Q2: Jessica Stauth – Seeking Alpha? Try Alpha Factors

Chat with Traders, Collaboration with Quantopian, Episode 2. Hey, what's up folks? Welcome to the second instalment of the mini-series in collaboration with Quantopian. On the previous episode, we discussed various biases that traders and investors can be affected by. On this episode, we begin the discussion by talking about the things that quants consider when determining their trading universe.

From there we explore alpha factors and risk factors, and then we finally go through an example of how all of the above are relevant in an actual trading strategy. Delaney is with us again for this episode, as he will be throughout the series. But we are also joined by Jess Stouth, who is the VP of Quant Strategy at Quantopian and has a pretty impressive background in the field.

Uh, this is the point in the series where things do become more specific to quantitative methods of trading and slightly less relevant to those who favour discretion. But still, there's no harm in listening if this is something that interests you. As the subject starts to get a little more advanced. I would encourage you to listen over this episode twice if need be and make use of the resources that we mention throughout the episode.

All of these have been curated for you at quantopian.com forward slash chat with traders. So anything that gets mentioned, you can find there all in one place. Also, if something doesn't make sense to you and you would like to ask a question, you can do so at quantopian.com forward slash questions. Uh we will do our best to tackle all questions on part six of this series. And before we get going, I'd like to say a massive thanks to Data Camp for sponsoring this series.

You know I often get asked by traders who want to learn how to code, where should I start? And honestly, I think Data Camp is the best place in my opinion. They specifically teach you how to work with data, manipulate it, visualize it, and run statistical research, all using programming languages Python and R, which are of course ideal for quantitative trading.

For more info and to start any one of their online courses today for free, visit datacamp.com. All right guys, let's skip to my discussion with Delaney and Jess. I hope you enjoy it and thank you so much for listening. All right. Well Delaney, how's things, Ben? What's new? Uh I mean that's a loaded question because we're always doing a lot of stuff here.

So uh I don't know how much time do we have for just that question alone, but a lot of stuff is the is the short answer. Um and uh we're we're just kind of busy all the time. Excellent. Keeping busy, I lock it. Now we also have Jess Steuth joining us for this episode. Jess, how are you?

Great. Thanks. Thanks for having me. No, my pleasure. Uh it's excellent to be speaking with you. So, I mean, tell us a little bit about your story, Jess. Share with us a bit about your background. I think you were originally a a physics and chemistry teacher, right?

Uh sh I did one year of teaching physics and chemistry, but I'll certainly take credit for that. Um so my background really was in basic research, um, did physics undergrad, but I went uh to UC Berkeley and did a PhD there in biophysics, um which is a very, very broad field. I ended up in a computational neuroscience lab. So I learned a bunch of analytic techniques there for analyzing complex large data sets.

um really enjoyed data analysis, didn't really enjoy uh all the data collection. And so really looked at quant finance as sort of the best possible world where you got to ask really interesting. complicated questions about uh large data sets. Um, but you didn't have to necessarily go through all the blood, sweat, and tears of putting grad students in a extremely high magnetic field.

Um, so that seemed attractive to me. So I took my first job in San Francisco actually for sort of a quant modeling research team. Um, so I became a quant analyst, got to learn a lot about the financial markets and specifically um equity stock selection models. So at that company, uh Starmine was called the time, I did things like building a stock ranking model that took inputs like how much uh short availability for certain stocks was there.

So I spent a few years there. That company was sort of a startup. They got acquired by a much larger company called Thomson Reuters. Um, so I spent a few more years uh there, kind of more broadly looking across their quantitative analytics and data. I got the chance to visit.

Quantitative asset managers, hedge funds, really around the world, which was an amazing experience. I got to see uh inside quant shops anywhere from, you know, two or three folks uh working as a small team up to, you know, desks of, you know, a hundred researchers or so. Um and uh through the course of that move from San Francisco back to the the East Coast in the US to Boston, and that's where I met John Fawcett, who was

sort of in the process of starting Quantopian, he had a back tester. Um he had a community that was really taking off. Um, so I just got really excited about, you know, sort of all of the cool stuff that Quantopian was doing in this space that I felt like I already You know, kind of understood reasonably well from sort of a professional side. Quantopian was sort of opening this up and blowing the doors open, I guess, to this.

You know, sort of really elite secretive kind of high tech field um and letting, you know, sort of anyone on the internet uh with access to a web browser. sort of have access to these same tools and data and analytics that, you know, I was previously seeing sort of be available to, you know, a very, very small select, you know, group of folks around the world. So joined Quantopian about three years ago.

And my role at Quantopian is VP of Quant Strategy, which uh entails running a little research team. Um, we evaluate algorithms on the Quantopian platform. that have been written by our community members. Um, and we're looking for strategies that, you know, beat the market, basically. Um, really to get a fine point on it, we're looking for strategies that

uh sort of have risk adjusted returns. And I think we'll probably end up talking a lot about that later on. And so we're we're looking to evaluate strategies and we then actually strike license agreements with folks in our community and we license their intellectual property and we make allocations um to their algorithms and then monitor how they're doing.

Excellent. Well that's a very cool story. I just want to go back to the startup that you worked at, which you mentioned there. Um sounds like a very interesting kind of business that you were working in there. So you were creating these uh these trading models essentially, but you weren't actually trading them. What were you doing with them? Were you were you selling them off to like larger firms? Like how did that work?

Yeah, great question. So we were building these stock ranking models and we had sort of a whole suite of models, anything from, you know, basic technical price-driven factors up to one of our flagship models was looking at analyst revisions. Um, so you know, there's a whole cohort of south side analysts uh reading stocks all the time.

And we had built a model that would ingest all of those estimates from the IBIS estimate service and basically make predictions about, you know, as analysts were changing their opinions, what did that mean for stocks? So The output of those models got productized in two ways from from that team. One was they would be sort of fed into a dashboard that portfolio managers could buy access to. So they would kind of be shown as like, you know

you know, red, green, and yellow lights, right? Like this stock looks like a buy or a or a hold or a sell based on our model scores. And then sort of more interesting to me was that we would actually sell a data feed of these one to a hundred rankings. So we'd sort of be delivering a file, you know, on a daily basis to quantitative managers. Every single stock would get a score every single day across all these models.

