96 The Simplicity of Trend Following with Katy Kaminski, Alex Greyserman & Roberto Osorio - 2of2

Katy

… What this tells us is that the simpler and the more robust and simple the solution is, the better.

So Alex and I had done some work on this as well, over long time horizons, the typical market based trend following approach tends to outperform. So one over N. So you allocate equal risk to every market. That particular strategy over long time horizons does very well because it basically has no view which market is going to do better than another. If you’re going to have a view, you need to have proof as to the contrary, and proof is hard to ascertain in a world where you’re stuck at 50/50, or 51/49.

Roberto

It’s the role of uncertainty that really makes mean variance not to do the job that some people expect that theoretical financial economists would expect it to do. Like Katy said, we don’t know what the expected returns are. At the most you make an informed bet that things are going to go up or down based on some statistical evidence of how they behave with respect to their past history, but it’s very hard to be any more precise than that. When you put that in a mean variance optimization engine small changes in expected returns yield very different answers.

Going back to the problem of simplicity of complexity, I think in my opinion, you should not introduce complexity gratuitously. I think Einstein said that we should be as simple as possible, but not simpler than granted by the problem – I’m paraphrasing now. Probably worse than he put it.

So the idea is that the components should be simple. The components of a model should be simple, but once you have new signals, you should introduce those signals in a way that’s not a Rube Goldberg machine, if you know what I’m talking about – those famous old comics where you have those very complicated problems to just draw a ball from point A to point B. We don’t want to do that. If you went to grab some effect you have to try first the simplest way possible, then you can introduce a bit of overlay of, let’s say granted complexity, that is granted that’s ensured by the evidence that you have, by the data evidence that this added complexity yields better results. This is always to be done very carefully because, again, you don’t want complexity for the sake of it.

Niels

No, absolutely. Now, in the old days, when you looked at performance reports of typical trend followers, you could almost guess how firms like Campbell, Dunn, ISAM, would perform if you knew the performance of just one of them. That’s not the case anymore. How is trend following involved in the, say past decade or so? Are there new trend following styles that have emerged that have created this more dispersed performance landscape?

Katy

Niels, I would just say that trend following has evolved just like any other industry out there. Back in the ‘70s when Campbell was formed, our founder was… basically that was when a computer was the size of a gymnasium or something like that. Really, in those days, it was about drawing pictures on paper and using your ruler and trying to figure out the trend, and slide rule. Eventually people figured out how to use computers, then eventually they used Excel, then they started coding MATLAB and doing all sorts of complex things. I just think it’s a natural progression of our industry.

What happens is we evolved, with time, to survive – very consistent with the adaptive markets hypothesis. As markets adapt, we compete. Those who don’t continue to compete and innovate don’t survive. So we’re constantly working on trying to find better ways to do what we do, and also trying to find new ways to improve our process. That’s what makes us all better, and that’s how we’ve evolved.

Alex

I’m not sure that it’s correct that the dispersion between trend following is actually increased. If you actually look, historically, at different trend following strategies, it’s been pretty similar over the years. The reason the dispersion may seem to increase is because a lot of CTAs now do other things besides trend. Somebody who you think may be historically has been a trend follower and now they do something else, and now that increases the dispersion between them and somebody else. If you look at pure trend followers, or you ask all of us to somehow extract just the trend following track records. You normalize it to the same volatility. We’ve done this study; you’re going to get dispersion right in line with your expected dispersion. When I say expected dispersion, you just take two return series with assumed correlation, if it’s 100%, with the same expected return you’re going to get zero dispersion. Meaning, the difference between returns, at .7 correlation, which is roughly the average, peer-wise correlation within the industry, I believe… it depends on the volatility assumptions, you should get something like about 20, 25 percentage points of so-called interquartile range between the managers.

If you look at the New Edge, or Barclays, or any historical data of track records, it’s going to be right on, spot on, in terms of the dispersion you’re going to get. Most people make the mistake of thinking that a .7, .8 correlation is so high that the returns are going to be very similar. But actually, unless your correlation is really high, I’m talking .95 high… .7 is high, but low. It’s seemingly high, but it’s already low enough that the performance is actually going to be all over the place.

