Information Cascades: When Rational Agents Ignore What They Know

A queue, one decision at a time

Imagine a line of people choosing between two restaurants, A and B, one at a time and in full view of everyone behind them. Each person privately receives a noisy hint about which restaurant is actually better — say, they overheard a review that is right 70% of the time — but they can also see which restaurant every earlier person in line walked into. This is the setup Sushil Bikhchandani, David Hirshleifer and Ivo Welch formalised in 1992 as an informational cascade, and it is the model behind the simulation on this page: agents deciding sequentially, each with a private noisy signal, each observing only the public actions — never the private signals — of everyone who moved before them.

The striking result is that this setup does not average out into the crowd slowly discovering the truth. Instead, after only a handful of decisions, the queue typically freezes into unanimous agreement — sometimes on the correct restaurant, sometimes, with real probability, on the wrong one — and everyone after that point ignores their own private evidence entirely.

Why two decisions in a row can silence everyone after them

The mechanics are pure Bayesian updating. The first person has nothing to go on but their own signal, so they follow it. The second person sees the first person's choice and combines it with their own signal — if the two agree, that is already weak public evidence in favour of one restaurant. The pivotal moment is the third person: if the first two both chose A, the public tally already favours A by a margin that a single private signal — even a fairly reliable one — cannot overturn. The Bayes-optimal move for the third person, and for every person after them, is to copy A regardless of what their own signal says.

private signal:   correct with probability p > 0.5, independent across agents
prior belief:     P(A is better) = P(B is better) = 0.5
public tally:     count(A chosen) − count(B chosen)

rule (symmetric binary cascade model):
  once |public tally| reaches 2 in one direction
  → posterior already favours that option more than
    any single private signal could reverse
  → Bayes-optimal action: copy the majority, discard your own signal
  → the cascade is now self-sustaining: it neither strengthens nor
    weakens, because no further private information gets revealed

Notice what is lost here: from the third agent onward, every action reveals nothing new about the underlying state, because everyone is copying rather than reporting. A hundred more people joining the queue adds a hundred more identical votes to the tally, but zero additional bits of independent evidence — the cascade has effectively stopped the crowd's collective learning process, even though the queue keeps growing.

Cascades are not the same as learning

It is worth being precise about what an informational cascade is not. It is not the same as a group slowly converging on the truth through accumulated evidence — that would require each new decision to still carry information, which cascades explicitly shut off. It is also a narrower phenomenon than herding in general: herding can arise from payoff externalities (it is directly better for you if others do what you do, regardless of information) or from reputational concerns (a manager copying peers to avoid looking foolish alone), whereas a pure informational cascade arises purely from the logic of inferring hidden information from observed actions.

The single most important property of an informational cascade is its fragility. Because the crowd's apparent consensus is built on very little actual information — in the simplest model, just the first two signals that happened to agree — it takes surprisingly little new evidence to break it. A single credible, strongly informative public signal, arriving after hundreds of imitators have already piled in, can outweigh the entire cascade and flip everyone's behaviour almost immediately. This is precisely why cascades in financial markets, fashion and technology adoption can reverse so abruptly: the visible consensus was never as evidentially deep as its size suggested.

Where this shows up outside the lab

The theory was originally motivated by hard-to-explain patterns of fads, IPO underpricing and sudden shifts in medical practice, but the same logic recurs across economics and social behaviour: investors piling into an asset because earlier buyers did, regardless of their own research; diners choosing the busier of two adjacent restaurants; voters coalescing around whichever candidate polls first show as ahead; depositors joining a bank run set off by a handful of early withdrawals rather than any actual news about the bank's solvency. In every case the mechanism is identical — sequential decisions, public actions but private evidence, and a tipping point after which imitation dominates independent judgement.

Frequently asked questions

Try it live

Everything above runs in your browser — open Information Cascade and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.