Almost everything in this series describes people reasoning badly. This article describes people reasoning perfectly and arriving somewhere considerably worse than they could have, which is a different and more troubling structure.
Key Takeaway
The definition, from the paper that named it: "An informational cascade occurs when it is optimal for an individual, having observed the actions of those ahead of him, to follow the behavior of the preceding individual without regard to his own information."[1] We simulated twenty people with private signals accurate 70 percent of the time. A cascade formed in every run, locking in at the fourth or fifth person on average. Only 3.4 of the 20 private signals were ever acted on. Pooled instead of queued, the same twenty signals reach 95.2 percent accuracy against 84.5.
Our Grades For These Claims
Applying the scheme from the first article in this series.
Grade A for the theoretical result. It is a formal model in a leading economics journal, and its central property is a matter of arithmetic which we reproduce below.
Grade C for cascades occurring in laboratories for the reasons the model gives. One 1997 experiment confirmed the model; a 2000 paper reached contradictory results on crucial issues; and a 2008 paper titled with a question mark reports further doubts.
Grade B that cascades are fragile, on a paper whose title asserts it and which we did not obtain.
Our position: the mechanism is real and demonstrable, and whether real people follow it for the modelled reasons is genuinely contested, which matters less than usual because the structural remedy works either way.
A Note On Method
Everything here is verified to August 2026.
We did not obtain any paper in this article in full. We have citations from reference lists, a critical paper's characterisation of the experimental literature, and one quoted definition.
The definition reaches us through a book-notes site quoting the paper[1], not from the paper itself. It is presented there as a direct quotation and is widely reproduced, and we flag its provenance rather than treating it as first-hand.
Our best source for the dispute is a 2008 paper in a peer-reviewed journal summarising the contradictory experimental results[2].
The simulation is entirely ours, uses a standard setup with invented parameters, and demonstrates a mechanism rather than reproducing anyone's data.
This article discusses collective decision making. It is not investment or management advice, and nothing in it concerns any market.
The Definition
The concept.
Bikhchandani, Hirshleifer and Welch published A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades in the Journal of Political Economy, 100(5), 992–1026, in 1992, DOI 10.1086/261849[3][7].
Their definition, as quoted: "An informational cascade occurs when it is optimal for an individual, having observed the actions of those ahead of him, to follow the behavior of the preceding individual without regard to his own information."[1]
Three observations, ours.
The word doing all the work is optimal. This is not a failure of reasoning. It is what a correct reasoner does.
Without regard to his own information is the consequence. The person's private signal is not overridden by emotion or conformity; it is correctly judged to be outweighed by what the actions of others imply.
And the paper's title lists what it claims to explain: fads, fashion, custom, and cultural change. That is an ambitious scope for a model of sequential guessing, and the ambition is part of why the paper matters.
The Mechanism
How it happens, set out plainly. Ours, describing the standard setup.
People decide one after another. Each has a private signal pointing one way or the other, imperfect but better than chance. Each can see what everyone before them did, but not what they knew.
The first person follows their own signal, because it is all they have. The second sees one action and has one signal; if they disagree, the second may follow their own.
By the time the third person arrives, two people ahead may have acted the same way. If that is worth more than one private signal, the third person follows regardless of what they observed.
And here is the trap. Because the third person acted on the crowd rather than their signal, their action carries no new information. The fourth person sees three identical actions but is really still seeing only two signals' worth of evidence.
One observation, ours: the cascade is self-sealing. Every subsequent action looks like confirmation and contains none, and nobody involved has done anything wrong.
We Simulated It
Because a claim about arithmetic should be checked. The simulation is ours, uses invented parameters, and comes from no paper.
Twenty people in sequence. Each receives a private signal that is correct 70 percent of the time. Each observes all prior actions. Each follows their own signal unless the prior actions outweigh it, which is the rational rule.
Across 200,000 runs.
A cascade formed in 100 percent of runs.
It locked in, on average, at the fourth or fifth person.
Two observations, ours.
