In this episode, we talk with Ernest Chan from QTS Capital Management.
Ernie, a physicist by training, is the founder of QTS Capital Management and author of a number of books on Quantitative and Algorithmic Trading.
In today’s episode, we talk about Ernie’s tail reaper strategy at QTS which seeks to profit from down moves in the S&P. We go into how Gamma Dealer Hedging and Forced Rebalancing exacerbates intraday volatility in the SPX in a way that makes the S&P one of the best places for a tail strategy to work.
We also go into the Evolution of Machine Learning and how it is applied to trading strategies including tail reaper and the appropriate way to use kelly criterion for investors – a lesson Ernie learned the hard way.
I hope you enjoyed this conversation as much as I did.
Listening options:
- Listen on Stitcher
- Listen on iTunes
- Listen on Spotify
- Android users can grab our RSS feed here.
- Listen on YouTube
Have comments about the show, or ideas for things you’d like Taylor and Jason to discuss in future episodes? We’d love to hear from you at info@mutinyfund.com.
Transcript for Episode 17:
Taylor Pearson:
Hello and welcome. I’m Taylor Pearson and this is the Mutiny podcast. This podcast is an open ended exploration of topics relating to growing and preserving your wealth, including investing, markets, decision making under opacity, risk, volatility and complexity.
Taylor Pearson:
This podcast is provided for informational purposes only and should not be relied upon as legal, business, investment or tax advice. All opinions expressed by podcast participants are solely their own opinions and do not necessarily reflect the opinions of RCM alternatives, Mutiny Fund, their affiliates or companies featured. Due to industry regulations participants on this podcast are instructed to not make specific trade recommendations nor reference past or potential profits and listeners are reminded that managed futures, commodity trading, forex trading and other alternative investments are complex and carry risk of substantial losses. As such, they’re not suitable for all investors and you should not rely on any of the information as a substitute for the exercise of your own skill and judgment in making such a decision on the appropriateness of such investments. Visit www.rcmam.com/disclaimer for more information.
Taylor Pearson:
In today’s episode, we talk with Ernie Chan. Ernie, a physicist by training is the founder of QTS Capital Management, and author of a number of books on quantitative and algorithmic trading. In today’s episode, we talk about Ernie’s Tail Reaper Strategy at QTS which seeks to profit from down moves and the S&P index. We’re going to how gamma dealer hedging and forced rebalancing in the current market exacerbates intraday volatility in the SPX in a way that makes the S&P one of the best places for tail strategy to work. We’re also going to the evolution of machine learning, and how it is applied to trading strategies, including earnings Tail Reaper Strategy. Finally, we jump into the appropriate ways to use Kelly criterion and how Ernie thinks about the appropriate use of leverage in a portfolio, a lesson Ernie learned the hard way. I hope you enjoy this conversation as much as I do.
Taylor Pearson:
So, Ernie I’d like to start, one of the things you’ve spoken about that I find very interesting and also very counterintuitive. How is it possible to have an investment strategy that has an arithmetic return of zero, but it also improves a portfolio just to give an example of an arithmetic average, I mean, if we started a strategy on day one with $100 in year one, then 10 years later the ending balance of that strategy is also $100. The arithmetic average over that time Period is zero. How can a strategy like that improve a portfolio?
Ernest Chan:
Well, such a strategy would need to be combined with another strategy that is maybe anti correlated with or maybe even uncorrelated with in order to decrease the overall risk of the portfolio. So, essentially the growth of wealth is a compounding process, right? We invest a million on first year and second year let’s say go up 10% become 1.1 million, but then, if it keep going at the same rate it is compounding it’s not just going to add extra fixed amount of dollars every year it’s supposed to add more every year. So compounding has a very interesting effect that the compounding growth rate of a portfolio depends on the risk of the portfolio. It does not depend only on the arithmetic mean of the portfolio.
Ernest Chan:
The lower the risk of that portfolio, the higher the compounded growth rate. And so if your strategy that has zero arithmetic mean is applied to another strategy that has a small positive arithmetic mean, but that are uncorrelated with your first strategy, the overall portfolio may have low risk because they are anti correlated, so some of the fluctuation in the return are smoothed out. So, instead of one month up and one month down, you might get both months to be up a little bit. And that’s actually better for compound growth rate. A compound rate depend on the fact that your net worth doesn’t suddenly have a sharp draw down. It loves consistency. It loves smoothness. And so if you return up positive every month, it beats having one month with a big positive return and another month have a big negative.
And that’s how you’re adding a zero return arithmetic returns strategy to another strategy that has a small positive return can actually bring about a higher compounded growth rate. T that’s essentially the solution to the paradox.
Taylor Pearson:
Yeah, I find that interesting, we talked about in terms of volatility, drag, if you’re able to combine two things, and it reduces sort of that volatility now, I mean, if you start with $100, and you lose 50% of it, now, you have to get 100% to get back to where you started, right? Those big draw downs are very painful. So maybe to go from there we’d love to hear just little bit about the origin of your Tail Reaper Strategy, sort of how it works at a high level.
Ernest Chan:
Sure. So we started our fund in 2011 and at that time we were trading a spot Forex Strategy only, nothing to do with Tail Reaper. It was a fairly high frequency market making strategy in the spot forex space and we had a great one in the first half of the year until even July we had a great run, we practically have no losing trades. And we were very perhaps overly confident in that strategy. And all of a sudden in the August of 2011, a black swan event happened. And that was the historical first downgrade of the US Treasury that never happened before in 200 plus years. And that hit us really hard. It caused 35% draw down in our portfolio given that we elaborate 14 times. So we were certainly shell shocked at a time, but we were thinking how are we going to protect ourselves from these kinds of terror events that never happened once in 200 years.
