Forecasting Domestic Air Passenger Demand in Türkiye: A Comparative Analysis of Machine Learning and Time-Series Approaches

Accurate forecasting of domestic air passenger demand can support transportation planning, infrastructure development, and operational decision-making. However, passenger demand may be influenced by economic, tourism-related, technological, and mobility-related factors, making forecasting challenging. This study develops a multivariate forecasting framework for domestic air passenger demand in Türkiye using monthly data from 2001 to 2025. SARIMAX is evaluated alongside ARIMA, SARIMA, ARIMAX, ANN, and LSTM using a common expanding-window walk-forward validation framework. Forecasting performance is assessed using MSE, RMSE, MAE, and R2, together with the Diebold–Mariano test to examine differences in squared-error-based forecast accuracy. An ablation analysis is conducted for the ANN model to examine changes in forecasting performance associated with excluding individual explanatory variables. ANN and LSTM are also evaluated over 20 independent runs to assess run-to-run variation. The results indicate that SARIMAX produced the lowest MSE (0.00264), RMSE (0.05142), and MAE (0.03205) on the normalized evaluation scale, whereas ANN produced the lowest MAPE on the original passenger-demand scale. SARIMAX also required less computational time than the neural-network models under the experimental configuration. Overall, the findings indicate that forecasting performance varied across accuracy measures and computational requirements, highlighting the value of complementary evaluation criteria when comparing alternative forecasting approaches.