Ernest Chan - Machine Learning and Trading Strategies

[Music] 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 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 in the s p index we go into 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 also go into the evolution of machine learning and how it is applied to trading strategies including ernie's 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 enjoyed this conversation as much as i did so ernie i'd love to start you know one of the things you've you've spoken about that i find very interesting and also very you know counterintuitive um is how is it possible to have an investment strategy that has an arithmetic return of zero but that also improves a portfolio just to give an example of an arithmetic average i mean if we started a strategy on you know day one with a hundred dollars of year one then ten years later you know the ending balance of that strategy is also a hundred dollars um you know it hasn't it has an arithmetic average over that time period say is uh is zero um how yeah how how can a strategy like that improve a portfolio well such a strategy would need to be combined with another strategy that it is you know maybe anti-correlated with or maybe even uncorrelated with in order to decrease the overall risk of the portfolio um so you know essentially the um the growth of wealth is a compounding process right it is not you know we invest uh you know a million on the first year and the second year let's say go up 10 become 1.1 billion but then you know if you keep going at the same rate it is compounding it's not just going to get you know at extra fixed amount of dollars every year it's supposed to add more every year so compounding has um has a very interesting effect that the um 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 the lower the risk of that portfolio the higher the compounded growth rate and so if your you know strategy that has zero arithmetic mean is applied to another strategy that has a you know a small positive arithmetic mean but that are uncorrelated with your first strategy the overall portfolio will may have a low risk because they are anti-correlated so some of the fluctuations in the return are smooth 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 growth depends on the fact that your net worth doesn't suddenly have a sharp drawdown it won't he it loves consistency it loves smoothness and so if your return are positive every month it beats having one month with a big positive return and another month have a big negative you know negative return and that's how you know adding a zero return arithmetic return strategy to another strategy that has a small positive return can actually bring about a higher compounded growth rate um you know that's essentially the solution to the paradox yeah i i find that so you know we talk about interior sometimes like volatility drag you know you want if you're able to combine two things and it reduces sort of that that volatility element if you yeah if you go 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 those big drawdowns are very painful um so maybe to go from there uh i would love to hear just a little bit about um you know the origin of your tail reaper strategy um you know sort of how it works at a high level sure so um we started our fund in 2011 and at that time we were trading a sport forex strategy only nothing to do with tail ripper it was a fairly high frequency market making strategy in the in the sport forex space and we had a great one in the first half of the year until you know even july we had a great run we practically have no losing trades and we were very you know 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 you know 200 plus years and that hit us really hard it it costs 35 percent drawdown you know portfolio you know given that we elaborate 14 times so you know we were you know certainly shells shocked at that time but we were thinking how are we going to protect ourselves from these kind of tail events that never happened once in 200 years um and we look and look and we came up with this strategy that turned into today's tale reaper strategy it wasn't called terrible at that time but essentially the same strategy we tried today and that is an intraday trend falling 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 done and hope it goes lower extremely simple and we only hold precision for a few hours a day and with this simple strategy obviously we add the spells and results over the years but that was the core idea with this extremely simple setup it had we've threw two you might call extremely turbulent quarters since then the first one was the 2015 fourth quarter chinese stock market meltdown which affected u.s equities greatly and we you know generated a decent return in that quarter and then the second test was the 2018 first quarter former garden where we had the opposite problem it's not that uh the economy is going poorly it's going too well and people were afraid that this fed is going to increase interest rate so you know some volatility level volatility exchange traded product actually went out of business in that in that quarter because of the surge in in volatility so we were happy with the strategy continue to run it and of course this year we had a truly unprecedented event happen so that is that is the core strategy it's a very simple transform strategy but what distinguishes it from other telehealth strategy is actually what we do to prevent it from losing money when the market is bullish and calm but that's that's a separate layer that we apply to it using machine learning you know i guess this first one to start in terms of the tail reaper talking about you know you use a trend following approach and you know trend falling is some variation of you know you buy something that's been going up there you sell something that's been going down i think you know most people think of