October 3, 2022 – On today’s show we welcome special guest, Delphia’s Andrew Peek. Delphia runs an algorithmic investing strategy powered by machine learning. 

On the show, Andrew discusses:

  • Predicting fundamental data using AI to drive investment alpha
  • Which fundamentals their algorithm trades on
  • How humans can incorporate machine learning into their investment process
  • What he finds most compelling about capital markets
  • And more

Welcome investors to The Absolute Return Podcast. Your source for stock market analysis, global macro musings and hedge fund investment strategies, your hosts, Julian Klymochko, and Michael Kesslering aim to bring you the knowledge and analysis you need to become a more intelligent and wealthier investor. This episode is brought to you by Accelerate Financial Technologies. Accelerate because performance matters. Find out more at accelerateshares.com.

Julian Klymochko: All right. Kicking off the podcast with Delphia Andrew Peak. Andrew calling in from New York today. How are things? I hear you have a new office out there.

Andrew Peek: Yeah, we do. We share it with a group called Fiverr. I think they’re a public company, if I’m not mistaken. It’s a fun group. I don’t know if you know the company, but pretty much anything you could pay $5 for, they facilitate.

Julian Klymochko: Yeah. Fun fact, I get the podcast process through Fiverr [laugh], so well aware of the company. I use it on a weekly basis and certainly recommended if you have any requirements or if you run a podcast. But I’d digress. Let’s kick things off. Getting into your background in startups, I read just on your profile, you previously built and sold a company to Shopify. Can you walk us through that experience?

Andrew Peek: Yeah, for sure. We made a pretty simple bet out of the gate. This was 2010, 2009, 2010. And if you recall back then most of the software developed project cost was spent on the development side, not the design side. And yeah, we were seeing a proliferation of apps, and so we made a simple bet that said the balance of the equation would change and that 50 cents on every development dollar would end up being spent on design. So, we built a software design company that bet ended up playing out nicely. We were sort of the number two software design company in the country, spun out a piece of collaboration software. It was a bit of a poor man Slack before Slack was around. And both those companies sold to Shopify in 2013 when there was a race to arms for design talent.

Julian Klymochko: And then after that you made a leap into big data, which is, you know, you see just a ton of applications utilizing big data for machine learning, artificial intelligence, and so on and so on. Can you talk to us about, you know, the appeal of big data and how it can be utilized to make a prediction and many different things, whether it’s sports or politics or even investing?

Andrew Peek: Yeah, for sure. The big benefit of selling to Shopify and sort of riding that roller coaster through IPO was you kind of get to sit back and say, okay, well I have enough energy left in the tank to do another, you know, big swing, and I’ve created enough of a cushion here to make it, you know, a moonshot to really go after something esoteric or hard to pull off. And so, I went on a three-year hunt looking for very interesting projects out in the wild. And I came across a research lab that had spun out of the University of Toronto, I believe in 2010 or 2011. And I met this lab in 2014. They had built a survey-based tool that helped calculate your alignment of the different candidates that were running in an election. And this is a 50-question tool. But people loved it. People would spend 10 minutes on this thing, and they’d answer very reliably because they ultimately wanted an accurate output. They really wanted this insight into their alignment to the different issues and candidates and all of that. And then they would tweet or share their result on social media. So now you had this really rich profile that you could connect to this time series of behavior. And this lab, it turns out, became exceptional at forecasting the election outcome, the entire distribution. And they were outperforming every major polling agency in every country they operated. And I took a fascination to this, and the labs founder, Cliff Van Der Linden, and he had a really simple way of expressing it. He was trying to help people use their own data to their own benefit and thought this was a really great application of that.

