Study 03 · statistical power

How many spins does wheel-bias detection need?

Power calculations for detecting a roulette pocket that lands 0.25 to 2 percentage points more often than the fair 1/37 baseline.

A “hot” pocket is easy to spot after the fact and hard to establish before the fact. This study calculates the observation count needed to detect one pocket whose true probability exceeds 1/37, first when that pocket was named in advance and then when all 37 pockets were searched.

2.703%
Fair pocket probability
80%
Target power
5%
False-positive rate
one-sided
1,105
Smallest scan-all sample
for a +2 point bias

Method

The calculation uses a one-sided normal approximation to a binomial test. The null isp₀ = 1/37; the alternatives add 0.25, 0.5, 1, or 2 percentage points. We solve for 80% power at a 5% false-positive rate. The scan-all column applies a Bonferroni correction of 0.05 / 37 because checking every pocket creates 37 chances to find an impressive fluctuation.

Seed
None — this is an analytic power calculation
Null probability
1/37
Power / α
80% / 5% one-sided
Multiple testing
Bonferroni correction across 37 pockets
True excessTrue pocket rateNamed in advanceScan all 37
+0.25 points2.953%26,792 spins63,283 spins
+0.5 points3.203%6,887 spins16,110 spins
+1 points3.703%1,812 spins4,165 spins
+2 points4.703%495 spins1,105 spins
+0.25 point bias
63,283 spins
+0.5 point bias
16,110 spins
+1 point bias
4,165 spins
+2 point bias
1,105 spins
Bar lengths share one scale within this figure. Exact values appear at right.

Why casual spin logs mislead

A full one-percentage-point excess is large in roulette terms: the pocket rises from 2.703% to 3.703%. Even then, naming the pocket before observation requires about 1,812spins. Searching the whole wheel raises that to 4,165. A subtler quarter-point bias pushes the scan-all requirement beyond 63,283 spins. A few hundred observations can suggest a hypothesis; they usually cannot carry the proof.

Limitations

These are planning numbers, not guarantees. The normal approximation is less exact than a discrete binomial power calculation, wheel conditions may drift during collection, and spins from a physical wheel may not be identically distributed. Bonferroni control is conservative. A real protocol should pre-register the wheel, pocket, collection rules, and stopping rule before looking at results.

Audit the numbers.The chart data is checked in as plain CSV; the combined snapshot is also available as JSON.

Sources and implementation references