Michael Krause - Evolving Long/Short Equity (S3E2)

Hey everyone, Corey here. Thanks for tuning in to another episode of Flirting with Models. If you're enjoying the show, I'd greatly appreciate it if you'd take a moment to rate, review, and most importantly, share with a friend. Word of mouth is how this podcast grows. And if you'd like to learn more about Newfound's platform of return stacked mutual funds, ETFs, and model portfolios, head over to ReturnStacks.com. Now on with the show. All right, three, two, one. Let's go.

Hello and welcome everyone. I'm Corey Hofstein and this is Flirting with Models, the podcast that pulls back the curtain to discover the human factor behind the quantitative strategy. Corey Hofstein is the co-founder and chief investment officer of Newfound Research. Due to industry regulations, he will not discuss any of Newfound Research's funds on this podcast. All opinions expressed by podcast participants are solely their own opinion.

This podcast is for informational purposes only and should not be able to do that. Basis for investment decisions. Clients of Newfound Research may maintain positions and securities discussed in this podcast. For more information, visit think Newfound. In this episode, I am joined by Michael Krauss, co-founder of Counterpoint Asset Management.

Our conversation covers two major topics. In the first half, we discuss some of the nuances of high yield bond timing and the subtleties of strategy construction. In the second half, we discuss long-short equity strategies. For listeners more interested in the technical, this is where the meat and potatoes of the conversation lie. We discuss Michael's evolution from regression to machine learning techniques.

the unintended consequences of accidental exposures, and managing risk through optimization while managing the risk of optimization. I hope you enjoy my conversation with Michael Krauss. Michael, thank you for joining me today. Excited to have you here on this episode. You have a really interesting background that I want to start with. One that's pretty unusual for the world of finance or just unusual in general. You actually started

an internet service provider at the age of fourteen, which you then went on to sell when you were nineteen. That is an entrepreneurial story I have to hear, regardless of what this podcast is gonna be about. But can you just take us back and tell me about what it was like starting a business? Back when you were fourteen.

Yeah, so I was a very enterprising young kid, I guess. Looking back, I don't think I give myself any credit. Maybe I shouldn't give myself any credit right now. That's being uh not humble enough, but long story. to a slightly shorter one. I at the time was an eager pre internet user. I call it pre internet because it's not what we know of it today.

We had a local internet access uh provider that we could get on. We had the Cleveland Freenet. The internet access provider that was available at the time was a Unix shell access. And what you could do on the internet at the time was pretty rudimentary. You could go do gophering. You could look at kind of the pre-web. This is 1994. You could do some basic internet email. You could do some FTP.

So it was a very different time. Usenet was a big thing, which nowadays we have message forums and and all that sort. So getting on the internet was an exciting end of war. There it wasn't even really much of a large community. Really, it was largely an academic thing and just starting to catch on with services like America Online. So I paint that picture.

At the time, I uh ran a BBS, a bulletin board service. I was just a 13, 14-year-old, nerdy kid. This was my way to meet people. And I basically had about 15 users within my small local bulletin board system in Cleveland, Ohio that were really kind of captive. They were excited about what I was doing. This was called the Exchange BBS back then.

It was a Mac BBS actually. I was a Apple person back then. I've turned my back since, but I do have an iPhone now and I do cynically wince at people who are who are Mac users now. for overpaying for their hardware. But aside from the point that that's a tangent in itself, I basically had about 15 users that were willing to pay ahead of time for service.

I went up to them and I basically said, Can you pay three months ahead of time and a and a startup fee? I got ninety dollars each from these fifteen people. I went to my brother. who is about 11 years older than me. He just graduated college. And actually he got into the mutual fund business just prior to this. So he was about 25. And I showed him, hey, look, I have people who are interested. They want to see my BBS.

become something more substantial, you know, be able to be a way to get on the internet. So it showed my brother Dan enough and he invested a few thousand dollars in what amounted to several computers, enough equipment to set up a dial-up internet service provider. how it worked was imagine two or three small homebrew Unix servers. We were running BSDI Unix at the time. We had Four dial up modems.

So that was uh downstream and then upstream our internet connection was a one single twenty-eight point eight K modem. That was how we could serve basically all of these people, all fifteen people and maybe a little extra demand. You know, like anything operation-wise, you don't have full utilization. So, you know, you you could make four modems stretch to, I don't remember the multiples, let's just say 40, 50 users.

on average without any one user generally getting a busy signal even at peak times. It's funny when I later went to do my MBA, years later, you know, your the introduction to operation research and you talk about queuing and utilization and bottlenecks and all this sort of stuff. And I realized I was intuitively dealing with a lot of these problems without any training. Anyway, long story short, I went to my brother. He funded the thing.

I basically taught myself how to be a systems administrator. I wouldn't characterize myself as a programmer per se. Now I have some programming and database skills. that I use day to day. But back then really I could kind of hack anything together. And that's kind of how I survive with tech. So I basically taught myself how to be a system admin. I

got the O'Reilly books, which I think we talked about. You were getting O'Reilly books too. I'm so impressed with you. And basically each book would be some sub-area back in the day. So TCP IP talking about the protocol stack and how everything was structured. how to administer a DNS bind server. You know, this is fundamentally a DNS server. How to administer an SMTP server. You know, all these protocols. So every single aspect I had basically teach myself.

And that was enough to go on. We started up and fast forward five years, cutting the story really short, we had about 8,000 customers. We had probably about thirty corporate clients who had various dedicated connections and and upstream where I told you it was originally a twenty eight point eight K connection, now we had multiple T threes coming in.

into the premises. And it grew from instead of the upstairs in my parents' house, which was basically my bedroom, it went to some actual office space. We had about 20 employees. It was quite an experience. I would say that there's a risk that this entire podcast could just become about you age fourteen to nineteen and we never talk about anything finance I actually think that's a lot more interesting, to tell you the truth.

I mean when you and I first started chatting, I think we were teed up to talk about some optimization stuff on factor portfolios and we started talking about backgrounds and you started telling me about this and it was bringing me back to the days of

mid nineties, I remember starting to learn about a lot of the protocols and people on message ports would say, Oh, just go read the RFCs, right? It's like, Yeah, why don't you go read the original document? And you could back then. You know, w I think we were talking about how In the early nineties, mid nineties, it still felt a lot more like cars seemed to be in the sixties, where they were mechanical and you could understand it.

