Probabilistic Pre-tactical Arrival and Departure Flight Delay Prediction with Quantile Regression

Airports plan their resources well in advance based on anticipated traffic. Currently, the only traffic information accessible in the pre-tactical phase are the flight schedules and historical data. In practice, however, flights do not always depart or arrive on time for a variety of reasons, such as air traffic flow management or reactionary delay. Because neither air traffic flow management regulations nor aircraft rotations are known during the pre-tactical phase, predicting the precise arrival and departure delay of individual flights is challenging given current technologies. As a result, probabilistic flight delay predictions are more plausible. This paper presents a machine learning model trained on historical data that learned the various quantiles of the departure and arrival delay distributions of individual flights. The model makes use of input features available during the pre-tactical phase, such as the airline, aircraft type, or expected number of passengers, to provide predictions of the delay distribution several days before operations. The performance of the model trained on operational data from Geneva airport is compared to a statistical baseline, providing evidence that machine learning is superior. Furthermore, the contribution of the various input features is quantified using the Shapely method, stressing the importance of the expected number of passengers. Finally, some examples are presented to illustrate how such a model could be applied in the pre-tactical phase.