From Cointegration to Causality: Applying Granger Causality Tests in Trading Strategies

The Distinction Between Cointegration and Causality

Cointegration and causality are related but distinct concepts that are often confused. Cointegration implies that two or more time series have a long-run equilibrium relationship. If they drift apart, they will eventually converge. However, cointegration itself does not tell us anything about the direction of this relationship. Does stock A move to correct a deviation from stock B, or does stock B move to correct a deviation from stock A? Or do both adjust simultaneously? This is a question of causality.

Granger causality, developed by Nobel laureate Clive Granger, provides a statistical framework for investigating this question. A time series X is said to Granger-cause another time series Y if past values of X contain information that helps predict future values of Y, beyond the information already contained in past values of Y itself. It is a test of predictive causality, not necessarily true philosophical causality. In a cointegrated system, the error correction term (ECT) represents the disequilibrium. The concept of Granger causality is intimately linked to how the variables in the system react to this disequilibrium, as captured by the adjustment coefficients in a Vector Error Correction Model (VECM).

The Granger Causality Test in a VECM Framework

When variables are cointegrated, the standard Granger causality test based on a VAR in levels is not valid. Instead, causality testing must be conducted within the VECM framework, which correctly incorporates the long-run relationship. The VECM for a bivariate system (Y, X) with one cointegrating vector is:

$\Delta Y_t = \alpha_Y (Y_{t-1} - \theta X_{t-1}) + \sum \gamma_{Y,i} \Delta Y_{t-i} + \sum \delta_{Y,i} \Delta X_{t-i} + \epsilon_{Y,t}$

$\Delta X_t = \alpha_X (Y_{t-1} - \theta X_{t-1}) + \sum \gamma_{X,i} \Delta Y_{t-i} + \sum \delta_{X,i} \Delta X_{t-i} + \epsilon_{X,t}$

There are two channels through which causality can occur in this system:

1. Short-Run Causality: This is tested by examining the joint significance of the lagged difference terms. For example, X Granger-causes Y in the short run if the coefficients $\delta_{Y,i}$ are jointly significantly different from zero. This is a standard Wald test on the coefficients of the lagged explanatory variables.

2. Long-Run Causality (via Error Correction): This is the more interesting channel in a cointegrated system. Causality is transmitted through the error correction term. If the adjustment coefficient $\alpha_Y$ is statistically significant, it means that Y responds to the disequilibrium from the previous period. This implies that there is a causal link running from the ECT (which is a function of both Y and X) to Y.

Therefore, in a VECM, X Granger-causes Y if either the set of $\delta_{Y,i}$ coefficients is jointly significant (short-run causality) or if the adjustment coefficient $\alpha_Y$ is significant (long-run causality). If both are significant, there is strong Granger causality.

A Practical Example: Spot and Futures Prices

Consider the relationship between the spot price of a commodity (e.g., crude oil, WTI) and its front-month futures price (CL). Economic theory suggests they are cointegrated, as the futures price must converge to the spot price at expiration. But what is the direction of causality? Does the futures market lead the spot market, or vice versa?

1. Cointegration Test: We first run a Johansen test on the log prices of WTI and CL and confirm they are cointegrated with one cointegrating vector, $\beta' = [1, -\theta]$. We find $\hat{\theta} \approx 1$, so the long-run relationship is simply $log(WTI_t) - log(CL_t) = Z_t$, where $Z_t$ is the basis.

2. VECM Estimation: We estimate the VECM:

$\Delta log(WTI_t) = \alpha_{WTI} (log(WTI_{t-1}) - log(CL_{t-1})) + ...$

$\Delta log(CL_t) = \alpha_{CL} (log(WTI_{t-1}) - log(CL_{t-1})) + ...$

Let's say the estimation yields the following adjustment coefficients (with t-statistics in parentheses):

• $\hat{\alpha}_{WTI} = -0.05$ (-1.5)
• $\hat{\alpha}_{CL} = 0.15$ (4.5)

3. Interpretation and Causality Analysis:

• The coefficient $\hat{\alpha}_{WTI}$ is not statistically significant (t-stat of -1.5 is less than the important value). This means that the spot price (WTI) does not significantly react to deviations from the long-run equilibrium. It is the "leader" in the long-run relationship.
• The coefficient $\hat{\alpha}{CL}$ is positive and highly significant (t-stat of 4.5). This means that the futures price (CL) strongly adjusts to correct any disequilibrium. If the basis was positive in the last period ($log(WTI{t-1}) > log(CL_{t-1})$), the futures price will tend to rise in the current period to close the gap. This is long-run Granger causality running from WTI to CL.

We would also perform Wald tests on the lagged difference terms. We might find that lagged changes in futures prices help predict current changes in spot prices (short-run causality from CL to WTI), reflecting the rapid flow of information in the futures market. However, the VECM analysis reveals the long-run anchor: the spot price leads, and the futures price adjusts.

Trading Strategy Implications

Understanding the causal structure is not just an academic exercise; it has direct trading implications.

1. Identifying the Source of Alpha: In our WTI/CL example, the finding that the spot price is the long-run leader suggests that signals originating from the physical oil market might be more valuable for long-term forecasting than signals from the futures market alone. A strategy might be designed to trade the futures contract based on signals derived from the spot market's deviation from the equilibrium.

2. Improving Pairs Trading Models: In a standard pairs trade, we treat the spread symmetrically. However, if we know that Stock A Granger-causes Stock B, it means Stock B is the one that does the adjusting. This has two implications:

   • Execution: When the pair diverges, we expect B to move more than A to restore the equilibrium. Our execution and profit expectations should be focused on the movement of B.
   • Risk Management: If we are in a trade and the spread continues to diverge, the problem is more likely to be with the behavior of stock B (the follower) than with stock A (the leader). This can help in diagnosing a breakdown in the relationship.

3. Impulse Response Functions (IRFs): A VECM allows us to generate IRFs, which trace out the reaction of each variable to a shock in another variable over time. By analyzing the IRFs, a trader can visualize the entire dynamic path of adjustment. For example, an IRF could show how a 1% shock to the spot oil price affects the futures price over the next 20 days, providing a quantitative estimate of the timing and magnitude of the response.

Conclusion

Granger causality analysis, when correctly applied within a VECM framework, improves cointegration from a simple statement about correlation to a directional map of predictability. It allows traders to dissect the dynamic interplay between cointegrated assets, identifying which assets lead and which follow in the dance back to equilibrium. This deeper understanding of the system's causal structure enables the development of more sophisticated and targeted trading strategies, moving beyond symmetric pairs trading to models that explicitly exploit the predictive power flowing from one asset to another.