WeatherBlend

Multi-model forecast blending for Sennen, Cornwall

Dry-window forecast

P(at least one N-hour dry block in 09:00–18:00 local). 3b champion plus Monte-Carlo challengers.

Trengwainton

2-hour dry window

Phase 3q — Gaussian copula MC over Phase 3c hourly P(wet)

Gaussian copula MC over Phase 3c's hourly P(wet) marginals — the 3p engine bound to the rich (non-orographic) 3c instead of 3o, for coastal locations without a 3o model. Single empirical Σ per station from train-split observed daytime wet/dry sequences. Captures within-day wet/dry autocorrelation an iid sampler misses; scored against the 3b champion at Sennen.

Target date (UTC) +24h +48h +72h Confidence Best start (UTC, P dry block)
100%100%99% confident dry (P=100% · run 4–9h) 13:00Z (96%)
100%100%100% confident dry (P=100% · run 7–9h) 12:00Z (97%)
—89%84% confident dry (P=89% · run 1–9h) 15:00Z (72%)
——99% confident dry (P=99% · run 4–9h) 08:00Z (90%)
Monte-Carlo P(an 2-hour dry block runs from each start hour) — one line per forecast horizon, soonest target day shown. The peak is the best time to set off; the height is how likely it stays dry. Each hour is its own probability, so the lines need not sum to the daily "any dry window" figure.

3-hour dry window

Phase 3b — lean (53 features)

53-feature LightGBM. Champion.

Target date (UTC) +24h +48h +72h Model agreement Confidence
98%99%99% 100% confident dry (P=98%)
98%99%99% 100% confident dry (P=98%)
—88%87% 50% confident dry (P=88%)
——99% 100% confident dry (P=99%)
Phase 3q — Gaussian copula MC over Phase 3c hourly P(wet)

Gaussian copula MC over Phase 3c's hourly P(wet) marginals — the 3p engine bound to the rich (non-orographic) 3c instead of 3o, for coastal locations without a 3o model. Single empirical Σ per station from train-split observed daytime wet/dry sequences. Captures within-day wet/dry autocorrelation an iid sampler misses; scored against the 3b champion at Sennen.

Target date (UTC) +24h +48h +72h Confidence Best start (UTC, P dry block)
98%98%97% confident dry (P=98% · run 4–9h) 13:00Z (94%)
100%99%98% confident dry (P=100% · run 7–9h) 11:00Z (95%)
—72%65% ambiguous (P=72% · run 1–9h) 14:00Z (59%)
——97% confident dry (P=97% · run 4–9h) 08:00Z (86%)
Monte-Carlo P(an 3-hour dry block runs from each start hour) — one line per forecast horizon, soonest target day shown. The peak is the best time to set off; the height is how likely it stays dry. Each hour is its own probability, so the lines need not sum to the daily "any dry window" figure.

4-hour dry window

Phase 3b — lean (53 features)

53-feature LightGBM. Champion.

Target date (UTC) +24h +48h +72h Model agreement Confidence
99%98%98% 100% confident dry (P=99%)
99%98%97% 100% confident dry (P=99%)
—71%80% 33% ambiguous (P=71%)
——97% 83% confident dry (P=97%)
Phase 3q — Gaussian copula MC over Phase 3c hourly P(wet)

Gaussian copula MC over Phase 3c's hourly P(wet) marginals — the 3p engine bound to the rich (non-orographic) 3c instead of 3o, for coastal locations without a 3o model. Single empirical Σ per station from train-split observed daytime wet/dry sequences. Captures within-day wet/dry autocorrelation an iid sampler misses; scored against the 3b champion at Sennen.

Target date (UTC) +24h +48h +72h Confidence Best start (UTC, P dry block)
95%95%94% confident dry (P=95% · run 4–9h) 13:00Z (93%)
98%97%96% confident dry (P=98% · run 7–9h) 10:00Z (94%)
—56%44% ambiguous (P=56% · run 1–9h) 13:00Z (47%)
——92% confident dry (P=92% · run 4–9h) 08:00Z (80%)
Monte-Carlo P(an 4-hour dry block runs from each start hour) — one line per forecast horizon, soonest target day shown. The peak is the best time to set off; the height is how likely it stays dry. Each hour is its own probability, so the lines need not sum to the daily "any dry window" figure.

5-hour dry window

Phase 3q — Gaussian copula MC over Phase 3c hourly P(wet)

Gaussian copula MC over Phase 3c's hourly P(wet) marginals — the 3p engine bound to the rich (non-orographic) 3c instead of 3o, for coastal locations without a 3o model. Single empirical Σ per station from train-split observed daytime wet/dry sequences. Captures within-day wet/dry autocorrelation an iid sampler misses; scored against the 3b champion at Sennen.

Target date (UTC) +24h +48h +72h Confidence Best start (UTC, P dry block)
82%83%88% confident dry (P=82% · run 4–9h) 12:00Z (80%)
97%92%91% confident dry (P=97% · run 7–9h) 09:00Z (92%)
—36%30% ambiguous (P=36% · run 1–9h) 12:00Z (32%)
——85% confident dry (P=85% · run 4–9h) 08:00Z (76%)
Monte-Carlo P(an 5-hour dry block runs from each start hour) — one line per forecast horizon, soonest target day shown. The peak is the best time to set off; the height is how likely it stays dry. Each hour is its own probability, so the lines need not sum to the daily "any dry window" figure.

6-hour dry window

Phase 3b — lean (53 features)

53-feature LightGBM. Champion.

Target date (UTC) +24h +48h +72h Model agreement Confidence
91%91%86% 83% confident dry (P=91%)
91%79%94% 100% confident dry (P=91%)
—35%25% 0% ambiguous (P=35%)
——96% 67% confident dry (P=96%)
Phase 3q — Gaussian copula MC over Phase 3c hourly P(wet)

Gaussian copula MC over Phase 3c's hourly P(wet) marginals — the 3p engine bound to the rich (non-orographic) 3c instead of 3o, for coastal locations without a 3o model. Single empirical Σ per station from train-split observed daytime wet/dry sequences. Captures within-day wet/dry autocorrelation an iid sampler misses; scored against the 3b champion at Sennen.

Target date (UTC) +24h +48h +72h Confidence Best start (UTC, P dry block)
73%74%82% ambiguous (P=73% · run 4–9h) 11:00Z (72%)
95%84%86% confident dry (P=95% · run 7–9h) 08:00Z (91%)
—27%20% ambiguous (P=27% · run 1–9h) 11:00Z (25%)
——75% ambiguous (P=75% · run 4–9h) 08:00Z (69%)
Monte-Carlo P(an 6-hour dry block runs from each start hour) — one line per forecast horizon, soonest target day shown. The peak is the best time to set off; the height is how likely it stays dry. Each hour is its own probability, so the lines need not sum to the daily "any dry window" figure.

Dry hour = all four 15-min readings ≤ 0.1 mm. Search bounded to 09:00–18:00 local.