Covariance Calculator

The magnitude is a fact about your units

“Is a covariance of 4,800 strong?” has no answer, and it is worth understanding why the question is malformed rather than just hard.

A covariance is measured in x-units multiplied by y-units. Height in metres against weight in kilograms gives a covariance in metre-kilograms. Record the same heights in centimetres and the covariance multiplies by 100. Nothing about the relationship changed. The number did.

Use the unit multiplier in the tool to see it directly. The four settings differ by a factor of a million overall, and the covariance tracks them exactly — while the correlation stays fixed to every decimal place shown. One of those two quantities is describing the data and the other is describing your choice of ruler.

So the only thing readable from a covariance in isolation is its sign. Positive, negative, or approximately zero. That is genuinely useful — it is the direction of the relationship — and it is all there is.

Zero covariance does not mean independence. It means no linear association. Take x values symmetric about zero and set y = x²: y is completely determined by x, and the covariance is exactly zero, because the positive and negative contributions cancel. Perfect dependence, zero covariance. The reverse does hold: independent variables always have zero covariance.

Pearson’s correlation is the covariance divided by both standard deviations, which cancels the units and confines the result to −1 to +1. It is the covariance made comparable, and it is almost always the number you want to report.

Where a raw covariance is the right quantity

Given all of the above, it is fair to ask why the covariance is computed at all rather than going straight to a correlation. There are three real answers.

Portfolio variance. The variance of a weighted sum is the sum of every pairwise covariance, each weighted by both assets’ shares. This is Markowitz’s result and the foundation of diversification: adding an asset with low or negative covariance to the rest reduces total risk even when that asset is individually volatile. Here the covariances are in squared-return units, they are summed rather than compared, and the units are exactly right.

The regression slope. The least-squares slope is the covariance of x and y divided by the variance of x. That division leaves y-units per x-unit — which is what a slope is, and what makes it interpretable where the covariance is not. The regression calculator goes further into what the slope supports.

The covariance matrix. Principal component analysis, the multivariate normal distribution, Kalman filters and Mahalanobis distance all operate on a matrix of covariances rather than correlations, because the matrix has to carry the scale information the analysis needs. Standardising to correlations first is a different analysis, not a tidier version of the same one — PCA on a covariance matrix and PCA on a correlation matrix give different components.

One identity that makes the matrix easier to read: the covariance of a variable with itself is its variance. That is why a covariance matrix has the variances down its diagonal, and why “variance” and “covariance” are the same operation with different arguments.

Sample or population, and why it matters less than you think

The choice between dividing by n and n − 1 is the same choice the variance calculator faces, for the same reason.

Dividing by n − 1 is Bessel’s correction, and it exists because the deviations are measured from the sample mean rather than the true mean. The sample mean is, by construction, the point that minimises those squared deviations — so they come out systematically too small, and dividing by a smaller number compensates. Without it the covariance is biased toward zero.

Use n − 1 when your data is a sample from something larger you want to describe. Use n when your data is the whole population. In practice that is nearly always the first case.

The gap closes quickly. At n = 5 the two differ by 25%; at n = 30 by 3.4%; at n = 100 by 1%. It matters for small samples and is negligible for large ones.

And for the correlation it does not matter at all. Pearson’s r divides a covariance by two standard deviations, and if all three use the same divisor it cancels exactly. That is one of several reasons the correlation is the more robust thing to report — it cannot be wrong because someone picked the other convention.

Reading it without overreaching

Three cautions, beyond the units.

It is a linear measure, and it inherits every limitation of one. A U-shaped relationship, a threshold effect, or a relationship that reverses direction will all show a covariance near zero. Plot the data; the tool draws the scatter above the numbers for exactly this reason.

It is as outlier-sensitive as the mean it is built on. A single extreme pair can set the sign of the whole covariance, because each point contributes the product of two deviations and products grow quickly. If one point is far out on both variables, the covariance is largely reporting that point. The outlier calculator covers how to find out.

Direction is not causation. A positive covariance is consistent with x causing y, y causing x, a common cause, or coincidence. The correlation page sets out the specific mechanisms that produce association without a causal link, and all of them apply here unchanged.

The practical summary: compute the covariance when you need it as an ingredient, report the correlation when you need to communicate, and never compare two covariances’ magnitudes to each other.

Why the variance of a sum is not the sum of the variances

This is the identity that gives the covariance its one genuinely standalone job, and it is worth stating on its own because the intuition it corrects is very strong.

Var(X + Y) = Var(X) + Var(Y) + 2⋅Cov(X, Y). The third term is the whole story. Variances only add when the covariance is zero, and nothing guarantees that.

With a positive covariance the sum is more variable than its parts suggest. Two investments that rise and fall together produce a portfolio riskier than adding their individual risks. Two project tasks that are both delayed by the same missing supplier produce a schedule with a longer tail than either estimate implies.

With a negative covariance it is less variable, and that is diversification stated exactly. Adding an asset that tends to rise when the rest fall lowers total variance even when that asset is individually volatile — the reduction comes from the cross term, not from the asset’s own behaviour.

Subtraction does not help. Var(X − Y) = Var(X) + Var(Y) − 2⋅Cov(X, Y), so the variances still add while the covariance flips sign. A difference between two positively correlated measurements is less variable than either; a difference between two independent ones is more variable than both. This is why paired designs are more sensitive, and it is the same fact seen from the other side.

Independence is sufficient but not necessary. Variances add whenever the covariance is zero, and zero covariance is weaker than independence — which makes this one of the few places where the weaker condition is genuinely all you need.

Sources and methodology

References for the definition and its portfolio use.

• NIST/SEMATECH e-Handbook — scatter plots, covariance and correlation
• Markowitz, “Portfolio Selection”, The Journal of Finance 7(1), 1952 — where summed covariances become portfolio variance
• R documentation for cov — which uses the n − 1 divisor by default
• Correlation Coefficient Calculator — the covariance made unit-free
• Variance Calculator — the covariance of a variable with itself
• How we calculate — methodology

Method. The covariance is computed from centred sums rather than the algebraically identical (nΣxy − ΣxΣy)/(n(n − 1)) form, which subtracts two large nearly-equal numbers and loses most of its significant digits on data far from the origin. The unit multiplier rescales the original values and recomputes, rather than scaling the covariance — the point is that the recomputation genuinely lands somewhere else. The suite asserts that scaling x by k multiplies the covariance by exactly k while leaving the correlation unchanged to twelve decimals, that the covariance of a variable with itself equals its variance, that y = x² on a symmetric x gives exactly zero covariance despite perfect dependence, and that the sample and population divisors differ by exactly n/(n − 1). That engine is verified on every change against 79 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.