The global commercial air transportation network is a complex and centrally managed system and as such highly vulnerable to disturbances. When flights are affected by such events, their impacts can easily propagate across the network, facilitated by shared airline and airport resources. Event propagation typically leads to system-wide inefficiencies and passenger inconvenience. Mitigating these cascading effects requires accurate and early prediction, optimally as soon as an event occurs. Focusing on enhancing tactical network planning, this paper proposes an event-based machine learning framework for the probabilistic prediction of arrival delays, explicitly incorporating real-time operational data from both ground and airborne phases of flight. The proposed framework was implemented using the CatBoost and LightGBM ensembles, both of which significantly outperformed established statistical models. A detailed comparison revealed a performance-efficiency trade-off: CatBoost yielded the highest predictive accuracy, while LightGBM trained up to seven times faster, offering the best computational efficiency with a minimal performance loss. SHAP analysis reveals that event features belong to the primary drivers of the model decision, confirming the framework's value across different scenarios.