In plain language — Volatility says how dispersed price moves are under a stated measure. It does not say whether the next move will be up or down, and it does not exist as one number detached from data, horizon and method.
In finance, volatility generally describes the variability of an asset's returns. A price that rises steadily may have lower volatility than one that alternates sharply between gains and losses, even if both finish at the same level. High volatility means greater dispersion in the measured coordinate; it does not automatically mean a bear market, high total risk or a profitable opportunity.
Before comparing two numbers, ask: volatility of what, calculated from which observations, at what frequency, over which window, with which estimator and in which unit? Without those coordinates, “20% volatility” is incomplete.
Prices, returns and direction
A standard deviation applied directly to price levels depends on the scale and trend of the series. Comparisons across periods or instruments therefore often use simple or logarithmic returns:
rₜ = Pₜ / Pₜ₋₁ − 1gₜ = ln(Pₜ / Pₜ₋₁)The definitions are not identical. They are close for small moves but differ over wide intervals and in aggregation. Dividends, interest, futures rolls and corporate actions must match the question being asked: price volatility, total-return volatility and the volatility of a continuous futures series are not the same object.
A common measure is the sample standard deviation of returns, with equal weights. Here n counts returns (at least two), not prices; r̄ is their arithmetic mean:
s = √[Σ(rₜ − r̄)² / (n − 1)]It is the deviations from the mean that are squared, not uncentred returns. Reversing all signs leaves dispersion unchanged: volatility does not indicate the direction of the next move. Two distributions can share the same standard deviation while differing substantially in skewness, tails, jumps and maximum loss.
Reproducible example: five prices, four returns
Use five hypothetical closes from consecutive sessions, at the same observation time and in the same currency, without dividends or other adjustments. The frequency is daily, the window contains four simple returns, weights are equal and the denominator is n − 1.
| Close | Price | Return from the preceding close |
|---|---|---|
| 0 | 100.00 | — |
| 1 | 102.00 | +2% |
| 2 | 102.00 | 0% |
| 3 | 104.04 | +2% |
| 4 | 104.04 | 0% |
For example, 104.04 / 102 − 1 = 0.02 = 2%. The four decimal returns are 0.02; 0; 0.02; 0. Their mean is 0.01, or 1% per day in the sample.
Deviations from the mean are +0.01; −0.01; +0.01; −0.01; their squared sum is 0.0004. Thus:
s² = 0.0004 / 3 = 0.00013333…; s = √s² ≈ 0.011547 = 1.1547%Dividing by n = 4 would give 1%: a different convention, not a rounding difference. Prices rose, but volatility measures dispersion around the mean. Four returns illustrate the calculation; they do not provide a reliable estimate of future risk.
“Historical” and “realized” require a convention
In market usage, historical volatility and realized volatility are sometimes used interchangeably for a measure obtained from past data. In the literature, realized volatility may refer more specifically to the square root of the sum of intraday squared returns. The sum is realized variance (RV = Σg²); its square root is realized volatility over the interval. This is different from the sample standard deviation, which subtracts the mean and divides by n − 1. Frequency and interval must be stated.
A reproducible specification states at least:
| Coordinate | Example choices | Why it matters |
|---|---|---|
| Series | close, total return, mid quote, continuous future | changes the measured object |
| Return | simple or logarithmic | changes composition and scale |
| Frequency | five-minute, hourly, daily | noise and information differ |
| Window | 20, 60 or 252 observations | responsiveness and stability differ |
| Estimator | close-to-close, range, realized variance | uses different information |
| Weights | equal, exponential or another rule | changes memory of the past |
| Annualisation | factor and calendar | supports comparison only under assumptions |
A 20-session estimate reacts faster than a 252-session estimate but is also more variable. Intraday estimates can be affected by microstructure, trading hours, missing data and overnight jumps. Parkinson's estimator uses the high and low; other estimators combine open, high, low and close. None is automatically superior outside its assumptions.
Annualising without hiding the assumptions
If s_d is a daily standard deviation, a common convention is:
σ_ann = s_d × √NN is the assumed number of daily periods per year, such as 252 sessions or 365 days. It must match the data frequency and calendar. In the example, 1.1547% × √252 ≈ 18.33%. The standard deviation is annualised, not the 1% mean.
The justification starts with the variance of a sum: add the variances and twice the covariances between periods. If increments have equal finite variances and zero covariances, variance over N periods is Nσ² and standard deviation is √Nσ. Normality is not required for this identity.
