Todd Hawthorne of Frontier Alpha Source Technologies joins IBKR’s Senior Trading Education Specialist Jeff Praissman to discuss the past, present and future impact Artificial Intelligence is having on the Investment Community.
Note: Any performance figures mentioned in this podcast are as of the date of recording (September 14, 2022).
Summary – IBKR Podcasts Ep. 39
The following is a summary of a live audio recording and may contain errors in spelling or grammar. Although IBKR has edited for clarity no material changes have been made.
Hi everyone, and welcome to IBKR podcast. I’m your host Jeff Praissman, Interactive Brokers, senior trading education specialist. It’s my pleasure to welcome Todd Hawthorne and discuss how artificial intelligence is changing the role of equity analysts and the portfolio manager.
Todd’s been training in managing equity and equity derivatives for over 20 years. He began his career at Morgan Stanley and most recently is in the process of launching Frontier Alpha Source Technologies, also known as FAST, which seeks to provide volatility as a source of yield for investors and uses AI and machine learning to analyze and project volatility manage risk.
Welcome, Todd thank you for joining us.
Thanks for having me, Jeff. I appreciate it.
I want to start this off, how did you get interested in AI? Obviously, you started off with Morgan Stanley as an equity and equity derivatives trader and it seems like you’ve evolved over time and really curious how your interest in AI got piqued and what you think is potential is any analyst and portfolio management area?
AI came to me some from sort of a roundabout way. In in 2008 I developed a volatility strategy which essentially turns equity into a type of fixed income and the strategies built around bottom-up fundamental analysis, that traditional equity analysis. But in fact, it uses a different asset class, uses volatility as an asset class. And as such, volatility has different return and risk characteristics than traditional stocks, and as such the ideal investment has different characteristics than traditional stocks do. So, this requires looking at that same fundamental analysis in a different way or through a different colored lens, and it turned out to be pretty difficult for a lot of stock analysts to do. It required change, and they don’t typically like that.
So, I started thinking how can I build my own analysts? What characteristics would that analysts have? What would they care about, and how would they change the way that they look at stock fundamentals to specifically address volatility as an asset class? And I was lamenting with this colleague of mine, Scott Grossman, who came out of AllianceBernstein and later Aeira Capital, which also uses AI as an investing platform, and he was successfully using AI in his long-short hedge fund as the primary source of equity analysis.
And in talking to him, we had this idea that AI could be effectively used to project volatility as opposed to stock returns. So in other words, you could ask the machine to give you a different answer, you’re not asking a machine to say what stocks are going to go up or down, but you’re asking a machine to give you its estimate of how volatile that stock will be going forward. And this led me down the path to trying to use AI in my own investment process, and it turned out that was a pretty effective question to ask AI.
AI is a term that gets thrown around a lot, I think artificial intelligence, and I think most people are familiar with the term. I’d be hesitant to say that most people are really familiar with how it actually works. In layman’s terms, could you kind of give our listeners an overview of really the difference between a computer or computing program using artificial intelligence and just say, an algorithm or a computer just crunching numbers and spitting something out?
Sure, quant portfolios, the idea of a quant portfolio, has been around since late 30s I believe, and the original quant portfolio was a way of taking a look at stock fundamentals, balance sheet, income statement, cash flow statement and saying that specific characteristics were going to come through a lot of different stocks and stocks with specific characteristics should exhibit specific return profiles. So, stocks with good return on equity, for example, will tend to do better than stocks with bad return on equity.
So that’s a fixed system, and what it presupposes is that whatever factors that you’re looking for in this case, return on equity, will be consistent no matter what the current market environment is like. So that is really what quant system suppose, they take all these factors and then they back test all these factors, all these static factors and they get a group of factors and then they look for those factors and invest and create a portfolio surrounding those factors.
