Current model · uk-monotonic-calibrated-spatial-v2

Model statistics and validation

These results compare the calibrated model with the previous HadUK-derived probability on 2021–2024 cell-Christmas results. Those four years were held back until the promotion decision, then added when fitting the operational version.

Brier skill

+36.42%

Reduction in mean squared probability error.

Held-out results

151,980

37,995 linked cells across four later Christmases.

ROC–AUC

0.961

Unchanged from the incumbent after monotonic calibration.

Calibration error

6.92%

Down from 10.21%.

Overall comparison

Lower is better for Brier score, log loss and calibration error. Higher is better for ROC–AUC.

ModelBrierLog lossROC–AUCCalibration error
Previous probability0.0216830.1302590.960610.21%
Current calibrated model0.0137860.0971970.96066.92%

Results by region

Brier skill remained positive in every standard UK region.

RegionBrier skill
Northern Ireland+54.24%
North West England+53.55%
Yorkshire and the Humber+52.37%
East Midlands+50.27%
East of England+49.41%
Wales+46.84%
West Midlands+45.80%
North East England+41.17%
South West England+38.03%
South East England+35.14%
London+34.27%
Scotland+25.39%

Spatial-block checks

The country was also divided into five broad geographic blocks. All five improved.

BlockCellsBrier skill
Block 17,560+30.67%
Block 29,430+37.64%
Block 38,849+27.61%
Block 45,103+47.03%
Block 57,053+47.08%

What the scorecard covers

Primary test

HadUK reconstruction

The 151,980 results are reconstructed falling-snow events from daily HadUK temperature and rainfall. They are not 151,980 physical snow gauges.

Spatial check

Observed snow lying

Against the observed December snow-lying field, spatial rank correlation is 0.728 across 37,523 cells.

External check

GHCN-D stations

The supplementary station set contains 430 Christmas snow-depth reports from 21 merged UK sites. It is independent but geographically sparse.

See How it works for the full method and limitations, or return to the forecast.