And as things have gone on, they become more and more electronic and there's more interfaces layered on top and you get further and further away from that understanding. And if you ask me to open the hood of my car at this point, it's like I don't want to touch anything because it's so electronic. I think what happened fundamentally is, and this is a general statement about technological progress we've made, is that everything has gotten much more high level, right? So now we abstract.

Now we to do even coding, right? You don't necessarily start with assembly language and these fundamental layers. You start on high abstract levels, object-oriented layers. And it is kind of oddly dissatisfying because you don't kind of get to get your hands dirty with what really makes things go. You're detached from it. Yeah.

There's certainly a bit of nostalgia to it, but not having to remember to free memory anymore is I'll wash my hands of that. But let's fast forward a bit because you did sell that business. And I know you made some major changes. You were thinking about being a professional pianist for a while, I believe, but you did eventually find your way towards the field of finance. So can you catch us up to speed a little bit how you ultimately made that transition?

The key theme here is I've not actually gotten away from my musical roots ever. You know, I studied piano. I play classical piano. If you see on my Twitter profile, I say I'm an amateur classical pianist and It's a big part of my life. After I sold the business, I went that direction, uh, went to music school, found some great teachers and kind of pursued it to its practical limits in terms of discovering, well, is this something I can actually do feasibly?

And I came to a conclusion, unfortunately, that really it wasn't what I was really cut out for in the long run, just as a sole thing, but it's something that's stayed very close to me. The firm that I created, Counterpoint, is actually named in homage of Bach. Bach was the master of Counterpoint.

Right. Most people uh think it's just a cool name for a business. It might ring a bell if you're a contrarian investor, right? Counterpoint, you know, a sophisticated investor. Maybe it resonates with people that way, but but really what it is is that.

What I did basically, I kind of f fell in love with finance accidentally. After I sold the business, I had a natural gas partnership. It was a private investment I made. It was at the end of So the end of the nineties, right when the gas markets were in a relatively depressed state. So what turned into was originally kind of a tax optimized investment, not particularly great on an ROI basis. It bred a fascination for all things economics and markets.

So in my early twenties, like a lot of people, I took up trading and the idea of researching to invest. And amongst my earliest trades, and it was probably the worst thing that could happen for that period, I had a great trade. I was watching the natural gas spreads. This is uh mid two thousand five, two thousand six. So this is calendar spreads, difference between contract month valuations, and saw some very unusual things relative to historical data. I just made some intuitive bets.

That paid off and they happen to be opposite of what the Amaranth hedge fund was doing. I don't know if that rings a bell. Does that? Oh yeah, absolutely. Absolutely. Okay, so I was just I was really an outsider in a market that I really had no idea about. Maybe I lucked into a decent trade, but I made some money and got the idea that, hey, this is something I really want to pursue. And not just for the sake of trying to make money, but because it genuinely fascinates me.

me. So that seed was sown and uh fast forward a few years in my twenties, I like I said, I studied music to not much practical end. I decided eventually late twenties, go back to school. I got my MBA, did my CFA credential.

And really set upon the goal to get myself into a position where I could be a fund manager someday. So really pursuing that vision I had. What happened later along, right, is finished my MBA, I actually got a job at the local utility, San Diego Gas and Electric, helping in the risk management. So here I was starting to utilize my tech skills, not just my fascination with the markets, but you know, I was doing things like programming VAR models from the ground up to replace some in-house tools.

And looking at all facets of energy risk. A really fascinating project was modeling And this was actually an internship I had with a connection I made during my MBA, but this is how I was involved with the utility. We're modeling uh gas storage, right? So fundamentally looking at the futures curve in the natural gas markets and

coming up with a fundamental model for what storage is worth. What's fascinating about that, right, is you can go through simulation processes, optimization processes to get to these solutions. In the end, the shape of the future curve, if you see it's like a zigzag in natural gas, that's a function of those very real storage fundamentals. The gas has to come in and leave the shape of that curve as a function of the storage dynamics in the market.

So I've fun again, I've kind of always tilted towards these kind of wonkish interests. So in doing the MBA, I'm time shifting back a little. I actually met my professor who we later collaborated with and he joined Counterpoint later as a co-portfolio manager. This is Joey Engelberg. He is an expert in behavioral finance. He's

very well published, especially in the meta-anomaly literature. So looking at questions, for example, when do pricing of many factor anomalies occur, you'd find answers like, oh, it's going to be in news and earnings related period. Right. So I met him actually doing my MBA. We later collaborated. And the funny thing is, when I joined the utility, I actually moonlighted and created an investment advisor called Counterpoint Asset Manager.

that was intended to really incubate some single factor equity anomaly strategy. So that's really how I started off in the business. And then fast forward, now we've created counterpoint mutual funds. I collaborated once again with my brother, who was my business partner in the internet service provider. So that gets us to twenty fourteen.

So I know, and we're going to talk a little bit about your factor work both in the single-factor space and the multi-factor space, but one of the first strategies that you launched at counterpoint and in fact still manage today was a tactical high yield bond rotation strategy. So I want to start there and talk about, well, what's the premise behind this strategy and really why would we expect this type of strategy to work?

So what I like about this strategy, and this really goes in stark contrast to equity factor or equity anomaly type strategies. You have simplicity on one hand. And I think that simplicity lends itself to a simple behavioral explanation, which in turn I think has really borne out the result, especially in a time right where anything quant has been

generally very challenging. If you're in the last decade, I'd have to say as a generalization, well, how about the last five years in particular, not really the last decade, but the last five years. If you've been doing any strategy that goes by long-term research and you allocate a portfolio effectively to aligned with that research, you've not been rewarded. The best funds out there are I would call anti-factor funds.

Right. So going back to the tactical high yield strategy, it's kind of unusual because during this time, it's actually one quantitative strategy that's done pretty well. through that environment. So j just a high level view what it is. The tactical high yield strategy is effectively trend following high yield index. And when it's above a moving average or some kind of trend indicator, you're invested.

And when you fall below that indicator or some kind of filter related to that indicator, you get out. And you're either in cash or you're in treasuries. These are all variations on the theme of the same conceptual idea. The thing about tactical high yield is that it's, you know, I'd say to generalize why would you invest in such a strategy and how would it compare to other trend strategies? It's basically really great way to get risk on exposure while reducing drawdown.