Along comes AI or machine learning and what machine learning or what the difference between quant, typical quant and machine learning is that the machine does indeed learn, and what it does is it goes and looks at all of these factors and currently uses what we call unstructured learning. And what unstructured learning is, is you don’t tell the machine what to look for, it learns the way a child, for example, learns. You don’t tell a child what the term mother means, they actually just learn that term on their own through experience, it’s called heuristic learning. And the machine will essentially take all of these factors and create a defective living correlation matrix where it tests again and again and again all of these factors until it finds the factors that currently matter to the outcome that you’re looking for, whether that be stock price improvement. or volatility. And so, you can look at it as a quant model that is constantly evolving and constantly being tested and constantly being compared to the existing state of the market that exists. So in a lot of ways, I think of it is quant 2.0.
With artificial intelligence, it’s almost like if you took an analyst and he was able to live for, say 500 years and able to constantly go through these problems and use his experience. It’s something that ideally a machine with, you know, increased computing power and artificial intelligence is able to do in the matter of days or hours or whatever the time frame is, but clearly at a much faster pace at being able to sort of learn through a process of just seeing every outcome that’s possible. Is that that kind of a good way to sum it up?
Yeah, repetition is the most basic answer. The machine learning game takes the idea of Malcolm Gladwell’s principle of 10,000 hours, right? So, if you take that principle, and that principle states that it takes roughly 10,000 hours to master any task. However, that 10,000 hours is hours of human time. For example, let’s say it takes 30 minutes to play a game of chess, well, 10,000 hours of chess would then be 20,000 games and if you played one game of chess per day, that would equate to something like 54 years before you mastered chess. A computer can play thousands and thousands of games in a single day. Also, assuming that the computer can quote-unquote learn at the same rate as a human through this heuristic process, this hands-on learning process, and given the fact that a computer can play, again, thousands of games in a day, it makes sense that the computer can reach mastery in a relatively short period of time.
Also, computers remember every game they’ve ever played, so unlike a human, which is likely to make the same mistake at least a couple of times before you learn. A computer doesn’t have that failing and can theoretically, learn faster or progress through that faster. So, you’re playing thousands of games of chess in a day and you’re learning faster by default. You have to reach mastery more quickly than a human under that timescale. In fact, this idea of unstructured learning or teaching computers to learn the way humans learn, really reached a pinnacle in about 2016. The game of Go is a Chinese game that is multiple times more complex than chess, and they always thought that there was no way that a computer would ever beat the Go grandmasters.
So in 2016, DeepMind, which was a Google venture, had this idea that rather than structured learning or giving the computer all of these examples of fantastic grandmaster games to learn from. They would give the computer absolutely no directions, only the rules of the game, and in an unstructured fashion it would learn by playing thousands and thousands and thousands of games against itself. And in a matter of about two weeks, that particular algorithm beat the prior algorithm, which was based on again, grandmaster play and all these rules, 100 times in a row and then went on to beat one of the greatest grandmasters, greatest human grandmasters in Go. Which was a feat that again, just a few years before that was thought to never have been possible.
So, this whole idea of unstructured learning has really taken the idea of AI into the mainstream, and not only that has brought it closer to the way in which humans process information.
And it would seem to from that example, and I assume it’s really not getting prejudiced by prior human –in this case, the Go example, starting fresh with the fresh tab there’s no influence or prejudice from previous grandmasters or their styles. The computer is free, for lack of a better word, to think through the entire process, whether this Go game or whether it’s looking at investment information or volatility or some sort of financial aspect and really taking it and really utilizing all its computer knowledge to process it and come up with a profitable solution.
100%. It’s really interesting, there was a Nobel Prize laureate, and his name was Daniel Kahneman, and he won the Nobel Prize for behavioral economics and one of the really interesting things that he proves in the course of his career was that humans have a way, and specifically stock analysts. He was looking at the stock market when he wrote this. Humans have a row of overestimating their excellence in predicting chaotic systems. So, if you are an analyst and you’ve been an analyst for 20 years, you have mathematically, statistically, an inflated opinion of your ability to predict stock prices going forward.
