Crypto Realized Volatility: Measurement Guide

Crypto realized volatility is an ex-post measure of how dispersed an asset’s observed returns were over a specified interval. In its simplest form, calculate returns from one price series, take their standard deviation (or sum squared returns for an intraday estimator), and state the sampling interval, window, venue, missing-data rule and annualisation factor. The result describes past variability; it does not establish future direction or guarantee that volatility will remain at the same level.

This distinction matters because crypto trades continuously across exchanges. Two analysts can use the same ticker yet obtain different numbers by changing the exchange, spot versus perpetual instrument, UTC day boundary, candle construction or treatment of empty intervals. A reproducible method therefore matters as much as the final percentage.

What realized volatility measures

The Bank for International Settlements describes realised (or historical) volatility as a model-free standard deviation of actual returns over a chosen window. It separates this statistical, observed-return measure from implied volatility, which is extracted from option prices and embeds market expectations and risk preferences (BIS, “Volatility concepts and the risk premium”). Realized volatility is therefore backward-looking by construction.

It is not the same as a candle’s high-low range. Range records the largest and smallest quoted or traded prices inside one candle and ignores the path between them. Realized volatility aggregates return changes at a chosen frequency, so a session with several reversals can have high realized volatility even when it finishes near its opening price. Conversely, a single smooth move can show a modest range-based statistic but still produce substantial return variance across finer observations.

Nor is realized volatility a forecast. A rolling estimate can be used as an input to scenario analysis or a risk limit, but any forecast requires additional assumptions or a statistical model. Research on Bitcoin finds that lagged realized variance can contain information about later variance, yet the relationship depends on horizon and model specification; that evidence does not turn a chart reading into a certainty (Hu, Härdle and Kuo, “Risk of Bitcoin Market: Volatility, Jumps, and Forecasts”).

Choose the series before the formula

Write a one-line data specification before downloading candles:

Venue and instrument: name the exchange, market (for example BTC-USD spot or BTC-USDT perpetual), and price field (trade, close, index or mark).
Quote currency: keep USD, USDT or another quote consistent. Converting later can add currency noise.
Clock and cut-off: use UTC or another declared timezone. A “daily” crypto candle is a convention, not a universal session.
Sampling frequency: select daily closes for a broad historical view, or 1-, 5-, 15- or 60-minute observations for intraday measurement.
Window: state the number of observations (for example, 30 daily returns or 90 days of five-minute returns).
Data cleaning: document duplicate timestamps, outliers, exchange outages and whether an interval with no trade is missing or assigned a carry-forward price.

These choices are substantive. A peer-reviewed study using 22 exchanges notes that it synchronised timestamps and used explicit outlier rules; it also shows that return behaviour differs between crypto-to-crypto and crypto-to-dollar pairs (“Cryptocurrencies and stablecoins: a high-frequency analysis”). Another study estimates five-minute log-return volatility separately for each exchange and finds that the same coin can have distinct realized-volatility paths across venues (“Heterogeneity in the volatility spillover of cryptocurrencies and exchanges”).

Calculate daily and rolling volatility

For prices P_t, the continuously compounded (log) return is:

r_t = ln(P_t / P_t-1)

For a rolling window of N daily returns, the sample standard deviation is:

s = sqrt[ Σ(r_i − r̄)² / (N − 1) ]

where r̄ is the window’s mean return. Some risk reports use the population denominator N; either convention is acceptable if disclosed and applied consistently. For a simple daily realized-variance estimate that assumes the mean is negligible, use RV = Σ r_i² and take its square root. High-frequency literature commonly defines realized variation as the sum of squared intraday returns. A Springer study of crypto markets, for example, computes one-day volatility as the square root of summed one-minute log returns squared (Digital Finance methodology).

For intraday data, divide a UTC day into equal intervals, calculate each log return, square and sum them, then take the square root. This captures the day’s path rather than only its close-to-close move. Finer sampling is not automatically better: very short intervals can amplify bid-ask bounce, stale quotes and price discreteness. The Bitcoin volatility literature therefore treats frequency as a trade-off between information and market-microstructure noise (Hu, Härdle and Kuo).

A rolling series repeats the calculation after moving the window one observation forward. A 30-day rolling estimate answers, “How variable were the latest 30 daily returns?” It does not mean the next 30 days will have that volatility. Overlapping windows also create serial dependence in the plotted estimates, so avoid interpreting every small wiggle as a new regime.

