Ask one seasoned analyst to predict an election and you get a confident, articulate, often wrong answer. Ask ten thousand people to stake something on it and the aggregate is eerily accurate. That gap — between the lone expert and the aggregated crowd — is one of the most durable findings in forecasting, and it is the reason prediction markets exist at all.
It is also the reason our public experiment treats the market price as the benchmark to beat, not a number to ignore. Understanding why the crowd is so hard to out-forecast is the first step to knowing when a model actually has an edge.
The Ox That Started It
In 1906 the statistician Francis Galton watched a country-fair contest where 787 people guessed the weight of a slaughtered ox. Galton expected the crowd to be hopeless. Instead, when he took the average of every guess, it came out at 1,197 pounds against an actual weight of 1,198 — off by a fraction of a percent. The butchers and farmers who "should" have known best did worse individually than the pooled guess of the whole fair.
That result feels like a magic trick, but it is just statistics. The crowd was not smarter than the experts. The averaging was smarter than any single guesser.
Why Averaging Works: Errors Cancel
Think of each person's estimate as three things added together: the true value, a personal bias, and some random noise. When you average many independent estimates, the random noise pulls in every direction at once and largely cancels out. What survives the averaging is the signal — the part of everyone's estimate that points at the truth.
Three conditions make this work. Estimates need to be diverse, so people are wrong in different ways. They need to be independent, so one loud voice does not infect the rest. And there has to be a sensible way to aggregate them. When those hold, the crowd's error shrinks faster than any individual can reliably manage.
When Crowds Get It Wrong
The magic breaks the moment errors stop being independent. If everyone reads the same headline, follows the same pundit, or copies the person next to them, their mistakes line up instead of cancelling. Herding, information cascades and groupthink all do the same damage: they correlate the errors, and correlated errors do not average away.
This is why bubbles and panics exist even though markets are usually good aggregators. The "wisdom" in wisdom of crowds is conditional. Remove diversity and independence and a crowd can be more confidently wrong than any single person in it.
Prediction Markets as Aggregation Machines
A prediction market is a wisdom-of-crowds engine with two upgrades. First, it weights opinions by conviction — people who are more certain commit more, so the price reflects not just how many believe something but how strongly. Second, it updates continuously, so new information gets absorbed into the price within minutes. The resulting number behaves like a probability, and a well-traded one is remarkably well calibrated.
That is exactly why a market price is a brutal thing to beat. It is not one opinion; it is thousands of them, weighted and refreshed in real time. Any forecaster claiming to outperform it is claiming to beat the aggregated, self-correcting judgment of a whole crowd.
What This Means for an AI Forecaster
A single model — however clever — is structurally a lone expert. It has one worldview, one set of blind spots, one way of being wrong. To beat a market, it needs a genuine informational or analytical edge, not just confidence. Most of the time it will not have one, and honesty means admitting that.
Our own scoreboard says so plainly: across the graded calls in our public experiment, the market currently leads our model 11 to 4. The crowd is winning, and we publish that rather than hide it. Every forecast is locked and Bitcoin-timestamped before the event, then Brier-scored in the open against the market price at the same instant. When the aggregated crowd beats a single model, that is not a bug in the experiment — it is the wisdom of crowds doing exactly what a century of evidence says it should. You can read every scored call at neuportal.ai/experiment.
Educational content — not financial advice, and not a betting tip.