Futures spread trading

[bone]

Drawdowns are random from what I can tell.

I try to keep my clients from replicating forward curve exposure in the same market sector: in other words, being long the TuF and long the NoB and long GE Z2-Z4 simultaneously for example.

Or being short ZW Z1-Z2 and being short ZC Z1-Z2 simultaneously would be another example. In those cases, I want my clients to choose the best set-up in terms of the model and risk/reward skew.

I haven't gotten involved with any Martingale strategies. In terms of capitalization - the volatility and trading range of the individual spread dictates the leverage and sizing.

Good stuff, great topics to discuss.

 

[rose]

Quote from Martinghoul:

Sure...

So the basic principle is rather simple. It ain't rocket engineering, that's for sure. Suppose we start with a time series that we believe exhibits some mean-reversion properties. Also, let's say that, in your latest observation, you establish that the model residual is, say, n (where, for simplicity, let's assume your model is some flavor of multiple regression). Next you fit some sort of a AR model to your time series (I actually do it to the residuals) and you establish the speed of mean-reversion, e.g. smth like a half-life. The last step is to obtain the volatility of the residual time series in the most bog-standard manner (stdev). Now you have pretty much everything you need to calculate what I call the ex-ante Sharpe over the period equal to the half-life. Your denominator, i.e. the return, is n/2 plus carry. Your numerator is the volatility from above (normalized to the half-life as well). Once you have the half-life Sharpe, you can "annualize" it to bring everything to a consistent basis. None of this is particularly scientific and it doesn't even have to be that rigorous, i.e. you can estimate your inputs whichever way you like, including waving fingers in the air.

Again, we can discuss the technical details of the specific steps separately, but my point here is more about a consistent framework for entries and exits. You can make the approach above systematic (not that I recommend it). Specifically, suppose you have a limited risk budget and you scan your opportunity set for trades, say, every day. As soon as you establish that either a) an existing trade's ex-ante Sharpe has fallen below a set threshold; or, b) a new trade exists that has a higher ex-ante Sharpe than an existing trade; you exit the existing trade and, in case b), strap on the new one. You can also make the size of the trade dependent on the Sharpe.

Now, obviously, this is all VERY simplistic and there's lots of devils in the details. However, I really like to use the above as a self-consistent starting point that you can refine to your heart's content (sky's the limit). As I mentioned, we can talk further about the gory details, including the "divergence" issues (if I understand what you mean correctly, there may be a way to deal with some of them).

I'd be very interested in learning more about this approach. I've had a stab at doing it - I've picked some Brent flies as that is my main market.

I picked a time frame where this particular market seemed to mean revert well - ZHM Brent since May 1st this year, chart is here:

http://quotes.esignal.com/esignalpr...chart.bardensity=LOW&x=53&y=13&chart.studies=

So by eye, and from trading it, I think it's range bound and am happy to scale buy from -17 and scale sell from -6. I started off by running a Dickey-Fuller test on the data. There were 59 observations (daily closing prices), t-stat was -1.68 and as that was greater than the 2%, 5% and 10% critical values it means I cannot reject the hypothesis that there is a unit root.

So far I think that means statistically it seems to be mean reverting, it's then up to judgement etc. to check that there is a logical reason behind it - which I believe there is. The R-Square value was only 0.04 - that is very low - but then I'm not sure it is relevant.

What would be the next step? Do I run another regression model on the components of the fly? Is it sensible to run a linear regression of the HM spread vs the ZH spread? Should I run a multiple regression model of the individual legs?

I trade flies a lot, and would really like to explore a more rigorous approach such as the one described above.

I think this fly (ZHM Brent) could be an interesting study, below is the same market but with the time frame opened out - obviously February and March began with some painful divergence!

http://quotes.esignal.com/esignalpr...rt.bardensity=MEDIUM&x=53&y=11&chart.studies=


Any and all help appreciated!

 

[Rationalize]

Quote from Martinghoul:

Sure...

So the basic principle is rather simple. It ain't rocket engineering, that's for sure. Suppose we start with a time series that we believe exhibits some mean-reversion properties. Also, let's say that, in your latest observation, you establish that the model residual is, say, n (where, for simplicity, let's assume your model is some flavor of multiple regression). Next you fit some sort of a AR model to your time series (I actually do it to the residuals) and you establish the speed of mean-reversion, e.g. smth like a half-life. The last step is to obtain the volatility of the residual time series in the most bog-standard manner (stdev). Now you have pretty much everything you need to calculate what I call the ex-ante Sharpe over the period equal to the half-life. Your denominator, i.e. the return, is n/2 plus carry. Your numerator is the volatility from above (normalized to the half-life as well). Once you have the half-life Sharpe, you can "annualize" it to bring everything to a consistent basis. None of this is particularly scientific and it doesn't even have to be that rigorous, i.e. you can estimate your inputs whichever way you like, including waving fingers in the air.