In the end, I can show you back tests that will outperform over the long run. But I think the realistic expectation is to, you know, basically get the same return level in the long run, but without the drawdown characteristics. And avoiding the drawdown, you know, helps you stay invested and sticking to a strategy as long as you don't look at non-

tactical market as your reference point and always compare yourself every minute to it. You have to look at the strategy as a separate investment that you're willing to be committing

So this is an area I know there are a number of firms that have offered over time strategies like this. And there seem to be conceptually very high level, very similar to your point, they're applying some sort of trend following technique on some sort of high yield index. They're moving either to cash or some sort of treasury position to preserve capital when trends look negative. But there are very subtle nuances in the way that they implement this strategy that seem to lead to some

significant disparity in the performance. For example, are you trend following on a high yield ETF versus looking at a mutual fund proxy versus looking at a basket of mutual Are you implementing with an ETF versus implementing with a selection of high yield fund managers?

And all those are gonna obviously create trade offs. The h the ETF you can trade every single day, the mutual fund not as easily. I wanted to get your thoughts on those nuances. Which ones do you think are really important for the effectiveness of this strategy? So

This is a loaded question. We have products that are in both areas, right? We have a fund which is the our mutual fund product that addresses primarily trading mutual fund vehicles, which on the surface of it sounds completely unsophisticated and fee layering and an absurd thing to invest in.

And then on the other hand, we have an ETF product that we launched, and that product only trades ETFs, you know, by have regulatory constraint and design constraint in the end. So to answer it, right, what you trade and how you execute. creates all of the disparity in results in these strategies.

But before you zoom in, right, if you go back from a very 30,000 foot view, these strategies are all pretty correlated as long as the manager stays systematic and doesn't start overriding his signal with discretionary calls. So now getting into the real thick of it, when you trade high yield, you're ideally, right, if you think from a model point of view, I want to get the pure asset class.

Right. And the best way to do it, you can't go out and just buy the high yield index. This is one of the problems that presents itself. The index doesn't exist. As with all bond indexes, everything is a replication. Some better than others. They can be very close to each other. So you have a continuum of available securities. You know, this is a funny thing because everyone's fixated on low fees nowadays. And especially in fixed income, I think they miss.

some of the subtlety, uh which are really uh the subtlety is becomes something not so subtle. It's actually a key factor aside from asset class. So when you buy mutual funds, for example, in high yield. Generally, as a group, they capture the broad high yield market better than anything. Right. So they get the garbage, the lowest credit quality. There's gonna be some exposure to you, don't get just a subset, which is a very liquid subset that you're biased to.

you get perhaps an increase in yield, of course, because you have exposure to lower credit quality in those indexes. And there's another perk when you enter a mutual fund, right? The pricing services to give a stable price, they price They try to price every security in the fund. Most of those securities don't trade on a daily basis. So there's a bit of stale pricing effect. It turns out to stabilize and lower the volatility of these funds.

versus what they may be if you had to dynamically trade every single bond within those portfolios every time you traded. So what that gets you is the ability to enter and exit. closer to midpoint of more pure representation of the asset class. So Trading neutral funds gives you this kind of pure exposure to high yield index, more and more idealized exposure.

Now the ETFs are the other side of the continuum, right? Let's start with HYG or J and K, any of these ETFs. So they have a liquid subset of high yields. that is not representative of the entire basket. It's one slice of the market. which by product design is a kind of a requirement, right? So the one of the reasons HYG became so successful is because the market makers, the APs

they could trade in and out of the underlying bonds and do the arbitrage against the actual ETF units. Right. So if one is mispriced versus the other, you have something to trade against and it it enables the market to have more liquidity. That liquidity, beget more liquidity, and suddenly it took a life of its own where the ETF trades at a bid ask that doesn't actually represent the real underlying bond liquidity.

So for example, if I wanted to go trade a representative basket in the HYG right now. Typically, I'd need to pay about 30, 40 basis points to get in and out of that real basket of bonds. So if I went to a dealer on the market, I'd probably have to pay that spread effectively. Whereas the participants at the ETF don't do it. They just pay whatever the bid ask is and they bear that cost.

by the risk and movement of the fund around its NAV, its idealized uh value that the intraday nav that you see. So one of the this is an area I've studied a bit at the request of some of my clients. And one of the things I always found fascinating about the high yield bond timing type strategy was trying to figure out where exactly it was able to harvest. It's excess return.

So for example, you you mentioned specifically the goal of trying to avoid those significant and prolonged drawdowns. I would sort of put that in the bucket of we're trying to avoid those periods of expanding credit sprints. But there's also the opportunity to reinvest at higher and contracting credit spreads, positive repricing that might be purely fundamental, economic, or sentiment driven.

And then there's even the opportunity to potentially harvest some excess carry in high yield versus sort of a corresponding Treasury benchmark during calmer market environments. So when you think full cycle and you look at a product like this, how would you think about where the majority of the access return comes from? Is it from the capital preservation? Is it from that repricing shock? Or is it from that, hey, there's some good cumulative bips we can earn along the way?

So this is a great question. I have some data I've prepared that really points to an answer actually, fundamentally where you're getting return from a strategy like this. So, you know, just to give you an idea, I took a very generalized version of the strategy. So this is not how we implement it. The details of the exact signal aren't really important when we're talking in broad stroke.

So just to give you an idea, if you ran the strategy, it's a very simple one where you took the 200-day moving average and you looked at the Morningstar high yield bond category. So this is a category that includes fees of exposure. You're able to, you know, you could generalize if I buy a few big funds within this category, in and out, I could pretty much track this category.

So if you take a two hundred day moving average of that category and what you do is you alternate between it and three to five year treasuries when the switch being above or below the two hundred day moving average. The general return difference over the long run with a simple back test like that, and this is 1990 to present, is you get about 10% return versus the index itself, where you get about 6.6%. So there's

a little more than three percent of outperformance coming from that. Now the thing to keep in mind, right, and this is why a question that I think has a quantitative answer In truth, we're going to have to kind of default to, you know, the why you invest in it to more qualitative reasons. And I'll get into that. So the neat thing about this, right, is

Again, why you have the strategy is again to avoid those large market routes and the damage that occurs. So I took just on that simple version of the strategy, I looked at the major signals. And if we look in the last decade, we may be in the middle of one of them. I would think we are intuitively. The signal is risk-off at the moment. But in the last most recent period, we've had two major signals. It's the financial crisis.

