Some people who listened to music on Spotify, the streaming service, used to complain that its shuffle button wasn’t random. They would press shuffle on a playlist and hear two or three songs by the same singer one after another. “Why isn’t your shuffling random?” they asked.
The engineers looked again. The shuffle was perfectly random, made by a method that gives every possible order exactly the same chance. Two songs by one singer side by side was simply something a random order does quite often. So the company did a strange thing. It made the shuffle less random, spreading each singer’s songs out evenly along the playlist, because that was what listeners thought random looked like. “We learned that they don’t like perfect randomness,” one of the engineers wrote in 2014.
The listeners weren’t foolish. They were doing what everyone does. Ask people to picture chance, and they picture things evenly mixed: heads, tails, heads, tails, with only the occasional pair. Real chance isn’t like that. It comes in lumps and streaks and empty patches, and it does so all the time, with nothing behind it. That is the one idea of this chapter, and it explains a great deal. Things that happen by chance look as if they were arranged. A rare thing becomes a common thing when enough people are around for it to happen to. And a coin, however long its streak, has no idea what it did last time.
Chance comes in lumps
The mathematician Theodore Hill used to start his probability class at Georgia Tech, in the United States, with a piece of homework. Students whose mother’s maiden name began with A to L were to toss a coin 200 times and write down the results. The rest were to make up 200 tosses and write those down instead. The next day he collected the sheets and sorted the real ones from the fakes, and he said he was right about 95 times in 100.
His trick was simple. Somewhere in 200 real tosses there is nearly always a run of six heads or six tails in a row. Work it out and the chance is about 97 in 100. Yet a person making up tosses almost never writes six of the same in a row, because it doesn’t look random. People switch too often. Across many studies, people inventing random sequences switch from heads to tails, or back, about 6 times in every 10 tosses. A real coin switches about 5 times in 10, which leaves room for long runs.
The made-up tosses look tidier, and that is how to spot them. The coin keeps making runs of six, seven, sometimes nine, and about four times in five, 100 real tosses have a run of six or more somewhere. People’s fakes stop at four or five. The lumps that make real chance look suspicious are exactly what makes it real.
The same thing happens in space as well as in time. In the summer of 1944 Germany began firing flying bombs at London. These were the V-1s, small pilotless planes carrying a warhead. A counter in the nose told each one when to dive, and as it dived its engine cut out. They couldn’t be aimed at a street: at first they landed anywhere in a circle about 30 kilometers across. Even so, it was often said in London that the bombs fell in clusters, some places hit again and again while others nearby were spared.
An insurance mathematician, R. D. Clarke, made the check. He took an area of south London of 144 square kilometers, cut it on the map into 576 small squares of a quarter of a square kilometer each, and counted how many bombs had fallen in each square. There were 537 bombs in all, a little under one per square. If the bombs were aimed, or the clusters meant something, some squares should have had far too many and others far too few. If they fell by chance, there is a formula that says how many squares should get none, one, two and so on.
The counts fit the formula almost exactly. Chance said about 227 squares should have no bomb at all, and 229 had none. It said about 31 should have three, and 35 did. Seven squares took four bombs each, and one took five or more, which is just what chance predicted. The clusters were real, in the sense that some neighborhoods really were hit far harder than others. But they were the clusters that bombs falling anywhere at all would make.
Look at a picture like that and the eye finds patterns at once: a crowded corner, an empty stripe. Nothing put them there. So a lump on its own, three cases of an illness on one street, four accidents at one bend in a month, isn’t evidence that something caused it. Something may have. But chance makes lumps too, and the fair question is whether there are more of them, or bigger ones, than chance alone would make. Clarke’s answer for London was no.
One in a million, every day
In February 1986 The New York Times reported on its front page that Evelyn Marie Adams, who managed a shop in New Jersey, had won the state lottery for the second time in four months. The chance of that, the story said, was one in 17 trillion.
Two statisticians, Persi Diaconis and Frederick Mosteller, later called that figure “the correct answer to a not-very-relevant question”. It is the chance that one particular person, buying one ticket in each of two particular draws, wins both. But Ms. Adams bought many tickets, many times. And the real surprise wasn’t that it happened to her. Nobody had picked her out in advance. Somebody somewhere winning twice would have made the same story. Two other statisticians, at Purdue University, worked out the chance that someone, among all the millions of Americans who buy lottery tickets, wins twice. They found it was better than even over seven years, and better than one in 30 over any four months. In their words, practically a sure thing.
Diaconis and Mosteller gave the idea a name: the law of truly large numbers. With a large enough number of people, any outrageous thing is likely to happen. Their sum was this. Call something a miracle if it happens to one person in a million on a given day. The United States then had about 250 million people, so 250 such miracles would happen every day there, close to 100,000 a year.
Nepal has about 30 million people. So a one-in-a-million coincidence happens to about 30 Nepalis every day, and to around 11,000 of them in a year. Each of those 11,000 will feel, quite rightly, that something extraordinary has happened. Some will tell their families, and a few of the stories will spread. Nobody passes on the millions of days on which nothing odd happened.
The everyday coincidences work the same way, on a smaller scale. You think of a cousin you haven’t spoken to in months, and that evening she calls. It feels like more than chance. But count the times you think of someone in a week, and the calls you get, and some of them are bound to land together now and then. The times they do are the ones you remember. The hundred times you thought of someone who didn’t call never felt like anything, so they were never counted. And the coincidence you noticed was one of a thousand that would have struck you just as much: the same cousin at a wedding, a stranger with your grandfather’s name, a bus ticket with your birthday on it. You weren’t waiting for that one. You were waiting for any of them.