If you look in the 80s, MINT and Campbell existed in the 80s, just between MINT and Campbell there was a fair bit of… maybe it’s easy to say now that it was the same in the 80s, but there were years, between MINT and Campbell and AHL and John Henry and whoever was around back then, that there was easily 20, 30 percentage point differences between terms. That includes the investors and doing their work and understand it.

Katy

We’ve actually done a little bit of research on this as well.

Niels

You did a white paper on this.

Katy

I wrote a white paper on it. It’s called Return Dispersion Counter Intuitive Correlation. One of the things that we discuss in this paper is that when trends, especially really strong trends occur, they are very idiosyncratic. It really depends on the exact allocation you have, different time horizons, different markets, and you can’t always predict exactly how that’s going to work. Sometimes it’s a difference… take the tulip crisis as an example, you wait two more days you lose it all, but if you’d been the lucky one that had the lookback window that got you to sell the day before, you look like a superstar compared to the guy who was going to wait one more day.

Roberto

Doesn’t your paper show that dispersion is higher when returns are positive? So yeah, when people, when everybody is following similar trends, they actually get vary collectively and intuitively they may get very different returns because it depends on the point of entry, the point of exit, how they diversify among different markets. The details of their algorithm, what is really their signal indicator?

Katy

It’s not a long release strategy. I think that’s the real challenge. If it was a long only strategy, correlation would be much better mapped to returns, but since it’s a function of how I’m going to decide to get in or out, you can have huge differences if you wait one more day, versus the day before, so that’s why we diversify, to try and smooth that out.

Niels

Speaking on diversification, you can talk about diversification on markets, models, data sources. Talk maybe a little bit about that. I think most people would say that they understand diversification of markets. They can make sense that if you trade oil, and you trade lean hogs, and maybe cocoa, and a few financial markets that that’s diversifying their returns, but what about the other kinds of diversification we can get? How important are they, actually in the overall…?

Alex

Trend itself is a diversifier. I know what you’re asking, but I’m going to be a little… This maybe gets into almost more than people need to… People get into analysis paralysis mode. It happened to me all day today, and all day yesterday, asking everything about everything. At the end of the day, most of the audience that is probably going to listen to this is going to need to step back and say, “Do I need this asset class? And how do I get into this asset class?”

Twenty, thirty years ago the access points weren’t as easy to find. MINT had a fund, Campbell had a fund, John Henry had these private funds. Now, I think, all three of us will do… Well some of us have mutual funds, some of us have managed accounts, and there’s almost no excuse not to have this in a portfolio.

So I need to mention a white paper. I have two. Katy, you haven’t seen these yet. We should probably agree that anything general should be for public consumption. So one of them talks about Geek Warning: The Black Litterman Model, which nobody really has to know about what that means other than I’m proud enough that I was able to convince one investor, who was on the edge, by showing them that what kind of implied views on trend that they actually have to have in order to have a zero allocation.

Everyday there’s actually an investment decision. So I’m talking to an investor and they have an investment decision. They have zero and trend. There’s no excuse not to invest anymore because they don’t have access. There’s mutual funds, there’s managed accounts, there’s all sorts of private vehicles, there’s all sorts of stuff. Tomorrow you can be invested, so you can’t have an excuse.

It’s transparent, we all have professional administrators, mark to market, there’s no operational excuse. But yet you have zero, so I worked out the kind of Black Litterman thing, which your readers can probably fall asleep trying to read about and say well, what kind of views and trend do you have to have in order for your optimal allocation to be zero? Have you seen this work here? So assume it’s an asset class, and assume you went through an investment decision making process and you came out with zero, as an optimal outcome. So what would you have to feed in for that to come out?

You have to have… because of the risk defining nature of this…

Katy

Did you include the correlation effects too?

Alex

Of course, so you either have to have one of two things have to happen in order for you to have an optimal outcome. You have to assume highly negative returns, which OK, if you think trend following is going to consistently lose 10% a year or something like that, then you shouldn’t invest, fair enough. Or you have to assume extremely high correlation to major asset classes, which none of us have displayed for any period of time in actual history, in 800 years, or anything like that, we have negative correlation to equities. I think most people do.

