Prediction markets vs polls

A poll asks a sample of people what they currently intend and reports that, with a margin of error. It is a measurement of a present state. A market asks anyone willing to risk money what will eventually happen, and reports the price at which those people stop disagreeing. It is a forecast of a future state.

This is why a poll showing 52 % and a market pricing 70 % is not a contradiction. The market has taken the poll, added turnout expectations, historical polling error, the time left to move, and everything else it knows, and produced a different kind of number.

Three things. It aggregates every information source rather than one instrument. It updates continuously instead of at publication intervals. And it charges people for being wrong, which is a meaningfully different incentive from being asked a question by a stranger.

The third is the one people underrate. A poll respondent bears no cost for a careless answer. A trader who prices something at 70 % that should be 40 % hands money to whoever noticed.

Markets are only as good as their liquidity. A market with three traders and $400 of volume is not the wisdom of crowds — it is the opinion of three people with a price attached. The aggregation argument requires an actual crowd, and most listed markets do not have one.

There is a second failure. Markets inherit the polling error rather than escaping it, because traders read the same polls. When the polls are systematically wrong in one direction, the markets built on them are usually wrong in the same direction — less so, but not independently.

As of 1 September 2026 our cache held 43,914 markets, 6,360 of them open and active, and 16,812 resolved ones in the corpus we score against. Volume across the open set is heavily concentrated: a small number of markets carry most of the money, and the rest sit on volumes where a single trade moves the price several points.

That distribution is the honest basis for the accuracy question. Aggregated across the liquid markets, prices are informative. Averaged across everything listed, they are not — and any study that treats all listed markets as equal evidence is measuring the wrong population.

It is also why we log every AI probability estimate against the market price at the time and score it later, rather than claiming accuracy in the abstract — and why our market accuracy study reports prices by range and topic, with the sample behind every figure.

Read the poll for the level and the market for the direction. A poll tells you where opinion sits with a stated uncertainty; a market tells you what changed today and by how much. The disagreement between them is the interesting signal, because it usually means the market knows something about turnout, timing or structure that the poll did not ask about.

When they diverge sharply and the market is liquid, the market is usually the better guess. When they diverge and the market is thin, the poll is.