Coin Flip Streaks and What They Mean

The log-2 rule

For n fair coin flips, the expected longest run of heads is approximately log base 2 of n. It is an approximation and a good one, typically accurate to within about one for any realistic length.

The intuition behind it: a run of length k has probability 1 in 2 to the power k of starting at any given position, and there are roughly n positions, so runs of length around log₂(n) are the longest you expect to see once.

The consequence is that streaks grow logarithmically. Doubling the flips adds one to the expected longest run — from 100 to 200 flips, expected longest goes from about 6.6 to about 7.6.

Why intuition is so far off

People asked to write down a fake sequence of coin flips reliably produce too few long runs. The typical fabricated sequence has a longest run of three or four where a genuine hundred-flip sequence has six or seven, and this is consistent enough that it is used as a teaching exercise.

The reason appears to be that people model randomness as local balance — they alternate to keep the running count near even, because a run feels like it is going wrong. Genuine randomness has no such tendency, so it produces runs constantly.

The practical consequence is that a run which feels impossible is usually ordinary. A student picked three lessons running out of thirty is a specific and calculable event, not evidence of a rigged picker, and the calculation is usually reassuring once someone does it.

Variance between sessions is large

The log-2 figure is an average, and individual batches vary widely around it. Ten batches of two hundred flips will typically produce longest runs from about five to about eleven, with a mean near seven and a half.

That spread is why judging a generator from a single batch is meaningless. A batch showing a run of eleven is not evidence of anything; a batch showing a run of four is not evidence of anything either.

Distinguishing a genuinely biased source from an unbiased one on streak evidence alone takes many batches, and even then a frequency count is a far more sensitive test than a streak length.

• Ten batches of 200 flips: longest runs typically range from 5 to 11.
• The mean tracks log₂(200) ≈ 7.6 closely; any individual batch may not.
• A single batch cannot distinguish a fair source from a biased one.
• For detecting bias, a frequency count is far more sensitive than a streak measurement.

Runs in a picker, not a coin

The same arithmetic applies to any repeated draw, with the base changed. For a class of thirty, the chance a specific student is picked twice in a row is 1 in 900, which sounds remote — but the chance that some student is picked twice in a row somewhere across a term of sixty lessons is close to certain.

That distinction between a specific event and any event is where most intuitions about randomness fail. The specific coincidence is rare; the class of coincidences is large, and something from it happens almost every time.

It is also why a no-repeat picker exists. Not because the independent draw is unfair, but because "picked twice running" is a thing people notice, remember and complain about, and removing it removes the complaint.

What a genuinely suspicious result looks like

Streaks are a weak test. If you actually want to check whether a source is fair, count frequencies rather than runs: tally the outcomes over several thousand draws and compare against what a uniform source would give.

A useful rule of thumb is that the typical deviation in a count of n draws is about the square root of n. Two hundred flips has a typical deviation of about seven heads, so anything from 93 to 107 is unremarkable and 85 would be worth a second look.

Ten thousand flips has a typical deviation of about fifty, so 5,150 heads is ordinary and 5,400 is not. Scaling that rule to whatever you are testing is the fastest available sanity check.

Frequently asked questions

Tools that use this

Related guides