Two Researchers Explain How Quants Are Going To Revolutionize Long-Term Investing

constantly evolving and adapting to the market. That point about it adapting, and of course that's the name of his book, Adaptive Market is the name of his hypothesis. Is really key because one of the main points that he made that I loved was this idea of hedge funds as sort of the R and D laboratory for all of the financial industry. Right, the funds are where innovative new techniques get to be sort of hashed out without doing usually too much damage, I guess, to the wider ecosyst

hopefully right, not always. There's certainly examples of hedge funds actually having done major damage from time to time. But ideally, you know what the evolution seems to be that some new idea sort of starts in the hedge fund world

and eventually makes its way to the broader world. I think the most obvious example of that that we could say these days is a lot of this sort of popular smart beta strategies e t f s that are built on things like momentum or value or other factors, sort of quantitative ideas that for many years were only

available to uh, you know, researchers at hedge funds. Right. So, these sorts of quantitative investment or trading methods were usually the purview of sophisticated hedge funds who had the time and resources to develop them. And then you had a bunch of e t f s who kind of caught

on and managed to replicate them. And now we can all trade lighthedge funds for zero percent fees, right, exactly right, And of course, once everyone can do it for very few fees, I think it's safe to say those strategies aren't going to produce the same returns, and hence the market is forced to adapt again. Right. Presumably the hedge funds are always trying to stay one step ahead as well, right, exactly so, which raises the idea of like what will

be the next thing. If anyone can sort of invest in a crude momentum strategy for virtually no fees, then that requires the people on the cutting edge, the people doing the R and D of this industry, to uh, you know, figure out what the next big thing it's gonna be. Do you know what the next big thing is going to be? Joe? Can you share it with

your fellow partner at Odd Thoughts LLC? Sadly and unfortunately to all of the odd Lots fan, I myself do not know what the next big thing in quantitative strategy or sort of advanced investing is going to be. But I'm hoping that our guests on today's episode might be able to shed some light. Who are there? Okay, so today we're going to be talking to John Elberg. He is the founder of Euclidean Technologies, a quant firm, as well as Zach Lipton, a professor at Carnegie Mellon University

in the Business School and expert on machine learning. They recently published a paper titled Improving factor based Quantitative Investing

by Forecasting company fundamentals. So what I think that means and we'll talk to them, is, you know, we talk all this stuff about price and computers and algorithms figuring out what signal we can get from price, But maybe the next generation can actually tell us something about the fundamental workings of the company itself, and maybe this could be sort of the next wave of where quant investing goes. And this sounds absolutely senating, Joe, let's bring them on.

John and Zach, thank you very much for joining us. Was that a reasonable characterization of sort of where your paper and where your research is taking things? Yeah, I

think it is so. So. First of all, machine learning has been kind of on a rocket ship of innovation for the last ten years or so, and with the advent of deep learning, you know, computers and machine learning have been able to do things that you know, historically have been very challenging, like image captioning and language translation. So we Zach and I, you know, a couple of years back, thought of the idea of collaborating to apply

deep learning to the problem of long term investing. So how did you actually go about doing that and what exactly do you mean by deep learning? That's exactly what I wanted to know. To deep learning sort of the rebranding of neural networks research to say I say I had some data about a company, right like I had

machine learning. We call a vector of features. But what we mean it's just like a list of attributes, each of which is somehow like be made into a numerical quantity, whether it's like their income, they're the number of assets whatever.

One way of deciding how to predict what the say, what the price will be or something, as we say, well, we're going to have this long vector of features, and then we're going for every single company, uh, you know, at every single time while this vector of features corresponding to the state of the company at some period of time, and then we'll have some target that we want to predict.

