How a cascade forms, and why it is rational. Suppose your own information points weakly one way, and you can see that several people before you chose the other. Each of those choices is evidence, because presumably each was made on some information. Once enough of them accumulate, the information you can infer from their actions outweighs your own single observation, and following them is the correct inference rather than a failure of nerve. The trouble is what happens next: because you followed the crowd instead of your own signal, your information never enters the public pool either. The next person sees one more choice and no more evidence. From that point the sequence can run indefinitely on the same small stock of genuine information, and every additional participant makes the crowd larger without making it better informed.
Two properties follow, and both are practically useful. The first is that a cascade can settle on the wrong answer, because only the earliest participants' information actually got in, and early participants are not selected for being right. The second is fragility: precisely because so little real information underlies the pattern, a modest piece of new public information can flip it. Bikhchandani, Hirshleifer, and Welch emphasize this, and it explains a familiar shape in markets, where a long, apparently stable consensus reverses faster than the change in fundamentals seems to justify. A crowd is not a heavy object; it is a thin one that looks heavy.
This is a different mechanism from the ones it gets confused with. Conformity, in the sense of not wanting to look foolish or to stand apart, is a social-pressure story and it exists, but it is not what these models describe. Extrapolating from a recent run of good outcomes is recency, a separate error about which evidence gets weight. Seeking out only supportive coverage is confirmation, an error in the search rather than in the inference. Herding is narrower and stranger: correct reasoning from an impoverished evidence base that the reasoning itself keeps impoverished.
A second and genuinely distinct channel is reputational. David Scharfstein and Jeremy Stein argued in "Herd Behavior and Investment," in the American Economic Review 80(3) in 1990, that a professional whose incentives depend on relative judgment has reason to follow consensus even when their own information disagrees, because being wrong alone is penalized more heavily than being wrong with everybody else. That mechanism has nothing to do with what anyone believes and everything to do with how they are evaluated, and it points the same direction. Their formal model was subsequently the subject of a published comment and reply in the same journal, so treat the channel as a well-known argument rather than a settled measurement.
Where it shows up in ordinary financial life. Fund flows follow performance, so the money most often arrives after the returns that attracted it. Asset classes acquire and lose respectability in ways that track visible participation more closely than any change in the underlying claim. And the question people actually ask, in a rising market or a falling one, is usually what other people are doing, which is exactly the observation the cascade model says is least informative once the crowd is large.
One honest limit on the concept. Herding is a mechanism, not a diagnosis. Observing that many people hold something does not establish that a cascade is operating, and a widely held view can be widely held because it is correct. What the model does is remove the comfort from the observation: agreement is informative only to the extent the agreeing parties reached their views independently, and visible agreement is evidence that they did not.
