Peter Kambolin & Alexei Chekhlov
Adam: You run a wide gamut of strategies. You have HFT strategies, which many people say are all about the brute force of technology, but you also do sophisticated trading based on science and deep research. How do you marry those two?
Alexei: Even though we trade a variety of different strategies, in reality we are -- probably like many others actually in the systematic trading space -- focused on exploiting two main themes: mean reversion and directional trend-following. This is probably something you would hear from virtually any systematic trader. You may categorise yourself into either one or the other. For historical reasons we have been more focused on mean reversion over the years, and that is marrying well with our focus on shorter average trade holding times. So even though it may look like we use a lot of sub-strategies, if you bracket them, a lot of them will be mean reverting and some of them will be trend-following.
Within these large brackets, we do split into individual ways of doing things. We believe in a certain universality of description of price changes, which allows us to apply similar sub-strategies to a wide range of markets and holding times. This is also similar to how many other systematic traders operate.
Adam: Your firm describes the programs as being high-frequency and contrarian in nature. So is the focus on mean reversion what you mean when you say contrarian, given that mean reversion by definition suggests that a market will revert?
Alexei: In reality, most of the strategies which are not too complex are driven by correlation.
Ultimately it's a question of studying a time series, maybe even building a time series yourself as well, because nobody forces you to use the market-given time series. Maybe you'd like to construct your own synthetic time series out of the individual pieces the market provides you. That is also something we are interested in, which brings us to the notion of market neutrality.
Adam: What do you mean by market neutrality?
Alexei: Normally, conventional systematic futures traders just look at individual futures markets, they try to focus on futures markets like gold, copper or the US 10-year Treasury note. But if you have a more statistical point of view, you can introduce a statistical hypothesis that there is a similar statistical description of price changes across a wide range of short-term observation times -- say, from one minute to a few days in duration -- and also across a wide range of markets. Similarity here is understood, just like it is in statistical physics contexts, that the same algebraic laws are describing the random price change fluctuations, with possibly some features for particular market case constants. Such a hypothesis greatly simplifies the seemingly indefinite complexity of the market description. Instead of trying to find some features for every timescale and market miraculous patterns and rules, we can apply the same description across a wide range of timescales and markets and only focus on fitting a few parameters to individual markets. Obviously, the most difficult part here would be to adequately test such a hypothesis.
Moreover, taking the next step, out of these markets, you can create a synthetic market which will not have a directional exposure to, let's say a benchmark everybody thinks of, either the S&P 500 index or an index of CTAs. You can actually be neutral to one or several of these benchmarks as a byproduct of such a view as well. One of our programs, Systematic Alpha Futures Program, is going along exactly that path. It's both focused on trading mean reversion and it is also market neutral.
Adam: When you talk about using analogies of physics such as fluid dynamics or fluid turbulence, could you give an example of that and how it might work?
Alexei: This subject currently is reasonably well covered in some well-known econophysics books. The random fluctuations which need to be compared are the price changes in time for finance and velocity differences in both space and time for fluid turbulence. In both cases Gaussian distribution function serves as a certain limiting case. Despite that fact, for important situations of strong turbulence and volatile liquid markets there have been observed strong persistent deviations from the Gaussian distribution. In both cases, those deviations lead to much higher probabilities assigned to the tail events, which in the case of finance manifests itself in the abundance of price shocks up or down. In the case of strong turbulence, depending on the space dimension of the problem, those structures lead to strong visible coherent vortices of various shapes. The study of this anomalous behavior in the case of turbulence over at least the last 60 years led to the creation of vast statistical techniques, which can be adapted to finance.
Obviously, as rigorous theory behind this does not yet exist -- it may never exist -- applying such hypotheses at the engineering level of rigor should rely on careful statistical testing against the historical data. What does help here, however, is that at short timescales there are a lot more events available in the recorded history of the liquid markets. One also has to be careful in applying these techniques, as finance although similar is not really statistical physics, simply for reasons of not having enough statistical events behind it. The total number of all events in the life of S&P 500 futures can be estimated as 109, compared with the approximate number of atoms in one gram of a simple gas, 1023.
Adam: And you're not alone in going down this path?