Again, we can discuss the technical details of the specific steps separately, but my point here is more about a consistent framework for entries and exits. You can make the approach above systematic (not that I recommend it). Specifically, suppose you have a limited risk budget and you scan your opportunity set for trades, say, every day. As soon as you establish that either a) an existing trade's ex-ante Sharpe has fallen below a set threshold; or, b) a new trade exists that has a higher ex-ante Sharpe than an existing trade; you exit the existing trade and, in case b), strap on the new one. You can also make the size of the trade dependent on the Sharpe.

Now, obviously, this is all VERY simplistic and there's lots of devils in the details. However, I really like to use the above as a self-consistent starting point that you can refine to your heart's content (sky's the limit). As I mentioned, we can talk further about the gory details, including the "divergence" issues (if I understand what you mean correctly, there may be a way to deal with some of them).

Will reply to this fully later.

In the mean time .. the distribution of the residuals, and the fitting of an AR model is .. interesting .. and I think, perhaps where the juice is.

The rest is just "work" by comparison, no?

 

[Martinghoul]

I'll be back a bit later to discuss, gents...

 

[Rationalize]

Quote from Rationalize:

Will reply to this fully later.

In the mean time .. the distribution of the residuals, and the fitting of an AR model is .. interesting .. and I think, perhaps where the juice is.

The rest is just "work" by comparison, no?

+ something along the lines of simple variance ratio tests .. to identify potential reversion timeframes .. occasionally finding several .. with different means .. for the same combination of instruments .. :eek: ..

 

[rose]

Quote from Rationalize:

+ something along the lines of simple variance ratio tests .. to identify potential reversion timeframes .. occasionally finding several .. with different means .. for the same combination of instruments .. :eek: ..

This is something that happens in energy at least, you can have the same strategy and have one range for intraday jobbing, and another range for longer term positions, bias is pretty much always going to be the same, but there are opportunities - obviously if you get caught out there are problems.

 

[Rationalize]

Quote from rose:

This is something that happens in energy at least, you can have the same strategy and have one range for intraday jobbing, and another range for longer term positions, bias is pretty much always going to be the same, but there are opportunities - obviously if you get caught out there are problems.

"obviously if you get caught out there are problems."

unless you only enter the shorter term trades in the direction of the longer term bias?

 

[rose]

Quote from Rationalize:

"obviously if you get caught out there are problems."

unless you only enter the shorter term trades in the direction of the longer term bias?

Ha, yeah - that's the idea - the problem is that a short term trade has morphed into a long term trade at a rubbish level, and it may hamper your appetite for the trade etc. at least it gives you something to job around

 

[Trader13]

Quote from rose:

I'd be very interested in learning more about this approach. I've had a stab at doing it - I've picked some Brent flies as that is my main market.

I picked a time frame where this particular market seemed to mean revert well - ZHM Brent since May 1st this year, chart is here:

http://quotes.esignal.com/esignalpr...chart.bardensity=LOW&x=53&y=13&chart.studies=

So by eye, and from trading it, I think it's range bound and am happy to scale buy from -17 and scale sell from -6. I started off by running a Dickey-Fuller test on the data. There were 59 observations (daily closing prices), t-stat was -1.68 and as that was greater than the 2%, 5% and 10% critical values it means I cannot reject the hypothesis that there is a unit root.

So far I think that means statistically it seems to be mean reverting, it's then up to judgement etc. to check that there is a logical reason behind it - which I believe there is. The R-Square value was only 0.04 - that is very low - but then I'm not sure it is relevant.

What would be the next step? Do I run another regression model on the components of the fly? Is it sensible to run a linear regression of the HM spread vs the ZH spread? Should I run a multiple regression model of the individual legs?

I trade flies a lot, and would really like to explore a more rigorous approach such as the one described above.

I think this fly (ZHM Brent) could be an interesting study, below is the same market but with the time frame opened out - obviously February and March began with some painful divergence!

http://quotes.esignal.com/esignalpr...rt.bardensity=MEDIUM&x=53&y=11&chart.studies=


Any and all help appreciated!

I'm hardly a veteran of spread trading flies, but I really like the concept of mean reversion trading when you can find the right setup. Expanding your chart to a longer timeframe reveals that this energy spread has some wide excursions from the mean for several months during the past year. Going with the principle that your trading time frame should be proportional to your risk tolerance, I would trade this market on an intraday basis, perhaps using 30 - 60 min bars.

 

[bone]

Quote from Trader13:

Going with the principle that your trading time frame should be proportional to your risk tolerance, I would trade this market on an intraday basis, perhaps using 30 - 60 min bars.

You have to be very careful and really do your scouting in terms of trading flys and condors on a HF basis - the slippage will eat you alive. Look at the actual order books for the individual legs in lieu of the charts (there is a latency artifact in the charts for the last print versus the actual bid and offer in the lower volume legs - makes the trading range appear larger than it really is).