neveragain, ive done tests on eft's, indices etc... found high winning rate, however lower profit per trade, but I guess you could you use size and scale down as you suggest as tail risk in those spreads are almost non-existant.
I think correlation is a good & simple tool to see how well two shares have moved in unison historically, but like many suggest a high correlation doesn't equal a good pair to trade, like you say yobo there is a difference between correlation and reversion to the mean probability, i have a few pairs that have above 85% correlation, however I wouldn't trade them as there 2 stocks in different industries. Also all statistical tests are done looking back, we have to trade the hard right edge of our charts and look into the future, looking at historical data is a good tool amongst other tools, however relying on it soley is like driving around with a 10yr old street map, things are still the same, however there are subtle differences.
There has too be a bit of common sense in pair trading, for example you may run tests over a pair for eg C/FNM and find highly predictive divergence/convergence oscillations, however you would be gambling to trade this pair going forward. Ive got no financials on my radar at the moment as I think the risk of an outlier event is elevated in that sector. Corporate scandals, concentrated business risk, lawsuits, currency movements, mergers, aquisitions, changes to dividend policy, capital raising, spinoffs, among many other things can affect stock prices going forward that no amount of backtesting will predict.
I think pair trading is a balance of science, art & application.
