Raw model output on the native 0.25° grid — no MOS, no bias correction, no post-processing. Scores understate operational forecast quality. Read the fairness statement →

CastCheck methodology v0.2 · data through 2026-08-30

All stations (mean of daily station errors) · GraphCast (GFS init) · lead day 0

ALL · graphcast_gfs · lead 0

Target day = initialization date + 0 days. This page is a permanent link: the numbers change as history accumulates, the address and the methodology version do not.

Scores use the latest model version segment, GRAP_v100, starting 2026-08-28 — days before that ran a different version of the model and are excluded rather than silently pooled. Of the scored days, 2 carry a QC flag on the observation.

JSON for this page CSV of the daily errors Cite this page

Every number for this combination

All four windows × both initializations × both interpolations × both variables. Errors are computed in °C and shown in °F; CI is a 95 % moving-block bootstrap interval. MAE debiased and skill, debiased remove the constant offset of the window first, so a station that the model always reads a degree too warm is not scored as if it had no skill — compare the two skill columns to see how much of the error is a fixed offset. n QC next to n counts days whose observation carries a QC flag.

VariableInitMethod Windown MAE °F MAE debiased Bias °F RMSE °F ±1 °F±2 °F±3 °F Skill Skill, debiased Period
tmax00Zbilinear30d2 2 QC5.60 [5.35, 5.86]0.26-5.60 [-5.86, -5.35]6.25 [5.93, 6.56]4% [0%, 9%]9%13%——2026-08-29 → 2026-08-30
tmax00Zbilinear90d2 2 QC5.60 [5.35, 5.86]0.26-5.60 [-5.86, -5.35]6.25 [5.93, 6.56]4% [0%, 9%]9%13%——2026-08-29 → 2026-08-30
tmax00Zbilinear365d2 2 QC5.60 [5.35, 5.86]0.26-5.60 [-5.86, -5.35]6.25 [5.93, 6.56]4% [0%, 9%]9%13%——2026-08-29 → 2026-08-30
tmax00Zbilinearall2 2 QC5.60 [5.35, 5.86]0.26-5.60 [-5.86, -5.35]6.25 [5.93, 6.56]4% [0%, 9%]9%13%——2026-08-29 → 2026-08-30
tmax00Znearest30d2 2 QC5.61 [5.34, 5.87]0.27-5.61 [-5.87, -5.34]6.41 [6.06, 6.75]9% [9%, 9%]11%17%——2026-08-29 → 2026-08-30
tmax00Znearest90d2 2 QC5.61 [5.34, 5.87]0.27-5.61 [-5.87, -5.34]6.41 [6.06, 6.75]9% [9%, 9%]11%17%——2026-08-29 → 2026-08-30
tmax00Znearest365d2 2 QC5.61 [5.34, 5.87]0.27-5.61 [-5.87, -5.34]6.41 [6.06, 6.75]9% [9%, 9%]11%17%——2026-08-29 → 2026-08-30
tmax00Znearestall2 2 QC5.61 [5.34, 5.87]0.27-5.61 [-5.87, -5.34]6.41 [6.06, 6.75]9% [9%, 9%]11%17%——2026-08-29 → 2026-08-30
tmin00Zbilinear30d2 2 QC2.78 [2.76, 2.80]0.80+0.58 [-0.22, 1.37]3.64 [3.57, 3.70]26% [26%, 26%]54%67%——2026-08-29 → 2026-08-30
tmin00Zbilinear90d2 2 QC2.78 [2.76, 2.80]0.80+0.58 [-0.22, 1.37]3.64 [3.57, 3.70]26% [26%, 26%]54%67%——2026-08-29 → 2026-08-30
tmin00Zbilinear365d2 2 QC2.78 [2.76, 2.80]0.80+0.58 [-0.22, 1.37]3.64 [3.57, 3.70]26% [26%, 26%]54%67%——2026-08-29 → 2026-08-30
tmin00Zbilinearall2 2 QC2.78 [2.76, 2.80]0.80+0.58 [-0.22, 1.37]3.64 [3.57, 3.70]26% [26%, 26%]54%67%——2026-08-29 → 2026-08-30
tmin00Znearest30d2 2 QC2.83 [2.79, 2.87]0.85+0.50 [-0.35, 1.35]3.78 [3.70, 3.86]26% [26%, 26%]52%65%——2026-08-29 → 2026-08-30
tmin00Znearest90d2 2 QC2.83 [2.79, 2.87]0.85+0.50 [-0.35, 1.35]3.78 [3.70, 3.86]26% [26%, 26%]52%65%——2026-08-29 → 2026-08-30
tmin00Znearest365d2 2 QC2.83 [2.79, 2.87]0.85+0.50 [-0.35, 1.35]3.78 [3.70, 3.86]26% [26%, 26%]52%65%——2026-08-29 → 2026-08-30
tmin00Znearestall2 2 QC2.83 [2.79, 2.87]0.85+0.50 [-0.35, 1.35]3.78 [3.70, 3.86]26% [26%, 26%]52%65%——2026-08-29 → 2026-08-30

+ model too warm · − model too cold · bias interval includes zero · ★ lowest MAE · = not distinguishable from the leader · ▼ significantly worse · n < 30 greyed and unranked · CI is a 95 % moving-block bootstrap interval.

