Does cointegration require the variables to be equal?

No. The stable relationship may involve a coefficient, intercept and other specified components.

Does a different upward response automatically prove asymmetric cointegration?

No. Nonstationarity, the long-run relationship and the relevant asymmetry hypothesis must be assessed within an appropriate model.

Can a linear equilibrium relationship have asymmetric adjustment?

Yes. Different restoring speeds can operate around the same linear relationship.

AI Glossary / Statistics and probability / Advanced

Asymmetric cointegration

Definition

Asymmetric cointegration describes long-run relationships between time series in which the adjustment process or the relationship itself permits different responses to positive and negative movements. In a common threshold formulation, departures above and below an equilibrium are corrected at different speeds. Other formulations allow increases and decreases in an explanatory variable to have different long-run effects.

What it means

Cointegration concerns a stable relationship among variables that may individually be nonstationary. Ordinary correlation between trending series is not enough to establish it. Asymmetry adds a question about direction or regime: does the system behave differently on opposite sides of an equilibrium or following increases versus decreases? These are distinct hypotheses. A threshold error-correction model can preserve a linear long-run relationship while making the return toward it nonlinear; a nonlinear ARDL model can represent a different form of asymmetry.

How it works

Specify the proposed relationship and examine the time-series properties of the variables. In a threshold approach, define an equilibrium deviation and a rule assigning observations to regimes. Model changes in the deviation using different adjustment coefficients, potentially with lagged changes and other justified terms. Test for a restoring long-run relationship using inference appropriate to the estimated residuals, threshold specification and sample. Then assess whether adjustment differs across regimes. Momentum-threshold rules use changes in disequilibrium rather than simply its level. In nonlinear ARDL approaches, positive and negative changes can instead be accumulated separately. The models, restrictions and tests should not be treated as interchangeable.

Three practical examples

Price-transmission research: investigate whether a gap between related prices closes differently when it is positive and when it is negative.
Interest-rate analysis: examine whether deviations between rates at different maturities exhibit direction-dependent adjustment under a specified long-run model.
Operational scenario modeling: compare the implications of upward and downward input-cost movements while testing whether an asymmetric specification improves forecasts beyond a symmetric baseline.

Think of it this way

Think of a thermostat that corrects an overly warm room quickly but an overly cold room slowly. Both responses can restore the same target, yet their speeds differ. The analogy illustrates asymmetric adjustment; it does not establish that any observed pair of economic series actually shares a stable equilibrium.

Worked example

Let an illustrative equilibrium deviation be e = y - 2x. Ignore new shocks and other dynamics for this calculation. Suppose the next change in e is -0.20e when e is nonnegative and -0.05e when e is negative. Starting from e = 10, the next change is -2 and the new deviation is 8. Starting from e = -10, the next change is +0.5 and the new deviation is -9.5. After three periods without new shocks, the positive deviation is 10 × 0.8 cubed = 5.12, while the negative deviation is -10 × 0.95 cubed = -8.57375. Both move toward zero, but at different speeds. The negative coefficient does not mean every observed change is negative: multiplying it by a negative deviation produces a positive correction. These stipulated dynamics illustrate adjustment only. They do not constitute a statistical test for cointegration, establish that x and y are individually integrated, or identify which observed variable supplies the correction in a full system.

Illustrative case study

Fictional case: Alder Procurement studies whether a supplier-price series and a material-cost index maintain a stable relationship. Analysts compare symmetric and threshold adjustment models, check the influence of contract changes and reserve later observations for forecast evaluation. A faster estimated correction on one side of the equilibrium is treated as a descriptive finding requiring uncertainty assessment. The team does not interpret it as proof of unfair pricing or as a guaranteed trading opportunity. This fictional case describes a modeling workflow rather than an empirical result.

Common misconceptions and mistakes

Asymmetric adjustment is not the same as asymmetric long-run coefficients. Different responses to positive and negative shocks can also arise from changing volatility, structural breaks, omitted variables or measurement choices. A significant ordinary regression between trending series can be spurious. Standard reference distributions for a familiar regression statistic may not apply to residual-based cointegration tests, especially when thresholds are estimated. Rejecting no cointegration and rejecting symmetric adjustment are separate inferential questions. Bounds-testing methods have assumptions and integration-order restrictions; their flexibility does not mean they cover arbitrary orders of integration.

Business use

Businesses can use these models to explore directional persistence in costs, prices and related economic indicators. Define the decision before choosing the model, and compare economic effect sizes with estimation uncertainty. Record lag selection, threshold selection and any transformations, because apparently small choices can alter regime membership or estimated adjustment. Test sensitivity to unusual periods and evaluate forecast performance out of sample. Evidence of a persistent statistical relationship does not establish a causal mechanism or ensure that the relationship survives a policy, technology or market-structure change.

Check your understanding

Q: In the worked example, which deviation is closer to zero after three periods?

Answer: A: The initially positive deviation, whose magnitude falls to 5.12 rather than 8.57375. This follows from the assumed adjustment factors and is not an estimated result from observed data.

Related terms

Cointegration
Nonstationarity
error-correction model
threshold autoregression
momentum threshold
nonlinear ARDL
structural break and spurious regression
regime-dependent dynamics and alternative explanations for apparent co-movement

Learn next

simulate the two-regime deviation process and add random shocks to see why individual observations need not move toward equilibrium every period. Then study the difference between testing a long-run relationship and testing symmetry, using the critical values or resampling procedure appropriate to the selected method.

Sources

Sources: Enders and Siklos, “Cointegration and Threshold Adjustment,” author-posted research record: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=243554 ; Walter Enders, threshold and momentum-threshold testing resources: https://sites.ua.edu/wenders/enders-granger-test-files/ ; Shin, Yu and Greenwood-Nimmo, “Modelling Asymmetric Cointegration and Dynamic Multipliers in a Nonlinear ARDL Framework,” institutional research record: https://pure.york.ac.uk/portal/en/publications/modelling-asymmetric-cointegration-and-dynamic-multipliers-in-a-n/ .

Previous: Assumed-density filter Next: Asymmetric multiprocessing