A Generalisable Machine Learning Framework for Taxi Time Prediction at A-CDM Airports

Accurate taxi time estimates are central to the Air- port Collaborative Decision Making (A-CDM) framework, where the sharing of timely information underlies network efficiency and predictability. Many airports still rely on historical averages as operational baselines. However, these values fail to capture the variability inherent to daily operations. This paper presents a generalisable machine learning model based on Gradient Boosted Decision Trees (GBDT) for both taxi-out (AXOT) and taxi-in (AXIT) operations, trained on 4.1 million flights (2022–2024) across six European A-CDM airports. The model draws on standard airport data sources, combining AODB operational airport records with METAR weather observations and aircraft characteristics from the Base of Aircraft Data (BADA). Evalua- tion against operational baselines (Standard Taxi Times) shows MAE reductions of 30–44% at regional airports and 1–6% at major airports, with typical mean absolute percentage errors of 17–20%. Feature importance analysis confirms the crucial role of airport infrastructure and operational factors, supporting the cross-applicability of this approach with minimal local calibration. The approach offers a practical path to improving the accuracy of shared operational information, strengthening both local decision support and network-wide flow management.