Which started actually for the high yield market as it turned down after the Bear Stearns route. Was that March of 2008? It turned down and started to see signals like that around June of 2008. the main re-entry into that signal would have been around April of two thousand nine. So you can see how this would have missed most of the downturn. For that period, right? By just being out, you missed a maximum drawdown of this index of about 27%.

But if you look at the missed negative return from exit to re-entry, so this means you're not gonna be able to catch the bottom. You're just gonna kind of add the idea of what you said, re-enter at a lower level. Your missed negative return on that was about 18.4%. That's very, very material in this long-term back test that I told you. There's another period, the 2015-2016 oil crisis.

So same thing, the index fell around 10% peak to trough of that kind of drawdown. But following that signal, that simple 200-day moving average, you would have missed about 3.7% negative total return. This includes dividends. T this is total return. I've just pointed out to you about twenty percent of missed downside over approximately a a decade period. This chalks up to a good portion.

of that 3% differential, that 3.5% differential in return. The other part is coming from, I would call it some noise around the signal as well. There's the part coming from the fact that when this version of the strategy I said you're in three to five year treasuries. So this is another kind of kind of neat conclusion, you know, when we're talking about looking at trend following.

Our findings are at least that it's not robust to every asset class, despite technical people loving the idea of trend following everything with some kind of moving average or crossover or or whatever you can invent. In fact, high quality fixed income is definitely not robust to these longer-term moving average signals. It's really hit or miss.

I think you're better off, you know, just putting your finger in the air and gambling. You well, you may be, maybe not. It depends how bad you are behaviorally and how how how how lacking in discipline you are. But the point here is that you're picking up a little bit from the avoidance of the downturn in high yield, but likewise from the updraft in having exposure to high quality duration. And in those periods, so this is what's interesting is if I trend follow

Treasuries directly, the result isn't so great. But if I trend follow high yield and say, hey, when I'm out of high yield, I'm going to be in treasuries, you actually do tend to have a somewhat robust signal. to picking up a little extra tailwind in treasuries. So basically the risk off for one asset class is a good signal for risk on in treasuries. So that complements. And that's intuitive, right?

The economy is crashing. You want something that does well with the Fed cutting rates and long-term growth prospects falling, and treasuries are your natural security.

So in contrast to this high yield bond timing strategy, I know you also run a couple of multi factor long short equity strategies. which I want to dive into because I know this is an area of thinking that you've spent a lot of time on and also have evolved your thinking on over the last half a decade. So I want to start with basics, which are

Let's just define what factors are you looking at? What are the sort of characteristics you're looking at? And when you talk multi-factor, are you talking bottom-up composite multi-factor? Are you or integrated multi-factor? Are you talking sort of a top-down sleeve-based approach? Okay, so what we're talking about, there's a many questions in that question. The place to start, I think, is, you know, what sort of factors are we looking at?

It's what everyone else is looking at to some degree, right? So you have the valuation related, the momentum related. to to give a little detail, you know, you could start we actually don't look at price to book. It's not terribly robust across uh many market segments in this last 20 years, but there are others, price to sales, price to earnings. Other related factors, cash flow to enterprise value, this sort of valuation idea. You want to capture a sense of value.

In the momentum bucket, right? You have short-term momentum, so two to six month momentum. You have seven to twelve month momentum. So you're decomposing effectively twelve month momentum. You exclude that first month, right? Because there's evidence of rebound.

So if you have cross sec and by the way, I should say everything is cross sectional we're looking at. So you're taking the universe of stocks today and you're ranking them versus looking at time series momentum, which is for our audience just looking at one asset or one asset class and trying to trend follow it effectively.

In the momentum category, there's related indicators or, you know, we can call them factors or indicators anomalies. There's a paper on one which looks actually at a simple moving average crossover. something like the twenty-one day over the 200 day, right? And if you look at a result like that and we put these through machine learning models so we can see the interactions and we see how they score and look like to each other and what looks like to proxy each other.

You know, you'll find that ones like that, which seem very technical on the surface, and again, you're cross sectionally ranking them, you know, they rhyme. They're very similar to you know, two to six month momentum is very, very similar to that. So The point is you can reinvent or recharacterize a lot of different fundamental effects and you know, they can have different names or different settings, but they're effectively telling you something similar.

So that's value and momentum. By the way, momentum can be, again, like I said, the return of the most recent month, whereas it signals not to buy the most recent winner or short the most recent loser, you'll get the opposite effect. You'll actually short the most recent winner and buy the most recent loser in that, right? So what we effectively do, well, before I go on, right, we also have sentiment categories. So looking at short interests, looking at

and revenue expectations, looking at analyst behavior. So analysts, for example, one thing they often do, and my co-manager Joey Engelberg, he wrote a paper on this, is that analysts are often way behind the curve. If they set very high price targets, Typically that's a predictor of negative returns. So high price targets relative to current price.

That's a a great predictor of falling short, right? Whereas often they're behind the curve and it's when they're setting a low price target and the stock just is pulling ahead of it. Those are the stocks actually that tend to do better over the long run. So Imagine all these different factors, you take them into one model, and then you create one composite score of them, and then you put a portfolio together.

So this composite approach of multi-factor is what we do rather than just taking a few separate factor portfolios of let's say momentum and value and then just combining them and you know netting them out. We don't go with that approach because we find There's better payoff over the long run to factoring the interactions between the factors, which these separate simple factor portfolios don't really do.

I know this is an approach that you've evolved significantly over the last several years. You started with More of a standard regression based approach. You now incorporate a lot of machine learning techniques into what you do. I was hoping you could take us back sort of to the start of the approach and talk us through some of the evolution and the thinking and really what drove these decisions. Right, so

In our process, we're a small firm as we started out. We have three people in our investment group. Not to be underestimated though, because a small group with the right data capabilities can do an awful lot, right? We launched our first fund, tactical equity fund, at the end of two thousand fifteen. And originally when we modeled it, we were assuming some operational constraints. First of all, we didn't have access to international names effectively, you know, with

having a small company, I mean a small fund. We had the problem to determine of, you know, how best to score, but really the toolbox we were working with at the time was a regression toolbox, right? And furthermore, about trying to get the most bang from your buck out of all the different factors. we kind of came to, and this is interesting actually, this goes to the end of the story. We came to a realization that the low volatility anomaly really dominated a lot of the long-term return.

from this exposure. So, you know, this is just out of our own testing, right? If you look at any single name data, it's really easy to come upon the conclusion that the most volatile names in the long run tend to do pretty horribly. Right. This is across almost any market out of sample, you'll see that. Staying short those names is another point of survival. another challenge. So what we came on when we launched the fund was we were trading in US only.

