Prediction Markets
Prediction markets trade contracts whose payoffs depend on whether a future event happens: a contract paying $1 if the event occurs and $0 otherwise, priced at 0.70, expresses a 70% market-probability estimate. The reading is testable — sort forecasts by probability and check how often each bucket actually occurs — and the 2026 cross-platform evidence below says it holds up in aggregate.
Information aggregation
Wolfers and Zitzewitz separate three functions the market performs: incentives to seek information, incentives to reveal it truthfully (an informed trader profits by correcting a misprice), and an aggregation mechanism that combines diverse opinions into one price. This is Hayek’s price system operating on beliefs about single scored events; knowledge-problem is the general account of prices as a communications channel for dispersed knowledge. Conditional contracts extend the instrument — bets on quantities given outcomes (oil prices conditional on a supply disruption, for instance) express the market’s expectations about joint distributions, means, and medians, while the usual problem of separating correlation from causation remains.
Calibration is the headline statistical property: among events forecast at p%, about p% occur. Calibration alone is a modest test — a constant forecast at the base rate is perfectly calibrated while carrying no information — so accuracy also requires resolution, and evaluation uses proper scoring rules such as the Brier score (Brier 1950; Murphy 1973, both cited in the PNAS study).
Evidence
Wolfers and Zitzewitz (2004) surveyed a range of prediction contexts — Iowa-style political markets, betting markets, internal corporate markets — and found market-generated forecasts “typically fairly accurate,” outperforming most moderately sophisticated benchmarks. They open with the 2003 DARPA Policy Analysis Market episode: a geopolitical futures market cancelled after public uproar, with the follow-on betting on the DARPA director’s own departure becoming a live demonstration of the market’s information function.
Dickerson et al. (PNAS 2026) evaluated approximately one million resolved contracts and questions that opened and resolved between October 2015 and November 2025 across four platforms — Polymarket, Kalshi, Manifold, and Metaculus, spanning real-money, virtual-currency, and reputation-based incentives — over twelve domains and time horizons from hours to years. Aggregate calibration held: events assigned a probability of p% occurred approximately p% of the time, across platforms and many domains. Fine-grained analyses found systematic over- or underestimation in several highly consequential areas. The authors’ own framing: large-scale aggregation of judgments contains meaningful information about real-world uncertainty, and says nothing about the reliability of any individual forecast. Disclosures: the data-pipeline author received prizes and grants from Manifold-related funds and later consulted for Manifold Markets (no role in study design or analysis); the underlying data pipeline, Project Themis, is open source.
Market design
W&Z’s design checklist: how orders match (continuous double auction, pari-mutuel pools, posted market makers, or Hanson-style market scoring rules for combinatorial outcome sets), contract specification and adjudication (their examples: the Iowa 1994 Senate market broke when a senator switched parties while votes were still being counted; an internal software market broke when the client changed the delivery deadline), whether real money is at stake, and whether enough diverse information exists to generate trade at all. Manipulation attempts on the markets they reviewed had little effect, though the incentive to manipulate grows with a contract’s use in decision-making.
Related
- knowledge-problem — the general case for prices as information channels
- theory-of-the-firm — market versus firm allocation; LLM-era task-market experiments
- adversarial-collaboration — another epistemic reform instrument with scored outcomes