Wisdom of the crowds

Wisdom of the crowds describes how a group’s judgment can outperform individual judgments when its members bring diverse information, think independently, draw on decentralized knowledge, and have a way to aggregate their conclusions. In this account, the crowd does not become wise simply by being large: its members must possess relevant information, and the conditions for combining that information must be present.

The conditions for collective judgment

The framework has four parts. Diversity of opinion means that each person contributes some private information, even if it is only an unusual interpretation of known facts. Independence means that judgments are not determined by the views of others. Decentralization lets people specialize and draw on local knowledge. Aggregation supplies a mechanism that turns separate judgments into a collective decision. When these conditions hold, individual estimates contain both information and error; averaging can cancel errors while retaining information. But aggregation cannot manufacture knowledge from nothing: a crowd without useful information may still reach a poor judgment.

A stock market’s response to the Challenger disaster illustrates how the framework can work. Traders were effectively asked how much less certain companies were worth after the accident, a question with an objectively correct answer. Even if no individual trader knew that Morton Thiokol was responsible, the combined judgments of traders could reflect scattered pieces of information and collectively register that responsibility. The point is not that each participant possessed the answer, but that the market could bring dispersed knowledge together.

The account also distinguishes aggregation from Hayek’s emphasis on decentralized coordination. Friedrich Hayek valued markets as systems in which people transmit local knowledge through prices, without a central authority attempting to direct them. The book accepts the force of that argument, especially for coordination, but contends that the difficulty of aggregating information can be overstated. Pari-mutuel betting, for example, uses a central mechanism to combine bets into shared odds, yet can produce a notably accurate picture of the future.

Markets as tests of the idea

Sports betting provides a setting where predictions can be compared with outcomes. At the Mirage, bookmaker Robert Walker adjusted the point spread when early bets revealed strong demand for one side. Before the 2001 Super Bowl, two early $3,000 bets on the Baltimore Ravens were enough to move the opening spread from two and a half points to three. The line was initially set by the bookmaker, but thereafter shifted in response to bettors’ actions. The aim was to balance wagers on both teams, while the resulting line also represented the crowd’s evolving judgment of the game.

The final spread is not a crystal ball: point spreads only weakly predict the exact scores of most NFL games. Yet gamblers find it difficult to beat the spread consistently. Favorites cover about half the time, underdogs the other half, and there is no obvious, persistent error such as home teams regularly winning more often than predicted. A documented exception offered a narrow opportunity: home underdogs in weeks fifteen through seventeen had historically been good bets. Even with such occasional patterns, the Mirage’s final line was described as the most reliable available forecast of NFL outcomes roughly three-quarters of the time. Studies of other major sports likewise found markets generally efficient; in horse racing, final odds tracked finishing order and gave reasonably good estimates of winning probabilities.

The Iowa Electronic Markets offer another comparison between market forecasts and conventional polling. In one contract type, the payoff corresponded to a candidate’s share of the popular vote: a George W. Bush contract bought for the 2000 election paid 48 cents when he received 48 percent of the vote. A study of forty-nine elections from 1988 to 2000 found that election-eve prices were off by an average of 1.37 percentage points in presidential elections, 3.43 points in other U.S. elections, and 2.12 points in foreign elections. The market generally outperformed major national polls, including months before election day. Of 596 polls released across the presidential elections from 1988 to 2000, the market price on the day of each poll was more accurate three-quarters of the time.

These results are notable partly because the IEM was small, with no more than about 800 traders, and did not resemble the electorate: most traders were men, and a disproportionate, though shrinking, share lived in Iowa. They were not simply reporting their own intended votes. The market’s forecasts of what voters nationwide would do were nevertheless more accurate than asking those voters directly. Its predictions also changed less wildly than poll results, tending to shift sharply in response to new information rather than fluctuating without it.

Diversity in groups and organizations

Markets often contain diversity without having to design for it: participants differ in risk tolerance, time horizons, investing styles, and information, and entry is generally open to anyone with money. Small teams and formal organizations cannot rely on size and open participation in the same way. There, a few biased members can dominate a decision, so cognitive diversity must be deliberately built into the group.