This could be a binary quantity like will the stock go up or down in the next you know, time unit of your choice, whether it's the next day or in the next month or in the next year. Or you could try to directly predict say the relative price, so like you know, the percent improvement or decrease based on sort of the available features. So one of the simplest ways you can make a model is you say, hey,

I've got a bunch of features. I'm gonna do is I'm gonna take a weighted some of these features the way like you'd calculate a score to see, like what's your risk of a heart disease. Maybe you take you know, well, four times your cholesterol plus two times your age minus one times you know, your amount of good cholesterol is something like this if you come up with some formula that's expressed simply as a weighted sum, so that would

be a linear model. Where deep learning make things different is that you have many different layers of computation that you basically are learning very complex patterns that maybe couldn't be expressed as a as a weighted sum. So maybe you're uncovering interactions between all of your features. Um. So, for example, if you want to learn to recognize a dog versus a cat in an image, there's no weighted sum of pixel values it's actually going to tell you

this because it's just the patterns too complicated. So in that case you need some some more like heavy duty machinery. So what you do in deep learning essentially is that you learn multiple successive transformations of your data such that after applying many such transformations, you know, could be two, four or five, ten, whatever, you come out at the end of a representation of your data where you actually can learn a very simple model on top of that.

So we sometimes call deep learning representation learning because it's what we're doing is we're both learning how to feature eye our data essentially, how to transform it and how to classify it at the same time. So one of the things in sort of traditional quantity, a lot of quantitative investing focuses a lot on price and sort of listening to your characterization. It seems like price and this is relatively speaking, of course, price is a fairly you know,

it's sort of easy idea to capture. So you can come up with some definition of what momentum is and then sort of say, okay, these stocks are experiencing momentum right now, or these stocks aren't, and then his history tell us the stocks are going to do next if

they sort of meet these characterizations. Your paper really looks at what can you do with this technology for sort of looking at future fundamental so looking at the sort of characteristics of the company and not just trying to see where prices going, but where those characteristics are going, so explain sort of what your research specifically attempts to uncover. So one thing that deep learning allows a researcher to do is look at kind of more raw features um.

Like Zach explained in the image case, you're looking at raw pixels. Now, if you think about most quant funds and most quant models, they the features that go into the model are highly engineered, and they include things like price and maybe book value, price divided by book value, price divided by earnings, and then maybe some momentum features. The interesting thing about deep learning is it allows you to potentially let it uncover what the best features are.

If you over engineer features, you may not find the ones that are best to predict what you're interested in predicting. So that, you know, allows you to potentially find features in the data that you wouldn't find through which traditional

feature engineering process. Yeah, and you know, to directly address your question, your point is that the very most obvious thing you could say, now, if I have this, I have this learning machine, I have a bunch of features, and I have to choose what am I going to predict? The very most obvious thing to try to predict is the price, because if you can actually do that perfectly, then you're done, right. If if you actually know which way the price is going to move in the next year,

then you can make the perfect choice. So the problem is that that's that's not so easy because the markets are quite capricious, right, Um, So one problem that we found is we actually did these models where we were trying to predict price directly. But among the other things that you have is that one, it's hard to learn models that do a good job of this that are

sort of robust across different time periods. So you might have like, hey, I'm going to train on these like decades of data and I'm going to try to directly predict the price. But then I come into periods of time where the markets behaving a little bit differently, and

we call this nonstationarity. Basically, like you're modeled, there's a great job of uncovering the pattern that's present in the data that you gave to the model, but that data is anchored to some period of time, and the future data that comes in, you know, the patterns changed a little bit, and so the kind of like function that

you've learned no longer does a great job. So so what we do instead of directly trying to predict the price, the idea that we had was to think, well, this core idea behind a factor model, generally right, is to just say, hey, I'm going to sort all the stocks according to some reason idea, Hey, the price of the company should be tied to its income, any company, and somehow it is justified by like it's the long term discounted cash as well. Let's just say a factor strategy

just something very simple. It says, well, let's just look at the current income divided by say the current price or current income divided by the current you know, market cap or enterprise value, some some notion of income and some notion notion of financial performance, and divided by some notion of company size and this, and then I'm going to sort the stocks according to this. The ones that come out highest are like most cheaply priced, so let's buy those. So the ideas to say, hey, well, what

