Thunder at the airport, or nearby?
Finished · 14 November 2026
The question. Should the thunder chance be for the spot, or for "somewhere within R km", and which R and which predictors give the most skill?
Why
A storm a few kilometres away is a storm for anyone planning a day, but a single airport only hears thunder within about 15 km; and a model at 0.5° cannot place showers to the town. So we tested defining the event as "thunder reported by any airport within R km" and forecasting it from the 3×3 neighbourhood of grid cells around the airport.
Data
Summer 2025, 101 airports, as in storm ingredients. Logistic regression, leave-one-week-out cross-validation, scored by Brier skill (BSS) against the event's base rate.
| Radius | Neighbouring airports (median) | Airports with none | Base rate | Thunder days |
|---|---|---|---|---|
| airport only | – | – | 1.6 % | 142 |
| 30 km | 0 | 53 of 101 | 2.4 % | 212 |
| 50 km | 2 | 21 of 101 | 3.6 % | 317 |
Results (BSS, cross-validated)
| Event, predictors | Day 0 | Day 3 |
|---|---|---|
| Airport, member share of the rule | +0.017 | +0.022 |
| Airport, best LI + log rain | +0.053 | −0.008 |
| 30 km, best LI + log rain | +0.066 | −0.006 |
| 30 km, member share | +0.031 | +0.035 |
| 50 km, best LI + log rain | +0.079 | +0.007 |
| 50 km, member share | +0.046 | +0.055 |
| 50 km, share + best LI + rain | +0.077 | +0.030 |
What worked
- 50 km. More events (317) and almost every airport with a neighbour, and the best skill at both leads.
- Different predictors by lead. At day 0 the ensemble-mean lowest best lifted index and the log of the largest 6-hour rain carry the information; by day 3 those continuous fields are too noisy and the simple share of members meeting the rule is better.
What did not
- 30 km: half the airports have no other airport that close, so the event is barely different from "at the airport".
- Continuous predictors at day 3: best LI + rain had no skill (−0.008 to +0.007); adding them to the member share made day 3 worse.
- Most-unstable CAPE as a third predictor added nothing to best LI + rain (+0.051 against +0.053 at the airport).
What we changed
The live thunder chance became "thunder within about 50 km", with day 0 calibrated on best LI + rain and day 3 (and later days) on the member share; the production fit (3×3 neighbourhood predictors) scored +0.082 and +0.053 on summer 2025. See three summers for what more data did.
Scripts: scripts/thunder/multi.py, scripts/thunder/area.py.