How accurate were prediction-market prices a day before the result?

The test

A price on a prediction market is meant to read as a probability: a contract at 30 cents pays a dollar if the outcome happens, so 30 cents says "about 30%". If prices mean what they say, then of all the outcomes priced near 30%, about 30% should happen. Forecasters call that property calibration. It can be checked directly, without trusting any single forecast: group many markets by price and count how often each group came true.

pdata records a market's price every time it changes, at most every 30 minutes, for every venue it tracks. For each yes-or-no market that settled in the window with a winner named by its venue, this study takes the last price recorded at least 24 hours before the venue closed the market, and again at least 7 days before. It counts a price only if the market had traded and its order book had a bid and an ask no more than 10 points apart — a working market, not a placeholder. The venue's own verdict decides the outcome.

A day before close

The prices tracked outcomes closely across the whole range. Markets priced between 40% and 50% averaged 44.8% and resolved YES 45.6% of the time. Markets at 90% or more averaged 96.2% and resolved YES 96.3% of the time. The 16,644 markets priced under 10% — long shots, 42% of the sample — averaged 2.2% and came true 1.4% of the time.

Where the prices were off, they were off in one direction. Long shots happened slightly less often than priced, and favourites priced between 70% and 90% won slightly more often: 84.5% on average in the 80–90% band against 87.4% resolved, outside the band's 95% range. Cheap outcomes priced a little too high and likely ones a little too low is the shape known as the favourite–longshot bias. Here it amounts to one to three points.

Calibration alone is a low bar: a forecaster who gave every market the same number, the overall share that resolved YES (27.3%), would be perfectly calibrated and useless. The Brier score — the average squared gap between price and outcome, where 0 is perfect — measures both together. The prices scored 0.110; always quoting the overall YES rate would have scored 0.199.

Polymarket and Kalshi separately

Each venue is calibrated to about a point on its own: Polymarket's 19,282 markets were 0.9 points off on average, Kalshi's 20,399 1.1. Their Brier scores differ more — 0.085 for Polymarket, 0.134 for Kalshi — but that mostly reflects which questions each lists, not how good its prices are. Polymarket's sample has more long shots — 21% of its markets resolved YES, against 34% of Kalshi's — and a set of long shots is easier to score well on than a set of close questions, whoever prices them. That is why each venue's score is set against its own always-the-average baseline below.

By category

Sport dominates the sample: 26,484 of the 39,681 markets (67%) were sports markets, which is what the two venues list. Across the categories with enough markets to report, the calibration gap stayed within a few points. The share of favourites that won varies much more, but that measures how lopsided a category's questions were, not how well they were priced: most election markets a day before close are long shots on candidates who are about to lose, so the favourite wins almost every time.

A category with few events is noisier than its market count suggests, because the markets in one event — the candidates in one race, the price bands in one weather question — move together. Commodities, for example, is 329 markets from 27 events.

One market per event

The 39,681 markets came from 7,939 events. Many events are ladders of linked markets — one per candidate, score or price band — where one leg wins and the rest lose together, so a study that counts every leg could look calibrated only because the cheap legs are easy. Keeping a single market per event, the one with the most trading volume a day before close, the result holds: across those 7,939 markets the gap averaged 1.3 points, with a Brier score of 0.150 against 0.237 for always quoting the average.

A week before close

Fewer markets have a working price seven days out — 6,469 here, from 1,405 events — because many markets in the window were listed less than a week before they closed. The ones that remain are longer-running and mostly lopsided: 74% were priced under 10%. Their calibration gap was 1.4 points and the favourite won 94% of the time. Bands between 10% and 90% hold 80 to 330 markets each at this horizon, so their 95% ranges are several points wide.

What this does not show

It covers eight days of settlements, 28 September 2026 to 5 October 2026, and two venues. Results from a sports-heavy week need not hold for a month of elections or a year of long-dated questions; pdata keeps a market's price history for only ten days after its event closes, so each edition of this study can look back about a week, and later editions will add windows rather than lengthen this one.

It only scores working prices. Leaving the filter off makes Polymarket look badly calibrated — markets quoted between 40% and 50% a day before close resolved YES about a fifth of the time — because many sports side markets a day before close showed 49.5% on a book whose bid and ask were far apart. That is a quote nobody had acted on, not a forecast, and the study excludes it. The result therefore describes markets people were trading.

"A day before close" is measured from the time the venue closed the market, which is not always when the outcome became known. In this window Polymarket closed its markets a median of 3.5 hours after their scheduled end, so for a Polymarket market the price was typically taken about 20 hours before it was due to finish; for a market that was settled late, less.

Six venues are not in the pooled figures. Predict publishes no settlement time — pdata records the moment it first sees a result, which can be well after the match — and when a Predict market has no bids pdata stores its asking price, so an untraded long shot can read 99%. Gemini reports trading volume only for whole events, so there is no way to tell which of its contracts had traded. Limitless and Manifold record no bid–ask spread, and pdata does not record a winner for Myriad or Opinion markets. Of the 10,710 Predict markets that settled in the window, 1,423 had a working price a day out; they are listed in the data file but not pooled.