Comparative Analysis of Statistical and Machine Learning Models for Air Passenger Demand Forecasting: Evidence from Domestic and International Passenger Segments
TL;DR
This study benchmarks four forecasting approaches—SARIMA, SARIMAX, Prophet, and XGBoost—on monthly domestic and international passenger traffic data from 2019 to 2024, using a 12-month hold-out test set and a 36-month projection horizon, revealing stark segmentation in model suitability.
Abstract
Accurate forecasting of air passenger demand is critical for aviation sector planning, yet model selection remains contested when data exhibit structural discontinuities and segment-specific behavior. This study benchmarks four forecasting approaches—SARIMA, SARIMAX, Prophet, and XGBoost—on monthly domestic and international passenger traffic data from 2019 to 2024, using a 12-month hold-out test set and a 36-month projection horizon. Model performance was evaluated using MAE, RMSE, and MAPE, with stationarity confirmed via Augmented Dickey-Fuller tests.Results reveal stark segmentation in model suitability. For international passengers, SARIMA achieved the lowest test MAPE (8.66%), substantially outperforming XGBoost (14.23%) and Prophet (57.65%), attributable to the strong first-order autocorrelation structure of post-recovery international traffic. For domestic passengers, SARIMA and SARIMAX both exhibited explosive forecast divergence (MAPE exceeding 11,900%), rendering them operationally unusable; XGBoost (MAPE 14.87%) emerged as the most reliable alternative by capturing non-linear interactions among lag features, rolling statistics, and annual trend components.The central finding concerns the role of exogenous variables: SARIMAX produced metrics identical to SARIMA for international passengers and marginally worse for domestic passengers, yielding no incremental predictive gain in either segment. This challenges the assumption that exogenous augmentation inherently improves seasonal time series models—such benefits are conditional on the relevance, stability, and non-redundancy of external regressors relative to existing seasonal components. These findings offer empirical guidance for segment-specific model selection in aviation demand forecasting and caution against uncritical adoption of SARIMAX when core seasonal structure is already sufficient.