Alexei: As we know, there has been a flood of engineers, physicists and mathematicians into finance since the early 1990s. But statistical description of turbulence dates back to as early as the beginning of the previous century. Starting from the 1930s, there was significant progress made, both by Russian and Western scientists, to produce the adequate description of the process with great success. Many problems became, at the engineering level, solvable. They were maybe not fully solvable from a rigorous mathematical standpoint, but at the engineering level they were solved. Evidence to that is abundant: cars and airplanes are being designed now with the explicit use of engineering turbulence models.
There are companies now that use this technology for the things we all use every day without fully realising that. I happened to work in that turbulence modeling field, and as I switched to finance, I tried -- as many others did actually -- to use this machinery to help design and describe the statistical properties of the financial time series.
We don't necessarily view our systems as exploiting some magical patterns; we more view them as statistical models for time series that may be described by probability density functions, with certain properties that may be fit by certain algebraic formulas, with certain properties of correlation functions for the time series, across the time series. This fundamental view doesn't really go much further once you really need to design the trading strategies. At that stage, we use many of the same techniques that many other systematic traders use.
"The study of this anomalous behavior in the case of turbulence over at least the last 60 years led to the creation of vast statistical techniques, which can be adapted to finance."
Adam: So it's more of a tool than a formula in and of itself. In other words, when developing your models and backtesting them, you would say, 'This set of equations helps us understand turbulence, and can it also help us understand this time series?'
Alexei: There are many exact examples why turbulence is analogous to price changes of, let's say, the S&P futures. Some of them are widely known. A quantitative analyst who doesn't come from these areas of science might use the wrong assumptions of how to build a trading strategy; he might be inclined to think that trend following and mean reversion is some kind of religion, and that you need to apply some miraculous filters and go through this almost prayer-like number of steps in order to achieve good results. This is not how a rational scientist should approach this, obviously. You can only answer questions if mean reversion or trend following are adequate descriptions of your time series by rigorously statistically studying it. You can rank typical behaviors based on their strength for various ranges of timescales. It makes your view of the markets more consistent and more like seeing order in it, similar to the Mendeleev's table of chemical elements. There's order there, and this order can be established by applying this methodology. This is how we view this.
Adam: What you were saying reminded me of how people view the Fibonacci sequence or some of the more esoteric technical theories such as Gann theory or Elliott Wave theory. They're looking for patterns in nature to be superimposed on the markets as a guide to the future, and some of them have an almost religious devotion to the theories. But you're not talking about using nature itself. You're talking about using the tools that are used to study nature and applying them to a new set of data, right?
Alexei: Yes. This has been our signature from the very beginning of our business. We started doing this in the early 2000s. At the time, there were not many traders focused on minutes and hours holding times. This range of timescales was largely perceived as noise then, but we loved this noise because we can actually study it and discover some predictabilities and build some strategies out of it.
But as you know, the number of entrants into this field has grown significantly and there is even a short-term trader index and various mean reversion traders categories or buckets. Obviously, the field became better known and more transparent. To some extent, I've been involved in this as well by teaching at Columbia University Mathematics Department for the last five years. There are over 100 students who graduate each year who study this in an academic way. There is this saying, 'The ones who don't do, teach'. We are involved in both -- doing and teaching.
Richard Flom (front), Andrey Manakov (back)
Adam: There's more competition in the mean reversion game, but your firm has been doing it for a while and it's still standing.
Alexei: Let me give you an example. Once you start talking about these fundamental issues, you discover that there are many smart people out there involved in the same. Now we're definitely not the only ones and being humble, probably not the smartest ones out there. But ultimately when you start being a practitioner, it comes down to you being disciplined and being consistent in applying this approach all the way through. I think if you align together at the start line, many similar players and look at them 10 years later, I think the ones who will have higher Sharpe ratios will most likely have not better but consistent models, consistent processes. You have a hypothesis; you continuously test your hypothesis against the ever-changing real data, to make sure your assumptions match your conclusions. Everybody now does backtesting, this is no longer a secret. But this is where the problems start because most often you visualise things that are not there -- you can imagine mean reversion in a random walk, for example. You can always over-fit a finite piece of financial data and produce some good-looking parameters, and it may be very tricky to really distinguish the signal from noise, especially since we live in this gambling-spirited industry. Our industry unfortunately is such that the average actual Sharpe ratio of the managed futures managers is below 0.5 and yet the average Sharpe ratio demanded by head hunters in the Bloomberg JOBS section is 6. Our strategies have Sharpe ratio above 1, but this is an actual live 12+ years track record, with thousands of trades taking place every year.
