How Granger Causality Tests for Predictive Precedence, Not True Causal Influence

In 1969, within the pages of the journal Econometrica, British econometrician Clive Granger proposed a mathematical workaround for one of the oldest philosophical traps in statistics. The trap was the impossibility of proving true causation from observational data. Granger's solution was to shift the goalpost. Instead of asking whether variable X physically brings about variable Y, he asked a narrower, highly specific question: does knowing the history of X improve our ability to forecast the future of Y?[1]

This concept, which Granger initially termed 'temporally related' but which the wider scientific community quickly branded 'Granger causality,' became a cornerstone of time series analysis. It provided a computationally simple way to measure information flow between two sequences of data points ordered chronologically. However, the adoption of the word 'causality' created a linguistic trap that continues to snare data scientists, economists, and medical researchers decades later.[5][6]

To understand why Granger causality is not true causality, one must look at the mechanics of the test itself. The procedure relies on Vector Autoregression (VAR). An analyst first builds a restricted model, attempting to predict the current value of variable Y using only its own past values, testing anywhere from 1 to 10 previous time steps, known as lags. They then build an unrestricted model, which includes both the past values of Y and the past values of variable X.[3][4]

The mathematical verdict relies on an F-test, which compares the predictive accuracy of the two models. If the unrestricted model—the one containing X's history—produces a statistically significant reduction in forecasting error, typically measured against a standard alpha threshold of 0.05, the test rejects the null hypothesis. The analyst concludes that X 'Granger-causes' Y. The past values of X contain unique information about Y's future that Y's own history does not already possess.[4][6]

This is a statement about predictive precedence, not physical mechanism. True causal inference, as defined in modern econometrics and epidemiology, requires counterfactuals: proving that an intervention changing X would physically alter the trajectory of Y. The Granger test requires no such intervention. It merely observes the sequence of events, making it highly vulnerable to confounding variables.[2][5]

The most common failure mode occurs when a third, unobserved variable drives both X and Y, but at different speeds. Imagine a scenario where a central bank secretly decides to raise interest rates. Bond markets (variable X) might react within 50 milliseconds, while consumer mortgage rates (variable Y) take 14 to 21 days to adjust. A Granger causality test would confidently declare that bond market movements 'cause' mortgage rate hikes. In reality, both are reacting to the central bank's hidden decision; intervening to artificially manipulate the bond market would not force mortgage rates to rise.[5][6]

A second major limitation is lagged reverse causality. The test checks whether X's lags predict Y, but it cannot rule out that the true causal arrow runs from Y to X with a delay longer than the analyst has modeled. If Y physically causes X, but takes 12 months to do so, and the analyst only tests a 3-month lag, the test might register X as predicting Y simply because X is reacting to an earlier, unmeasured value of Y that has not yet fully worked through the system.[3][4]

Furthermore, the mathematical validity of the test rests on strict assumptions about the underlying data. Both time series must be stationary, meaning their statistical properties—like mean and variance—do not change over time. If the data contains trends or seasonal cycles, the test can produce spurious correlations, falsely identifying predictive precedence where none exists. Analysts must often difference the data—subtracting the value at time T-1 from the value at time T—to achieve stationarity before running the test, a step that can obscure long-term relationships.[3][4]

Despite these limitations, the test remains a powerful tool when used for its original, intended purpose: feature selection in forecasting. In fields ranging from macroeconomic nowcasting to machine learning, knowing which variables contain early signals of a target outcome is immensely valuable, regardless of the underlying physical mechanism.[5][6]

The danger arises only when the output of a forecasting filter is presented as an epidemiological or policy proof. 'Of course, many ridiculous papers appeared,' Granger noted in his 2003 Nobel lecture, referring to studies that assumed his statistical filter proved physical mechanisms. The mathematical reality remains unchanged: temporal precedence is a necessary condition for true causality, but it is never a sufficient one.[2][5]