If Two People Flip Until One Gets 3 Heads First, Who’s Actually Favored?
When two players each flip their own fair coin once per round, racing to see who accumulates 3 heads first, neither player is a true 50/50 favorite: an exact tie — both reaching their third head on the very same round — happens 13.58% of the time, leaving each player only a 43.21% chance of outright winning, not 50%.
A worked example with numbers
Model the race as two independent sequences of coin flips, one per player, where each player’s “finishing round” is the round on which their third head appears — a value following the negative binomial distribution with r = 3 successes and p = 0.5 per flip. Because both players draw from the identical distribution, the probability that player A finishes strictly before player B equals the probability that B finishes strictly before A, by symmetry; what remains is the probability that both finish on the exact same round, which is not automatically small just because it is a tie. Summing the squared per-round finishing probabilities across every possible finishing round gives a tie probability of exactly 13.580246913580243%, or almost precisely 11/81 as a fraction. With the tie probability fixed, each player’s outright win probability is (100% − 13.58%) ÷ 2 = 43.21%, exactly matching the earlier headline figure and confirming the three numbers add to 100%.
How to set the rule before the result
To verify this kind of race probability independently, either simulate it directly — run many thousands of paired trials, each flipping two virtual coins per round until one side reaches 3 heads, and tally wins, losses, and ties — or compute it exactly using the negative-binomial formula P(finish on round k) = C(k−1, 2) × 0.5^k for k = 3, 4, 5, and so on, then sum P(tie) = Σ P(k)² over every possible k and derive each player’s win share as (1 − P(tie)) ÷ 2. Both approaches should converge to the same 13.58% tie figure and 43.21% win figure; a simulation landing meaningfully outside that range after tens of thousands of trials indicates an error in how the race’s stopping rule was implemented, not a flaw in the underlying probability. As a further check, the three round-target probabilities in the table below should always sum to within rounding error of 100% for every race length — a quick arithmetic sanity check that catches a sign error or an off-by-one in the summation loop before trusting any single figure it produces.
Common mistakes that change the odds or the process
The most common mistake is assuming that because each individual flip is a fair 50/50 event, the overall race must also be exactly 50/50 between the two players — that reasoning skips over the possibility of an exact tie entirely, treating the race as if it must always produce a clean winner. A second mistake is assuming a tie is a vanishingly rare edge case not worth accounting for; at 13.58%, an exact tie in this race happens more often than most people’s intuition for “simultaneous tie” would suggest. A third mistake is assuming the tie probability stays the same regardless of the target number of heads — the tie probability actually shrinks as the target rises: races to 1, 2, 3, and 5 heads have tie probabilities of 33.33%, 18.52%, 13.58%, and 9.76% respectively, so a longer race to more heads is both more decisive and closer to a genuine 50/50 split between the two players, not further from it.
Where this method stops being appropriate
This model assumes both players flip simultaneously, once per round, with fully independent fair coins and no round in which one player could flip more times than the other; a variant where players take alternating single turns instead of simultaneous rounds is a different race structure with its own separate calculation, not interchangeable with the simultaneous-rounds model used here. It also assumes the “win” condition is reaching exactly 3 heads first, with a tie explicitly possible and left unresolved rather than broken by some tiebreaker rule; a version of the game with an explicit tiebreaker — such as an extra sudden-death round whenever both players finish simultaneously — changes each player’s true win probability back toward, but not exactly to, 50%, since the tiebreaker round itself still carries its own small further chance of a repeated tie.
How the random source fits into the rule
Wolfram MathWorld’s negative binomial distribution reference documents the negative binomial distribution’s formula used above for the round on which a player’s r-th success occurs, the same distribution this race’s finishing-round calculation is built on for both players independently.
Tie and win probabilities for races to different head counts
| Race target | P(exact tie) | P(each player wins outright) |
|---|---|---|
| 1 head | 33.33% | 33.33% |
| 2 heads | 18.52% | 40.74% |
| 3 heads | 13.58% | 43.21% |
| 5 heads | 9.76% | 45.12% |
Every row’s three-way split — tie, player A, player B — sums to 100% exactly, since a tie and one of the two players winning outright are the only mutually exclusive outcomes the race can end in. Racing to a single head is the most tie-prone version, essentially reducing to “did both players flip the same result on round one,” while racing to more heads gives the underlying variation more room to separate the two players before they can land on the exact same finishing round.
A second case: what happens when one player starts one flip ahead
Suppose player A has already recorded 1 head before the race officially starts, needing only 2 more, while player B starts from zero and needs the full 3. This is no longer a symmetric race, so the tie-probability shortcut used above no longer applies directly; instead, each player’s finishing-round distribution must be computed separately — A’s from a negative binomial with r = 2, B’s from r = 3 — and compared round by round. Working through the same summation approach with these two different distributions gives A a win probability of about 65.62%, B about 21.48%, and a tie probability of about 12.90% — a clear head start, but nowhere near a guaranteed win, since B can still catch up across enough additional rounds before A completes the remaining 2 heads.
Why a short observed race is weak evidence either way
Suppose two players run this race only once and player A wins. That single result carries almost no information about who the “better” player is — there is no such thing as coin-flipping skill, and a 43.21% true win probability means A is expected to win a minority of races run this way over the long run, not a rare one. Confusing a fair race’s built-in tie mechanism for an “unfair” or “broken” result is a version of the same reasoning error covered elsewhere on this site for repeated dice or coin outcomes: judging a single trial, or even a handful of trials, against a probability that only becomes visible over many repetitions.