Airlines require delay forecasts that translate into scheduled actions because late arrivals can propagate through aircraft rotations. This study links airport-network forecasting with flight-level schedule-buffer design. Using 2024 Bureau of Transportation Statistics on-time records for 49 continental U.S. hubs merged with hourly METAR weather, we construct a leakage-controlled airport-hour panel and represent airport dependence with geographic and scheduled-flow graphs. A spatiotemporal propagation network estimates the destination-hour arrival-delay risk, split-conformal calibration audits residual uncertainty, and a rotation-aware newsvendor rule converts the forecast into a scheduled-arrival buffer based on idle slack, primary delay, and same-tail downstream exposure. In the held-out November-December operations, STPN achieves the lowest airport-hour MAE among the recurrent and graph baselines, although the RMSE and F1 are mixed. Static conformal intervals are sharp but show undercoverage during the holiday shift; a rolling conformal variant restores near-nominal coverage. Under literature-based delay-cost assumptions and signed latent delay-exposure accounting, the policy shifts the median buffer from 9.0 to 22.9 min and reduces a retrospective proxy fleet delay-exposure cost by 16.0% across 650,091 mapped test flights, with positive savings in 33 of 36 cost scenarios.

Keywords

Flight delay prediction, graph neural networks, conformal prediction, schedule buffers, and airline operations.