if I told you so. We actually know that this does pretty well in back testing whether or not the patterns will hold in the future. But you know, many people have made a lot of money for many years, so there's an idea of if you knew the income, this is a good thing, a reasonable thing to try

to do. Our question that we asked, Unfortunately, um, John, because he's actually in finance and I'm not, has this really great set of like industry grade tools that unlike most academic papers that look at like one stock over a short period of time or something, we actually had, you know, forty plus years of financial data and can actually simulate like an applausible that guess what's going on. We said, well, what if you did a factor model,

but someone gave you a crystal ball. So basically, instead of dividing the current income divided by the current enterprise value, someone gave you next year's income, and so you sorted the stocks according to next year's income divided by the current enterprise value something like this. So you're you're able to peek into the future. You know how the company will be performing next year, and you're saying, is how is its next year's performance? Is that based on next

year's performance? Is its current price? Is it currently priced cheaply or not? So it's what we call like a clairvoyant factor model. Like you don't actually have such a crystal ball, but if you, you know, give us some license and you imagine that you did what would have happened if you went back in history and you had this crystal ball and you traded based on a clairvoyant factor model, and it turns out that the clairvoyant factor

model just crushes it. So it does really, really well and and not surprisingly, the more clairvoyant the model is. So if it if it knows the performance of the company six months out versus now, or twelve months out versus six months out, it keeps getting better and better and better. So what we decided was, well, maybe trying to predict price directly as a bit you know, subject to you know, a kind of fickle market, but the patterns present in the fundamental reporting data itself is more stable.

So in our method what we do is instead of just trying to predict a return, we try to predict actually the fundamental reporting data itself, just so we're given um these these features for like a trailing window of of time corresponding to the company's like financial reporting, and then we're trying to predict what they're going to report next year. And then based on what they're going we think they're going to report next year, we sort the

companies according to a value factor. So in essence, you can pick out of that future prediction the components of the factor model. Let let it whether it's a few future predicted earnings, and you can take that out of the future predicted fundamentals, divide that by current enterprise value and and sort and then you have basically a factor model which you are using. Instead of trailing twelve months earnings, you're using the future predicted earnings by the deep learning

the deep neural network. So, as I understand it, the deep learning or the neural networks are used primarily to forecast the future fundamentals based on historic performance. Is that right, historic fundamentals? Yeah, okay, So walk us through how you actually develop an application that's able to do that, Like what are those neural networks looking at and what sort of information are they drawing in other than you know, past predictive data to make those forecasts. There's two parts

of that. One is the data that we use, and then two is the technology we use to build the deep you know, neural network models. So on the data side, what you use is historical fundamentals on all companies you know that have ever you know, been listed in the US for the past fifty years, and so what a historical fundamentals mean? What it means earnings, book value, anything you can find on an income statement and balance sheet

going back in time. In addition to fundamentals, we also use as inputs to the to the model, you know, momentum over you know, one month, six months, twelve months. So then you know, if you think of it as like a big you know, spreadsheet table where each row is a point in time for a specific company, and then you can think of sequences going back through time. You know IBM in March of and then all of its fundamentals in one row, plus it's moment at them,

and then that going back five years and time. So those sequences, both the fundamentals and the momentum are fed into a neural network and uh and and all of those sequences for all companies and all time are fed into a neural network and are trained to predict what the fundamentals will be, you know, one time step out

in the future. So just to sort of summarize it all up, you know, it's like, if you have all these strategies, if you have all these funds chasing things like earnings, quality, earnings, growth, momentum, all kinds of stuff like that, your goal is to anticipate today when those funds are going to be buying tomorrow. Is that a fair way to characterize it. I think that's a fair

way to characterize is it. I think what we're really just doing is trying to build a better a better factor model, A better factor model in the sense that you know, as Zach explained, if you had a clairvoyant model where you actually knew what future fundamentals were and could plug that into a factor model, you do substantially

