Robert Carver, the author of "Systematic Trading," "Smart Portfolios," and "Leveraged Trading," maintains a lengthy discussion thread on elitetrader.com. This thread provides valuable insights into fully automated futures trading. After dedicating a significant amount of time, I've thoroughly read through the thread from start to finish and extracted some useful portions (up to page 286).
Q is the question and A is the answer from Robert Carver.
Q: In the time frame between Jan 2014 and Jul 2014 the PL advances pretty fast. Maybe you could adjust the strategy in a way that the rest of the months advance at the same pace.
A: The only way to do that would be to boost the Sharpe Ratio whilst keeping the same skew (which would be lovely, if I knew how...) or to make the skew negative (which would improve Sharpe but isn't the kind of trading I like doing). Positive skew trend following will always have periods when it works, and longer periods when it doesn't.
page 1
I don't have discrete entry and exit. Each market has a continuous signal; when it goes to zero I'll not hold a position. Signals are a mixture of trend following, breakout and Carry (or rolldown/contango if you prefer). The predominance of trend following means positions will tend to be cut when prices move against us, so its an implicit rather than an explicit stop loss.
page 2
Q: In my past i used work for an affiliated firm of ahl so i was somewhat familiar with their product. Ahl, Aspect and the like would have gradually building and tapering positions in various markets over multi systems like what you are doing. But such firms had a sharpe of maybe 1.0. You have a sharpe of twice that. What would you say you are doing differently from them?
A: Not much differently (I don't know if you are aware of this but I used to work for AHL). AHL also had a very high Sharpe last year. On a risk adjusted basis we're probably pretty even. I will soon post my simulated returns. The average SR from 1980 onwards is about 0.9. So in a similar ballpark.
page 5
Q: Outside of "is the resultant P&L positive," what metric(s) matter most to you?
A: In order: 1. Realised volatlility versus expectations; 2. Costs versus expectations; 3.Skew, and related statistics like average gain to loss; 4. Last, and least, Sharpe Ratio; as over short periods of time this has the largest variance.
page 7
A: The short answer to your question is that with my capital of £400,000 the total nominal size of my positions (using price, not nominal value, so for example a US 20 year bond is worth $160K rather than $100K) is £5.86 million, which is a ratio of 14.6:1.
A: I'm currently using £127,000 of margin on a £400,000 account, or an average of 2.2% of the notional leveraged value of £5.9 million.
page 8
A: I won't lie to you, I don't like losing money! But if you're a medium term trend follower you spend a higher proportion of your time in drawdown than you do not in drawdown….
page 8
Success in systematic trading is mostly down to avoiding mistakes: over complicating things, being too optimistic about likely returns, taking too much risk and trading too often. I will help you avoid these errors. This won't guarantee large profits, but it will make failure much less likely.
page 9
Obviously some chunky losses on equities, but also short crude didn't play out well today. Yesterdays expected risk was £4554, so this was a 2 and a bit sigma day, or roughly one in two months loss. NOT so easy… How to put this into perspective? Since I began writing this thread I've made about 24K, or call it 16K a month. Annualised that would still be a 50% return. My conservative expectation is to make about 12% a year (SR of 0.5).
This is just trading. Unless you're exceptional, or you're running a highly negative skew strategy which one day will blow up in your face, you don't make money every day or every week.
page 9
I think in backtest my max DD is around 37%. This is probably optimistic, but on the other hand it doesn't allow for kelly reduction in capital as DD happens. Also reduction in risk as DD happens also reduces margin. So if you're paying 50% margin, and you lose half; by then you'd only need half your original margin (25% of capital).
page 13
So yesterday, for a few hours, I broke through my HWM and out of drawdown. A drawdown lasting a little less than 4 months is relatively short if you look at the backtest, but it's a different matter when you're living through it.
page 16
The length of drawdown is also a factor to consider
Q: I'm amazed by your graph (in prior post) showing the expected and actual risk. You mentioned that your "risk" calculations are based on standard deviation of the daily returns. I assume that the calculation of expected and actual risk is a measure of standard deviation of returns based on some look-back period, is that correct? If so, what is the duration of the look-back that you use?
