Definition
Two nonstationary series are cointegrated when some nonzero linear combination of them is stationary. Cointegration describes a long-run equilibrium relation even though the individual levels may wander. It is a statistical property that requires a specified sample, variables, deterministic terms and test procedure.
How it is calculated
The residual y_t minus beta times x_t is stationary, I(0), while the individual series are integrated of order one, I(1).
| Symbol | Meaning | Unit |
|---|---|---|
| y | first level series | level units |
| x | second level series | level units |
| beta | cointegrating coefficient | y units per x unit |
| a | intercept | y units |
| e | residual | y units |
-
Estimate the long-run relation between the two level series.
-
Form the residual by subtracting the fitted intercept and slope contribution from the first series.
-
Test the residual for stationarity with the appropriate nonstandard critical values.
Interpretation
Two drifting paths can still be tied together: they may separate temporarily but a particular spread does not keep accumulating drift. The tie is about the combination, not about either level series being stable on its own. A test result can change with sample window and deterministic specification.
Assumptions and conventions
-
The order of integration and deterministic terms are assessed rather than assumed blindly.
-
Stationarity is a process property; the test provides sample-dependent evidence.
The residual has y's units; beta has y units per x unit.
Limitations and common misreadings
-
Correlation of levels is not sufficient evidence of cointegration.
-
A stationary residual does not make the original levels stationary.
References
- Robert F. Engle, C. W. J. Granger. (1987). Co-Integration and Error Correction: Representation, Estimation, and Testing. Econometrica. Theorem 1 and Section 2. doi:10.2307/1913236. Source

