Coefficient of Variation Calculator

What the coefficient of variation measures

The coefficient of variation compares the spread of a data set to its own average. Divide the standard deviation by the mean and multiply by 100, and you get a percentage that says how large the wobble is relative to the level. A series of measurements that averages 500 g with a standard deviation of 5 g has a CV of 1 percent. Another series that averages 20 g with a standard deviation of 0.4 g has a CV of 2 percent, even though its standard deviation is smaller in absolute terms.

That is the whole point of the statistic. Standard deviation carries the units of the data, so a spread of 5 g and a spread of 5 minutes cannot be compared. The CV divides the units away, which lets you line up a lab assay against a delivery time, or a stock's volatility against its return. The NIST Dataplot reference page defines it as the sample standard deviation divided by the sample mean and notes that it should be used for data measured on a ratio scale, where zero means none of the thing being measured.

A CV is not a better standard deviation, just a rescaled one. It throws away the unit and the level, so a reader cannot rebuild the original numbers from it. Report both when the level matters: a mean of 10.2 with a CV of 2.68 percent tells a different story from a mean of 1,020 with the same CV.

Sample or population: the denominator matters

Two formulas differ by one in the denominator. The sample standard deviation divides the squared deviations by n - 1, and the population standard deviation divides by N. The n - 1 version is the default when your numbers are a sample from something larger, because dividing by n understates the spread of the wider group.

The gap is small for large data sets and large for small ones. With the five readings used below, the sample standard deviation is 0.2739 and the population standard deviation is 0.2449, so the two CVs are 2.68 percent and 2.40 percent. At n = 5 that is a 12 percent difference in the answer, which is enough to fail an acceptance limit that sits between them.

Software defaults differ, which is a common source of confusion. Spreadsheet functions STDEV.S and STDEV.P split this way, and many online calculators quietly use the population version. If a reported CV does not match yours, check the denominator before hunting for a typo in the data.

A worked example with five readings

Take five repeat measurements from a scale: 10.2, 10.5, 9.8, 10.1 and 10.4. The sum is 51.0, so the mean is 51.0 / 5 = 10.2. The next step is the deviation of each reading from the mean, then the square of each deviation.

The deviations add to zero, which is a useful check: if they do not, the mean was computed wrong. The squared deviations add to 0.30. Divide by n - 1, which is 4, and the variance is 0.30 / 4 = 0.075. The square root of 0.075 is 0.2739, the sample standard deviation. Finally, 0.2739 / 10.2 = 0.02685, and 0.02685 x 100 = 2.685, so the CV is about 2.68 percent.

Read that number as a statement about the process. A CV of 2.68 percent on a 10 g target means a typical reading sits within about 0.27 g of the mean, so a tolerance of plus or minus 0.5 g would be met most of the time but not always. The same CV on a 1,000 g target means a spread of 27 g, which is a different quality problem in absolute terms and exactly the same one in relative terms.

Why the units cancel

Convert the five readings from grams to milligrams by multiplying each by 1,000, and the mean becomes 10,200 while the standard deviation becomes 273.9. The ratio is unchanged: 273.9 / 10,200 = 0.02685, so the CV is still 2.68 percent. That is the property that makes the statistic portable across units and across scales of the same quantity.

Multiplying every value by 100 gives the same result, and the table below makes the point plainly. Both columns hold the same five readings, one in grams and one in centigrams, and both produce a CV of 2.68 percent.

The cancellation only works for scales that start at a true zero. Grams, meters, seconds, dollars and kelvin all have one. Fahrenheit and Celsius do not, and the next section shows what goes wrong there.

Comparing things that are not alike

The CV earns its keep when two data sets share no units. A lab comparing two balances is the clean case. Scale A weighs samples around 500 g with a standard deviation of 5 g, so 5 / 500 = 0.01 and the CV is 1.0 percent. Scale B weighs samples around 20 g with a standard deviation of 0.4 g, so 0.4 / 20 = 0.02 and the CV is 2.0 percent. Scale B has the smaller standard deviation and the worse relative precision, and only the CV shows that.

Finance uses the same arithmetic with returns. A fund with an average annual return of 8 percent and a standard deviation of 12 percent has a CV of 150 percent, because 12 / 8 = 1.5. A second fund returns 15 percent with a standard deviation of 18 percent, so its CV is 120 percent. The second fund swings more in absolute points and less per point of return, which is the comparison a risk-adjusted reading is after.

Daily life offers the same contrast. Someone walking 8,000 steps a day with a standard deviation of 1,200 has a CV of 15 percent. The same person at 180 lb with a standard deviation of 9 lb has a CV of 5 percent. Step counts are four times as variable relative to their level as body weight, which matches the intuition that a routine is easier to hold steady than a habit.

A rough reading guide, not a rule from any standards body: under 10 percent is tight for physical measurements, 10 to 25 percent is ordinary for human behavior, and anything above 100 percent means the spread is as large as the mean itself. Fund returns live above 100 percent routinely because returns hover near zero and can be negative.

Where the CV misleads

The first failure is the offset scale. Four daily highs of 50, 59, 68 and 77 F average 63.5 with a standard deviation of 11.619, so the CV is 18.3 percent. Convert the same four readings to Celsius and they become 10, 15, 20 and 25, with a mean of 17.5 and a standard deviation of 6.455, which gives a CV of 36.9 percent. Same weather, double the CV, because 0 F is not the absence of warmth.

The second failure is a mean near zero. A set of readings with a mean of 0.5 and a standard deviation of 0.5 has a CV of 100 percent, and a tiny shift in the mean moves that number by a lot. When the mean crosses zero the CV flips sign, since 0.4 divided by -2 is -0.2. NIST notes both problems on its reference page and adds a third: a CV cannot be used to build a confidence interval for the mean.

The third failure is small samples. With two values the sample standard deviation is half the gap between them, so a CV from two readings is mostly noise. Five readings give a usable first look, and twenty or more give a figure worth putting in a report. The fourth failure is bounded data such as a score from 0 to 100, where the ceiling compresses the spread and makes a high-scoring group look more consistent than it is.

Reading a CV in a lab report

Laboratories usually call the same number the relative standard deviation, or RSD. The NIST note on relative standard deviation spells out the difference: the RSD divides by the absolute value of the mean, so it is always positive, while the CV keeps the sign of the mean. For positive data the two are identical, which covers most measurements a lab makes.

A typical report reads something like: six replicates, mean 10.2 mg/L, standard deviation 0.27 mg/L, RSD 2.7 percent. The RSD is the number compared against the method's acceptance limit, and the standard deviation is the number used to judge a single result against a specification. Keeping both in view avoids the trap of thinking a method is precise because its RSD is small while the actual values are far from the target.

Accuracy and precision are separate questions, and the CV only speaks to the second. A method that reads 10 percent high every time has a CV near zero and a bias of 10 percent. That is why a calibration check comes before a precision number in any competent report, and why the OpenStax chapter on measures of spread treats standard deviation as a description of variability rather than a statement of correctness.

Mistakes that change the answer

The last row is the one that catches analysts. Averaging a CV of 1 percent from 100 readings with a CV of 20 percent from 3 readings gives 10.5 percent, which describes neither set. The correct pooled approach combines the squared deviations and the counts, so the larger sample dominates. For two sets of 100 and 3 readings the pooled CV lands near the larger set, not halfway between the two.

One more habit is worth building. Write down which formula was used and how many values went into it, next to the CV itself. A CV without a sample size and a denominator is not reproducible, and two analysts comparing notes will otherwise spend an afternoon on a 0.3 percentage point gap that comes down to STDEV.S versus STDEV.P.