Information Economics
The Wisdom of Crowds: Why the Average Guess Beats the Experts
The Wisdom of Crowds is the empirical observation — and a provable statistical fact — that the average of many independent guesses is often more accurate than almost any single guess, including the expert's. Aggregate the estimates of an uninformed crowd about the weight of an ox or the number of jellybeans in a jar, and the mean lands astonishingly close to the truth. The magic is not in any one head; it is in the arithmetic of cancellation. Under the right conditions, individual errors — some too high, some too low — offset one another, leaving the shared signal to shine through. But those conditions are fragile, and when they break, the crowd turns into a mob.- Famous exampleGalton's ox, 1906: 787 guesses, mean 1,197 lb vs true 1,198 lb
- Popularized byJames Surowiecki, The Wisdom of Crowds (2004)
- Core theoremDiversity Prediction Theorem (Scott Page)
- Key relationCrowd error = Average individual error − Diversity
- Four conditionsDiversity, Independence, Decentralization, Aggregation
- Main failure modeCorrelated errors / herding (loss of independence)
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
Galton's ox: the founding experiment
In 1906 the statistician Francis Galton visited a country fair in Plymouth, England, where villagers paid sixpence to guess the dressed weight of an ox. Galton — no democrat, and expecting to prove the mob foolish — collected 787 legible tickets from butchers, farmers, and clueless townsfolk alike. He computed the median of the guesses: 1,207 lb. The ox's true weight was 1,198 lb, an error of under 1%. When the full dataset was later re-examined, the mean of the guesses came to 1,197 lb — within a single pound of the truth.
What stunned Galton was that the collective estimate beat essentially every individual, including the local cattle experts, even though the crowd contained many wild guesses. This is the seed of the whole idea: you do not need any single accurate person; you need many independent people whose errors point in different directions. The crowd's answer emerges from cancellation, not consensus.
The mechanism: errors that cancel
Model each person i as reporting a guess gᵢ = θ + εᵢ, where θ is the true value and εᵢ is that person's error. The crowd's estimate is the average ḡ = θ + ε̄, where ε̄ is the average error. If errors are unbiased (average to zero) and independent, then by the law of large numbers ε̄ → 0 as the crowd grows. More precisely, if each error has variance σ², the variance of the crowd's error is σ²/n. With n = 100 independent guessers, the crowd's expected error is 10× smaller than a typical individual's (since standard deviation shrinks by √n).
The two enemies of this result are bias and correlation. If everyone systematically overestimates (bias b ≠ 0), averaging cannot help — ε̄ → b, not 0. And if guesses are correlated with pairwise correlation ρ, the crowd's error variance floors out at ρσ² no matter how many people you add. This is why independence is not a nicety but the load-bearing assumption. Add a million people who all read the same wrong tip sheet and you have one guess repeated a million times.
The Diversity Prediction Theorem: a clean identity
Scott Page (2007) gave the intuition an exact algebraic form. For squared error, the following is a mathematical identity — always true, no assumptions needed:
Crowd Error = Average Individual Error − Diversity
Written out: (ḡ − θ)² = (1/n)∑(gᵢ − θ)² − (1/n)∑(gᵢ − ḡ)². The middle term is how wrong people are on average; the last term (the variance of the guesses around their own mean) measures how much they disagree. Because Diversity is subtracted, the crowd is always at least as accurate as its average member, and it gets strictly better the more the members disagree.
This is the counterintuitive punchline: disagreement is an asset. A crowd of clones (zero diversity) is only as good as one person. A crowd of thoughtful-but-different people can beat every one of them. Ability and diversity are, in a sense, on equal footing — which is why a diverse group of amateurs can out-predict a homogeneous group of experts.
A worked example you can check by hand
Suppose the true count of jellybeans is θ = 1,000. Four people guess:
- Alice: 700 (error −300)
- Bob: 1,300 (error +300)
- Carol: 900 (error −100)
- Dan: 1,500 (error +500)
The average guess is (700 + 1,300 + 900 + 1,500) ÷ 4 = 1,100, a crowd error of +100 — squared error 10,000. Now check the identity. Average individual squared error = (300² + 300² + 100² + 500²) ÷ 4 = (90,000 + 90,000 + 10,000 + 250,000) ÷ 4 = 110,000. Diversity (variance of guesses around 1,100) = (400² + 200² + 200² + 400²) ÷ 4 = (160,000 + 40,000 + 40,000 + 160,000) ÷ 4 = 100,000. And indeed 110,000 − 100,000 = 10,000 — exactly the crowd's squared error. The crowd (error 100) beats three of the four individuals and ties Carol, the single best guesser. Notice the biggest disagreements (Dan) helped, because his overshoot balanced the undershoots.
