Loading…
Every source's real, saved forecast against the bracket Polymarket settled on. Snapshots are taken every hour on GitHub's servers and never edited afterward.
The newest snapshot: what every source says for each upcoming market, next to the crowd's favorite. Rows where Sunny points at a different bracket than the crowd are marked.
*Morning-of backtest numbers are flattered: the model archive can include runs issued after 8 AM local. Trust the day-before figures most.
Each Polymarket market asks for the highest (or lowest) temperature at one named station, usually an airport, over one calendar day in that city's time, midnight to midnight. A source is right when its number, rounded the way the market rounds (nearest whole degree, halves up), falls inside the bracket that won. Within one also counts the bracket next to it.
Day before = each source's latest snapshot taken at or before 12:00 noon local time the day before. For a Tuesday market in Chicago that's Monday noon Chicago time: 12 hours before the day starts and 36 hours before it ends. The snapshot used is never more than 18 hours older than that noon.
2 days before = the same rule at noon two days before (36 hours before the day starts, 60 before it ends).
Every source is judged from the same snapshot, so nobody gets a fresher forecast than anyone else. Only days that hadn't started yet are scored, because hour-by-hour forecasts only cover the hours that are left.
Sunny: the exact number the app shows (the app's own server). Apple: WeatherKit, the Weather app's data, daily high and low. Google: Google Weather API, hour by hour, turned into the calendar-day high and low (Google's own daily low runs 7 AM to 7 AM, so it isn't used). AccuWeather: through Microsoft Azure Maps, hour by hour. National Weather Service: api.weather.gov hourly forecast (US cities). German Weather Service (DWD): its station forecast for the exact settlement airport (45 of 53 cities). Hong Kong Observatory: its official 9-day forecast (Hong Kong settles at the Observatory). Yr, Open-Meteo, Tomorrow.io, WeatherAPI, Visual Crossing, OpenWeather: hour by hour. Every source is asked about the settlement station's own coordinates.
The crowd: the highest-priced bracket on Polymarket in that same snapshot, only when the bracket prices add up to about 100% (quiet markets show placeholder prices and are skipped).
Apple, Google and most apps keep no record of past forecasts, so they're scored only from the day logging began (Nov 10, 2026). Weather.com, Weather Underground and TV stations have no public feed. The 3-week backtest uses the archives that do exist: the weather models, the National Weather Service's issued forecasts, and Polymarket's price history.