Now those folks, you might say, you know, why would they want to buy a stock ranking model? That's their job. But it turns out, you know, as we'll dig into, there are so many steps in the process of going from raw data to, you know, a a portfolio that you want to buy. um and trade. So there were a lot of folks that were interested in buying sort of these Lego blocks from from that company, Starmine. Um and and I should say they're certainly not the only type uh of company that would s that sold

stock selection factors like that. Um one other big one was called QSG. Um and they've been acquired by one of the other large um, you know, uh one of the other large data vendors. So there's another number of companies like their like this out there. Um and their value sort of is, you know, in that

range where they're going from unstructured or raw data, um, processing it, cleaning up cleaning it up, productizing it, and reducing it down to a more simple factor. Okay. Now I'm not sure if this is gonna be a good question or not, but

Was it was anyone ever put off by the fact that you were selling these feeds to multiple firms or multiple funds? Like was that ever an issue? Yeah, so folks would ask, you know, okay, um, why am I gonna buy this model from you if there is any predictive signal in it that would be arbitraged away by other people using it? Um so that's like a definitely a common sort of objection to purchasing like an off the shelf factor.

Um, and you know, we could answer that in a number of different ways. You know, one thing we would say was, you know, so there are, you know, how many people directly looking at analyst estimates and trading off of that information, either systematically or not. And compared to sort of that number of people in the market analyzing that data.

you know, there's actually a relatively small number of people that would be, you know, necessarily like buying these these models from us. So sort of one thing we would say was like, sure, in the limit, if we sold this model to every single market participant. Then everyone has the same information. But to the extent that not everyone can afford this model, so we would make it, you know, it's funny because.

You're in a business when there where you know your product almost sells better the more expensive you make it. Um so so there's an interesting um you know, sort of supply and demand curve, I think. And and actually, you know, I think it's a great question because it doesn't just apply to, you know, those factors that that we were selling then. I see it actually in um sort of the boutique data vendor space.

Just more broadly, even today, you have this interesting tension between, you know, if you think you have data that Very, very valuable. Um, you're trying to figure out how valuable is it and how do you maximize your revenue on it? Do you try to sell it for a very high price point to a very small number of people?

um, you know, or or do you sort of risk selling it, you know, more broadly at a lower price? It's it's uh, you know, not a solved problem. It's a really interesting question, I think. And if I could just

Well first of all, like obviously this term factor is being thrown around pretty liberally right now and and I you know I think we should actually get to like defining what that is. Um but I I also wanted to make the point that uh Something that that was interesting to me when I started figuring out, you know, how data sets are actually used in finance is there's kind of this sense that if you develop a stock prediction model that

The only thing you can do with it is use it to try and buy and sell stocks to make money on that, on those trades. But that's actually not the case. And there's a lot more uses, a few of which we'll get into today, for you know, being able to score every stock.

Um and uh just as a quick example If you know that some other firm is located closer to an exchange than you are and can make trades faster, and you know that they're using a certain uh factor to trade, then if you can get a hold of that factor you might actually use it as an indication not to trade yet, because if you know they're going to trade on something like that.

then you might want to say, Well, I know that they're already going to trade on it. I know that they're going to trade faster than I can. So why would I just throw in this order, which is just going to get behind them in the queue? So like there's There's more sophisticated uses of this data than just using it to try to directly make money based on trades. Right, right. Okay. Yeah. I mean that's that's a fair point.

Um, well Jess, I know you're gonna be an awesome fit for the topics we're gonna get into uh during this episode. Um just quickly, last episode we covered uh biases. So anyone listening to this, if you haven't heard that first episode, make sure you go back and check that out.

Uh it was a very interesting episode and uh Delaney you might find this funny. It actually inspired me to uh go out and get a copy of Fooled by Randomness by uh Nassim Taleb. So um I've just started reading that. I'm sure you've probably read that at some point. Uh I actually haven't read that book yet. I mean I I've actually I've taken a a a course in the past which covered a lot of You know, kind of this notion of structural uncertainty and tail risk and how different

It was actually a course offered by a philosophy department, so it was really fascinating because beyond just like the numbers behind it, we got into what does risk even mean and and how different people think about it. And and we we did read a lot of um Taleb's stuff.

for that for that course. Uh but what I find interesting is there's actually a little bit of a of a of a war going on right now, I think, where if you look at someone else who's also very high profile right now in that field, which is Nate Silver. Um he's actually been in a little bit of a feud with with Nassim Taleb recently. So I I think it's uh what's actually also fascinating is to kind of see how they fight each other on on on these concepts.

Okay. Very interesting. I'm I'm not familiar with um who was it? Nate Silver. Is that who you said? Nate Silver. He's a little of a little bit of a celebrity in the US now because he's a he's a modeling guy. Who um started a business based on being able to predict political elections and had predicted the last two federal elections in the state. like incredibly well. Um and also maintains a lot of uh sports predictions and his he runs a website called five thirty eight uh and has a book called

uh the signal and the noise. And I think those are kind of great resources for just learning more about the notion of statistical rigor in general. Very cool. I'm gonna check that out for sure. Now this episode we're covering various factors in a quant workflow, uh in particular alpha factors and risk factors. And we're also gonna go through some examples of these as well.

Um before we do so though, let's start with universe selection. So I guess the first question I'll ask around this is what are some of the ways that quants decide which products or markets to include in their universe for a particular strategy or portfolio. Sure. So um there are a lot, there's a lot of ways to think about this. I'm going to take the attack of answering this.

with respect to what I'm gonna call sort of a cross-sectional quant equity strategy. So that's the type of strategy that that we're sort of seeking in our community and looking for for allocation. So If you're going to do something where the design of your um prediction is to take a set of stocks. and compare them all against each other to try to find out.

which of these stocks do I want to own and which of these stocks do I actually think may go down in value? Then really what you're looking for is to find a big basket of stocks that are all relatively comparable to each other. So, you know, what are things that make stocks comparable to each other? Are they all treating in the same market? Are they all to some broad extent liquid? So they can all, you know, you can sort of exchange one for the other without sort of really undue um conflict.

And then sort of well, liquidity covers a lot of it. There's a lot of different ways to look at how liquid a stock is. So one thing people will do is say, Um, okay, maybe I'd want to choose to select a universe of stocks that are all large cap or large and mid-market capitalization, or at least exclude something like micro cap stocks.

