One report gives an asset's volatility as 1%; another gives 16%. Lin suspects one tool is broken. A difference that large seems impossible to ignore, but choosing the smaller number would not explain why the reports disagree.
Check the time unit next to each figure first. Four daily returns will let us calculate a standard deviation by hand, then examine what changes when a daily figure is reported on an annualized basis. The goal is to discover what a number measures before deciding whether to trust or compare it.
What you will be able to do
- Calculate a mean, deviations, variance, and sample standard deviation for a small return series.
- Specify a volatility estimate's return definition, observation window, frequency, and annualization assumptions.
- Distinguish historical, forecast, and implied volatility, and explain what low volatility has not established.
Choose returns before choosing a formula
Here, historical volatility means the standard deviation of returns over a specified window. Standard deviation measures how far observations are dispersed around their mean.[1] First decide which returns enter the calculation. The standard deviation of price levels measures something different.
Changing the unit in which a security is quoted can multiply all its price figures. Returns describe relative changes and are more useful for comparing objects with different price levels. Even then, simple returns and logarithmic returns are different inputs. We use simple returns throughout this chapter rather than silently switching between definitions.
Corporate actions and cash flows also belong in the input check. A split recorded as a crash or an outside deposit recorded as a gain contaminates the series before the statistics begin. Preserve original materials and the processing method. G10: dividends, splits, and adjusted prices supplies the detailed input checks.
All series below are separate synthetic teaching inputs: equally spaced daily simple total returns, with no outside deposits or withdrawals and no additional fees or taxes. Four observations keep the arithmetic manageable. They are not enough to establish that a real strategy is stable.
Calculate one standard deviation from beginning to end
The first four daily returns are −1%, 0%, 1%, and 0%. Express them as decimals: −0.01, 0, 0.01, and 0. Their sum is zero. Divide by four observations and the sample arithmetic mean is zero as well.
The mean describes the average level. To describe dispersion, subtract it from each observation. Because the mean is zero here, the deviations remain −0.01, 0, 0.01, and 0. Adding those deviations would give zero: the positive and negative values cancel. That cannot distinguish this series from a series that never moves.
Square each deviation before adding. Each nonzero square is 0.0001, so the sum of squared deviations is 0.0002. For sample variance, divide by the number of observations minus one. Four observations give a divisor of three, producing approximately 0.00006667.
Take the square root to obtain sample standard deviation: approximately 0.008165 as a decimal, or 0.8165% per day. Variance uses squared units; taking the square root returns the measure to the scale of the returns themselves.
The result describes dispersion in these observations. It is not a daily loss allowance, and it does not promise that tomorrow's return will fall within 0.8165% of the mean. A descriptive statistic does not by itself supply a probability model or a guaranteed range.
Why divide by three rather than four? We chose a sample estimate intended to describe a broader return process, with the mean estimated from the same observations. If the task is instead to describe these four values as the complete population, dividing by four gives a population standard deviation of about 0.7071%. The definitions answer different statistical tasks. Choose the matching calculator or spreadsheet function and record the choice.
Under relevant conditions, such as independent and identically distributed observations, the sample-variance divisor produces an unbiased variance estimate. That does not make its square root an unbiased estimate of standard deviation, or establish that financial returns satisfy those conditions. You need not prove this here. You do need to recognize that a software default is a method choice.
Now double every return: −2%, 0%, 2%, and 0%. The mean remains zero. Squared deviations become four times as large, and their square root becomes twice as large. The sample standard deviation is approximately 1.6330%. This controlled change shows how the statistic responds to larger movements without changing their pattern.
A shared endpoint does not imply a shared experience
Consider two independent synthetic value paths, each starting at 100. The flat path continues through 100, 100, and 100. The moving path goes through 110, 99, and 100. Both finish where they started.
The flat path's three returns are zero, so its sample standard deviation is zero. The moving path has returns of 10%, −10%, and approximately 1.0101%. Its dispersion is not zero. Moving from 99 to 100 requires a gain of 1 divided by 99, not 1%. The denominator does not change just to make the endpoint calculation convenient.
Standard deviation also does not preserve the full time order. Rearranging the same return observations leaves their dispersion unchanged, even though the timing of a fall from a peak or the point at which money is needed can change. Keep drawdown, cash needs, and liquidity beside volatility rather than expecting it to describe the entire path.
This gives you two separate questions to ask of a performance report. Did the return series vary substantially? And could the account meet its obligations along the actual route? A standard deviation addresses the first question within its specified inputs. It cannot replace the second.
How 1% can become roughly 16%
Suppose a different, longer sample has a daily standard deviation of 1%. A reporting convention that assumes 252 trading days per year multiplies it by the square root of 252. The result is approximately 15.87% annualized volatility. The number 252 is a conversion convention here, not a claim that every market and calendar year contains exactly that many trading days.
Why a square root? If per-period returns have equal variance and are uncorrelated, the variance of their sum increases in proportion to the number of periods. Standard deviation then increases with the square root of that count. Sharpe's discussion of aggregation explicitly assumes added returns and zero serial correlation.[2]
Actual compounded simple returns are not the sum of daily returns. Square-root annualization is therefore usually a standardization convention or approximation, not an unconditional identity for the standard deviation of a year's compounded return. Using logarithmic returns does not remove the need to examine correlation and variance stability.
Successive returns may respond to shared influences and be correlated. Volatility can also change between regimes, making the recent window a poor description of the whole year. Multiplying by a familiar square root can hide those structures. Two decimal places in the output do not make an annualized estimate more certain than its input sample.