trend following especially in sort of the commodity's future space the ctas tends to be you know longer term trend followers that are using you know one month two month uh year-long sort of um timings in terms of when they're when they're going you know how they're going long or short you know i know you're doing um you know purely sort of an intraday approach and so it's you know a matter of hours as opposed to a matter of you know days or weeks or months but yeah i'd love to just comment on sort of your you know your approach to to sort of trend following and i want to get into you mentioned the machine learning and how you think about sort of the the risk element there but maybe just start with with trend following in general in terms of the approach you take and how it's maybe different from what people typically think about when they hear the term trend following that's right yeah i mean traditionally um yes most cta transforming strategy uh has a faller trend that has been established over multiple months a very favorite time frame would be one year would be the typical time frame for transforming but we choose to be intraday because we also observe that trend has been getting shorter and shorter um that's that's you know one of the reasons why i think it's advantageous to have a much shorter um transform you know much shorter time frame transforming strategy but also our transforming strategy take advantage of a certain market dislocation or you know market opportunity that most other transforming strategies are not utilizing and that is the phenomenon of forced rebalancing of level portfolio so you know to explain it better uh you know imagine that you are uh holding a um three times leverage uh etf that is based on the s p 500 um if you know one day the market goes up you know two percent all of a sudden your etf is losing six percent not two percent but six percent because of the leverage and if you don't do anything at the end of the day that etf will be over level because your equity has dropped due to the big loss um and uh you know the equity has dropped much more than the market value of your portfolio has dropped so you're over level and as the um as a sponsor of that etf you have to sell some holdings in order to reduce the leverage and it's the same applied when the market went up a lot so because of all these leverage instruments in the marketplace not just etf but you know 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 there are another phenomenon which is gamma hedging so many option market maker or you know market maker for um foreign swaps let's say the a big bank sell if hedge fund equals three swap and then they have to hatch the gamma of that volatility swap they also have to trade in the same direction as the market movement to to limit their risk to hedge the risk and all these hedging and rebalancing activities drives the the trend uh it exacerbates the market moves and that's what we're trying to capitalize it but this all kind of hedging like this only occurs in your day you know right after the market close is a new day nobody cares anymore and so we do not expect that the trend to last beyond the market close we have to capture it intraday and that is one of the main driver of the of the offer of this strategy and so you think this strategy is very like it is based on i mean for lack of a better term sort of current market microstructure in a way it's part of the way in which um you said like gamma dealer hedging and then also um different leverage players aren't saying the portfolio insurance analogy is interesting right you think i was there's there was a lesson that got learned there and maybe that has that lesson just been forgotten or you know why what is sort of the rationale of like why why they're why things are behaving that way why players behave that way under certain incentives or how do you look at it well i don't know that the lesson is forgotten but we have a new class of instruments that is such as the level etf that they have no choice but to do portfolio insurance it's mandate it's it is in the prospectus of the offering that they have to do it so they have no choice and then i guess the next sort of area i want to get into you know you mentioned that the way you approach this is using exclusively you know s p futures e-minis um is your instrument so obviously tail risk can show up in you know lots of different places it can show up in equity markets in other countries and bonds and uh forex whatever you know why why sort of e-minis exclusively what is it about that that makes sense to you or how do you think about it yes so this phenomenon is most prominent in the s p because of the large amount of derivatives that are tied to the s p um you know you they are not an equal amount of notional derivatives exposure to um the you know the tax index or to the nikkei and so forth i mean there are some naturally but nothing comparable to that tie to the s and p and similarly um for the leverage etf um the the aum of leverage that are tied to european or asian indices are minimal compared to what's tied to the s p even within the u.s market the amount of level utf 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 very well in other markets uh that's interesting i think i said i'm sure probably using this term improperly but sort of a market microstructure function there of how sort of things evolve around the um the s p i guess next i want to talk a little bit about you know machine learning you've written a number of books on machine learning you've been very involved in the field um you know going back into the uh the 1990s and i want to talk sort of about how you you think about incorporating machine learning into your your trading approach and sort of the tail reaper strategy um specifically but maybe just to start you know how would you explain machine learning you know in general to um you know someone that isn't familiar