Ernest Chan:
And we look and look and we came up with this strategy that turned into Tail Reaper strategy. It wasn’t called Tail Reaper at that time, but it’s essentially the same strategy we trade today. And that is an intraday trend following strategy that trade nothing but the E-mini futures. Very, very simple, buy when the market is already up, and hope that it goes up further and short when the market was down and hope it goes lower. Extremely simple and we only hold position for a few hours a day. And with this simple strategy, obviously we add [inaudible 00:07:47] resource over the years but that was the core idea.
Ernest Chan:
With this extremely simple setup, we’ve see two, you might call extremely troubling quarters. The first one was the 2015 fourth quarter Chinese stock market meltdown, which affected US equities greatly and we generate a decent return in that quarter. And then the second test was the 2018 first quarter where we had the opposite problem it’s not that the economy’s going poorly, it’s going to well, and people were afraid that this fade is going to increase interest rate. So some volatility level, volatility
exchange traded portal actually went out of business in that quarter because of surge in volatility. So, we were happy with the strategy continue to run it and of course, this year, we had truly unprecedented event happen. So that is the core strategy, it’s a very simple, transforming strategy, but what distinguishes it from other Tail head strategy is actually what we do to prevent it from losing money when the market is bullish and calm. But that’s a separate layer that we apply to it using machine learning.
Taylor Pearson:
I guess this first one to start in terms of the Tail Reaper we’re talking about your user trend following approach and your trend following or some variation of you buy something that’s been going up there, you sell something that’s been going down, I think most people think of trend following, especially in the Commodity Futures space, the CTA is tends to be longer term trend followers that are using one month, two month, year long sort of timings in terms of when they’re going, how they’re going long or short. I know you’re doing purely sort of an intraday approach. And so it’s a matter of hours as opposed to a matter of days or weeks or months. But yeah, I’d love to just comment on sort of your approach to sort of trend following and I want to get into you mentioned, machine learning and how you think about sort of the risk element there maybe just start with trend following in general and sort of the approach you take and how it’s maybe different from what people typically think about when they hear the term trend following.
Ernest Chan:
That’s right. Yeah, I mean, traditionally, yes, most CTA trend following strategy follow a trend that has been established over multiple months. A very favorite timeframe would be one year, would be the typical timeframe for trend following, but we choose to be intraday because we also observe that trend has been getting shorter and shorter. That’s one of the reasons why I think it’s advantageous to have a much shorter timeframe trend following strategy, but also our trend forming strategy take advantage of a certain market dislocation or market opportunity that most other trend following strategies are not utilizing. And that is the phenomenon of forced rebalancing of levered portfolio.
Ernest Chan:
So to explain it better, imagine that you are holding a three times levered ETF that is based on the S&P 500. If, one day, the market goes up 2%, all of a sudden, your ETF is losing 6%, not 2%. 6% because of the leverage. And if you don’t do anything, at the end of the day, that ETF will be over levered because your equity has dropped due to the big loss and the equity has dropped much more than the market value of your portfolio has dropped. So you’re over levered. And as the sponsor of that ETF, you have to sell some holdings in order to reduce the leverage.
Ernest Chan:
And it’s the same apply when the market went up a lot. So, because of all these levered instruments in the marketplace, not just ETF but many portfolio managers have the same risk management mandate, whenever the market goes down, they have to sell some holdings. It’s like portfolio insurance in the old days, in order to reduce risk and maintain your leverage. And that actually, that sort of risk management drives the market to trend in the same direction of the move. And that’s what we’re trying to capitalize on. And similarly, another phenomenon, which is gamma hedging, so many option market maker or market maker for swaps, let’s say, a big bang sale, had funded for a volatility swap, and then they have to hedge the gamma that follows a swap. They also have to trade in the same direction as the market movement to limit their risk, to hedge the risk, and all these hedging and rebalancing activities drives, the trend, it exacerbate the market moves.
Ernest Chan:
And that’s what we’re trying to capitalize on. But this all kind of hedging like this only occurs in your day. Right after the market closes it’s a new day, nobody care anymore. And so, we do not expect the trend to last beyond the market close, we have to capture it intraday. And that is one of the main driver of the offer of this strategy.
Taylor Pearson:
And so you think this strategy… it is based on for lack of a better term or sort of current market microstructure in a way it’s part of the way in which you said like gamma dealer hedging, and then also different leverage players, it’s the portfolio insurance analogy is interesting, right? Because there’s a lesson that got learned there and maybe has lesson just been forgotten? Or why? What is sort of the rationale why things are behaving that way? Why players behave in that way, are there certain incentives or how do you look at it?
Ernest Chan:
Well, I don’t know that the lesson is forgotten, but we have a new class of instruments that is such as the levered ETF that they have no choice but to do portfolio insurance. It’s mandate, it’s in the prospectus of the offering that they have to do it so they have no choice.
Taylor Pearson:
And then there is the next sort of arrow and again to you, you’ve mentioned that the way you approach this is using exclusively S&P futures, E-minis as your instrument so obviously [inaudible 00:14:37] can show up in lots of different places it can show up in equity markets in other countries and bonds and forex whatever. Why sort of E-minis exclusively? What is it about that that makes sense to you? Or how do you think about it?