A: I use a lookback of about one month for measuring standard deviations, and a longer lookback of about 6 month for correlations (and using weekly returns).
page 17
Compute real-time risk (standard deviation) and obtaining the target risk could be a valuable approach for updating strategies. (Emphasize risk assessment rather than focusing solely on gains)
Extreme values of momentum don't translate into larger trends (think - dead cat bounce).
I use a simple cap to get round this. However AHL use something where larger signals are actually downweighted. Note for very large signals (bigger than 1.5- AHL scale their forecasts so they have a standard deviation of 1.0) the forecast after applying a response is actually getting smaller as we get more confident. There is some statistical evidence that this makes sense, but it is quite weak and doesn't seem to be around in every asset class or speed of momentum. Also it leads to weird behaviour - when a trend strengthens you start closing your position; then when it weakens again you open your position up before closing it again. Which is why I didn't keep this feature when I built my own system.
page 20
A significant signal doesn't necessarily translate into a major trend. It seems to be a nonlinear process, somewhat resembling a logarithmic function. Once a certain threshold is crossed, signals are considered equally significant. In fact, large signals might even have a counterproductive effect.
Q: You mentioned in your profile that you mainly focused on fundamental strategies during your tenor at AHL. How do these strategies usually work (ie what variables do they look at)? There's a lot more about price based systematic trading compared to fundamental based macro models…so I am more inclined to hear your experience in that area. What are the typical sharpe ratios and returns and how do they compare from a correlation perspective with price based systems?
A: Mostly it was taking ideas from economic theory and trying to apply them. So for example you have things like PPP in currencies (google it if you're not familiar), taylor rule in interest rates, cochrance-piazzesi in bonds.
A: Then you can also use classic value indicators but aggregated. So for example you could buy cheap PE countries and sell expensive. Although differences in accounting, taxes, and systematic investor biases mean things can stay expensive a long time (think Japan). Ideas like Shillers PE are interesting.
A: You can also look at value across asset classes, eg the Fed model. Although you need to add inflation and a few other things to it to make it reasonable.
A: A lot of these things are correlated with carry. For example bonds tend to do well when the yield curve is steep. But a simple carry method will give you the same answer.
A: Holding periods tend to be long, so sharpe ratios lower (well below 1.0). In isolation you wouldn't look at these things twice, but they do add something to a basic technicals + carry model. But there is a lot of work involved in building them, and then in getting clean data. So it's something that only an institution would probably bother doing.
page 21
So for the last tax year (which was unusually good in terms of performance) at least the figures are 45% win rate, 1.92 win:loss rate.
page 26
Q: Can you elaborate on why you roll manually? What are the problems you encounter in automating this task?
A: To be clear the rolling is automatic (my execution algo issues either a spread, or two individual leg orders), but the decision to roll is manual. This is because there are multiple things to consider as http://qoppac.blogspot.co.uk/2015/05/systems-building-futures-rolling.html discusses. It would be very complicated to code up all these considerations into an automated decision making process, which would save very little time (a few minutes a month).
page 34
Q: Why do you favour commodity futures versus other instruments, and why are commodity futures the principal instruments traded by large systematic hedge funds?
A: 1.huge liquidity (more important for multi billion dollar CTA's than me); 2. they trade on exchange. The advantages of this are too numerous to list and would require a separate post; 3. they are very cheap to trade (slippage and commissions for a given level of risk); 4. easy leverage: margin requirements are low; 5. cheap available long history of data; 6. path dependence; In my professional career I spent nearly all my time working on futures models (with some dabbling in equities, and fairly serious work on OTC interest rate products), so makes sense for me to it now
page 34
Q: Do you have any thoughts on what can be done to limit drawdowns, apart from running on a lower vol target?
A: Drawdown distributions are a function of sharpe ratio, volatility target and the autocorrelation and skew of the return distribution. So if for example you do something like equity curve trading (see fuller analysis here http://qoppac.blogspot.co.uk/2015/11/random-data-evaluating-trading-equity.html) on something like an option selling strategy (which has positive autocorrelation of returns, and negative skew), then you'll reduce the drawdown but at the expense of the average return (there is no free lunch).
page 41
Q: In case I missed it, could you please give one more time exact definitions of 'breakout' & 'momentum'?
