How accurate is Aurorama?
Aurorama's chance for a dark, clear hour comes from a model fitted to 18,633 reports from Aurorasaurus, a citizen-science project: 15,546 sightings and 3,087 reports of a clear sky with no aurora, 2014–2025. To test it fairly, we fit it on half the days and scored it on the other half, which it never saw (8,686 reports).
The short answer
- It tells good nights from bad ones. Take a real sighting and a clear-sky miss at random: the model gave the sighting the higher chance 84% of the time (AUC 0.84; 50% would be guessing).
- 37% better than our first, textbook model (log-loss 0.555 → 0.350; Brier 0.148 → 0.108).
- Adding light pollution clearly improved the test scores; the live solar wind improved them slightly. Both are in the model.
Are the percentages right?
Each dot groups held-out reports by the chance the model gave them; its height is how many of those reports were real sightings. Dots on the diagonal mean the percentages can be taken at face value; bigger dots hold more reports.
From about 56% up, the chances match what happened (8,009 reports). Below 20%, the model is cautious: when it said 6–15%, about 32% of reports were sightings (only 186 reports, so a low number is worth a look outside if the sky is clear).
People report a sighting far more readily than a miss (about 15× here), so the chart compares like with like: the model's chance adjusted for that, against the reports.
What this can’t tell you
- Clouds. The test uses reports made under a clear sky. The cloud forecast (Open-Meteo) adds its own uncertainty, especially more than a day ahead.
- Space weather forecasts. The test uses the Kp that actually happened. NOAA's Kp forecast for the coming days is less certain, and sudden storms can arrive with little warning.
- Who reports. Citizen reports come mostly from North America and Europe, and from people already looking up.
Aurora is never guaranteed: treat the chance as odds, not a promise. Data: Aurorasaurus web observations 2014-08 to 2025-08 (CC BY-NC 4.0), GFZ Kp, NASA OMNI.