Anyway, the consistency of applying and checking every step of the way that your assumptions are in sync with what the financial reality right now is, this is I think what makes the biggest difference. It takes discipline and this is what differentiates, 10 years later, the winner from a loser.
Fact file Systematic Alpha Management
When: The company was formed in 2007 by Peter Kambolin and Dr Alexei Chekhlov.
Who: Kambolin is CEO and has worked for a number of buy-side firms on Wall Street. Chekhlov heads up research and graduated from the Moscow Institute of Physics and Technology before earning his PhD in applied and computational mathematics from Princeton.
Funds: The company's flagship Systematic Alpha Futures Fund is market neutral and based on mean reversion. The Systematic Alpha Multi Strategy Program features uncorrelated strategies, some market-neutral strategies and directional trend-following strategies. It is not 100%, market neutral.
Performance: The futures fund made money in almost every year of the past decade. The biggest gain, based on Class A shares with standard leverage, was a rise of nearly 17% in 2008. In 2011, the Class A shares lost nearly 12% but that was followed by a gain of more than 10% in 2012 and more than 9% in 2013. The multi-strategy fund has been running for two years, with an 11.5% appreciation in 2012 and an estimated 4.4% gain for last year. SAM says the standard deviation of the returns is 7-8% and the Sharpe ratio is above 1 for both funds.
How: The company has a four-person research team with backgrounds in scientific research. It has built its own analytical software which it says can backtest millions of records of tick data. SAM uses drawdown as its main risk measure and has built a method of optimising risk to more than 200 sub-models. It begins deleveraging once a drawdown reaches 5%.
In our view, if there is a secret to the trading model's development, it is the model's consistency or self-consistency. If the model is rational, which means its signals are based on exploiting an objectively existing price anomaly, then all models exploiting the same anomaly at similar timescales and on similar markets will be having highly correlated returns. What will differentiate these models from one another, however, is the average expected return per unit of risk. What will produce that higher risk-adjusted return is how the models are consistent with the assumptions behind them. One of the simple consistencies is that between the assumed effective transaction costs while optimising for the model parameters and the actual transaction costs. The actual transaction costs are those that will be achieved out-of-sample, given your particular trade sizes, capital under management, order latency, commission and exchange fee levels, et cetera.
"In our view, if there is a secret to the trading model's development, it is the model's consistency or self-consistency."
If your trading robots split your full orders into liquidity blocks and your hypothetical backtesting software is assuming a single block execution, you have to take this additional latency into account as well. One will easily notice that your strategy actual versus hypothetical returns decay is very strongly correlated to the consistency of your hypothetical and actual transaction costs. Another, maybe even more important, consistency is related to overfitting of your parameters to a particular finite historical price path.
Adam: How does that translate into your own performance. I would imagine you wouldn't see a lot of volatility but that it would generate a fairly steady performance.
Peter: If you look at our returns in the Systematic Alpha Futures Program, there are periods, like for example in 2013, when we make money very consistently for six, seven, eight months in a row while the volatility of the returns stays fairly low. However, there will be other periods, when all of a sudden we have either a sharp negative or a strongly positive month and rarely we have drawdowns. In fact, in our 13-year history we only had two drawdowns that were 10% or larger. The consistent returns that we typically generate are a result of the high percentage of profitable trades: in our case about 60-65% of individual trades are positive. So, most of the time, our models work consistently well. But there are other periods when something happens in the market place -- and of course, after doing it for so long, we now know what these periods typically are -- when our models go through a difficult time. The good thing is that these periods are rare; they happen on average once in five years, in terms of the drawdowns. In terms of the negative monthly returns, only 27% of months are negative, which means that 73% are positive.
The problem with the investor psychology is that when you're doing well while others are doing poorly, everyone loves you. But on the other hand, because our strategies are so unique and uncorrelated, our drawdowns happen during times when everyone else could be doing fine, and if the market goes up, and let's say the indices go up and hedge funds go up and we have a drawdown, that's when investors lose patience and say, 'Well, this strategy doesn't work.' We've experienced such feelings towards us a couple of times, but at the end of the day we fully recovered from both drawdowns that we had and those investors that penalised us by redeeming ultimately penalised themselves, because had they stayed invested they would have done very well.
