Phase 3 · nocturnal migration forecast per city

Forecast of the density of birds in nocturnal flight over each city for the coming nights, made from the weather forecast alone: there is no radar in any of these cities. Both models are the phase 2 ones retrained without the local climatology, which is the configuration validated by leaving whole radars out of the training: in that scenario they captured 34 % of the heavy-passage nights against the 10 % that chance would give, with a median area under the curve of 0.77.

What the alert level means. The model gives a continuous number, and what counts as a lot in Sevilla is not the same as in Bilbao, nor the same in Sevilla in March as in Sevilla in October. For each city the model has been run over the weather archive from 2021 onwards at its own coordinates, and the levels are cut at the percentiles of its distribution around the same date: moderate above the median, high above the 75th percentile and very high above the 90th percentile, that is the busiest night in every ten. The window is the 21 days either side of the night, so the threshold climbs into the peak of the passage and falls away from it. That way the alert means the same thing everywhere and at any date: «tonight far more is passing than is normal here at this time of year».

The first verification of the published forecasts, in September 2026, is what forced this: with a single threshold for the whole season, 65 % of the nights at the peak came out as an alert and the level no longer told the nights apart. The moving window is the fix.

Nights with a high or very high alert

Full forecast

The thresholds in force for these nights

Computed over the predictions of the 2021-today weather archive at each city's point, on the 21 days either side of each date. The birds/km² are the cuts of the intensity model; the probability is the cut of the heavy-passage classifier, which is what decides the alert. Other dates of the year have their own cuts, higher in the middle of the passage and lower at its edges.

Limitations

• The forecast degrades with the days: the first night runs on an almost closed analysis and the seventh on a six-day forecast. The model was trained on analyses and short-range forecasts, so the distant nights are worse than the validation figures suggest.
• None of these cities has a radar within useful range, so there is nothing to compare the forecast against the next day. The phase 2 validation is the only guarantee, and it was done precisely by leaving whole radars out to imitate this situation.
• The forecast density is that of the air volume over the city, not the number of birds that approach the lights. The exposure piece is missing, and that is the artificial-light ranking.
• Without the flight speed from the radar processor, part of the autumn density in the south may be insect and not bird. It is the same limitation as in phase 2.