Multi-model forecast blending for Bonehill Rocks, Dartmoor
Forecast vs truth
Side-by-side view of what each active blender predicted vs what actually happened.
Temperature lines run forward past "now" because the blend projects 24/48/72h ahead;
truth (ERA5, METAR, rainfall) stops at the last observation.
Temperature — Phase 2b lean (13 features)
Six per-model temperatures, their mean/std/range, and cyclical hour/day-of-year encodings. The original champion.
Temperature — Phase 2c rich (88 features)
Adds per-model dew point, RH, cloud {total/low/mid/high}, wind speed/dir/gusts, surface pressure, plus cross-model aggregates. Trained to challenge 2b.
Temperature — other versions
Versions with no training metadata on disk — typically pre-2b experiments left in the manifest.
Precipitation — P(wet) vs observed rainfall
P(wet) is the blender's probability that the next hour sees ≥ 0.1 mm. The rainfall line is the same 4-of-4 hourly aggregation used for verification — partial hours are dropped to avoid flipping wet↔dry at the boundary.
Bellever Dartmoor
Phase 3a (lean)
Phase 3c (rich)
Three-way comparison — +24h lead
Dartmoor Nr Hexworthy
Phase 3c (rich)
Princetown
Phase 3a (lean)
Phase 3c (rich)
Three-way comparison — +24h lead
Dry window — predicted probability vs (eventual) observation
One row per target UTC day × window length. The "observed" column is blank for dates beyond the last full rainfall day. Rolling Brier skill scores are on the verification page.