No disrespect taken, but I do think you are missing the point. If a trade moves towards 3x stddev than I think you have an obvious exit point, because - in my opinion- this will be due to some clear fundamental change in the relative fundamentals of the companies. They won't be 'pairs' anymore and you should be out of the trade. The practical problem is, rather, what happens when the ratio keeps 'snaking' along the 2x std dev and taking the mean and bollinger bands higher. This can go on for a while, so that even when it goes back to the mean or a 1x stddev, you are facing a heavy loss. If you stick purely to stddev than this approach will hurt you in this scenario. I was suggesting incorporating a % from mean measure in conjunction with stddev rather than replacing it. It will help deal with the above issue, and it will help keep you out of trades where the stddev has narrowed considerably short term, only to suddenly breakout against you after you placed the trade. I'm interested in formulating a trading algorithm which could be used to deal with the above scenario. Anyway, good luck with what you decide to do.
Hope I'm not barging in, as I find this thread interesting. total_keops, Just curious. What exact series is in the above plot (spread, moving hedge ratio, one series)? And what exactly are you using as the mean in the bottom chart? 1) One fixed spread mean calculated over the entire time window snapshot? (which case you would have look forward bias) 2) A running mean (like MA)? I think a lot of some of the conclusions mentioned about 'railing' against 2 sigma have to do with how the above is being considered. If you divide up the range into one fixed parameter mean for a training region, and keep that value as a standardized z reference for an out of sample region, I think you will see a completely different viewpoint on the entire discussion (which is more inline with what academic has been discussing).
I think you're right. I'm looking at a standard deviation that's calculated over a moving window of about 2 years. When the spread pushes past 2 stddev, it doesn't catch up so much as it does for you guys. Obviously, I'm looking at pairs with a long term cointegration (years) and I'm looking to hold onto trades for days up to months.
Above is a spread with the ratio calculated as the OLS beta (hedge ratio from cointegration) over a period of 167 trading days. In that case the spread is HCBK - 1.3*SUSQ for all the period. edit: the ratio is calculated over the entire period. I dont have a forward bias because I dont do backtesting. Below the mean used is 20. The Stdev is a rolling 20 days period. And yes, the parameters will change the decisions.
Out of that trade with a profit. Sold NBR@17.80 Cov BPZ@ 6.97 I didnt wait for the signal because of the upcoming earnings. Shortly after I covered BPZ it decided to tank nicely without me lol. So what, a gain is a gain.
Have you ever tried to backtest this system? I'm curious as to what long term results you get and why you use a 20 day moving average and 20 day moving stddev.
No backtest, none. I trade mostly discretionarily so a backtest does not really make sense as it includes many trades I would not have taken. It is good if you trade fully automated. I guess many highly profitable trades occured in the last year and could skew your average win in $ to the upside.