Calibration · updated weekly
Do our scores hold up?
Every night we publish a 0–100 sky score. A score that never faces reality is just marketing — so each week we replay our own forecasts against what the atmosphere actually did, using ERA5 reanalysis (a measurement-grade record assimilating satellite and ground observations). Same cities, same scoring rule, no cherry-picking.
93% of nights we scored 80+ were actually clear
Cloud forecast reliability
The falsifiable part of a sky score is its cloud forecast. Below: when we predicted a given cloud level for the dark hours, how often the night turned out genuinely clear (observed weighted cloud ≤ 20%, same low/mid/high weighting the live score uses).
| Predicted cloud | Nights | Actually clear | Observed cloud (mean) |
|---|---|---|---|
| 0–10% | 1295 | 5.1% | |
| 10–30% | 835 | 19.2% | |
| 30–60% | 695 | 38% | |
| 60–100% | 295 | 64% |
Across all 3,120 verified nights, the predicted weighted cloud missed the observed value by 6.6 percentage points on average.
Method — and the honest limits
- What is compared. For each of 52 cities (fixed public coordinates), each night's astronomical-darkness window is computed locally (Skyfield + JPL DE440s). The published cloud forecast for those hours is scored with the exact live rule (sky-score-v2-20260715), then re-scored with the observed cloud from ERA5. Moon inputs are identical on both sides — every difference you see comes from cloud.
- Lead time. The headline table verifies the shortest-lead published forecast (≤ 1 day). Since 2026-10-12 we also archive every day's 0–7-day forecast for all sampled locations so longer leads can be verified against the same nights; that table appears once every required lead (1, 3, 7 days) has at least 200 verified nights — until then we won't imply scores 3–16 nights out are as reliable as tomorrow's.
- "Observed" means reanalysis. ERA5/ERA5T assimilates satellite and ground observations into a physically consistent record. It is measurement-grade — not another forecast — but it is not a person standing in a field, and it lags a few days, which is why the period ends 2026-11-06.
- Why we verify the cloud claim, not just the score. A cloudless full-moon night scores low for deep-sky targets yet is genuinely clear — so a low score is often the Moon doing its job, not a missed forecast. The table above isolates the part of the score that can actually be wrong.
- Skipped nights are counted, not hidden. 0 nights lacked forecast, observation or an astronomical dark window and were excluded as such.
Forecast archive: Open-Meteo Historical Forecast API (CC BY 4.0) · Observations: ERA5/ERA5T reanalysis via Open-Meteo Archive API (CC BY 4.0). Data: calibration.json (CC BY 4.0) — the exact file this page is built from, method notes included.