better than what you could achieve with a value factor model. Today, we're not like directly considering the psychology of the other players in the market in this particular approach, right, No, sure, but it's essentially saying, like, maybe the way to characterize it is, if you want to invest on some fundamental factor like earnings quality or earnings growth, bottom line, is better to look at future twelve month results rather than trailing twelve months. You look at the trailing, but you're

trying to predict the future. Like, so those two components, right, you could say, like one is we have the component that is trying to predict the future fundamentals. You know, imagine that I came for the future, and I got out of my time machine, and I gave you the earnings reports from the future. Right, So so the first thing you need is how do I get an approximate time machine, right, which in our case is a predictive model that has a good guess about what the future

will look like. The second thing is you still need a way of executing on the strategy ones I. You still need a way to decide which stocks to buy, right, So, based on based on this future information, Like, it's possible that if I if I come from the future and I give you the earnings report, and I tell you what the future income will be, what, it's possible that the income is going to go up with the stock price is going to go down, you know, like say

it's an Apple and like they made a lot more money, but it was also like announced that they had a major plant failure in the iPhone fourteen or whatever they're up to is going to be delayed. So these two components are are a little bit modular, Like we could come up with m. John I think is more the domain expert, so I'm I'm more the machine learning guy. Like I'm sure John could come up with you know, a million other ways that you might imagine that someone

would try to execute on this information. In our case, what we're doing is we've adapted a factor model to work with this kind of future guess. So one other example, so so again, in our case, what we're doing is taking the predicted future fundamentals and feeding that into a value factor model. But you could imagine using let's say the deep neural networks said, you know, a company is going to do a hundred million, but consensus estimates in

in in earnings. Let's say, but consensus estimates said it's gonna do seventy five million in in earnings. Well, you know that might be you could you can imagine devising a strategy around that where you'd want to go, you know, bet on those guys and ones where consensus estimates are above what the deep neural network is predicting, you'd want

to bet against. Right, assuming the current price is pricing and that that's a really you know, John, you shouldnt give away so goods, that that's a really good idea. So are these kinds of machine learning driven predictive models the future of investing? You think is that the way that we're heading. I think what this paper showed is that there's a lot of potential in using deep learning

to long term investing. I think that there's been some debate about whether, you know, deep learning, which requires a lot of data um to to to build successful models, um whether in finance there's enough data, or whether you even need this these kinds of complex models in finance, I mean a lot of quant people feel, you know, linear simple factor models are the best route to go, And I think what we showed here is that if you're trying to predict price changes, that might be true.

But if you decompose the problem into first trying to predict fundamentals and then later you know, through a factor model or some other method, trying to use those predicted fundamentals to predict price, deep learning has a lot of potential and does does substantially better at predicting future fundamentals than than what you could do with a linear model. There's a sort of a technical reason to recommend the way we've cast a problem also without going too far

into the weeds. Basically, uh, you think really really powerful machine learning models, deep neural networks. The thing that you worried about is John was talking about how people people agonize over what can you bring us to bear on long term investing because you don't have as much data right as if you were looking at the you know, micro second kind of trade frequency, then you'd have, you know,

trillions of trade examples or something you get on. But if you if you're looking at you know, your your time tick is I have a data point you know, once per month or once per year suddenly, and I only have thousands of stocks, not millions of stocks. You

don't have such a huge amount of data. Um, So what you worry about is that a model given given a super powerful model, like a super overpowered model, and then not too much data, that there's a propensity for the models to do what we call overfitting, which is the model basically it does a really good job of memorizing the training data it's seen, but it learns kind of a spurious pattern that doesn't generalize to future data

that it hasn't seen. So one cool thing about the way that we're casting the problem is that we're not just trying to predict the factor of interest. We're actually trying to predict all the factors in the future. And this means that the model has to simultaneously get the income right, and get the assets right, and get the debt right, and get all these different factors that are available. So John was a fifteen target factors that we have

