Hi all, Am about to undertake a research project to identify correlated pairs of US stocks that would be suitable for statistical arbitrage (or pairs trading.) The aim of my project will be to formulate rules that will identify pairs whose spreads (not prices) exhibit "steady and even" mean reversion characteristics. Does anyone have any advice on what identification criteria I should use, what correlation measures will be the most effective (and over what time horizon), or any other criteria that I could take into account in the pairs identification process or any practical advise on the best way to undertake such a project. Any help will be much appreciated Thanks very much for your time and efforts.Nikk.

I'm not an expert for US equity, but in general terms from my experience I would thing about:- whether you care about sectors or not, i.e. do you want to impose clusters on the stocks before pairing?- what you want to mean revert (return, cumulative return, var, ect. relationships)- plain correlation might not be the way forward, lot's of other distances to look at (distance between ret, var, etc paths under various metrics)...- don't get too hang up on statistical test- entry and exit levels are usually not symmetric- make sure you have clean data and if it is to be traded on, tradeable prices (so not the close...)- time horizon is tricky, try to find a horizon that is stable with respect to shifting it through history

Hi Zed,Thanks very much for your thoughts. They have been very helpful Am probably going to impose industry constraints on the pairs they will increase chance of correlation + is an easy way to limit my sample size and avoid sector risk. Will also try to maintain a zero beta balance to avoid significant sector risk. Wnat to formulate identification rules that will help to identify pairs that are likely to have a mean reverting ratio of prices, within a particular range and exhibit mean reversion on a regular basis. Am looking into cointegration tests to identify pairs, but am not sure how to achieve this on such a large scale. I know that it is probably going to be impossible to formulate any hard and fast rules to achieve this end, but I just wanted to know if anyone has any suggestions?

Make fundamental jugdement before you go datamining on the scope of securities and how you pre-filter. If you just run correlations on everything against everything your results will be extremely biased to random results showing up leading to wrong assumptions.One filtering method is to run correlation against sector index for different time intervals. Maybe you could then run a statistical hypothesis that correlations with the index are independant of time intervals. Those stocks who fail the hypothesis you could run through a cointergration model. Just a thought.As a guideline I would also look at the implied volatilities of the stocks you are comparing, to see if you have similar behaviour in future expectations, not just historical data. Z

I do not recommend looking at ratios of prices... These things can behave quite nasty...Grouping stocks is fine, as long as you have enough reason do believe that these groups do exist for real (sounds trivial, but ever tried to back-out industries from return data?).Cointegration is a nice method, but I'd like to repeat my advice, do not get into testing too fast. I know econometrics tend to cover the world knee-deep in tests, but in my experience, that's not how it works.Make sure the series you are looking at has good properties, or at least know the properties. Remember, you not only care about things mean-reverting, you also care about the speed of the mean reversion, the maximum drawdown on the path, etc...I like ziggy's idea about using a second indicator like implied vol. You could run a whole batch of models and require a certain number to agree on a pair.