When does network information pay off? Delay propagation for arrival delay prediction and decisions before and after departure

Airline and air cargo operators base rebooking and dispatch decisions on arrival time estimates. This study asks when network delay features, evaluated mainly for departure delays, pay off for arrival decisions. We derive personalized PageRank features from airport, airspace, and multilayer graphs to estimate arrival delay quantiles before and after departure. We decompose their value against a strong baseline into load (how much delay the network carries) and propagation (where it spreads relative to a flight). A precautionary intervention rule with break-even cost ratios turns this value into decisions. Hypotheses from one month of EUROCONTROL data are tested under pre-specified rules in six confirmatory weeks from 2023 and 2019: predictive value is confirmed in both years, pre-departure decision value in 2023, mainly in March. Before departure, network features reduce the upper-quantile loss in every test week, by 1.4–7.6%, also beyond recent delay levels. Once departure delay is known, their added predictive value is negligible. The propagation component improves the forecast in every week and cuts false alarms outside Christmas with missed delays nearly unchanged. In exploratory analyses, the load component adapts the intervention rate when weeks deviate from training, avoiding false alarms in lighter and missed delays in heavier weeks. There, in high-load hours, network features achieve almost half the missed-delay reduction of the observed departure, about one hour earlier, with more interventions. Under this rule, they pay off in all test weeks if a missed delay costs 5–25 false alarms (point estimates). Network information belongs before the decisive observation.