Thunder: does place or time of year matter?
Planned · 15 November 2026
The question. Does the thunder calibration improve if it knows the region (e.g. south-east England against north-west Scotland) or the day of the year?
Why
Thunder is far commoner in south-east England than in north-west Scotland, and commoner in July than in June or late August; the live model knows neither, so it may be over-confident in the north-west and early in the season.
Plan
- Regions from the site's own areas (the regions file the synopsis uses), as offsets on the live model's intercept, penalised (ridge) so that a region with few thunder days stays close to the shared model.
- Day of year as one smooth seasonal term (a sine and cosine of the date).
- Data: summers 2023–2025 (the live model's data).
How it will be judged (fixed before running)
Leave-one-week-out cross-validated BSS and AUC, by summer. Adopted only if BSS rises in all three summers and no region's reliability gets worse; otherwise reported here as not helping.