Uncertainty quantification of departure delay considering network properties and conformal prediction framework

Prediction of flight departure delay is an important task in air traffic flow management. In most cases, prediction is done for point estimate, and confidence is often overlooked. In this study, we use the Conformal Prediction (CP) framework to predict average departure delay and quantify its uncertainty using network properties, temporal factors, and lagged delays for major domestic hub airports in Japan. We investigate two core questions: which network-centric features most influence average departure delays using random forests, and how the CP framework quantifies uncertainty around those predictions. Random forests reveal that lagged delay drives average delay along with network measures such as in- and out-degree and betweenness centrality, and seasonal monthly effects strengthen predictive accuracy. Adding network properties increased the predictive capability. We then apply the CP framework to Japanese domestic airline flight delays (2018-2021), evaluating prediction intervals across uncertainty levels. Results highlight significant variations in adaptability and reliability among methods. At tight coverage targets, all methods face difficulties adapting quickly to delay fluctuations during-COVID phase, but it was easier at the pre-COVID phase. However, at moderate uncertainty levels, the Aggregate Adaptive Inference model effectively balances predictive reliability and responsiveness to changes in delay patterns. The fully adaptive approach further demonstrates how balancing interval precision with responsiveness can offer practical benefits, while less flexible methods exhibit challenges in handling abrupt delay spikes. These insights emphasize the importance of method selection based on operational context and acceptable uncertainty, providing practical guidelines for managing aviation delays in dynamic environments. Our study introduces CP methods in air traffic delay modeling and provides practical guidance for choosing uncertainty quantification strategies.