Logarithmic returns add over time; simple returns compound by multiplying 1 + r. For simple returns, the formula remains a scaling convention, not an exact identity for the dispersion of the compounded annual return.
Autocorrelation and unequal variances require additional terms. Clustering alone does not invalidate every square-root-of-time rule, but projecting a recent estimate over a year needs care: a conditional forecast must specify how variances evolve. Jumps, calendar and tail risk still require assessment. 18.33% is neither a predicted move nor a guaranteed interval.
Standard deviation, range and ATR are not synonyms
Volatility can be described by different families of measures:
- return standard deviation measures dispersion around the mean of the selected series;
- realized variance sums squared returns over an interval;
- range-based estimators use highs and lows, sometimes open and close;
- Average True Range (ATR) aggregates true range,
TR = max(H − L, |H − C_previous|, |L − C_previous|), in price units;HandLare the bar's high and low, andC_previousis the preceding close. Length, initialisation and averaging method must be stated; - implied volatility is backed out of an option price through a model and is not a statistic calculated solely from past returns.
An ATR of two currency units is not 20% annualised volatility. To compare instruments, an absolute range may need to be normalised by price; even then it remains a different measure from standard deviation. Implied and realized volatility keeps the two coordinates separate.
Regimes describe; they do not predict
Volatility often clusters: turbulent periods may follow turbulent periods and quiet periods may persist. This property motivated conditional models and regime filters, but it does not make the next value certain. An estimate uses already observed information and can react late to a jump.
Labels such as high volatility, low volatility, compression and expansion only make sense relative to a stated distribution, horizon and benchmark. What is high for one stock may be ordinary for another asset. Compression does not guarantee a breakout; expansion does not reveal direction.
Volatility is also not liquidity. A market can show small observed returns while its order book is thin and gap risk is high; another can be volatile and deeply liquid. Spread, depth and executability need separate measures.
Risk and position sizing
Volatility can help put exposures with different scales on a more comparable basis. A simple approach reduces quantity when an estimated volatility rises:
indicative exposure ∝ risk budget / estimated volatilityThis is a calibration principle, not a maximum-loss guarantee. The estimate can change, returns can contain tails and gaps, correlations may rise and a liquidation may occur at worse prices. Volatility targeting therefore requires leverage limits, stress tests, costs and rebalancing rules.
Likewise, adapting a stop to ATR does not turn ordinary range into an objective invalidation threshold. Stops and position sizing must start from the position's economic risk, payoff and liquidity. Volatility is one coordinate among several.
Common error — Comparing “20%” and “25%” without knowing whether they are implied or realized, at what horizon, over which window and with which annualisation. Decimal precision cannot repair a missing definition.
Checklist for reading a volatility measure
- Does the series contain prices, returns or total returns?
- Are returns simple or logarithmic?
- Which frequency, window and timestamp were used?
- Which estimator is used and how does it treat gaps, overnight moves and missing data?
- Is the number daily, annualised or expressed in price units?
- Which annualisation factor and calendar apply?
- Is it realized, a statistical forecast or a model-implied parameter?
- Which tails, jumps, asymmetries and liquidity risks does it omit?
- Does the comparison use the same convention on both series?
- Does the decision remain robust if the estimate changes abruptly?
Sources
U.S. Commodity Futures Trading Commission, CFTC Glossary — Volatility — institutional definition and distinction between historical and implied volatility.
The Options Industry Council, Volatility & the Greeks — historical volatility, implied volatility and option sensitivities.
National Institute of Standards and Technology, Measures of Scale — standard deviation and robust measures of dispersion.
Michael Parkinson, The Extreme Value Method for Estimating the Variance of the Rate of Return, The Journal of Business, 1980 — a range-based estimator.
Torben G. Andersen, Tim Bollerslev, Francis X. Diebold and Heiko Ebens, The Distribution of Realized Stock Return Volatility, Journal of Financial Economics, 2001 — construction and distribution of high-frequency realized volatility.
Cboe Global Indices, Cboe Volatility Index Methodology — one forward-looking measure derived from option prices under a published methodology, not a universal volatility measure.
National Institute of Standards and Technology, Propagating Error — variance of sums and the covariance term.
J.P. Morgan/Reuters, RiskMetrics — Technical Document, 4th ed., 1996, §5.2.2, pp. 85–88 — time scaling and limits of volatility projections.
Federal Reserve Board, Expected Stock Returns and Variance Risk Premia, 2007 — realized variance from intraday squared returns.
Fidelity, Average True Range, Calculation section — true range definition and ATR aggregation.