Image: Michael Barera, CC BY-SA 4.0A coin is not quite fair, and we know why
A coin toss is the textbook example of a fair chance, the thing with probability one half that every probability course starts from. It is not quite that. A coin caught in the hand lands on the same side it started on slightly more often than half the time, about 51 per cent, and this is known not from folklore but from a physical theory published in 2007 and an experiment of 350,757 tosses published in 2023.
The physics
Persi Diaconis, Susan Holmes, and Richard Montgomery, two statisticians and a mathematician at Stanford University and the University of California, Santa Cruz, set out to describe what a coin actually does in the air. A tossed coin spins about an axis, and the common picture is that the axis lies flat in the plane of the coin, so that heads and tails alternate cleanly as it turns. Real thumbs do not do that. The axis of rotation is almost always tilted a little out of the plane of the coin, and a coin spinning about a tilted axis wobbles, or precesses, so that the side that was up at the start spends slightly more than half of each turn facing up.
They worked out the equations of this motion and measured real tosses with a high-speed camera and with a coin fitted with a ribbon to record its rotations. The result was a prediction: for a coin tossed and caught, the probability that it lands with the same side up as it started is about 0.51. Heads against tails, averaged over coins that start randomly, stays at 0.50. The bias is in the start, not in the metal. The paper, Dynamical Bias in the Coin Toss, appeared in SIAM Review and is free to read.
One percentage point is a small effect, and the authors said that to detect it with confidence would take about 250,000 tosses, which was more than anyone had done. The largest previous dataset was the 10,000 tosses that John Kerrich made while interned in Denmark during the Second World War, in which he recorded 5,067 heads, and Kerrich had not recorded which side started up.
The experiment
In 2023 a group of 48 people, led by František Bartoš, then a doctoral student at the University of Amsterdam, did the tosses. They used coins of 46 currencies and denominations, recorded the starting side and the landing side of every toss, and filmed themselves to allow checking. The total came to 350,757 tosses.
The coin landed on the same side it started on in 50.8 per cent of them. The 95 per cent confidence interval runs from about 50.6 to 50.9 per cent, which excludes one half comfortably and sits close to the predicted 0.51. Heads landed up in 50.0 per cent of tosses, as the theory said it should once the starting side is averaged out. The preprint has the data by person and by coin.
The variation between people is the most interesting part of the data. Some tossers showed almost no same-side bias; a few showed a much larger one, approaching 60 per cent, presumably because their tosses wobbled more. The bias is a property of a thumb, not of a coin, and it was the same for a Czech crown and a United States quarter.
What it changes
For a decision at home, nothing, as long as nobody looks. If the person calling the toss does not know which side is up when the coin leaves the thumb, the 51 per cent cannot be used, and heads or tails is as even as it ever was. If they do look, the advantage is real but small: over 1,000 tosses, betting on the starting side wins about 508 times instead of 500. Bartoš and colleagues point out that this is about the size of the casino's edge in some games, and that casinos make a living from edges of that size. The remedy is simple: hide the starting side, or let the coin fall on the floor, which adds a bounce the theory does not cover.
For the statistician the interest is elsewhere. The coin toss is the model for the idea that probability is a stable long-run frequency, and here is a case in which the frequency is stable, measurable, and not the number everyone assumed. The probability of one half was never a law of nature. It was a good approximation that came from a symmetry argument, heads and tails look alike to the physics, and the symmetry argument is correct about heads against tails and silently wrong about start against finish, because the toss breaks that symmetry. Unit 3 of the Statistics Masterclass uses the experiment to ask what the probability of a coin toss is, before it uses the coin for anything else.
How the question was settled
Notice how the knowledge was produced. A theory made a quantitative prediction, 0.51, and stated how much data would be needed to test it. Sixteen years later a group collected that much data, in public, with the procedure written down in advance and the videos kept. The result matched the prediction and the confidence interval said how closely. This is a good small example of how an empirical question is settled, and the number of tosses was the price of settling a one-point difference: the standard error of a proportion near one half from \(n\) tosses is \(0.5/\sqrt{n}\), which for 350,757 tosses is 0.08 of a percentage point, small enough to separate 50.8 from 50.0 by ten standard errors. A hundred tosses, with a standard error of 5 points, could never have seen it.
That is the arithmetic of Unit 5, and the reason polls need a thousand people and a one-point effect needs a third of a million coin tosses. Diaconis, who is the research spotlight of Unit 3, has also worked out how many times a deck must be shuffled, with the same mixture of physics, mathematics, and a magician's eye for what hands actually do.