So, you’ve put in your 10,000 hours and you value that 10,000 hours in a way in which you think can determine the outcomes, but statistically you really can’t. You’re overstating your excellence. Add to that he also did some really interesting things with consistent information versus inconsistent decisions. So what does that mean? The analyst you are today is not the analyst you are tomorrow, so given the same set of information, you’re actually likely to make a different decision based on that same information today, as you will be tomorrow.
So, all of that is just a way to point out that humans’ guts or intuition are not all that accurate, and the consistency of the decisions that we make is not all that consistent. AI and machine learning has that same base of 10,000 hours, but it doesn’t have those two failings, which is really, really important when you’re a portfolio manager. It still makes errors, and even if it’s only as accurate as your human analyst, but you know that it’s going to be inaccurate in a very specific way and that inaccuracy is going to be consistent. You can use that in portfolio creation to great effect to manage risk and that is something that you cannot do with its human counterparts.
You know, as far as AI’s influence currently on the industry and then its potential down the road for investing in portfolio managing, fair to say that it will still always require a human being to make a decision based on what the computer is telling us? Or do you foresee a point where there will not be a need for any sort of investment manager, and we’ll just let the computers run the world?
We’re ways from that and let’s say, it depends on your time, right? If we’re talking the next 100 years, my answer is going to be very different than if we’re talking the next five years.
In 1950, Alan Turing came up with a test of machine intelligence before there was any machine intelligence. And basically, what the idea was is you take two people and you put them in separate rooms, and they have a conversation over text, a text-based conversation, and if the human asking the questions in one room can tell that the person in the other room is not a person, but a computer, then the computers failed the Turing test. So again, 1950, this idea was put forward and the first time the Turing test was successfully passed was in 2014. It was a theoretical, I think 14-year-old Ukrainian boy who had a name, and he was out there having conversations as a chat bot, and no one could tell that it wasn’t a human. So again, 1950 to 2014. For it to take that same time frame and compare that to the language of financial investment, which it is just another language, it’s another way to communicate and think about and organize ideas. Maybe in 100 years we will have the ability to completely replicate the human thought process in terms of portfolio construction and risk management and what have you. And we’ll have a hive mind that performs all those tasks. Over the next five years that’s not going to happen.
So that kind of leads me to ask though, even in both scenarios really, whether or not it’s a human analyst or portfolio manager or investor utilizing an AI program or fast forward 100 years down the road and possibly it is just the machine doing it. Do you feel there’s a danger of like a herd mentality in really both situations? Everyone’s sort of either may be coming up with the same — the programs are coming up with the same data, whether or not they’re investing on their own in the distant future or in the not distant future where analysts and managers and traders are using these programs. Do you think we kind of get into a spot where everyone starts coming up with the same thing because the computers are all so much further ahead of what we could come up with and we’re all kind of bumping into each other again, sort of like a high frequency trading situation?
AI is a quant system first and foremost. It is a living quant system, so it’s constantly evolving but it is a quant system and even Graham and Dodd, back in 34 or whatever the year was, is a quant system. It’s a way of quantifying and looking at characteristics of stocks. The difference being that the factors which we’re looking at that create the output are not fixed within AI. They fluctuate as the machine is running this type of living correlation matrix machine quote-unquote decides which factors are important today, as opposed to those fixed factors or algorithms that we’ve used in the past.
The decision that the AI makes is based on a desired outcome as opposed to a static quality. So what does that mean? You’re telling the computer what the pro forma outcome that you desire is, as opposed to back testing a bunch of factors that have a correlation to the outcome that you’re looking for. That’s a slightly different but very important distinction in that number one: you can change what that desired outcome is. So, in quant systems we’ve had a lot of times in the market where we call it, trying to put the elephant through the keyhole where every single client fund is constantly running these back tests on all these factors, and one or two factors tend to bubble up to the surface and then everyone is overweight those factors which seem highly correlated to our performance. And then there is a tectonic shift of some sort. Those factors stop working, and then the quant funds have to sell their stocks, and it’s the elephant through the keyhole syndrome and those stocks go down disproportionately because everyone is trying to get out at the same time.