A: "momentum" is an exponentially weighted moving average crossover; "breakout" is as defined pretty much as ROC (see here https://qoppac.blogspot.com/2016/05/a-simple-breakout-trading-rule.html)
page 45
Q: Interestingly, in backtest, the skew on the daily returns of the trend following strategy is negative, while the skew on monthly returns are positive. Do you see similar results on your trading? Why would this be the case?
A: Yes. This is actually a really interesting result. Basically the skew for a trading strategy will only become apparent at frequencies in line with it's holding period. So if you're intra day trading, and trend following, you'd expect to see positive skew at daily horizons. But with a holding period of weeks the positive skew doesn't kick in until you're looking at monthly results. At a daily frequency you end up with skew that is in line with the skew of the typical position you're holding.
page 51
The measurement time resolution of skewness should match with the holding time resolution.
My maximum risk is 400K. If I lose money below that I degear. But if I make money above that I don't increase my capital at risk (more http://qoppac.blogspot.co.uk/2016/06/capital-correction-pysystemtrade.html). This means effectively my trading account acts like a hedge fund with a 100% performance fee. Any money made above the high water mark gets paid to the investor.
page 53
Q: I notice there is some oblique discussion of systematic option strategies on this thread (including the very useful link to Palaro's research - thanks for that). It seems from this and also from browsing various academic papers online that the bulk of research surrounding systematic vol trader is centered mostly around systematic short selling strategies. I am wondering if you had any links/book recommendations/guidance on building more fully rounded systematic vol strategies - i.e. ones which not only build on systematic short selling, but also look at stuff like momentum in implied vols, trend in implied/realized spreads, trend in vol term structure, trend in skew etc as signals when to buy/sell implied vol or inform cross-sectional vol strategies? I have not come across anything related to this online or in systematic trading books and any guidance would be appreciated as I look to build something like this out.
A: I am afraid I don't know of any book or website that addresses this. Most (good) options traders seem to be discretionary in nature.
page 53
Q: what would you see as your "edge" in these strategies? As in: why would a one-man show be able to consistently run at a Sharpe ratio of around 0.9 (or even higher)
Q: what would you see as your "edge" in these strategies? As in: why would a one-man show be able to consistently run at a Sharpe ratio of around 0.9 (or even higher)
page 55
Q: what's on your research agenda for the next 12 months? I mean signal-wise (not refactoring of code).
A: https://www.elitetrader.com/et/threads/fully-automated-futures-trading.289589/page-55#post-4343411
Here, the author has listed numerous strategies that he intends to develop, consolidating them here for potential future reference.
Q: I think I've read that the amount of institutional money trading trend following strategies have hit a record high, and on a similar note smart beta strategies are also very popular these days. How would you identify if this space is crowded or not?
A: I find these kinds of questions a bit meaningless if I'm being honest. I think "crowded trades" are a problem when you are trading quickly, are trading mispricings, are trading relative value, have high leverage and so on. This creates an environment in which crowded trades are obvious (eg the value premium being squeezed out in equities, interest rate premia vanishing in fx carry) and also dangerous. None of these things characterise institutional trend following, so it's hard to identify if the space is crowded, and also hard to know what you should do differently. If anything trend following could be self reinforcing the more people are doing it. Because it's relatively slow 90% of traded volume in most asset classes will still be done by people doing other strategies.
page 58
Q: Do you apply a portfolio position size scaling multiplier that is a function of cumulative drawdown?
A: Only in as much as I use Kelly, i.e. scale my positions according to my account size. So position size is proportional to (max capital at risk (400K fixed) - drawdown)
page 59
Q: What is the rationale behind not fitting system parameters, but fitting capital allocation? Is fitting capital allocations not similar to fitting system parameters as it involves the same data snooping biases? How precisely does your methodology differ in determining the capital allocations?
A: The logic is to design trading rules that "ought" to work, and then through capital allocation make sure we don't give capital to rules which are really terrible (and statistically significantly terrible) or 99% correlated with something we already have. Essentially we get away from explicit (through formal parameter search) and implicit (through trying multiple options i.e. data snooping) overfitting and are left with tacit overfitting (we only ever try rules we know will work before we even sit down at the computer).
page 61
Q: Hi everyone, I am comfortable with trend following, but am seeking to diversify through some mean reversion. I know strictly speaking carry is a form of mean reversion, but does anyone knows of some good sources or strategies for mean reversion on futures?
