Granger Causality

Granger causality tackles a very specific question: does knowing the past of series $X$ help you predict series $Y$ better than $Y$'s own past already does? If yes, we say "$X$ Granger-causes $Y$." The name is unfortunate, because it is not about causation in the everyday sense. It is about predictive usefulness with a time ordering: $X$'s history contains information about $Y$'s future that $Y$'s own history did not already carry.

Clarence "Clive" Granger's insight was to make a slippery idea testable. Cause is philosophically messy, but "does adding these extra lags shrink my forecast error?" is a clean statistical question you can answer with a regression and an F-test.

The two-regression test

Fit two models of $Y_t$. The restricted model uses only $Y$'s own past:

$$Y_t = a + \sum_{i=1}^{p} b_i\, Y_{t-i} + u_t .$$

The unrestricted model adds lags of $X$:

$$Y_t = a + \sum_{i=1}^{p} b_i\, Y_{t-i} + \sum_{j=1}^{p} c_j\, X_{t-j} + e_t .$$

Here the $b_i$ are the weights on $Y$'s own past, the $c_j$ are the weights on $X$'s past, and $u_t, e_t$ are the leftover errors. The null hypothesis is that all the $c_j$ are zero, meaning $X$'s past adds nothing. You test it with an F-test comparing the two models' leftover error (residual sum of squares, RSS):

$$F = \frac{(\mathrm{RSS}_{\text{restricted}} - \mathrm{RSS}_{\text{unrestricted}})/p}{\mathrm{RSS}_{\text{unrestricted}}/(n - 2p - 1)} .$$

If adding $X$'s lags cuts the error by a lot more than chance would, $F$ is large, the p-value is small, and you conclude $X$ Granger-causes $Y$.

Model	Predictors of $Y_t$	What it captures
Restricted	$Y$'s own lags only	how well $Y$ forecasts itself
Unrestricted	$Y$'s lags + $X$'s lags	whether $X$'s past adds forecasting power

The test is simply: does moving from the top row to the bottom row shrink the forecast error by more than luck would?

Worked example

You have $n = 120$ months of a factor's returns ($Y$) and a sentiment index ($X$), and you use $p = 2$ lags. Fitting the restricted model (returns on their own two lags) leaves $\mathrm{RSS}_{\text{restricted}} = 100$. Adding two lags of sentiment lowers it to $\mathrm{RSS}_{\text{unrestricted}} = 88$. Then

$$F = \frac{(100 - 88)/2}{88/(120 - 5)} = \frac{6.0}{0.765} \approx 7.8 .$$

With $p = 2$ and $115$ degrees of freedom, an $F$ of $7.8$ sits well past the $5\%$ critical value of about $3.1$, so you reject the null: sentiment's past does help predict this factor's returns. Note what you have and have not shown, you have shown predictive usefulness with the right time ordering, not that sentiment causes returns.

The causation trap

The word "causality" oversells what the test delivers. Three ways it misleads:

• A hidden common driver. If some third variable $Z$ moves $X$ first and $Y$ a bit later, $X$ will Granger-cause $Y$ even though $X$ does nothing to $Y$. This is the classic confounder, the same problem Correlation vs Causation warns about, now dressed in a time index.
• Anticipation reverses the arrow. Markets are forward-looking. If traders foresee an event, prices move before the fundamental does, so the price Granger-causes the fundamental, the arrow points the wrong way relative to true cause.
• It is only linear and only about the lags you chose. A nonlinear link or the wrong lag length can hide a real relationship or manufacture a fake one.

Practical cautions

• Stationarity first. Run the test on stationary series. On trending or unit-root data it inherits the spurious-regression disease and rejects far too often. Difference the series, or if they are cointegrated use the error-correction form.
• Lag length matters. Too few lags miss the relationship; too many burn degrees of freedom and add noise. Pick $p$ with an information criterion and check robustness.
• It is a joint test. You are testing all the $c_j$ at once, so report the F-test, not the individual t-stats, which multiple testing would render unreliable.

In quant work Granger tests show up when screening lead-lag relationships, does one asset's move precede another's, does an alternative-data feed lead returns. Useful for building forecasts, dangerous the moment someone reads "causality" literally. Pair it with Correlation vs Causation and, for genuine causal claims, instrumental variables.