The four conditions — and how each one breaks
Surowiecki's 2004 synthesis distilled reliable crowd wisdom into four requirements, each with a failure mode you can name in real markets:
- Diversity of opinion — people draw on different private information and models. Breaks under groupthink and monoculture (e.g., the entire mortgage industry using the same ratings and models before 2008).
- Independence — my guess isn't contaminated by yours. Breaks under information cascades: once early movers show their hand, later people copy rather than reveal, and the correlation term
ρσ²dominates. This is the engine of asset bubbles. - Decentralization — people can specialize and use local knowledge. Breaks under central control that filters or overrides dispersed information (a critique Hayek made of central planning in 1945).
- Aggregation — some mechanism (a mean, a market price, a vote) actually combines the guesses. Without it, dispersed knowledge is just noise.
Crucially, the conditions can trade off with each other perversely: the more people talk to one another (which feels like it should help), the more independence they lose — turning a wise crowd into a herd.
Where it shows up: markets, forecasts, and Wall Street
The most powerful aggregator humans have built is the price. A market price is a continuously updated weighted average of countless private estimates, which is why the Efficient Market Hypothesis is essentially the wisdom of crowds with money on the line — dollars weight the votes, and arbitrage punishes the wrong. Prediction markets operationalize this deliberately: the Iowa Electronic Markets have often beaten polls at forecasting U.S. elections, and after the 1986 Challenger disaster, stock prices identified Morton Thiokol (maker of the failed O-rings) as the culpable contractor within minutes — long before the official commission did.
Firms exploit it too: Google and other companies have run internal prediction markets to forecast launch dates and demand; Netflix's 2006–2009 \$1M Prize showed that ensembles — averaging many diverse models — beat every standalone algorithm, a machine-learning echo of Galton's ox. But the same architecture produces the pathologies. The 2008 financial crisis and the 1990s dot-com bubble were failures of independence and diversity, not stupidity: correlated bets by sophisticated experts all leaning the same way. When the errors stop cancelling and start compounding, the crowd is at its most dangerous precisely when it feels most confident.
| Dimension | Crowd Average | Single Expert |
|---|---|---|
| Source of accuracy | Cancellation of independent errors across many people | Deep individual knowledge and models |
| Best case | Near-perfect on estimation tasks with a true numeric answer | Excellent on structured, causal, in-domain problems |
| Worst case | Herding, bubbles, information cascades | Overconfidence, blind spots, single point of failure |
| Cost & scalability | Cheap once you can poll many people | Scarce, expensive, hard to replicate |
| Key requirement | Diversity + independence of guesses | Genuine expertise in the relevant domain |
Frequently asked questions
Does the crowd always beat the expert?
No — and this is the most common overclaim. The Diversity Prediction Theorem guarantees the crowd beats its own average member, not that it beats the single best member. On numeric estimation tasks with unbiased, independent guesses, the crowd usually does beat almost everyone including experts. But on problems requiring deep causal reasoning, or where the crowd shares a systematic bias, a genuine expert can dominate. Crowds are strong at estimation, weak at insight.
Why is disagreement (diversity) a good thing?
Because in the identity 'Crowd Error = Average Individual Error − Diversity,' diversity is subtracted. More disagreement means people are making errors in different directions, so averaging cancels more of them. A crowd of people who all think alike is no better than one person. Paradoxically, adding a somewhat-wrong person who errs in a fresh direction can improve the group's accuracy.
What is the single most important condition?
Independence. The variance of the crowd's error is roughly σ²/n only when guesses are uncorrelated; if the pairwise correlation is ρ, the error variance can't fall below ρσ² no matter how many people you add. Once people copy each other — information cascades, herding, everyone reading the same news — you effectively have far fewer independent guesses than heads. Lost independence is what turns a wise crowd into a bubble.
Should you use the mean or the median?
Both aggregate, but they fail differently. The mean is efficient when errors are symmetric but is dragged by extreme outliers (one person guessing a billion jellybeans wrecks a mean, not a median). Guesses are often right-skewed on a linear scale, so many practitioners take the median or the mean of log-guesses (the geometric mean), which handles multiplicative errors better. Galton himself reported the median first.
How is this related to the Efficient Market Hypothesis?
The EMH is the wisdom of crowds applied to asset prices, with dollars as weights. A price aggregates countless private estimates, and traders who guess better profit while worse ones lose, so information gets impounded into the price. The link also explains EMH's failures: when independence breaks (herding, cascades) or diversity collapses (everyone using the same model), prices can drift far from fundamentals — bubbles and crashes.
Can this idea be gamed or manipulated?
Yes. Because aggregation and independence are mechanical, an adversary can flood a poll with correlated fake guesses (bots, astroturfing) or seed an early cascade that biases everyone downstream. Prediction markets fight this with skin in the game — manipulators lose money — but even markets can be cornered. The defenses are the same as the theory: preserve diversity and independence, and weight by demonstrated accuracy where possible.