with what are some common common examples you know the types of problems that could be in trading or otherwise that you know machine learning is is good and bad at and then you know how it compares to i think people hear other terms around like ai like deep learning or neural nets and those sorts of things that you can maybe paint to sort of a picture for for how all that fits together so i think machine learning is a technique which allows us to examine a large number of potential predictors whether of the market or of um you know what the the speaker is going to speak next or you know or to predict where whether pedestrian is going to cross in front of a car you know in terms of in self-driving car contacts right so um traditionally uh you know in investment or in traditional quantity finance we use is 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 earning ratio you look at book to price ratio book market ratio whatever you know four or five factors and you try to use that to predict the return of a song you can say that it's 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 right you know well maybe not value at the moment but you know we know that um you know after careful in-depth research many finance professionals believe that these four variables are likely to be predictive of future return so you actually don't need the machine machine to tell you what variables are important you know already know that all you need is a actual magnitude prediction based on those four variables that is a kind of 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 you cannot possibly examine you know no finance professors no matter how genius he or she may be can examine one thousand variable in depth to see if any one of them can predict the stock market it's two too many variables and that's where machine learning can help because there are established algorithm that can handle one thousand or ten thousand 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 i think you know one uh one topic you've spoken to a lot and i think one of the reasons you uh you know went through a period of your career where you weren't actively involved in machine learning is this idea of um of overfitting right you know if you have a i guess the example i like is i think there's there's a very strong correlation between uh nicolas cage movies released in a year and swimming pool deaths right you know so you could say you know when when nicholas cage releases more movies you should be long swimming pool fatalities and when he releases less you should be you know short swimming pool fatalities but obviously you know that's that's a spurious correlation that's uh you know towards just cleaning data how maybe you know kind of looking back at the history of machine learning in general and as it relates to overfitting in particular you know uh how how has that changed over time you know how are the approaches changed why does it sort of make more sense now than maybe it did in the 90s or the 2000s yes so um indeed you know the way to reduce overfitting has been one of the most active research topic 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 but it it you know the whole reason where you know neural network wasn't able to be wasn't really commercialized until the last few years is this overfitting problem i mean you'll never it's not new it has been around for at least i don't know 25 30 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 to reduce overfitting to pass data that the performance out of sample performance suddenly had become tolerable and and the performance on you know business and commercial problem actually become adequate and so you know i i would say most of the advances in machine learning in the last 20 years is is focused on overcoming this particular problem uh and you know for example we started in the early days when i was in moon sunny in the early 2000s um or actually late uh late last minute unfortunately uh it was on uh you know we we like to use regression trees and decision tree to make prediction but that that invites overfitting so you know gradually people change that approach to and and adapt it to call random forest where you basically create many random trees and average them so averaging you know and in an ensemble of learners is one way to overcome um overfitting you introduce randomness essentially in order to overcome overfitting you have to introduce randomness in the model sounds intuitive but that that is the case that's the same as the dropout technique where you randomly punch holes into the neural network in the random force you create many many random trees based on different randomized set of data in order to generate noise in them in the data and create models that are uh that are you know fitted to those noise but you average them so that you won't be um dependent on that noise so all these techniques have been uh you know perfected or refined in the machine learning community to today where the problem of overfitting has been greatly reduced compared to you know 10 20 years ago and become usable this techniques but that is for general machine learning of course with respect to finance in particular there's one realization that we're 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 market is going to go up or whether the stock is going to go up and because you know if there is a strong signal it is going to be immediately availa charged away by you you know by default usually you won't find a strong signal because if if it's so obvious it would disappear it's like the famous economist joke that if you see the market is so efficient if you see a ten dollar bill sitting on the pavement of economists won't pick it up because they don't believe that it would actually fix it someone else should have picked it up already that's right so yeah i mean the same applies to applying machine learning to uh to finance is that if the opportunity so office