Glossary
Cointegration
A property of prices that each wander but whose combination, a spread, is stationary and returns towards a mean.
Engle and Granger (1987) defined cointegration: each series needs differencing to become stationary, yet a linear combination is already stationary. Their two-step test estimates the combination by regression, then tests the residual for a unit root with adjusted critical values.
Two trending series can look related in a regression without being cointegrated, so a unit-root test on the spread is required.
Example: an AR(1) spread with coefficient 0.95 and daily noise of 0.5% has a standard deviation of 1.60% and a half-life of 13.5 days.