Probability · Statistics
Pick a colour at random from nine and every colour has the same claim on the next pick: one in nine, 11.11%, forever. What changes with every draw is not the chance but your knowledge of it — a tally that starts as noise and narrows, slowly, onto that number.
Draw as often as you like. The disc shows the last colour, the strip the last few, and each bar leans away from one in nine as that colour runs hot or cold.
last draw · —
Each bar starts at one in nine and grows towards the side it is running; a full half-width is two percentage points either way.
Every colour’s running share, from draw fifty to the latest. The shaded band is ±2 standard errors around one in nine — where a fair tally should sit most of the time.
The axis begins at the fiftieth draw and stops at 20%: earlier than that a single lucky colour sits at 20% or 30%, and the band is wider than the frame.
Nine wandering lines, one destination. The band narrows as 1/√n — four times the draws to halve it — so the last stretch to 11.11% takes far longer than the first. A line that leaves the band has wandered two standard errors from fair, which nine colours will manage every so often by chance alone.
Counts and shares from the draws so far, with a 95% interval for each share and the probability the next draw assigns to that colour.
| Colour | Count | Share | 95% interval | Next draw |
|---|
The draws are independent, so a colour that has not appeared for a while is no more likely to appear next, and the current leader is no more likely to stay ahead. The tally records what happened; it does not tilt the next draw. A run of nine reds is unusual but it is not evidence of a loaded wheel — nine of them happen about once in 98 runs of nine.
The spread of a proportion is √(p(1−p)/n), so at 100 draws a fair colour’s share typically lands within about six points of 11.11%, at 10,000 within about 0.6. Because the width falls as the square root of the count, accuracy is expensive: a hundredfold increase in draws buys a tenfold tightening, no faster.
Each 95% interval misses the true value about once in twenty experiments, so nine of them together are expected to produce a miss roughly one time in two. Overlapping intervals are therefore weak evidence that two colours differ, and a leader at 13% after 200 draws is ordinary. The last column stays at 11.11% on purpose: the model for the next draw does not move with the sample, though a Bayesian estimate, (count+1)/(draws+9), does.