Does high correlation imply cointegration?

No. Similar trends can create high sample correlation without a stationary linear combination.

Must the difference y minus x be stationary?

No. The stationary combination can involve other coefficients and deterministic terms.

Does cointegration prove that one series causes the other?

No. It describes joint time-series structure; causal claims require additional reasoning and assumptions.

AI Glossary / Statistics and probability / Advanced

Cointegration

Definition

Cointegration is a relationship in which nonstationary time series share a linear combination that is stationary, under the relevant integration framework. In the common I(1) case, each series requires one difference to become stationary, but a particular combination of their levels is already stationary. It represents a stable statistical relation among drifting series, not merely a high correlation or a guarantee of causal connection.

What it means

Two series can wander over time while their deviation from a shared relationship remains comparatively stable. For example, y may track twice x plus a constant, with a stationary residual. The levels need not return to fixed individual means, and the coefficient linking them need not be one. Cointegration describes their joint stochastic structure. Trending series can also appear strongly related without cointegration, so a visually convincing plot or high regression R squared is not enough to establish the property.

How it works

Examine the time-series properties and specify relevant intercepts, trends, seasonal features and lag structure. A common two-step approach estimates a levels relationship and then tests whether its estimated residual has a unit root using a cointegration-specific reference distribution. The usual null is no cointegration. System methods can assess multiple cointegrating relationships among several series. When cointegration is supported and other assumptions are suitable, an error-correction model combines short-run changes with adjustment to the previous period's deviation from the long-run relation. Structural breaks and changing measurement definitions can undermine a relationship that appeared stable in the historical sample.

Three practical examples

Demand planning: investigate whether two nonstationary business measures share a stable long-run relation before building a joint forecasting model.
Process analysis: examine linked cumulative indicators whose levels drift but whose economically meaningful combination may remain stable.
Forecast diagnostics: distinguish a potentially informative levels relationship from a spurious regression between unrelated trending series.

Think of it this way

Think of two boats moving with the same current while a flexible line limits their separation. Their positions relative to shore can drift, while their separation follows a more stable pattern. The analogy concerns shared movement and relative deviation; it does not prove that one boat drives the other or that the line can never break.

Worked example

Construct a hypothetical process where x_t = x_(t-1) + u_t, with independent zero-mean innovations of finite positive variance. Let z_t = 0.5 z_(t-1) + e_t be a stationary process initialized in its stationary distribution, with suitable independent innovations. Define y_t = 5 + 2x_t + z_t. Both x and y contain the random-walk component, but y_t - 2x_t - 5 = z_t is stationary by construction. The cointegrating relationship therefore uses coefficients 1 and -2, with intercept 5. If the current deviation z is 8 and future innovations have conditional mean zero, expected deviations after one, two and four periods are 4, 2 and 0.5. Actual deviations need not follow that smooth path because new shocks continue to arrive.

Illustrative case study

Fictional case: Rowan Forecasting observes that two operational series rise together over several years. Before using one to predict the other, analysts investigate whether the relationship survives appropriate time-series testing and a later evaluation period. They document a reporting-system change that alters one series' definition and avoid treating the full record as automatically homogeneous. A model using a stable historical relation is compared with simpler benchmarks and reviewed after deployment. This invented case illustrates disciplined forecasting research, not a claim that any particular pair of business measures is cointegrated.

Common misconceptions and mistakes

Correlation and cointegration answer different questions. Independent random walks can produce misleadingly strong levels regressions. A stationary difference is only one special case; a weighted combination may be required. Standard unit-root critical values are not automatically valid for residuals from an estimated cointegrating regression. Failure to reject no cointegration is not proof that no relationship exists, especially in short samples. Cointegration also does not identify a causal effect, guarantee profitable decisions or ensure stability after a structural change. Searching many combinations without accounting for selection can overstate evidence.

Business use

Businesses can use cointegration analysis when long-run relationships among nonstationary measures have a substantive justification. It can help retain information lost by modeling changes alone, while avoiding unsupported levels regressions. Record data revisions, time alignment and transformations, and compare forecast performance on genuinely later data. A statistically stationary residual can still be volatile enough to make operational use unattractive. Assess the size and timing of prediction errors, the costs of acting on them and whether the hypothesized relationship remains plausible as the business changes.

Check your understanding

Why is the constructed example cointegrated even though neither level series has a fixed mean?

Answer: Subtracting twice x and the intercept from y removes the shared random-walk component. The remaining z process is stationary by the stated construction. Cointegration concerns that combination rather than stationarity of each original level.

Related terms

Stationarity
Random walk
Structural break
Vector Autoregression
unit root
integrated process
Engle–Granger test
Johansen test
error-correction model
spurious regression

Learn next

distinguish deterministic trends from stochastic trends and practice differencing a random walk. Then study residual-based cointegration testing and why its reference distribution differs from an ordinary residual check. Examine error-correction models, lag selection and structural-break diagnostics before applying the idea to operational forecasts.

Sources

Clive W. J. Granger, Nobel Lecture: https://www.nobelprize.org/uploads/2018/06/granger-lecture.pdf statsmodels, coint documentation: https://www.statsmodels.org/stable/generated/statsmodels.tsa.stattools.coint.html Clive W. J. Granger, Nobel lecture page: https://www.nobelprize.org/prizes/economic-sciences/2003/granger/lecture/

Previous: Cohort study Next: Cold languages