Post

Models disagree by different amounts on different days. Does the disagreement tell you anything about how wrong the forecast ends up? 258 verified days, 21 cities, ERA5 truth. Seven models: ECMWF, GFS, UKMO, ICON, GEM, and the two AI runs, AIFS and AIGFS.

Sort the days by how much those seven disagreed one day out. The third that agreed most: the consensus missed by 1.27°F on average. The third that argued most: 2.38°. Same direction at 3, 5 and 7 days — 1.35x to 1.69x.

Before believing it: r = 0.37. Real, and loose. But it isn't one maverick — drop any single model and it holds, 1.48x without AIFS through 2.07x without UKMO. Spread is informative, not decisive.

The edges do have regulars. AIFS was the coolest of the seven on 71 days and the warmest on none; UKMO warmest on 47, coolest on 6. Bias runs AIFS −1.5°F to UKMO +1.1°, ECMWF nearest zero — and the median beat all seven, 1.78° against ECMWF's 1.83.