How this forecast works

Start with the sky, not the map

The aurora follows Earth's magnetic field, and the magnetic pole sits a long way from the geographic one. So a place's latitude on a map says little about how much activity it needs. Whitehorse is 60.7°N on paper and about 63° magnetically, which is why it sees the aurora at levels that a map-based table calls a non-event.

Every location here has its own threshold, the level of activity at which the aurora becomes reachable from that spot. For the twenty tracked destinations that threshold comes from magnetic latitude and from seasons of watching that particular sky. For any other point you search, it is estimated from the magnetic field model and from the nearest calibrated locations, and the page says so.

Activity: what the sun is doing right now

Geomagnetic activity is read two ways. For this moment, a live estimate updated every minute, smoothed over the last quarter hour so a single noisy reading cannot swing the page. For the hours ahead, NOAA's three-day forecast, and the page says which one it is using.

Alongside that sits the solar wind itself, measured about 1.5 million kilometres sunward of Earth. The direction of the magnetic field it carries decides whether the door into Earth's field is open. When it turns south, the forecast rises before the activity index has caught up; when it turns north, it falls. That reading is averaged over ten minutes, because a single minute of solar wind is noise.

Darkness: how far below the horizon the sun sits

The aurora needs a dark sky, and dark is not a switch. The forecast uses the sun's actual depth below the horizon, minute by minute, and lets the number through gradually as the sky deepens from civil twilight into true night. Civil twilight is an honest zero, whatever the activity. Deep twilight lets a strong display through at a reduced value. Past full night the sky costs nothing.

In midsummer at high latitude the sky never gets dark enough, and the page says that rather than printing a small number that could not be true.

Cloud: four models, and a warning when they disagree

Cloud is the input most often wrong, and the one you can check yourself by looking up. Three things are done about that. Cloud is counted by what it actually blocks: thin high cloud dims the aurora a little, a low overcast hides it, and the number you see is that effective figure, which can be lower than a weather app's total.

The forecast uses four independent weather models and averages them, which is more reliable than any one of them. When the four disagree badly about the hour that matters, one calling it clear and another overcast, the page says so under the headline instead of printing a single confident number. And a forecast overcast is never allowed to zero a night on its own, because up here the cloud forecast is wrong often enough that a small signal is more honest than none. If the sky does clear, the sky-clearing alert reacts to the real cloud, not the forecast.

The moon: how full, and how high

A bright moon washes out a faint aurora. How much depends on more than the phase: a full moon sitting just above the ridge has its light crossing a great deal of air on the way in and brightens the sky far less than the same moon high overhead. The forecast weights moonlight by the moon's altitude using a published sky-brightness model, and a moon below the horizon costs nothing at all.

One formula, everywhere

The headline, the hour-by-hour bars, the five-night outlook, the evening heads-up and every push alert are the same computation. They cannot disagree with each other, because there is nothing for them to disagree about. If a bar says 26% for an hour, the alert for that hour fires on 26%.

Tonight means the best hour still to come

The big number is not this instant. It is the best the night still has left, from now until dawn. This second's value is shown underneath, with the time the peak is expected. That is the number a person deciding whether to put a coat on actually needs, and it cannot collapse just because the sky is still deepening when they look.

Checked against what people actually saw

Every forecast can be answered. Under the number there is a question, can you see it right now, with three honest answers: yes, nothing yet, and camera only. Each answer is stored with everything the forecast said at that moment, and a good share of them come from guides who are out every night the tours run, in the sky this forecast was built around.

Those reports are compared with the forecast, level by level. Where a given percentage came true more often than it said, the number is raised; where it came true less often, it is lowered. The adjustment is applied to every number on the site, and it is rebuilt as the reports accumulate. When it is in effect the page says so, under the headline. This is the part no amount of clever modelling replaces, and it is the reason the reporting buttons are free and unobtrusive.

What this forecast will not tell you

Nights four and five of the outlook carry no percentage. Darkness and the moon are exact for any night, so those are shown; activity and cloud that far out are not forecastable to a useful accuracy, and putting a number on them would be theatre.

When a live feed is down, no reading is invented. The page says the data is estimated, and alerts pause rather than fire on a guess. A night that the forecast calls poor is reported as poor. Honesty about a quiet night is what makes a good night believable.