Certainly, as we speak today, in this month of January, every CTA has displayed a negative correlation to major asset classes. Let’s just call that zero on the average. So it is, when I pointed this out to them, I actually got this guy to invest because, basically the decision, the implied views in this asset class is so ridiculous. OK, so then I follow with another white paper called All Portfolios Need Trend, again we can make this available for sale, or for free. You probably haven’t seen this one yet, where we took all, not all, but most available basic asset classes that people can have: stocks, equities, private equity, real estate; a bunch of indices from Bloomberg or somewhere, and we created 1 million permutations of investments that people might have, trend not included, that somebody might have i.e. 40% real estate, 30% private equity, 30% bonds. OK, that’s one possible portfolio that somebody has. A million permutations of all kinds of things in the investor world. Then we said OK, what’s the optimal allocation to trend from every one of those?

Katy

What was your optimization?

Alex

We put up a histogram. We did all kinds of minimum variance, and maximum verification and all the geeky stuff and just averaged. The point is not the geekiness, the point is that if you look at the histogram of the outcomes, the lowest outcome allocation to trend is something like 20%. The average is something like 35%. The highest is maybe higher. Very few investors will actually have the discipline to do this, but zero or something low? There’s just no excuse not to have that. At the end of the day we have to cut past all the analysis paralysis, and which one of the three of us is better? Any one of us is better than nothing.

Niels

Did you also not write a paper, maybe it’s part of this, something like The Cost of Not Having Trend Following? Or is that a different paper?

Alex

Was that in the book? I have a paper called The Cost of Not Having Trend Following in Your Portfolio, but maybe we’re starting to write another book already.

Niels

Absolutely. Do you have anything to add to this, Roberto?

Roberto

I guess the message is very clear then, when you put trend following and diversification together, the message is trend following CTAs are always diversifying instruments for people who have an optimal portfolio. I just want to make clear that we talked before about the entry barrier, and I just want to clarify that we were talking about the entry barrier for starting a business as a trend following CTA. The entry barrier for an investor is very low now because there are mutual funds that provide trend following CTA exposure, so really there is no excuse for a regular investor, even middle class families, to have some exposure to trend following CTAs.

Niels

Sure. I want to… Do you have anything you want to add, Katy?

Katy

I was just going to add to it, what Alex said, in our book we talk a lot about the concept of divergent strategies. Those strategies that are cutting losses and following winners, so looking for movements. If you really think about the philosophical reason for the solution in Alex’s example, if most of the long only portfolios that you mentioned really are convergent portfolios. They are having a view about your asset allocation, and they tend to be thinking about finding risk premium, which is a convergent strategy. Anytime you have that philosophy, adding a strategy that has a different philosophy is going to diversify how you react to different environments and that is why trend following matches so nicely with a lot of the more fundamental approaches out there. Because you’re really trying to deal with risk in a different way, and that naturally compliments these more traditional strategies. So that’s an empirical example, but there’s also philosophical reasons for why that’s the case.

Niels

Sure, absolutely. Now at the heart of trend following, of course, we’ve touched on this, is generation of signals, sizing of positions, exit strategies, and overall risk management. Now, I wanted you, Roberto, to kick off and talk about different ways that you generate signals apart from the coin flipping we talked about. Give me some examples for the audience to understand when we’ve used these terms, so that we’re not sounding too geeky about it, what does this really mean?

Roberto

Well, this really comes down to looking at best history and it is really brief then, summarizing the phrase buy the winners, sell the losers. I was going from divergent to convergent for a while: buy the winners, sell the losers. That’s what trend following is. The details make all the difference. You can have different lookback periods to define what is a winner, or a loser. You have the different ways of massaging your signal, of defining your momentum indicator, and different stop loss or other types of exit strategies. So that’s pretty much the signal definition. It’s hard to go into more detail than that without really getting into a lot of geekiness.

Niels

Sure, that’s fine. Now, Katy?

Katy

I always turn back and say, trend following is a heuristic, so it’s just a rule. So the way that you think about defining a trend following strategy is just making sure you define the parameters of that rule. In the simplest sense, the signal, in terms of going long and short – you can say, OK, I’ll look over a month. If the average is increasing over, let’s say three months, things look good and you go long. You do that until you finally see that you start to lose in that position. So really you could go through many permutations of all these simple heuristics, but at the base it’s really just defining rules to diligently get in and get out of positions. So, for investors its exactly what they do already in their portfolios.