that we're trying to predict. So so in this case, this sort of like this is this is what we call multitask learning and the machine learning literature. And one nice effect of multitask learning as that has a generalization effect in that it's it's harder to fit a spurious hypothesis because you have to come up with a representation that is good for task one and also good for

task two, and also good for tast three. And the probability that you come up with a pattern that's that's good for solving all of these tasks that is not the true pattern is much smaller than if you're only like trying to solve one task, where it's easier to

just kind of memorize those data points. So we have like essentially sixteen times as much training data and in some relevant sense, so I have to ask in the abstract of your paper or in the intro, you say that, um, with this approach, you can improve your annual returns pretty substantially over a standard factor model. In a bad test. Seventeen point one percent versus a fourteen point four percent,

just pretty big beat. But as we know, and as there's a lot of people pointed out, there's a lot of strategies that seem to work in academic papers and then when they're put into pract is, they don't seem the results don't seem to arrive as easily. John, in your firm, are you seeing the results of your research that on paper look very compelling actually play out in

the market. So so this this paper, uh, we we have not put this model to test, so to speak, in in a fund yet, but you know, we're very interested in in in doing that. I will I will add though here that many of the back tests that are done in the industry are done where you just run, you know, thousands of back tests on a data set over some time period ten years, twenty years, thirty years, and there's no out of sample testing, meaning that they don't then take that and then apply it to a

new data set. One thing that machine learning. One technique that is used in machine learning to prevent overfitting, and that we do here is we train them or we build the model on one data set and then test it at a sample on another on another data set during a different time period, and the results we present there are at a sample out of sample, always being sort of ahead right in the future, you could. So the model is the model at every given time. It

is trained on the path data. So we're simulating like what if, you know, if you train the model back then based on that it was only available up to that point. I think more broadly, there's a good question there of um, it's hard to say which which patterns are just you know, especially I think with short term investing, it's very obvious that any any pattern that exists on a scale of seconds is something that could be sort

of traded away. It's not as clear and I believe, I mean, John can speak more to it, right, But I believe part of the ethos of long term investing is very much that rather than interacting in a place where the price most price movements are due to the behavior of the high frequency traders, when you're in the long term space, the price movement is more tied to the actual fiscal performance of the company, and that's maybe

a more durable pattern. So Joe and I were talking about financial players and how quickly they adapt to new markets and new situations. At the beginning of this episode, from your respective viewpoints, how fast are these sorts of technologies and models and applications being developed, And for how long would something like, you know, a clairvoyant factor predicting model actually give you an edge four until someone else, maybe an E t F came along and copied it.

I think that's a hard question to answer, because again it gets back to how you would use this model. Right, So in the in the paper we give one very specific example. We use the deep learning normal network to predict fundamentals and then we plug that into one kind of factor model, right uh, in particular operating income predicted operating income over enterprise value. But as I suggested, you could use it to you know, figure you know, figure

out whether consensus forecasts are good or bad. Um. So, you know, I think that just saying in general, deep learning applied to you know, investing is going to get used and then a year later is going to be arbitraged away. Miss is the point that, look, you know, you can use deep learning in a myriad of ways to attack the problem of long term investing and presumably trading as well. To address your question about how quickly

is this kind of technology getting adopted? UM my sense and based a little bit on an outsider's view as an academic machine learning person, um talking to collegues who have either gone into fintech or who flirted with it

or tried to recruit me into it. The sense that I get is that actually, obviously a lot of people aren't talking about what they're doing, right, But my sense is that there's a lot of people doing this kind of stuff in the high frequency space, not maybe on the scale of you know, fractions of seconds, but but on on a pretty short time scale. And the reason why is because, um, right, it's it's easy to collect

a lot of data. If the patterns are very different, um, a year from now, Well, you just you have enough data. Like if you're trading like at the scale of months or years, then you have to look back twenty years, right, you have to look back thirty If you're trading at the scale seconds, then your whole universe could be formed by the previous four days. There's a very fast cycle of development. So if you're in it and you just wanna you don't you don't have any kind of strong

