Correlation and cointegration
Correlation describes how two assets moved together day to day. It says nothing about whether they stay tethered — which is the property a pair trade depends on.
Two questions that sound like one
"Do these two move together?" turns out to be two separate questions, and conflating them is the most common way a pair trade goes wrong.
The first: when one moved yesterday, did the other move with it? That is correlation.
The second: do they stay tethered, so that when they drift apart they are pulled back? That is cointegration.
A pair can score well on the first and fail completely on the second. That combination is exactly the case that looks attractive on a screener and loses money.
What correlation measures here
Correlation on this site is Pearson's coefficient computed on daily log returns — not on prices. The distinction matters more than it sounds.
Correlation on price levels is close to meaningless in a trending market. Any two assets that both rose over a year will show a correlation near 1, because both series went up and to the right. It is a statement about the shared trend, not about any relationship between them. Run it on two unrelated assets during a bull market and you will still get 0.9.
Correlation on returns asks a stricter question: on the days one moved up, did the other move up too? That is a statement about co-movement, and it survives the trend being removed.
Reading the value:
Above 0.7 — the legs moved together consistently. The regression fit has something real underneath it.
0.4 to 0.7 — partial. Much of each leg's movement is its own. The residual will be noisier and the hedge less effective.
Below 0.4 — most of what the regression is fitting is noise. β will still come out as a number, and the z-score will still be computed, but they are describing a relationship that is barely present.
Negative — the legs moved in opposite directions. This shows up together with a negative β, and it means there is no hedge to construct.
What correlation misses
Here is the failure case, and it is worth being concrete about.
Take two assets. Every day, both move in the same direction by similar percentages — correlation near 0.95. But one of them also drifts upward a little faster, week after week. Not enough to break the daily co-movement, just a persistent grind.
Correlation stays high the whole time. The gap between them widens without limit.
A pair trade on that relationship loses money continuously while every co-movement statistic looks healthy. The daily correlation is measuring the wrong horizon: it sees each day's move and never sees the accumulated drift.
What cointegration adds
Cointegration is the property that says the gap itself is stable. Formally: two series, each of which wanders on its own, are cointegrated if some linear combination of them does not wander — it stays around a fixed level, and when it strays it is pulled back.
That linear combination is the spread. Cointegration is precisely the statement that the spread is mean-reverting, which is the property a pair trade actually depends on. Not that the legs move together — that they stay tethered.
The standard tests are Engle–Granger, which fits the relationship and then tests the residual for a unit root, and Johansen, which handles more than two series at once.
What this site measures, and what it does not
Being direct about this, because the difference is the point of the article.
This site does not run a formal cointegration test. No Engle–Granger, no Johansen, no augmented Dickey–Fuller statistic, no p-value on stationarity. Nothing here should be read as a claim that a pair is cointegrated in the statistical sense.
What it measures instead are three properties that a cointegrated pair would exhibit, each reported separately so you can see which one is doing the work:
Correlation on log returns — whether the legs co-move at all.
Half-life from an Ornstein–Uhlenbeck fit — whether deviations in the spread were historically pulled back, and how quickly. This is the closest thing here to a reversion test: a half-life that is estimable and shorter than the sample is evidence the residual behaved like a mean-reverting series over the window. It is evidence, not a test with a confidence level attached.
Spread z-score — where the residual sits right now relative to its own recent range.
A pair with high correlation, an estimable half-life well inside the sample, and a stable β has the fingerprints of a cointegrated relationship. A pair with high correlation and no half-life estimate is the failure case described above: co-moving, untethered, and drifting.
Read that way, the half-life column is not a secondary detail. It is the column that separates the two questions this article opened with.
A note on what any of this can establish
All of these statistics are computed over a finite past window and describe that window. Cointegration tests included — a formal test would give you a p-value on the hypothesis that the residual was stationary in the sample, and relationships between crypto assets break for reasons no historical test can anticipate.
The honest framing of every number on this site is descriptive: this is how the pair behaved, over this window, on this timeframe. What it does next is a separate question, and not one a screener can answer.
You can see all three metrics side by side for any pair in the screener.
These pages describe how the site computes its metrics. They are not trading advice and not a recommendation to enter any position.