Accountability · Model skill · 30 days

How well do the models predict UK chase days?

Every day, chaseit captures forecasts from four models (UKV, ECMWF IFS, DWD ICON-EU, GFS-ENS). Every day after, we score those forecasts against what actually happened. Here’s the full record — public and unedited. The same data drives the skill caveat on /outlook.

Models

0–1 day lead · per-region · verified vs observed

UKV
41%exact
89% within ±1
16 days · 194 region-days
calls a fraction of a category too high on average
↑ 90-day · 4 weeks
best: south_east 53%
worst: north_west 26%
ECMWF IFS
64%exact
99% within ±1
19 days · 278 region-days
calls a fraction of a category too high on average
↓ 90-day · 4 weeks
best: london 78%
worst: north_west 56%
DWD ICON-EU
87%exact
100% within ±1
19 days · 278 region-days
calls category roughly right on average
↓ 90-day · 4 weeks
best: wales 100%
worst: east_midlands 78%
GFS-ENS most-likely
39%exact
100% within ±1
9 days · 178 region-days
calls a fraction of a category too high on average
↓ 90-day · 2 weeks
best: wales 41%
worst: south_east 35%
Met Office DataHub
100%exact
100% within ±1
8 days · 160 region-days
calls category roughly right on average
→ 90-day · 2 weeks
best: east_anglia 100%
worst: yorkshire 100%

How this is computed

Each model produces a categorical forecast (none through high) for every chase region, every six hours. The day after each forecast date we look up what actually happened in each region: the highest-severity public report submitted by chasers, combined with any Met Office warning in force.

We score each region’s prediction against that observed level on a three-step scale. hit_score = 1.0 for an exact category match. 0.5 for a forecast within one category of observed. 0.0for anything further off. Hit rate above is the share of region-days that scored 1.0; partial rate is the share that scored 0.5 or better. Bias is the mean difference between predicted and observed rank — positive means the model over-called.

Methodology caveats

·No reports doesn’t mean no severe weather — a quiet day with nobody chasing scores like a true quiet day.
·Met Office warnings are domain-wide; observed-level scoring per region is approximate.
·Backfilled IFS and ICON data uses Open-Meteo historical-forecast; UKV history is from prism’s stormfront snapshots.

Raw data

The full machine-readable record is at /api/public/skill. CORS-open JSON, 30-day window, no authentication. Cached for five minutes.

How this is computed

Methodology caveats

·No reports doesn’t mean no severe weather — a quiet day with nobody chasing scores like a true quiet day.

·Met Office warnings are domain-wide; observed-level scoring per region is approximate.

·Backfilled IFS and ICON data uses Open-Meteo historical-forecast; UKV history is from prism’s stormfront snapshots.