This may resolve Lin's original apparent contradiction. A 1% daily figure and a rounded 16% annualized figure could be consistent. But that explanation remains conditional until the return definition, frequency, window, and method are checked. If those differ too, the time conversion alone does not reconcile the reports.
Window length and observation frequency are different fields
“Daily returns over the past year” may contain hundreds of observations. “Monthly returns over the past year” contains about twelve. Both descriptions cover a year, but they do not have the same frequency. A monthly estimate cannot use a daily conversion factor merely because the desired output is annualized.
Nor does converting a four-day demonstration into annualized units make it a reliable estimate of annual risk. The conversion changes the reported scale, not the amount of evidence. Always keep the observation count visible next to the annualized result.
For another window on the same object, suppose two daily returns are both 0%. Their sample standard deviation is zero, while the earlier four-day sample produced 0.8165%. The object has not changed identity; the window has changed. Selecting only a quiet period can produce a reassuring-looking number without establishing a quiet future.
Infrequent quotes or stale valuations can also flatten a record. Filling missing values with zero may replace “no usable observation” with “no price movement.” Check valuation and missing-data policies before interpreting low volatility. A blank input and an observed zero are different claims about what is known.
What the statistic leaves unanswered
Standard deviation includes both upward and downward deviations from the mean. It does not separately report extreme downside loss, bankruptcy probability, or the discount required to sell urgently. High volatility does not imply certain permanent loss, and low volatility does not establish protection of principal.
Return to the underlying rights, the purpose of the money, and failure scenarios when deciding whether a risk can be borne. A small historical dispersion cannot tell you whether cash will be available in full on a payment date. That question depends on more than the variation in reported returns.
An event absent from the sample is not removed from the future merely because the sample has been summarized accurately. Distinguish a faithful calculation from a complete account of possible outcomes. The first is necessary for using a statistic; it does not automatically deliver the second.
Historical, forecast, and implied describe different origins
Historical volatility is calculated from returns that have already occurred. Forecast volatility estimates future variation using a model and assumptions. Historical data may be an input to that model, but its forecast remains an estimate rather than an observed fact.
Implied volatility is obtained from option-market prices together with a pricing method and other inputs. CME distinguishes information inferred from options from realized changes already observed.[3] The estimate also reflects pricing and risk-compensation considerations. It is neither a guarantee about future variability nor a forecast that the underlying price will rise.
Options on the same underlying can have different implied volatilities at different strikes and expirations. Ask which type, horizon, and method a displayed volatility number represents. This chapter does not require option-pricing calculations; recognizing the three origins is enough for the current task.
Recalculate before checking the answers
Explanation question: Why does a zero sum of deviations fail to show that the first return series has no volatility?
Error question: A report calculates a 2% monthly standard deviation from twelve monthly returns, then multiplies by the square root of 252 to annualize. Identify the method mismatch before doing any arithmetic.
Write your reasoning first. For the explanation question, positive and negative deviations around the sample mean cancel. Squaring and summing retain their magnitudes instead. For the error question, the report mixes monthly observations with a daily factor. Under the corresponding square-root convention, multiply 2% by the square root of twelve, obtaining approximately 6.93%. Disclose the assumptions, including dependence between observations. The result still does not guarantee next year's realized risk.
Changed-condition problem: Three separate teaching daily returns are −2%, 1%, and 4%. Calculate their mean, deviations, and sample standard deviation. What should the record say if only one usable observation remains?
Work through the steps before continuing. The mean is 1%. The deviations are −3, 0, and 3 percentage points. In decimals, their squared deviations sum to 0.0018. Dividing by two gives a variance of 0.0009, whose square root is 0.03: a 3% daily sample standard deviation.
With one observation, the sample divisor is zero. Mark the result as unavailable because the sample is insufficient. Reporting zero volatility would invent an estimate rather than acknowledge that the selected calculation cannot be performed.
Save a volatility checking card
Choose a fictional return series or one with a reliable source. Record the data date, beginning and ending window, observation count and frequency, simple or logarithmic return definition, dividend and cost treatment, missing-data policy, sample or population divisor, and calculated result. If annualized, include the factor and its assumptions.
Use the four-value example to check a tool before extending it to a longer series. An AI assistant can explain one row's mean and deviation; check the units yourself. A 1% return is entered as 0.01 in decimal notation. Entering 1 without changing the unit convention makes a result one hundred times too large.
Add one question the metric cannot answer, such as whether an asset can be converted into enough usable cash before a required payment. That unfinished question belongs on the card, so the convenient number does not quietly displace it.
Restore the time unit and the method
Trace a volatility figure back through the return inputs, processing, window, frequency, divisor, and annualization. Compare results only after those choices are aligned or their differences explained. A disagreement can reveal different questions rather than a broken calculator.
Then restore the wider risk picture. Missing observations, stale values, and failures absent from history are not resolved by a standard deviation. Keep drawdown, liquidity, and permanent-loss questions in the record. That is how a useful measure of dispersion remains useful without becoming a promise of safety.
References
- NIST: Trimmed Standard Deviation. Only the opening ordinary standard-deviation definition and sensitivity to extreme values are used; this chapter does not trim observations.
- William F. Sharpe: The Sharpe Ratio. 1994; “Time Dependence,” including additive aggregation, serial correlation, and compounding limitations.
- CME Group: Understanding the CVOL Index. Realized changes versus option-implied volatility; conceptual context only.
Sources checked October 30, 2026. AI-assisted draft; independent AI review against the accepted Chinese text is complete; human financial review and target-reader testing remain pending. Every numerical series is original synthetic teaching material.