Realised volatility is computed from returns over some period. Simple enough — except that the period, the sampling interval and the scaling to an annual figure are all choices, and reasonable choices produce meaningfully different numbers for the same market.
The window
Short windows are responsive and noisy; long ones are stable and slow. Neither is correct in general — the right window is the one matching the horizon of the decision. Measuring 90-day realised volatility to size a position you will hold for two hours answers a question you did not ask.
The honest practice is to compute several and look at whether they agree, which is what a volatility cone does systematically.
The sampling interval
Daily closes are the convention. Higher-frequency sampling captures intraday movement that daily closes miss entirely — an asset that swings violently and closes flat looks calm on daily data and is not.
Higher frequency also picks up microstructure noise: the bid-ask bounce inflates measured volatility at very short intervals without any real movement occurring. So finer is not simply better, and the interval should be stated alongside the number.
The annualisation factor, which is where crypto trips people
Scaling a period volatility to an annual figure multiplies by the square root of the number of periods in a year. Equity convention uses roughly 252 trading days.
Crypto trades every day. Using 252 on a market with 365 days produces a number about 20% too low, consistently, invisibly, and in a way that makes every comparison against an equity figure wrong in the same direction.
This is a close cousin of annualising a funding rate with the wrong interval: an arithmetic constant, carried across from another market, producing an authoritative figure that is wrong by a fixed multiple.
Reporting it honestly
Window, sampling interval and annualisation basis, stated with the number. Without those three, a realised volatility figure cannot be reproduced or compared, which makes it a claim rather than a measurement.