So a lot of the way that I would think about universe selection would be to say, okay, let's consider if I can buy US stocks, right? Or or any market stocks, publicly traded stocks. Um, I actually tend to think of it a little bit like what do I want to exclude that are going to be

um stocks or treated instruments that are going to be in that set that don't really rank very well against everything else. So some common universe selection techniques we see people apply would be getting rid of illiquid stocks. Um actually like if you just look at sort of every ticker listed on the US stock exchange, for example, um, and you go into, you know, Quantopian's back tester and just ask for everything.

You're going to not only get stocks actually, you're also going to get ETFs. Um, you're also gonna get ADRs. ADRs are uh depository receipts that are an instrument trading on one exchange where that company's actually located in another country. So you could sort of look through those things and say, hmm, you know, maybe an ETF.

exchange traded fund that itself holds, you know, another couple of hundred stocks or it holds futures or holds commodities. Maybe I wouldn't want to treat that in an apples to apples comparison way with you know, Apple or Microsoft or or IBM. So it's a lot I the way I think about it for this type of ranking workflow that that we think a lot about. is really trying to get to a sample of uh choices, tradable instruments that are kind of comparable to each other.

And and I think another important point there is that a lot of it is trying to make the real world implementation of your strategy as close to your your mathematic, you know, uh formulas as possible. What I mean by that is like if your math works out to purchase

ten thousand dollars worth of this stock um and that stock is going through a very illiquid time right now, you're not gonna be able to do that. And then how will that throw off your portfolio risk? How will that throw off your strategy in general? Um and you know, so if you're if you're if you're trading on stuff that's illiquid or you're trading on anything that's weird uh or maybe is prone to bankruptcy, like super microcapped or volatile stock.

It just means that in the future when you try to implement this strategy, you might get a model that looks great on paper.

but, you know, is just not gonna trade well at all. So really what universe selection does is not only makes it like Jess said in my like a comparable, you know thing when you're when you're building these models, but also just saves you time so that less of the time you build a model, it looks great, but then you try to trade it and it turns out you can't because of real world constraints like you know, liquidity.

Okay, so how should a trader or a a quant think about the size of the universe? Is this dependent on maybe the amount of capital that they have? Does that factor into it? Yeah, maybe speak a little bit about how a trader should think about the size, like how many stocks or products or whatever it may be to include in their in their universe. Uh yeah, so I think this is pretty capital dependent. Um and in general the way I think about it is let's say that you have some predictive

uh model, some predictive methodology? Well, you want to make as many independent predictions as possible. Uh and it's for some of the reasons I think we discussed a bit in the previous

uh in the previous episode. Um and we actually have a full lecture on this if you go to Quantopian.com slash lectures. There's a lecture called uh position concentration risk, which does a bunch of simulations to really show you that, you know uh if you if you have a slight edge over random, you should really try to exploit that edge as much as possible by applying your prediction to every single thing for which your prediction is valid.

And the more things you apply it to, the more diversification you're gonna get in your portfolio, the less risk you're gonna face overall. Um but of course, like you said. If you don't have that. uh capital, then the you know, amount of stuff you can the amount of positions you can enter is gonna go down because you're gonna be paying some kind of fixed trading cost for each position. And at a certain point the ratio of trading costs to profits is going to be too high.

So effectively the way I think about it and maybe Jess if if you if you have another perspective you can jump in, but the way I think about it is you want to always be invested in as many things as possible. The constraints being what can your model predict on and how much capital you have available for allocation.

I I think that's right. I think the only um thing I maybe would add is you can sort of also draw a distinction between the size of the universe that you're ranking or that you sort of want to look in and then the size of the portfolio that you sort of choose to trade.

So you might decide, okay, I've got, you know, a relatively small amount of capital I want to deploy today. So I'm stuck, you know, with a portfolio where I'm really only going to be able to buy and and sell short, you know, ten or twenty names, stocks or instruments on each side.

But I still want to look at and evaluate a universe of let's say the top five hundred most liquid instruments. So you can take a a large universe, rank it, and then you can say, cool, now in theory, maybe I've ranked this universe like one to five hundred, where one is my favorite stock to own and five hundredth is my favorite stock to sell short. And then you can kind of just decide as you build your portfolio of longs and shorts.

kind of how far in do you come from that edges of that distribution? And so like Delaney said, the further you come in from the edges of the distribution, the less position concentration risk that you have. And so, you know, sort of the smoother the returns that you're likely to see are gonna be. Um, but you have again this fixed cost where you know each individual stock that you trade has probably some cost associated with it. Um, and at you know, small portfolio values.

You get rounding issues with the, you know, you're not going to be buying round lots. Um, not only are you not gonna be buying round lots, you might have a really small portfolio size. you're not even gonna be able to afford, you know, round numbers of shares, right? You'll get these weird rounding errors. Um so I think, you know, there's just a little bit of a distinction um or or can be.

if you think about it, um, between the universe that you look in to find your portfolio and then how big of a portfolio you decide to trade.

Okay, just to pick up on something you said there, Jess, I think that was a really awesome answer, by the way. Um when you spoke about rounding, what are you actually referring to there? Sure. So usually when I think of, and I'm I'm maybe like skipping ahead even past the end of this session, but when I think of building like a portfolio, the way I think of it is.

Let's say we we don't know how many dollars we invest in this portfolio. It's sort of like a unit portfolio. Then instead we're thinking, well, what percent of my portfolio is going to be, you know, in any one stock? And so I'm gonna come up with sort of a a target portfolio, right, based on my model. It's gonna say, I wanna be, you know, ten percent in Apple, ten percent in Microsoft, ten percent in IBM, however.

Um, so then you go to like actually implement that and that then you're multiplying your percent weights that you want in your portfolio by the amount of dollars you have to invest in your portfolio. So let's pretend I have a million dollars. And I want to put 10% in each of 10 stocks. Great, I have$100,000 to put in each of those 10 stocks. And let's say the stock price is, you know, fifty dollars. Well fine. Then I you know

Ma that does, you know, that divides in evenly. But what if I want to invest like$1,000 in each stock, right? If I instead have only$10,000 to invest. Well, now I have$1,000 to invest in each stock. What if that stock itself costs? You know, two hundred and fifty dollars a share. And I can afford four.