And what this lab taught me was that what I had previously assumed was, you couldn’t catch up in the data race. The, you know, big tech platforms had run away with it. They had too much surface area, they were collecting too much behavioral data. But what his sort of little Petri dish proved to me was when you have this self-declared profile data that can run many attributes deep on the individual, and you can join that to behavioral data, there is a different level of forecasting that can be done. And so, I sort of loved that insight. I didn’t have much of an affinity for the election space, didn’t think it was a very good business model, but I sort of fell in love with the founder and the lab itself and that insight. And I immediately went looking for a context of bigger context where it could be applied

Julian Klymochko: And what bigger context compared to politics than financial markets. I mean, we’re talking about trillions and trillions of value traded, and there’s this notion that AI and machine learning can more accurately forecast or predict, you know, future capital markets movements, whether it’s stock prices, commodities, interest rates, FX, currencies, et cetera. Obviously, humans have been trying to do that for basically hundreds of years, if not longer. What makes you believe that AI and machine learning and these big data sets, how can that more accurately predict these future events, future prices in capital markets compared to humans?

Andrew Peek: Yeah, that’s a good question. I don’t know that you could just say it’s one or the other. So, in fact, there’s a graveyard of tech companies that have assumed you can throw you know, tens of thousands of data scientists around the world at a problem and somehow beat the market.

Julian Klymochko: Right.

Andrew Peek: Right. I don’t need to name names of them, I’m sure we’re all familiar with sort of a couple of these efforts. But you’re not excused from having let’s just say a framework, an investment framework in this case that hangs together and makes sense and is unique. I think that is still very much part of the equation, but there is merit to say that this toolkit that you’re alluding to machine learning and ai, it is extremely suitable for at least part of the problem.

Julian Klymochko: Hmm.

Andrew Peek: It is going to surface non-linear relationships far better than you and I will be able to and can ingest copious amounts of data. The key is, where and when is it responsible to go looking for these non-linear relationships? What’s sort of the foundation that lives beneath that that makes that something you can hang your hat on? And just a little bit of an aside, but the history of quantitative investing, which is to say the history of using machine learning to predict stocks generally was just about pointing a model at predicting price.

Julian Klymochko: Mm-Hmm.

Andrew Peek: Because you had an observation of a stock’s price every day, so you had all this training data that is sort of a prerequisite for training a model. But when you put it into the wild as it happens, many things move a stock’s price that the machine will never have seen before.

Julian Klymochko: Right.

Andrew Peek: Right. And so, it’s just not so simple as like, I’ve got this, you know, this grenade launcher, I’m going to point it at this target and hope for the best. You have to be a little more astute than that. Now, my partner Jonathan, our CIO, he was the managing director of the Quan Equity Group at CPPIB Canada Pension Plan. And his big insight was if you took that ML toolkit and you pointed it at fundamentals, you were in far better shape because the degrees of freedom around predicting a fundamental, we’re about an order of magnitude less, which is to say a machine has an order of magnitude less chances of overfitting itself. And so that just kind of goes to show that even one basic insight that can get you huge gains. And now having said all of that and even if you talked to Jonathan or anybody else who’s gone deep on ai, eventually it’s not about the toolkits. It is eventually about what you feed into it. And so even with a responsible framework, you know, the bet for us is really on data,

Julian Klymochko: Right. But your point being that ultimately AI and machine learning are a toolkit that an investor can utilize within their process. It doesn’t do it for them, but it can help augment their process and ultimately hopefully improve their results. Now you took this technology, you’re building a company around it. How does Delphia work and ultimately what’s a mission that you’re working towards?

Andrew Peek: Yeah, so we borrow the mission, if you will, it’s inspired by Cliff in the original research lab, which is to say you know, we believe that data’s going to unlock financial prosperity. And so, we’re trying to inspire people to invest their data with us. So, you can think of us like an asset manager that manages both data and dollars, and we use the former to improve the ladder, the returns on the ladder. Now, yeah, go ahead. Sorry.

Julian Klymochko: No, you go ahead.

