copulas
/ COP-yuh-luh /
Dependence and individual behaviour are two genuinely different things. One variable might be normal, another heavily skewed, yet the interesting question is how they move together. A copula is a device that separates these two concerns cleanly: it captures the pure dependence structure between variables while leaving each variable's own distribution to be specified separately.
The trick rests on the probability integral transform: if you feed each variable through its own cumulative distribution function, the result is uniformly distributed on (0, 1), regardless of the original shape. A copula is simply a joint distribution on the unit square (or cube) whose individual margins are all uniform — it describes how those transformed-to-uniform versions of the variables are coupled. Sklar's theorem is the headline result: every joint distribution can be written as a copula applied to its marginals, and that copula is unique when the variables are continuous. So you can mix and match — pick any marginals you like, pick any copula for the dependence, and glue them together.
This separation is invaluable when the margins and the dependence have different natures: model each variable's distribution however suits it, then choose a copula to describe co-movement, including tail dependence (the tendency to crash together) that a single correlation number cannot express. Copulas became notorious in finance because a too-simple Gaussian copula badly understated the chance of many assets failing at once during the 2008 crisis — an honest reminder that the copula you choose is a strong and falsifiable modelling assumption, not a neutral technicality.
To model insurance losses on a house and its garage, you might keep each loss with its own heavy-tailed marginal (most years zero, rare years huge), then bind them with a copula that has strong tail dependence so they tend to be huge together (the same fire hits both). A Gaussian copula, by contrast, would let big losses occur almost independently — and underprice the joint catastrophe.
A copula glues chosen marginals to a chosen dependence; the tail-dependence choice is consequential.
A copula is a modelling choice, not a fact about the data — picking one with the wrong tail behaviour (the Gaussian-copula error of 2008) can drastically misstate the risk of joint extremes.