Airline operations are highly sensitive to day-ofoperations disruptions, especially during aircraft turnaround, where local events may accumulate and propagate across the network. Accurate prediction of departure delay is therefore essential for tactical operational support and disruption management. This paper proposes a novel event-driven machine learning framework for real-time probabilistic prediction of departure
delay using operational event information inferred from historical flight data. Delay uncertainty is represented through conditional quantiles, directly supporting tactical planning and control actions in Airport Collaborative Decision-Making (ACDM)
environments. Predictions are updated dynamically as operational conditions evolve, including weather-related, airspace, carrier-, and airport-related events. Experimental results show that the proposed event-driven setting outperforms a conventional non-event-driven setting across all evaluated models. Among the tested methods, the Deep Neural Network (DNN) achieves the best predictive accuracy, with an Overall Mean Pinball Loss (OMPL) of 7.89 min, while CatBoost provides the best trade-off between predictive performance, runtime, and interpretability. Finally, SHAP analysis is used to identify the main drivers of the predictions and support more informed real-time operational actions.