That it forms every time is a property of this setup rather than a general claim; with a different signal strength or decision rule it would form less often. What is general is that it forms early.
And the fourth person is the striking number. Whatever the twentieth person knows, the outcome was effectively decided before the fifth.
What It Discards
The cost, and it is the number we would put on a wall. Ours.
Of twenty private signals, an average of 3.4 were ever acted on.
16.6 signals per run were discarded, not because anyone suppressed them, but because acting on them was the wrong move for the person holding them.
Three observations.
The information existed. Sixteen people each knew something relevant and correct seventy percent of the time.
It was never destroyed, only never entered. Each of those people could have told you their signal if asked, and nobody asked in a way that made answering worthwhile.
And it is invisible from inside. A room where twenty people agree looks like twenty people agreeing. It cannot be distinguished, by observation, from three people agreeing and seventeen deferring.
The Comparison That Matters
What the same group could have achieved. Our simulation, same parameters.
One person alone, acting on their own signal: 70.0 percent correct, by construction.
The twentieth person in the queue, having watched nineteen others: 84.5 percent correct.
The same twenty signals pooled and decided by simple majority: 95.2 percent correct.
Two observations, ours.
The gap between 84.5 and 95.2 is the entire cost of the sequence. Same people, same information, same accuracy per person. The only difference is whether they spoke one after another or at once.
And that gap is larger than the gap between one person and the queue. Sequencing captures some of the available benefit and throws away more than it captures.
The Honest Reading
What our own numbers do and do not show. Ours, and this section exists because the temptation is to overstate.
Three points.
The queue beat the individual. 84.5 against 70.0 is a real improvement, and someone who follows a crowd is on average better off than someone who ignores it. Cascades are not a disaster relative to going alone.
They are a disaster relative to what was available. The right comparison is not the individual but the pooled group, and against that benchmark the queue loses more than half of the available gain.
And 15.5 percent of cascades in our runs settled on the wrong answer, with everyone downstream agreeing confidently. A wrong cascade looks exactly like a right one from inside.
Nobody Behaved Badly
The property that makes this different from everything else in this series. Ours.
Three observations.
There is no bias in the model. No overconfidence, no conformity pressure, no fear of dissent. Every participant weighs the evidence correctly.
Which means the usual remedies cannot help. Telling people to think independently does not apply, because they did. Telling them to be brave does not apply, because nobody was afraid.
And it distinguishes this from the eighth article's finding about suppressed disagreement. There, dissent was discouraged. Here there is nothing to discourage, because the dissenter has correctly concluded that their own view is outweighed.
Fragility
The property that makes cascades unstable, and a paper devoted to it.
A paper in Experimental Economics is titled Fragility of information cascades: an experimental study using elicited beliefs, and states that it "examines the occurrence and fragility of information cascades in two laboratory experiments"[4].
We did not obtain its results and report its title and framing only.
Two observations, ours, following from the mechanism rather than from that paper.
A cascade rests on very little information, in our simulation about three signals. So a small amount of new public evidence can overturn it, because it is not outweighing much.
Which produces the characteristic pattern: long stability then sudden reversal, with no proportionate change in the underlying facts. A practice everyone follows for years is abandoned in a season.
A Mirage?
The experimental dispute, reported because the theory being elegant does not make it true.
A 2008 paper titled Informational cascades: A mirage? summarises the state of the evidence: "Experimental research found contradictory results regarding the occurrence of informational cascades. Whereas Anderson and Holt (1997) confirmed the model of Banerjee (1992) and Bikhchandani et al. (1992) through lab tests, Huck and Oechssler (2000) came to contradictory results on crucial issues."[2]
It then states its own contribution: "This article presents experimental evidence supporting further doubts concerning 'Bayesian' informational cascades: just under two thirds of all decisions are characterized" by something our source does not reveal, truncating mid-sentence[2].
Two observations, ours.
We report the existence of that finding without its content, because the sentence that would tell us what two thirds of decisions were characterised by is cut off.