people would have already traded on it you don't need to wait for the machine to find it so one realization uh one particular good application of machine learning financing that has been talked about in the last few years particularly by dr you know marcus lopez de prado since his new book since his book was published was the technical matter 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 a little um less fear of arbitrage because everybody trading strategies are different and the machine is trying to learn from the trader's own track record rather than from a public data set where everybody can arbitrage on and that is a many people have found to be a much better use of machine learning in finance than just directly to predict the market 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 uh you mentioned sort of like out of sample performance just defining you know maybe giving an example of of out of sample performance and what that um you know use case or what that would mean and essentially you're using it and then yeah i'm also i'm very interested in this like the random forest idea of of using an ensemble of approaches you know i guess we talk a lot about um ensembles and diversification in terms of of investing but how does that um just could you speak to sort of that random force like what you know if there's any you know simplified examples that you might use in you know one of your books or when you speak about it how is that approach sort of different from what was done historically yeah so the random forest approach is essentially randomizing the data so you know originally you know you have let's say a thousand rows of data and you use that a thousand row to fit one decision tree that's it and you know and that tree can certainly uh is is going to capture all the non-repeatable patterns in that 1000 rows of data you know capture every wrinkle of that data and they think that wrinkle is going to repeat itself but for random forest they will what is called sample with replacement form that one thousand row and create let's say 100 different data sets these data sets are replicas of the original sample but they some of them are some of these data repeated because you're sampling with replacement so um you know the distribution of data will likely remain the same because after all they're sampled from the same data set but they will not be exact replicas of the original data set 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 another tree will be on the second replica until you get 100 trees or train on some different sampling of the original data set and so um and you don't trust any one of them you will trust only the average prediction from that 100 trees and so this introduction of 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 path 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 you know it will it will need to appear in multiple data sets for it to be picked up as a repeatable pattern so that's that's how random force are able to reduce overfitting now the other thing which you talk about uh which is that uh you know what about this um you know coincidences that happened in the past and that does not you know just would not occur again and in the investment context that's a very famous um that's a very well-known problem so um you know we uh let's say we are selecting investment managers right and we look at the past track record you know let's say you have one ten thousand investment managers that you're interviewing uh to manage your investment well you know it it could be that you know this could be just ten thousand start for monkeys and one of them and i have done the simulation one of these monkeys would you know if if you give them um you know three years time one year track record let's say you only look at one year track record and you have ten thousand monkeys to try to generate this one year trade record one of them is going to generate a sharp ratio higher than two that's no no problem if you're smoking in the past it's very easy to find this monkey that has a sharp ratio of two now uh and that is the same problem with machine learning you know if you if you look hard enough you're going to find a pattern that is like huge sharp ratio in the past but how are you going to make sure that it's going to perform in the future so that retires 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 um even with out of sample chances are you know out of one million monkeys you're still going to have one that can pass the in-sample test and still do well enough sample so in machine learning we have this technique called course validation which is essentially dividing the data into many parts and you will look at and you will at each situation you will exclude one part and leave that as 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 accept the model so that is a way to prevent this kind of problem where where where you you can essentially find a sharp ratio 2 model just by random so and that's the second advance uh in in machine learning to prevent overfitting and then to complicate it even a little more let's just say you had a thousand random tree samples and you're looking for a high probability trade of a high expected value isn't it possible sometimes that given the where the market's at in the current state you could have you know 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 uh well yes i mean typically we we train the machine learning model regularly um just to take a new data and every time you take a new data or retrain the model with a different random seed when you know initial random configuration you get completely different set of models so one you know 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 i mean going back to the sorry going back to the meta labeling is this a good example i think about um dan rasmussen and bird dad like the idea of metal labeling is like if you just want to know if the