Smart tourism demand forecasting using machine learning and behavioral travel data

Tourism demand forecasting is central to smart destination governance, but model comparisons can be overly optimistic when feature construction, hyperparameter tuning, or model selection uses information unavailable at forecast time. This study develops a time-aware machine-learning forecasting model that is reproducible and leakage-free, based on provincial tourism data in Thailand on a monthly basis. The data sources consist of 30,800 long-format observations for 77 provinces, 5 regions and 50 months (January 2019 to February 2023). The province-month records of the long-to-wide transformation were 3850. Eliminating a 12-month historical feature requirement reduced 924 target rows (24.0%), leaving 2926 modeling observations and 183 predictors created from calendar, spatial, regime, lagged, rolling, and historical growth information. Eight regression techniques were tested for total tourist arrivals, tourism revenue and tourist stay ratio. Expanding-window validation was used to select hyperparameters and model families only within the pretest period, and July 2022 to February 2023 was kept as an untouched final test. Random Forest, LightGBM and Gradient Boosting were chosen for model fits of arrivals, revenue and stay ratio, respectively, with final-test R² values of 0.8456, 0.6834 and 0.3327, respectively. Friedman tests revealed significant differences overall between the evaluated models for all three targets, while paired Wilcoxon tests (with Holm’s adjustment) showed no significant superiority of the selected models over their alternatives across the 6 rolling origins. Direct forecasts also exhibited increasing prediction errors along the forecasting horizon, from one to eight months. The global feature importance and the local SHAP explanations were connected to forecasts with recent tourist arrivals, domestic travel, revenue persistence, temporal regime and destination size. What the contribution is about is not the absolute superiority of a single model, but an auditable information-set design, time-aware validation, multi-target and multi-horizon evaluation and interpretable decision support.