And note the quotation marks the authors put around "Bayesian". Their doubt is specifically about the reasoning process the model specifies, not necessarily about whether people herd.
The Right Reasons
The sharpest version of the objection, which is in a title.
The paper referred to above is Huck, S., and Oechssler, J. (2000), Informational cascades in the laboratory: do they occur for the right reasons?, Journal of Economic Psychology, 21, 661–671[2].
We did not obtain it and report the title only.
Three observations, ours.
The question is exactly right and it is the one this series keeps arriving at: a pattern can appear for reasons other than the ones proposed. The twenty-sixth article found a chart reproducible from noise; the thirty-first found a decline reproducible from time pressure.
Here the alternative is ordinary conformity. People might copy because copying is comfortable rather than because they have correctly computed that the evidence favours it.
And the practical consequence is smaller than it looks. If people herd for the wrong reasons, the information is still lost, and the structural remedy below works either way.
Two Different Things
A distinction the literature makes and popular usage collapses.
A reference list identifies Çelen, B., and Kariv, S. (2004), Distinguishing informational cascades from herd behavior in the laboratory, American Economic Review, 94, 484–497[4].
We did not obtain it and report only that the distinction is treated as requiring a paper to establish.
Two observations, ours.
Herding describes an outcome: everyone doing the same thing. A cascade describes a specific mechanism for it, in which private information stops being used.
Which means observing that everyone agrees tells you nothing about why, and a paper exists specifically because separating the two in the laboratory is difficult. Outside a laboratory it is harder.
Four Citation Variants
A sourcing note, continuing this series' running count.
Across our sources, the 1997 experiment's first author appears as both Anderson, L. R. and Anderson, R.[4][5]. The 1992 herd behaviour paper's page range appears as both 797–817 and 797–818[2][4]. One source spells the second author of the cascade paper Hishleifer[2]. And one bibliographic record titles it Cultural Change in Informational Cascades where every other source has as[3].
One observation, ours. That is four variants in a single article's sourcing, and it brings this series' running count to fourteen. None of them changes anything. The accumulation continues to be the point.
The Bridge To Averaging
How this connects to the thirty-sixth article, and it is the same result seen from the other side. Ours.
That article reported that averaging independent estimates cannot lose, and that the gain comes from errors pointing in different directions.
Three consequences.
A cascade is the systematic destruction of that independence. After the fourth person, the remaining judgments are not errors pointing different ways; they are copies of the same judgment.
Which is why our pooled figure of 95.2 percent and our queued figure of 84.5 come from identical inputs. Independence is not a nice-to-have in aggregation; it is the entire source of the benefit.
And a reference list points at the same connection from the literature's side, identifying Golub, B., and Jackson, M. O. (2010), Naive learning in social networks and the wisdom of crowds[6], which we did not obtain but whose title places the two phenomena in one frame.
Where This Happens In A Firm
The application. Ours, untested, and offered as recognisable situations rather than findings.
Four settings where decisions are made in sequence with visible actions and invisible reasons.
The partner meeting where views are given round the table. The fourth speaker has heard three and may reasonably conclude their own reading is outweighed.
Software and vendor selection. What is visible is which firms adopted a product. What is not visible is whether each evaluated it or copied the last one.
Pricing and engagement terms across a profession. Everyone can see what others charge and nobody can see the analysis behind it.
And technical positions on unsettled questions, where the visible fact is how others have filed and the private signal is your own reading of the rule.
One observation. In every case the observable is the action and the unobservable is the reasoning, which is the exact condition the model requires.
The Remedy Is Structural
What actually works, and it follows from the mechanism rather than from any study. Ours.
Four measures.
Collect views before any are shared. Written and simultaneous, so that nobody's judgment is formed after seeing another's. This is the same instruction the thirty-sixth article gave for a different reason.
Ask for the reason, not the conclusion. A cascade forms because actions are visible and signals are not. Requiring people to state what they observed rather than what they concluded puts the signal back into the room.
Reverse the speaking order deliberately, so the most junior or least committed speaks first and the person whose view carries most weight speaks last.