Okay, what if the next stock um, you know, is actually I don't know what Google is these days, but what whatever, you know, like they split actually. Okay, so let's assume that they're still a thousand dollars. Yeah, so let's assume then that Google's like eight hundred dollars or or something that's, you know, eight hundred dollars a share. Well now

Maybe my portfolio wants to have some position in that, but I can only buy I can't buy 1.1 shares. Um actually uh you know, I can even simplify this to say You get down to a one of the cons, one of the rounding issues is you get down to a level where even if it wasn't that expensive to trade, you really ideally want like fractional shares. That's like a whole separate topic of conversation, but a lot of folks

um think that's a really interesting spot for sort of retail fintech innovation. There's a company in the US called Motif where they let you buy like these really small little representative portfolios where you say, Oh, I only have, you know, whatever, five hundred dollars to invest.

Well, you'd kind of want to invest that in like a diversified portfolio of stocks and they'll sort of figure out a way to get you into those fractional shares. But but that's something that that is sort of a tricky problem for, again, like small retail level portfolios. Okay. This is interesting. So just to be clear, and I I think I understand why this might be an issue, is because if you want to be uh evenly invested in these ten companies, you know, a as the example goes here.

you're not going to be able to do so. if uh the the actual dollar amount of the shares doesn't equal the ten percent that you want to be invested. So you're like not evenly invested even though you want to be. Is that um is that

What you're getting at? Yeah. Exactly. Because, you know, I mean, so if you're like a total O C D nerd like I am, I wish like every stock had to be like, you know, one share is like ten dollars roughly and they have to like You know, you don't actually splits are a nightmare, so you don't really want splits, but the problem imagine uh emerges, right, that

Some stocks are penny stocks, some stocks are a couple dollars a share, and some stocks are very expensive. So um if you have small amounts of capital, then when you go to try to split your portfolio up into many, many, many investments. Exactly. You run into these rounding errors where you sort of can't get, you know, seven point three percent exposure to this one stock because you can't, you know, own a fractional share, for example.

Yeah, and it it gets back into what I was saying earlier, just like an infidelity between the math and the real world implementation because If you tell your broker or your if your algorithm tells your broker, hey, I'd like to be ten percent invested in each of these ten names, and then your algorithm doesn't really have any guarantee that anything close to that is going to happen, uh, we've just kind of

We're probably throwing away our entire edge because the amount of noise in what we might actually buy and sell is gonna be so great compared to kind of any edge we've attained by our through our math that we're i i it's it's really tough to to get anything done.

Right. Okay. Now just while we're still on this topic, are there any common mistakes that stand out to you which you often see traders and quants making when determining their universe? Like just not doing it at all. Is the biggest one. So like not having a step that defines what am I willing to consider in my model. So um, you know, if you're basically just saying like, hey, I'm going to look at

I'm going to look to find um what's, you know, let's say you're looking at like a technical price driven signal. So you just like sort everything based on last five days returns. Um, but you don't even consider that that thing that's sort of to the top is like a triple levered ETF, right? You just buy it or sell it. Um, so just the worst mistake I think you can make in universe selection is not doing it.

Yeah, and I would say that's kind of uh I think typical of people who are now entering the quant field as, you know, more and more opportunities to do so are arising these days.

Um but you know, people who are coming in from maybe more traditional discretionary trading background and they're entering the quant field and they're saying, Let's apply my normal thing, which is I'm gonna, you know, try to do some research and here's what I used to do research and I n oh my gosh, now with a computer and Excel, I can actually look at every stock every day and I'm oh my gosh, this one has such a great ratio.

Um I'm gonna grab it. But you know, that there's just no thought to do that that universe selection step in advance. Okay. Yeah. No, I think those are really great points. That's awesome. Now

We're about to get into factors. So let's tackle, first of all, what exactly is a factor in a quantitative sense? So I mean, I I think that and again, I I said this before, finance is one of the fields that because of the amount of secrecy of all the companies working, um, I think that there is a a pretty big kind of Uh

not disconnect uh inconsistency in in the terms that are used by various different people to mean various things. Um so, you know, when I define a factor, I think this is a pretty broad definition of a factor that's accepted by by most, you know, kind of uh professional quant quant teams, but different people might have different words for what I'm saying. Um

A factor can mean one of two things usually. It can mean an alpha factor, which is what Jess has been talking about with kind of a stock selection model. An alpha factor is Any way that you assign a score to each stock at you know regular intervals. And that score is supposed to be representative of which stocks you would like to own more and which stocks you would like to own less. The score doesn't have to be the expected return out of each stock.

or the expected future price. Um, it just has to be representative of like better or worse. Um and then the other uh way that people will kind of the other thing that people refer to as a factor is something known as a risk factor. Um and when you get into it, a risk factor is literally just a return stream. It's a time series of of returns. And we'll we'll talk about that, I think, um a little bit after we we talk about alpha factors a little more.

Yeah, absolutely. And I mean when we talk about fact is um as such. Are these factors that apply to individual strategies or is this something that is only applicable on a portfolio level? So let me sort of replay like one sort of high level uh

way of thinking about an alpha factor versus a risk factor, like Delaney was just saying, um, that helped helped me. And I mean, I've been like thinking about these things for a long time. Um, and and still like you'll come up with new here are new ways people describe these things that are really helpful. So Um in any case, a factor is like a single number per stock per day or per minute, right? So it's it's like a piece of of data. Like Delaney said.

It could be a return stream, um, but it doesn't actually like have to be a return stream. It's it's a single number, the way I think of it, it's a single number per day per stock that's gonna tell you something. An alpha factor. uh that number should be correlated with or predicting the direction of future returns, such that you should be able to use that factor to pick stocks that you think are gonna go up or down.

um in some way, right? Feed that into some model and and it should actually have power of predicting the direction of your returns, the sign of your returns. Whereas a risk factor, and we'll talk about what they, what some of them are. is something that y you're at best you hope is going to potentially be predictive of sort of the volatility.

of the returns going forward. So if we say something ha is risky, that means we don't know what its price is going to be tomorrow, but we expect sort of the variance between today's price and tomorrow's price. to be relatively large. If something's very low risk.

That means I don't know whether it's going to go up or down tomorrow, but I expect it to if it does, go up or down like by a relatively small amount. And all of these things are relative, right? So that's kind of part of why universe selection is the right first step because These factors tend to be only useful on a relative basis. So that's the distinction I the way I think of.