How many people do you know that say, “I’m going to get out when this happens, or I’m going to buy this month and then when I buy this month I’m going to wait until this happens. All we’re doing is trying to be consistent, parsimonious, and thoughtful, in terms of the way that we design these heuristics and then to put them together in a way that makes sense. So really we’re trying to systematize what investors are doing themselves.

Roberto

You don’t have to find the best statistical evidence that your particular choice of parameters works, you have to be very careful about not overfitting which goes back to that theme of complexity that we talked about. You don’t want your system to be more complex than it needs to be.

Niels

Sure. Now earlier we talked about that a lot of people think that the entry point is the most important of what we do, but in fact, I think we all agree that that’s not the case, so position sizing or management of the position during the trade and exit, how do they stack up? What do you feel, if anything, might be most important? What do you think Alex?

Alex

Yes, probably some of the original trend following systems that you get from text books are very binary in nature. You’re looking at data, you look at some past amount of data, you’re possibly looking at a breakout of some technical factor. You get in, you buy 100 contracts or whatever the number is, and you stay until trend comes out.

The more modern versions of trend following are much more continuous in nature. You evaluate signals every day, you adjust positions every day. This month I’m sure, in our case and other CTAs, there’s a lot of position adjustments due to risk increasing or decreasing, mostly increasing, whatever the case may be. So actually managing risk is probably more…

When I was being hired that’s what Larry told me that we’re just, what we are is just risk managers. Everyday your just handed a position and you’re trying to manage it because your risk increases or decreases during the trade. So in order of importance, in my mind, it’s exits, position management, entries are actually the least important. In fact, if you look at the corollary or the opposite of something of my exercise of that random entry with a trailing stop, if you actually use any kind of professional trend following entry – some kind of moving average or something like that, and you just hold for a fixed amount of time, you make very little money. The forecasting nature of a trend. In fact, the third or fourth thing I did was take a trend following… This is now 25 years ago… take a trend following system and delay it by a few days, or five days, or… you can’t delay it by five years, but if you delay it by a small number of days, it makes no difference to the result. It just shows you how unimportant in terms of precision of timing is the…

Basics, just look at some history, take a ruler out. If it’s pointing up trends are going up; and if it’s going down, trend is going down. If I ask anyone of my investors to guess whether we are long or short any market, unless it’s been going totally sideways, nobody is going to be surprised by the direction of the position, or at least shouldn’t be surprised.

Niels

Sure, sure. We talk about some of the key elements in trend following and we talked about parameters inside those things, but how do we figure out what’s then the most robust parameters to use? Roberto, why don’t you…?

Roberto

The approach we have done is to use what we call meta-parameters which will determine your choice of specific parameters. You don’t want to pick… Let’s say you have a certain set of parameters, end parameters, you don’t want to pick the set that performed the best in a given lookback period, which would be a number of years, or in case of some people, it could be the whole available data history – it could be several decades. Because, of course, whatever performs best, and the best is not weighted, almost surely is not going to be the best performer in the future. It might provide some indication about what will perform better in the future.

So you can choose a number of parameters, around the best performers. You can use some correlation rule for that. So that involves another level, another hierarchy of parameters. How many of the individual parameter sets are you going to pick? What is the rule that you’re going to invoke to pick those parameters? This process we call a walk-forward process, can be done every month, or every week. The important thing is that your final system is robust with respect to the meta-parameters. So if you choose slightly different values, or even somewhat different values of those meta-parameters you ought to make sure that your system has a similar performance. The sharpe ratio should not be different than .1 or .2 from the choice that you actually make at the end.

Niels

Sure, OK.

Katy

I think, to add to that, in the sense that I was just talking about this from the Campbell perspective for example. We follow a pure scientific process. Every investment idea that we have is derived from an investment thesis. That thesis has to be validated, and then that thesis has to be peer reviewed. So through that process, it’s about understanding parameter stability. It’s about understanding, really, if this is an investment idea that makes sense, and that is the principle that we follow through and through to the end of the process. Even with an investment thesis seems to not work, we go back and we review the investment thesis. So what you can see here is that it’s not really a question about picking parameters, it’s about following a systematic scientific approach to make sure that you’re not just trying to match the past; you’re trying to have a view that you can validate in many different ways.

