Does a hot streak mean anything?

It is the most natural question to ask of a grid like this one: that window has been winning lately — is it hot? Across 2,221 weekday × hour windows there is something there, and it is 0.91 of a percentage point. The reason this site has no streak feature is not the size of it. It is where it does not go away.

The lags are the finding

Condition each window's up-rate on what the previous week did in that same window, and the mean z comes out at 0.277. Positive, and small. Now do it again at two weeks' lag: 0.26. And at four: 0.246.

Where it lives

The effect is not spread evenly, and the pattern is the second reason not to build on it.

Almost all of it is in the metals. Take silver out of the whole calculation and the site-wide mean falls from 0.277 to 0.214, and the 0.91 points to 0.7. ⚠ Silver is also the one market whose data we tell you not to trust before 2006 and the one market whose effects we measure as genuinely decaying — so the group carrying this result is the group we already qualify most heavily. In the indices and in crypto there is nothing at all.

Every market

Across the whole grid, 201 windows clear |z| ≥ 2 against 101.1 expected by chance, and 45 clear |z| ≥ 3 against 6. Real, then — roughly twice noise at the looser bar — and still worth under a point.

The strongest single window anywhere is NZD/USD at that market's Fri 21:00 UTC: it closed up 56.9% of the 144 weeks that followed a winning one, against 28.1% of the 217 that did not (z = 5.48). ⚠ One window out of 2,221, at 21:00 UTC, which is the hour this site tells you to distrust.

Why there is no streak feature here

• It does not decay, so it is not memory. A product built on "this window is hot right now" would be selling a property the data says the window has all the time.
• 0.91 points does not clear a round trip in any market we hold. It is smaller than the margin every rule on our board already runs on.
• The group carrying it is the group we trust least. Remove silver and most of it leaves with it.
• A recency filter would make the sample smaller, not the signal bigger. Conditioning on last week throws away most of the history that made the window worth looking at, which is how a 300-sample window becomes a 60-sample one and every extreme reading comes back.

Full method in section 4 of the evidence page, and every claim we have tested — including the ones that held — on the reports index.

More on how we decide what counts

Every claim we have tested, including the ones that held, is on the reports index.