Coefficient of Variation

Three situations. First, when the mean is close to zero — dividing by a near-zero μ makes CV blow up or become unstable, even if σ is small and the data is well-behaved. This happens for temperature in Celsius or Fahrenheit, financial returns, and any data centered near zero. Second, when the data is on an interval scale rather than a ratio scale — Celsius and Fahrenheit have arbitrary zero points, so the same physical temperature gives different CV values depending on the unit. Third, when the data can take negative values — CV assumes the mean represents a "typical" positive magnitude, which is not true for signed data. In all three cases, use σ directly, use the coefficient of quartile variation (based on the median and IQR), or transform the data first.

Nothing — they are the same quantity with two names. CV stands for "coefficient of variation" and is the more common term in statistics and finance. RSD stands for "relative standard deviation" and is the standard term in analytical chemistry, pharmaceutical QC, and clinical laboratory medicine. Both are calculated identically: σ/μ × 100%. The RSD terminology emphasizes the "relative" nature of the quantity — it is a standard deviation expressed relative to the mean — while the CV terminology emphasizes that it is a coefficient (a dimensionless multiplier). Journals and standards organizations often have preferences: the USP, EP, and ICH all use RSD in pharmacopeial monographs, while statistics journals prefer CV.

It depends entirely on the field. In analytical chemistry, a CV below 1% is excellent precision, 1–5% is acceptable for most methods, and above 15% indicates a problem. In biology and medicine, CVs of 10–30% are common for biological measurements (weight, blood pressure, cholesterol) because natural variation between individuals is high. In finance, CVs of 100% or more are normal for stock returns — the volatility exceeds the average return. In manufacturing, tolerances are often set at 1–5% of nominal. There is no universal "good CV" — you must compare against what is expected in your domain and against comparable datasets. The CV scale bar in this calculator uses a general-purpose scheme (≤10% very low, 10–20% low, 20–33% moderate, 33–50% high, >50% very high), but treat these as rules of thumb, not standards.

Variance σ² has squared units (kg², $², etc.) while the mean μ has linear units (kg, $). Dividing variance by the mean gives a quantity with mismatched units (kg²/$ or similar), which is meaningless as a ratio. Standard deviation σ has the same units as the mean, so σ/μ is dimensionless — that is what makes CV work. The square root in the definition of σ is precisely what restores the units and makes the ratio meaningful. Using variance instead of standard deviation in the "CV" formula would give a quantity with units of (mean unit), not a pure number, and would not be comparable across scales. So the choice of σ over σ² is forced by the requirement that CV be dimensionless.

Small samples give unreliable CV estimates because both σ and μ are estimated with high uncertainty. At small N, the estimate of σ is biased downward (using N−1 partially corrects this, but the estimate is still noisy), and rare extreme values may or may not appear in the sample. For CV estimates that are stable enough for comparison, you generally want at least 10–20 samples, and often more. The standard error of CV itself is approximately CV/√(2N) for normally distributed data, so a CV of 10% estimated from N = 10 has an SE of about 2.2% — meaning the "true" CV could be anywhere from 6% to 14%. As N grows, this uncertainty shrinks as 1/√N. If you are comparing CV values across groups, make sure the sample sizes are comparable — a CV from N = 5 and one from N = 500 are not directly comparable.

Yes, and it frequently is. CV > 100% simply means the standard deviation exceeds the mean. This is common for non-negative quantities with wide ranges: incomes (many low, few very high), city populations, biological measurements with high natural variation (serum triglycerides, rainfall), and financial returns (which can be negative). There is no mathematical upper bound on CV. When CV > 100%, the mean is no longer a very representative center of the distribution — the data is likely right-skewed or heavy-tailed, and you may want to consider reporting the median and interquartile range instead, or taking logarithms before computing variability metrics.