Multi-model forecast blending for Sennen, Cornwall
Sea state models
Significant wave height — quantile-LGB over the five wave models with a
conformal-calibrated (CQR) 90% band. Hs MAE in metres, lower better;
band coverage nominal 0.90.
Quantile-LGB (q05/q50/q95) over the five wave models, split-CQR
calibrated to a 90% band. Trained 2026-08-02 on
era5_ocean reanalysis (training truth; the Sevenstones
buoy is the verification truth below). Trained on lead-unlabelled hindcast; per-lead calibration planned once the offset_day archive matures (~2026-09).
Hs MAE (cal slice, vs era5_ocean)
0.108 m
90% band coverage (cal slice)
0.901 (nominal 0.90)
Training / val / cal rows
31,142 / 3,893 / 3,893
Training window
2022-01-01 → 2026-06-10
Features
51
Buoy verify history (15 runs)
Mon + Thu rolling Hs MAE (m) + empirical 90%-band coverage vs the
Sevenstones Light Vessel buoy, per 24h actual-lead bucket (≤24h = leads 0-23h, …).
The buoy sits ~28 km W of the forecast cell, so absolute MAE carries a location
mismatch — read the trend, not the level. Red cells flag drift: band coverage
well below the nominal 0.90 (MAE is never a drift trigger here, for the same
location-mismatch reason).