A: In the simplest case you would want to determine if a time series is showing mean reverting behaviour over a time window and then act accordingly. There are a bunch of ways for accessing the mean reverting properties of a time series e.g by using the Hurst exponent https://en.wikipedia.org/wiki/Hurst_exponent
page 131
FWIW the SG Trend Indicator, which is a fully-disclosed mechanical strategy, was -0.60% yesterday, +5.67% mtd.v
page 143
Q: Is it a good use of time to try to build a more accurate trend filter? Say using machine learning techniques and using more factors/features to redue noise/detection? If not why so?
A: No, a bad use of time IMHO. I'd be surprised if you could get significantly better out of sample performance from the ragtag bunch of simple trend rules I use myself (I've seen people try, and fail, at this). You're better off diversifying your system to different instruments and different styles of trading.
page 148
Achieving that Sharpe （SR 1.5+） with daily strategies is possible (and with intraday trading maybe even up to 2-3 sharpe). But like you brought up, the trick is all in the execution/details.
Strategies that are under 1 Sharpe are easy to find in the public domain (such as GAT's trend and carry signals or other alternative beta stuff), but ones that are higher than that are much more closely guarded and probably worth multiple millions. The best thing you can do is to build a framework that lets you backtest strategy ideas quickly and rapidly innovate and iterate. It comes down to eliminating bad ideas quickly. You say each of those ideas would take you multiple months to test, but for comparison I can probably go though each of those in less than a week using my system. Try to design a flexible framework that can let you do that.
page 149
Q: You know we were talking about strategies with no negative years? I just noticed that Winton hasn't had any down years. I'm not sure what they're doing that makes such a big difference. Here's the result of my constant capital back test from 2000.
A: From what I understand, Winton has only about 50-60% of their risk in trend following, about a quarter in carry and the rest in newer strategies, so they are becoming more and more of a hybrid trend-follower as opposed to a pure trend follower. They also trade long/short market neutral equities in their original flagship fund. So you’re not necessarily comparing apples with apples if you’re comparing them with a pure trend follower….
page 151
Q: do you have any recommendations for reading on the topic of trading options?
A: Taleb dynamic hedging
A: Funny, I am not super-keen on that book, it's very much oriented towards old school market making and it pretty dated. My initial reaction is almost always something like Volatility Trading by filthy, which is not perfect either.
page 172
Q: I'm curious to learn what kind of intraday system you have in mind. I guess that you will publish about it once it has reached an, in your opinion, sufficient level of maturity.
A: Basically mean reversion, initially on individual futures and then at some point on spreads / flys. Holding period will be a few days. The key point here is the execution, essentially you place limit orders eithier side of the 'fair value' so you're always filled passively (unless you're using stop losses). That requires a different kind of backtest to my current 'assume you always pay the spread and get filled at the close the day after the signal is generated'. Instead you need to work out whether the market moved through your limit during a particular time segment (in which case you assume you get filled at the limit price), and obviously with more frequent intraday data you can avoid missing the situations when the market touched your price and then moved away again.
A: will probably use stop losses for which the logic should be something like (conservatively) assume if we traded through the stop that it was executed at the worst level [although it's possible again that the price touched an even worse level in the intra- time period between samples], or (aggresively) exactly at the stop level plus the spread, or somewhere in between.
page 186
Q: Just curious if anyone has seen this article before by @globalarbtrader previous firm, https://www.man.com/maninstitute/volatility-is-back-better-to-target-returns-or-target-risk In the article, it discusses how dynamic volatility switching may bring up the sharpe. In particular this para, In this second layer of risk sizing, we look to improve on the first layer by introducing a faster measure of volatility (with a half-life of just 12 days). We don’t actually use this faster measure unless it starts to diverge materially from the slower measure and this will typically only happen when market conditions are going through material change. In those scenarios, we use the faster measure, and de-risk much faster. Just wondering if this switch here is a binary switch or continuous scaling kind of switch (e.g. weighted short term and longer term volatility)
A: I hadn't seen this article previously, but you mentioning this reminded me that I modified the volatility calculation in my system compared to what is used in the book "Systematic Trading". In the book the volatility is based on a 25 day lookback period. I noticed some undesired behavior with this once an instrument goes from a period with high volatility to normal/low volatility. I decided to modify my system to not use a lookback period of only 25 days, but to use the average of 10 days, 18 days and 25 days. The effect of this is that when volatility spikes up this is immediately visible and taken into account. When volatility reduces the averaged calculation tracks it better than using the 25 days average only.
