Hot and Cold Roulette Numbers: Why Tracking Does Not Work
Every roulette lobby has a board of hot and cold numbers, and every serious player should know what it actually tells you: nothing about the next spin. Each spin of a single-zero wheel is an independent 1-in-37 draw, so past frequency carries zero predictive power. This article walks through the math, the history of the one era when tracking really did work, and why provably-fair on-chain roulette closes that door for good.
Spin provably-fair roulette →What a hot and cold board actually shows
A hot and cold display is just a frequency table over a recent window — usually the last 100 to 500 spins. Numbers that landed more often than average are marked hot; numbers that landed less often are marked cold. That is a true statement about the past. The mistake is reading it as a statement about the future. The board is descriptive, not predictive: it summarizes spins that have already been paid out, and the wheel keeps no memory of them. Casinos display these boards because players enjoy them and because they encourage action, not because they contain an edge. If hot numbers really did keep hitting, the operator showing you the board would be handing you its own bankroll.
Every spin is an independent 1-in-37 event
On a single-zero wheel there are 37 pockets, and on a fair wheel each has an identical 1/37 chance — about 2.70% — on every spin, regardless of anything that came before. Independence is the whole story. The number 17 hitting three times in an hour does not make it more likely to hit again (the hot-hand reading) and does not make it 'due' to go quiet (the gambler's fallacy reading). Both readings assign memory to a device that has none. The probabilities reset completely at every spin, which is why the straight-up payout of 35:1 against true odds of 36:1 produces the same 2.70% house edge on a hot number, a cold number, and a number that has never appeared on the board at all.
The law of large numbers versus short-run clusters
The law of large numbers says that over a very long run, each number's observed frequency converges toward 1/37. What it does not say is that short samples look uniform — in fact, genuinely random data is required to be lumpy in the short run. Over 370 spins the expected count per number is 10, but the natural spread means some numbers will land 16 or 18 times while others land 3 or 4, purely by chance. Streaks and gaps are not anomalies to be explained; they are exactly what randomness looks like at small sample sizes. A perfectly even distribution over a few hundred spins would actually be suspicious. So when the board shows a scorching number, you are not looking at a signal — you are looking at ordinary variance doing what variance does.
When tracking DID work: the biased-wheel era
There is one honest chapter in the tracking story, and it belongs to physical wheels. In 1873, British engineer Joseph Jagger had clerks record tens of thousands of spins at Monte Carlo and found a wheel whose worn mechanics favored a small set of numbers; he won a fortune before the casino began swapping wheel parts nightly. A century later, Gonzalo Garcia-Pelayo and his family logged spins across Madrid casinos in the early 1990s, ran the data through a computer, and exploited measurable pocket biases until casinos upgraded their equipment and courts confirmed he had broken no law. Note what made these wins possible: a physical defect, thousands of recorded spins, and rigorous statistics — not a hunch about hot numbers. The moment the bias was repaired, the edge evaporated.
Why bias cannot exist in provably-fair roulette
On-chain roulette has no rotor to wear down, no frets to loosen, no dealer signature. The outcome comes from a deterministic function of a server seed and your client seed. On Llamabet the SHA-256 hash of the server seed is committed on the Sui blockchain before you bet, and the seed is revealed after the round, so you can recompute the result yourself and confirm the house could not have altered it. The commit-reveal scheme flips the Jagger problem on its head: instead of needing 30,000 observations to detect a bias, you can verify every single spin directly, and anyone can audit the revealed seeds to confirm the distribution is uniform. There is no hidden physical process left for tracking to discover.
What number tracking is actually good for
Entertainment — and that is a legitimate answer. Betting your birthday, riding a number that just hit, or fading the coldest pocket on the board are all fine ways to choose among bets that carry an identical 2.70% edge on a single-zero wheel. Tracking gives the session texture and stories; it just cannot give it an edge. The disciplined version of playing roulette is choosing a stake you are comfortable losing (Llamabet's minimum is 1 SUI), knowing the edge going in, and treating any pattern on the board as decoration. If a strategy seller claims their hot-number system beats independent spins, they are selling you the gambler's fallacy with extra steps.
Frequently asked questions
Do hot numbers in roulette hit more often?
No. A hot number is simply one that hit more often in a recent sample. On a fair single-zero wheel every number has the same 1/37 chance on the next spin regardless of history, so hot streaks are ordinary short-run variance, not a signal. Physical wheel bias once made tracking profitable, but that cannot exist in RNG roulette with an on-chain commit-reveal.
Is a cold roulette number due to hit?
No — that is the gambler's fallacy. Spins are independent, so a number that has not appeared in 200 spins still has exactly a 1/37 chance on the next one. The law of large numbers describes convergence over enormous samples; it never obligates the wheel to catch up in your session.
How do I know on-chain roulette is not biased?
Because you can check it. Llamabet commits the SHA-256 hash of the server seed on Sui before you bet and reveals the seed after the spin. Recomputing the outcome from the revealed seed proves the result was fixed before your wager, and the public record of seeds lets anyone verify the distribution is uniform.