Uh, you know, as we start to talk about what some of these things are, there's often like sort of almost an ideological or philosophical debate that can happen with people like is that thing an alpha factor or is that thing a risk factor? Um, and the the definition that I like a lot is an alpha factor should be predicting the sign of the future returns and the risk factor, you know, at best hopes to predict the variance of the future returns.

So and in addition to that, so I I think it might be helpful to actually just like walk through a concrete example of this. Because at you know, at least when I was kind of figuring out this stuff, one of the hardest things was just getting my head around like, what is this thing actually? What is it, what is it, what does it do, what does it mean?

Um and one thing I just wanted to say is that uh We're gonna set up um for like any of the content that we talk about or or or point to on the podcast, uh, we'll set up and we'll have a we'll have it so that that's all listed out so everybody can kind of go and check it out, interact with it, any you know, notes that I I say, hey, you should check this out. We'll try to get it in there. Uh the link for that is going to be quantopian.com slash chat with traders, no spaces.

Um, and and we'll have that all set up by the time that this this podcast airs. So like as I'm talking about this stuff, you know, we'll we'll do our best to make sure that there's actually concrete examples that you can follow along with if you want to pause this podcast and actually like try implementing this in code and see what happens. But like a concrete example of this might be something like uh market cap.

Right. So market cap can be used, I mean, both as and I arguably maybe not a very good one, but both as an alpha factor and as a risk factor. And and what I mean by that is the following. So Every company at every point in time has a market cap.

And so in that sense, it is an alpha factor. Again, like is it predictive of future returns? Well, it was, you know, uh one of the original Fama French factors uh went which if for people who aren't familiar, Fama French Um Fama is a uh Fama and French are both professors um at University of Chicago in Dartmouth, and they came up with uh uh basically factor modeling as a whole, they really kind of s started um and they developed a factor model that explained like a large percentage of

uh returns in the market and one of the original factors was market cap. Now of course That's so old now that, you know, I don't I haven't done any work recently. I haven't checked if market cap is still explanatory of future returns. But um the idea is that That's an alpha factor. You just say every stock has a market cap at every point in time. So you could set up a portfolio where you weighted everything in your universe by its market cap. and then went long or short on the on the biggest

uh cap companies and the smallest cap companies. And you could be long market cap, which would mean you'd be long the biggest companies and short the smallest companies, or you could be short market cap, which would mean you'd be short the biggest companies and long the smallest companies. And on the flip side, now what you can do is you can say, okay, well, let's say that you theoretically had that portfolio running.

Right. And you had been long the big companies and and short the small companies. And ignoring for now like transaction cost and slippage, let's just say what's the return on that portfolio? over time. And and that return stream is used as the risk factor. For market cap, because the idea is that that's a well-known factor, and it's kind of a fundamental thing that moves along in the market.

If your trading strategy, you know, whatever your trading strategy is, it could be an algorithm, it could just be you manually placing trades. If your trading strategies return stream has dependence or correlation with this well-known market cap return stream, Then that means that you are exposed, you're dependent on it, you are carrying some of the risk from that risk factor into your returns.

I I don't know if that's helpful in kind of trying to understand maybe the split between an alpha factor and a risk factor, but the general idea is that you can convert one to the other fairly easily because a risk factor is just what are the returns Had you traded on this alpha factor, and what is you know your some other strategies dependence on this on this known risk factor?

Maybe what might be helpful is if we could talk about how how a quant may go about actually seeking out an alpha factor. I mean I know that that might be um that might be a whole nother conversation on its own, but maybe if we could just sort of briefly uh run through that, that might be helpful. Yeah, I I I think we can I think we can sort of talk about that at at a high level and then I can give a specific example that I already mentioned from from the work I did at Starmine. So

Um, you want to find an alpha factor. What that means is you're saying, I want to find a piece of data that I can know about stocks today that predicts whether they're gonna go up or down tomorrow. So the really the most straightforward way of doing that, uh, well, the I think best and most straightforward way is first to develop some hypothesis that you want to test.

Intuition, economic rationale. What's a thing I could know about a stock today that might tell me if it's gonna go up or down tomorrow? So an example that I mentioned earlier is, okay, well what if I know today how many, like what fraction of the shares of a given company have been borrowed. And lent out for short sale. So the economic rationale for that factor being predictive would be would go like this.

Uh shorting stocks is expensive and you know relatively uh tricky to do. Uh you pay interest when you short a stock, and not everybody knows how to do it. So people that are out there in the market shorting stocks. are s what we might call smart money. This is the hypothesis. So if I could know today for every company traded on a given exchange, what fraction of its shares have been sold out short?

I might develop the hypothesis that would say companies where, you know, uh there's a lot of their shares that are sold short, those companies are in trouble. And I think they're gonna go down. If not tomorrow, soon. Companies with almost no uh shares sold out short, those companies are probably fine. So I might want to hold those. So that's sort of your first step as you develop

a hypothesis about a piece of data that you could have today. Then you basically need to test that. So one of the things that my research team at Quantopian has built is a couple of open source Python libraries for looking at this type of stuff. The one we built for analyzing alpha factors is called alpha lens. So the one of the ways that people would test this type of thing would be to say, okay, great. So let me take my factor. That's what I know today.

So that's how many, you know, what fraction of shares are sold short, for example, today. Now let me run just m maybe basically like a correlation, right? Like a rank correlation or some sort of regression against tomorrow's price.

So let's pretend I I can take tomorrow's price and roll it in today or take the price five days from now. And I basically just say, you know, is there a correlation? If I sorted my universe and ranked stocks and I shorted the ones with the most short interest, it's called. Uh and long to the ones with the least, then how well does that predict or rank stocks based on their returns one day ahead? Five days ahead, ten days ahead, what have you?

And so I actually can get like a single number or a distribution estimate that says how predictive, like what amount of tomorrow's returns can I explain today with this single number. That's the basic idea. And like I guess another way you the the very last step is is you say, okay, like I had a hypothesis, I tried it, there's no predictive power, which is usually what happens.

I have to discard that and start over again. Or, you know, you you either get lucky or you've been working on something for a long time, you finally find something where, yes, there is predictive power. So a lot of it is like, you know, what I would consider any kind of research scientific process where

You develop a hypothesis and then you set about really trying to disprove it until you find something that you know you really kind of can't, can't discard or disprove, it actually seems to to hold up.

