A single-zero roulette wheel offers about ten standard bets: a single number, two numbers, a row of three, a block of four, red or black, a dozen, and so on. They pay very different amounts, from 35 to 1 down to 1 to 1. Yet every one of them has exactly the same expected value: −1/37 of the stake.
That follows from one line of arithmetic. It's also a clean example of expected value, exact fractions, and how little the mean tells you about a random variable. The code blocks below form one script: paste them into a single .mjs file in order and node prints the output shown under each.
The setup
A European wheel has 37 pockets: the numbers 1 to 36 plus a single 0. Each spin lands on one pocket uniformly at random.
A bet covers c pockets and pays p to 1. If the ball lands in one of your c pockets you get your stake back plus p times the stake. Otherwise you lose the stake. Per unit staked, the net result is:
- +p with probability c / 37
- −1 with probability (37 − c) / 37
The one-line derivation
E = (c/37)·p − (37 − c)/37
= (c·p + c − 37) / 37
= (c·(p + 1) − 37) / 37
Now look at the standard payouts. In every case, c · (p + 1) = 36:
| Bet | Covers c | Pays p to 1 | c · (p + 1) |
|---|---|---|---|
| Straight | 1 | 35 | 36 |
| Split | 2 | 17 | 36 |
| Street | 3 | 11 | 36 |
| Corner | 4 | 8 | 36 |
| Six line | 6 | 5 | 36 |
| Dozen, column | 12 | 2 | 36 |
| Red/black, odd/even, high/low | 18 | 1 | 36 |
The payouts are set as if the wheel had only the 36 numbered pockets, p = 36/c − 1, which would make every bet exactly fair. Every standard bet size (1, 2, 3, 4, 6, 12, 18) divides 36, so that payout is always a whole number. But the wheel also has the zero, and the payouts ignore it. So:
E = (36 − 37) / 37 = −1/37 ≈ −2.70%
for every one of these bets, whatever c is.
In code, with exact fractions
Floating point would print -0.02702702702702703 and hide the structure, so the script keeps the numerator and denominator as integers and reduces them:
const gcd = (a, b) => (b === 0 ? Math.abs(a) : gcd(b, a % b));
const frac = (n, d) => {
const g = gcd(n, d);
return `${n / g}/${d / g}`;
};
// Numerator of the expected net gain per unit staked; the denominator is `pockets`.
const expectation = (covers, payout, pockets) => covers * (payout + 1) - pockets;
// [name, pockets covered, pays N to 1]
const bets = [
["straight", 1, 35], ["split", 2, 17], ["street", 3, 11], ["corner", 4, 8],
["six line", 6, 5], ["dozen", 12, 2], ["column", 12, 2],
["red/black", 18, 1], ["odd/even", 18, 1], ["high/low", 18, 1],
];
console.log("bet covers pays c·(p+1) 37 pockets 38 pockets");
for (const [name, c, p] of bets) {
console.log(
name.padEnd(10),
String(c).padStart(6),
`${p}:1`.padStart(5),
String(c * (p + 1)).padStart(8),
frac(expectation(c, p, 37), 37).padStart(11),
frac(expectation(c, p, 38), 38).padStart(11),
);
}
bet covers pays c·(p+1) 37 pockets 38 pockets
straight 1 35:1 36 -1/37 -1/19
split 2 17:1 36 -1/37 -1/19
street 3 11:1 36 -1/37 -1/19
corner 4 8:1 36 -1/37 -1/19
six line 6 5:1 36 -1/37 -1/19
dozen 12 2:1 36 -1/37 -1/19
column 12 2:1 36 -1/37 -1/19
red/black 18 1:1 36 -1/37 -1/19
odd/even 18 1:1 36 -1/37 -1/19
high/low 18 1:1 36 -1/37 -1/19
The same formula covers other wheels. An American wheel has a second zero, 38 pockets, and every bet in the table becomes (36 − 38)/38 = −1/19 ≈ −5.26%. A triple-zero wheel gives (36 − 39)/39 = −1/13 ≈ −7.69%.
The one bet that breaks the pattern
The American table has one extra bet, the five-number "basket" (0, 00, 1, 2, 3), which pays 6 to 1. Five doesn't divide 36: a payout of 36/5 − 1 = 6.2 isn't a whole number, so it's rounded down to 6, and c · (p + 1) = 5 · 7 = 35.
console.log("basket:", frac(expectation(5, 6, 38), 38), "=", ((expectation(5, 6, 38) / 38) * 100).toFixed(2) + "%");
basket: -3/38 = -7.89%
That makes it the worst standard bet on the American table, and the formula shows why at a glance.