And ask explicitly whether anyone deferred. The question is not whether people agree; it is whether anyone agreed against their own reading. That is answerable and rarely asked.
One observation, ours: none of these requires anyone to be braver or smarter, which is appropriate, because in the model nobody was cowardly or stupid.
What To Do
Take views in writing and simultaneously. Our simulation gives 95.2 percent accuracy pooled against 84.5 in a queue, from identical inputs.
Assume the outcome is decided early. In our runs the cascade locked in at the fourth or fifth person out of twenty.
Ask what people saw, not what they think. Cascades form because actions are visible and private information is not.
Do not read agreement as confirmation. A room of twenty agreeing cannot be distinguished by observation from three deciding and seventeen deferring.
Expect reversals to be sudden and disproportionate. A cascade rests on very little information, so modest new evidence can overturn it.
Ask whether anyone agreed against their own reading. That question is answerable and almost never asked.
Do not treat this as a failure of courage. In the model every participant reasons correctly, so the remedies aimed at conformity do not apply.
Note that following the crowd still beats going alone. The queue reached 84.5 percent against 70 for an individual, and the loss is measured against the pooled alternative rather than against isolation.
The Limits Of This Analysis
Several caveats matter. This article discusses collective decision making and is not investment or management advice; nothing in it concerns any market. Everything is verified to August 2026. We did not obtain any paper discussed in full. The definition reaches us through a book-notes site quoting the 1992 paper, not from the paper itself. We did not obtain the 1997 experiment, the 2000 critique, the 2008 paper beyond two passages, the fragility paper, the paper distinguishing cascades from herd behaviour, or the social networks paper, and report titles, citations and characterisations only. The 2008 paper's own finding truncates mid-sentence in our source, so we report that it found further doubts without being able to say what they were. The simulation is entirely ours, uses a standard setup with invented parameters including a signal accuracy of 70 percent and a specific decision rule, and demonstrates a mechanism rather than reproducing anyone's data; that a cascade formed in every run is a property of these parameters and not a general claim. We found four citation variants across our sources and report them without resolving any. The section on where this occurs in a firm and the structural remedies are our own reasoning, untested, and the connection drawn to the thirty-sixth article in this series is our own.
Frequently Asked Questions
What is an information cascade?
How much information gets lost?
How much does that cost?
Is it just conformity?
Why do practices reverse so suddenly?
What do I actually change?
References
- Book-notes site reproducing a quotation attributed to Bikhchandani, Sushil, Hirshleifer, David, and Welch, Ivo, authors of A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades, that an informational cascade occurs when it is optimal for an individual, having observed the actions of those ahead of him, to follow the behavior of the preceding individual without regard to his own information. Note: a book-notes site quoting the paper, not the paper itself. The quotation is presented there as direct and is widely reproduced elsewhere; we flag its provenance rather than treating it as first-hand. goodreads.com
- Publisher record for Informational cascades: A mirage?, Journal of Economic Behavior and Organization, 67(1), 193–199, on experimental research having found contradictory results regarding the occurrence of informational cascades; on Anderson, L. R., and Holt, C. A. (1997), Information cascades in the laboratory, The American Economic Review, 87, 847–862, having confirmed the model of Banerjee, A. V. (1992), A simple model of herd behavior, The Quarterly Journal of Economics, 107, 797–817, and Bikhchandani, S., Hishleifer, D., and Welch, I. (1992), Journal of Political Economy, 100, 992–1026, through lab tests; on Huck, S., and Oechssler, J. (2000), Informational cascades in the laboratory: do they occur for the right reasons?, Journal of Economic Psychology, 21, 661–671, having come to contradictory results on crucial issues; on informational cascades in laboratory experiments also having occurred in the case of delayed decision-making so that subjects could benefit from observing others' actions; and on the article presenting experimental evidence supporting further doubts concerning Bayesian informational cascades, with just under two thirds of all decisions being characterized by something our source does not state, truncating mid-sentence. Note: a publisher record; our principal source for the experimental dispute. We obtained the abstract only, which truncates before stating the paper's own finding. This source spells the second author of the 1992 paper "Hishleifer". sciencedirect.com
