What the terms mean — UKMO, persistence, climatology, MAE, ETS, Brier skill, ±q90
UKMO
The Met Office deterministic model (ukmo_seamless via
Open-Meteo) — the forecast being scored.
Persistence
The laziest possible forecast: "the weather will be what it was
24 h ago" (for lead d days, the observed value 24×max(d,1) h before the
target hour, from the same truth source). Beating it is the minimum bar for any model.
Climatology
"Typical weather for the date": the mean over 2024–26 for that
station, day-of-year and hour of day. No weather information at all — only the calendar.
A model earns its keep only while it beats this.
MAE
Mean absolute error — average size of the miss, in the variable's own
units. Identical to CRPS for a single-valued forecast.
ETS
Equitable Threat Score for the selected yes/no rain event (the "Rain
event" buckets above; "Any" = ≥0.1 mm in the hour): the fraction of event hours
correctly forecast, after subtracting the hits random chance would get. 1 = perfect,
0 = no better than chance. The bucket names follow the Met Office intensity bands —
slight <0.5 mm/h, moderate 0.5–4, heavy >4 (torrential would be ~8+, too rare
here to score). Heavier buckets are much rarer (any rain: ~23% of hours, heavy:
~0.2%), and skill collapses with severity. Climatology has no ETS line:
its rain "forecast" is the historical frequency, which never crosses the 50%
yes/no threshold — it would never forecast rain, a degenerate ETS of 0 by construction.
The honest probabilistic comparison is the Brier-skill panel, where climatology is the
zero line the model must beat.
Brier skill
How much better the forecast's probabilities are than always
quoting the base rate (~23% chance of rain). Positive = adds information; negative =
a hard yes/no call is worse than the base rate, because it is overconfident.
±q90
Conformal interval half-width: 90% of outcomes fall
within forecast ± q90. Calibrated on 2024 only, then verified on held-out 2025–26
(the coverage numbers in the tooltip).
95% CI
Bootstrap confidence interval on the curve (resampling whole
station-days, so correlated hours don't fake precision). Computed at UK scope; with
~700k forecast/obs pairs per lead they come out at ±0.005–0.015 — too narrow to draw
as a visible band, so they live in the hover tooltip and each chart's data table.
A band will appear automatically wherever a CI is ever wide enough to see.