The coin keeps no score.

The gambler's fallacy is the conviction that an independent random process owes a correction — that after five reds, black is overdue; after a streak of heads, tails must be coming to restore balance. It feels like common sense, and it is wrong, because a fair coin and a fair wheel have no memory. Each spin is sealed off from every spin before it: the probability of heads is one half on the first toss and one half on the thousandth, whatever happened in between.

What misleads us is a half-remembered truth. The law of large numbers really does promise that the proportion of heads approaches one half as the flips pile up. But people quietly upgrade that into a promise that the coin will actively repay its debts in the short run — that a deficit of tails will be made good soon. It won't. As the hero shows, the proportion converges precisely because the growing pile of new, fair flips dilutes early imbalances, not because the coin ever tilts to cancel them. The gap between heads and tails typically grows; it just grows slower than the total.

The single most useful picture in this whole subject is the divergence between two curves that people merge into one. Track a fair coin for a couple of thousand flips and plot both the proportion of heads and the raw surplus of heads over tails.

On the 18th of August 1913, the roulette wheel at the Monte Carlo Casino landed on black twenty-six times in a row. As the streak lengthened, players crowded the table betting fortunes on red, certain that such a run could not continue — that red was overwhelmingly due. They were feeding the red curve, and it never came. The casino took millions, and the gambler's fallacy earned its other name: the Monte Carlo fallacy.

Both fallacies feed on the same defect: our mental model of randomness is too tidy. Ask people to write down a “random” sequence of coin flips and they alternate far too much, sprinkling in the odd repeat but instinctively breaking up any long run. Genuine randomness has no such manners.

This is the soil both fallacies grow in. The gambler sees a long run and feels the universe must correct it. The fan sees a long run and feels a force must be sustaining it. Neither has noticed that runs of that length are simply what fair processes produce all the time. In a couple of hundred coin flips, a streak of seven or eight is entirely ordinary.

The gambler's fallacy has a mirror image. The hot-hand fallacy is the belief that success breeds success — that a shooter who has just made several baskets is genuinely more likely to make the next. In 1985 a celebrated study by Gilovich, Vallone and Tversky examined real basketball data, found that a made shot was not followed by a higher success rate, and concluded the hot hand was a cognitive illusion: fans pattern-matching on noise. For thirty years this was the textbook verdict, a triumph of cool statistics over warm belief.

Then, in work published around 2018, Joshua Miller and Adam Sanjurjo found that the famous study had itself fallen into a statistical trap — a beautifully hidden one. Suppose you take a finite sequence of coin flips, find every flip that immediately follows a run of heads, and measure how often those flips are heads. You would expect 50%. It isn't.

The mechanism is pure selection. In a finite sequence, conditioning on “the previous flips were heads” changes what's left to be sampled, nudging the next flip's expected rate below one half — a cousin of regression to the mean. Because the original analysis compared real shooters against an unbiased 50% benchmark instead of this lower, correct one, it understated the hot hand. Reanalyses that fix the bias find the effect is modest but real. The people debunking one statistical illusion had been caught by another — which is the most this site could ever hope for from a single example.

The lottery of “due” numbers. The gambler's fallacy is quietly industrial. Lottery players favour numbers that haven't come up “in a while”; roulette boards helpfully display recent results to nudge exactly this instinct; slot machines feel “ready to pay.” Every one of these leans on a memory the machine does not have.

The two errors are siblings. Notice that the gambler and the hot-hand believer make opposite predictions — reversal versus continuation — from the same raw material, a streak. What unites them is a refusal to first ask whether the underlying process even has a memory. A roulette wheel does not; a fatiguing athlete, a warming engine, a market with momentum might. The error isn't believing in streaks or doubting them; it's not checking which world you're in.

And the benchmark can lie too. The hot-hand saga adds a second, subtler warning, the one this compendium keeps returning to: even when you correctly treat events as random, the way you select which events to measure can bias the very baseline you compare against. “No better than chance” is only meaningful if you've worked out what chance actually predicts under your selection rule.

Get the first part wrong and you bet on red at Monte Carlo. Get the second part wrong and you spend thirty years sure the hot hand is a myth. The honest move is the same as everywhere else in the compendium: pin down the process that generated the numbers before you trust the story they seem to tell.