Ernest Chan:
Yes, so this phenomena is most prominent in the S&P because of the large amount of derivatives that are tied to the S&P. There are not an equal amount of notional derivatives exposure to the DAX index or to the Nikkei and so forth. I mean, there are some naturally but nothing comparable to that tie to the S&P. And similarly, for the leveraged ETF, the AUM of leveraged that are tied to European or Asian indices are minimal compared to what’s tied to the S&P. Even within the US market, the amount of leveraged ETFs tied with small cap index is much lower than the AUM tied to S&P so it really works best in the S&P and not fare well in our markets.
Taylor Pearson:
That’s interesting, we should probably using this term improperly treated the market microstructure function their house where things evolve around the S&P, I guess next I want to talk a little bit about machine learning, you’ve written a number of books on machine learning, you’ve been very involved in the field going back into the 1990s, I want to talk to you about how you think about incorporating machine learning into your trading approach and sort of the Tail Reaper Strategy, specifically, but maybe just to start, how would you explain machine learning in general to someone that isn’t familiar to… do you have some common examples, do you have the types of problems that could be in trading or otherwise that machine learning is good and bad, and then how it compares to I think people hear other terms around like AI, like deep learning and neural nets. And those are just things that you can maybe paint to sort of a picture for how that fits together.
Ernest Chan:
So I think machine learning is a technique which allow us to examine a large number of potential predictors whether of the market or what the speaker is going to speak next or to predict whether a pedestrian is going to cross in front of a car in terms of in a self driving car context, right. So, traditionally in investment or in traditional quantitative finance, we use a very simple… you can call it machine learning, which is our good old linear regression fit, right? If you have a factor model, you look at price earnings ratio, you look at book to price ratio, book to market ratio, whatever four or five factors, and you try to use that to predict the return of a song. You can say that is machine learning, but that’s not the best use of machine learning because it is only looking at a handful of variables. And those variables, you already pretty much know they are predictive, well maybe not value at the moment but we know that after careful in depth research, many finance professional believe that these four variables are likely to be predictive of future return.
Ernest Chan:
So you actually don’t need the machine to tell you what variables are important, you already know that all you need is a actual [inaudible 00:18:17] to prediction based on those four variables. That is a kind of fun, naive form of machine learning, I would call it. The true, important and use of machine learning is when you have 1000 variables, and most of them you have no idea if it’s going to be predictive of the future, because you cannot possibly examine… no finance professors, no matter how genius he or she may be, can examine 1000 variables in depth to see if any one of them can predict the stock market it’s too many variables. And that’s where machine learning can help because they are established algorithm that can handle 1000 or 10,000 variables, and select those that are actually useful and get rid of those that are not in building a predictive model. So that in essence is I think how machine learning can help, both in investment and in any other field is the ability to examine a large number of predictors.
Taylor Pearson:
I think we know one topic, you’ve spoken to a lot, and I think one of the reasons you went through a period of your career where you weren’t actively involved in machine learning is this idea of overfitting. If you have… the example I like is that there’s a very strong correlation between Nicolas Cage movies released in a year and swimming pool deaths, so you could say when Nicolas Cage releases more movies, you should be long swimming pool fatalities and when you release this last you should be short showing four fatalities, but obviously that’s a spurious correlation that’s first you have to clean the data. How, maybe kind of looking back At the history of machine learning in general, and as it relates to overfitting, in particular how has that changed over time? How are the approaches change? Why does it sort of make more sense now than maybe it did in the 90s or the 2000s?
Ernest Chan:
Yes. So indeed the way to reduce overfitting has been one of the most active research topics in the machine learning community over the past 10/20 years, everybody’s aware of that problem. And it’s not just in finance, although it’s particularly problematic in finance, the whole reason where neural network wasn’t really commercialized until the last few years is this overfitting problem? I mean, it’s not new. It has been around for at least, I don’t know 35/40 years, it’s not a new concept. But it is only in the last 10 years, when a technique called dropout was invented by Professor Hinton among others, where you are deliberately punching holes in the neural network to reduce overfitting to pass data, that the out of sample performance suddenly becomes tolerable and the performance on business and promotional problem actually become adequate.
Ernest Chan:
And so I would say most of the advances in machine learning in the last 20 years is focused on overcoming this particular problem. And for example, we started in the early days when I was in [inaudible 00:21:43] in the early 2000. Actually, late last millennium, unfortunately, it was on… we like to use regression trees and decision tree to make prediction but that invites overfitting. So gradually people change that approach and adapted to call random forests where you basically create many random trees and average them. So averaging in an ensemble of learners is one way to overcome overfitting, you introduce randomness.
Ernest Chan:
Essentially, in order to overcome overfitting, you have to introduce randomness in the model sounds intuitive, but that is the case, that’s the same as the dropout technique where you randomly punch holes into the neural network. In the random forest, you create many, many random trees based on different randomized set of data in order to generate noise in the data and create models that are fitted to those noise but you averaged them so that you won’t be dependent on that noise.
