Futures & Market Research Glossary

Cointegration is a statistical property of two or more nonstationary time series that share a stable long-run relationship. In the common case where each series is integrated of order one, I(1), the individual price levels may wander over time, but the series are cointegrated if a non-zero linear combination of them is stationary, I(0). If a vector of series is denoted by X_t, cointegration therefore requires a vector of coefficients β such that β′X_t is stationary even though the components of X_t are not stationary in levels.

This distinction separates cointegration from correlation. Correlation measures co-movement over a specified sample or window; cointegration concerns whether nonstationary series share stochastic trends in such a way that a particular linear combination does not drift without bound. Two markets can therefore be highly correlated without being cointegrated, while cointegrated series need not maintain a consistently high short-horizon correlation. Cointegration is also more specific than simply observing that two prices sometimes converge after diverging.

The stationary combination is the key object

For two price series, the relevant relationship is generally not their raw difference. A relation such as y_t − βx_t may be stationary for an estimated coefficient β even when y_t − x_t is not. The cointegrating vector therefore determines the relative weighting that removes the shared nonstationary trend. In systems containing more than two I(1) variables, more than one independent stationary combination can exist; the number of such relationships is called the cointegration rank. Johansen's multivariate framework explicitly estimates and tests this rank rather than reducing the problem to a collection of pairwise relationships.

Cointegration also has a dynamic implication. Under the conditions of the Engle–Granger representation theorem, cointegrated variables admit an error-correction representation: short-run changes can respond to the previous period's deviation from the long-run relation. This does not mean that every deviation must close quickly or that one particular variable must do the adjusting. The speed and location of adjustment are empirical properties of the system.

A relationship can be statistical without being permanent

A cointegration result is conditional on model specification and sample evidence. Deterministic terms, lag choices, sampling frequency and the assumed integration properties of the inputs can affect formal tests. More importantly, a historical stationary combination can cease to describe the data after a structural break in the relationship. Research on regime-shift cointegration shows that conventional tests that assume constant coefficients can lose substantial power when the long-run intercept or slope changes. For market research, this makes stability testing economically important: evidence that a spread was cointegrated over one period is not proof that the same hedge ratio represents a durable relationship in another regime.

Cointegration is therefore most useful as evidence of a constrained long-run statistical relationship, not as a directional trading signal or a guarantee of mean reversion. Its practical value comes from identifying combinations of nonstationary variables that behave stationarily and then testing whether that relationship remains economically interpretable and statistically stable.