Okay. Now at the risk of sounding like a total amateur myself, um, is an alpha factor just a fancy term for a trade signal? I mean, so the thing about alpha factors is I might argue that a trade signal is just a less quantity way of thinking about an alpha factor, right, in the reverse direction. But the thing about alpha factors is yes, they just they encapsulate

a pricing model, right? Each each alpha factor really just has this sense of can you predict future motion of stocks? Are this gonna go up or is it gonna go down? So in the sense that you had a trade signal at every point in time and you either bought or sold based on that trade signal. Well if you just took that trade signal and applied it to every stock, now you have an alpha fact.

I mean The notion of an alpha factor is just taking A predictive model, a trading signal, some forecast of returns, and just applying it to everything, and then looking at cross-sectionally how the entire universe does rather than like any smaller portfolio stock. Okay. Okay. Now just because you might have a a quality alpha factor, right? There there's something in that, what you've what you've come up with

That doesn't necessarily mean you're going to be profitable when you implement that because an alpha factor is still totally separate from how you actually manage a trade or manage a position. Um am I right in saying so?

Yeah, absolutely. So, you know, it depends a lot on sort of what's the end strategy that you're trying to achieve, but the alpha factor, let's say even if it is predictive, um, there are a lot of other steps. for a quant um in implementing sort of a a a portfolio um you know managed strategy that come come after that. Um sort of

without even kind of getting into the whole process, one very common pitfall or failure mode is that your alpha factor is predictive, right? It does sort of contain a forecast about tomorrow that's true. Um, but the edge is so small that once you actually try to go into the market and trade that. um you can't overcome transaction costs. So on paper, it looks good. But if you don't model the costs, the friction in the market, realistically what you pay to trade it.

you actually can't achieve a profit. So one thing that people try to do, although, you know, it's it's very challenging, is they try to sort of have that alpha factor, you know, have a forecast of sort of how much you expect the returns to go up or down.

If you think you can achieve that, which is like really a little bit tough to have a lot of confidence in that, but if you think you can achieve that, then you can you think you have a forecast that says, I think this stock is going to go up by, you know. whatever, point one percent tomorrow. And then you also know, I know it's gonna cost me, you know, 0.1% to make that trade.

you have zero profit left over. So sort of one common failure mode is you find an alpha factor, but it's a weak alpha factor. So on its own, you can't exceed transaction costs. The thing that that sort of cross-sectional quant strategies try to do often is they try to say, okay, cool, I found one alpha factor. It's weak on its own. It's probably not good enough to overcome costs.

So let me look and try to actually find and sort of roll up many different uncorrelated weekly predictive signals, which does get into sort of another conversation, but that's sort of the direction that, you know, especially, you know, in Today's modern market where factor analysis has been, you know, well understood for

you know, decades certainly. Um, and some of these single factors just aren't sort of strong enough to overcome, you know, implementation costs anymore. The sort of modern quant approaches tend to actually say, great, let me sort of know and understand that this individual alpha factor is is relatively weak, uh weakly predictive. But if I can find lots of independent and very different sources of weak predictions.

and I can wrap them up together in an intelligent way, now I much more likely to have something that's A, gonna overcome transaction costs and B gonna work. sort of all the time. So the other risk with with alpha factors is that they're predictive, but not very consistently. So maybe you have an alpha factor that works really well.

you know, and I and I think traders tend to think this way too, like maybe I have a trading signal that I think works really well in bull markets or bear markets or very volatile markets, you know, or not so volatile markets. And if you think you have found signals that actually are sort of more intermittent in their um in their value. then it's an even stronger case for aggregating them up together with other signals that might work well at different times. So I would think of it that

you know, finding one good alpha factor is great, um, but is just one step of the process, you're probably gonna need to find several. And then, you know, I don't think it's necessarily in scope what we're talking about today, but There are a lot of challenges to then putting that into a portfolio that you can trade profitably in the market. I think of it as the most interesting fun step of the process.

Um, but it's certainly not the end of the process. Yeah. And I think that actually that that's kind of a summary of of what we're gonna be talking about in the next podcast, which is all the considerations that you might actually go through and reasons for combining many weekly

predictive alphas into one big, you know, monster alpha. And there's actually kind of sophisticated things you can do. You can you don't just have to average them together. Um but you know the general intuition is just like just said, if you've got

uh you know, 10 things and each of them has some signal and and some noise, well as you average more and more things together, um, you know, the noise, because it could be up or down is going to start canceling itself out, and then you're gonna be left with a stronger and stronger signal.

Uh and this is just something that's, you know, ubiquitous to any kind of statistics or predictive modeling or stati you know, uh uh data science. It's just you're constantly trying to average together weekly predictive things to get it a stronger signal. Uh and but again we'll we'll talk about that a lot more uh in the next podcast. Okay, Jess has to leave.

Thank you very much, Erin. It was really fun. And I'm sorry I have a hard stop. I have to get on a bus to head home. No, no stress. It was awesome speaking with you, Jess. Thanks very much for doing this. Cheers. Yeah, no problem. Thanks a lot.

So we's we spoke there a lot about alpha factors um and you briefly touched on risk factors. I think it might be interesting if we could pull that apart a little further. So Maybe what is a risk factor? I mean, that's a pretty basic question, but maybe a good point to start. So what is a risk factor? Maybe how does it differ from an alpha factor?

Uh how does a risk factor differ from risk management in a general sense? Maybe if you just wanna speak about that a little bit? Sure. So I mean risk factors are probably one of the main techniques in modern risk management. Um because your risk is often measured or quantified by risk factors. And uh so what I mean by that is again, like any alpha factor

uh will produce portfolio weights, you know, and and and from portfolio weights over time you can say, what would my returns have been? You know, were I able to trade this portfolio? Again, like Jess said, in the real world there's tons of of considerations like uh transaction cost and slippage and liquidity and you have to figure out all this stuff when you're saying, does my alpha factor actually translate into into returns? Um and in fact

Honestly, that's really what back tests are for. And I so I think I might have mentioned in the last podcast that uh you know professional quants really don't spend any time backtesting or or not any, but don't spend that much time back testing. And and and the reason for that is a lot of what they do is hypothesis work where they're like just said, they're testing a hypothesis.