And effectively the fund did a blend, it does still does a blend of tactical but also multi-factor exposure. And uh one thing we did was We basically maintained a short book that was a small amount when we were risk on. So imagine something like a hundred percent long and then against a fifteen percent short book when risk on.

Right. And then when there would be a tactical switch, so imagine you get a signal to get out of the markets, like the ten month moving average on the S P, something like that. Then we would rotate the portfolio into a more market neutral targeted portfolio, but dominated by volatility filters. So we were effectively looking at 70% long, 35% short. Now we're getting into a discussion of construction and how to effectively run these strategies.

And the one thing to know about this, right, is we weren't optimizing on industry exposure. We weren't saying, well, if our short basket is full of pharmaceuticals and oil stocks, we're gonna do terribly much about it. I mean there was some discretionary trimming, but there wasn't anything terribly systematic about what we were doing. It was pretty much an ad hoc approach to try to manage risk.

So we're optimizing on beta and hoping that beta estimate is somewhat robust. I say hoping because we know estimating beta presents this whole set of problems. So in the end, what happened? We launched a fund and it was trial by fire immediately. There was a sell-off in early 2016. It was associated with that same oil market crash. And our short book was dominated by oil stock.

And I'll cut to the chase, when we were still market neutral, we were effectively risk off in the fund, we had an overweight exposure to short oil, effectively. And being a small fund, what's nice is we can have small cap exposure. There's a lot more inefficiencies in the small cap space. That's a whole nother area and unto itself. But the consequence of that naturally when you're not optimized is that you have accidental exposures that you're not hoping. Right.

In this case, imagine being short, it wasn't a large portion of the portfolio. Actually, when I look back, it might have been 10% of the short of the total portfolio exposure, maybe 8%, something like that was actually energy short. But no, actually, I correct myself. It was 4%. Now my memory's coming back. But those shorts on average for a four-day period double. Right. Imagine being short, short name. Well, we've just seen it.

If you're not managing risk in this environment, you've seen some incredible moves where you have sectors just going up a hundred percent in three or four days. The perils of long short investing lead you to this. So I think we saw about a 400 basis point move. over this period in our risk our market neutral portfolio effectively that it wasn't dollar neutral. We're still it was only thirty five percent short.

Again, 70% long, but it didn't matter because we weren't long enough oil stocks on the long portfolio to offset. So that long story gets you to a conclusion of, okay, we have to go to the drawing board and really see how can we really improve this so we can live with this in all states of the world. What that got to in the end was by mid-2017, uh we came upon a few things. Продолжение следует...

the way to run long short, the real right way, with a dollar neutrality, an industry, uh well, actually a sector neutrality, industry, sub industry, subsectors, that becomes a little tougher. But by running with sector neutrality, And also even optimizing sometimes on some other factors. For example, in our energy portfolio, even despite being uh sector neutral, we actually optimize against price movement of oil to have zero beta to oil price.

So that way we're not long a bunch of refiners and short all of the producers, because you can again have a disastrous result. Right. So again, you know, running a market neutral portfolio is about Setting up the optimizer to get what you're really after. It's the factor exposures. It's not the accidental industry exposures or industry momentum that really can wreak havoc

It can help you, but Murphy's Law, you launch a fund, you're only going to get the bad output of what happens. And and that certainly is how it feels more often than not. It's funny. I was just reading a paper earlier today that was talking about momentum based investing and using residual momentum, right? So you're

basically regressing out all these other factors. And one of the things they talked specifically about was regressing out a macro oil factor. Basically saying if you don't use that as a factor, normally you would use market beta and value and size and maybe some other style factors if you want, but they started adding in these macroeconomic factors, saying there's certain of these factors that if you're not careful, you can find yourself way offside.

In either your short or your long book. And the short book can get you in particular. So I like this idea of accidental exposures. I think that People I find, at least it's my hypothesis that portfolio managers don't tend to be wrong in what they know. It's they're wrong in what they don't realize they're betting on.

It's what you don't know what will hurt you. Right. Exactly. It's these accidental exposures that you don't figure out you have until far too late whether it's Specification risk, timing risk, style risks that you have. You didn't have as, you know, all of a sudden your momentum was very sensitive to value.

or whether it's your a particular sector. So I think these accidental exposures are really important. Let's talk about this optimization based approach because it's easy to say, okay, we use optimization, but optimization is a huge field and there's a lot you can optimize over. So maybe you can tell us a little bit about, okay, let's narrow it down, how do you think about the optimization problem? Right.

If you go out of the box with a straightforward Markowitz optimization, right, and you leave it relatively unconstrained, you give it the universe, let's say you have a thousand stocks in the universe, it's gonna give you a thousand weights, right? It's gonna you're gonna have

insignificant weights for a lot of the portfolio. And if you don't constrain, for example, position size, you'll end up with absurdly large positions. And there's a lot of different ways about trying to solve these problems, right? Some managers, some larger funds by necessity have to go with cap weighting schemes, right? So now you're effectively constraining the solution to kind of a function of market capitalization.

Smaller funds like ours, which were more nimble, we have targeted an equal weight approach, right? So the idea isn't to have every stock in the universe in the portfolio. It's really to get the best of the top decile, the best of the bottom decile, and you mix them together in a way that really solves for a lot of the other problems.

And with that, we can have smaller counts of securities. Now, I don't think having a concentrated portfolio and factor investing is what we're after here. So I'm not after having a thousand names. I'm I'm sorry, I'm not after having a 30 pure play uh long and 30 pure play short names. That's not the intention. But having an equal weight exposure really to kind of get a result well enables you to get a result.

