Capacity and competition aware machine learning for airline route passenger traffic forecasting and operational decision support

Abstract Accurate route-level realized passenger traffic forecasting is important for near-term airline operational monitoring, utilization assessment, and competitive decision support. This study develops a capacity- and competition-aware machine-learning framework at the carrier–route–month level. The methodological design combines exact calendar-month lag and target construction, training-only feature selection, chronological validation, temporal baselines, feature-family ablation, and multiple operational prediction tasks. Missing months and the excluded disruption period are not bridged when constructing predictors or targets. Using U.S. BTS T-100 records, linear, ensemble, boosting, and neural models were evaluated on chronologically later observations. XGBoost achieved the strongest primary one-month passenger-traffic performance, with a WMAPE of 11.91%, RMSE of 1,434.92 passengers, and R 2 of 0.9742, reducing WMAPE by 16.72% relative to the last-month naive benchmark. It also achieved 8.37% WMAPE and R 2 of 0.8309 for next-month load-factor forecasting. XGBoost and HistGradientBoosting produced nearly identical leading utilization-state classification results, with approximately 73.37% accuracy and a macro-F1 score of 0.7413. Capacity variables provided the largest improvement over traffic history, while competition and co-airport variables produced smaller, task-specific gains. Reviewer-requested robustness analyses showed a 9.39% relative WMAPE increase when contemporaneous information was removed, regime-dependent rolling-origin performance, monotonic degradation on a common multi-horizon sample, and near-nominal split-conformal marginal coverage of 89.58% and 94.79% for nominal 90% and 95% intervals. These findings support calendar-aligned gradient boosting primarily as a near-term passenger-traffic forecasting and utilization-screening tool.