They're trying to see if this alpha factor spreads out returns, uh, you know, if it can predict future returns historically. Uh and then once they've gotten something that they're pretty sure does based on a lot of rigorous statistical analysis. Then the last step usually is okay, now let's see how it survives real market conditions. And that's usually what a back test is for. Back tests are usually done to simulate a variety of trading conditions because everybody's gonna have different.

prices they pay for trading and everything. Um but you know, that that's I think what what at least and I know what uh for a lot of what we do and we're like simulating whether a strategy will do well in this brokerage versus this brokerage. That's what we look at back tests for. But so the general idea is like ignoring all of those constraints, let's just say you were kind of

theoretically able to enter into this portfolio based on this alpha factor, well now you have a return stream, right? And the idea is that return stream represents the returns that are coming from that alpha factor. And so the alpha factor could be a momentum factor. So you could look at returns coming from momentum. It could be market cap.

Uh it could be some fundamental value factor where you're trying to estimate the value of a company based on a few of its, you know, um metrics like cash flow and revenue and everything. Uh so but the idea is once you have that return on the portfolio, uh when the portfolio's depending on the alpha factor, uh well then you kind of know like Had I been trading that alpha factor, I would have gotten these returns. So if your strategy

And there's a variety of ways to do this. Some people use covariance techniques, some people use uh multiple regression models. I think multiple regression models are the most common. Uh and in the Quantopian lecture series we actually cover all this. We actually have

uh lectures that touch on multiple regression. We have lectures that touch on spearman rank correlation, which is what Jess was talking about earlier. We have lectures that touch on uh like risk factors. So it's all there if you want to go check it out. Basically the idea is that let's say that, you know

Let's say that there is the the uh erin factor and I knew what your portfolio looked like every day and I knew what returns were coming out of it, right? And I was trying to sell you a strategy And you know, I was saying, Yo, uh Aaron, you should invest in my strategy and you looked at the, you know, uh correlation between My strategy and your strategy, your you know, your returns, and you found that there's actually a 100% correlation.

Would that be something that you would be interested in investing in? Probably not, because there's nothing new. It's just exactly the same as what you're doing, and you could just port more money into your strategy and not pay any transa you know any management. So the same thing applies for risk factors. So let's say that now we have, you know, the the the Aaron factor and the Erin risk factor, which is the returns on your portfolio. Well

When I'm evaluating my strategy, let's say that your portfolio is a very well-known portfolio. I'm going to look for um whether or not my new algorithm or my new strategy has correlation. in its return stream with your return stream. Because if it does, really what that's saying is it seems like some of what's going on in my strategy is really just this already known thing. And so, for example, uh one of the most common risk factors.

is the market, the broad market in the US, the S P five hundred, you know, traded as the spy. Uh and and it would be you would also probably look at, you know, the Australian broad market in Australia. Uh and so if you're looking at the the broad market and you're looking at the correlation between your returns and the broad market returns, if there's a high dependency, then it means that

Your strategy can basically be expressed by putting most of the money into the broad market and then do something else with the rest of the money. Does that make sense? Mm-hmm.

So the general idea is If you have a high dependency with the broad market, not only is it really not that different from investing in the broad market and then, you know, doing something else with a l with the rest of the capital, but also it means if the broad market goes off a cliff, you know, there's a strong risk that you're also gonna go off a cliff.

So in general, quants want to make their new uh strategies as independent of of well known factors as possible, both for the reason that it's more attractive to investors. um because they're not just redoing something that's already known, but also the fewer things you're dependent on, the fewer sources of risk you have in your returns. If I'm dependent on a lot of different things, only one of those things has to go off a cliff for me to go off a cliff. So I just thought that the

You know, the the the lower my dependence is on these risk factors, the better generally a return stream is considered to be. Okay. So let's let's start uh put this all together. Now, let's walk through an example. So I know Uh you guys there at Quantopian uh put together an example about uh female CEOs as a potential strategy.

Would you like to maybe walk us through that and then we can kind of pull it apart and speak about what part of that is the alpha factor, where's the risk factor, um, and that sort of thing. But maybe just walk us through that example to begin with and and we'll take it from there. Yeah, sure. So the the female CEOs project was a really good piece of work by um Karen Rubin, who's our uh VP of product here. Um and you know

She is an exceptional product manager in in that, you know, she's actually really willing to get her hands very dirty with the the the the stuff that she's building. And, you know, as part of that, she actually put together a full piece of of of quant research, which is I I thought was pretty impressive, uh, you know, given that a lot of product managers do not Do not take nearly as many steps into their own product. What she did is she came up with a hypothesis. And her hypothesis

Uh and this is actually backed by a a decent amount of finance and uh economic research in the last, I'd say, five to ten years, maybe more. But her hypothesis was that Investing in companies uh in the broad market. So uh investing in companies Uh and I think she used as her universe I think the Fortune one thousand maybe. I have to I'd have to go back and check it again. We'll have we'll have links to this. You can go check it on your own at at Quantopian.com slash

chat with traders um and you can see all the work. But uh the idea is she started with the universe and then she said within that universe at any given time I want to be invested in companies with female CEO. And so her hypothesis is that Being invested only in companies with female CEOs will yield better returns than being invested in the entire universe equally.

That's so that's that's the hypothesis. And so that's where a quant would start, right? They'd say, here's my here's my new hypothesis, here's what I think is going on. And it usually comes from some interesting insight or broad economic understanding. And and again, what we talked about in the previous episode was really like ways to avoid overfitting your models and and and trying to avoid bias. And you know, this is

This is an example of, I think, a good way to do things rather than trying a ton of different things and saying, oh, this thing happens to work. I'm not sure really why. But so here you start from like a really um solid hypothesis, which is that companies that are currently being run by women uh will will outperform uh you know the broad market in general.

Uh and maybe the reasons for that could be many. Uh I think that there's some arguments I don't know, I honestly don't know and I don't endorse any of these arguments, but there's some arguments that Um because there's sexism in the indo in all industries. uh, you know, the bar for women to meet to get promoted is higher. And so um they have to work harder and be better to get CEO positions as as opposed to comparable men. Um and and so therefore, you know.

Uh companies being run by women are just going to be run better. And then there's also actually a lot of work which has recently shown that. Kind of social environments of all men tend to make kind of more irrational and riskier decisions. And so, you know, obviously you want.

companies to make, you know, kind of low risk, rational decisions. And so uh if your board is is you know has some women in it, that might be a a a a really good sign that the company is going to make less risky decisions. And then finally, you know, maybe the theory is that Just the fact that a company is

is is in you know going to elect a a female CEO is a good sign that that company is is just, you know, it's it's free of more bias. You know, it's a more progressive company. They're they're smarter in how they work. So There's lots of different reasons that you could maybe hypothesize that that this effect would be true. But the general idea is that at the end of the day, you've got a hypothesis.