Andrew Peek: Yeah, so the business model we’ve wrapped around this is an interesting one. As I mentioned, Jonathan developed this sort of breakthrough in forecasting, and so you could kind of call this a modeling advantage. We think we have a modeling advantage today. Now, any modeling advantage in the history of, you know, active trading generally gets decayed away. Eventually the field catches up to what you’re doing. People start modeling the world in the same way. And so, we’re very explicit with, you know, we think we’ve got a 5, 6, 7-year horizon or decay rate depending on how you want to talk about it.

Julian Klymochko: Mm-hmm.

Andrew Peek: And in that time, by the time we get closer to the end of that, it’s going to put the focus on our relative forecasting advantage. And so therefore we want to ensure that we’ve got access to proprietary data that others don’t. So, then the question becomes how do we construct a business model that will bring about proprietary evergreen data? We want this to be continuous, and we made a really interesting decision. It’s controversial for sure. The alpha that comes off of our stock selection. So, one machine learning stack, one set of alphas, but we partition the alpha, We do two different portfolio constructions. One for your accredited folks. This is a long, short implementation. Master feeder came in, domiciled hedge fund, you know, two in twenty, all of that. The other is we take some of that alpha on sort of 500 names and we put it on top of the S&P 500, and we give it away to retail investors for free. And we say, contribute your data through our mobile applications to improve the strategy. So, they’re getting alpha on top of batá, they’re getting it for free. It’s a little bit of a premium on the S&P, and we’re asking to contribute data to improve the stock selection. And as it happens, that stock selection improves everybody’s alpha. That’s the theory. And the fees paid on one side are redistributed to those contributing data on the other.

Julian Klymochko: Okay. I just wanted to get more into details on how the investment algorithm works and how the investment strategies are structured. So, you have two strategies. One long short, I assume market neutral is, this North American equities only, and then you have the S&P, plus long short equity overlay. Like what are the exposures? Are you trying to eliminate the beta? Is it like dollar? Market neutral? How is the strategy set up and what sort of returns are you targeting?

Andrew Peek: Yeah, so let’s start with the long, short. Fortunately, I’ve got some very recent performance characteristics in front of me so I can reference them, but yeah, it’s intended to be dollar and beta neutral.

Julian Klymochko: Okay.

Andrew Peek: So, I think on an annualized basis, we’ve got about eight basis points of beta in there. We hedge out our exposure to well-known factors like value, quality, momentum. If I recall correctly, those exposures are somewhere around 12 basis points in that territory. And, you know, we launched this April 1st of 21, so that’s 17 months of track record to date. And looks like we are sitting at about just north of 80% gross returns in that time and 65% net if you take the average fee. So that works out to about a 3.2 sharp. And that’s the long, short fund. Now the long only implementation is a very new implementation. I just mentioned, the one that I mentioned for retail. That’s actually just rolling out the door I think in real time as I speak. Previously we put the alpha on top of a sector bet, a consumer staples, consumer discretionary sector. [Inaudible 00:13:49-13:57] it’s tough to argue relative performance to retail investors. I’m not sure it doesn’t goes over quite as well. And so that’s why we decided to move the benchmark to the S&P 500. And that strategy is now kind of rolling up the door, as I said.

Julian Klymochko: So, I wanted to understand the investment process better. You just discussed kind of two different ways you can apply machine learning to stock selection. I mean, first you can just utilize historical prices as your input and try to predict off that. Or you’ll also mention utilizing historical fundamentals, predicting fundamentals, and using those to make asset allocation decisions, pick stocks, et cetera. So ultimately, how does your algorithm work? Is it a mix of both or are you more focus on predicting the fundamentals, which then drive stock selection?

Andrew Peek: Yeah, we’re predicting fundamentals. In fact, specifically we’re predicting fundamental surprise.

Julian Klymochko: Okay.

