Granger Causality Test: Finding Predictive Lead-Lag Signals

Key Takeaways

• The Granger causality test adds lags of series X to an autoregression of Y, then uses an F-test to determine whether those lags improve the forecast, measuring statistical predictive content, not economic causality.
• A bivariate VAR of daily VIX changes and SPY returns from 2010–2020 produced an F-statistic of 8.0, rejecting the null that VIX changes fail to predict SPY returns at the 1 percent level.
• The most dangerous mistake is running Granger tests on non-stationary series; if either series has a unit root, the standard F and chi-square distributions do not apply and spurious causality appears regularly.
• Granger significance is necessary but not sufficient for a live trading signal; the relationship must survive out-of-sample testing and transaction cost analysis before capital is committed.

What It Is

The Granger causality test asks: does X contain information, in its past values, that improves the one-step-ahead forecast of Y? Clive Granger introduced the formal framework in 1969. The test is cast as a comparison between a restricted autoregression of Y on its own lags and an unrestricted autoregression that also includes lags of X.

If adding X's lags significantly reduces the residual variance, X is said to Granger-cause Y in the statistical sense. The reverse direction is a separate test.

The Intuition

Correlation between two return series tells you they move together contemporaneously. It does not say anything about lead-lag. If small-cap returns move before large-cap returns by a day, you have actionable information. If they move on the same day, you do not. Granger causality is the formal way to answer the lead-lag question.

The word "causality" is misleading. The test can only tell you that past X helps predict Y statistically. A confounder (a third variable that drives both) will produce a positive Granger test without any true causal link. Always check the economic story before trading it.

How It Works

Consider two stationary series X_t and Y_t. Specify a bivariate vector autoregression of order p:

Y_t = a_0 + sum_{i=1..p} a_i * Y_{t-i} + sum_{j=1..p} b_j * X_{t-j} + e1_t
X_t = c_0 + sum_{i=1..p} c_i * X_{t-i} + sum_{j=1..p} d_j * Y_{t-j} + e2_t

To test whether X Granger-causes Y, impose the null H_0: b_1 = b_2 = ... = b_p = 0. Compare the unrestricted RSS_U to the restricted RSS_R from the same regression with X's lags dropped. The F-statistic is:

F = ((RSS_R - RSS_U) / p) / (RSS_U / (T - 2p - 1))

Under the null, F follows an F(p, T-2p-1) distribution. A chi-square form from the likelihood ratio is also common. The test on the second equation, with the roles of X and Y swapped, checks the reverse direction.

Worked Example

Suppose you suspect that changes in the VIX lead daily SPY returns by a day or two. Build a two-equation VAR of dVIX and SPY returns with p = 3 lags on daily data from 2010 to 2020 (T = 2,517).

The restricted regression of SPY returns on three lags of itself gives RSS_R = 0.4120. The unrestricted regression adds three lags of dVIX and gives RSS_U = 0.4081. Compute:

F = ((0.4120 - 0.4081) / 3) / (0.4081 / 2510) = 8.0

An F statistic of 8.0 on (3, 2510) degrees of freedom has p-value well below 1 percent, so you reject the null that VIX changes fail to predict SPY returns. By contrast, the reverse test with SPY return lags on the right of the dVIX equation might yield F = 12, also significant. VIX and SPY Granger-cause each other, which is common for closely linked series.

A strategy built on this result should be skeptical: a signal that works in-sample on 10 years may still be spurious or swamped by costs. Granger significance is necessary but not sufficient for a live signal.

Common Mistakes

1. Running Granger on non-stationary data. If either series has a unit root, standard F and chi-square distributions do not apply, and you can find "causality" that is a statistical artifact. Difference first or use the Toda-Yamamoto procedure on the level VAR with extra lags.

2. Treating statistical Granger causality as economic causality. The test measures forecast improvement, not mechanism. Two variables driven by the same lagged macro factor can each Granger-cause the other without any direct channel.

3. Overfitting the lag order. Too many lags lowers power and inflates false positives. Select p using AIC or BIC on the VAR and check residual autocorrelation.

4. Ignoring structural breaks. A regime that starts in 2020 can make Granger tests pass or fail depending on whether the sample spans it. Rolling-window Granger tests help expose when the lead-lag relationship actually holds.

5. Using daily close-to-close data when the effect is intraday. Many lead-lag relationships resolve within hours. A daily-resolution test washes them out. If you suspect a minutes-scale effect, run the test at the matching frequency.

Frequently Asked Questions

Q: What is Granger causality in simple terms? Granger causality tests whether knowing the history of series X improves your forecast of series Y beyond what Y's own history provides. If adding lagged values of X significantly reduces the forecast error for Y, X is said to Granger-cause Y, meaning it contains predictive information, not that it literally causes Y to move.

Q: How does the Granger causality test affect investment decisions? It formalizes the search for lead-lag relationships that could generate trading signals. If small-cap returns consistently Granger-cause large-cap returns one day later, you can build a signal that watches small caps today to predict large caps tomorrow, with a proper statistical basis rather than anecdotal observation.

Q: What is a real-world example of a Granger causality test in trading? Changes in the VIX index were tested against SPY daily returns from 2010 to 2020. The F-statistic of 8.0 with three lags rejected the null that VIX changes have no predictive content for SPY returns at the 1 percent level. The reverse test also showed significance, confirming bidirectional Granger causality between the two series.

Q: How can investors avoid the biggest Granger causality mistake? Always test each series for stationarity using the ADF test before running Granger causality. If either series has a unit root, difference it or use the Toda-Yamamoto procedure on the levels VAR with extra lags to obtain valid test statistics. Standard F-distribution critical values are invalid when applied to non-stationary data.

Q: How is Granger causality different from true economic causality? Granger causality only measures whether past values of X help predict Y; a third variable driving both can produce a positive Granger test with no direct channel between X and Y. True economic causality requires a plausible mechanism, controlled experiment, or instrumental variable analysis, none of which the Granger test provides.

Sources

1. Granger, C.W.J. (1969). "Investigating Causal Relations by Econometric Models and Cross-spectral Methods." Econometrica 37(3), 424-438. https://www.sonoma.edu/users/c/cuellar/econ411/granger.pdf
2. Toda, H.Y. and Yamamoto, T. (1995). "Statistical Inference in Vector Autoregressions with Possibly Integrated Processes." Journal of Econometrics 66(1-2), 225-250. https://www.sciencedirect.com/science/article/abs/pii/030440769401616M
3. Hamilton, J.D. (1994). Time Series Analysis. Princeton University Press, Chapter 11. https://press.princeton.edu/books/hardcover/9780691042893/time-series-analysis
4. Lutkepohl, H. (2005). New Introduction to Multiple Time Series Analysis. Springer. https://link.springer.com/book/10.1007/978-3-540-27752-1

Disclaimer

This article is educational content only and is not financial advice. Nothing here is a recommendation to buy, sell, or hold any security. Consult a licensed advisor before making investment decisions.