You don’t tend to have that with human analysts because their portfolios are constructed along slightly different lines and they have slightly different outcomes that they’re looking for, right. You have growth funds, you have value funds, you have momentum managers, they all categorize stocks slightly differently. And the great thing about AI is you can essentially build your analyst, your AI analysts to look for the specific things that you want out of your portfolio. So it’s more, — it’s less concentrated in the way that traditional quant funds are concentrated because you’re not just looking for a specific factor. You’re more creating an analyst, a synthetic analyst that can comb through the data, the mounds of data in the way that you want it to.
So, I equated to that children’s toy which was the Build-A-Bear workshop where you went in, and you could build your own teddy bear. And the teddy bear was built along the lines of whatever the child wants. It was customized to that child where AI is a synthetic analyst that’s customized to the way in which you, the portfolio manager, need or want it to process that information. So from that POV, there’s less risk of hurting.
The AI platforms are really able to sort of look at the portfolio entire situation, not just look at stock ABC and say it’s undervalued, but they’re also able to look for the overall outcome that the analyst is looking for there. It’s just going to get there faster and more than likely do a much more efficient job of finding the best stocks or best instruments for that portfolio, that’s going to get that overall goal that the analyst is looking.
Yeah, Jeff, you said something very interesting actually. You said ABC stock is undervalued. That is actually a pretty loaded statement. And it’s loaded for the following reason, if ABC is undervalued, it’s undervalued relative to something. Is it undervalued relative to itself? Is it undervalued relative to the market? It can’t be undervalued in an absolute sense because “under” means it has to be under something. In its most basic sense, AI and this whole revolution of unstructured learning, what it does is it pits stocks against each other and comes up with a relative winner. That’s a really important point and you can think of it like that Mad Max movie where Tina Turner says “two men enter, one man leaves.”
I think that was Beyond Thunderdome, right?
Beyond Thunderdome, that’s right. What’s interesting about this is it doesn’t tell you how much a stock is going to go up, only that it’s likely to outperform the other stocks in that group.So you’ll get a portfolio of relative outperformers, if that’s what you’re looking for. And unstructured learning, by its very nature, it has to be relative. In that DeepMind experiment, in that Go experiment, it could use any move that it wanted to. And so it had to choose that move over all the other possible moves, and it did that through thousands and thousands and thousands of repetitions but it was always relative to all the other moves that it could have possibly made.
Human analysts, on the other hand, they often focus on pro forma or future stock performance, and we call those in the industry price targets. In my opinion, and I think Daniel Kahnemen would agree with me, this is the least accurate thing that fundamental analysts do. Not only that a stock would go up more than another stock but then a stock will go up by X amount within a given time frame. I’ve never met an analyst that is repeatably and reliably good at this. It’s just too difficult and too chaotic a system to be efficient at.
AI, what it does is it creates deciles, and those deciles are based on again that desired outcome of the model. For example, if the goal of the model is future price appreciations we’ve been talking about, then quantile 1 should outperform quantile 2, and so on. And how this gets implemented is really up to the portfolio manager and the specific strategy that is employed. It’s not that AI is really creating a portfolio and doing the duties of the portfolio manager itself. What it’s doing is “quantiling” or organizing information so that the portfolio manager can then apply that information to their specific strategy, and it helps them to create the portfolio and most importantly, it frees up the portfolio manager to do what he or she does best. And theoretically that should be investing in the themes that they find interesting, whether those are growth themes, whether those are value themes, what have you, and following that that thread, that investment thread through their portfolio. But they don’t have to sift through the mountains and mountains of data, and they don’t have to quantile all these stocks. They have a picture of what the current investment universe looks like, and they can use that to then paint whatever picture in that portfolio that they’re looking for.