Andrew Peek: And so, you can roll fundamentals up to sort of into three buckets. You’ve got your sales surprise, your EBIT margin surprise, and your net income surprise. And each of those top line fundamentals can be unpacked into a set of KPIs. And generally, a discretionary investor is in the business of trying to predict what the KPIs will be, and that informs their expectation of the fundamentals and therefore they’re discounted cash flows. And so, they have their own objective function with which they’re looking at the market. And then you’ve got quants and quants really aren’t in the business of prognosticating, they’re more in the business of reacting to information. So, it’s more about, okay, well we’ve got ground truth that’s been established using earnings report. How does a quant expect the field of fundamental investors to react in light of that new information?

And can they update their positions prior to any manual adjustments to models thereafter? And so, we’re trying to arbitrage that dynamic. We’re trying to use a quantitative toolkit to do actual prognostication in advance of ground truth on these fundamentals. And so, we’re sort of taking positions and harvesting as they converge towards the reveal of the surprise if we’re correct. And so just to kind of illustrate how different this looks from a quant perspective, you know, post quant winter, most hedge funds like two Sigma alike, their average hold time for a position just really collapsed down to days, I think it’s about seven days at two Sigma. at Delphia conversely, we hold a position for four months on average.

Julian Klymochko: Okay.

Andrew Peek: Which is an obvious demonstration that we’re trying to leapfrog over the short horizon where markets behave like a voting machine.

Julian Klymochko: Yep.

Andrew Peek: And we’re trying to place bets in sort of this mid horizon where they behave more like a weighing machine. So, we’re using ML to predict these fundamentals and then a much simpler technique, we’re using linear regression to map that back to returns. And we’re doing so at different horizons, seven different horizons, which is seven quarters into the future.

Julian Klymochko: In terms of predicting fundamental surprise. Just so I have this accurate, the algorithm makes its fundamental prediction and then your fund puts on a position either long or short, if that fundamental prediction utilizing ML is say quite a bit different than the consensus. Would that be a correct interpretation?

Andrew Peek: That’s it.

Julian Klymochko: Yeah.

Andrew Peek: And so, where machine learning becomes helpful is its job is to go look for non-linear relationships in the data. But rather than hang our hat on any one non-linear relationship, we’re looking for sort of a consensus to emerge or agreement to emerge amongst many non-linear relationships using the model.

Julian Klymochko: Right.

Andrew Peek: So, the model’s looking for these non-linear relationships, but the target is fundamentals, which are a variable far more stable than price, but still co integrated with price.

Julian Klymochko: Yeah. With the thought that over time fundamentals will drive the share price over the holding period. Now, are there any market environments in which the algorithm works better than expected? And conversely, are there adverse market environments where it’s like, you know, these predictions just aren’t working, or the predictions are working, but the stock price reactions aren’t what we think?

Andrew Peek: Yeah, there definitely are. So, there are moments in history where fundamentals matter more to the equation and moments in time where fundamentals just get tossed out the window. And in those particular moments, in the latter moments, you really have to believe that sanity prevails in the long run. Now, again, I’m not talking about the now casting environment. In the now casting, it’s extremely noisy. Prices are moving around on account of all kinds of information that have nothing to do with fundamentals. I’m really talking about, you know, beyond that now casting moment, can we rely on fundamentals? And you might be surprised by the moment where fundamentals kind of went out the window, you know, the model weathered the great financial collapse pretty well all things considered. And even March 2020, we were only down a point in the back test at least. The greatest dip in the back test occurred when the vaccine news came out when we finally had a vaccine. And it was in that moment that all of a sudden fundamentals got deprioritized writ large across the market.

Julian Klymochko: Yeah. That was really tough for quant investors. We run a few long short quant strategies and I remember that day specifically because it was one of the most painful

Andrew Peek: [Laugh], there you go. Fortunately, we weren’t live yet, but nevertheless, that was a big hit. And you know, we’ve seen days like this too since we’ve been live. You just have to rely on the fact that, again, sanity prevails, fundamentals do matter. Economic outcomes do matter. You could think of it like the strategy makes money by stack ranking companies correctly in a peer set, and that whole peer set can move up and down with whatever macro events or market conditions are going on. Really what we want to do is we want to get the order correct and the delta between them correct. And there’s just more payout for the top of the order than there is for the bottom. So, if you get number seven and eight backwards, not the end of the world, you get number one and two backwards, you’re probably not making money.

