Cointegration

Cointegrated series are driven by a shared underlying random walk component. While each individual series is non-stationary (I(1)), a linear combination of them is stationary (I(0)). This implies that the series are bound together by a long-run equilibrium force. For example, a pair of stocks in the same sector may each follow a random walk, but their price ratio reverts to a mean. The shared trend represents the fundamental economic driver, while the stationary residual captures temporary pricing errors.

The defining mathematical property is the existence of a cointegrating vector that produces a stationary residual. If y_t and x_t are both I(1), but u_t = y_t - βx_t is I(0), the series are cointegrated. This residual u_t represents the equilibrium error or spread. Stationarity of this spread is tested using the Engle-Granger or Johansen procedures. A stationary spread is mean-reverting, providing the statistical foundation for pairs trading strategies.

Cointegration implies a dynamic adjustment process captured by the Vector Error Correction Model (VECM). The VECM decomposes changes into short-term dynamics and a long-run equilibrium correction term. The error correction term (the lagged residual from the cointegrating regression) measures the deviation from equilibrium. If the spread widens, the error correction term forces one or both variables to adjust back toward the long-run relationship. This is formalized in the Granger Representation Theorem.

Cointegration is fundamentally different from correlation. Correlation measures the short-term, contemporaneous co-movement of returns, which can be unstable and spurious. Cointegration captures a long-run structural relationship in price levels, even if short-term returns diverge. Two assets can be highly correlated but not cointegrated (if their price spread drifts apart permanently), or cointegrated but have low correlation (if they adjust to equilibrium with lags). Cointegration implies a deeper economic linkage.

The Johansen procedure is the standard multivariate test for cointegration. Unlike the two-step Engle-Granger method, it can identify multiple cointegrating vectors in a system of n variables. The test uses maximum likelihood estimation on a VAR model and computes two statistics: the Trace statistic and the Maximum Eigenvalue statistic. These test the null hypothesis of r cointegrating vectors against alternatives of n or r+1 vectors, respectively, providing a rigorous rank identification.

The canonical application of cointegration in quantitative finance is statistical arbitrage via pairs trading. The strategy involves:

• Identifying two cointegrated assets (e.g., KO and PEP).
• Calculating the spread as the residual from the cointegrating regression.
• Executing a long-short trade when the spread deviates beyond a threshold (e.g., 2 standard deviations).
• Exiting when the spread reverts to its mean. The cointegration relationship provides the statistical guarantee that the spread is mean-reverting, forming the strategy's theoretical edge.

A prerequisite concept. A time series is stationary if its statistical properties—mean, variance, autocorrelation—are constant over time.

• Why it matters: Cointegration arises specifically from non-stationary I(1) variables that share a common stochastic trend.
• Testing: The Augmented Dickey-Fuller (ADF) test is the standard gatekeeper. You must confirm I(1) status before testing for cointegration.
• Risk: Ignoring non-stationarity leads to spurious regression with artificially high R-squared values.

A trap for the unwary. Occurs when two independent random walks appear to be highly correlated in a standard OLS regression, despite having no economic relationship.

• Diagnostic: High R-squared combined with very low Durbin-Watson statistic (indicating serially correlated residuals).
• Solution: If residuals from a levels regression are stationary (tested via Engle-Granger), the relationship is genuine cointegration, not spurious.
• Granger & Newbold (1974) first formalized this pitfall.

The multivariate standard for detecting cointegration. Unlike the two-step Engle-Granger method, it handles multiple cointegrating vectors simultaneously.

• Methodology: Uses a Vector Autoregression (VAR) framework and maximum likelihood estimation to test the rank of the impact matrix.
• Output: Two test statistics—Trace and Maximum Eigenvalue—to determine the exact number of cointegrating relationships (r).
• Advantage: Avoids the arbitrary normalization and small-sample bias of single-equation methods.

The canonical market-neutral application of cointegration. Identifies two historically linked assets and trades the spread when it deviates from the equilibrium.

• Signal Generation: Calculate the Z-score of the spread residual. Enter long/short when the spread exceeds ±2 standard deviations.
• Critical Metric: The half-life of mean reversion—how long it takes for the spread to revert halfway back to the mean.
• Risk: Cointegration breakdown during regime changes (e.g., M&A announcements) requires dynamic recalibration or stop-losses.

A distinct but complementary concept. Tests whether past values of X help predict future values of Y beyond the information in past Y alone.

• Distinction: Cointegration implies a long-run equilibrium; Granger causality implies short-run predictive power. They can exist independently.
• VECM Integration: In a VECM, Granger causality can flow through both the lagged differences (short-run) and the error correction term (long-run).
• Warning: Does not imply true structural causality—merely temporal precedence.