if their terms are different, there's not much point in them talking to each other unless they at least ask, what is their term of forecast? Now, obviously, you don't have that problem with the AI system, because it's all at the same term, and it's just learning how to slice and dice that term. Or you can have multiple terms in there, and then you can layer on top intelligence, which might be heuristically designed or through AI or whatever, to then mathematically solve that complex term-based optimization problem of trade-off of moving between opportunity over different terms and trading costs and friction and all the rest of it. So, I'm not surprised that humans struggle to communicate with each other about complex things. Look how long it takes us to get to what we're thinking in questions and being specific and all the rest of it. This is why these problems are solvable.

Stefan Wagner: 19:21 That's why we had to build shortcuts.

Jacob Ayres: 19:23 Yeah, and it's why things are solvable. Solving these problems requires going to code and data and numbers and facts and statistics and science.

Stefan Wagner: 19:32 Now, I could talk on for hours now, but I think one last question a little bit sort of is, so is there something you would like to debunk about AI as an industry?

Jacob Ayres: 19:45 That it's a fad? that it would ever be a fad. Artificial intelligence, for me, if you're an alien and you come to this planet and you look at what's happened with biological intelligence on this planet, the thing that you would realize is, number one, is that the collecting of information through senses and sending it back through a central nervous system to a brain is… the most successful thing that life has devised for survival and evolution and progress, and the human being the central thing to that story. But it's all through All through the evolution of life on this planet, life has been driving to collect more information through senses and get it back into the brain and process it. Machine learning, data science is the same thing, but it's us putting it into these other human things that we do. So understanding ourselves, understanding society, understanding the economy, understanding the markets.

So the idea that getting information, processing it, learning on it, and building models, information, which is essentially a compression of data into a bite-sized form that you can glean useful information, it would never be a fad. Like, nature just look around us, tells us it's not. You know, this is just us beginning to embed that process into our machines to use them for us. The next thing is, if we think about artificial intelligence, for it not to become artificial general intelligence, there has to be some supernatural component to human intelligence, which is veering you towards a religious-like belief.

So, if you're with Einstein and you believe that God doesn't play dice, then everything around us is discoverable, workable, solvable, re-creatable. And in that case, yes, intelligence. We can recreate human intelligence. So that's really the question. Is there something supernatural about our intelligence? But I think also with AI, we won't develop it to be like a human. We've got 8 billion humans. Why do we need more humans? We're more likely to develop it to be useful technologies and tools for us. So at the moment, all AI is simply like tools, a bit like a car. We never chopped our legs off when we got a car. We still get in them and we still drive. But when I walk up the stairs or walk into this room, I use my legs because it's better. The same so far is true of AI.

I do think there's a possibility of someone running an experiment where they give it the lifelong learnings, embed it with all of our best algorithms, sights and senses, and let it walk and brown and talk. That has, I believe, a potential to develop artificial general intelligence. But nothing that's happening at the moment has the potential to do it, because nobody's allowing it to live in this planet. Because to form a world map around you of everything so that you can grasp all of this, you need to see that. LLM models like chat GPT has only seen text-based data. So it can never understand the world around it. It just understands text-based data. It doesn't really understand what it's doing. It's using statistical probability Bayesian and all the rest of it to be able to tell you back and do very intelligent tasks for us. But it's a tool.

Hassan Salamony: 22:48 It's an incredible tool. Exactly. And I think just to add to that, I think the other thing to debunk is this idea of a black box and not understanding AI, and that being a blocker to its usage or reliance, particularly in the world of finance. I think finance is coming around to that now, but it's been a real blocker up till this point, because if you can't understand the outputs, the results, then I can't rely on them.

But I think we're reaching that threshold now whereby people realize that there's too much data, there's too many things going on out there. You need tools like this to make sense of it all, and that's when you bring it home. It will be something that people use, rely on as an additional lens to everything that they're doing, become more and more commonplace in everyday life as we're starting to see. So I think it's very much here to stay, and it will be something that people rely on increasingly going forward.

Stefan Wagner: 23:39 Thank you. Thank you very much. Excellent. Thank you. Awesome.

Jacob Ayres: 23:42 Thanks, Stefan