Roulette Spin Tracker
Log real spins and get a real answer, not a colorful grid.
This tests whether a logged sample of spins is statistically distinguishable from a fair wheel. It cannot tell you whether any specific wheel is physically biased, and even a genuine statistical signal is not, on its own, a reason to bet differently.
European, single zero (37 pockets). Switching wheels clears your logged spins.
Nothing yet. Each spin you log lands here, newest first.
Fair wheels with a pocket that looks significant
85.0%
Testing every one of this wheel's 37 pockets individually at the ordinary 5% significance level, a perfectly fair wheel throws up at least one pocket that looks significant 85.0% of the time, purely from testing that many things at once (31.1% at the stricter 1% level). That is why this tool tests every pocket together instead of highlighting whichever one has hit the most.
Chi-square test against a fair wheel
0/ 185
Log some spins to begin
- The most extreme 5% of samples a fair wheel produces
- Your sample is marked here once it reaches 185 spins, not before.
From a signal to a bet
Even a genuine, sustained signal is not a betting edge on its own. Detecting a bias large enough to beat a 35:1 straight-up payout takes roughly 291,076 spins on that one pocket at standard statistical power, which is upward of 7,277 hours of logging at 40 spins an hour. No amount of hand-logged data on this page gets there.
Fewer spins needed does not mean a bias is easier to find. Each row prices a different physical claim about this wheel: whether one named pocket has drifted, or whether the wider set of pockets a bet covers has drifted collectively, a coarser and far larger distortion. A smaller number below means that broader claim is cheaper to confirm, never that finding any bias got easier.
| Bet | Spins to detect | Hours logging |
|---|---|---|
| Straight-upprices one named pocket drifting | 291,076 | 7,277 |
| Splitprices the 2 covered pockets drifting, collectively | 141,456 | 3,536 |
| Cornerprices the 4 covered pockets drifting, collectively | 66,646 | 1,666 |
| Dozen or columnprices the 12 covered pockets drifting, collectively | 16,772 | 419 |
| Even-moneyprices the 18 covered pockets drifting, collectively | 8,460 | 211 |
Even the cheapest row above, even-money, is asking whether this wheel has drifted to an edge it would have to be conspicuously broken to have, and confirming that here would still take roughly 211 hours of logging.
The multiple-comparisons note above is about testing every pocket at once. Testing five bet types here is a separate multiplicity of the same kind, so read any one row that looks unusually low the same way: a reason to keep logging, not a finding.
Written and reviewed by Alexander Korsager · last reviewed September 27, 2026 · corrections
The idea
What this tool shows
Several roulette sites let you log the numbers coming up at a real wheel and show hit counts, hot and cold coloring, and percentage breakdowns by color, parity, range, column and dozen. None of them run a test on the data they collect. The most careful of them prints a static disclaimer next to the coloring anyway, and the least careful brand the identical mechanic a "prediction system" or a "bias detector", which tells the reader the pattern is something to act on.
This tool logs the same spins, by click or by pasting a list, on a European (37 pockets), American (38) or triple-zero (39) wheel. What it does differently is run a chi-square goodness-of-fit test against a fair wheel on every entry, report the result as a p-value beside the sample size it was computed from, and refuse to report one at all below the sample the test actually needs. Below 185 spins on a European wheel (190 American, 195 triple-zero), the expected count in each pocket is too small for the chi-square approximation to mean anything, so the page shows the count and the gap to the floor instead of a number that looks precise and is not.
The sentence that explains why hot numbers look real in the first place is the one no tracked competitor states: on a perfectly fair European wheel, testing all 37 pockets individually at the ordinary 5% significance level, there is an 85.0% chance that at least one of them looks significant anyway, purely from testing that many things at once. A wheel with one obviously "hot" number after a session of play is not a wheel behaving unusually. It is what a fair wheel looks like almost every time you check 37 things.
Even a genuine, sustained chi-square signal would not be a betting edge on its own.Detecting a bias large enough to beat a 35:1 straight-up payout, an edge of about 0.075 percentage points on one pocket, takes roughly 291,000 spins at that pocket under standard statistical power assumptions. At 40 spins an hour, live, that is upward of 7,000 hours of logging. "Track enough spins and you'll find the edge" is not a realistic plan; it is a description of a sample size nobody hand-logs.
FAQ
Questions players ask
Is my roulette wheel biased if one number keeps hitting?
Almost certainly not, and the reason is arithmetic rather than reassurance. Testing 37 pockets individually at the ordinary 5% significance level, a perfectly fair European wheel throws up at least one pocket that looks statistically significant 85.0% of the time, just from testing that many things at once. One hot number after an evening of play is closer to the expected outcome of a fair wheel than to evidence against one. This tool's chi-square test looks at every pocket together instead of hunting for the biggest outlier, which is the test that actually answers the question.
How many spins do I need to log before the test means anything?
185 on a European wheel, 190 on American, 195 on triple-zero. That comes from the standard chi-square validity rule: every pocket needs an expected count of at least 5, and a European wheel has 37 of them (5 x 37 = 185). Below that floor this tool will not print a p-value at all, because a precise-looking number computed from too little data is worse than no number.
Could I ever collect enough spins to find a real, bettable bias?
In principle, but not by hand. Detecting a bias big enough to beat a straight-up bet's 35:1 payout, an edge of roughly 0.075 percentage points on one pocket, takes about 291,000 spins on that pocket at standard statistical power (80% power, one-sided 5% significance). At a brisk 40 spins an hour at a live table, that is more than 7,000 hours of logging one wheel. The chi-square test on this page can tell you your sample is not yet inconsistent with a fair wheel, or that it has become inconsistent with one, but it cannot get you to a bettable edge on a hand-logged sample. Real wheel-bias advantage play, historically, relied on physical measurement of the wheel and thousands of casino-recorded spins, not a player's clicker.
