Why do only 23 people give a 50% chance of a shared birthday?

Because you count pairs, not people. Twenty-three people form 253 distinct pairs, each with a 1/365 chance of matching, so the expected number of matches is 253/365 ≈ 0.69. Working the exact complement gives a match probability of 50.7% at 23 people — just over even odds.

How many people are needed to be almost certain of a shared birthday?

The curve climbs fast once you pass 23. The probability of at least one shared birthday reaches 97.0% at 50 people, 99.0% at 57 people, and 99.92% at 70 people. It never hits 100% until you exceed 365 people, but it is effectively certain long before that.

What is the chance someone shares MY specific birthday?

Much lower. With 23 other people the chance that at least one matches your particular birthday is only 1 − (364/365)^23 ≈ 6.1%. To reach even odds on your specific date you need about 253 people, because now there is only one target birthday instead of 253 pairs.

Quant interview prep guides

The birthday paradox, for quant interviews

Twenty-three people, even odds of a shared birthday. The number surprises everyone once. Interviewers want to see whether you can rebuild it — cleanly, from the complement — and whether you know why it isn’t the question people think it is.

For candidates practicing complements, approximation, and collision probabilities · Updated July 21, 2026

What the question is testing

“In a room of 23 people, what’s the chance two share a birthday?” The interviewer is not impressed that you remember the answer is about a half. They are watching for three moves: that you attack the complement instead of the messy union, that you understand the count that actually drives the surprise, and that you can approximate the answer without a calculator. Get those and the exact number falls out; miss them and you’re stuck adding up overlapping events.

Fix the model first: $365$ equally likely birthdays, people independent, no leap days. State those assumptions out loud — they are the parts an interviewer can perturb later.

Work the complement

Computing “at least one shared birthday” head-on means a union over every pair — inclusion–exclusion hell. The complement is a clean product. Seat people one at a time; each new person must dodge all the birthdays already taken:

\Prob(\text{no match}) = \prod_{k=1}^{n-1}\left(1 - \frac{k}{365}\right) = \frac{365!}{365^{\,n}\,(365-n)!}

Then $\Prob(\text{match}) = 1 - \Prob(\text{no match})$ . Plug in $n = 23$ and the product lands at $0.4927$ , so the match probability is 50.7%. One person short, at $n = 22$ , it’s 47.6% — under even odds. Twenty-three is genuinely the smallest group that tips past a half.

Each column is the probability that everyone’s birthday is distinct. It slides under 50% for the first time at 23 people — the moment a shared birthday becomes more likely than not. n = 23 → 49.3% distinct → 50.7% match.

Why so few? Count the pairs

The whole “paradox” is a miscount. People imagine 23 birthdays racing against 365 days and expect a tiny number. But a match is about pairs of people, and pairs grow quadratically:

\binom{n}{2} = \frac{n(n-1)}{2} \;\Rightarrow\; \binom{23}{2} = 253

Two hundred fifty-three pairs, each a $1/365$ chance of matching. The expected number of matching pairs is $253/365 = 0.6931$ — not coincidentally, that’s $\ln 2$ . Once you have on the order of one expected collision, “at least one” is a coin flip. The people scale linearly; the opportunities scale like $n^2$ .

Left: the pair count

\binom{n}{2}

grows quadratically — 23 people already make 253 pairs. Right: at 23 people, the chance some pair matches (50.7%) dwarfs the chance someone matches your birthday (6.1%). Different question, different denominator.

The approximation to keep in your head

On a whiteboard you don’t evaluate a 23-term product. Treat the pair matches as rare and roughly independent — a Poisson count with mean $\lambda = \binom{n}{2}/365$ — so:

\Prob(\text{match}) \approx 1 - e^{-\binom{n}{2}/365} \approx 1 - e^{-n^2 / (2\cdot 365)}

At $n = 23$ , the exponent is $-253/365 = -\ln 2$ , giving exactly $1 - \tfrac12 = 0.500$ — the approximation nails the real 50.7%. Inverting it gives the rule worth memorizing: to reach even odds among $d$ equally likely days you need about $n \approx 1.177\sqrt{d}$ , where $1.177 = \sqrt{2\ln 2}$ . For $d = 365$ that’s $1.177\sqrt{365} \approx 22.5$ , i.e. 23. The same $\sqrt{d}$ scaling is exactly the hash-collision / birthday-attack estimate interviewers in crypto and systems roles are really after.

The same numbers as a curve. It crosses 50% at 23, then flattens into near-certainty: 97.0% at 50, 99.0% at 57, 99.92% at 70. Steep where it matters, then boring — the signature of a probability driven by a growing pair count.

