When do data-driven models improve runway-configuration forecasts? Multi-horizon operational evidence from São Paulo/Guarulhos

Runway-configuration forecasts can support coordination, but state persistence may inflate accuracy while changes remain undetected. We establish when data-driven models add value at São Paulo/Guarulhos International Airport, where two reciprocal configurations make the current state a strong reference. We integrate 87,540 15-minute movement, schedule, and METAR intervals from 1 January 2022 to 30 June 2024 into 33 predictors under stated availability rules. Forecasts at 15, 30, and 60 min are selected by expanding-window validation in 2023 and evaluated on a fixed January–June 2024 future test period, comparing majority, persistence, climatology, and a wind rule with logistic regression, random forest, and gradient boosting, plus three adapted comparators. Locked-test balanced accuracy for the selected models is 0.971, 0.959, and 0.940, against 0.961, 0.942, and 0.904 for persistence; 95% intervals for the paired gains exclude zero. Value concentrates at 60 min, where transition recall reaches 0.373 at precision 0.507 and 2.07 false alerts per operational day, against 0.220 and 1.70 for the wind rule; at 15 min it improves transition F1 over the wind rule only by accepting lower precision. At 60 min, resampling variants that raise recall do so only by increasing alert burden; a transition-F1 threshold improves precision and recall while reducing alert burden, but reduces eligible-event coverage. Reliance shifts from the current state mainly toward time of day, and less toward wind, as lead time grows, consistent with complexity becoming defensible only at the longest horizon. The evidence concerns forecast quality; capacity, delay, workload, and safety outcomes are not measured.