Thousands of traders, one opinion.

Simple assumes nothing

The reason a crowd beats an expert is that people are wrong in different directions. The errors cancel out and what survives is signal. This only works when the mistakes are independent.

Most of the money on Polymarket now belongs to bots. Bots built on similar models, reading similar sources, trained on overlapping data. They do not talk to each other, and they do not need to. Given the same inputs they produce the same answer, so they buy at the same moment and sell at the same moment.

A market like that looks busy and liquid. What it contains is one opinion held with a great deal of money, wearing a crowd's costume. If the opinion is wrong, nothing inside the market corrects it.

Moderate assumes you know what a market is

Crowd aggregation rests on error independence. Correlate the errors and the effective number of independent participants collapses. A market with ten thousand traders whose errors correlate at 0.9 has the statistical power of a handful of people.

Two consequences follow. Price quality degrades in a way volume cannot reveal, because volume rises while independence falls. And the market becomes fragile: one news headline parsed the same way by every model moves every agent in the same direction at once, with no natural counterparty on the other side. Individual rationality produces collective fragility.

The usual story about bots on Polymarket is efficiency. 37% of agents are profitable against 7 to 13% of humans, so the smart money is automated and the market is better for it. The correlation reading takes the same numbers and concludes the opposite. Both stories fit everything published so far, which is why this needs measuring rather than arguing.

Technical state of the art and the gap

Under a standard aggregation model the variance of the crowd estimate scales with 1/n_eff, where n_eff = n / (1 + (n-1)ρ) for pairwise error correlation ρ. As n grows, n_eff tends to 1/ρ. The ceiling is set by correlation and not by participation, so growth in agent count buys nothing once ρ is bounded away from zero.

Everything empirical here is open. Nobody has estimated ρ across the agent population on any venue, and the data is sitting there: Polymarket trades are on-chain, wallets are clusterable, and agent wallets are partially identifiable through timing regularity and API signatures. A first result is a time series of ρ, and the question worth answering is whether it rises as agent share rises.

Then the mechanism question. If ρ is high, what restores independence? Sybil resistance does not help, because these are distinct entities with correlated priors rather than one entity with many keys. Candidates: reward schemes paying for being early and contrarian rather than merely correct, diversity requirements on market maker counterparties, or subsidising human participation as decorrelating noise. All speculative, none tried.

Note the interaction with meta-markets. Trader correlation is a parameter in the recursive tower, and correlated agents are the most plausible mechanism for pushing that tower out of convergence and into oscillation.