How our aurora score works

How we built it

All-sky cameras at Kilpisjärvi and Sodankylä in Finland and Abisko in Sweden photographed the sky through more than ten winters. The Finnish Meteorological Institute runs the Kilpisjärvi camera, and the University of Oulu's Sodankylä Geophysical Observatory the other two. The Finnish Meteorological Institute published how bright the aurora was in these records, averaged over the sky above each camera. From these records we worked out how much aurora you can see with each mix of aurora activity, darkness and moonlight.

For cloud, we compared years of archived MET Norway forecasts with what the all-sky cameras at Tromsø and Skibotn saw on 818 nights. This shows how often aurora can still be seen when the forecast says, for example, that low cloud covers half the sky.

Inside the score, the model first works out your chance of seeing aurora with your own eyes in that hour. The score then adds how bright the aurora should be. To check the method, we tested that chance. We rebuilt the model eleven times, each time without one winter, and had it predict that winter's dark, moonless hours. We compared it with the simplest forecast we could think of, one fixed number. In the other winters, the cameras saw aurora in about half of the dark, moonless hours, so that forecast says “about 50%” for every hour. The model's error, measured as log loss (the usual measure for forecasts of chances), was 36% lower than that fixed number's. Over all those hours, the cameras recorded 93 aurora hours for every 100 the model expected.

Svalbard lies under the polar cap, where the aurora behaves differently and can shine at midday during the polar night. The score there has its own curve, built on camera records from Svalbard. Every source we used is on our data sources page.

Why Kp alone is not enough

Northern Norway lies under the auroral oval, the aurora's usual path, so aurora can be overhead at Kp 1 or 2. Whether you see it depends more often on cloud, darkness, moonlight and light pollution than on Kp, and the same Kp gives very different scores at different spots. Here are our scores at three spots, with Kp as the only measure of aurora activity. The night is clear and dark, without moonlight, at 22:30 Norway time in mid-December:

• Finnvikdalen near Tromsø scores about 24 at Kp 1, 82 at Kp 3 and 95 at Kp 5.
• Tromsø centre, with the town's light pollution, scores about 24 at Kp 3 and 49 at Kp 5.
• Løpsmarka near Bodø, further south, scores about 4 at Kp 1 and 70 at Kp 3.

In our camera replays, cloud alone picked the hour with aurora about as often as aurora activity alone, 78 and 77 times in 100. Kp alone did so 71 times in 100. The model, which weighs both, did so 89 times in 100.

How it compares

We replayed past winters at all-sky cameras in Finland, Sweden and Norway, with inputs rebuilt as the score would have had them at the time. Then we compared the model with simpler ways to forecast aurora. For each winter, we rebuilt the model and each simpler forecast without that winter and had them predict it. Most comparisons below use the model's chance of seeing aurora, as in the check above. The alerts use the score itself.

• Which nights to go out. We took 787 nights from seven winters at three of the cameras. Each method rated every night with what it knew at 18:00, and we counted how often the cameras saw aurora on the nights it rated highest. Across all the nights, the cameras saw aurora on about one in four. Of the nights in the model's top fifth, they saw aurora on 54%, and of those in its top tenth on 64%. Of the nights in the top fifth of NOAA's Kp forecast, they saw aurora on 25%, about as often as on any night. A rule from Kp and MET Norway's cloud forecast came closer to the model, at 50%. For the nights before spring 2022, we rebuilt NOAA's forecast from its known errors, because we found no archive of the forecasts it issued.
• Against Kp alone. We compared the model with a forecast from the Kp index and the time of night. Over 6,716 clear hours on 1,072 nights, the model's error (log loss) was about 30% lower (0.33 against 0.48), and lower in each of the 12 winters. About two thirds of that gain came from the live reading of the magnetometers, and the rest from how the model handles moonlight and twilight. In dark, moonless hours the model's error was about 24% lower, and about 95% of that gain came from the magnetometers. They stand at or near the cameras, so at a spot farther away the gain may be smaller.
• Where Kp goes wrong. When Kp was 1 or less, the cameras saw aurora in 11% of the hours. The Kp-only forecast gave 28% on average, and our model 21%. With a bright moon high in the sky, the cameras saw aurora in 32% of the hours. The Kp-only forecast gave 50%, and our model 35%. When activity was high, at Kp 4 or more, the cameras saw aurora in 98% of the hours. The Kp-only forecast gave 86%, and our model 95%.
• Against Kp and the cloud forecast. When we took one hour with visible aurora and one without, the model ranked the right hour higher 89 times in 100. A forecast from Kp and the same cloud forecast did so 83 times in 100 (AUC 0.890 against 0.825, over 13,693 hours). The model did better in each of the 7 winters with archived cloud forecasts. Without the magnetometers, the model was still better at telling which hours were most likely to bring aurora. But the chances it gave, as numbers, were less accurate than the simpler forecast's.
• Alerts. This comparison reuses winters the score was built on, so it flatters the score. In the replay at the Tromsø all-sky camera, “Good chance tonight” (a score of 60 or more) went out about 27 times a season. Of these alerts, 9 in 10 were followed by aurora on the camera in the same hour or the next two. Simpler rules were right less often. “Kp 3 or more, with low and medium cloud under 50%” was right 7 in 10 times and “Kp 4 or more” 8 in 10. An alert that looks only at the month and the time of night was right 4 in 10 times. On two in three nights when the camera saw aurora in a dark, clear hour, no alert went out. The camera picks up fainter aurora than your eyes can see.
• Cloud. The way the score reads MET Norway's cloud forecast picks out clear hours better than the forecast's total cloud cover (AUC 0.803 against 0.738). We checked this over about 150,000 dark hours at seven airports and three camera sites. The gain comes mostly from how it treats thin high cloud. Night-time reports rarely record thin high cloud, so part of that gain may come from how the reports are made. In hours with little high cloud the two were close (0.806 against 0.797).
• Against NOAA's OVATION model. Many aurora apps show OVATION's chance of aurora overhead. We could not replay this comparison, because NOAA does not keep OVATION's past output and only a few files from 2020 onward survive in public archives. This winter's test makes it hour by hour, on clear, dark, moonless hours, against the OVATION value the score read at the time.

Two parts were not rebuilt for each winter. The moon and twilight terms were fitted on the same cameras, and the cloud reading on two other cameras in the same winters. Rebuilt without each winter, the cloud reading gave the same result. This winter's test remains the fairest measure.

How we test it

We recorded the test plan before the 2026/27 season, with timestamps anyone can check. The test runs from the night of 10 October 2026 to the night of 30 April 2027. The model's parameters stay fixed for that time, apart from fixes the plan allows in advance. The forecasts keep updating as new data arrives.

The test uses the all-sky cameras at Kilpisjärvi and Abisko, two of the three the model was built on, so it checks the model on new nights at the same places. The calibrated camera at Sodankylä, the third, was taken out of service in 2021. It compares the model's chance of seeing aurora with a forecast from Kp alone, for the current hour and the next 12 hours, and with a forecast from Kp and the cloud forecast. It also checks what the model adds to NOAA's own OVATION value. Finally, it compares our forecasts for the next three nights with a simpler forecast, and our alerts with a rule built on Kp and cloud.

We will publish the results, good or bad. Reports follow in April 2027 for the nights to 31 March, and by 31 May 2027 for the rest of the season. We will publish the plan itself with the April report. The camera comparisons follow once the Finnish Meteorological Institute releases this winter's camera data.