Search every window. See how many are noise.

How the noise number is calculated

The grid

Every weekday × hour bucket in every selected market, pooled across all months and all years — the same buckets the dashboard reads, on the same UTC bars.

The null

For each bucket we keep its real sample size and ask: if this market had no hour-of-week structure at all — if every hour simply matched the market's own overall up-rate — how often would it clear the filter you just set? Summing that probability over every bucket searched gives the expected number of hits from noise alone.

The baseline

Every figure here — z, the gap, the direction filter — is measured against that market's own up-rate across all its hours, not against a flat 50%. In FX the two are the same question: every baseline sits within a point of 50. In the equity indices they are not. NAS100 and SPX500 close up 52.3% of every hour they trade, so scored against 50 their buckets all look bullish because the index is. Measured against their own baseline, not one of the four indices produces a single significant window — which is exactly what /evidence already concluded by a different route.

Accuracy

Computed analytically (normal approximation to the binomial) so it updates as you type. It lands a little above the 4,000-draw simulation behind /evidence — 15.3 against 13.6 at our default settings — and we keep the higher number rather than the flattering one.

What it is not

It is not a p-value for any individual row, and a result that beats the noise count is not thereby a trade. It answers one question: is this shortlist longer than a coin-flip universe would have handed me?