market's going up tomorrow that's maybe too difficult but if you can niche down to very specific things like i believe uh verdad looks for if a company is is paying down its 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 as a part of metal labeling is like niching down to a very specific um you know example that you're looking for and that gives you a broader set of where you could have a you know probabilistic you know higher expected value trade um actually i would interpret it as a machine learning on a human constructed strategy instead of directly on the market right 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 the marketing go up or down you know everybody has the same data right i mean i don't believe that my machine learning model is better than citadels or better than renaissance technologies or better than the goldman sachs no no i i'm i don't think so so if we are all trying to predict the market go up i think when amazon probably had a better chance of predicting accurately where the market goes 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 ripper 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 you know 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 and you're i know you've talked about before you know you you almost think of it more as a um you know correct me from as almost like a risk management tool the actual you know you're using as you said a fairly simple trend following intraday trend following strategy and then you're incorporating the machine learning as uh yeah for like a risk management or sort of how do you think about it and it's not it's not generating the trade idea so to speak like you said you're not just training it on market data it's uh it's helping sort of improve the whole the strategy could you yeah speak a little bit more to that right so um you know 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 it's a very simple trend for pretty much technical indicators are the inputs right we are not looking at non-farm payroll numbers we are not looking at how how gold performed that's that day in order to generate this trend foreign signal but that's exactly what we should look at in the machine learning layer because well one for one thing you know our model typically perform poorly when there's a bull market because there's no tail risk to hatch in a bull market very very low volatility is unlikely to have a tail moves and so one of the variables that we would want to look at is volatility you need always manifest forms implied volatility you know historical volatility guards follow whatever you can come up with measure volatility and you know and and in the original model it is too complicated to 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 variables globally you know it could be that um exchange rates might affect the probability of this strategy it could be commodity price would affect it who knows what we don't know we don't know a priority what variables are important so we throw it all in to 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 used you know i think successfully you know in the last year and so you're almost on i'm maybe i'm oversimplifying this incorrectly but you know you the the initial idea is just as you said sort of a simple trend following strategy then you're passing it through this layer of the machine learning that's kind of saying like you know is you know do we want to sort of assess the problem the position sizing or the probabilities you know based on this this simple model and that we're using that to sort of adjust based on our past performance and the med labeling technique yes that's essentially it yes um and yeah i guess you know sticking on the the tail approach i think you know one common conversation i've had talking with quantitative investors about tail strategies in general is well if a tail strategy you know a tail risk even happens once every five years and you have 100 years of data you've only have 20 data points so how you know sort of how can you take a quantitative approach to uh you know tail risk is just there's just not enough data how you know would you sort of comment on that and how you think about it yeah so um our strategy it's always called a tail ripper strategy actually trade more than what one would expect a tail strategy to trade at least in the base model um so in the base model i would say that you know 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 you know one third of the time every year right so that's why exactly in our opinion you trade too often but that's good because it provides 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 over abundance of trace and to tell us which one is actually going you 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't really learn from those trades but fortunately we internally run a our base model still we still have our internal base model to run and that generates hypothetical traits for our machine learning model to to run on and that is generating sufficient data for training yeah it's interesting uh there's always used to be it's you know how how far out into the tail counts is the tail you know it's like at what point do you draw the line or like you know this is the tail and uh and this isn't but uh you know i find people tend to get hung up on that and then jason i think you're going to say something yeah part of that i think to pick it back when you were saying is that if people are trading tail risk put options it's very different than monetization schedule that where when 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 you're not necessarily as long as it keeps running intraday you're you're you're just trend following that move and you you don't need to necessarily monetize your positions because iv is expanded or you're worried about your theta necessarily especially on an intraday trade yes it has a completely different profile from holding put because paul input