WeatherBlend

Multi-model forecast blending for Bonehill Rocks, Dartmoor

Models

Held-out test performance for each active blender, against ERA5 (temperature) or EA rainfall (precipitation, dry window). Temperature scores are MAE in °C; precipitation and dry-window scores are Brier — lower is better in both cases. Test slices are time-split (walk-forward): train → validation → test, in that order, with no data leakage across the boundary.

Temperature

Metric: Test MAE (°C).

Version Phase Trained Lead Blend score RMSE Bias Best single Best-single val Test rows
v2026-04-23_132453_phase2c 2c +24h 0.508 0.692 +0.10 temp_ecmwf 0.711 2219
+48h 0.646 0.853 +0.23 temp_icon 0.830 2214
+72h 0.793 1.112 +0.28 temp_icon 0.924 2206
v2026-04-21_201231_phase2redo 2b_redo +24h 0.560 0.789 +0.14 temp_ecmwf 0.711 2219
+48h 0.675 0.912 +0.19 temp_icon 0.830 2214
+72h 0.805 1.125 +0.24 temp_icon 0.924 2206

Precipitation — Bellever Dartmoor

Metric: Test Brier.

Version Phase Trained Lead Blend score RMSE Bias Best single Best-single val Test rows
v2026-04-23_154405_phase3c 3c +24h 0.130 0.215 +0.97 precip_mf 0.207 2959
+48h 0.148 0.215 +0.89 precip_ecmwf 0.226 2955
+72h 0.158 0.215 +0.87 precip_ecmwf 0.253 2951
v2026-04-23_071842 3a +24h 0.136 0.215 +0.81 precip_mf 0.207 2959
+48h 0.163 0.215 +0.71 precip_ecmwf 0.226 2955
+72h 0.171 0.215 +0.69 precip_ecmwf 0.253 2951

Precipitation — Dartmoor Nr Hexworthy

Metric: Test Brier.

Version Phase Trained Lead Blend score RMSE Bias Best single Best-single val Test rows
v2026-04-23_154459_phase3c 3c +24h 0.149 0.230 +0.91 precip_mf 0.214 2961
+48h 0.163 0.229 +0.86 precip_ecmwf 0.217 2958
+72h 0.176 0.229 +0.85 precip_ecmwf 0.253 2954

Precipitation — Princetown

Metric: Test Brier.

Version Phase Trained Lead Blend score RMSE Bias Best single Best-single val Test rows
v2026-04-23_154433_phase3c 3c +24h 0.136 0.210 +0.89 precip_mf 0.214 2962
+48h 0.152 0.210 +0.79 precip_ecmwf 0.223 2959
+72h 0.160 0.210 +0.76 precip_gfs 0.255 2955
v2026-04-23_071934 3a +24h 0.137 0.210 +0.78 precip_mf 0.214 2962
+48h 0.160 0.210 +0.69 precip_ecmwf 0.223 2959
+72h 0.171 0.210 +0.66 precip_gfs 0.255 2955

Dry window — Bellever Dartmoor — 3-hour

Metric: Test Brier.

Version Phase Trained Lead Blend score RMSE Bias Best single Best-single val Test rows
v2026-04-23_192737_phase3d_shape 3d-shape +24h 0.056 0.082 +1.05 has_dry_window_mf 0.032 127
+48h 0.071 0.082 +1.11 has_dry_window_gfs 0.056 127
+72h 0.070 0.082 +1.09 has_dry_window_gfs 0.056 127
v2026-04-23_192259_phase3d_calibrated 3d-calibrated +24h 0.058 0.082 +1.04 — 0.000 127
+48h 0.073 0.082 +1.11 — 0.000 127
+72h 0.063 0.082 +1.11 — 0.000 127
v2026-04-23_101107 3b +24h 0.057 0.082 +1.05 has_dry_window_mf 0.032 127
+48h 0.067 0.082 +1.07 has_dry_window_gfs 0.056 127
+72h 0.069 0.082 +1.11 has_dry_window_gfs 0.056 127

Dry window — Bellever Dartmoor — 4-hour

Metric: Test Brier.

