When Crowds Go Wrong: Bubbles, Cascades, and the Madness of Herds

The same crowd, two very different days

In 1906, a crowd of 787 fairgoers at an English livestock exhibition guessed the weight of an ox to within about one percent of the truth — the founding demonstration of the wisdom of crowds, which we tell in full in the story of Galton’s ox. Two and a half centuries earlier, crowds of equally ordinary Dutch traders bid the price of rare tulip bulbs to extraordinary heights before the market collapsed in a single February. Same species, same cognitive equipment. What flips the switch?

The difference is not intelligence, and it is not greed. It is structure. Galton’s fairgoers wrote their guesses privately on cards, each drawing on their own knowledge, with no view of anyone else’s number. The tulip traders watched each other bid. The moment judgments become sequential and observable, a crowd stops being a collection of independent measurements and becomes a chain of inferences about inferences — and chains like that can carry a society-sized error on almost no information at all. This article is about how that happens: the mathematics of cascades, the psychology of groupthink, the historical wreckage, and the structures that keep groups on the right side of the switch.

The machinery of a cascade

In 1992, Sushil Bikhchandani, David Hirshleifer, and Ivo Welch published “A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades” in the Journal of Political Economy — one of the most influential models in modern social science. Abhijit Banerjee independently published “A Simple Model of Herd Behavior” in the Quarterly Journal of Economics the same year. The setup is austere: people choose between two options, one at a time, in public. Each person holds a private signal — a noisy hint about which option is better — and can see every previous person’s choice, but not the signal behind it.

Watch the logic unfold. The first person follows her own signal — it is all she has. The second person sees the first choice and his own signal; if they agree, easy, and if they conflict, his signal at least keeps him at a coin flip. But the third person who sees two identical choices ahead of her faces arithmetic: two signals’ worth of evidence (inferred from the choices) against her one. Even if her private signal screams the opposite, the rational Bayesian move is to follow the crowd. And here is the poison: because she ignored her signal, her choice reveals nothing. Person four sees three identical choices, but only two of them ever contained information. The cascade is now self-sealing — every subsequent person imitates, no one’s private information enters the public pool, and the run of conformity grows without a single new fact behind it.

Three properties fall out of the model, each confirmed in laboratory experiments by Lisa Anderson and Charles Holt (American Economic Review, 1997), where a majority-color urn game reproduced cascades — including wrong ones — with real people and real payoffs:

• Cascades are rational. No individual is behaving foolishly. Each imitation is a defensible inference from the available evidence. The failure is collective: the group’s information is trapped in heads that no longer act on it.
• Cascades are frequently wrong. If the first two signals happen to mislead — perfectly possible with noisy information — the entire subsequent crowd locks onto the wrong option. The bigger the crowd, the more impressive the error looks.
• Cascades are fragile. Because the conformity rests on two or three signals, not thousands, a small injection of public information — one credible dissenter, one visible contrary fact — can shatter a cascade overnight. This is why fads collapse as suddenly as they form, and why the model’s authors titled it after fashion.

That fragility is the tell. Genuine consensus, built from many independent judgments, is robust to a lone dissenter. Cascade consensus is a house of cards wearing the costume of a fortress. The full model, its assumptions, and its applications get a deeper treatment on our information cascades reference page.

A tour of the wreckage

Tulip mania, 1636–37 — and the cascade about the cascade

The canonical bubble story: at the peak of the Dutch tulip trade in the winter of 1636–37, contracts for rare bulbs changed hands at prices rivaling houses, before the market collapsed in February 1637 and — legend says — ruined a nation. Modern scholarship tells a more interesting story. Historian Anne Goldgar, working through the Dutch archives for her 2007 book Tulipmania, found that the trade was confined to a fairly small network of merchants and craftsmen, that documented bankruptcies were few, and that the wider Dutch economy sailed on largely undamaged. Prices for rare bulbs genuinely spiked and crashed; the nationwide devastation is myth.

Why does the myth persist? Because stories cascade too. Goldgar traced how the lurid version descends substantially from moralizing pamphlets and from Charles Mackay’s 1841 Extraordinary Popular Delusions and the Madness of Crowds, repeated from author to author for centuries — each writer trusting predecessors rather than checking the archives. The most famous cautionary tale about herding is itself a textbook information cascade: sequential adoption of a claim based on observing prior adopters, with almost nobody consulting the primary evidence.

The South Sea Bubble, 1720

The South Sea Company’s stock rose roughly from around £128 in January 1720 to nearly £1,000 by that summer, propelled by debt-conversion schemes, aggressive promotion, and — crucially — the visible spectacle of neighbors getting rich, before collapsing back to earth by year’s end. Parliament’s subsequent investigation uncovered bribery reaching into government. As an information environment it was cascade-perfect: purchases were highly observable, the company’s actual prospects were nearly unknowable, and each new buyer’s decision leaned on the inference that so many prior buyers couldn’t all be wrong. They could, because most of them were making the same inference.

