What does it mean to say oil prices Granger-cause airline stock returns?

## The definition Clive Granger's idea, from 1969, is deliberately operational: **X Granger-causes Y if past values of X help predict Y better than the past of Y alone.** Nothing more is claimed. Formally, take the model `y_t = a + b_1*y_{t-1} + ... + b_p*y_{t-p} + c_1*x_{t-1} + ... + c_p*x_{t-p} + e_t` and test the joint null `c_1 = c_2 = ... = c_p = 0` with an F-test (or the equivalent Wald test). Rejecting the null means the lags of X add predictive content, and X is said to Granger-cause Y. Applied to oil and airlines: if last week's oil prices sharpen the forecast of this week's airline returns beyond what airline returns' own history gives, oil Granger-causes airline returns. ## Why the name misleads Granger himself preferred to talk about *temporal precedence in predictability*, and the confusion he predicted is exactly what interviewers probe. Three distinct failure modes: **Omitted common driver.** Suppose global demand conditions move both oil and airline profitability, and oil markets happen to reprice a day faster than equities. Oil will Granger-cause airline returns even though the causal arrow runs from demand to both. The test sees only the two series it is given; it cannot condition on anything outside the model. **Anticipation.** Financial series price expectations, not just realised events. If traders expect fuel costs to fall next quarter, airline shares can move *before* oil does, and the test will report airline returns Granger-causing oil prices — the reverse of the physical mechanism. Any forward-looking market can invert the apparent direction. **Sampling frequency.** If the real transmission happens within a day and you have monthly data, both series move within the same observation and neither precedes the other; the test sees nothing, or attributes the ordering to noise. Conversely, aggregation can create apparent precedence where none exists at the underlying frequency. ## Practical requirements - **Stationarity.** The standard F-test assumes stationary inputs. Run it on two drifting price levels and the test statistic no longer has its usual distribution; you get inflated rejection rates for the same reason a levels regression between two wandering series looks significant. The usual practice is to test on returns or differences. - **A unit-root-safe alternative.** If you must work in levels, the Toda-Yamamoto procedure fits a model in levels with extra lags equal to the maximal order of integration and tests only the original lags; that Wald statistic has a standard asymptotic chi-squared distribution regardless of whether the series have unit roots or are cointegrated. - **Lag length matters.** The conclusion can flip with the number of lags included, so choose the lag order by an information criterion and report the sensitivity rather than picking the length that gives the answer you wanted. - **Direction is not exclusive.** Testing both directions can find that each series Granger-causes the other. That is feedback, and it is a perfectly legitimate finding — it simply means each carries predictive information about the other. ## What it is good for Despite the caveats, the test earns its place. It is a disciplined way to ask whether a candidate leading indicator actually carries usable predictive content beyond the target's own history — which is precisely the bar a feature must clear before it belongs in a forecasting model. Framed as "does adding this series improve my forecast?", it is honest and useful. Framed as "does this series cause the other?", it overclaims. ## Interview framing The strongest answer states the definition in forecasting terms in one sentence, then names at least one concrete way the test can point the wrong way — a common driver, or anticipation reversing the order — and finishes with the stationarity requirement. Candidates who simply say "it proves causality" or "it proves nothing at all" both miss: it is a real, testable, useful claim about predictive precedence.