Weather Market Probabilities — Polymarket temperature brackets

For every active Polymarket daily high/low temperature market, this actor returns the market bracket, a model probability for that bracket, the market's best ask, and the difference, in one row per bracket. It is a research and monitoring dataset for people who trade, study or report on weather prediction markets.

What you get

One row per bracket market: slug, question, city, date, kind (high/low), bracket (e.g. 34C, 76-77F, ≥84F), low, high, unit, bounds, forecastC, sigmaC, modelPrimary, modelCrossVal, divergenceC, ensembleStdC, ensembleMembers, probability, probabilityCalibrated, marketYesPrice, bestBid, bestAsk, diffVsAsk, liquidity, endDate, stationNote, url, scrapedAt.

How the probability is computed

1. Forecast: an hourly temperature series for the market date from a per-city regional model (HRRR for US cities, ICON-D2/EU for Germany and Europe, JMA MSM for Japan, ECMWF IFS elsewhere), reduced to the daily max (high markets) or min (low markets) in the city's local day.
2. Uncertainty: the ECMWF ensemble (50 members) spread for the same day, floored at 2.0 °C, inflated by 15 % for max/min markets and further when a cross-validation model disagrees by more than 1 °C.
3. Bracket probability: normal distribution with continuity correction at bracket edges; open-ended brackets ("X or higher") use the half-line form so mass is not truncated.
4. probabilityCalibrated: a Platt correction fitted on 3,301 resolved markets (a = 0.528, b = −0.297). Raw model probabilities were over-confident by 17–20 percentage points on our own historical bets; the calibrated column is the one to compare with prices.

Calibration note (read before using diffVsAsk)

On 36,778 resolved Polymarket weather markets, Yes prices sit within ~1.7 points of realised frequency and a group's Yes prices sum to ~1.08. A positive diffVsAsk is therefore usually the market's margin, not an edge. This dataset gives you a documented, reproducible model probability next to the price; it is not betting advice and does not claim an edge.

Input

• cities (optional): city slugs as in market URLs; empty = all.
• dates (optional): YYYY-MM-DD; empty = all active market dates.
• maxItems: stop after this many bracket rows.
• includeKalshi: off by default (Kalshi Data Terms restrict commercial use).

Pricing

Event	Price
result	$0.005 per bracket row ($5 per 1,000)

Charged only for rows with a computed probability. If a city has no coordinates or the forecast source fails, its markets are skipped and counted in the run summary.

Limits

• Forecast is a grid point at the city coordinates (LaGuardia for NYC), not the exact reporting station; stationNote states the local-day resolution rule.
• Typically 200–300 active bracket markets across ~20 cities; one forecast bundle (3 requests) per city/date, one order-book request per market.
• Open-Meteo commercial plan key is required for commercial use of the forecast source (set the OPEN_METEO_API_KEY secret); the free endpoints are for evaluation only.

Changelog

• 0.1: initial release — Polymarket temperature markets, regional models + ECMWF ensemble sigma, Platt calibration, CLOB best ask.