Roberto

That’s very important there. You have to have a story about what you’re trying to say because otherwise it becomes data mining, which you want to avoid all the traps of.

Niels

Yeah.

Alex

We have to think about this also at a high level. I come back to my class, again. Every fall I have people do a project on… many projects, one of them is to come up with a trend following system. People think, well I asked them that because I’m trying to get some free ideas from academics. That’s not the case. The case is just to see the traps that… for me to see the traps that they fall into and then to knock them down. So this past fall, they do a project and they present. This one kid came up and had like 20, 30 years of data, and he came up and shows me a sharpe ratio of 2.5, which in our industry is unheard of, so how did you get that? Well, I give them a data set of 100 markets and he said, “Well, I ran this trend following system on these 30 markets.”

“OK, why is it 30 markets?”

“Well, because the other 70 markets don’t trend,” he said.

Right, because obviously he filtered out after the fact. The next guy comes up and he’s got another sharpe ratio of 1.5 or 1.8 or something. “How’d you get that?”

“Oh, I have a slow system on bonds,” which of course worked well for the last 20, 30 years.

“How’d you do that?”

“Well, because that’s what worked.”

They go on, and on, and on, and on, and they probably pull me up on the internet and say, “Well, the professor’s a dummy, he can’t even get a sharpe ratio of 1, and we get a sharpe ratio of 2 or 3.” So what happens is, I think we all have to realize as scientists that we actually have a trivially small amount of data. Even if you’re not technically data mining, we’ve had a 20, 30-year period of basically one-way markets in bonds – not really, necessarily one-way market in stocks, but certain biases. If you put a science hat on, this is like a trivially small amount of data compared to what we might have done in science.

So sometimes you’re just tempted to data mine the past and think that it worked. It’s true that slow systems in bond worked in the last 22 years. I can’t say would not have worked. So we do a lot of synthetic data creation, and things like that to try to avoid the traps of actual data, even though we should be using actual data. It’s amazing how every single one of my students gets a higher sharpe ratio in their piece of research than we’ve actually delivered in real life.

Roberto

The other point I’d like emphasize from this scientific viewpoint is that we have critical advice. People, every member of the team to criticize strongly, vehemently every new idea that comes by. Everybody should have a high incentive of trying to debunk any new ideas. Nobody should try to take it personally. When you come up with some new idea, you should be ready for it to be destroyed by someone else until you have very, very strong evidence that it works.

Niels

Yeah, no absolutely.

Katy

Maybe strong evidence that it works sometimes – most of the time.

Roberto

That’s what you want.

Katy

Sometimes, yeah.

Roberto

Do you want to talk about risk control?

Niels

Yeah.

Roberto

Alex mentioned, I think…. If you look at the history of trend following CTAs, I think the initial, and I mean ‘70s, ‘80s approach, was to have risk control based market by market. A situation where you don’t even consider the intermarket correlations, you just normalize your exposure by some measure of each market’s volatility. So this evolved in more recent times. Everybody is looking at correlation. You have either a target for your value at risk, or new portfolio volatility and you’re going to make your exposures to each market fluctuate according to that target. I think few people really do mean variance because, for the reasons we discussed, that this is a very unstable process. Really the “one over N” approach seems to work better in the trend following regime.

A new development, I think, is to have your risk target, and that’s what we’re doing right now, your risk target fluctuates with time, according to the conditions of the market. You can go up or down if you think there are more trends or fewer trends. It also depends on the correlations. Typically, when you have more correlations – inter-market correlations – we are going to decrease the value at risk target. There is also some scientific evidence for that in the way you can optimize your portfolio according to your expectations and your variance.

Niels

Sure. I want to be mindful of the time because I know we, today, have some constraints on time, but I want to touch on a couple of things, maybe more from a 10,000-foot level.

Roberto

30,000-foot.