beliefs about finance. You're just a machine learning person throwing your hammer at financing. Then going in the high frequency space gives you or are the comparatively high frequency space gives you like the sandbacked box to just really quickly test stuff validated, see if it works. My feeling and when I've talked to friends who are doing this kind of stuff, but what we're doing is that I think almost no one that I've talked to out of a lot of people doing this stuff with finance, is looking

at the same kinds of time scales. And John might be able to to speak to that because he might actually be deeper and the I mean, he's definitely differ in the finance community than I am. But my sense is people doing deep learning for finance, and there are many, Um it's on the rise, but they're not necessarily looking at it in the same way, and certainly very few on as long a time scale. Yeah, I mean, I think if you look, you know, the a q r s and the d f as of the world, which

are you know, these huge you know, quantitative shops. They do. They certainly do long term investing, but um, there's not a lot of evidence that there's a ton of machine learning deep learning going on there. But you know, I think if if if stuff is a successful, you know it's likely to be a opted, so probably won't be true forever. John Elberg and Zachary Lipton, that was a fascinating conversation, so much to think about and wrap our

heads around. Really appreciate you both coming on. Thanks for having us. Thank you guys. Well, Tracy, we didn't really plan it that way, but I really do think that was sort of the perfect follow up to Andrew Lot left. No, Joe, you're supposed to pretend we did plan it that way so everyone will think we're really working so good. I mean, like we should continue this series on quantitative strategies and new ways to evolve to beat the market. Let's continue

this continue. Yes, absolutely, Okay, in all seriousness, Yes, it was fascinating. I really like the idea of well, who doesn't like the idea of a clairvoyant robot who can predict how well a company is going to do in the future and then apply that to a factor based investment model. If someone comes back in time and they're like giving me hints on what the stock market is gonna do, it's like just give me the winning stocks. You know what I'm saying, Like, if you're time traveling,

don't like be a tease, just give me the winning stocks. No, But in all seriousness, hey, I felt like several times in that conversation, it's just like the level that they're operating and thinking about the market on is so like high above anything that you and I like typically talk about it on a day. Like several times I felt like I had to catch my breath speak for yourself, Joe, because it's just like absorbing all of that, and you know,

obviously there's probably lots that I didn't get. But then the other thing I really thought that last point was very interesting about time. So obviously going back to the adaptive framework for thinking about markets. You know, if there is a strategy that works over a day and you can get it, and you can you just have to back test four days or whatever, it's very easy to see, Okay, this works, this doesn't. Let's go with the thing that works and then everyone can sort of figure out the

things that work and then it doesn't work anymore. But this idea that maybe a quantitative approach to long term fundamental investing, you don't get that sort of instant feedback on whether it's working as fast, and so maybe winning strategies might prove to be a bit more do Yeah, and presumably it's much more difficult to actually develop them

and see them evolved. It's like, um, I guess it's like if you bred successive generations of like rabbits, right, Like it takes like a year two, well less than a year, you could breathe like a hundred generations in a year. Yeah, right, Or I guess, like, you know, it's like laboratories use mice and because they do, like

they can get so many so fast. But if you had to sort of, you know, breed hippopotamus is you wouldn't know for a much longer period of time whether you're down the road you had sort of created the master hip hop. Does that make sense? Wow? This past? Yeah, Okay, let's leave it at master hippo. Okay, alright. This has been another episode of the Odd Thoughts Podcast. I'm Tracy Alloway. You can follow me on Twitter at Tracy Alloway and

I'm Joe Why Isn't All? You can follow me on Twitter at the Stalwart, and you can follow our guests on Twitter John Elberg is at at John Elbert. There's just one L in that. Zachary Lipton is on Twitter at Zachary Lipton, and you can follow our producer Sarah Patterson on Twitter at Sarah pat with two teams. Thanks for listening,