Daily error, last 90 days

Daily forecast minus observed error in °F for GraphCast (GFS init) at ALL, lead day 0, 00Z bilinear tmax+10+50-5-102026-08-29: -5.86 °F2026-08-30: -5.35 °F2026-08-292026-08-30°F, forecast − observed
Forecast − observed, °F, 00Z bilinear daily maximum. The dashed line is zero; the thick line is a centred 7-day mean, drawn only where the whole window is present. Gaps are days with no score, never interpolation. Hover a point for its date and value.
The same series as a table (2 days, newest first) · download CSV
Climatological dayError °F
2026-08-30-5.35
2026-08-29-5.86

Error distribution

Distribution of the signed daily error in °F for GraphCast (GFS init) at ALL, lead day 0-10 to -9 °F: 0 days-9 to -8 °F: 0 days-8 to -7 °F: 0 days-7 to -6 °F: 0 days-6 to -5 °F: 2 days-5 to -4 °F: 0 days-4 to -3 °F: 0 days-3 to -2 °F: 0 days-2 to -1 °F: 0 days-1 to 0 °F: 0 days0 to 1 °F: 0 days1 to 2 °F: 0 days2 to 3 °F: 0 days3 to 4 °F: 0 days4 to 5 °F: 0 days5 to 6 °F: 0 days6 to 7 °F: 0 days7 to 8 °F: 0 days8 to 9 °F: 0 days9 to 10 °F: 0 daysdashed: P50 |e| 5.6 · P90 |e| 5.81 °F-100+10 °Fn = 2 days
All scored days for 00Z bilinear daily maximum, 20 equal bins. The dashed marks are the 50th and 90th percentiles of the absolute error, mirrored on both sides. Hover a bar for its bin and count.
Bin counts as a table
From °FTo °FDays
-10.00-9.000
-9.00-8.000
-8.00-7.000
-7.00-6.000
-6.00-5.002
-5.00-4.000
-4.00-3.000
-3.00-2.000
-2.00-1.000
-1.00+0.000
+0.00+1.000
+1.00+2.000
+2.00+3.000
+3.00+4.000
+4.00+5.000
+5.00+6.000
+6.00+7.000
+7.00+8.000
+8.00+9.000
+9.00+10.000

Month by month

00Z bilinear daily maximum, newest first.

Monthn MAE °FBias °F
2026-0825.60 -5.60

Availability

ModelScored period DaysCoverage
GraphCast (GFS init)2026-08-29 → 2026-08-30 2
GraphCast (GFS init)2026-08-29 → 2026-08-30 2

Compared with other models

Paired bootstrap on the days both models cover (90d, 00Z, bilinear, daily maximum). Negative means GraphCast (GFS init) has the lower MAE. "Distinguishable" means the 95 % interval excludes zero.

Other modelCommon days MAE difference °FDistinguishable
ECMWF AIFS Single 2 +2.11 [2.01, 2.20] yes
Aurora (GFS init) 2 -0.20 [-0.21, -0.18] yes
Aurora (IFS init) 2 +0.62 [0.35, 0.90] yes
FourCastNet v2 (GFS init) 2 +1.24 [1.13, 1.35] yes
FourCastNet v2 (IFS init) 2 +1.42 [1.39, 1.45] yes
NCEP GFS 2 +2.67 [2.57, 2.78] yes
GraphCast (IFS init) 2 +0.66 [0.42, 0.90] yes
ECMWF IFS HRES 2 +1.93 [1.66, 2.20] yes
Pangu-Weather (GFS init) 2 +1.86 [1.73, 1.99] yes
Pangu-Weather (IFS init) 2 +1.94 [1.93, 1.95] yes

Cite this page

ShortCastCheck, ALL graphcast_gfs lead 0, methodology v0.2, accessed 2026-08-31, https://castcheck.zifanzhang.com/station/ALL/model/graphcast_gfs/lead/0/

LongCastCheck (2026). Station-level verification of raw weather-model 2 m temperature forecasts: ALL (All stations (mean of daily station errors)), model graphcast_gfs, lead day 0. Methodology version 0.2, schema version 0.1, data through 2026-08-30. Accessed 2026-08-31. https://castcheck.zifanzhang.com/station/ALL/model/graphcast_gfs/lead/0/

Accessed2026-08-31 (data through 2026-08-30; page built 2026-08-31T03:52:06+00:00)

JSON CSV all scores methodology v0.2 licence