Niels

Or 30,000-foot level, if you prefer that. So, in terms of education, some investors that I’ve come across would say, “Well, I’ve already got one trend following in my portfolio, so I think I’m covered, I’m good.” We talked about that correlation between managers doesn’t mean that, necessarily, the performance will be the same, but are there a few simple guiding points we can share with our audience to say, “Well, maybe you need to consider at least three or five trend followers in order to have some level of diversification among trend followers, or how should investors go about this?

Katy

I guess there’s a simple answer, right? An investor should choose three, because there’s three at this table (laugh). Just as a joke there, but we talk about this in this white paper that I mentioned earlier about return dispersion and counter intuitive correlation. I’d say it depends on how diversified you are. The more diversified you are in the space, the better your risk adjusted sharpe ratio is going to be, or your risk adjusted performance. I think the real reason this type of question, especially, comes up is a lot of people think that they can maybe have a really simple exposure to trend following. They get the correlation, but they’re not going to get the performance.

I tend to tell investors that more diversification doesn’t hurt you. You’ll see that in any particular year. There are certain months where there are inflection points. Some of us may have more commodities exposure than others, or others may be more short term. Really, to weather difficult storms in markets, the more tools you have in your tool chest the better off you’re going to be. Each of us has our own set of ways of doing things. Some of them are similar, but really it’s about the risk management decisions that we make, and the type of markets, and the allocations we make, and how those differ, that’s going to provide some diversification – if you include different managers in a portfolio. We see allocators doing that as well. Some of them look for specifically to pair us up with other managers, depending on what return profile they’re trying to find. Some of them are looking for, maybe a robust one manager solution. It really depends.

Niels

Sure.

Alex

We have graph in the book. It’s a number of mangers on the X axis, and dispersion on the Y axis and the answer is, you should get… If you have one manager, I don’t have it in front of me, but if you pick the right one, it will be the right one, but you could easily experience 10%, 15% variability in your returns. It’s like picking one stock. So there’s evidence from how many stocks do you need to minimize variance. So once you get to 3 or 4, or three of us and maybe you spin out your own CTA, Niels, that makes 4 (laugh), then once you get past that it just becomes diversification. Like Buffet would say, “Change the V to a W.” So the answer is, not one, not even us or probably any of the ones here. Two is tight, unless you have very specific mandate and you’re trying to plug in a very specific hole. Three or four tends to be, in practice, the right number.

Katy

Actually, I can go back out to a point that we brought up in the book that I think is a really good way to communicate this to investors. I talk about heuristics, we talked about parameters, if you think about any individual manager, what we are is a collection of heuristics that we’re putting together and parameter sets. So the more parameters you have, the more diversified you are. At a certain point the marginal contribution of more parameters doesn’t add as much, so it’s the same idea as a stock portfolio. The more stocks you add, the better your diversification. In our space the more heuristics and the more parameters that you cover, the more likely that you have a robust exposure to the asset class.

Roberto

I just wanted to add that, yes, you are going to increase your diversification by picking three or four parameters in the sense that even though they have correlations of .7 to .8 you might increase the information ratio of your portfolio if they are good performers. So you need to make sure that the process of choice is very effective. You need to make sure that you’re picking the best CTAs around and you decide that, not only by the best performance, but you have to know what the process followed is. You have to know the research team and make sure that they have a scientific statistical approach and rigorous mechanism to develop and improve their systems.

Niels

Sure, absolutely. On this not let’s stop the conversation for now. Let this be the beginning of a series of master classes with some of the best trend followers in the business. I appreciate your willingness to share your insights, your experiences, and all the work that you do. The book I mentioned in the beginning is, of course, Trend Following with Managed Futures: The Search for Crisis Alpha, written by Katy and Alex and also I’m grateful to you, Roberto, for joining us today. Of course you can find, as a listener, all the details of today’s conversation on the show notes page of this episode of Top Traders Unplugged. I hope we’ll speak again soon and continue this conversation.

Katy

Thank you, Niels.

Ending

Thanks for listening to Top Traders Unplugged. If you feel you learned something of value from today’s episode, the best way to stay updated is to go on over to iTunes and subscribe to the show so that you’ll be sure to get all the new episodes as they’re released. We have some amazing guests lined up for you, and to ensure our show continues to grow, please leave us an honest rating and review on iTunes. It only takes a minute, and it’s the best way to show us you love the podcast. We’ll see you next time on Top Traders Unplugged.