Really, what that is is an alpha factor, right? Because your alpha factor is now a score of one for companies that currently are being run by a female, and zero for companies that are currently being run by a male. And that's that's your alpha factor. You're saying wait, you know, wait my sort companies based on female versus non-female. And so in this case, all the female ones go to the top and all the male ones go to the bottom. and uh you know see does that spread out returns at all?

Um, you know, do the returns on the female companies at any given point in time exceed the returns on the male companies at any given point in time? And so the way that uh Karen tested that is she actually ran a simulation of saying at every point in time, just Be invested in uh the companies that are being run by by women. And she you she found that she did actually significantly beat. uh the the broad market as a benchmark. Um and since then she's actually done gone back

Um and again we'll we'll link to all this. Uh she's gone back and you know done this kind of redone this study and made many improvements and picked better benchmarks. And and each time, uh basically her strategy has still managed to to outperform whatever whatever benchmark she's chosen. So it seems like there is something there. There is something in at least in the 2003 to current time frame, it seems that there is still an effect.

Um, in which being invested in companies run by women does, you know, uh exceed the performance. Of being invested in companies that are run by men. And the thing that I love about this Alpha Factor is that it's also something where you could feasibly investment. Invest in it yourself. Uh and it's not, I mean, obviously like I can't really in make any statements of informed risk, but just from an intuitive perspective, you know, as long as you

or invested in a f in in a few different companies run by women, um, you know, that's n you're not really doing anything too crazy or exotic with that, right? You're just kind of that's a method of of picking stock. Absolutely. So in this example you highlighted where the alpha factor lies. What about the risk factor? Can you talk about where that kind of fits into this example?

Sure. So now in this case we have this return stream. We have the returns of the strategy of let let's call it the The The the female risk factor, right? We have the returns of a portfolio of holding just female-run uh companies over time. And so now what we could do is as we're developing other factors, other strategies, we could look at those. uh factors risk exposure to the returns on uh the the female portfolio.

So uh that's how you would then transform this into a risk factor because like, okay, well you might say, well, maybe in five years this is just well known and everybody trades on it. So I don't wanna be dependent. I don't wanna have any correlation with this factor. um you know, then you can start using it in the other direction as a risk factor rather than than trying to trade in it yourself. Uh and then

We can also go the other direction and we can say, okay, we have this return stream of uh female CEOs, and what are its risk exposures? So we could take like a set of You know, any number of known risk factors. And then we could look for dependencies between our female portfolio and those known risk factors. And we could say,

Oh, interesting. It seems like companies that are run by women have exposure to these other known risk factors. Or maybe it doesn't have exposure to these other known risk factors, and this is just a completely new thing that we've discovered. So that's kind of the approach that you would take to either turn it into a risk factor or use risk factors to try to evaluate what is this new thing that you've just constructed.

Excellent. Okay. Well yeah, I mean this is definitely a really good example and I've actually watched uh that video. Was it was it Karen uh who gave the talk about this uh this particular um strategy? I I think there's a few different videos as well because she's given the

different time seasons also actually talk to a few different reporters about it because as you can imagine this is a pretty hot topic issue. Yeah, so I mean definitely um I'd encourage anyone who's listening to this right now to um check out those links which we'll put at quantopin dot com forward slash

Chat with traders. Um, Delaney, we're pushing on over an hour now. Is I mean, is there anything else you'd like to add on this subject? Um, I think I think at the at the end of the day we we talked about a lot of really conceptual stuff today. Um and I have found that a lot of this stuff is Hard to get unless you actually kind of

either muck with it a bit or get your hands slightly dirty or, you know, actually try to see it in practice. So all I'm gonna say is again, just like, I strongly recommend that if people have kind of listened to it and maybe they're like, I don't quite get that part. Or um that part made sense, but I don't get how you'd use it. Uh I just I recommend going to again that link you mentioned, quantopian.com slash chat with traders.

And and we'll try to link the relevant lectures examples. And I think just seeing this stuff visually. it can be a lot more helpful. So like for instance, the way I would probably listen to this podcast if I were trying to learn this stuff is maybe um, you know, listen to it and then uh go back and actually try to look at some of these examples. uh that that we posted, but maybe also listening through it again, like pausing it whenever you don't understand something and actually trying to

to to to read through and and look at the plots and just see physically like, oh, okay, I get it. You're you're you're you're sorting stocks this way and then that mean your portfolio weights are this, et cetera. Absolutely. Yeah. I mean, if you don't get it, just look it up and uh dig deeper for sure. So Uh next episode, episode three, what are some of the subjects we're gonna be covering uh then?

Sure. So I mean I like I said, I think Jess Jess kind of previewed it um in one of the answers she gave. And uh the general theme is, okay, well, you know, now you have some alpha factors. what do you do with them? Because like Jess said, individual alpha factors I think will tend not to be very good. So uh one of the things that that quants just accept is that

Individual pricing models or individual trading signals tend not to work very well. They don't tend to have much edge over random. And that edge can also go in and out depending on market conditions. Or factors can be predictive in one sector but not another sector, for example. So the idea is you want generally to combine many different individual alpha factors.

Into one prediction, which is actually quite good. And so what we're going to discuss is Um all sorts of ways that you could look at, you know. basically kind of combining them together. I'll try to go back into the original kind of stats theory behind this and describe how this is not, you know, not just a thing that you done in finance. It's really just done in every single predictive modeling field there is.

And uh we'll actually uh try to talk about uh you know, some interesting techniques you can use because Um, this is actually a pretty interesting application of, you know, and I think I've mentioned that I'm kind of dubious of machine learning and finance.

because a lot of people use it in ways that that won't necessarily work in the long run, but there's are some really powerful ways to apply it. And this is an interesting um application of machine learning, uh, which is maybe using machine learning to try to at any given point in time select the best factors to be training on then.

uh as an example. So those are some of the things we'll we'll try to discuss in the next podcast. Awesome, man. Sounds like a good episode. Let's leave it at that for now. Delaney, thank you very much for doing this, man. I appreciate you taking the time and um we'll speak again very soon. All right, thank you, appreciate it.