page 235
Q: I noticed that some of you are using MongoDB (via Arctic) to store your time series data. What's the disadvantage/ advantage over say Kdb+ (Q) which seems like a commonly used DB in the quant space? At the moment, I'm not well verse Kdb+ but plan to look into it.
A: Kdb+ is a more industrial strength solution which is used by a lot of HFT type people. It's licensing is complicated and it's generally more palaver to run than a simple mongoDB instance (which you can run with fancy distributed sharding and what not, but out of the box for simple instances). Kdb is also columnar not no-sql. I find mongoDB without the Arctic layer very handy for storing non time series information; basically I just store everything as a dict and write .as_dict() and .from_dict() methods for all my objects. Basically for low frequency traders Kdb is overkill; in fact even Mongo is unneccessary given I ran the first iteration of my trading system for over five years on SQLite.
page 251
Q: I once read the book written by the firms's CIO - "the Dao of Capital." It was a misterious book, with full of philosophical things. The book talks very little of technicals, but it says that you need to buy deep out of the money put option of S&P500, with about 40% implied volatility and with 2 months maturity, and to keep rolling. There is a bloomberg artcle that says the firm gained 3600% return in the crisis 2020 March, which I guess you already know.
A: Buying deep OTM puts is likely the essence of the strategy, although in reality they probably run something more sophisticated than the recipe outlined in the book. The firm did gain that in March 2020 but the media made it out to be more than what it is in reality, as always. There is a 2020 Q1 (written on April 7) leaked letter to investors: "Based on your required invested capital at the start of the year, in March 2020 you experienced a +3,612% net return on capital; year-to-date you have experienced a +4,144% net return on capital" Impressive numbers? However: As in my last Decennial Letter to you, we will show this portfolio effect by updating the performance of the hypothetical Universa “risk mitigated portfolio,” which pairs our actual net performance (monthly administrator-provided net returns, using yours from your start date, expressed as returns on a standardized capital investment) with an SPX position (a realistic proxy for the systematic risk being mitigated). The weightings between the two are 3.33% and 96.67%, respectively, as per the weightings we have always recommended for a fully “tail-hedged” Universa risk mitigated portfolio."
page 262
Q: How do you define 'slippage'? Sorry this has probably already been mentioned before. It may be 'unfair' to use midpoint, especially for fat tick sized futures e.g. if bid-ask is 100(20 lots)/101(1 lot), the 'fair' value is much closer to 101, clearly much easier to sell at 101 than buying at 100. Some people then decide to use a volume-weighted average on BBO as 'fair px'. IMO this is more accurate, but HFTs are also known to mix up their sizings to confuse people, so not perfect either.
A: So in general, I'm way over my head here, my execution algo is also a rip-off of the Rob's algo described here: https://qoppac.blogspot.com/2014/10/the-worlds-simplest-execution-algorithim.html Basically I issue only limit orders, I start on the best side and will cross the spread if either: 1. too much time passed (we're talking tens of minutes); 2.the price(the side of the spread I'm currently at) moved against me; 3.there's an unfavorable bid\ask volume imbalance (i.e. if I'm buying and bidSize > 5AskSize or I'm selling and askSize > 5BidSize). So of course I'm not trying to compete with HFT, the only valid question for me is whether doing this simple algo is better than just sending market orders all the time. It seems to be better, because as I understand, with a market order I'll pay the full spread every time for sure, and with the algo I seem to do a little better on average, and that's really all I'm after here.
A: This is indeed what I do with my trading software: if the spread is more than one tick I place a limit order at midpoint. If this order isn't filled after one minute waiting I modify the price to cross the spread.
page 266