Michael Kesslering: So, a couple of things here because this is all very interesting. I guess first is there any capacity constraints that you have for the long, short strategy specifically? Any sort of size that it would no longer work or there would be too much decay? And then as well on for sector specific, do you look into sector specific KPIs? So, say the KPIs would be very different. We’re based in Calgary for an energy company, an EMP energy company based on production versus a consumer company that’s in maybe consumer packaged goods. How do those KPIs and your model differ for those?

Andrew Peek: Great question. So, the first one, very simply, yes, there are capacity constraints and long short space for us. It is higher capacity than most. So, if we are just working with a universe of U.S. equities. Our expectation is around 1 to 1.2 billion. However, we do know this works in global equity space, Europe, Japan, Australia, Asia X. So, by adding equities to the pool, we think we can max it out around 2.5. To your second question, this is a really interesting one. So, we do have a nested modeled approach where we’ve got, you know, company specific models inside of industry level models inside of a global model. But it’s a really, what you’re pointing to is actually a really hard problem. So especially when you think about point in time. So KPIs, let’s just use Netflix for example, right?

Netflix sales is going to have a longer reporting period than any given Netflix KPI that rolls up to it. So, you can think of KPIs as having a half-life that’s different than a fundamental in many ways. So, what that means is that at any given moment in time when you’re trying to do an assessment of what these KPIs were and a compression exercise to feed that signal into your sales estimate, right? It’s not the same approach kind of one year to the next, or five years to the next. And that exercise is something we’re actually spending a lot of time on right now doing by hand. Because we believe there’s a whole other tier we can get to if we can marry sort of depth of understanding on a per name basis with a cross section of 3000 names, right? So, we benefit a lot from the cross section right now. And if we can behave even more like a fundamental investor, that’s another way to say it then we can optimize our bets accordingly.

Julian Klymochko: And now how much does algorithm utilize the platform’s user data? Because obviously you mentioned someone like, you know, a company like Netflix obviously highly reliant on consumer behavior, but there’s various B2B business models or commodity producers or things of that nature where perhaps consumer attitudes or use would have a much smaller effect. So, does the platform user data augment other fundamental data that you utilize? Or is it, you know, some sort of balance or split between them?

Andrew Peek: So right now, it doesn’t augment anything. The panel’s too small. We only started collecting panel in earnest, you know, we’re less than a year in that front. So really the investment in sort of the proprietary data is an investment in our future.

Julian Klymochko: Oh, okay.

Andrew Peek: However, yeah, we don’t derive alpha from it today, however, we collect it in a manner from which we expect to do model training and research.

Julian Klymochko: Okay. So that will be something that perhaps improves the algorithm at a later date. So that makes a lot of sense. Now, from a personal perspective, the capital markets can be fun and frustrating and, you know, generate a whole host of emotions. Now for you, what do you find most compelling about the financial markets? And conversely, what do you find most frustrating?

Andrew Peek: Feedback loop’s incredibly addictive. So, it’s just a fun space. Because everyone’s walking around with a theory, a hypothesis, you know, high conviction in a lot of cases. And yet the feedback we cannot be argued with, it’s right there in front of you. And so, it’s humbling, it’s fun and you really have to be prepared to update your priors regularly. You can’t have a lot of hubris, is what I’ve quickly realized. And not that I feel like I did coming in, but I’m enjoying the rate of change, let’s just say. But I think for me, there’s still sort of one level higher that I appreciate, which is, you know, predicting or modeling the world is just such an addictive activity.

Julian Klymochko: Mm-Hmm.