The trap: “shares my birthday”

A favourite follow-up: “so with 22 other people, is there a good chance someone shares your birthday?” This is a different question, and the honest answer is no. Now there is a single target date, not 253 free-floating pairs:

\Prob(\text{someone matches you}) = 1 - \left(\tfrac{364}{365}\right)^{n}

With $n = 23$ that’s $1 - (364/365)^{23} \approx 6.1\%$ . To get even odds on your specific birthday you need roughly $365\ln 2 \approx 253$ people. The gap between 6.1% and 50.7% at the same group size is the entire lesson: the surprise comes from any pair matching, not a fixed one.

Assumptions and variants

Non-uniform birthdays. Real birthdays cluster (September peaks, almost no Feb 29s). Uniformity actually minimizes the collision probability, so any real-world skew only raises the match chance — meaning 23 is a conservative, upper-bound count for reaching 50%. You can assert this without recomputing.
General $d$ and hash collisions. Swap 365 for any $d$ and the even-odds threshold is $\approx 1.177\sqrt{d}$ . This is why a 64-bit hash collides around $2^{32}$ items, not $2^{64}$ .
Triple birthdays. If they ask for three people sharing a day, the threshold jumps to about $88$ people. Flag it as a known result rather than deriving it live — the naive Poisson shortcut understates it, and the whiteboard version isn’t worth the time.

Common mistakes

Adding pair probabilities. “253 pairs times $1/365$ ” gives 0.69, which people quote as 69%. That’s an expected count, not a probability — at $n = 30$ the same sum is $435/365 = 1.19 > 1$ , which should end the argument. Use it inside $1 - e^{-\lambda}$ instead.
Answering the wrong question. Confusing “any two match” (50.7%) with “someone matches me” (6.1%). Always pin down whether the target birthday is fixed.
Fighting the union head-on. Inclusion–exclusion over 253 events is correct and hopeless. Reaching for the complement first is the signal the interviewer is grading.

Practice the pattern

The birthday problem is the canonical “work the complement, then approximate” question — and the $\sqrt{d}$ collision scaling shows up everywhere from hashing to matching trades. Drill complements, Poisson approximations, and collision estimates in the LeetQuidity curriculum, or take the free calibration to see where your reasoning breaks under time.

Take the calibration View the curriculum

Quant interview prep guides

The birthday paradox, for quant interviews

For candidates practicing complements, approximation, and collision probabilities · Updated July 21, 2026

What the question is testing

Fix the model first: $365$ equally likely birthdays, people independent, no leap days. State those assumptions out loud — they are the parts an interviewer can perturb later.

Work the complement

\Prob(\text{no match}) = \prod_{k=1}^{n-1}\left(1 - \frac{k}{365}\right) = \frac{365!}{365^{\,n}\,(365-n)!}

Why so few? Count the pairs

The whole “paradox” is a miscount. People imagine 23 birthdays racing against 365 days and expect a tiny number. But a match is about pairs of people, and pairs grow quadratically:

\binom{n}{2} = \frac{n(n-1)}{2} \;\Rightarrow\; \binom{23}{2} = 253

Left: the pair count

\binom{n}{2}

The approximation to keep in your head

On a whiteboard you don’t evaluate a 23-term product. Treat the pair matches as rare and roughly independent — a Poisson count with mean $\lambda = \binom{n}{2}/365$ — so:

\Prob(\text{match}) \approx 1 - e^{-\binom{n}{2}/365} \approx 1 - e^{-n^2 / (2\cdot 365)}

The trap: “shares my birthday”

\Prob(\text{someone matches you}) = 1 - \left(\tfrac{364}{365}\right)^{n}

Assumptions and variants

Non-uniform birthdays. Real birthdays cluster (September peaks, almost no Feb 29s). Uniformity actually minimizes the collision probability, so any real-world skew only raises the match chance — meaning 23 is a conservative, upper-bound count for reaching 50%. You can assert this without recomputing.
General $d$ and hash collisions. Swap 365 for any $d$ and the even-odds threshold is $\approx 1.177\sqrt{d}$ . This is why a 64-bit hash collides around $2^{32}$ items, not $2^{64}$ .
Triple birthdays. If they ask for three people sharing a day, the threshold jumps to about $88$ people. Flag it as a known result rather than deriving it live — the naive Poisson shortcut understates it, and the whiteboard version isn’t worth the time.

Common mistakes

Adding pair probabilities. “253 pairs times $1/365$ ” gives 0.69, which people quote as 69%. That’s an expected count, not a probability — at $n = 30$ the same sum is $435/365 = 1.19 > 1$ , which should end the argument. Use it inside $1 - e^{-\lambda}$ instead.
Answering the wrong question. Confusing “any two match” (50.7%) with “someone matches me” (6.1%). Always pin down whether the target birthday is fixed.
Fighting the union head-on. Inclusion–exclusion over 253 events is correct and hopeless. Reaching for the complement first is the signal the interviewer is grading.

Practice the pattern

Take the calibration View the curriculum