Version Phase Trained Lead Blend score RMSE Bias Best single Best-single val Test rows
v2026-04-23_192825_phase3d_shape 3d-shape +24h 0.075 0.130 +1.10 has_dry_window_ukmo 0.048 127
+48h 0.104 0.130 +1.15 has_dry_window_mf 0.072 127
+72h 0.112 0.130 +1.17 has_dry_window_mf 0.072 127
v2026-04-23_192345_phase3d_calibrated 3d-calibrated +24h 0.075 0.130 +1.06 — 0.000 127
+48h 0.111 0.130 +1.18 — 0.000 127
+72h 0.115 0.130 +1.18 — 0.000 127
v2026-04-23_101150 3b +24h 0.076 0.130 +1.10 has_dry_window_ukmo 0.048 127
+48h 0.103 0.130 +1.16 has_dry_window_mf 0.072 127
+72h 0.114 0.130 +1.18 has_dry_window_mf 0.072 127

Dry window — Bellever Dartmoor — 6-hour

Metric: Test Brier.

Version Phase Trained Lead Blend score RMSE Bias Best single Best-single val Test rows
v2026-04-23_192908_phase3d_shape 3d-shape +24h 0.130 0.200 +1.19 has_dry_window_ecmwf 0.128 127
+48h 0.162 0.200 +1.23 has_dry_window_ecmwf 0.152 127
+72h 0.159 0.200 +1.24 has_dry_window_gfs 0.200 127
v2026-04-23_192428_phase3d_calibrated 3d-calibrated +24h 0.111 0.200 +1.20 — 0.000 127
+48h 0.143 0.200 +1.03 — 0.000 127
+72h 0.153 0.200 +1.12 — 0.000 127
v2026-04-23_101214 3b +24h 0.131 0.200 +1.17 has_dry_window_ecmwf 0.128 127
+48h 0.158 0.200 +1.21 has_dry_window_ecmwf 0.152 127
+72h 0.160 0.200 +1.21 has_dry_window_gfs 0.200 127

Dry window — Princetown — 3-hour

Metric: Test Brier.

Version Phase Trained Lead Blend score RMSE Bias Best single Best-single val Test rows
v2026-04-23_192951_phase3d_shape 3d-shape +24h 0.047 0.069 +1.06 has_dry_window_ukmo 0.032 127
+48h 0.053 0.069 +1.05 has_dry_window_icon 0.056 127
+72h 0.061 0.069 +1.09 has_dry_window_gfs 0.063 127
v2026-04-23_192510_phase3d_calibrated 3d-calibrated +24h 0.053 0.069 +1.05 — 0.000 127
+48h 0.053 0.069 +1.06 — 0.000 127
+72h 0.056 0.069 +1.09 — 0.000 127
v2026-04-23_101238 3b +24h 0.047 0.069 +1.06 has_dry_window_ukmo 0.032 127
+48h 0.052 0.069 +1.05 has_dry_window_icon 0.056 127
+72h 0.060 0.069 +1.09 has_dry_window_gfs 0.063 127

Dry window — Princetown — 4-hour

Metric: Test Brier.

Version Phase Trained Lead Blend score RMSE Bias Best single Best-single val Test rows
v2026-04-23_193036_phase3d_shape 3d-shape +24h 0.067 0.099 +1.07 has_dry_window_ukmo 0.087 127
+48h 0.079 0.099 +1.10 has_dry_window_icon 0.095 127
+72h 0.088 0.099 +1.06 has_dry_window_mf 0.111 127
v2026-04-23_192554_phase3d_calibrated 3d-calibrated +24h 0.072 0.099 +1.03 — 0.000 127
+48h 0.087 0.099 +0.86 — 0.000 127
+72h 0.074 0.099 +1.13 — 0.000 127
v2026-04-23_101310 3b +24h 0.067 0.099 +1.06 has_dry_window_ukmo 0.087 127
+48h 0.077 0.099 +1.10 has_dry_window_icon 0.095 127
+72h 0.089 0.099 +1.08 has_dry_window_mf 0.111 127

Dry window — Princetown — 6-hour

Metric: Test Brier.

Version Phase Trained Lead Blend score RMSE Bias Best single Best-single val Test rows
v2026-04-23_193122_phase3d_shape 3d-shape +24h 0.117 0.183 +1.14 has_dry_window_ecmwf 0.135 127
+48h 0.136 0.183 +1.17 has_dry_window_ecmwf 0.159 127
+72h 0.148 0.183 +1.16 has_dry_window_gfs 0.198 127
v2026-04-23_192636_phase3d_calibrated 3d-calibrated +24h 0.115 0.183 +1.09 — 0.000 127
+48h 0.136 0.183 +1.18 — 0.000 127
+72h 0.143 0.183 +1.12 — 0.000 127
v2026-04-23_101336 3b +24h 0.124 0.183 +1.15 has_dry_window_ecmwf 0.135 127
+48h 0.132 0.183 +1.15 has_dry_window_ecmwf 0.159 127
+72h 0.150 0.183 +1.22 has_dry_window_gfs 0.198 127