Ernest Chan:
So, all these techniques have been perfected or refined in the machine learning community today where the problem of overfitting has been greatly reduced compared to 10/20 years ago and become usable techniques. But that is for general machine learning. Of course, with respect to finance in particular, there is one realization that were much more recent, which is that if you are trying to use machine learning to predict the market directly, the chance of success is low because of the low signal to noise ratio in the market. Everybody whether it’s your machine or human want to predict whether tomorrow markets going to go up or whether stock is gonna go up. And because if there is a strong signal it is going to be immediately arbitrage away by default, usually you won’t find a strong signal because if it was so obvious, it will disappear. It’s like the famous economist joke that if you see the market is so efficient, if you see a $10 bill, sitting on the pavement of economists won’t pick it up because they don’t believe that it actually exists.
Taylor Pearson:
Someone else would’ve picked it up already.
Ernest Chan:
So yeah, I mean, the same applies to applying machine learning to finance is that if the opportunity is so obvious, people would have already traded on it, you don’t need to wait for the machine to find it. So one realization, one particular good application of machine learning to finance that has been talked about in the last few years, particularly by Dr. Marcus Lopez de Prado, since this book was published, was the technical meta labeling where instead of actually trying to predict whether the market is going to go up, we want to use it to predict whether a particular trade is going to be profitable not based on a trader’s own past history, and that has much lower signal to noise ratio and also that has less fear of arbitrage because everybody’s trading strategies are different. And the machine is trying to learn from the traders own track record, rather than from a public data set where everybody can arbitrage on. And that many people have found to be a much better use of machine learning in finance than just directly to predict the market.
Taylor Pearson:
And yeah, I want to come back and talk a little bit about meta labeling, but maybe if you just wouldn’t mind for the listeners, sort of like out of sample performance, just defining maybe giving an example of out of sample performance and what that… a use case or what that would mean. And essentially you’re using it and then yeah, I’m also very interested in the random forest idea of of using an ensemble of approaches as we talk a lot about ensembles and diversification in terms of investing. But could you speak to sort of that random forest like if there’s any simplified examples that you might use in one of your books or when you speak about it, how is that approach sort of different from what was done historically?
Ernest Chan:
So the random forests approach is essentially randomizing the data. So originally we have, let’s say 1000 rows of data and you use that 1000 rows to fit one decision tree. That’s it and that tree is going to capture all the non repeatable patterns in that 1000 rows of data, capture every wrinkle of that data, and they think that wrinkle is going to repeat yourself. But for random forests, they will what is called sample with replacement from that 1000 roll and create, let’s say 100 different data set. These data set are replicas of the original data sample, but some of this data repeated because you are sampling with replacement. So the distribution of data will likely remain the same because after all, they are a sample from the same data set, but they will not be exact replicas of the original dataset. So they created some noise, which has the same distribution as the original data. And you will train one model on one of this replica and the other tree will be on the second replica until you get 100 trees or train on some different sampling of the original data set.
Ernest Chan:
And so you don’t trust any one of them, you will trust only the average prediction from that 100 trees. And so this deliberate introduction of randomness is able to overcome the fact that you put too much faith in the past because now the past is no longer the same past, you have 100 different paths, and that apparently have allowed the machine learning algorithm to not focus on some features that are unique to that data set. It will need to appear in multiple data set for it to be picked up as a repeatable pattern. So that’s, how random forests are able to reduce overfitting.
Ernest Chan:
Now, the other thing, which we’ll talk about, which is that what about this coincidences that happened in the past, and that would not occur again. And in the investment context, that’s a very well known problem. So let’s say we are selecting investment managers, and we look at their past track record. Let’s say you have 10,000 investment managers that you’re interviewing to manage your investment. Well it could be that this could be just 10,000, stock phone monkeys, and one of them and I have done the simulation one of these monkeys, if you give them three years time… one year track record, let’s say you only look at one year track record and you have 10,000 monkeys to try to generate this one year track record. One of them is going to generate a sharpe ratio higher than two. That’s no problem if you’re just looking in the past, it’s very easy to find this monkey that has a sharpe ratio of two.
Ernest Chan:
And that is the same problem with machine learning. If you look hard enough, you’re going to find the pattern that is like huge sharpe ratio in the past, but how are you going to make sure that it’s going to perform in the future so that we ties to the out of sample test. But the problem is, in machine learning, you actually have more than 10,000 monkeys, you could have 100,000 monkeys that you’re looking for the pattern. And so one specific out of sample test is no longer enough because even with an out of sample, chances are out of 1 million monkeys, you’re still going to have one that can pass the in sample, test and still do well.
Ernest Chan:
So in machine learning, we have this technical cross validation, which is essentially dividing the data into many parts. And you will look at and you will, at each iteration, you will exclude one part and leave that as the out of sample. So you essentially created many different out of sample data set. And the monkey has to perform well in, on average on all this out of sample data set before you set the model. So, that is a way to prevent this kind of problem where, where, where you can essentially find the sharpe ratio to model just by random So, and that’s the second events in machine learning to prevent overfitting.
Jason Buck:
And then, to complicate even a little more, let’s just say you had 1000 random tree samples and you’re looking for a high probability trade of have a high expected value. Isn’t it possible sometimes that given the weather markets and the current state, you could have a hundreds of those random trees light up, and that exhibits a positive expected value trade for you. So you’d make that trade, but then maybe a year from now, it could be a separate several hundred trees that light up that symbol an expected positive value trade for you. So it’s not like you’re looking for the same random trees or the same probabilistic paths to light up it can be different paths at different times given different markets. Is that correct?