What does a low p-value actually mean here?
That the entered sample would be unusual to see from a perfectly fair wheel, nothing more. It does not mean the wheel is biased, and on its own it is not even good evidence of that: with this many pockets tested, a low p-value shows up on fair wheels too, at the base rate the test is built to allow. Treat a single low reading as a reason to keep logging, not as a finding, and re-check the sample-size floor before reading anything into any per-pocket count.
Why does this tool refuse to show hot and cold numbers the normal way?
It shows the counts, but subordinate to the test result rather than in front of it, because leading with a colorful grid is exactly the design choice that makes an ordinary fair-wheel fluctuation look like a finding. The page's own hard rule is that no percentage or hit count outranks the chi-square result in position or size, and no page copy anywhere on it calls a number "due", "hot" in a predictive sense, or biased on the strength of a raw count.
Methodology
How this is calculated
A chi-square goodness-of-fit test against a uniform distribution: for n logged spins spread over p pockets, the expected count per pocket is n / p, and the statistic is the sum over every pocket of (observed minus expected) squared over expected, with p minus 1 degrees of freedom.
The p-value is the upper-tail probability of a chi-square random variable at that many degrees of freedom exceeding the observed statistic, computed from the regularized incomplete gamma function rather than read off a table, so it is exact to floating-point precision rather than interpolated.
The test is withheld below the standard validity floor of an expected count of at least 5 per pocket (185 spins European, 190 American, 195 triple-zero). Color, parity, range, column and dozen breakdowns are plain counts and percentages of the same logged sample, carrying no test of their own.
Two further figures are fixed illustrative calculations rather than readings of a reader's own data: the 85.0% figure is 1 minus (1 minus 0.05) to the power of the pocket count, the chance that testing every pocket individually at 5% turns up at least one false positive on a fair wheel; and the roughly 291,000-spin figure is a one-sample proportion power calculation (80% power, one-sided 5% significance) for detecting the gap between a fair wheel's per-pocket probability and the break-even probability for a 35:1 straight-up bet, which is the first row of the five-bet ladder.
The ladder's other four rows generalize the same formula to a bet's covered-pocket count and payout rather than pricing a second calculation.
Common misreadings3
- A pocket with the highest count in your session is not "due" to hit less, and it is not confirmed "hot" either. Every spin on a fair wheel is independent of every other one, whatever has already landed.
- A p-value under 0.05 is not a 95% chance the wheel is biased. It is the chance a fair wheel would produce a sample this extreme or more, and with many pockets tested, samples this extreme happen to fair wheels at a known, non-trivial rate (the 85.0% and 31.1% figures above).
- Passing the sample-size floor is not the same as having enough data to bet on. It is the minimum for the test's own maths to hold; the amount of data it would take to find a bettable edge is two to three orders of magnitude larger depending on the wheel, roughly three on the default European wheel, per the FAQ above.
Assumptions4
- Every logged spin is an independent, honestly reported outcome. The tool cannot detect a mis-keyed number, a dealer signature the entry sequence does not capture, or spins entered out of order.
- The chi-square approximation to the true multinomial distribution is only accurate once every pocket's expected count is at least 5, which is why the sample floor is enforced rather than advisory.
- The 85.0% (and 31.1% at the 1% level) false-positive figure assumes independence across per-pocket tests, which is the standard simplification for this kind of illustrative figure and is not itself a claim about this tool's own omnibus test, which tests every pocket jointly rather than one at a time.
- The roughly 291,000-spin figure is the straight-up row of the bias-detection ladder on this page: one named pocket carrying the entire bias, priced against its 35:1 payout. The ladder prices four more common bet types (split, corner, dozen or column, even-money) the same way, and a lower spin count further down it prices a coarser claim about the wheel, several pockets drifting collectively rather than one, never an easier way to find the same bias.
Limits3
- A statistically significant result is evidence the sample is unusual for a fair wheel. It is not proof of a mechanically biased wheel, which historically has been established by physical inspection and casino-scale spin counts, not a player's own log.
- This tool has no way to verify what a reader enters. A batch paste of numbers is trusted at face value, the same way any self-reported dataset is.
- The color, parity, range, column and dozen breakdowns carry no test of their own. A percentage that looks lopsided in one of them is not evidence by itself; only the chi-square result on the full pocket distribution is.
Evidence
The chi-square goodness-of-fit test is standard inferential statistics; this page's implementation is independently checked against six published chi-square critical values (df 36, 37 and 38 at the 5% and 1% significance levels, cross-referenced against the NIST/SEMATECH e-Handbook of Statistical Methods' chi-square table) rather than against a competitor's tool, since none of the tracked competitors run this test at all.
The sample-size floor (expected count of at least 5 per cell) is the standard rule of thumb for chi-square validity, stated in the same reference.
The one-sample proportion power formula behind the roughly 291,000-spin figure is the standard normal-approximation power calculation for a one-sample test of a proportion, and math.test.ts re-derives the figure from the formula rather than pinning a copied constant, plus the other four rows of the bias-detection ladder at their real payouts, and asserts the ladder is monotonically decreasing in spins on every wheel this tool offers.
The competitor mechanic described in this page's introduction (hot/cold coloring with no underlying test, in some cases branded as predictive) is search-snippet confidence rather than a verified page read, and no specific competitor's behavior is asserted as fact anywhere in the rendered copy.
Gambling can cause serious harm. The National Problem Gambling Helpline is free, confidential and open 24/7, call or text 1-800-MY-RESET (1-800-697-3738), or chat at ncpgambling.org/chat.
Read up
Go deeper
Related tools