you are um uh you know you you're not putting on a trade essentially you're 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 turmoil 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 and as part of the as you have your azure have that movement intraday in the market let's just say the market's selling off and you're short the s p um do you have any of the sort of trailing stops or you ratchet up the stop behind as as the move grows larger and larger or how do you think about monetizing or getting out of the trade well there is definitely a stop because not every trend is a trend you thought it's a trend and then it turned into mean reversion so when it does that we get out so yes there's a stop loss for sure but one of the um i think crucial benefit of a trend following strategy is that there should be no limit to the upside uh in you know so for example in the last few months we have days that were you know 10 greater than 10 move per day and in the past history there are days that move over 20 percent the u.s the equity index and so let's say we um get into a position when the market is up one percent or down one percent moves more often down one percent um if we say oh you know we are going to have a profit target of you know just make two percent and be out well you know as the market keeps sliding you get out at when the market is minus three percent and you make two percent you might be happy but that's not that's not a conflict strategy because the contract strategy 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 percent um so when how do we decide to get out well as we dis because of the rationale for this strategy because of the particular market dislocation um or opportunity that we try to capture we always exit at or before the close of the cash market because that's when rebalancing a portfolio stops you know people have to rebalance by the cash market close to maintain the leverage that they are allowed to have and so after that nobody cares you know you make money or lose money frankly nobody cares not in the legal document that you cannot lose money after that time so so they are fine with that and so so so we also exit at that same time as all these are and that's i'm glad you brought back up almost that that end of the day function because you know when we're talking about you brought up earlier like creating robust models of robust uh ways of implementing trades i think historically when ctas were fairly algorithmic maybe before machine learning or ai you know came onto the scene is historically a robust uh algorithm would mean they could apply it to multiple markets but if you apply your 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 of day dynamics with etfs then it actually is a robust model and it only can apply to the s p that that's right so yes i have been brought up you know since my days in a big investment bank so you know your 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 rationale because i believe that every market has its uniqueness you know you cannot trade the corn market the same way as you trade the gold market it just doesn't make any sense you know just look at the corn futures they don't have a confused every month that expire right because of seasonality how can you attack the corn market using the same model as you attack the gold market which is essentially financial futures not not even the commodity future so it has never made any sense to me that someone would you know say that this model has to work on all 52 futures for it to be considered acceptable i have always preferred to specialize in particular market i want to develop a special model for gold special model for corn special model for you know equity index and and in this particular case the specific market opportunity is only exists in the s p index not in corn not in gold not in anything else so we will exploit it um 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 to identify diversification invest in all of them you know just allocate some to this strategy which is good at what it does and it doesn't claim that it can work everywhere else exactly and so part of that too before somebody else tries to run out and do this at home because we brought up uh you know gamma dealer hedging or and the etf structures at the end of the day and then yeah mark it on close orders which can come from act 40 funds or other places is 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 median reversion scenario until until you get up to the extremes when 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 you can also mean revert intraday on you and i i think you've talked about historically you know which months have been you know 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 you know most of time it's going to mean revert on you even intraday before we get those breakout trends that 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 at what under what condition you get in uh that has all been added over the years but you know particularly the major advance was machine learning which is you know we are not just talking about s p market we need to look at globally all the other markets how they behave in in recent days before we decide if today is going to be truly a trend following day or mean reversion date and so um yes so the the art is not to uh decide when to trade because you know but decide when not to trade when do you reject the simple strategy suggestion and then if you think about you know he 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 you know that changes over time or maybe in the last year uh year plus you you've been maybe trading primarily through the end of the day or right you know right into that market close and maybe a few years prior to that you were trading at market open and for the first few you know the first hour of the trading day how does that change throughout time or in different markets um we we don't