Are Prediction Markets Accurate? Evidence vs. Polls

Prediction markets are accurate more often than most alternatives, but they are not oracles. Across decades of study — most notably the Iowa Electronic Markets, which have traded on US elections since 1988 — market prices have generally landed closer to final outcomes than major polls, especially weeks or months out. Markets have also missed high-profile calls, carry documented structural biases, and tend to fail exactly when polls fail, because both lean on the same underlying information. The honest summary: a well-traded market is among the best forecasting tools available, and still fallible enough that the price should be the start of your analysis, not the end of it.

Why Would Traded Odds Beat Polls in the First Place?

Polls and markets answer different questions. A poll asks a sample of people what they think or intend to do. A market asks anyone with capital what they are willing to pay to be right. Three mechanisms give the second approach its theoretical edge.

What do incentives change?

A poll respondent pays nothing for a careless answer, a socially convenient one, or a refusal to pick up the phone. A trader pays for every mistake. That does not make traders smart, but it filters participation: people tend to size up in markets where they believe they know something, and the cost of being wrong punishes cheap talk over time. Uninformed flow still exists — it is what pays informed traders to show up and correct the price.

How does a market aggregate information?

A poll is one input. A market price can absorb every available input at once: the polls themselves, early-vote data, fundraising reports, court calendars, candidate travel schedules, local knowledge that never reaches a national survey. It also weights that information by conviction. A trader confident enough to put serious size behind a view moves the price more than someone idly clicking. In theory, the result is a running consensus probability that no single poll, model, or pundit can replicate on their own.

Why does continuous updating matter?

A poll has a field period measured in days and publishes with a lag. A market reprices in seconds. When a debate, a ruling, or a data release hits, odds move before the next survey is even in the field. That speed is the core of how W.E.T. covers market-moving news: the question is never just what happened, but what the odds did when it happened.

What Does the Academic Evidence Actually Say?

Two bodies of work anchor the literature, and both are worth reading in the original.

Wolfers and Zitzewitz's 2004 survey, "Prediction Markets," in the Journal of Economic Perspectives, laid out the framework most people still use: under reasonable conditions, the price of a binary event contract can be read as a crowd-implied probability, and across the settings they reviewed — elections, economic data releases, other public events — those prices generally compared well against alternative forecasts. The same paper flagged the caveats the field still wrestles with, including miscalibration at extreme probabilities and the limits of thinly traded contracts.

The Iowa Electronic Markets are the long-running natural experiment. Operated by the University of Iowa's business school as a real-money research market since 1988, the IEM has traded on every US presidential cycle since. Berg, Nelson, Rietz, and coauthors compared IEM prices against a large set of major national polls across multiple election cycles and found market prices closer to the final vote shares than the polls in most comparisons — with the market's edge generally larger at longer horizons, where polling is noisiest. The exact magnitudes vary by study and by election, which is why the qualitative finding is the one worth carrying: sustained, modest, repeated outperformance. Not clairvoyance.

The limits of that evidence deserve equal billing. The IEM is small, stakes are capped, and its traders are self-selected. Results from vote-share markets do not automatically generalize to every contract listed on modern venues like Kalshi and Polymarket, where market quality ranges from deep and competitive to nearly untraded. The fair reading is that markets are competitive with, and often better than, polls at forecasting elections — not that any price on any board is trustworthy.

What Are the Famous Hits — and What Do They Actually Prove?

What happened in 2024?

In the final weeks of the 2024 US presidential race, prediction markets — Polymarket most visibly, with Kalshi newly live for US election contracts — priced Donald Trump as a clear favorite while major poll-based models described something close to a coin flip. Trump won, and the "markets beat the polls" narrative hardened almost immediately. Some of it is deserved: markets moved earlier and further than survey averages, and anyone who disagreed was free to take the other side at a profit if the price was wrong.

But the hedge matters as much as the headline. A single binary outcome cannot validate a probability — a model that says 50/50 is not refuted by either result, and a market near a Trump favorite "wins" the narrative while telling you almost nothing about calibration from one observation. Documented large single-account buying moved Polymarket's price during that stretch, and venues diverged from one another, a reminder that a price reflects flow as well as belief. Score 2024 as a point for markets. It is not a proof.

Where have markets been flat wrong?

Brexit, June 2016: betting markets priced Remain as a heavy favorite into referendum day even while polls showed a near-even race. That is the uncomfortable case where markets were arguably worse than the polls they were supposed to improve on. The 2016 US election is the other canonical miss — markets favored Clinton, alongside nearly every poll and model. Markets aggregate available information; when the available information is wrong, the price is wrong with it.

When Do Prediction Markets Fail?

Thin liquidity

The aggregation argument assumes people are actually trading. In a thin market, the last trade may be hours old, the spread wide, and a modest order enough to swing the displayed probability by several points. Volume and depth are part of the forecast. Across the events W.E.T. tracks, liquidity varies enormously — headline political contracts trade deep, while niche events trade by appointment.

Longshot bias

Prediction markets inherit the favorite-longshot bias documented across betting markets: low-probability outcomes tend to trade above fair value, near-certainties slightly below. Lottery-style demand for cheap tails, the capital cost of shorting them, and bounded payoffs all contribute. Practical translation: a 3-cent contract is not automatically a 3% event.

Manipulation attempts

Manipulation is real and documented — researchers identified a single large trader who appeared to prop up Mitt Romney's Intrade price for weeks in 2012. The academic evidence on historical and modern episodes suggests the effects tend to be transient in liquid markets, because holding a price away from fair value is a standing subsidy to everyone trading against you. But transient can still matter when a screenshot of the odds becomes a news story, and thin markets are far cheaper to bend.

Correlated errors

Markets and polls are not independent forecasts, because traders trade on polls. When polling carries a systematic error, markets tend to inherit it — 2016 being the clearest case. The diversification you get from watching both is smaller than it looks.

Self-fulfilling concerns

Odds are now media objects. When campaigns, donors, and journalists treat a market price as evidence about the race, there is at least a plausible channel for the measurement to feed back into the thing being measured — morale, coverage, money. The empirical size of that effect is unsettled, and honest coverage says so rather than waving it away.

How Should You Actually Use Market Odds?

Treat a market price as an incentive-weighted consensus estimate: usually the best single number available, never a guarantee. In practice that means checking liquidity before trusting a probability, comparing venues — divergence between Kalshi and Polymarket is itself information — watching how odds respond to catalysts rather than staring at levels, and discounting extreme tails for longshot bias. If you are new to the mechanics, start with what a prediction market is, then how the odds actually work, and keep our evergreen prediction markets guide close. All of this is informational context, not financial advice — event contracts carry loss, liquidity, and platform risk.

Where to go next: watch live odds, volume, and upcoming catalysts on the event dashboard, and bring your reads to the W.E.T. community, where the accuracy debate plays out one market at a time.