- Bibliographic database record for Bikhchandani, Sushil, Hirshleifer, David, and Welch, Ivo (1992), Journal of Political Economy, 100(5), 992–1026, DOI 10.1086/261849, with access to the full text restricted; and its reference list identifying Becker, Gary S. (1991), A Note on Restaurant Pricing and Other Examples of Social Influences on Price, Journal of Political Economy, 99(5), 1109–1116, and Welch, Ivo (1992), Sequential Sales, Learning, and Cascades, Journal of Finance, 47(2), 695–732. Note: a bibliographic record; access restricted and we did not obtain the paper. This record titles it "Cultural Change in Informational Cascades" where every other source has "as". ideas.repec.org
- Publisher record for Fragility of information cascades: an experimental study using elicited beliefs, Experimental Economics, stating that the paper examines the occurrence and fragility of information cascades in two laboratory experiments; and its reference list identifying Anderson, L., and Holt, C. (1997), American Economic Review, 87, 847–862; Banerjee, A. (1992), Quarterly Journal of Economics, 107, 797–818, DOI 10.2307/2118364; Bikhchandani, S., Hirshleifer, D., and Welch, I. (1992), Journal of Political Economy, 100, 992–1026, DOI 10.1086/261849; and Çelen, B., and Kariv, S. (2004), Distinguishing informational cascades from herd behavior in the laboratory, American Economic Review, 94, 484–497. Note: a publisher record; we obtained the framing sentence and reference list only, and none of the paper's results. This source gives the 1992 herd behaviour paper's page range as 797–818 where reference 2 gives 797–817. cambridge.org
- Publisher record for a herding experiment in endogenous time, Experimental Economics, whose reference list identifies Anderson, R., and Holt, C. (1997), Information Cascades in the Laboratory, American Economic Review, 87, 847–862; Banerjee, A. V. (1992), Quarterly Journal of Economics, 107, 797–817; Bikhchandani, S., Hirshleifer, D., and Welch, I. (1992), Journal of Political Economy, 100, 992–1026; and Bikhchandani, S., Hirshleifer, D., and Welch, I. (1998), Learning from the Behavior of Others: Conformity, Fads, and Informational Cascades, Journal of Economic Perspectives, 12, 151–170. Note: a publisher record; citations only. This source gives the 1997 first author as "Anderson, R." where others give "Anderson, L. R." cambridge.org
- Academic preprint reference list identifying Bikhchandani, Sushil, Hirshleifer, David, and Welch, Ivo (1992), Journal of Political Economy, 100(5), 992–1026; Golub, Benjamin, and Jackson, Matthew O. (2010), Naive learning in social networks and the wisdom of crowds, American Economic Journal: Microeconomics; and Granovetter, Mark (1978), Threshold models of collective behavior, American Journal of Sociology. Note: citations only. We obtained none of these papers; the social networks title is cited here only because it places cascades and crowd wisdom in a single frame. arxiv.org
- Academic survey reference list identifying Bikhchandani, Sushil, Hirshleifer, David, and Welch, Ivo (1992), A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades, Journal of Political Economy, 100(5), 992–1026, and (1998), Learning from the behavior of others: Conformity, fads, and informational cascades, Journal of Economic Perspectives, 12(3), 151–170; together with later work including Bohren, J. Aislinn (2016), Informational herding with model misspecification, Journal of Economic Theory, 163, 222–247. Note: a survey's reference list, used to confirm the canonical citations and to record that the literature continued. We obtained none of the works listed. arxiv.org
This article discusses collective decision making and is not investment or management advice. No paper discussed was obtained in full, and the definition quoted reaches this article through a site quoting the paper rather than from the paper itself. The simulation is the authors' own, uses invented parameters, and demonstrates a mechanism rather than reproducing anyone's data. Four citation variants were found across sources and are reported without resolution.