actually see a big move uh or a a consistent pattern over the years what is the best time to enter um i i think that it is more of a question of what is the best regime where this strategy would work then you know what what is the best time of the day to enter so the time of the day to enter um you know we we believe that we have consistently find better energy conditions but that better entry condition actually is applicable to five years ago as now it's just that we only discover it now but it doesn't mean that they suddenly appear now and wasn't working five years ago um so yeah so that that sort of uh drift of of the entry condition actually was not was not discernable from what what we we observed and then over the last decade or two it looked like uh most of the down moves in the s p were happening during the intraday session and then 20 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. um what do you kind of attribute to this as other managers that maybe just traded short intraday weren't able to quite to have much uh as high of returns in q1 yeah yeah so um yeah i think it is you know been established that mo the biggest part of the daily move is during the um overnight session that is no nobody can dispute that it's just plain 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 you know continuing that trend and that is for the simple reason that most of the market liquidity in the u.s 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 this 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 happen overnight we only get that on those severe panic moments they are offered to be captured they are trained to be captured during the regular trading hours and because we are leveled we are we are not um you know trading with one times leverage on the s p we're trading about four point five to five times average a little bit of return during the regular mile even if only one third of the moves of the daily move happened during regular market hours whereas two-thirds happen 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 uh high return despite only trading a few hours a day how do you think about uh you mentioned earlier the 4x strategy you're using in uh 2011 you know maybe the kelly optimal leverage was 14x or so i don't know if that was the the kelly number but um yeah i guess the the idea of kelly criterion is is very interesting obviously in like a a game like uh blackjack or roulette you know post something where the the odds are known it's a lot easier uh to sort of employ that strategy you know talk to poker players for example often say like why and i use kelly but i use half kelly because maybe i don't you know there's some things that i can't predict and so i don't know the exact odds i'm not able to to do it um you know and i've seen you yeah you know many investment traders like oh you should run this at you know 20x leverage is the kelly optimal um approach but yeah you know how do you think about i guess just just what the appropriate role of leverage is in general and then maybe as it relates to uh to kelly or whatever else you used to think about it yeah so um i think when one applied 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 a very large number you know if you you you trade and you you you think that the returns of your trades is normally distributed and you had some edge in your trading you know you have maybe average return 10 your cal your cali formula will be your kelly leverage will be very high but because it is based on an erroneous unrealistic distribution there are some traits in your strategy or in the real in in the underlying market that is going to be outlier six sigma 10 sigma maybe 20 sigma and those can are not taken into account by california and so that gives you a if if carry formula for example uh to take into account these 10 sigma events that happened this year and you know it would have a you know he would not assume a leverage greater than i don't know um it cannot possibly have a leverage of greater than 10 because that would immediately bankrupt the account you know if if a daily move is 10 and you leverage 10 you know you're done you know at the end of the day you wipe out the accounting zero so uh and you know if if you look at historically i think during the black monday i think in 1980 89 87 the market moved 20 drop 20 in one day so if you take into account those days your your your your s p strategy cannot be leveled more than five times so you know clearly just plugging a continuous you know cut plucking a gaussian assumption into california lead 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 it can be you know probably two times worse than that right you whatever you observe is the worst it's not the worst you know the worst is really in the future and that will be probably two times worse than what you have seen so far and that really is the limit that you you that that that really sets the limit of your leverage it makes me wonder though like you know we think about the origins of kelly criterion and it came from athletic or eleatory um environment i mean casino games dice throwing et cetera where you you knew your return you knew your variance you know 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 a six sigma or 20 sigma event that's not in your back test so is it can you really apply it to markets on a walk forward basis when you don't know your return you don't know your variance and you don't know your absorbing barrier but you know are we are we fooling ourselves or by the additions of you know if i if i change my my my worst case drawdown and i put a stop loss there are we adding are we bolting on too many things to kelly to where it's no longer kelly is that a fair comment on on trying to apply kelly's criteria into a to a infinite game versus where it was meant to be applied in finite games yes so um i for me i regard the number that come out of category whether you assume a