We backtested 1,000 strategies that we knew contained no signal at all. More than half the time, the best of them had a Sharpe ratio above 1.0. A simulation study of selection bias, the expected maximum Sharpe ratio, and why a parameter sweep flatters you less than you fear.
