Pairs need an economic prior before a p-value

Test two hundred price series against each other and some will look cointegrated. At a five per cent threshold you expect ten false positives before a real relationship appears. A pairs screen that reports every one of them is a random number generator with a spreadsheet attached.

Crossview applies three defences together, not one at a time.

An economic prior comes first

A candidate has to be in the leader's sector, or carry a 30 day correlation above a floor. Two unrelated names that happen to share a drift do not get tested. This is the cheapest filter and the one that removes the largest number of nonsense pairs.

Multiple testing is corrected, not ignored

Every leader is tested against its own candidate set, and the p-values in that set are corrected with the Benjamini-Hochberg false discovery rate. The count of tests is recorded on the pair, so a reader can see how much correction was applied. A p-value of 0.01 means one thing when it came from four tests and something else entirely when it came from four hundred.

Economic significance is required as well

Statistical significance is a low bar. Each surviving pair also carries a hedge ratio on winsorised log prices, a spread z-score, an estimated mean reversion half-life, a lead-lag estimate and a suggested shape. A pair with a p-value of 0.001 and a half-life of nine months is not a trade, it is a curiosity.

What this costs

Fewer pairs. Some genuine relationships are excluded because the sector prior does not reach them, and some are excluded because the correction is strict. That is the intended trade. A short list you can check beats a long list you cannot.

What the numbers on screen mean

The spread z-score is the current dislocation, in standard deviations from the spread's own recent mean. The half-life is how long the spread has historically taken to cover half the distance back. The pair scan is rebuilt daily from a fresh hand-off, and a stale scan is left out rather than served. If the pairs panel is empty, the hand-off did not arrive that day.

The honest limit

Cointegration is estimated on a window, and relationships break. A pair that passed every test for six months can stop working next week, and nothing in the screen will warn you before it does. The z-score tells you where the spread is, not whether the reason for it still holds.