Originally posted by man
hallo to everbody,
I am new to ET and this is my first post. I find the discussion on pairstrading pretty interesting, since I've been doing some research in that field for several months.
Le me give you a brief overview of what I did. First thing was that I tried to find a pure mathematical/statistical approach to the whole subject. so i took the s&p 500, formed all possible pairs (125.000) and tested a simple startegy using bollinger bands. i did that to have statistical material to play with. so i was able to sort the results by the different GICS-levels. it was interesting then to see that there were sectors that performed quite well, like banks with a SharpeRatio for the whole sector in the area above 1.2. it was no surprise to find that boring industries with rather homogenous companies like oil and financials did best in a test with all possible pairs trading. unfortunately it became obvious that the edge was diminishing at the end of the nineties.
so we entered next step which was to select the pairs instead of taking all of them.
we used two hand full fundamental data like pe ratio, priceToBook, sales to debt ratio and stuff of that kind and tried to improve the strategy by using stops, some filtering and take profits. we came to portfolios that seemed not loose their edge. but we still were in low sharpe ratios of about 0.9. and we had several weaknesses in our model.
first we assumed that the s&p members were the same in the mid nineties as they are now. this is not the case. since 1993 there were about 300 changes within the structure of the s&p.
second we had a natural survivor bias. bankrupt and unlisted stocks are not included in the analysis.
third we did our fundamental analysis at the end of the data. so we were kind of prophets for the similarity of two companies.
now to get rid of these weaknesses we went one step back and gave up the fundamental comparison and said we test weekly which pairs out of the universe of all pairs which are in the same subindustry had the best sharpe ratio in recent time and traded this for the next couple of weeks. doing this we got rid of our fundamental comparison problem. we could test a sharpe of about 1 with this dynamic retraining. still we have survivorship bias in the fundamentals. we try to get rid of that by building a database that contains all changes and all data about all stocks in the sp1500, since we feel it is better to have more stocks for each sector than the sp500 offers.
the actual reason for me posting here is that i am still unsatisfied with our strategy. i tested quite different things from correlationDependentBands to modifiedBollingers with filters in volaDifferences, shortTermCorrelationFilters, RSI-DifferenceFilters and other stuff. i find it very strange not to be able to efficiently trade highly similar and correlated pairs consistently. it seems that correlation and tradeability do not go hand in hand. at least with my current tools. it sems to me that it is hard to find one strategy - no matter how flexible you build it - that works consistently for longer on a daily basis.
we are still considering whether we should trade the pairsstrategy i mentioned, since the results seem rather low for the effort put in.
sorry for this long inital mail. i hope i am right placing this in this thread.
peace
