Copula-Based Pairs Trading

Classic pairs trading assumes the spread between two cointegrated stocks reverts in a predictable, roughly linear way. But two stocks can be tightly linked in normal times and behave completely differently during a crash — correlation alone doesn't capture that. A copula separates each stock's own return distribution from the dependence structure between them, so the strategy can model exactly how likely it is for one stock to be unusually cheap relative to the other, without assuming the relationship is symmetric or constant across the whole distribution.

How the signal works

Fit each stock's marginal return distribution, transform both series to uniform "percentile" scores, then fit a copula (Clayton, Gumbel, or Student-t are common choices) that captures how those percentile scores move together — including whether they cling together more tightly in the tails than in the middle. The trading signal comes from the conditional probability: given today's percentile for stock A, how extreme is stock B's percentile relative to what the copula says is typical? A large mismatch signals divergence worth trading.

Worked example

Stock A sits at its 80th percentile of daily returns while stock B sits at only its 20th percentile — a 60-point gap. If the fitted copula says this joint outcome has less than 2% probability given their usual co-movement, the strategy shorts A and buys B, betting the atypical divergence closes. A plain-correlation approach might not flag this if the historical correlation coefficient looks unremarkable, since correlation only summarizes average co-movement, not this kind of tail-level divergence.