Andrew Peek: Right? Trying to guess the future and being in a relatively better position than the next person is just so, I don’t know, as a game it just gives me a lot of pleasure. And so, I’m very much enjoying the predictive side of the equation and the feedback from supporting that

Michael Kesslering: Then moving away from investing in capital markets a little bit. Now you’re a repeat founder, two different industries that you’ve been involved in. Do you have any advice for future founders or entrepreneurs in your path?

Andrew Peek: I think generally going eyes wide open into whatever the nature of business you’re trying to start. So, you know, Delphia is not a small build, it’s a large build. It was quite obvious to me this was going to be a venture back business and, you know, we’d go many rounds deep. And so that ends up looking like a very binary bet at the end of the day, right? It could be incredibly successful. It may not pay itself back. And just making peace with your probable outcomes in advance of going down this path, I think is really important. If you’re looking for something that’s got a little bit more of a linear growth curve to it or you’re looking for something where, you know, it’s more lifestyle oriented and you don’t necessarily have to answer to VCs. I think it’s really good to get the criteria on the table on day one. Because in my experience, there’s so many ideas, there’s so many different business ideas you can pursue. That’s not the shortage. Where we lack skill as entrepreneurs is the matching exercise. Which of the ideas in my midst am I going to be intrigued by, want to work on, satisfy the conditions I’m doing this for? I think people could stand to spend a lot more time in that sort of self-awareness phase and reflecting on what’s a good match for them.

Julian Klymochko: Andrew, you approach it very mathematically and unemotionally, which is fitting for a quant investor, [laugh]. Because you never want to get emotional and override the models when it temporarily stops working. In any event prior to letting, you go here. One last fun question that I like hearing from CEOs and founders is what time do you wake up and what’s your favorite productivity hack?

Andrew Peek: I woke up at 5:00 AM today, which is pretty typical. 5-5:30. An hour is not made equal. Again, this is a self-reflection exercise. So, if you ask me to write up a blog post, for example. It would take me four hours if I started at 9:00 PM. It’ll take me 45 minutes if I start at 6:00 AM. And so, you really can choose how to slot your work into your day in an extremely efficient way once you understand when you’re prone to being in a creative space, when you’re able to do sort of multitasking around monotonous things, you know, when you’re best suited to take in information as in from a meeting. And I’m religious about how I structure my day so that, you know, all hours are not made equal.

Julian Klymochko: That’s really good point, yeah. I find myself personally most productive and effective between five and eight in the morning, but after say eight at night, just complete disaster [laugh] in terms of productivity. So that makes a lot of sense. Thanks for coming on the show today. A lot of interesting insights into Delphia, the platform, how it works. Super cool that you are rewarding users for providing data, which ultimately you hope leads to improvement in the algorithm at a future date. So, making those investments for future alpha and wish you the best of luck.

Andrew Peek: Thank you very much. Appreciate you gentlemen, having me on.

Julian Klymochko: All right, take care. Bye everybody.

Thanks for tuning in to the Absolute Return Podcast. This episode was brought to you by Accelerate Financial Technologies. Accelerate, because performance matters. Find out more at www.AccelerateShares.com. The views expressed in this podcast to the personal views of the participants and do not reflect the views of Accelerate. No aspect of this podcast constitutes investment legal or tax advice. Opinions expressed in this podcast should not be viewed as a recommendation or solicitation of an offer to buy or sell any securities or investment strategies. The information and opinions in this podcast are based on current market conditions and may fluctuate and change in the future. No representation or warranty expressed or implied is made on behalf of Accelerate as to the accuracy or completeness of the information contained in this podcast. Accelerate does not accept any liability for any direct indirect or consequential loss or damage suffered by any person as a result relying on all or any part of this podcast and any liability is expressly disclaimed.


Want to learn about the investment strategies and techniques used by hedge fund managers to beat the market? Download Reminiscences of a Hedge Fund Operator by investor, Julian Klymochko
Terms and Conditions apply
Download Free Ebook