Ernest Chan:
Well, yes, I mean, typically, we retrain the machine learning model regularly, just to take a new data. And every time you take a new data or we train the model with a different random seed, when a new initial random configuration, you get completely different set of models. So one test of the robustness of your machine learning strategy is that to see if the performance will vary drastically, when you take in new data and retrain the model. Hopefully it doesn’t.
Jason Buck:
Going back to the meta-labeling, is this a good example I think about Dan Rasmussen and Verdad. The idea of meta-labeling is like, if you just want to know if the markets going up tomorrow, that’s maybe too difficult. But if you can niche down to very specific things, like I believe Verdad looks for, if a company is paying down, it’s a loan structure, they can predict that using machine learning that pay down of the loan structure, and that’s where they look for out performance of those stocks. So is it part of meta-labeling is like niching down to a very specific example that you’re looking for and that gives you a broader set of where you could have a probabilistic higher expected value trade.
Ernest Chan:
Actually, I would interpret it as machine learning on a human construct a strategy instead of directly on the market. So, it’s not so much narrowing it to a particular niche, but it is really using private data instead of public data for learning, because if you are applying machine learning to public data, like whether marking go up or down, everybody has the same data. Right? I mean, I don’t believe that my machine learning model is better than Citadel or better than Renaissance technologies are better than Goldman Sachs. No, I don’t think so. So if we are all trying to predict the market go up, I think randomness on top we have a better chance of predicting accurately where the market go.
Ernest Chan:
But that’s not what I’m trying to use my machine learning to do for me. I have a simple strategy in this case, Tail Reaper, intraday, transforming strategy. Maybe Renaissance has one, maybe not, I don’t care but because they don’t have my exact data that I’m going to learn from and with that, I’m going to apply machine learning to learn from my private data to learn whether my strategy do well or not for a particular day. And so that no one else is learning from my model only I am. And I’m trying to only beat my own base model. I’m not trying to beat the market, I’m trying to improve my own strategy using all these features. And that has a much better success rate than if you’re trying to beat the market because everybody’s trying to beat the same market, but very few people hopefully only me are trying to beat my own model, so to speak.
Taylor Pearson:
I know you’ve talked about before you almost hear it more as almost like a risk management tool, the actual you’re using, as you said, a fairly simple intraday trend following strategy, and then you’re incorporating the machine learning as efforts for a risk management or sort of how do you think about it? It’s not generating the trade ideas, so to speak, like you said, you’re not just training it on market data, it’s helping to improve the strategy. Could you just speak a little bit more to that?
Ernest Chan:
So, like I said, the main use of machine learning is to look at variables that you have not taken into account in your original strategy because in the original strategy is a very simple transforming pretty much technical indicators are the inputs, right? We are not looking at non [inaudible 00:35:23] payroll numbers, we are not looking at how code performed that day in order to generate this trend following signal. But that’s exactly what we should look at in the machine learning layer. Because for one thing you know our model typically perform poorly when there’s a pool market. Because there’s no tail risk to hedge in a pool market, very low volatility it’s unlikely to have a tail move. And so one of the variable that we would want to look at is volatility, you need all its manifest forms, in high volatility the historical volatility [inaudible 00:36:00] or whatever you can come up with measure volatility.
Ernest Chan:
And in the original model, it is too complicated take into account all these different forms of volatility. But machine learning can do that easily. And so we will enter that and also many other variable globally. It could be that exchange rates might affect the profitability of this function, it could be commodity price would affect it, who knows what. We don’t know what variables are important. So we throw it all in the machine learning, and then it will try to learn how this large set of predictors can be used to reject or accept your original models trade. And that was what we use I think successfully in the last year.
Taylor Pearson:
Maybe I’m oversimplifying this incorrectly but the initial idea is just as you said, a sort of a simple trend following strategy and you’re passing it through this layer of the machine learning that’s kind of saying, like do we want to assess the position sizing or the probabilities based on this simple model that we’re using that to sort of adjust based on our past performance in the meta labeling technique?
Ernest Chan:
Yes, that’s essentially. Yes.
Taylor Pearson:
And yeah, I guess you’re sort of sticking on the tail approach, I think one common conversation I’ve had talking with quantitative investors about tail strategies in general is, well, if a tail strategy, tail risk happens once every five years, and you have 100 years of data, you’ve only have 20 data points, so how can you take a quantitative approach to tell us there’s just not enough data? How would you? Would you try to comment on that and how you think about it?
Ernest Chan:
Yeah, so our strategy so is called a Tail Reaper Strategy actually trade more than what one would expect a tail strategy to trade At least in the base model. So in the base model, I would say that we have at least a quarter to one third of the days, we have trades, which is actually, in my opinion, far too often, as you pointed out catastrophe or global crisis don’t occur one third of the time, every year. So that’s why exactly, in our opinion we trade too often, but that’s good, because they provide enough data point for us to learn from, for the machine learning to learn from. And so the machine learning is going to learn from this overabundance of trades and to tell us which one you might as well skip doing it. And so after we applied the machine learning layer, the number of trades become far fewer and so you can really learn from those trades. But fortunately, we internally run our base model. See we still have our Internal base model to run and that generates hypothetical trades for our machine learning model to run on, and that is generating sufficient data for training.
Taylor Pearson:
Yeah, it’s interesting. There’s always used to be how far out into the tail counts as the tail, it’s like what point do you draw the line where like this is the tail and this isn’t and I find people tend to get hung up on that. And Jason, I think you’re going to say something.
