Radar Perene / Archive / science

What cointegration is: when two series walk on a leash

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

Science

A woman leaves the bar late at night with her dog on a leash. She wanders with no particular direction; the dog sniffs its way left and right, equally aimless. Nobody can predict the next step of either one. And yet an observer watching from across the street knows one thing for certain: the two of them never drift farther apart than the length of the leash. The economist Michael Murray used that scene, in a classic 1994 article, to explain one of the most useful — and least translated — concepts in time-series statistics.

Cointegration is the property of two or more series that individually wander with no predictable path, but whose difference — or some combination of them — remains stable over time. Each series can go anywhere; the distance between them cannot. It is the invisible leash that turns two aimless walks into a long-run pair.

The term barely exists in Portuguese outside theses and formula-heavy econometrics material. That absence has a cost: "these two series move together" is one of the most repeated phrases in market commentary, and the person saying it almost never distinguishes between a genuine leash and a coincidence of routes.

What correlation cannot see

The most common confusion is treating cointegration as a fancy synonym for correlation. The two concepts answer different questions — and the difference is exactly where the mistake lives.

Correlation measures whether two short-term movements tend to occur in the same direction: when one rises, does the other usually rise? Cointegration asks something else: is there a relationship between levels that survives time — an anchor that pulls the two back whenever they stray too far?

The answers can diverge in both directions. Two series can be highly correlated day to day and drift apart forever, like two boats rocking on the same waves while sailing to different continents. And two series can show low correlation in their daily moves and still never lose each other — the woman and the dog rarely step at the same instant, but the leash is still there.

There is a known aggravating factor on this trail: series that merely grow over time produce sky-high correlations with one another despite no relationship at all — the spurious regression, the classic false alarm of market statistics. Cointegration was born, in good part, as the formal antidote to that alarm: a criterion for separating "genuinely move together" from "happened to grow in the same era".

The test the folklore never runs

The formalization has an address: Robert Engle and Clive Granger published in 1987 the framework that makes it possible to test whether a leash exists — and Granger received the 2003 Nobel Prize in Economics, in part, for that contribution. The mechanics of the test belong in the machine room; its spirit fits in one sentence: estimate the long-run relationship between the two series and check whether the deviation around it behaves like something that always comes back, or like something that leaves for good.

Market folklore is full of pairs that "move together": index and currency, sector and interest rate, an asset and its trading-floor neighbor. What almost never accompanies the claim is the test. This house has been through the experience of applying the ruler to its own collection: the working paper Brazilian Intramarket Relationships, public on Zenodo, retested a set of relationships between segments of the Brazilian market that custom treats as reliable gears. In numbers: of the relationships examined, a single one survived the retest — and the study's second version openly revises a finding the first one upheld. Most of the advertised leashes, when pulled, were attached to nothing.

Long-run relationships of this kind continue to be tested in the house's research — the detail of applied method, series by series, belongs to the workbench, not to the public article. The principle, though, is publishable: "they move together" is a hypothesis, not an observation. Hypotheses get tested.

Why the concept matters for someone who only reads

The market reader does not need to run a cointegration test to benefit from the concept. What matters is the question it installs: when a text claims two series walk together, is that a claim about short-term jolts or about a long-run anchor? Did the author test the anchor, or photograph a season of similar routes?

The question has practical consequence, because the two readings fail in different ways. Trusting a short-term correlation as if it were a leash leads to expecting reunions that never happen. And ignoring a real leash leads to treating as a definitive rupture what the archive records as a temporary separation. In both cases the error is born of a single word — "together" — used without saying in which of the two senses.

Frequently asked questions

Is cointegration the same as high correlation?

No. Correlation measures the synchrony of short-term movements; cointegration measures the existence of a relationship between levels that persists over the long run. Highly correlated series may not be cointegrated, and cointegrated series may show low correlation day to day.

Do two cointegrated series guarantee that the gap between them always closes?

The test says that, in the sample examined, separations behaved as temporary. Out of sample, the relationship can weaken or disappear — economic relationships change regime, and a statistical result is not a contract.

Do I need the mathematics to use the concept?

To read research, no. The relevant question for the reader is whether the author distinguished "similar movements" from "a tested long-run relationship" — a distinction of method, visible without a single equation.

Is cointegration useful for trading pairs of assets?

There is a literature on strategies built on cointegrated pairs, and it lives with every problem already described on this trail: flattering sample windows, relationships that expire, costs the simulation forgets. This article describes the concept; it does not recommend a use.

The leash is the exception; the rule is the aimless walk. The next step on the trail examines the series that owes nothing to anyone — not even to its own past: Random walk: yesterday's price does not predict tomorrow's →

House readings: today's note, in the Daily · the precedents, in the Atlas.

Testing whether a specific pair of series carries a real leash is workbench material — the kind of exercise the house conducts on request.