Our response to such investor reaction to the drawdowns was an introduction of the Systematic Alpha Multi Strategy Program that we launched back a couple years ago. There, we blend together highly uncorrelated strategies, market neutral contrarian strategies and directional trend-following strategies, with the idea to mitigate the drawdowns that will come in the future. So when we speak to clients, if we deal with an experienced investor who understands the properties of low correlation, who appreciates our consistent returns, and who can handle drawdowns, we recommend our original market neutral program. To those investors who are looking for a more diversified program, or the program that will have less likelihood of a large drawdown, we recommend the multi strategy program. However, that program is no longer market neutral 100%, and it is no longer zero percent correlated to the benchmarks.
So currently we offer two products to our investors , and each one has its own investor base.
Adam: When you talk about the times when you have had drawdowns, are those typically related to market regime change, or is it something else you can identify?
Peter: Regime change is something that will impact the returns. Back at the end of '07, when we had our first drawdown, there was a regime change from the market being highly bullish to the market being bearish and going into recession. The second drawdown we had was in 2011 and there was a major regime change in a particular country, Japan, after the earthquake and the nuclear disaster that followed. That was a regime change, but country specific, and then there was a regime change in Europe related to the debt crisis and credit risk and all that. At the same time, our correlation to the direction of the equity markets is zero. So, if there is no regime change but let's say some event takes place and the markets have a healthy 10% correction, it doesn't mean we'll have a drawdown. It may actually be a positive environment for us because volatility will spike, correlation between the markets most likely will remain quite high and for us, being market neutral, as we rely on markets being correlated to one another, actually it could be a positive development.
Adam: So it sounds like it all has to do with the nature of a regime change, such as whether it's one that has been long anticipated and has finally happened or, for instance, a black swan event.
Peter: If it's a healthy correction but the regime is more or less the same, then it's a good environment for us. If it's a regime change in a major way, yes, it could be a challenging environment. The good thing is that our models are adapting. For example, after a major regime change starting in December of 2007, it took us a couple of months to adjust to that new regime, and starting February 2008, the models started producing very solid returns and 2008 turned out to be our best year on record.
Adam: You're a fund that puts a lot of effort, time, computational power and brain power into statistical analysis. Do you do a certain amount of fundamental analysis to keep an eye out for those regime changes?
Peter: I would say that before 2011, before our second drawdown, we were not doing it at all. We were just relying on our statistical analysis. Now we are more vigilant, I would say. There's no direct input; we don't have a fundamental factor that we would use when we place trades, but we would be more willing or likely to intervene or to change things quicker if we detect another regime change going forward.
"If it's a healthy correction but the regime is more or less the same, then it's a good environment for us."
Adam: In other words, if, for example, the new Federal Reserve chief said something that changed the market's view entirely, you would be more likely to make immediate changes than you would have been, say, five years ago?
Peter: Well, no. The immediate trade that we
will place will be more or less the same.
I don't think we will change anything if one day something like that happened. If we feel and see and determine that this regime change is undergoing and it's been a month or two and there are multiple confirmations of that, most likely we will start acting and doing something. But we will not wait three to six months. Day to day, we will not intervene because a lot of market moves are so-called knee-jerk reaction moves -- markets move up and then they move down and then they move up again and actually these are the opportune times for us. Yes, in one or two occasions we will be correct intervening but in eight others we will be wrong intervening, so if we intervene all the time we would potentially kill the arbitrage that we're looking to exploit in the first place.
Alexei: We're not talking about cancelling a particular trade or doubling up on a particular trade. This is not really at all how it's done. I think it's better to get a sense of it by example. Let's look at Japan over the last two years. Statistically at the end of 2012 through early 2013, a major policy shift had been put in place by the new prime minister -- which is known as Abenomics -- and that was a huge difference to what Japan had been doing since 1980. And he put into place a significant currency devaluation, significant tools to ignite the economy and he signaled left and right that this would be going on for some time. So this is a very clear signal of a regime change in Japanese assets such as Nikkei 225 or the Japanese yen that will most likely undergo a change from one equilibrium to another equilibrium and the policy will probably stick. But it will take some time for this new equilibrium to be established. So it's not a question of shielding one trade or doubling up on another trade; this is not how it's done. It's more a question of okay, this doesn't seem to be really statistical right now, and maybe you should abstain from trading this until the arrival of statistical evidence that the new equilibrium has been established.