What Are Prediction Markets? Event Risk, Information Aggregation, and the Meaning of Market Probabilities 

Prediction markets are usually described as mechanisms that aggregate dispersed beliefs into probabilities. That description is useful, but it is incomplete. A prediction market is first a market for state-contingent claims: participants trade exposure to a contractual event using unequal amounts of capital, information, conviction, risk tolerance, and strategic intensity. The resulting price may become a forecast, and increasingly it also becomes a public information product consumed by journalists, investors, policymakers, and individuals who never trade. This paper asks a more basic question than whether prediction markets are accurate: what kind of economic institution produces the probability, and what can that probability actually tell us about the information behind it?

Using a selected universe of 500 Polymarket markets, approximately 14.6 million transactions, and 1.25 million wallet addresses, the paper studies the relationship between quoted probabilities and the trading structures that generate them. The empirical design constructs as-of wallet-flow matrices before each market observation and uses spectral diagnostics to characterize common dependence across traders. The resulting structural panel contains 802 market-horizon observations.

Three findings matter. First, similar probabilities can arise from very different trading environments. Within narrow price and horizon groups, markets display substantial heterogeneity in detectable common trading structure. A scalar probability therefore does not uniquely identify the process that produced it. Second, the raw number of detected factors is strongly related to market size: the rank correlation between wallet count and factor count is approximately 0.90, showing that a larger number of observed components cannot be interpreted mechanically as a larger number of independent information sources. Third, this dependence does not disappear when matrix width is held fixed. In fixed-size experiments, the median observed factor count is 3 for 32-wallet matrices, compared with a median null mean of about 0.45, and 4 for 64-wallet matrices, compared with about 0.61 under temporal-shift nulls. These results indicate persistent cross-wallet structure, but they do not identify its economic origin. Common public news, shared models, imitation, arbitrage, automation, or informed trading can all generate similar statistical patterns.

The paper therefore proposes a functional interpretation of prediction markets. They perform four distinct but overlapping roles: they are markets for event risk, mechanisms of information aggregation, forecasting technologies, and increasingly public information products. None of these functions implies the others automatically. A market can generate an informative forecast without representing average beliefs; broad participation need not imply informational independence; and a contract price can be economically meaningful without being universally well calibrated.

The central conclusion is simple: prediction markets are best understood as event-risk markets that endogenously produce probabilistic signals. The probability is an important output of the institution, but it does not by itself reveal representativeness, independence, informational breadth, or forecast quality. Those are separate empirical questions.