How close were prediction-market prices to outcomes? Calibration of Polymarket and Kalshi one day and one week before close

A day before they closed, Polymarket and Kalshi prices matched how often outcomes happened to within about one percentage point. We take 39,681 binary markets from 7,939 events that settled between 28 September and 5 October 2026, read each market's price 24 hours and 7 days before its venue closed it from a third-party record of quotes, and compare it with the venue's verdict. Across ten price bands, the share resolving YES differed from the band's average price by 0.96 points on average (market-weighted). The prices' Brier score was 0.110, against 0.199 for always quoting the sample's YES rate. The side priced above 50% won 83.4% of the time. The remaining error has the shape of the favourite–longshot bias: long shots resolved YES slightly less often than priced and favourites priced 80–90% slightly more often, by one to three points. Keeping one market per event gives a gap of 1.25 points. The result holds only for quotes from traded markets with a bid–ask spread of at most 10 points; without that filter Polymarket's 40–50% band resolved YES 22% of the time. We document this and six other data traps that change the answer if ignored. Prepared by pdata.world from its record of prediction-market quotes. Supplementary files: calibration.json (every band and group reported in the note) and coverage.json. Web version: pdata.world/research/prediction-market-accuracy. Drafted with the assistance of an AI model (Anthropic's Claude); numbers checked by script against the supplementary files.