That's more in line with all the research, effectively. Most research is done on equal weighted portfolios. It's not on cap-weighted portfolios. And you know, it's kind of funny when you talk about this, you're you're kind of jogging my memory to a lot of papers that I've seen, and it's something I don't see talked about. It's not in the FinTwit world, certainly. It's the fact that this equal weight

versus cap weight as it applies to neutral portfolios, long short portfolios, it's not a very discussed topic. People aren't talking about, well, if we look at the seminal momentum study or the value study, how are we weighting the portfolio? There's a lot of ways to skin the cat and almost every factor you can look at on multiple dimensions. You could look at it versus a cap weighted versus value weighted dimension. You might get different outcomes. Optimization is a really

Big area. Lots of academic papers written on it. And there's all sorts of ways in which you can optimize, all sorts of things you can optimize for, all sorts of ways you can design your constraints, your functional form and target that you're trying to optimize over. When you're looking at building the optimization problem.

What are the big muscle movements that you're trying to hit? If you were to sort of itemize like your big constraints that you're trying to focus on or the big targets, what are they? The guiding philosophy around choosing the fight you're gonna fight with the optimizer is around Trying to avoid unness or unintended exposures in the factor space, right? So those unintended bets, right, are concentration.

industry exposures that are unintended and as well then country or again larger sector asset class exposures. Now, what can you do? Right. There's so many dimensions that we optimize on. The Around concentration, right? We target equal weighted portfolios. So there's a discussion right there about how do you optimize for equal weight using a a mixed integer optimization rather than a straight quadratic approach.

Is gonna let you solve for that, right? That lets you address some of the concentration issues that you get because now you're capping your positions. on sector and industry, right? It's pretty straightforward. You can take a sector classification. And likewise constraints. So your net of a sector is zero or or you allow some wiggle room, right? If you have some other constraints that are or if you have that constraint binding and you choose, you know, you decide other things are more important.

Likewise. uh beta, your market exposure, right? So it's about unintended factors, getting rid of them. So as I said earlier, we do a little in the energy space with getting rid of oil exposure, oil price exposure within the energy sector. That's number one. But likewise uh trying to get rid of market exposure the proper way, right? So how do you which goes into not only how do you run the optimizer, how do you estimate beta.

And how do you do that across a lot of different assets, right? That it creates all sorts of challenges directly. There are really two inputs to most portfolio optimizations, the return side and then the covariance. I want to actually go back to the return side a little bit because we talked about some of the machine learning.

techniques that you've utilized, or at least on the surface, we mentioned that you use them. You talked about some of the factors and characteristics you look at, but we didn't really dive into at all the machine learning techniques that you've explored. I'd love to get some color on how you've tried to apply different machine learning approaches to the return side of the equation. Right, so you're gonna love this. You're gonna love this. The best result is an ensemble.

What we found, right, if we walk through you asked this question earlier, right, about evolving from regression approaches all the way to a machine learning approach. What they have in common, right, is if you start with a regression approach, in the paradigm of more advanced machine learning algorithms, you have regressor approaches, right? So they start to look alike, actually. You know, the idea is you're

taking a predictor and you're putting that through the model and then you get a score, some kind of prediction based on it. And often a regression approach is used within these other models, whether it's a gradient boosted tree or random forest. One of the things, right, is a big deal about, you know, in the machine learning space, it's how you prepare your data. You don't just give the model prices and hope you'll get something out of it.

Something we found when we were implementing originally is that a lot of research out there was put out by computer science type people, the people who developed a lot of these machine learning algorithms, and often very little experience in the finance. Right. So they how to prepare the variables and how to really give it something to work with, to give those models something to work with is

a big thing. It's a very big thing. And those papers which would try to do something, throw some prices through a neural network or through throw some technical patterns through a neural network, would just miss the boat entirely. They you're basically trying to predict a shape. with some kind of other input shape and you're leaving so much on the table. Maybe in a high frequency vein, you could pull something off there. I'm sure anything can be done.

Right. But you know, one thing we learned is early on when we had, you know, when we were set up to do regression approaches only, something we would do, we found, for example, that effects within different factors were nonlinear, right? You would have You would have I'm going to talk about the asset growth anomaly. This is basically where companies that do acquisitions, a lot of CapEx, those companies tend to underperform relative to companies that just

don't do much with their balance sheet uh in terms of growing it too quickly. So if you do a cross sectional rank and you look at those asset growers, the acquirers, all these companies that basically are often reckless with their shareholders, those companies tend to perform poorly in the long run. Now the opposite, right? The people who don't asset grow, well

They don't perform poorly, but they don't outperform either. So it's not a buy indicator, right? It's only a short indicator. You could say it's a buy indicator of what to avoid, but it's immeasurable at that point because you want to somehow get a pure exposure to the asset grower. So you need to short it naturally. The point of that, right, is that instead of having a linear regression model, which will often have a weak effect.

Imagine you break up the problem into deciles, or even you just say top and bottom decile only. I'm going to ignore everything in between, and you create a dummy variant. Right. That represents, oh, is this a top decile asset grower or a bottom decile asset grower? Imagine doing a three-factor regression with top, bottom decile asset grower, bottom decile asset grower, and then the market. You got to get rid of the market factor.

And what you're gonna get often, you might have a statistically insignificant effect for the bottom dust salt, the companies that don't grow assets. But you will have a strong loading to the high decile asset grower. What that informs, right? How does this translate? It's how you prepare your variables is everything, as it is with a regression, as it is with machine learning approaches.

And the neat thing about a machine learning approach though is now imagine you have all these different factors, especially tree-based approaches. They can look at how all these different factors interact with each other. You can achieve the same thing with a regression manually, where you can construct joint dummy factors. You know, you can say, hey, what happens when there's a high asset grower uh times a high volatility name?

And that's a joint dummy, and then you have a fourth factor you regress upon. That can tell you that, but here's the problem. How many different combinations are there? And at a certain point, how much are you trying to squeeze blood out of a turnip, right? So you have constraints, what a researcher can do.

Now you can put all those factors into a machine learning model like a gradient boosted tree. You run an output and it'll show you the interactions. You can get a ranked report showing, oh, this is how it works. So you know, when you think like I think AQR did a paper about momentum in Japan, the finding that

Japanese names don't exhibit much momentum versus other markets that do. Maybe there's a behavioral explanation about it, people who trade in those markets versus US markets. I don't know what the explanation is, but You plug that into a tree bottle, you see the interaction. You've basically hired a PhD to do research by brute force for you. Now the question is.

