Are prediction markets accurate?

What prediction market calibration means#

A single outcome cannot grade a probability. When a market says 70% and the event does not happen, that is not evidence of failure: 70% shots are supposed to miss three times in ten. So researchers grade prediction markets the way meteorologists get graded, on calibration. Collect every contract that ever traded at a given price, wait for the events to resolve, and count. If 60 cent contracts win about 60% of the time and 85 cent contracts win about 85% of the time, the prices mean what they claim to mean.

That is the yardstick, and it comes with a built-in warning. Calibration is a property of averages over many markets. It says nothing about whether the specific price in front of you is right. Everything below is evidence for the first half and the known exceptions to the second.

The calibration evidence#

The oldest dataset belongs to the Iowa Electronic Markets, a small real-money exchange the University of Iowa has run for research since 1988. In the Berg, Nelson and Rietz data, polls taken within five days of a presidential election missed the final two-party vote split by 1.62 percentage points on average. The market's error over the same days was 1.11 points. A modest edge, but it repeated across five straight elections.

Modern venues now have their own audit. In 2025, economists Bürgi, Deng and Whelan pulled more than 300,000 contract prices from Kalshi's API, covering 12,403 events across politics, economics, and entertainment, and checked every price against what actually happened. Prices tracked winning frequencies well through most of the range, and they grew more accurate day by day as markets approached resolution, with a sharp final-day improvement.

The mechanism is not mysterious. Being wrong costs money, so bad information gets priced out instead of amplified; a pundit pays nothing for a confident miss, while a trader pays the full stake. And a market reprices in seconds, without waiting for anyone to field a survey. If prices-as-probabilities is new to you, how prediction markets work is the five-minute foundation. This page assumes it.

Both studies also found places where prices bend away from the truth. Those are below, and they matter just as much.

Prediction markets vs polls#

The cleanest comparison on record is the same Iowa study. Berg, Nelson and Rietz lined up 964 national polls from the five presidential elections between 1988 and 2004 and asked, for each poll, a simple question: on the day this poll left the field, which was closer to the eventual vote split, the poll or the market price? The market won 74% of the time. At horizons beyond 100 days it significantly beat the polls in every one of the five elections, which is exactly when a forecast is worth the most.

The result makes sense once you see the asymmetry. A poll is one input: a snapshot of whoever answered this week, filtered through someone's turnout model. A market price consumes the polls and then everything else, fundraising numbers, early votes, debate reactions, whatever broke an hour ago. Traders who ignore good polls lose money, and so do traders who ignore everything that is not a poll.

Markets are downstream of polls plus the rest. Downstream cuts both ways, though. When every input shares a blind spot, the price inherits it, and that is the honest way into 2016.

Prediction market accuracy in elections: 2016 and 2024#

On the final weekend of the 2016 campaign, Reuters surveyed where the venues stood. Betfair had Hillary Clinton at 83%. PredictIt had her at 81%. Even the academic Iowa market, the one with the celebrated track record, gave her 71%. Donald Trump won. The markets missed alongside the pollsters, largely because the polls they leaned on were off in the same direction in the same states.

There are two fair readings of that night. The defense: an outcome priced around 20% arriving is something a calibrated system must deliver regularly, and one landing proves nothing. The critique: markets are supposed to aggregate more than polls, and in 2016 that extra information was not worth much. Both are true. What 2016 settles is narrow but useful: prediction markets do not manufacture knowledge nobody has. When markets and polls drink from the same well, they get sick together.

2024 read differently. Polling averages showed a dead heat to the end, and the markets disagreed: on election morning Kalshi priced Trump at 57 cents and Polymarket at 62, with Polymarket alone handling more than $3.6 billion on the presidential race. Trump won. The caveat that keeps the story honest sits in the same coverage: PredictIt priced the identical race as a coin flip that morning, more than ten points from Polymarket. The venues could not all have been right, so a single cycle cannot crown the method either. What 2024 unmistakably did was make these prices mainstream: more money, more attention, and more people treating a contract price as the forecast of record.

The favorite-longshot bias#

The best measured flaw in market prices lives at the extremes, and it is old. Griffith found it at American racetracks in 1949, and it has appeared in betting data around the world since, with very few exceptions. Snowberg and Wolfers put numbers on it using a decade of US racing: betting every horse at odds of 100 to 1 or longer lost about 61 cents on the dollar, while betting every favorite lost only about five and a half. Longshots are priced as if they win far more often than they do. Favorites are priced as if they win slightly less often than they do.

Event contracts inherit the same tilt. In the Kalshi data, buyers of contracts at 10 cents and under lost more than 60% of their money, while contracts above 50 cents won a little more often than their prices implied and returned a small profit before fees. The bias showed up across politics, economics, and entertainment, across small trades and large ones, and across every volume level the authors checked, so liquidity alone does not scrub it out.

Snowberg and Wolfers' tests point at the cause: misperception rather than a taste for risk. People overweight tiny probabilities when a big payout is attached, the same instinct that sells lottery tickets. The adjustment for a reader of prices is blunt. Treat a contract under 20 cents as claiming more than it can deliver, and treat a heavy favorite as slightly better than it looks. How to read prediction market odds covers the arithmetic.

The interest drag on long-dated contracts#

The second bias is less famous and easier to explain with a bank account. Money committed to a contract typically earns nothing while it waits. Page and Clemen showed what that does to prices using a large set of InTrade transactions: markets close to expiration were reasonably well calibrated, but prices on events months or years out were squeezed toward 50 cents, near-certainties too cheap and longshots too dear.

The logic is capital cost. Buying a 90 cent favorite that resolves next year locks up a lot of money for a small, slow payoff, so traders who value their capital walk away and the price sags below fair value. The gap looks like free money, and mostly is not: by their analysis, only a trader with an unusually low discount rate earns excess returns exploiting it. The authors even proposed the obvious fix, paying interest on traders' balances.

The practical lesson survives any fix. A far-off near-certainty trading at 88 cents is often not a considered forecast of 88%; part of the gap is rent on parked cash.

What a calibrated average means for a trader#

Every finding above describes thousands of markets at once, and none of it certifies the one price on your screen. That gap is where trading happens. The same election traded more than ten points apart across venues on the same morning, and the longshot bias survives at every volume level, so the aggregate can be honest while a given segment stays reliably wrong. A short checklist for reading any single price:

• Check the volume first. Calibration studies measure markets people actually trade. A price with fifty dollars behind it is one person's opinion wearing decimals.
• Discount cheap longshots. The shelf under 20 cents has claimed more than it delivers since 1949, and it still did in the newest Kalshi data.
• Expect drag on far-dated contracts. Prices get squeezed toward 50 cents by capital costs, not by information.
• Ask whether related markets agree. Real news tends to move a whole family of markets. Until the neighbors confirm it, treat a lone spike as one trader's opinion.

Those habits are also where edges start: the biased segments and thin books are the places a careful trader can beat the posted price. Prediction market strategies works through which of those edges hold up.

The cheapest way to test all of this#

Reading calibration studies settles what markets can do. It does not settle whether you can beat them, and that question has a cheap answer. PaperPicks mirrors live prediction market odds on iPhone with a $100 paper bankroll: real prices, real settlement when the underlying markets resolve, and a permanent server-scored record with no resets and no real money ever. Trade against the consensus for a month and your record will say which side of the calibration line you are on. Download it on the App Store and find out for free, against the same prices the research keeps grading.