Annualize with an explicit convention

Annualisation puts estimates from different sampling intervals onto a common scale under a square-root-of-time convention:

annualized volatility = volatility per interval × sqrt(number of intervals per year)

For daily crypto returns, many researchers use sqrt(365) because the market operates around the clock. A Bitcoin exchange study explicitly annualises daily realized volatility with this factor (Price Discovery of a Speculative Asset). If your organisation uses 252 observations for comparability with traditional assets, label it clearly; the percentage will be lower even though the underlying returns are unchanged.

For five-minute bars, there are 288 intervals per 24-hour day, so a 365-day convention uses sqrt(288 × 365). Do not annualise a statistic already expressed as a one-day sum with a second daily factor. Also distinguish annualised variance from annualised volatility: variance scales by the number of periods, while volatility scales by its square root.

Worked example: comparing two assets

Suppose you collect 30 UTC daily closes for two spot pairs from the same exchange and quote currency. After computing 29 log returns, Asset A has a sample standard deviation of 0.032 (3.2%) and Asset B has 0.021 (2.1%). Using 365-day annualisation:

Asset A: 0.032 × sqrt(365) ≈ 0.611, or about 61.1% annualised volatility.
Asset B: 0.021 × sqrt(365) ≈ 0.401, or about 40.1%.

Within this identical design, A experienced roughly one-and-a-half times B’s return dispersion. That supports a relative risk statement for the observed month. It does not say A will fall, outperform, or remain 61.1% volatile. Check whether one large jump drives the result: recompute with and without the observation, report both, and investigate the timestamp and market event rather than silently deleting it.

Now imagine calculating A from five-minute candles on one venue and B from hourly candles on another. The comparison is no longer controlled. Align the instrument, venue scope, quote, timezone, frequency, window length and annualisation before drawing conclusions. If a composite index is used, preserve its constituent and weighting methodology because aggregation can smooth venue-specific jumps.

Realized versus implied volatility

Historical or realized volatility is computed from what prices did. Implied volatility is inferred from option prices and represents a risk-neutral, forward-looking market price for variance. CME explains that its CVOL indexes derive expected 30-day risk from option prices across an implied-volatility curve (CME CVOL methodology overview). The two measures answer different questions:

Realized: “How variable were returns over the selected past window?”
Implied: “What volatility level is embedded in traded options for a stated future horizon?”

A realized-implied spread can be informative for research, but it is not automatically a trading signal. Maturities, sampling conventions, risk premia, liquidity and jumps may differ. Compare like with like: a 30-day implied measure with a carefully defined 30-day forward realized outcome, not with an arbitrary 90-day historical chart.

What the measure can—and cannot—support

Used carefully, realized volatility can help compare assets, size stress scenarios, set monitoring bands, evaluate whether a backtest used realistic risk, and identify periods of clustered movement. It can also reveal data problems: a sudden venue-only spike may indicate a bad tick, thin liquidity or a contract-specific event.

It cannot establish direction, causation or a guaranteed range of future prices. Standard deviation is not a forecast distribution, and crypto returns can be fat-tailed and jumpy. A low recent estimate can coexist with material tail risk; a high estimate can fall quickly after a shock. Treat the number as one descriptive feature alongside drawdowns, liquidity, leverage, funding, gaps and concentration.

A reproducible research checklist

Record the exact venue, symbol, contract type and quote currency.
Save the data retrieval date, source URL, timestamps and candle-generation rules.
Declare UTC cut-off and sampling interval.
Choose log or simple returns and the variance denominator.
Set the rolling window in observations and calendar days.
Define missing, duplicated, stale and outlier observations before calculation.
Select an annualisation factor (365, 252 or another justified value) and show the formula.
Publish both raw-period and annualised results, with units as percentages.
Run sensitivity checks for a second frequency, window and venue.
Interpret the output as observed variability, then state explicitly what remains unknown about the future.

Following this framework makes a volatility chart auditable. Readers can reproduce the series, understand why another venue differs and decide whether the statistic is fit for their research question. The discipline is simple: define the data first, calculate consistently, and keep measurement separate from prediction.

Sources and further reading

Volatility concepts and the risk premium — Bank for International Settlements, 14 September 2014
CME Group Volatility Indexes (CVOL) — CME Group, Updated 2026
Cryptocurrencies and stablecoins: a high-frequency analysis — Digital Finance, Springer Nature, 2022
Heterogeneity in the volatility spillover of cryptocurrencies and exchanges — Financial Innovation, Springer Nature, 2023
Risk of Bitcoin Market: Volatility, Jumps, and Forecasts — arXiv, Updated August 2021
The Effects of the Introduction of Bitcoin Futures on the Volatility of Bitcoin Returns — arXiv, 8 June 2019

Editorial note: This article is general educational information, not personalized financial, accounting, tax, or legal advice. Product capabilities and obligations can change; verify current facts and consult a qualified professional where needed.