WeatherBlend

Multi-model forecast blending for Bonehill Rocks, Dartmoor

Dry-window forecast

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

Bellever Dartmoor

2-hour dry window

Phase 3p — Gaussian copula MC over Phase 3o hourly P(wet)

Gaussian copula MC over the rich+orographic Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences (daytime-shape autocorrelation is lead-independent for truth, so per-lead Σs pool into one). Captures within-day wet/dry autocorrelation that an iid sampler misses. Bake-off Brier 0.1064 aggregate on the 9 (station, lead) Bonehill cells.

Target date (UTC) +24h +48h +72h Confidence Best start (UTC, P dry block)
99%99%99% confident dry (P=99% · run 4–9h) 15:00Z (98%)
100%100%100% confident dry (P=100% · run 9–9h) 14:00Z (98%)
—63%70% ambiguous (P=63% · run 1–7h) 15:00Z (44%)
——100% confident dry (P=100% · run 7–9h) 13:00Z (97%)
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
100%73%96% 100% confident dry (P=100% · τ=36%)
100%100%100% 100% confident dry (P=100% · τ=36%)
—41%55% 50% confident wet (P=41% · τ=40%)
——99% 100% confident dry (P=99% · τ=44%)
Phase 3p — Gaussian copula MC over Phase 3o hourly P(wet)

Gaussian copula MC over the rich+orographic Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences (daytime-shape autocorrelation is lead-independent for truth, so per-lead Σs pool into one). Captures within-day wet/dry autocorrelation that an iid sampler misses. Bake-off Brier 0.1064 aggregate on the 9 (station, lead) Bonehill cells.

Target date (UTC) +24h +48h +72h Confidence Best start (UTC, P dry block)
98%97%98% confident dry (P=98% · run 4–9h) 14:00Z (97%)
100%99%100% confident dry (P=100% · run 9–9h) 10:00Z (97%)
—48%47% ambiguous (P=48% · run 1–7h) 14:00Z (36%)
——99% confident dry (P=99% · run 7–9h) 12:00Z (96%)
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
94%42%67% 100% confident dry (P=94% · τ=32%)
100%100%99% 100% confident dry (P=100% · τ=32%)
—41%39% 33% confident wet (P=41% · τ=45%)
——91% 100% confident dry (P=91% · τ=40%)
Phase 3p — Gaussian copula MC over Phase 3o hourly P(wet)

Gaussian copula MC over the rich+orographic Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences (daytime-shape autocorrelation is lead-independent for truth, so per-lead Σs pool into one). Captures within-day wet/dry autocorrelation that an iid sampler misses. Bake-off Brier 0.1064 aggregate on the 9 (station, lead) Bonehill cells.

Target date (UTC) +24h +48h +72h Confidence Best start (UTC, P dry block)
96%93%96% confident dry (P=96% · run 4–9h) 13:00Z (95%)
99%98%99% confident dry (P=99% · run 9–9h) 12:00Z (96%)
—33%35% ambiguous (P=33% · run 1–7h) 13:00Z (24%)
——99% confident dry (P=99% · run 7–9h) 11:00Z (94%)
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 3p — Gaussian copula MC over Phase 3o hourly P(wet)

Gaussian copula MC over the rich+orographic Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences (daytime-shape autocorrelation is lead-independent for truth, so per-lead Σs pool into one). Captures within-day wet/dry autocorrelation that an iid sampler misses. Bake-off Brier 0.1064 aggregate on the 9 (station, lead) Bonehill cells.

Target date (UTC) +24h +48h +72h Confidence Best start (UTC, P dry block)
77%84%82% ambiguous (P=77% · run 4–9h) 12:00Z (81%)
98%97%99% confident dry (P=98% · run 9–9h) 11:00Z (95%)
—23%27% ambiguous (P=23% · run 1–7h) 12:00Z (19%)
——97% confident dry (P=97% · run 7–9h) 10:00Z (93%)
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
67%13%81% 50% confident dry (P=67% · τ=46%)
93%100%79% 83% confident dry (P=93% · τ=46%)
—7%5% 17% confident wet (P=7% · τ=38%)
——55% 83% confident dry (P=55% · τ=48%)
Phase 3p — Gaussian copula MC over Phase 3o hourly P(wet)

Gaussian copula MC over the rich+orographic Phase 3o's hourly P(wet) marginals. Single empirical Σ per station, fit on train-split observed daytime wet/dry binary sequences (daytime-shape autocorrelation is lead-independent for truth, so per-lead Σs pool into one). Captures within-day wet/dry autocorrelation that an iid sampler misses. Bake-off Brier 0.1064 aggregate on the 9 (station, lead) Bonehill cells.

Target date (UTC) +24h +48h +72h Confidence Best start (UTC, P dry block)
63%76%74% ambiguous (P=63% · run 4–9h) 11:00Z (69%)
96%92%98% confident dry (P=96% · run 9–9h) 10:00Z (94%)
—17%21% confident wet (P=17% · run 1–7h) 11:00Z (15%)
——94% confident dry (P=94% · run 7–9h) 09:00Z (92%)
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.