How Accurate Are Election Markets? The Graded Record

We graded every settled Kalshi and Polymarket election contract we track against its own venue's settlement. Here is what the price ladder actually did.

GhostLeg, 2026-08-08, 6 min

Calibration numbers in this post were computed from GhostLeg's graded election-market record as of Thursday, August 7, 2026. The record grows as races settle; the politics board carries the live state.

Accuracy is a ladder, not a number

Ask "how accurate are election markets" and most answers reach for a single percentage. That number is close to meaningless. A market that prices an outcome at 10 cents is supposed to be wrong about it nine times out of ten. The honest question is whether the prices mean what they say at every rung of the ladder: do 10-cent contracts resolve yes about 10% of the time, 50-cent contracts about half the time, 90-cent contracts almost always?

That property is called calibration, and it is a graded, empirical question. GhostLeg tracks every election contract on Kalshi and Polymarket across the races on our politics board, and when a contract settles, we grade it against its own venue's settlement. No vibes, no cherry-picked wins. Here is what the record shows.

The ladder, graded

The table below is Kalshi's settled election record in the modern era, graded at the 31-to-90-days-out horizon: for every contract, what did the market price say one to three months before settlement, and what happened? This horizon is the honest one to lead with. It is far enough out that the market is genuinely forecasting, not just reading a called race off the news, and it carries our largest graded sample: 3,547 contract observations.

Read it top to bottom: the hit rate climbs at every single rung. Cheap longshots die (0.8% of sub-10-cent contracts came in). Coin-flip prices split roughly like coin flips (41.7% and 60.9% either side of the midline). Heavy favorites deliver (99.2%). The market's price ladder tracks reality all the way up.

That is what "election markets are accurate" actually means, and it is a stronger claim than any single percentage: a month or more before settlement, the price was already a usable probability.

What it looked like on real races

Calibration tables are abstract. Settled races are not. Two from our permanent record pages:

New Jersey Governor 2025. On the final day of trading, Kalshi closed Mikie Sherrill at 99.0 cents. Polymarket closed her at 85.5 cents, still holding Jack Ciattarelli at 14.5. Both venues settled the race for Sherrill. Same race, same day, same eventual settlement, and a 13.5 point gap between two markets that were both, in the binary sense, right. Which price was better is exactly the kind of question a graded record answers over many races rather than one.

President 2024. The largest settled dataset in our record. On election day, Polymarket's final close had the eventual winner at 58.0 cents against 42.2 for the runner-up: a favorite, not a certainty, which is what a contested national race should look like. The race page charts the full two-venue price path.

Both pages are permanent: every settled race we track becomes a fixed public record of what the markets priced and what happened.

How we grade, and what we deliberately do not claim

The methodology matters more than the headline, so here it is plainly:

• Each venue is graded against its own settlement. Kalshi and Polymarket have different resolution rules (official outcomes with media-call acceleration versus network-consensus or certification), so a contract's grade comes from its own venue's terms, never cross-applied.
• Horizon bands, not one snapshot. A price the night before settlement and a price three months out are different forecasts. Our record grades both, separately. The near-settlement bands look better for the obvious reason; the table above deliberately leads with the harder, further-out horizon.
• Every band carries its sample size. A 100% hit rate on 8 contracts is noise, not insight. Where a band is thin, treat it that way.
• The Polymarket side of the record is younger and thinner in our modern-era grading, and its historical backfill does not carry the volume data needed to filter thin, illiquid contracts the way we would want before quoting its ladder with confidence. So we do not. Polymarket's graded record is published on our race and record pages with its sample sizes, and its headline calibration will get the same treatment here once the record supports it. What we will not do is construct a marketing number from a dataset we just told you has a known gap.
• No "X% accurate" superlatives. The product's own record pages show graded results with sample sizes; that is the claim, in full.

This discipline is the point of the whole exercise. An accuracy analysis you can trust has to be one that was capable of returning a bad answer.

What this means if you follow elections

Three practical takeaways from the graded record:

1. Treat prices as probabilities, not predictions. A 70-cent contract is not "the market says it will happen." It is the market saying roughly three in four. The ladder above is what licenses that reading.
2. The far-out price already carries signal. You do not have to wait for election week; the 31-90 day horizon is already well calibrated on the big sample. Watching how prices move between now and November is watching real information arrive.
3. When two calibrated markets disagree, that gap is the story. Both venues can be well calibrated in aggregate and still disagree sharply on a specific race, which is happening right now on House control. That is the subject of our companion piece on the current Kalshi-Polymarket gaps.

The graded record grows as races settle and lives on the politics board; it is built on the same measurement infrastructure that powers the GhostLeg Data API.

FAQ

How accurate are election prediction markets? Graded across thousands of settled contracts, well calibrated: contracts priced 0-10 cents a month or more out resolved yes 0.8% of the time, 50-60 cent contracts 60.9%, 90-cent-plus contracts 99.2% (Kalshi, 31-90 day horizon, n=3,547). Accuracy is the whole ladder, not one number.

Are prediction markets better than polls? They answer different questions: a poll measures opinion at a moment; a market prices an outcome continuously, absorbing polls among many inputs. What the graded record shows is that market prices function as usable probabilities well before election day.

Which is more accurate, Kalshi or Polymarket? Our record grades each venue against its own settlement rules. Kalshi's modern-era ladder is above; Polymarket's graded record is published on our race pages with sample sizes, and we will publish its headline calibration when the sample supports it without caveats we cannot close.

Where can I see a race's full price history? Every race page on the politics board charts both venues' daily prices, and settled races like New Jersey Governor 2025 stay up as permanent records.

GhostLeg reports what prediction markets price. Prices are not predictions or recommendations. For entertainment purposes only. Past performance does not indicate future results.