Does A Cause B? Granger's Modest, Powerful Answer

"Causation" is a word that has kept philosophers busy for two thousand years and has never been fully pinned down. It is also a word that economists, and traders, desperately need. Does money supply cause inflation? Does order flow cause price moves, or does price move first and drag order flow along? Does the VIX lead equities, or follow them?

In 1969, Clive Granger did something clever. Rather than try to solve the philosophy, he redefined the question into something data could actually answer, then built a test for it. The concept now carries his name, and it is both one of the most used and one of the most abused tools in applied time-series work.

The problem: correlation everywhere, direction nowhere

If two series move together, you have learned almost nothing about which one is driving. Correlation is symmetric. It has no arrow. Yet almost every interesting economic and financial question is about the arrow.

Worse, in economics you cannot run the experiment. A chemist who wants to know if X causes Y adds X and watches. An economist cannot double the money supply in one universe and hold it fixed in a parallel one. You are stuck with the historical record, a single run of the tape, and you have to extract direction from it somehow.

Granger's move was to reach for the one asymmetry that history does give you: time. The future cannot cause the past. Whatever else is murky, that much is solid. So build your definition of causation on top of the one arrow you can trust.

The key idea via analogy: the weather forecaster's test

Suppose you are trying to forecast tomorrow's temperature in London. You have a good model that uses only London's own past temperatures. It does a decent job.

Now someone hands you a second dataset: yesterday's barometric pressure readings. You add them to the model. If your forecasts get meaningfully better, then pressure contains information about future temperature that temperature's own history did not already have.

In Granger's language, pressure Granger-causes temperature.

That is the whole definition. Strip away the machinery and it reduces to a single sentence: A Granger-causes B if the past of A helps you predict B better than B's own past alone can.

The test is built directly on this. You run two forecasting models for B. One uses only B's own lagged values. The other uses B's lagged values plus A's lagged values. Then you check, statistically, whether the second model's errors are significantly smaller. If they are, A carries predictive information about B. If they are not, A adds nothing you did not already have.

And critically, the procedure is directional. You can run it the other way round: does the past of B help predict A? You get four possible answers, and they are all interesting:

• A helps predict B, but not vice versa. A one-way arrow.
• B helps predict A, but not vice versa. The opposite arrow.
• Each helps predict the other. Granger called this feedback, and it is extremely common in economics.
• Neither helps predict the other. They are, for forecasting purposes, unrelated.

Why it mattered

• It made "which came first" an empirical question. Before Granger, arguments about whether money drives output or output drives money were largely theoretical shouting matches. Afterwards, you could bring data.
• It is the beating heart of vector autoregressions. When Sims built the VAR framework a decade later, Granger causality tests became the natural way to interrogate the resulting system: which variables actually matter for which?
• It is genuinely useful in trading research. Does the futures market lead the cash market? Does one exchange's price discovery lead another's? Does credit spread widening lead equity drawdowns? These are all Granger causality questions, and answering them is directly actionable for anyone building a signal or thinking about execution.
• It escaped economics entirely. Neuroscientists use it to study which brain regions drive which. Climate scientists use it on temperature and carbon series. It became one of economics' most successful intellectual exports.

The honest limitations, and they are severe

Granger himself was careful. Most of the people who cite him are not. This is the section that matters most.

• It is a statement about prediction, not about mechanism. Granger causality is a claim about information flow in time, not about physical or economic causation. A famous illustration: the barometer falling Granger-causes rain. The barometer does not cause rain. The weather system causes both, and the barometer just gets the news first. Granger causality is entirely compatible with there being no causal link at all. Granger knew this and said so. The unfortunate choice of the word "causality" has been misleading people ever since.
• A missing third variable can invent an arrow. If some unmeasured factor Z drives both A and B, but hits A a little sooner, then A will Granger-cause B, and the relationship is pure illusion. In markets, where nearly everything is driven by a common shock arriving at slightly different speeds in different venues, this is not an edge case. It is the default situation.
• A missing third variable can also hide a real arrow. If you leave out the variable that actually carries the information, your test can conclude nothing is happening when plenty is. The test only ever sees the variables you gave it.
• It is sensitive to sampling frequency. A relationship that operates in milliseconds will be completely invisible in daily data, and one that operates over years may look like noise in daily data. Change the frequency and the answer can change.
• It assumes linearity in its standard form. If A affects B only during crises, or only above a threshold, a standard linear test can easily miss it.
• Non-stationary series make it lie. Run Granger causality on two trending, wandering series and you can get significant results out of pure spurious regression, which is the trap Granger himself, with Newbold, would document five years later.

The one-line takeaway

Granger replaced the unanswerable question "does A cause B?" with the answerable one "does A's past improve my forecast of B?", which gave economics a rigorous, directional, testable tool, and simultaneously created a lasting confusion because he gave that tool a name it does not deserve.