GraphCast vs IFS HRES, verified against ERA5

WeatherBench 2, 2020: 732 initialisations (00 and 12 UTC), lead times 1-10 days, 5.625° grid

spatial_region
GraphCast vs IFS HRES, verified against ERA5
globaln.extratropicstropicss.extratropicseurope
rmset2m
msl
ws10

How to read this

Each cell is one comparison across forecast lead time, earliest on the left: 10 boxes from T+24 to T+240. Colour shows the difference between graphcast and ifs-hres, and its intensity the size of that difference relative to ifs-hres. Hover a box for the exact value, or click a cell for the full series.

error metrics: worse no change better

activity metrics: less active no change more active

A framed box is significant — its 68% interval on the difference excludes zero. An unframed box is not. 149 of 150 boxes are framed. The frame is one bit; hover a box for the strongest level it reaches, up to 99.7%.

A blank grey cell has no data at all, which is deliberately distinct from a cell whose difference happens to be zero. A hatched box is a lead time with no value inside a cell that otherwise has data.

What this card cannot tell you

Intervals from a moving-block bootstrap over forecast cases (blocks of 20 cases, 2000 resamples, seed 0). The difference is paired: it is taken per case before averaging, with one resample shared by both sources, so their common error cancels.
Consecutive forecasts share a weather system. Blocking reduces the resulting over-marking but does not remove it: measured at roughly 8% false positives against a nominal 5% on AR(1) synthetic data, and no block length reaches nominal.
This card shows 150 simultaneous comparisons. Isolated cells mean little; coherent blocks mean a lot.
99% of boxes are marked significant. That is high enough to be worth double-checking how the interval was resampled, rather than read as 99% confidence in the result.

3 rows x 5 columns; 15 of 15 crossings populated; 150 boxes. Drill-down data 7 kB (1.9x compressed).