How does AI process [intangibles]? To me the numbers make sense like it can go through these scenarios. It’s obviously doing at a much higher level than a human can. It can look at historical patterns, future patterns as you just mentioned. It can certainly compare stocks against each other, instruments against each other to find the best overall outcome for what the analyst is looking at. Is it possible, as of right now, to process the human element of analysis, like the opinion of the leadership of the company, the products in the pipeline, stuff that’s not going to be necessarily reflecting any sort of balance sheet?
It may be reflected, especially products in the pipeline, may be reflected in sort of current stock price. Is it able to process, say like a CEO’s track record, a new CEO comes on board to company ABC? This CEO has a track record of turning around companies and say ABC is sort of overvalued, I guess you could say right now, or undervalued and lagging its competitors. CEO Mr. John Smith comes in, he has his great track record. Is artificial intelligence at this point in time able to sort of process that element or would it just sort of process the stock bumps up from this new person leader coming on board and it would just use those numerical values as part of its processing.
Yeah, that’s exactly the point. I think it does not process any of the intangibles at all. In my opinion, this is what humans do best. They create filaments of attachment that wind their way through the investing portfolio. I have a friend of mine, Jim Stovell, who runs Stovell AI Systems, and he’s had an AI building investment platform for over 10 years that they call Night vision. And Night Vision is really good at specific things. For example, identifying stocks that could underperform over a very specific shorter duration and for this purpose it works extremely well and what Jim often says is that the purpose of his Night Vision is that it frees up portfolio managers to do what they do best, which is develop those intangibles and develop broad themes that can be used to create the overall portfolio.
So for example, electric cars electric cars have implications in batteries in lithium and semiconductors and software and cloud storage and connectivity etc. These are really, really broad reaching themes that are not going to be represented in the balance sheet or the income statement of Tesla. So, there’s no way that the AI can infer any of those connections. It’s really up to the human portfolio manager to do that. So, we’re not there yet. I mean, maybe in 100 years, as I mentioned, we’ll have a Turing test for a portfolio manager in one room and an AI in another, and we won’t be able to tell the difference. But for now, the intangibles and the picture that is painted by the portfolio manager of the themes that they’re investing in is really the purview of humans and humans only. So it’s good I have a job for another couple of years and I’m happy about that.
I was going to say, yeah, it’s good to have some job security for a little while then.
What is the cost of entry to use AI? I mean is this something that really only the big players right now can afford? Or is it something that obviously, depending on the level of it, is accessible to a smaller shop or an individual… I want to say investor, but individual trader or money manager? Or is it something again, like just really the big players can use at this point in time?
We’re at the point in time where AI is becoming more and more relevant and also more and more affordable. If you were to look at the costs on a sort of ad hoc basis, the costs are very multi fold. The first cost is knowledge. You have to get a data scientist who not only understands data and statistics and programming, but who understands the end goal which is typically picking stocks that go up. And then there is the development in time of the algorithm that you’re using. You have to back test that algorithm and not least you have to have the compute power and the compute cycle costs. So that’s a lot of money and if you were to put all those pieces together, you’re looking at round numbers, probably $1,000,000 a year.
However, within the last five years or so, compute has become a lot cheaper, cloud compute has become reality, AI algorithms are now hosted in the cloud and are shareware, so you can get them for free, and this has made it possible for a lot of companies to begin using AI in an investment context. Companies like Boosted.ai for example, they have become a reality over the last three or four years and they in essence are a soup to nuts provider of AI compute back-end cloud storage. You can run a bunch of diverse algos on their site and you can back test them and then you can create a live portfolio based on those back tests. So, it really is a soup to nuts AI platform, where you can create an analyst that looks at the world through your specific lens and delivers that AI solution to you on a daily basis. And those costs are about 1/10th of what it would cost you to build it ad hoc.
