A good example of the sample size thing: I am currently "live" testing (ie not back testing but recording each day) a method which, after 133 races, is winning 43.8% on turnover. Promising, yes. But I know from past experience that I need far, far more than that.

Preferably, you want a sample of 10,000. That's a bit unrealistic though. IMO 1000 is bare minimum, but I've seen methods change their profile after more selections than that. 3000 is a number where, from my experience, it does seem to settle down. So, I'll be waiting on 3000 selections before I get too excited about this.