Imagine that percentage is a parameter p that you want to optimize. So you'll get some historical data and run montecarlo simulations. Letting the data show us the way is the obvious way to go, but we are assuming a lot of things without even knowing.
We already know that returns do not follow a normal distribution. The problem is to find a suitable distribution to fill this gap. If the tail is "too heavy", the option market couldn't even exist at all (at least without another pricing model), and the past wouldn't inform us enough (the sample average is not an unbiased estimator for the population mean in some cases, for instance). So finding the "right tail" (goldilocks principle) is one option.
Remember the p parameter? We could take the worst event ever (in the backtest) and plug some value p0 in order to survive a two-fold drawdown (damn it, let's make it ten-fold). If Mandelbrot was right in the 60's, for example, and the prices do follow some kind of "stable Paretian distribution", we're screwed anyway... in the long run (it's important to mention).
But I'm speaking from a theoretical point of view... the things I'm saying works only asymptotically. In the "real world", people trade for some years and some get rich exploiting market "inefficiencies" (the person in the right place and time with the right approach "almost surely" exists).