Jason Buck:
Yeah, part of that. I think to piggyback what you were saying is that if people are trading tail risk put options, it’s very different than monetization schedule that where you’re trading intraday, the E-mini futures, the S&P futures, you can follow that move from a small move to a big move to a huge move, and as long as it keeps running intraday, you’re just trend following that move. And you don’t REALLYU necessarily monetize your positions because IV is expanded or you’re worried about your data necessarily, especially on an intraday trade.
Ernest Chan:
Yes, it has a complete different profile from holding put because holding you’re not putting on a tray essentially, you are holding a position. It’s like a buy and hold position, essentially. Whereas here we are on a as needed basis. If there’s no term all[inaudible 00:40:20] no moves, don’t trade, no position and no trades. And that saves a lot of premium decay that way, especially with the help of the rejection of a trade by machine learning.
Jason Buck:
Let’s just say the market selling off and you’re short the S&P. Do you have any of that sort of trailing stops or you ratchet up the stop behind as the move grows larger and larger, or how do you think about monetizing or getting out of the trade?
Ernest Chan:
Well, there is definitely a stop because not every trend is a trend. You thought is the trend and then it turned into mini version. So when it does that, we get out, so yes, there’s a stop loss for sure. But one of the, I think, crucial benefit of a trend following strategy is that there should be no limit to the upside. So for example, in the last few months, we have days that were greater than 10% move per day. And in the past history, there are days that move over 20% the equity index. And so let’s say we get into a position when the market is up 1% or down 1% moves more often down 1%. If we say oh we are going to have a profit target of just make 2%. Well as the market keeps sliding, you will get out at when the market is minus 3% and you make 2% you might be happy, but that’s not a contract[inaudible 00:42:03] strategy because the [inaudible 00:42:04] strategy is supposedly supposed to have unlimited upside and limited downside, but unlimited upside, so, we do not impose a profit target. So if the market ended up, down 20% we will make 19%.
Ernest Chan:
So how do we decide to get out? Well because of the rationale for this strategy, because of the particular market dislocation or opportunity that we try to capture, we always exit at or before the close of a cash market, because that’s when rebalancing a portfolio stops. People have to be balanced by the cash market close to maintain the leverage that they are allowed to have. And so after that, nobody cares, you make money or lose money, frankly, nobody cares, it’s not in the legal document, that you cannot lose money after that time. So they are fine with that. And so we also exit at that same time, as all these others.
Jason Buck:
I’m glad brought back up almost that end of the day function because when we’re talking about you brought up earlier, like creating robust models or robust ways of implementing trades, I think historically when CTAs were fairly algorithmic, maybe before machine learning or AI came onto the scene is historically a robust algorithm would mean they could apply it to multiple markets. But if you apply your E-mini strategy or S&P strategy to multiple markets, it actually doesn’t work across multiple markets. And previously, a CTA would have thrown that out historically, because they wanted to see it work across multiple markets. But I think what you’re hinting at and I want to maybe dive into a little more is that because you have these certain dynamics of the S&P markets with the gamma dealer hedging, or the market on close or end to day dynamics with ETFs then it actually is a robust model and it only can apply to the S&P.
Ernest Chan:
That’s right. So yes, I have been bought up since my days in a big investment bank, your strategy has to apply to 52 futures and all work on all of them before we can trade it, I said, I have never understood that question now. Because I believed that every market has its unique. You cannot trade the corn market the same way as you trade the gold market. It just doesn’t make any sense. Just look at the corn futures, they don’t have a computer every month that expired right, because of seasonality, how can you attack the corn market using the same model as you attacked the gold market, which is essentially financial futures, not not even the commodity futures. So, it has never make any sense to me that someone would say that this model has to work on all 52 futures for it to be considered acceptable. I have always prefer to specialize in particular market. I want to develop a special model for gold, special model for corn, special model for equity index and in this particular case, the specific market opportunity only exists in the S&P index, not in corn, not in gold, not in anything else. So we will exploit it.
Ernest Chan:
If a investor say, “Oh no, but that’s undiversified,” I say well, you are not forced to only invest in our strategy. There are lots of different strategies, CTAs funds out there that you can invest with. Well, if you want diversification, invest in all of them, just allocate some to this strategy, which is good at what it does and it doesn’t claim that it can work everywhere else.
Jason Buck:
Exactly. And so, part of that too, before somebody else tries to run out and do this at home, because we brought up gamma dealer hedging or in the ETF structures at the end of the day and then market on close orders, which can come from [inaudible 00:45:56] funds or other places, those things don’t necessarily always exacerbate the move or accelerate the move because they can be mean reverting especially gamma dealer hedging can create a mean reversion scenario until you get out to the extreme and that rubber band breaks and then accelerates the move. So when you have all of these competing forces, it doesn’t necessarily mean that move is going to trend and continue. It can also mean revert intraday on you and I think you’ve talked about historically which months have been where maybe volatility picked up and so you were able to get into more trades, but then it mean reverted more than was expected. But that’s just kind of what’s expected from the model and the most of time it’s going to mean revert on you even intraday before we get those breakout trends.
Ernest Chan:
That is very true. And so what I described, the strategy was actually kind of our first version back in 2012. As time goes on, many bells and whistles have been added. When do you get out. When do you get out early. Should you get in more importantly, at what time you get in, under what condition you get in? That has all been added over the years. But particularly the major advance was machine learning which is we are not just talking about S&P market, we need to look at globally, all the other markets how they behave in recent days, before we decide if today is going to be truly a trend following day or mean reversion date. And so the art is not to decide when to trade because but decide when not to trade, when do you reject the simple strategy suggestion?
