The limitation of machine-learning based models in predicting airline flight block time

Abstract

This study presents three different machine learning (ML) models to estimate the flight block time for commercial airlines. The models rely only on explanatory variables that airlines would know when they were planning their schedules several months prior to the actual flight. Historical Actual Block Time (ABT) data is collected for one-way flights in seven airport pairs in the domestic U.S. market for 2019. The variability of the ABT and its components for these airport pairs is presented. The main features that affect this variability are investigated. The results confirm that seasonality, aircraft type, departure/arrival hour, and airport congestion are significant variables in partially explaining ABT variations. Overall, the results show that the considered ML models have limited capability in predicting the ABT. The prediction accuracy of these ML models is compared against a benchmarking scenario, where the median value of the historical ABT is used as an estimate for the block time. It is found that the median-based approach provides better performance compared to the ML models. A sensitivity analysis is performed to evaluate the risk of flight delay against different levels of block time padding. The possible trade-off between the block time padding and the expected flight delays is presented. Finally, the study evaluates and compares different airlines' block time padding strategies in the different airport pairs. Results show significant discrepancies among airlines concerning setting the scheduled block time (SBT) in the same airport pair. However, it is difficult to confirm whether airlines are purposely adopting padding strategies, or they lack the ability to optimize their block time padding against expected on-time performance.