Thunder: one summer or three?
Finished · 15 November 2026
The question. Does training the thunder calibration on three summers (2023–2025) instead of one make it better, or just different?
Data
Summers 2023, 2024 and 2025 (1 June – 31 August), 276 00z runs, all 31 members, against airport thunder within 50 km: 26,583 station-days at day 0 (1,033 with thunder) and 25,721 at day 3 (1,033). The 2023 and 2024 runs were downloaded for this; the 2025 ones were already in use. Same model as before (see thunder nearby), fitted by the production calib_fit.
Results
| Summer 2025 only | Summers 2023–2025 | |
|---|---|---|
| Runs | 92 | 276 |
| Day 0 BSS | +0.082 | +0.149 |
| Day 0 AUC | 0.866 | 0.892 |
| Day 3 BSS | +0.053 | +0.104 |
| Day 3 AUC | 0.856 | 0.852 |
| Day 0 coefficients (intercept, best LI, log rain) | −4.870, −0.508, 1.179 | −4.849, −0.692, 1.054 |
| Day 3 coefficients (intercept, logit share) | −2.539, 0.716 | −2.609, 0.719 |
What worked
More data. The coefficients barely moved, yet skill on held-out weeks nearly doubled: with one summer the weekly cross-validation was fitting on few events, and each held-out week was judged by a noisier fit.
What did not
Nothing failed outright, but the day-3 ROC area did not improve (0.856 → 0.852): more data made day-3 chances better calibrated, not better at ranking days. Ranking at day 3 is limited by the forecast itself.
What we changed
The live thunder calibration is the three-summer fit (from 15 November 2026), and the weekly refit keeps adding summers as the archive grows.
Caveat
All three summers are scored by cross-validation within the same period; how it does on summer 2027 is the real test, and the Verification page will show it.