gaussian distribution or some more sophisticated fat tail distribution you know take your pick as the upper limit of leverage okay so if you you should not level any anywhere near that preferably you are going to be half or even lower so that is sort of um the maximum it establishes maximum which is good right we we want to know the maximum because amazingly there are certain level etf out there that plainly exceed the kelly leverage i have written a blog post some years ago already i said this etf are prime for extinction because they have plainly exceeded the caddy leverage at three times and sure enough some of them have gone out of business so i i mean so that that is sort of a sanity check so we do not use it to we do not use kelly leverage nowadays i mean i 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 average if our leverage is higher than what caddy suggests definitely something is wrong with that level that's a great way to think about it because once you get past full kelly you it's just a matter of time before you're you're you're bust or bankrupt so but that leads me to something i think about far too often right that's always on my mind is 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 you know doing a short volatility strategy with vixx and 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 on tail reaper um the math kind of like no longer applies so it's like so i'm wondering how you discretionarily because when you have in your own personal commodity pool operation you're combining multiple strategies so i'm just wondering how you know you have to almost take off your your math hat and your in your m in 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 if i can maybe maybe not kelly but maybe i use kurtosis or cliche curves and i have a an end point for my absorbing barrier i can get a better idea of that mean reverting strategy but then i have to pair that with like 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 how you think about you know combining these uh convergent and divergent strategies into an overarching portfolio yes so um you're clearly the first concern that we would have in this combination are 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 a attractive mean return so after we sorted that out after we imposed a constraint that it has to be conflicts we now need to work on increasing the average return and that's when we add other strategies such as short volatility strategy but you do not add so much short volatility strategy that it overcome the convexity of the overall portfolio so that's that's how we look at it and so whenever we add strategies um we always kept an eye on whether the uh you know it added you know it neutralized completely or you know the the the tail hedge property of tail ripper and we keep the allocation below that based on for example historical um track record or or even back test um with what you just said i wonder too if you know a lot of times when you combine uh mean reversion and divergent strategies especially if you have the convexity on like the tail reaper or divergent strategy is it is it wrong to kind of assume that they you would net out that like that that one over like the 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 worry about them negating each other as much as one has much farther to run and can accelerate into that complexity well yes if you're running a mean referral strategy that is true because um you know mean reference strategy um you know it can lose money but you know it's it's you know usually you can you can apply sub-laws so you can de-lever a meaningful strategy we start to run into a long and deep drawdown but there are short implied policy strategies short option strategy that can actually explode many times to the downside so um and we we need to keep a close eye on that to not to over allocate to those even though it might have an attractive return uh during the bullish months yeah so i should have clarified a linear mean return return reversion strategy yeah linear that's right that combination yeah linear we are okay with yes well i think ernie that was great to speak with you i've covered a lot today but just for anyone that's interested in learning more about you or you know qts or tale ripper what's sort of the best place for people to get in touch with you or find out more well i have a um the fund has a website qtscm.com uh that listed uh all our track record and uh some you know education material but actually the you know i i also write books and publish many articles and papers and the best place to get the entire output that i have you know publicly available uh my personal website epchen.com so that has everything it has links to all the things that i'm doing and it's a good place to start great well thank you very much ernie thank you taylor thanks for listening if you enjoyed today's show we'd appreciate 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 to 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 mutinyfun.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 mutinyfun.com and jason is jason efun.com or you can reach us on twitter i'm at taylor pearson m e and jason is at jason mutiny to hear about new episodes or get our monthly newsletter with reading recommendations sign up at mutinyfun.com newsletter

Channel: Mutiny Funds

Views: 2,291

Rating: 4.9540229 out of 5

Keywords: Ernest Chan, QTS Capital Management, investing, stock market, portfolio diversification, mutiny fund, taylor pearson, jason buck, long volatility, hedge fund, ergodicity, VIX, volatility, futures trading, options trading, futures and options, finance, markets, economy, trading, investment portfolio, investment podcast, trading podcast, investment strategy, tail risk hedging, tail risk, trading strategies, machine learning, Quantitative and Algorithmic Trading, Ernie Chan

Id: NBDh7fvN53g

Length: 62min 17sec (3737 seconds)

Published: Thu Oct 15 2020