Jason Buck:
And then, if you think about I hate to use the term average because like you’d say the average time of trade is a few hours. Do you find in the last… that changes over time or maybe in the last year, year plus, you’ve been maybe trading primarily through the end of the day or right right into that market close and maybe a few years prior to that you’re trading at market open and for the first hour of the trading day, how does that change throughout time or in different markets?
Ernest Chan:
We don’t actually see a big move or a consistent pattern over the years. What is the best time to enter? I think that it is more of a question of what is the best regime where this strategy would work, then what is the best time of the day to enter. So the time of the day to enter we believe that we have consistently find better entry conditions, but that better entry condition actually is applicable to five years ago as now. It’s just that we only discovered now, but it doesn’t mean that it suddenly appear now and wasn’t working five years ago. So, that sort of drift of the entry condition actually was not discernible from what we observe.
Jason Buck:
And then over the last decade or two, it looks like most of the down moves in the S&P were happening during the intraday session and then late 2019, 2020 that kind of flipped and a lot of the down moves were in the overnight session, yet you guys were able to have a great Q1 of 2020 when we kind of attribute to this as other managers that maybe just traded short intraday, weren’t able to quite to have much as high returns in Q1.
Ernest Chan:
I think it has been establish that the biggest part of the daily move is during the overnight session. Nobody can dispute that. It’s just playing numbers. But interestingly however, when there is a true tail move, the tail move never stops when the regular market opens. It always going to end up continuing that trend and that is for simple reason that most of the market liquidity in the US can only execute, especially if you’re a stock investor. If you’re not if you’re a future investor, maybe you can unload your futures overnight. But if you are a cash equities investor, the only way you can lighten up your portfolio in any sort of size is during the regular trading hours. And so when there’s true panic sets in, and when portfolio need to really rebalance, it has to happen in the regular hours and we don’t care that on average, yes, most of the return happened overnight. We only care that on those severe panic moments they are offered to be captured, they’re trying to be captured during the regular trading hours.
Ernest Chan:
And because we are levered, we are not trading with one times leverage on the S&P we’re trading about 4.5 or 5 times the average. A little bit of return during the regular hour, even if only one third of the daily move happened during regular market hours, where as two third happened overnight, because of that five times leverage we are going to recover the entire daily move and more than the entire daily move because our leverage. So that’s how we can capture this high return despite only trading a few hours a day.
Jason Buck:
How do you think about, you mentioned earlier the Forex Strategy you were using in 2011 maybe the Kelly optimal leverage was 14X or set I don’t know if that was the the Kelly number but I guess the the idea of Kelly criterion is very interesting. I was seeing a game like blackjack or roulette post something where the the odds are known. It’s a lot easier to sort of employ that strategy and talk to poker players for example often say like, I use Kelly, but I use half Kelly, because maybe there’s some things that I can’t predict and so I don’t know the exact odds, I’m not able to do it. I’ve seen Yeah, yeah, me many investments are you like, Oh, you should run this at 20x leverage is the Kelley optimal approach. But yeah how do you think about I guess just what the appropriate role of leverage is in general and then maybe as it relates to Kelly or whatever else you used to think about it?
Ernest Chan:
Yeah. So I think when one apply Kelly formula to a continuous market, like finance, there’s a big problem in what distribution we assume, if you assume normal distribution of returns, Kelly typically come out to be very large number if you trade and you think that the returns of your trades is normally distributed, and you had some edge in your trading maybe average return 10% your Kelly leverage will be very high. But because it is based on an erroneous unrealistic distribution, there are some trades in your strategy or in the underlying market that is going to be outliers, 6 sigma, 10 sigma, maybe 20 sigma and those are not taken into account by Kelly formula and so that gives you… If Kelly formula for example, to take into account this 10 sigma events that happened this year, it would not assume a leverage greater than I don’t know, it cannot possibly have a leverage of greater than 10 because that would immediately bankrupt the account.
Ernest Chan:
If a daily move is 10% and you leverage 10% you’re done at the end of the day, you wipe out, the accounting is zero. So, if you look at historically, I think during the Black Monday I think in 1989/87, the market moved 20%, dropped 20% in one day. So if you take into account those days, your S&P strategy cannot be levered more than five times. So you’re clearly just talking a continuous cut plucking a Gaussian assumption in your Kelly formula leads to disaster and that’s what I’ve learned over the years. So what we do now is to always look at the worst day in history, that this strategy can happen and assume that you can be probably two times worse than that, or whatever you observe as the worst is not the worst case. The worst is really in the future. And that will be probably two times worse than what I’ve seen so far. And that really is the limit that you… that really sets the limit of your leverage.
Jason Buck:
It makes me wonder, though, when we think about the origins of Kelly criterion. It came from a logic or Elliott story[inaudible 00:55:12], environment, I mean casino games, dice throwing, etc. Where you knew your return, you knew your variance but when we apply that to markets, I mean you can look at your back test and you can know your return and your variance. Like you were just alluding to, you could see at 6 sigma or 20 sigma then that’s not in your back test. So, can you really apply it to markets on a walk forward basis when you don’t know your churn, you don’t know your variance and you don’t know your absorbing barrier. But are we fooling ourselves or by the additions of if I change my worst case draw down I put a stop loss there are we adding are we bolting on too many things to Kelly where it’s no longer Kelly? Is that a fair comment on trying to apply Kelly criterion to an infinite game versus where it was meant to be applied in finite games.