A rule that does change the expectation
Some single-zero tables use "la partage": when the ball lands on 0, an even-money bet (red/black, odd/even, high/low) gets half its stake back. The zero outcome is now −½ instead of −1:
E = 18/37·(+1) + 18/37·(−1) + 1/37·(−½) = −1/74 ≈ −1.35%
// Work in halves of a unit: +2 on a hit, −2 on a miss, −1 on zero.
console.log("la partage:", frac(18 * 2 + 18 * -2 + 1 * -1, 37 * 2));
la partage: -1/74
Half the edge, on those bets only.
Same mean, very different spread
If every bet has the same expectation, the choice of bet changes only the spread. The standard deviation of one spin, per unit staked:
// Standard deviation of the net result of one spin, per unit staked.
function sdOf(covers, payout, pockets) {
const q = covers / pockets;
const mean = q * payout - (1 - q);
return Math.sqrt(q * (payout - mean) ** 2 + (1 - q) * (-1 - mean) ** 2);
}
console.log("bet sd of 1 spin sd of 1,000 spins");
for (const [name, c, p] of [bets[0], bets[3], bets[5], bets[7]]) {
const sd = sdOf(c, p, 37);
console.log(name.padEnd(10), sd.toFixed(4).padStart(13), (sd * Math.sqrt(1000)).toFixed(1).padStart(18));
}
console.log("expected net after 1,000 spins:", ((1000 * -1) / 37).toFixed(2));
bet sd of 1 spin sd of 1,000 spins
straight 5.8378 184.6
corner 2.7947 88.4
dozen 1.4044 44.4
red/black 0.9996 31.6
expected net after 1,000 spins: -27.03
After 1,000 one-unit spins the expected net is −27.03 units on any of these bets. On red, the standard deviation of the total is 31.6 units, so most results land within a few dozen units of −27. On a single number it's 184.6, so results scatter over hundreds of units in either direction.
That's also why a simulation needs a lot of spins before the −1/37 shows through. A million spins of each bet, each with its own seed:
// Small seeded PRNG (mulberry32), so every run prints the same numbers.
function mulberry32(seed) {
return function () {
seed |= 0;
seed = (seed + 0x6d2b79f5) | 0;
let t = Math.imul(seed ^ (seed >>> 15), 1 | seed);
t = (t + Math.imul(t ^ (t >>> 7), 61 | t)) ^ t;
return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
};
}
const spins = 1_000_000;
const exact = -1 / 37;
console.log(`exact mean ${exact.toFixed(5)}`);
console.log("bet simulated standard error distance");
[bets[0], bets[3], bets[5], bets[7]].forEach(([name, c, p], i) => {
const random = mulberry32(i + 1); // an independent stream for each bet
let net = 0;
// Which c pockets a bet covers doesn't matter; call them pockets 0 … c−1.
for (let s = 0; s < spins; s++) net += Math.floor(random() * 37) < c ? p : -1;
const mean = net / spins;
const se = sdOf(c, p, 37) / Math.sqrt(spins);
console.log(
name.padEnd(10),
mean.toFixed(5).padStart(9),
se.toFixed(5).padStart(15),
`${((mean - exact) / se).toFixed(2)} SE`.padStart(9),
);
});
exact mean -0.02703
bet simulated standard error distance
straight -0.03293 0.00584 -1.01 SE
corner -0.02854 0.00279 -0.54 SE
dozen -0.02756 0.00140 -0.38 SE
red/black -0.02516 0.00100 1.86 SE
All four estimates are within two standard errors of −1/37. But on the single number, a million spins pin the mean down only to about ±0.0117 (two standard errors), which is more than 40% of the size of the edge itself.
Combining bets doesn't help
Expectation is linear. If you place several bets on one spin, the expected net is the sum of their expectations: −1/37 of the total staked. The same holds across spins even when each stake depends on earlier results. Every individual bet has conditional expectation −1/37 of its stake, so the expected total is −1/37 of the expected amount wagered (as long as that amount is finite). No staking system changes that ratio; it only changes how much gets wagered and how the results are spread.
If you want to try other bets, stakes and wheels without running the script, our small team built a roulette odds and payout calculator that shows each bet's chance, payout and expected value, plus streak odds, for European, American and triple-zero wheels.
Takeaways
- Expected value per unit is (c · (p + 1) − pockets) / pockets.
- Standard payouts make c · (p + 1) = 36, so every one of those bets on a 37-pocket wheel is −1/37.
- The American basket breaks the pattern (5 · 7 = 35) and is worse, −3/38.
- The bets differ in spread, not in mean, and it takes many samples to see the mean through the noise.
Top comments (0)