Interesting post about the correlation trend. And to further the discussion.... Quick question for anyone in the know... Excel has a forcast calculation based on linear regression. To use this function I believe one would use the current mean pair ratio as the value you want to base off of to predict a future value, and then plug in the range of the correlations and the range of the pair ratios to get a predictive value? Anyone else doing this or have experience with excel and statistics to confirm I am using the right data points and ranges etc. Thanks.
Just a follow up to my previous post. Current pair ratio of cf/pot is .877, the mean is .74. and two standard deviations away from the mean is .85. Also, as you noted, the correlation of the two has been trending down and the mean ratio has been trending up. Applying linear regression the predicted pair ratio is .822 using the last 30 days of trading. Interesting enough, the predicted value is 1 standard deviation back towards the mean indicating a mean reversion. I'll make this point again, but everything is dependent upon the timing of your trade and the time frame of your trade. The window of opportunity to lock in profits could easily pass you by with this one as the longer term trend is clearly favoring cf long and pot short. But one has to remember that pair trading by definition means taking profits quickly and when you have them.
I use a number of probability sets for helping me generate signals (a couple different standard deviations are in there) and one thing you need to consider is that they're constantly changing. This is why I prefer a discretionary system over purely mechanical because even when the same conditions generating a buy/sell are triggered at different times, the probability sets are not quite the same (even though the numbers are) and you have to adjust for that....just something to consider, don't fight a big move just because some indicator is a screaming buy.
I am wondering how this fits into an overall risk/reward profile? If you enter a trade when the pair is 2 stdev's away from the mean (for example) is your profit exit goal at 0 stdev's? 1 stdev? Now if the pair continues to diverge I understand you add layers (new trades) but without an ultimate hard stop loss point I am not sure where you could/would set your profit target.
Well that's the million dollar question and varies with every trader! I don't use fixed layering but do trade all time frames and using your example you wouldn't be looking at a full retracement to the mean. I don't use fibonacci's but you could use that as a example, 25/50% retracements. Here's a key factor in my trading and that's that I'll adjust my position size according to the probability of the trade. Some setups are just better signals then others so I'll increase size and scale in and out as the spread moves around. Very often i'll put on 1/2 or even 1/3rd a layer and take mini scalps on an intraday basis or maybe hold that position size as it's trending in my favor for the longterm....there's a lot of artwork in pairs trading and we're all different
yobo ive done some simple testing on the forecast function and I didn't find anything of predicitive value, however as GGSAE suggests a combination of different measurements might work. Hey Emilo at what point of stdev would you take a loss if you've entered at 2stdev, do you take a loss if nothing has changed(no news released) it seems others here will scale into the pair if it deviates more. Interested to hear how other pair traders trade management, do you guys have equal $ amounts on each side, or beta adjust, or any other risk controls.
"yobo ive done some simple testing on the forecast function and I didn't find anything of predicitive value, however as GGSAE suggests a combination of different measurements might work." THe forcast function is exactly that a forcast function based on linear regression. if you plug in the right numbers it will forcast a value. I plugged it in to forcast a pair ratio given the trends of the pair ratio mean and correlation. Actually kind of interesting when you do it for a significant sample size of pairs. You do not get what you would expect all the time. Anyway sitll holding strong on the cf/pot trade. I danced around with CF last Friday capturing some small profits on scalps but at the end of the day balancing out the pair with a doubling down. Still expecting the spread to come back in a bit in the short run.
If your strategy is defined as entering when the spread is at 2stdevs with setting a stop loss at 2.5 stdevs (just as an example) then setting a profit target at or near the mean makes sense (risking 1/2 stdev to make 2 stdevs). However it seems if an entry at 2stdevs is good then >2stdevs for entry could be even better. Adding a layer here (spread>2stdevs) makes sense if you still believe (for a reason you have tested, fundementals, technicals,etc.) that the pair is still going to converge. I guess it just comes down to the system you have developed and tested out.
How is the correlation of the pairs being determined? I understand statistical correlation and how it is calculated, that is not my question, my question is more geared towards what variation (if there is any) of price is being used. Is it strictly price, variation of price from previous close to current close (last), or variation of price from open to close (last)? Or some other variation that i haven't thought of. As you can imagine any one of these could be "correct" but i am having a hard time finding what websites use for their correlation.