Ernest Chan:
Yes. So for me, I regard the number that come out of Kelly criterion whether you assume a Gaussian distribution or some small sophisticated fat tail distribution, take your pick as the upper limit of leverage, okay. So, you should not lever anywhere near that, preferably you are going to be half or even lower. So, that is sort of the maximum. It establishes a maximum, which is good, right? We want to know the maximum because amazingly, there are certain levered ETFs out there that plainly exceed the Kelley leverage.
Ernest Chan:
I’ve written a blog post some years ago already, I said this ETF are prime for extinction, because they have plainly exceeded the Kelly leverage at three times. And sure enough, some of them have gone out of business. So that is sort of a sanity check. So we do not use Kelly leverage nowadays. When I was younger, and more naive I use it but no longer, we no longer use it to set our leverage. But we do look at it to set our maximum effort if our leverage is higher than what Kelly suggests definitely something’s wrong with that.
Jason Buck:
That’s a great way to think about it once you get past full Kelly it’s just a matter of time before you’re bust or bankrupt. So that leads me to something I think about far too often right? That’s always on my mind and I think that you have to work with this as well because not only do you have Tail Reaper, you have other strategies like your VIX timer and combining them into Chimera. And so you can kind of apply Kelly and look at Kelly when you have mean reverting strategies when you’re doing a short volatility strategy with VIX and that’s more of a mean reversion strategy. But then when you combine it with a divergent strategy, like a trend following strategy on Tail Reaper, the math kind of no longer applies. So I’m wondering how you discretionarily could when you have in your own personal commodity pool operation, you’re combining multiple strategies.
Jason Buck:
So I’m just wondering how you have to almost take off your math hat and your machine learning hat and you have to start thinking discretionarily about worst case scenarios or how do I combine two strategies where one might… maybe not Kelly, but maybe I use Kurtosis or Clichy curves and I have an endpoint for my absorbing barrier, I can get a better idea that mean reverting strategy, but then I have to pair that with a Tail Reaper strategy that’s divergent and may not trade for months, but then has huge convexity in its returns to balance out the mean reversion. I’m really curious about how you think about combining these convergent and divergent strategies into an overarching portfolio.
Ernest Chan:
Yes, so clearly the first concern that we would have in this combination, that whether the tails… where the overall portfolio is still contracts, right. So that’s an overarching concern, but just because the overall portfolio is contracts doesn’t mean that it has attractive mean return. So after we sorted that out, after we impose the constraint that it has to be convex, we now need to work on increasing the average return and that’s when we add other strategies such as shortfall as a strategy but you do not add so much shortfall as a strategy that it overcome the convexity of the overall portfolio.
Ernest Chan:
So that’s how we look at it. And so whenever we add strategies we always keep an eye on whether it neutralize completely or the tail hedge property of Tail Reaper and we keep the allocation below that based on for example, historical track record or even back test.
Jason Buck:
With what you just said, I wonder too, if a lot of times when you combine mean reversion and divergent strategies especially if you have the convexity on like the Tail Reaper and divergent strategy, is it wrong to kind of assume that you would net out that like divergent strategy with that convexity is always going to overwhelm the mean reversion strategy if combined in the right proportions, but you don’t have to necessarily worrying about them negating each other as much as one has much farther to run and can accelerate into that complexity.
Ernest Chan:
Well, yes, if you are wanting a mean reversion strategy that is true because mean reversion strategy it can lose money but usually you can apply a stop loss or you can delever a mean reversion strategy will start to run into a long and deep draw down, but they are short in ply policies strategies, short option strategy that can actually explode many times to the downside and we need to keep a close eye on that to not to over allocate to those even though it might have an attractive return during the bullish months.
Jason Buck:
So actually to clarify that, a linear mean return reversion strategy-
Ernest Chan:
Yeah, linear that’s right. Linear we are okay with.
Taylor Pearson:
Well, I think Ernie that was great speaking with you, covered a lot today but just for anyone that’s interested in learning more about you or QTS or Tail Reaper, what’s your the best place for people to get in touch with you or find out more?
Ernest Chan:
Well, I have a website qtscm.com. That lists all our track record and some education material, but actually I also write books and published many articles and papers and the best place to get the entire output that I have publicly available my personal website, etchan.com. So that has everything. It has links to all the things that I’m doing, and it’s a good place to start.
Taylor Pearson:
Great. Well, thank you very much, Ernie.
Ernest Chan: Thank you, Taylor.
Taylor Pearson:
Thanks for listening. If you enjoyed today’s show, we’d appreciate it if you would share this show with friends and leave us a review on iTunes as it helps more listeners find the show and join our amazing community. To those of you who already shared or left a review, thank you very sincerely. It does mean a lot to us. If you’d like more information about Mutiny fund, you can go to mutinyfund.com. For any thoughts on how we can improve the show or questions about anything we’ve talked about here on the podcast today drop us a message via email. I’m taylor@mutinyfund.com and Jason is jason@mutinyfund.com or you can reach us on Twitter. I’m @taylorpearsonMe and Jason is @jasonmutiny. To hear about new episodes or get our monthly newsletter with reading recommendations, sign up at mutinyfund.com/newsletter.
