Which weather model is most accurate in your city?

What one scored day looks like

The most recent day we settled, whichever way it went — not the one that flatters the models.

14 days out, 4 models put Oklahoma City's Thu Sep 17 high anywhere from 74° to 99°; by the final call they had settled to a median of 97°. It hit 99° — ICON landed closest, off 0.3°.

Each bar is the full range of the models still forecasting that day, in °F; the rule is what the OKC observations recorded. This one opened 25° wide and closed to 4°. Across every day we have scored, the range narrows even though more models arrive as the day approaches — a median 11.6° at 13 days out with four models, 5.8° on the day with seven. See this day's full receipt →

More scored days

The next-most-recent settled days across the archive — records, not rankings: each bar is the models' opening range, the rule is what happened.

• Washington · Thu Sep 17

  4 models opened 81–89° at 13 days out; it hit 87° — GEM closest, off 0.2°.

  Full receipt →

• Tampa · Thu Sep 17

  4 models opened 86–94° at 13 days out; it hit 92° — AIGFS closest, off 0.1°.

  Full receipt →

• St. Louis · Thu Sep 17

  4 models opened 71–109° at 13 days out; it hit 99° — GEM closest, off 0.4°.

  Full receipt →

• Seattle · Thu Sep 17

  4 models opened 66–77° at 14 days out; it hit 77° — GFS closest, off 1.0°.

  Full receipt →

How the scoring works

Each model's 1-, 3-, and 7-day-ahead forecasts are captured from its actual archived runs (nothing is reconstructed after the fact), then verified against each city's truth — its own station's observations where one is pinned, the ECMWF IFS analysis otherwise — once the day has happened. Bias is the mean signed error of the predicted daily high — positive means the model ran warm. The typical miss (MAE) is the average absolute error. Cities join these pages as they join The Spread's archive.

Two things you should know about the yardstick. Where a city is pinned to its own station, models are verified against that thermometer's observed highs outright, and the city page names the station. Everywhere else, truth is the ECMWF IFS analysis — ECMWF's operational product, a grid rather than a thermometer, while two of the seven models scored here are ECMWF's own. Every registry city has been pinned to its station since 2026-08-22, so the grid is the yardstick only for places no thermometer serves, and those pages say so.

The Spread itself is in closed beta — these verification pages and the Updates log are public while the rest of the product is gated. If you want in, ask for an invite — say where you forecast and I will send you one.