covariance and correlation

Covariance and correlation measure whether two random quantities tend to move together. When one is above its average, is the other usually above its average too, or below? If big losses on one policy line up with big losses on another, they share risk and amplify each other; if one rises while the other tends to fall, they offset and steady the total. These two numbers put a value on that co-movement.

Covariance, Cov(X, Y) = E[(X - E[X])(Y - E[Y])], is positive when the two rise and fall together, negative when they move oppositely, and zero when there is no linear tendency either way. Its size, though, is hard to interpret because it carries the units of both variables multiplied together. Correlation fixes that by dividing covariance by the two standard deviations, producing a clean, unit-free number always between -1 and +1: +1 is perfect lockstep, -1 is perfect mirror-image, and 0 is no linear relationship. A correlation of 0.8 signals strong positive co-movement; -0.3 a mild offsetting tendency.

For actuaries this is the heartbeat of diversification. Variance of a sum is Var(X) + Var(Y) + 2·Cov(X, Y), so positive covariance inflates portfolio risk while negative covariance dampens it — which is why insurers spread across uncorrelated lines and regions and why correlated catastrophe risk is so feared. Two warnings are essential. First, correlation is not causation: two series can move together because both follow a third driver, with no direct link. Second, zero correlation does not mean independence — correlation only sees straight-line relationships and can read exactly zero even when two variables are strongly but non-linearly dependent, a blind spot that has burned more than one risk model.