Calibrated but Not Sharp: Stated Probabilities in Five-Minute Crypto Markets on Polymarket

Luis Phillip Gruener
Independent

Abstract
When a prediction market shows a high chance, does the event happen that often, or is the price fooling you? This paper studies that for five-minute, USDC-denominated up/down contracts on Polymarket, across Bitcoin, Ether, Solana, and Dogecoin, and reports three findings. First, the prices are calibrated. Using a clean logger that records the implied probability of every active market at a fixed moment about a minute before close, whether or not anyone traded it, I obtain 14,808 contracts over sixteen days, and stated probabilities match realized frequencies to within about one percentage point across the range, including in the extreme tails. The slope of realized outcomes on stated probability is 1.01, and three small-sample methods built for sixteen day-clusters, a wild-cluster bootstrap, an Ibragimov-Muller test, and randomization inference, all fail to reject the value of 1.0; the intervals are wide enough that only a favorite-longshot bias larger than a few percent in slope terms is ruled out. This is reassuring but unsurprising, since a liquid market should price roughly like this. Second, and more useful as method, a separate large but selected dataset of about 209,000 trade entries shows how easily this can be gotten wrong: treating those correlated trades as independent manufactures a statistically significant favorite-longshot bias that vanishes under dependence-robust inference and is absent from the clean data. Third, calibration is not sharpness. The posted probabilities hug fifty-fifty even when the near-term outcome is largely decided. A costless rule that asks only which side is ahead at the snapshot calls the final direction about 81 percent of the time against the market's 64 percent, with more than double the Brier resolution, and this survives rebuilding the rule on a second, independent exchange feed. I therefore report the discrimination of the posted quote as low relative to freely available information, and stop short of calling it trader underreaction, because the logged value is a non-executable quote on a thin market and I have no spread or order-book data to rule out microstructure. Two limits bound everything: sixteen days on a single venue, and a deliberately extreme five-minute horizon.

Keywords: prediction markets, calibration, sharpness, favorite-longshot bias, market efficiency, cryptocurrency, Polymarket, cluster-robust inference

JEL Classification: G14, G12, C12, D84

Suggested Citation:

Gruener, Luis Phillip, Calibrated but Not Sharp: Stated Probabilities in Five-Minute Crypto Markets on Polymarket (June 01, 2026). Available at SSRN: https://ssrn.com/abstract=6863546 or http://dx.doi.org/10.2139/ssrn.6863546