Guide

Are Prediction Markets Accurate? What the Evidence Shows

Prediction markets, where people trade contracts that pay out based on whether an event happens, are often accurate but not infallible. Decades of research show market prices are frequently well calibrated and tend to match or beat polls and expert forecasts, especially over longer horizons. But accuracy degrades in thin markets, around genuine surprises, and where bias or manipulation creeps in. Prices are probability estimates, not guarantees.

What "accurate" actually means here

A prediction market price is read as a probability. A contract trading at 70 cents implies roughly a 70 percent chance the event happens. Accuracy is not whether the favorite won a single time; it is whether those probabilities hold up across many events.

Researchers measure this with calibration and scoring rules. Calibration asks: of all the things priced near 70 percent, did about 70 percent actually occur? Two common scores, the Brier score and log-loss, reward forecasts that are both correct and appropriately confident, and they punish predictions that are confidently wrong. These are the same metrics used to grade weather forecasts and poll-based models, which makes cross-comparison meaningful.

The key takeaway: well-calibrated does not mean always right. A good market should be wrong about 30 percent of the time on its 70-cent contracts. Being surprised occasionally is expected, not a failure.

The academic case that markets work

The foundational survey by economists Justin Wolfers and Eric Zitzewitz (published in the Journal of Economic Perspectives in 2004, with later follow-up work) argued that market-generated forecasts typically outperform moderately sophisticated benchmarks across elections, sports, and business outcomes. Their theoretical work showed that, under reasonable assumptions, prices reflect a wealth-weighted average of traders' beliefs, which is part of why prices can be treated as probabilities.

The longest-running evidence comes from the Iowa Electronic Markets, a small not-for-profit academic exchange run by the University of Iowa since 1988 under a no-action letter from the Commodity Futures Trading Commission. In a study by Berg and colleagues comparing the market against 964 polls across five US presidential elections from 1988 to 2004, the market was closer to the final result about 74 percent of the time, with the edge growing at longer forecasting horizons where polls are noisiest. This is the recurring basis for the claim that markets beat polls.

Where markets beat polls

Markets have structural advantages over polls in specific situations. They aggregate continuously rather than as snapshots, so they update the instant news breaks instead of waiting for the next survey. They put money behind opinions, which can discourage cheap talk and reward people who do real homework. And they fold in factors a single poll cannot, such as turnout, momentum, and how separate events interact.

The edge is largest when polls are weakest: far ahead of an event, on questions with no good polling at all (a CEO departure, a court ruling, a product launch date), and where many information sources must be combined. Internal corporate markets at firms like Google have produced useful forecasts that would be hard to obtain any other way, though research on those markets also documented identifiable biases.

Where markets fail

Accuracy is conditional, and several well-documented failure modes recur.

Thin markets: With few traders and little money at stake, a single participant can move the price, and the signal becomes noise. Niche or obscure contracts are the least reliable.

Favorite-longshot bias: Across betting and prediction markets, traders tend to overpay for unlikely "longshot" outcomes and slightly underprice heavy favorites. This means very low and very high probabilities are often the least trustworthy zone.

Surprises and tail events: Markets, like people, are poorly calibrated on rare events. A genuine shock the crowd did not see coming will be mispriced, by definition, until news arrives.

Manipulation and whales: Large traders can distort prices. In the 2024 US election, a French trader known as "Théo" placed over $28 million on a Trump win on Polymarket, using several accounts, and reportedly made tens of millions; while he was correct, the episode showed how a single actor can dominate a market and fueled questions about whether prices reflect crowds or capital.

The Vanderbilt comparison: bigger is not more accurate

A study from Vanderbilt University by Joshua Clinton and TzuFeng Huang examined roughly 2,500 political markets and about $2.5 billion in volume across Polymarket, Kalshi, and PredictIt during the final five weeks before the 2024 US election.

Measured with log-loss and Brier scores, the platforms diverged sharply: the researchers reported PredictIt as the most accurate at about 93 percent, Kalshi at about 78 percent, and Polymarket, by far the largest, the least accurate at about 67 percent. They found that many of Polymarket's national markets showed negative serial correlation, a fingerprint of overreaction and noise trading rather than steady, informed updating, and noted cases where mutually exclusive outcomes moved in the same direction.

The lesson runs against intuition: PredictIt's small, capped-stake design (an $850 per-contract limit during the 2024 cycle, since raised) appeared to produce calmer, more accurate prices than Polymarket's far larger stakes. Liquidity and volume alone do not guarantee accuracy. These are single-cycle findings on one election, so they should be read as evidence, not a final verdict.

How to read a prediction market sensibly

Treat a price as a probability with error bars, not a prophecy. A 65 percent contract means the other outcome is a realistic one-in-three event.

Check the depth before trusting the number. High volume and many distinct traders make a price more credible; a thin, low-volume market can be moved by one person. Be most skeptical at the extremes, where favorite-longshot bias and poor tail calibration tend to bite hardest. Watch for sudden moves with no news, which can signal a large trader rather than new information. And compare across venues and against polls or models: when independent sources agree, confidence is warranted; when they diverge, that disagreement is itself information.

Crowdtells reads markets this way, as one signal among several, never as advice.