The Bay of Pigs, 1961: when the cascade happens around a table

Market bubbles involve strangers; the same dynamics operate among colleagues who admire each other — and there the psychology gets an extra twist. In April 1961, the Kennedy administration launched a CIA-backed invasion of Cuba by roughly 1,400 exiles at the Bay of Pigs. It collapsed within three days. The advisory group that approved it was, by any measure, brilliant. Psychologist Irving Janis made the case the centerpiece of his 1972 book Victims of Groupthink, asking how such a group approved a plan whose flaws were visible in advance — some advisers, Arthur Schlesinger among them, later admitted to doubts they never pressed in the room.

Janis’s answer was groupthink: when cohesion and the desire for unanimity override realistic appraisal, the group hears far more agreement than actually exists. He cataloged eight symptoms in three clusters:

Note the difference from a market cascade: the tulip trader genuinely came to believe prices would rise. The groupthink victim still privately disagrees — and hides it. Jerry Harvey’s Abilene Paradox (1974) pushed this to its absurd conclusion with the story of a family that drives four hot, dusty hours to Abilene for a dinner that, it emerges afterward, not one of them wanted: groups can converge on an outcome that no individual member prefers, because everyone mistakes everyone else’s silence for enthusiasm. Where a cascade destroys information through rational imitation, groupthink destroys it through self-censorship — a theme we explore across group settings in our guide to group decision making.

The Bay of Pigs story has a second act that proves the failure was structural, not personal. Eighteen months later, facing the Cuban Missile Crisis, Kennedy deliberately re-engineered how the same circle of advisers deliberated — Janis himself drew the contrast. Robert Kennedy was assigned to play devil’s advocate, the group split into subgroups that developed options independently before comparing them, outside experts were brought in, and the president absented himself from sessions so his preference could not anchor the room. Same people, different structure, and a deliberation that surfaced and stress-tested alternatives instead of ratifying the first plan on the table.

Dot-com and the housing bubble: cascades with modern plumbing

The NASDAQ Composite peaked at 5,048.62 on March 10, 2000, and lost roughly 78% of its value by its trough in October 2002. The bubble’s fuel was familiar — each funding round, IPO pop, and magazine cover was a public signal that so-many-smart-people-can’t-be-wrong — amplified by a genuinely new technology whose fundamentals were hard to value, which is precisely when private signals are weakest and imitation most tempting. The 2008 housing crisis added a structural accelerant: the near-universal belief that US house prices do not fall nationally was a correlated error — shared by borrowers, lenders, rating agencies, and regulators alike — and securitization propagated that single shared assumption through the entire financial system. When errors are correlated, aggregation doesn’t cancel them; it concentrates them. A diversified portfolio of bets that all secretly depend on the same belief is not diversified at all.

Cascades at internet speed

Social media is a cascade machine by construction: sharing is sequential, visible, and effortless, and verifying a claim is costly — the exact parameter regime the 1992 model identifies as cascade-prone. The landmark empirical study is Soroush Vosoughi, Deb Roy, and Sinan Aral’s “The spread of true and false news online” (Science, 2018), which analyzed roughly 126,000 verified story cascades on Twitter across eleven years. False news spread significantly farther, faster, and deeper than the truth, and falsehoods were about 70% more likely to be retweeted — an effect driven by humans, not bots, apparently because false stories were more novel and more emotionally arousing.

A newer worry sits on the horizon. A 28-author perspective led by Jason Burton (Nature Human Behaviour, 2024) examined how large language models may reshape collective intelligence — and flagged homogenization as a core risk: when millions of people consult the same few models, trained on overlapping data, their independent judgments quietly correlate. A society that outsources its thinking to one oracle is, in cascade terms, a crowd of one. The individual answers may be good; the diversity that makes aggregation work is what erodes — a concern that connects directly to automation bias and trust calibration in human-AI systems.

A field guide: cascade, groupthink, or herd?

“Herding” is the umbrella term for convergent behavior; the mechanisms underneath differ, and so do the fixes. A quick taxonomy:

Real disasters usually braid all three. A housing analyst in 2006 faced a cascade (everyone’s models said prices rise), groupthink (the desk’s culture punished pessimism), and payoff herding (careers are safe in the consensus). Which is why the remedies below attack structure rather than exhorting individuals to be braver.

How to break a cascade

Every effective countermeasure does one of two things: it restores independence to individual judgments, or it moves private information into the public pool. The classics:

Structure as the cure: separating contribution from observation

Look back at the cascade model and the failure has a precise location: the moment a person sees others’ choices before committing their own judgment. That is the joint where structure can intervene. Argumentree is built around exactly that intervention for collaborative decision making: participants contribute arguments independently and asynchronously — recording their reasoning before the room converges — with anonymous contribution available where status pressure is the risk. Arguments then meet the group as structured pro and con branches rated on their merits (helpfulness, clarity, accuracy, completeness), so a position’s support reflects its evidence rather than its arrival order.

And because the full audit trail preserves why each position was taken, the group never faces the cascade’s core blindness — choices visible, reasons invisible. A decision record with the reasoning attached is the organizational equivalent of publishing the private signals: later participants, and later decisions, inherit the information instead of just the conformity. The crowd stays a measuring instrument instead of becoming an echo chamber.