Is this a robust factor or just something that's a product of data mining? So again, now this goes back to the discussion we had earlier, right, about tactical high yield. We have this strategy that seems robust over time. Why the heck does it work? And I think it's important that their behavioral explanation, right? The idea with trend in general, right, is that people take time to absorb the truth. Things don't play out instantly in the economy. And this benefits trend followers now.

the same goes in picking factors. I think you have to have a economic explanation, which is why, right? This is I think your biggest argument, why to be very, very careful, I think. and almost dogmatic about asking these questions about value. Has anything changed? Should we ditch value suddenly? Have people's behavior impulses changed?

No, I think people overthink it. Basically the reality is we're in a time, we're in a wave of enthusiasm about stocks that has rhymes with history, and we have a history of making those errors c collectively. But that's what factor investing is all about. It's about taking a disciplined side against that, you know, investing in behaviorally consistent characteristics about the markets that will sustain in the long run.

So I wanna talk about the second really important input as well, the covariance matrix, which I know you've spent a lot of time working with. For listeners who are maybe a little less familiar with the whole optimization process, one of the problems with the covariance matrix in traditional mean variance optimization. is what that naive mean variance optimization is going to do is basically break down all of your securities into these independent principal portfolios.

And each of those portfolios is going to be ranked from the one with the most variants to the one with the least. And traditional mean variance optimization is then going to take all these independent portfolios and lever them up so they all introduce the same amount of risk into the portfolio, which ultimately means that the independent portfolio with the smallest variance. gets levered up the most.

And when you look at it from a mathematical perspective, what you find is that that is most likely to be the part of the covariance matrix that is just complete and random noise. So you're ultimately jacking up your noise and reducing your signal. And so there's all sorts of techniques of dealing with this.

shrinkage, eigenvalue clipping. And Michael, I know this is an area you've spent a lot of time on. So I wanted to get your thoughts as to what's been effective for you, what hasn't, maybe leave a little bit of breadcrumb for our listeners who are exploring this area.

This is a fun area and I I do love tinkering with the optimizer and often struggling with some of these problems. One of the things we've found, right, is that you can test a lot of ideas, right? And you know, from a high level, think scientific method, you have kind of a controlled setup and you try something, you control it, you test it against your baseline, and you have a way to kind of measure what is better than what I had before, right?

So something we can do very easily, right? When we're trying to create our covariance matrix, which by the way, for our listeners, covariance is effectively a combination of volatility characteristics. and correlation characteristics and interasset correlation characteristics of the portfolio you're trying to construct.

Right. When you look at the problem, right, you're trying to measure essentially what can be improved upon. So in correlation, I can compare my model result of correlation. So talking about that one subcomponent. to an actual after the fact realized correlation. So I can take Apple and Google and run a sample correlation, you know, on their history. And we know that history is the actual, right?

So now we can go and test prior to knowing that actual history and create different models, right? So you can have a model of correlation. Well, let's go back to covariance, but essentially it's all connected. You can have a model of covariance. That is simply by prior history, you can have a model of covariance that's based on taking a bunch of factors that all the members of the covariance matrix are sensitive to.

And again, constructing estimates of covariance on those factors. Another thing you can do a little more cutting edge is do that same exact process of looking at inner asset correlation, but instead of a regression approach effectively.

you plug it into a machine learning model. And again, those factors that are, you know, the assets are jointly sensitive to will give you a loading. So we've been able to test all these approaches. Now You mentioned, Corey, you have shrinkage, you know, which is effectively about trying to get rid of some of the noise in the matrix. That causes estimation error because the optimizer is very sensitive to small perturbations, small differences that can come from noise.

So there's another method, it's a denoising method that's been put out there. I read about recently. If you basically take you show the eigenvalues of a matrix, the eigenvalues are you're basically mapping your principal components. Eigenvalues are very high numbers. Those are telling you, hey, it's normally the market factor is that top eigenvalue. And then you you go down the list.

And most of the eigenvalues of the matrix are garbage. It's all noise, right? So the idea is you take those lower values, you normalize them, you take the average of them, you kind of reset them. This is a denoising method. But all these get at kind of the same thing. Let's reduce the noise and try to show the best of our signals so the optimizer can make its best decision.

So here's kind of the neat findings we've made around this. Number one, if you have other constraints that are perhaps more important than the covariance matrix. You're going to often get a result that's hard to determine, or even kind of it shows you the changes you're making to the covariance matrix are kind of insignificant.

Which sounds counterintuitive. How can you not if you improve your matrix, how can you not reap the benefit of it? But the reality of it is, and I'll give an example in our investment process. We basically take in our universe of, let's say, fifty five hundred liquid enough stocks globally to trade, we take top and bottom deciles of those.

within market segment when we rebalance and that's all we feed the optimizer, right? So in effect, you might have a thousand names going into the optimizer. We don't have all fifty five hundred going in. So now we've said, hey, out of those 1,000 names, maybe I only want to make portfolios out of half of them or even less than half of them, maybe a third of them.

The reality is that with an equal weight constraint Right, I can often I'm gonna end up with the same decisions because what's dominating is the expected returns, the other side of the formula versus the covariance made. All of this is said with guarded optimism. If we can improve the covariance matrix, we'll we're gonna reap a benefit from it. Well, it's kind of interesting. We have fed perfect

hindsight or perfect foresight correlations. So basically taking the correlations that were realized, cheating effectively, into our correlation model to test. Hey, how does this improve things versus a non-cheating method where you're looking only at hind site factor returns or whichever method you're using and finding you're not going to get much results. So It can be a lot of consternation depending about how your process works.

about something that may not be all that valuable. I say this in a guarded manner, right? Especially if you're if you're thinking about any one pair trade, you want to improve your correlation the best you can. We try our best.

But the point being, especially if you have a very constrained model in other places the characteristics of the benefits you're looking for, unless you're optimizing on a daily time frame where your tolerances are very tight, you may not see the result empirically in your testing, which is kind of a counterintuitive conclusion.

