Granger Causality

The fundamental axiom of Granger causality is that the cause must precede the effect. The model strictly tests whether lagged values of variable X provide statistically significant information about future values of Y. This temporal ordering is non-negotiable; if X does not occur before Y, it cannot Granger-cause Y. This principle distinguishes it from contemporaneous correlation, where two variables move simultaneously without a clear directional lead-lag relationship.

Granger causality is not about perfect prediction but marginal improvement in forecasting accuracy. The test compares two models:

• Restricted Model: Predicts Y using only past values of Y (autoregressive terms).
• Unrestricted Model: Predicts Y using past values of both Y and X. If the unrestricted model yields a statistically significant reduction in forecast error (measured via F-test on residual sum of squares), X is said to Granger-cause Y. This is purely about information content, not mechanistic causation.

The standard Granger causality test assumes the underlying time series are covariance stationary—meaning their mean, variance, and autocorrelation structure are constant over time. Non-stationary data can produce spurious regression results, falsely indicating a causal relationship. In practice, analysts must first difference the data or apply cointegration tests. If two non-stationary series are cointegrated, a Vector Error Correction Model (VECM) must be used instead of the standard VAR-based Granger test.

Granger causality is a data-driven, predictive relationship, not a structural causal mechanism. It cannot distinguish between:

• Direct causation (X → Y)
• A common confounding variable Z driving both X and Y
• Reverse causality where Y actually drives X but with a longer lag Clive Granger himself emphasized this limitation. For true causal identification, combine Granger tests with instrumental variables, Directed Acyclic Graphs (DAGs), or randomized experiments. In financial markets, a leading indicator may Granger-cause returns without having any economic causal link.

The test is inherently asymmetric and bidirectional. The hypothesis 'X Granger-causes Y' is tested independently from 'Y Granger-causes X'. It is entirely possible to find:

• Unidirectional causality: X → Y but not Y → X
• Bidirectional causality (feedback): X → Y and Y → X
• Independence: Neither causes the other In high-frequency trading, feedback loops are common—order flow Granger-causes price changes, and price changes Granger-cause subsequent order flow, creating a dynamic microstructure loop.

The choice of lag order (p) is critical and non-trivial. Too few lags omit relevant historical information, causing omitted variable bias. Too many lags introduce multicollinearity and reduce the test's power. Selection is typically automated using information criteria:

• Akaike Information Criterion (AIC): Balances fit and complexity
• Bayesian Information Criterion (BIC): Imposes a heavier penalty for extra parameters, favoring parsimony In financial applications with tick data, cross-validation on out-of-sample predictive performance often supplements these theoretical criteria.