Why Sensible People Stampede: Informational Cascades

Here is a comfortable thought that this paper takes away from you.

When you see a crowd all doing the same thing (all buying a stock, all queuing outside a restaurant, all convinced a technology is the future), you assume it means something. Surely all those people cannot be wrong. Even if each individual is only slightly informed, the aggregate of thousands of independent judgments should be reliable. That is the wisdom of crowds.

Bikhchandani, Hirshleifer and Welch showed that under very ordinary conditions, the crowd is not aggregating anything. The vast majority of the people in it have thrown their own information away, and the entire stampede may be resting on the private opinions of the first two or three people who happened to go first.

And, crucially, every single person in the crowd is behaving perfectly rationally. No biases required. That is what makes this paper so unsettling.

The problem: everyone can be rational and everyone can be wrong

The setup is a beautifully simple sequential decision problem.

People decide one at a time, in order. Each person has to make a binary choice: adopt or reject, buy or don't buy. Each person has:

A private signal. Their own information, which is imperfect but genuinely informative. Say it is right 60 percent of the time.
The observed actions of everyone before them. They can see what people did, but not why. They see the choices, not the signals behind them.

That second detail is the entire engine of the paper. You can see what people did. You cannot see what they knew. In real life this is almost always true: you see that the fund bought the stock, not the research that made them buy it. You see a queue, not the reviews each person read.

Now watch what happens.

The key idea via analogy: the two restaurants

Two restaurants sit side by side. Both are empty. Each is equally likely to be the better one, and every diner has done a little research, so each has a private hunch that is right more often than not.

Diner 1 arrives. She has no one to observe, so she follows her own signal. Her research says restaurant A. She goes in. Now A has one person, B has none.

Diner 2 arrives. His private signal says B. But he can see that someone chose A, and he can rationally infer that this person had a signal favouring A. So he now has two pieces of evidence of equal strength: his own signal (B) and the inferred signal of diner 1 (A). It is a tie. Say he breaks the tie by following his own hunch and goes to B. Fine. One and one.

Diner 3 arrives. Her signal says A. She sees one person in each. The evidence from the crowd is a wash, so she follows her own signal, and goes to A. Now it is A: 2, B: 1.

Diner 4 arrives. And here is where the trap springs shut.

His private signal says B. But he looks at the restaurants and sees two people in A and one in B. He rationally infers that the crowd is holding roughly two signals for A and one for B. His own single signal for B is not enough to overturn a two-to-one evidence advantage.

So the rational thing to do is ignore his own information entirely and go to A.

And now the catastrophe. Diner 5 arrives. She sees three in A and one in B. But notice: diner 4's choice contains no information whatsoever. He would have gone to A regardless of what his own signal said. His action is a pure echo of the crowd, and it added no new evidence to the pile.

But diner 5 cannot tell the difference. From the outside, a person who joined A because his signal said A and a person who joined A because he was copying the crowd look completely identical. So she updates as if diner 4's choice were informative, becomes even more confident in A, ignores her own signal, and joins A too.

And so does everyone after her, forever.

That is an informational cascade. From diner 4 onward, nobody's private information enters the public pool. Every subsequent person is copying, and every subsequent copy makes the herd look stronger, which makes the next person copy harder. The crowd grows enormous and its information content stays frozen at whatever the first three people happened to know.

Now the punchline: what if diners 1 and 3 were simply wrong? Their signals are only 60 percent accurate. Two unlucky draws is not unusual at all. The entire restaurant is now packed on the strength of two coin flips that came up badly, and hundreds of people with correct private information about restaurant B are sitting in restaurant A, having rationally discarded what they knew.

Two properties follow immediately, and both are important:

Cascades are fragile. Because the whole edifice rests on very little actual information, a single credible new piece of information (a famous critic reviews restaurant B, a short seller publishes a report) can shatter it instantly and flip the entire crowd. This explains why fads and fashions and market manias can reverse with stunning speed for no proportionate reason. There was never much holding them up.
Cascades are inefficient by construction. The crowd's aggregate information is tiny, even though it looks overwhelming. The wisdom of crowds requires independent judgments. A cascade destroys independence, and once independence is gone the crowd is not wise, it is just large.

Why it mattered

It explains herding without stupidity. This is the crucial contribution. You do not need investors to be irrational, panicky, or biased. Everyone in the cascade is doing exactly what a Bayesian statistician would do with the information available. Rational individual behavior produces collectively idiotic outcomes. That is a much stronger claim than "people are dumb", and much harder to argue with.
It gives bubbles a mechanism that survives scrutiny. Efficient market defenders will say a bubble requires irrational investors. This paper says no: it requires only that people can see each other's trades but not each other's reasons, which is a description of every market that has ever existed.
It explains the fragility of manias. The speed of a crash has always been puzzling under rational models. Cascade theory says the herd was always informationally hollow, so it takes very little to collapse it.
It has an uncomfortable warning for quants. If your signal is that a lot of smart money is buying something, you may be looking at genuine information, or you may be looking at the fourth diner. You cannot tell them apart from the outside, and that is not a limitation of your data, it is a theorem.

The honest limitations

The sequential, one-at-a-time structure is stylized. Real markets have people acting simultaneously, repeatedly, and with the chance to revise. Cascades are less clean in that world, though the intuition survives.
Binary choices and identical signal quality are strong assumptions. In the basic model everyone's signal is equally good and the choice is yes or no. If some people have much better information, or if people can express how strongly they believe (which in markets they can, by trading larger), cascades become harder to start, because a big confident trade reveals more than a small one. Prices, which aggregate the intensity of belief and not just its direction, are a genuine partial defence against cascades, and this is arguably the most important limitation for finance specifically.
It competes with a simultaneous rival. Banerjee published a very similar herding model in the same year, and there is a long-running question of how much of the credit belongs where. Both are excellent.
It is hard to test. In the field, you cannot see people's private signals, so proving that a real-world herd was a cascade (rather than a crowd that genuinely knew something) is close to impossible. The laboratory evidence is supportive, but the model's most dramatic claims are, by their nature, resistant to direct confirmation.

The one-line takeaway

Bikhchandani, Hirshleifer and Welch showed that when people decide in sequence and can see each other's actions but not each other's reasons, everyone will rationally start ignoring their own information and copying the crowd, so a stampede of thousands can rest entirely on the private hunches of the first two or three people, and can be shattered by a single credible whisper.