So this is a long, uh, long circular answer to your question, but yes, you can use all these approaches. We've even tried the machine learning approaches. We get a better result. We get lower, you know, mean squared error where we take basically

factors, relationships, given pairs in our correlation matrix. We have an output that is a better estimate. It's not as good as perfect foresight, but still you get very marginal improvement. I find that If you're running a relatively constrained optimization with small enough universe, a sub-universe that you optimize, I think other parts of the process are going to dominate. So what you're looking at, I may be able to make a result and it shows, you know, lower volatility or lower variance.

in my back test, but you know, I don't know how much to believe it. It could be just a function of noise and error because ultimately remember, I'm running a quarterly rebalance model, but I'm evaluating on a daily timeframe. And, you know, even a correlation number over three months uh is going to be somewhat disconnected from a daily timeframe observation.

To your point, it's been shown and established that different weight constraints are effectively shrinkage techniques in their own right. So when you move from totally unconstrained mean variance to constraints on mean variance, even before you start applying these techniques to the covariance matrix, you've already done it in effect through the optimizer. It It's always interesting to see sort of the marginal benefits of all these other numerical techniques.

that can be potentially mitigated through other actions you can take. And in your case, it's the optimization process and the constraints you place that move you a long way along the curve.

I would say that that and expected returns are amongst our most important inputs. You know, this is what we've empirically discovered. We got a lot out of shrinkage, right? There were improvements definitely to be had that were very easy to.

procure. But you know, when you started putting real firepower on the problem and machine learning model or even cheating effectively just to set a baseline for what your result could look like if you got it perfect, when you see how little it moved the needle past

what shrinkage gave you, it's kind of depressing because naturally you think, oh, if I can unleash, you know, all the firepower on the world on this problem, I can get a lot more out of it. No, it's not the case. You know, the reality is Uh hey, what's explaining our portfolio returns are really the factor movements. It's not our shortcomings in the optimization problem.

So I very selfishly have to ask this next question, which is you've done quite a bit of research into portfolio stability and the idea of rebalance timing luck, which is a real passion topic of mine. So I'd love to hear what some of your discoveries were. Okay. I have a I'm like you, I have a real gripe with everyone who evaluates returns out there of anything. It's obviously there's hindsight bias plays into things.

But you know, I often see people post charts about almost anything and try to compare a factor portfolio. They'll look at a few percent difference. and make it into something. And the reality is often what they're looking at is noise when you're making minute comparisons, whatever it may be. So and that's borne out from our experience, for example.

in doing our research, our testing on these multi-factor portfolios, we found when you rebalance, whether it's which month of the quarter. So it told you we have a quarterly model. When you rebalance, can have a very, very different impact in the short term. The shape of the long-term test is very similar, but I can have variance within one year that could be plus or minus 10%.

You could have a lot of error thrown into any single incidence right within the model. Again, the long-term picture is a very different one. It's one that says, okay, it's gonna be plus or minus 10 this year, but overall the picture is the same picture. The point being, you know, we realize there's kind of an inevitable tracking error from your ideal, but you realize your ideal is kind of formed out of some randomness. So what we do.

And I think we've talked about this before. You have a kind of a tranched approach to rebalancing. And we do exactly the same thing. And it's really to replicate our testing. So we basically when we run our test, we run with the first month, the second month, the third month of the quarter as our kind of rebalanced turning period to test our ideas.

just to make sure everything is consistent, everything works. And we see these large differences, but how do we get at making the real life portfolio kind of look like our back test? Well, what you do if effectively is you stagger your rebalance. So for us.

something we practically do. It's very simple. We imagine we have multiple market neutral portfolios, re-rebalance. Imagine on a quarterly one. And then in the mid-quarter, and we'll have half of our exposure in those independent separate market neutral portfolios. And then forty-five days in, you have the other half of the portfolio, a separate, newly optimized market neutral portfolio. So we're rerunning the model.

We're getting more recency for this period. Things go stale the minute you rebalance, especially with faster moving factors like momentum category factors. So with that, having staggered rebalance is a very practical solution. It gets you to a more robust answer.

And it mitigates, I think, some of the noise from timing luck. But I I think it's the point to remember, even with our high yield strategies, we run effectively a single factor trend model. I don't put terribly much weight in any one signal. Our investors may. I think it's all noise in the short term. It's a gamble. The long run, very different picture, right? So we say it'll all average out in the long run. And I think that's a prudent investor has to look at it.

So going forward, what are you really excited about researching? Oh, researching. I thought you were gonna ask, what am I excited about with respect to the markets, but with researching? Well you can answer that as well. Let's start there. The big thing we all want to know, especially for those of us in the factor space, we're clearly on the wrong side of the trade in the short term, right? Living through the value is trash route and anti-factor

is good, period. It's a hard thing to do. I'm very excited to see a turn. like we have in multiple times in history, right? So I would like to experience that just as a proving point. Maybe there's a little ego involved. You put so much work into something. You want to see a good outcome, but that does terribly excite me.

As far as research, I do love testing these ideas. You know, actually it's funny, you mentioned the idea of clipping eigenvalues and denoising uh covariance matrix. I love these little side projects. We test these ideas all the time. We have, for example, an ensemble approach we use in a lot of what we do, not just the factor side, but the tactical side.

It's always nice to see what new models, what new algorithms can come to light that we can add to the pool of essentially committee members voting at the table. And the same goes, what are the new signals? For what it's worth, on an ongoing research basis.

We don't add a new factor very often. And that's not for lack of trying. We try a lot of things and it's just hard to get any improvement after after you've exhausted so much of what's out there. Right. So I think it all fundamentally boils down to we're seeing results in our testing that's consistent with the academic research.

And I'd be suspicious of factors that come out of the blue that teach us something totally different than what we're already seeing. But it's always fun to test these new ideas. That's what keeps things interesting, I guess. You always have a fun pool of ideas to play with. Keeps the game going, right?

Well Michael, it's been a wide-ranging conversation, but one that I've had a lot of fun chatting with you about. Thank you very much for joining me. Uh if the listeners are interested in finding you, finding out more about your research and following you, where can they do that?

I think a good starting point is on our website. We have counterpointmutual funds.com. We have a factor scoreboard. And the neat thing about that, right, is that For most advisors or professionals who are not running the factor portfolios themselves and don't have their own data sets, they often don't have access to see what the real academic kind of long short portfolio returns are.

So there we post that. That's live and daily. It essentially comes out of the models, the factors that are in our model, and it shows broadly what the categories of factors are doing. So we hope that's a valuable resource for people out there who are just. I'm curious to get a little more wonkish about things. Well thanks again for joining me, Michael. Thank you, Corey. I appreciate your time. Thanks for having me.