UTC --:--

📈 Forecast Timeline

🔬 Cascade Decomposition

Stage 1: Aurora Occurrence

Space Weather Driven

0%aurora probability100%
×

Stage 2: Clean View Given Aurora

Cloud Cover & Moonlight

clearovercast
=

Product: Clean Auroral Hour

Combined Prediction

0%visibility100%

🧪 Methodology

Aurora Hunter decomposes visibility forecasting into two physically distinct stages, multiplied into a single calibrated probability.

1

Aurora Occurrence

A calibrated XGBoost model trained on multi-station all-sky camera archives, driven by live NOAA solar wind and geomagnetic data (Kp, Bz, Newell coupling) in AACGM magnetic coordinates.

2

Observation Conditions

A calibrated logistic regression estimating whether an occurring aurora is actually observable, from cloud cover (Open-Meteo) and moonlight conditions.

×

Visibility Probability

The product P(visible) = P(occurring) × P(clear | occurring) gives a well-calibrated probability you can act on.

0.941
ROC-AUC · Tromsø test (2019–2020)
0.915
ROC-AUC · Kiruna 2024
0.892
ROC-AUC · Skibotn (withheld station)
+0.11–0.24
vs. Stage 1 only
Scope of this prototype

Retrospective validation used archived OMNI solar-wind data and ERA5 cloud reanalysis at three Nordic stations (magnetic latitude 65–67°). This site runs the same models on real-time NOAA SWPC feeds and cloud forecasts, so its skill is expected to be lower than the published scores. Probabilities refer to a clean auroral camera view, used as a proxy for good naked-eye viewing; outside the validated latitude band and above Kp 5 they are extrapolations.