You will see a volatility reading described as sitting in the ninetieth percentile. It sounds precise. It is incomplete, because a percentile is a rank within a set and the sentence has not told you what set.
The window changes the answer
Ninetieth percentile of the last thirty days is a statement about the last month. Ninetieth of the last three years is a statement about a market regime. A reading can easily be extreme against a calm month and thoroughly ordinary against a long history that includes a crisis.
This is not a subtle difference. Across a long enough window, the top decile is dominated by a handful of genuinely violent episodes, and clearing it means something. Across thirty quiet days, clearing it means the market got slightly less boring.
Short windows drift
A rolling window that is too short does something worse than being unrepresentative: it adapts. After a fortnight of elevated volatility, a thirty-day window has absorbed the new level and stops calling it elevated. The ranking quietly recalibrates to whatever has recently been normal, which is precisely when you wanted it not to.
Longer windows resist that, at the cost of responding slowly to a genuine regime change. There is no window that is right for both, which is an argument for publishing the window rather than picking one and hiding it.
What to do with this
When a reading is quoted with a percentile, find the window before you weight it. If the source does not state it, treat the number as decorative. And prefer sources that carry the window in the data itself, because that is the only version that stays correct when the number is copied somewhere else.