Coefficient of Variation Calculator

It needs a true zero, and temperature proves it

The CV is taught as the answer to “you cannot compare a standard deviation in kilograms with one in seconds”. Dividing by the mean cancels the units, so the result is a pure number. That much is true.

What cancels the units does not cancel the zero. Take five temperatures: 50, 60, 70, 80 and 90°F.

In Fahrenheit the CV is 22.59%. Convert to Celsius — the same five temperatures, described differently — and it is 41.61%. Convert to Kelvin and it is 2.99%. Three answers, a factor of 14 apart, for physically identical data.

The reason is the denominator. The standard deviation is unaffected by shifting a scale — it measures spread, and moving everything by a constant does not change the spread. The mean is affected, because it is a location. So the ratio changes whenever the zero moves, and on Fahrenheit and Celsius the zero is a historical accident.

The rule is that the CV needs a ratio scale — one where zero genuinely means “none of it”, and where “twice as much” is a meaningful statement. Length, mass, duration, count, money, concentration and Kelvin all qualify. Celsius, Fahrenheit, calendar years, IQ, standardised test scores and any index rebased to 100 do not.

The Kelvin figure is the only defensible one of the three, because 0 K is the real absence of thermal energy. It is also uninformative — 2.99% makes a 40-degree spread sound like nothing, which is the honest consequence of measuring variation relative to absolute zero.

It breaks near zero, well before it breaks at zero

The second failure is more common than the scale trap, because it arrives gradually.

The CV divides by the mean, so it grows without bound as the mean approaches zero. Its undefined point is a single value, but its unusable region is a whole neighbourhood: when the mean is within a standard deviation or two of zero, small changes in the data produce large changes in the CV. Two samples from the same process can report wildly different CVs.

Data that straddles zero is worse, not just noisier. Positive and negative values cancel in the mean while both contribute to the spread, so the ratio can be enormous, tiny or negative for perfectly ordinary data. Profit and loss, temperature changes, returns that can go either way, and any difference between two measurements all fall into this class.

A negative CV means the mean is negative. That is all it means. It is not a measure of anything, and it certainly does not indicate negative variability, which does not exist. Software that returns it is doing arithmetic, not statistics.

The tool warns on both conditions — the straddle and the proximity — because neither is visible in the output number. In both cases the right report is the standard deviation on its own, in the data’s own units, with the mean stated alongside so a reader can form their own ratio if they want one.

Where it does real work

Used on the right scale it is genuinely valuable, and in some fields it is the primary quality measure.

Laboratory precision, where it is called the relative standard deviation. Repeat an assay on the same sample and the RSD summarises reproducibility in a way that is comparable across analytes at wildly different concentrations. Acceptance limits are usually written directly in CV terms — under 15% for a bioanalytical method, tighter near the middle of the range.

Manufacturing consistency. A 0.1 mm standard deviation is excellent on a 100 mm part and unacceptable on a 1 mm one. The CV puts both on one scale, which is why process capability work leans on it.

Finance, as risk per unit of return. An investment averaging 8% with a 12% standard deviation has a CV of 1.5; one averaging 3% with a 4% standard deviation has 1.33 — less volatile per point of return despite the lower absolute risk. This is the Sharpe ratio’s logic inverted, and it fails precisely when average returns approach zero, which is when people most want it.

Comparing variability across groups with different averages. Rainfall in a wet region and a dry one, spending across income brackets, response times on fast and slow endpoints. The CV asks “which is more variable relative to its own size?” and that is often the real question.

A connection worth knowing. For lognormal data the CV is √(e^σ²− 1), depending only on σ and not on the mean at all. That is why the CV is stable across lognormal populations of very different sizes, and it is the theoretical reason it works so well on incomes, concentrations and durations — the quantities that are lognormal in the first place.

Reporting it without inviting a wrong reading

Four habits, each preventing a specific misreading.

Say which standard deviation. Sample (÷ n − 1) and population (÷ n) give different CVs, and the gap is material at small n — at n = 5 the sample version is 12% larger. The standard deviation calculator covers why the correction exists.

Give the mean and the standard deviation alongside. A CV of 30% could be 3 on a mean of 10 or 3,000 on a mean of 10,000, and those are different situations. The ratio discards the information that tells a reader which one they have.

Do not compare CVs across different measurement scales. This follows from the first section but is worth stating as a rule: two CVs are only comparable when both quantities sit on ratio scales. Comparing a CV of weights with a CV of test scores is arithmetic without meaning.

Do not treat CV thresholds as universal. “Under 10% is good” is a field-specific convention, not a fact. A 10% CV is poor for an analytical balance and outstanding for biological replicates. Compare against what is typical for your measurement, the way you would with a correlation coefficient.

The geometric CV, for data that is genuinely multiplicative

When data is lognormal — incomes, concentrations, durations, particle sizes — there is a version of the coefficient of variation that fits the process better than the ordinary one.

The ordinary CV of lognormal data is √(e^σ² − 1), which depends only on σ and not on the mean at all. That is already a useful fact: it is why the CV is stable across lognormal populations of very different sizes, and it is the theoretical reason the statistic works so well on exactly the quantities it is most used for.

The geometric CV is simply e^σ − 1, computed from the standard deviation of the logs. It reads directly as a multiplicative spread: a geometric CV of 0.30 means values typically fall within a factor of about 1.3 of the median, up or down, symmetrically on the log scale.

That symmetry is the reason to prefer it. The ordinary CV puts a fixed percentage band around the mean, which on right-skewed data reaches below zero long before it reaches the top of the distribution. The geometric version has no such problem, because a multiplicative band cannot cross zero.

It is standard in fields that log their data as a matter of course. Analytical chemistry, pharmacokinetics and microbiology quote geometric CVs, and comparing one against an ordinary CV from another source is comparing two different quantities that share four letters.

The check for which you have: if the source reports a geometric mean, the CV beside it is almost certainly geometric too. If it reports an arithmetic mean, it is the ordinary one.

Sources and methodology

References for the statistic and its scale requirement.

• NIST/SEMATECH e-Handbook — measures of scale, including the coefficient of variation
• Stevens, “On the Theory of Scales of Measurement”, Science 103, 1946 — the nominal/ordinal/interval/ratio distinction this page turns on
• FDA bioanalytical method validation guidance — an example of acceptance limits written directly as CV thresholds
• Standard Deviation Calculator — the numerator, and the sample-versus-population choice
• Lognormal Distribution Calculator — where the CV is √(e^σ² − 1) and depends only on σ
• How we calculate — methodology

Method. The CV is computed as the standard deviation over the mean, with the sample or population divisor as selected, and quoted as both a ratio and a percentage. Two conditions are checked that the number itself cannot show: data straddling zero, where the mean cancels while the spread does not, and a mean within a few standard deviations of zero, where the statistic is unstable rather than merely undefined. The unit toggle recomputes from the original values rather than rescaling the CV, which is the point — the CV is not a rescalable quantity. The suite asserts that the same five temperatures give 22.588%, 41.609% and 2.985% in Fahrenheit, Celsius and Kelvin, that shifting a dataset leaves the standard deviation unchanged while changing the CV, and that the CV of lognormal data equals √(e^σ² − 1) independently of the mean. That engine is verified on every change against 73 assertions. The count and the per-case breakdown are published on the formula verification page.

Related calculators

Where this goes next:

More in Math, or browse all calculators.