Frost and fog: can the raw chances be trusted?
Finished · 13 November 2026
The question. How good are GEFS's raw member shares for frost and fog at airports, and how much does calibrating them against airport reports help?
Data
00z GEFS runs from January–March and October–November 2026 against METARs from airports in the UK and Ireland: 9,559 station-days at day 0 and 9,295 at day 3. Events: frost0 = minimum temperature 0 °C or below, frost5 = −5 °C or below, fog1 = visibility under 1 km, fog2 = under 200 m. Raw chance: the share of members reaching the step. Calibrated: frost by a Gaussian for the airport's lowest temperature of the day (EMOS: the mean from the ensemble-mean minimum, the share of members with clear, calm conditions, the land share and the height above the model's ground; the variance from the ensemble variance and the clear-and-calm share), fog by logistic regression on the share of members with fog, the members' highest humidity, lowest wind, the share that are clear and calm, and the land share of the cell. Skill is cross-validated by 7-day blocks.
Results (Brier skill score)
| Event | Raw, day 0 | Calibrated, day 0 | Raw, day 3 | Calibrated, day 3 | Events (day 0) |
|---|---|---|---|---|---|
| Air frost (≤ 0 °C) | +0.42 | +0.52 | +0.33 | +0.41 | 1,128 |
| Hard frost (≤ −5 °C) | −0.01 | +0.20 | +0.09 | +0.17 | 148 |
| Fog, under 1 km | −1.34 | +0.15 | −0.64 | +0.10 | 446 |
| Fog, under 200 m | −0.49 | +0.11 | −0.20 | +0.06 | 277 |
What worked
Calibration, for every event and both leads. GEFS's 2 m temperature is several degrees too warm on clear, calm nights, so raw frost shares were too low; the calibration's shift and spread fix most of that.
What did not
- Raw fog shares. At −1.34, the raw "visibility under 1 km" chance was more than twice as bad as always forecasting the usual frequency. A 0.5° model cannot resolve fog, and its visibility field is closer to a diagnostic guess.
- Hard frost, raw: no skill at all at day 0 — the warm bias pushes the members above −5 °C on the very nights it matters.
What we changed
Frost and fog chances on the site are always calibrated. Frost is trained on October–April only (summer days add nothing but non-events), and the weekly refit keeps re-fitting both against new airport reports (see Verification).
Caveat
Airports are flat, open and often a little warm; frost hollows and fog-prone valleys are under-represented, so the calibration is fairest where the land looks like an airfield.