Radar Perene / Archive / science

Correlation is not causation — and markets confuse the two

◦ Index methodology v2.2 (working papers with DOI). See the methodology.

Science

At five in the afternoon of any trading day, financial news delivers a ready-made explanation: stocks fell with the dollar, rates rose because of someone's remarks, the index gave back gains in reaction to a data release. The sentence takes minutes to write. The study that would support the "because" would take months — and is almost never done. Every day, then, the reader is served instantly manufactured causality about moves that are only known to have been simultaneous.

Correlation is the tendency of two variables to move together; causation is one of them moving the other. The first can be measured in any spreadsheet in seconds; the second requires study design — a prior hypothesis, control of common factors, tests across distinct periods — and frequently remains undemonstrable.

The distinction sounds like a classroom point. Market practice shows it is not: it separates the reading that informs from the one that merely tucks the day into a story.

The routine confusion in the Brazilian market

The most famous pair is the Ibovespa and the dollar. In many periods the two series show visible negative correlation — weak equity days tend to coincide with weak real days. From there to the headline "dollar drags stocks down" is a logical leap the data does not authorize: both series respond to common factors — global risk appetite, foreign flows, fiscal perception — and their coincidence reveals neither who pushed whom, nor whether anyone pushed at all.

The same holds for rates×stocks, oil×Petrobras, "foreign flow"×the weekly rally. In all of them a documentable correlation exists in certain windows; in none of them can the day's headline know the direction of the arrow. And there is the detail habit conceals: market correlations change sign over time. The pair that "always moved together" decouples without notice — and yesterday's explanation becomes a quiet embarrassment.

It helps to name the three traps statistics catalogs. The first is the third factor: ice cream sales and drownings rise together every summer, and the common cause is the heat — in markets, the "heat" is usually global risk appetite. The second is reverse causality: even when an arrow exists, it may point the opposite way from the headline. The third is pure coincidence: among thousands of series, some will move together for years with no relation at all — that is arithmetic, not mystery. The five o'clock headline rules out none of the three; study design exists to rule them out one by one.

What research does differently

Serious research does not renounce the causal question — it treats it with the respect its difficulty demands. Three disciplines make the difference, and none is exotic.

First, the hypothesis comes before the data: the expected relation, its direction and its mechanism are declared before looking at the result. Second, control: one hunts for the third factor that could move both series at once — in the Brazilian case, the global environment is usually the prime suspect, and ignoring it condemns the study. Third, robustness: the relation must survive across subperiods, other windows, other measures. A correlation that only exists between 2016 and 2019 is not a law; it is an episode.

Applying that full design to a specific pair — with chosen controls and documented tests — is bench work, the kind that does not fit in a concept article. But the ruler is public and any reader can use it: whenever someone claims that A moves B, the three questions are always the same. Did the hypothesis come first? Was the third factor sought? Does the relation survive outside the window that displayed it?

An episode from the archive

The house doctrine for the currency×stocks pair is written in the entry on the dollar as thermometer: the exchange rate is treated as a gauge of the environment, not as a cause of other prices — the coincidence is recorded, the why is left unsigned. The archive keeps a test of that discipline: in December 2024, the dollar left its own historical pattern, and that week's headlines offered chains of causes complete with names and timestamps. The house record, reread today, describes the move, the environment and the rarity — and stops there, in the episode that stayed on file. That week's confident causes aged the way such causes usually age; the descriptive record did not.

This is not decorative modesty. It is the practical consequence of the definition at the top: the correlation was in the data; the cause was not. Writing only what the data contained is what lets an archive be reread years later without embarrassment.

Frequently asked questions

Does high correlation never indicate cause?

It indicates something worth investigating. Persistent correlation is raw material for a hypothesis — the error is not noticing it, but announcing it as a cause without the work the word demands.

Why does the news insist on "because of"?

Because causal narrative is the format readers expect from news, and the deadline does not accommodate a study design. The defect is structural to the genre, not individual bad faith of whoever writes it.

How does one test causality in markets, in practice?

With design: hypotheses declared in advance, common factors controlled, subperiods compared, and — when possible — events that work as natural experiments. Even so, a good share of causal questions in macro remains without a definitive answer.

What about two series that move together for decades?

Longevity strengthens the correlation, not the cause. Two series can follow the same third factor for decades — and long correlations have also come undone when the regime changed.

---

If correlation deceives along the axis of time, the next defect on this trail deceives along the axis of the sample: series that only show the survivors — Survivorship bias.

House reading: the record without manufactured cause is in today's reading, in the Diário; the coincidences time has judged are among the precedents, in the Atlas.

Putting a specific pair of series through this study design is the kind of exercise the house conducts on request.

Characters: Dollar