Birthdays make the point neatly. In a class of 40 students, the chance that two share a birthday is about 9 in 10. It sounds too high until you notice that it isn’t asking whether someone shares your birthday. It asks about every pair in the room, and 40 people make 780 pairs. Math, Going Further works it out in Will two people in your class share a birthday?.
The coin has no memory
The Spotify engineers explained their listeners’ mistake with lunch. Suppose you toss a coin every day at work to decide between Thai food and Indian, and it says Thai four days running. It feels as if Friday must be Indian’s turn. It isn’t. The coin has no memory and doesn’t know what it said on Monday. Its fifth toss is a fresh toss, with the same even chance as the first. Believing that a run must soon be balanced out has a name, the gambler’s fallacy, because it has emptied so many gamblers’ pockets.
It emptied a good many in Italy in 2005. Italy’s Lotto then drew five numbers out of 90 on each of ten “wheels”, named after cities, and players could bet on a number coming up on a particular wheel. In May 2003 the number 53 came up on the Venice wheel, and then it didn’t come up again. Draw after draw it stayed away, and people became sure it was due. The longer it was missing, the more they bet. By one newspaper’s count, Italians bet about 3.2 billion euros on 53 in eighteen months. Newspapers reported debts, ruin and even deaths linked to the craze.
At each draw, the chance of 53 coming up on Venice was 5 in 90, about one in 18, and it stayed one in 18 however long the number had been missing. The balls in the drum didn’t know it was late. It finally came up on 9 February 2005, after 182 draws. A gap that long is very unlikely for any one number on any one wheel. But there were 90 numbers on each of ten wheels, drawn week after week for years, so somewhere, sometime, a long gap was bound to turn up. It was the law of truly large numbers again, and it happened to land on 53.
This is the coin with no memory, and it holds for coins, dice, lottery drums and shuffled songs, things built so that one go can’t affect the next. A great deal of life isn’t built like that. Tomorrow’s weather does depend on today’s. A student who has worked hard all term is more likely to do well in the next test, not less. And if a coin comes up heads twenty times in a row, the sensible thought isn’t that tails is due, or that heads is on a streak. It’s that this may not be a fair coin. Whether something has a memory is a question about how it works, and sometimes the honest answer is that nobody is sure. Which brings this chapter to its surprise.
The hot hand
Basketball players talk about being “hot”: making a few shots in a row and feeling that the next one will go in too. Coaches believe it, players believe it, and fans believe it. In 1985 three psychologists, Thomas Gilovich, Robert Vallone and Amos Tversky, set out to see whether it was real.
They studied the shooting records of two professional teams, and they ran a careful experiment of their own with 26 players from the men’s and women’s teams at Cornell University. Each player was to take 100 shots from a distance at which they made about half. If players got hot, they should have hit more often just after a run of hits than just after a run of misses. On average the players made 49% of their shots after three hits in a row and 45% after three misses in a row. A gap of four points was small enough to put down to chance. The three psychologists concluded that the hot hand was an illusion, people seeing streaks in chance just as the Spotify listeners did, and they called the belief “a powerful and widely shared cognitive illusion”.
For about thirty years this was one of the best-known findings in psychology, the standard example in books about the mind of people finding patterns where there are none. The psychologist Daniel Kahneman, in his 2011 book Thinking, Fast and Slow, wrote that “the hot hand is a massive and widespread cognitive illusion.” Players and coaches kept rejecting the finding, and that was usually taken as a case in point.
Then two economists, Joshua Miller and Adam Sanjurjo, found a mistake, not in the players but in the arithmetic. It is subtle enough that it got past nearly everyone, and simple enough to check with three tosses of a coin.
Toss a coin three times, and look only at the tosses that come straight after a heads. What share of those tosses would you expect to be heads? Half, surely, since the coin has no memory. Now try it. Three tosses can come out eight ways, each as likely as the others. In two of them, TTH and TTT, no toss follows a heads, so there is nothing to look at. In the other six, the share of heads among the tosses after a heads comes out as 1, ½, 0, 0, 1 and 0. The average of those six is 5 in 12, about 42%, not half.
Nothing is wrong with the coin. The bias comes from the way the counting is done: picking out the tosses that follow a streak, within a short run of tosses, and then averaging over players or sequences. In a short sequence, a streak of heads has used up some of the heads, so what follows it leans toward tails. The effect shrinks as the sequence gets longer, but it doesn’t vanish at 100 shots. For a player who makes exactly half his shots, with no hot hand at all, Miller and Sanjurjo worked out that the method the psychologists used should show him about 8 points worse after three hits than after three misses.
The Cornell players weren’t 8 points worse. They were 4 points better. So in the very numbers that had proved there was no hot hand, the gap between shooting after hits and shooting after misses was about 12 points wider than a player with no hot hand would show. Miller and Sanjurjo published this in 2018. Their conclusion was that the old result was reversed.
Is the hot hand real, then? In shooting contests, with nobody defending, later studies have found signs of it. In real games it is harder to tell, because a player who seems hot gets guarded more closely and takes harder shots, and the studies disagree about how big the effect is there. What is no longer in doubt is the slip in the arithmetic. For decades, experts told the public that people see streaks in chance that aren’t there, and they were right about that in general. But their own picture of what a short run of chance should look like was wrong, in exactly the direction this chapter has been about.
None of this rescues the gambler’s fallacy. The lottery drum still has no memory, and the Spotify listeners were still wrong about their shuffle. What it shows is how slippery chance is. It fools fans and gamblers, and now and then it fools the people who study how chance fools us. The fix was the same as always: someone checked the sum again.
Found a mistake? Please tell us.