Political scientist Scott Page tested the value of diversity with computer-simulated groups of ten or twenty problem-solving agents, each possessing a different set of skills. Some agents were highly effective individually and others less so. Yet groups containing both strong and weaker agents almost always outperformed groups composed only of the strongest; choosing a group at random and letting it work could match or beat the strategy of carefully selecting the best individuals to work alone together. The result does not make intelligence irrelevant: successful groups included high-performing agents, and none were ignorant. Rather, intelligence by itself does not guarantee varied approaches. People selected for similar strengths may bring overlapping tools, leaving the group with less collective capacity than a more varied team.

Collaboration and scientific knowledge

The global investigation of SARS shows collective intelligence operating through organized collaboration rather than a market. After reports of a severe respiratory disease in Guangdong, Hanoi, and Hong Kong, the World Health Organization issued a warning and organized an effort to identify the disease’s cause. On March 15 and 16, 2003, it contacted eleven laboratories in countries and territories including France, Germany, Japan, the United States, Hong Kong, Singapore, and China. The laboratories agreed to collaborate and began work on March 17.

Researchers held daily teleconferences, posted electron-microscope images, analyses, and test results on a WHO website, and exchanged virus samples. Working in parallel, they considered and eliminated possible causes. Scientists at Hong Kong University and the Centers for Disease Control separately isolated candidate viruses on March 21; subsequent tests detected a coronavirus in samples from many SARS patients, and laboratories in Germany, the Netherlands, and Hong Kong began sequencing it. In early April, infected monkeys at a Netherlands laboratory developed SARS. By April 16, about a month after the collaboration began, the network was confident that the coronavirus caused the disease.

Cynthia Goldsmith, an electron microscopist at the Centers for Disease Control and Prevention in Atlanta, was the first to spot the coronavirus. But identifying a candidate was not the same as establishing that it caused SARS. That conclusion depended on weeks of work across laboratories, including tests that ruled out other viruses. The discovery therefore belonged to the group: no single researcher or laboratory could claim the whole achievement, which the WHO account described as a collective discovery. The network had no director assigning each laboratory’s tasks or controlling the exchange of information. The WHO convened it, while the laboratories shared data and met daily; beyond that, they worked out how to divide the work themselves.

This collaboration exemplifies science as a collective enterprise extending beyond joint projects. Scientists publish results and make data available for scrutiny, allowing others to challenge conclusions or build on them. Failed hypotheses also contribute by showing other researchers where not to look. Scientific knowledge therefore depends not only on accumulated discoveries but on contemporaries exchanging and evaluating work.

Trust, acceptance, and limits

A finding becomes part of scientific knowledge through a community’s scrutiny and acceptance, not solely because one person announces it. The coronavirus caused SARS before the WHO made that conclusion public; in scientific practice, however, the claim became established through evidence that other researchers could assess and use. The subsequent work on tests and vaccines proceeded on the basis of a scientific consensus that the virus was a coronavirus.

That collective judgment relies on trust as well as reproducible experiments. Most scientists do not replicate most other researchers’ work; they rely on reported data and methods while pursuing new questions. A failed attempt to reproduce a well-established experiment may lead a scientist to question their own technique before doubting the original result. Replication and the exposure of fraud remain important, but scientific work would stall if every researcher had to retest every finding and every supporting procedure. The resulting system rests on an implicit confidence that the community can distinguish credible claims from unreliable ones.

The book also applies the idea to a proposal that provoked a sharp conflict between forecasting and public legitimacy. The Policy Analysis Market was intended to collect predictions about the Middle East and was to be open to the public. Senators Ron Wyden and Byron Dorgan led the campaign to end it, denouncing the proposal, and the market was killed amid public outrage. The author argues that public participation need not make a market’s forecasts less accurate: it could draw in information from people whom the intelligence community might not hear, and offer a way around bureaucratic and political pressures that shape analysts’ conclusions. The controversy nevertheless shows that a mechanism’s potential predictive value does not settle whether people will find it acceptable.

Wisdom of the crowds is therefore a conditional argument, not a claim that collective judgment is infallible. Markets can aggregate informed estimates, diverse groups can solve problems better than uniform groups of top performers, and scientific communities can establish discoveries no one participant proves alone. Each case depends on mechanisms that preserve useful differences while making their contributions available to others. The idea’s central contrast is between knowledge held in separate places and knowledge made usable through aggregation, collaboration, and scrutiny.