So you know, were not quite at the level where your average investor runs their own AI analytical platform, but any moderately sized investment platform can play in AI and have an AI component in their strategy for less than the cost of just hiring another sector analyst. So to me, that’s pretty democratic in terms of what the costs actually are, and they’re only going to come down from there. The flip side of that is if you’re the director of research inside even a moderately sized fund, and you have all of these sector analysts that are reporting to you, your biggest desire and their biggest complaint is that there is too much to look at. The amount of data that is out there that is required – that you’re required to come through as an analyst before an investment decision is made is gigantic. We’re in the midst of a data explosion right now. Data is everywhere, and it’s growing exponentially.
Without some way to organize and categorize that data? It’s pretty much impossible to be an efficient and effective investment analyst, enter AI. So, a lot of ways in which I believe directors of research are thinking about this is I have all my sector analysts which are humans and they’re making decisions, but I have to have a data wrangler. What is that data wrangler? Someone who can process and comb through the mounds of data quickly, efficiently, who doesn’t sleep, who didn’t have three margaritas the day before, who always makes the same decision, who I don’t have to pay a bonus to and it’s relatively cheap. That’s pretty much the definition of AI and I think you over the next five years, you’re going to see AI have an analytic place at the table specifically for that purpose of just categorizing, organizing data, and to be able to present that to the analysts so that they can apply that human side of following that that investment thread through Tesla for example.
So really just instead of a replacement, it’s more of an addition to the team. It’s not replacing anyone on the team, it’s an addition to help everyone on the investment team do their job better. Essentially, give them as you said, less noise to look at it and make sure that they’re looking at the data that’s relevant for them to perform their jobs and make the best decision they can as far as going forward on the investments.
Absolutely. To use another sci-fi reference, The Six Million Dollar Man “we can rebuild him better, stronger, faster,” right? So, you have an analyst who is just a human, and they can only process so much information and there’s only so much time in the day. And you can essentially graft this bionic analyst onto your sector analysts, which allows them to just process more information and characterize more information more efficiently and allows them to be better at their job. That really is the current application for AI that is most applicable in my estimation.
Well, Todd, this has been great, thank you so much for joining us. This has been a pleasure and it’s been very informative.
I want to remind everyone that you can find all our podcasts on our website, under Education, scroll down to IBKR podcasts or on Spotify, Apple Music, Amazon Music, Podbean, etc.
Look Todd, we’re looking forward to your latest venture Frontier Alpha Source Technologies and we would love to have you back as well to continue this conversation or any other items you’d like to discuss.
Great Jeff, I really appreciate it. Thanks for having me it was a lot of fun. Look forward to speaking to you again.
Todd has been trading and managing equity and equity derivatives for over 20 years. He began his career in equity derivatives at Morgan Stanley before moving over to the buy side where he pioneered the use of volatility as an asset class in liquid alternatives and buywrite portfolios. Todd designed and managed the Redwood Alternative Yield Strategy as a managing director at Boston Partners. The Redwood strategy was unique in that it uses equity volatility and deep in the money buywrites as a source of yield and as a fixed income replacement in this time of suppressed yields. Most recently Todd is in the process of launching F.A.S.T. Frontier Alpha Source Technologies which seeks to provide volatility as a source of yield for investors and uses A.I. and machine learning to analyze and project proforma forward volatility and manage risk.
Disclosure: Interactive Brokers
The analysis in this material is provided for information only and is not and should not be construed as an offer to sell or the solicitation of an offer to buy any security. To the extent that this material discusses general market activity, industry or sector trends or other broad-based economic or political conditions, it should not be construed as research or investment advice. To the extent that it includes references to specific securities, commodities, currencies, or other instruments, those references do not constitute a recommendation by IBKR to buy, sell or hold such investments. This material does not and is not intended to take into account the particular financial conditions, investment objectives or requirements of individual customers. Before acting on this material, you should consider whether it is suitable for your particular circumstances and, as necessary, seek professional advice.
The views and opinions expressed herein are those of the author and do not necessarily reflect the views of Interactive Brokers LLC, its affiliates, or its employees.
Disclosure: Hedge Funds
Hedge Funds are highly speculative, and investors may lose their entire investment.