A rolling-origin ensemble forecasting framework for tourism revenue: elastic net and CatBoost – the case of Türkiye (2002–2024)

Purpose This study develops and evaluates a multivariate ensemble forecasting framework for tourism revenue that combines Elastic Net and CatBoost under rolling-origin validation, using Türkiye (2002–2024) as an illustrative case to generate medium-term reference scenarios. Design/methodology/approach Tourism revenue was the dependent variable, with international visitors, Blue Flag beaches, ministry-certified facilities, travel agencies, museums/historical sites, and museum-site visitors as explanatory indicators. The multicollinear multivariate structure was modeled using Elastic Net and CatBoost and evaluated via rolling-origin (expanding-window) cross-validation using RMSE, MAE, and R2. A benchmark ARIMA model was also tested under the same evaluation scheme for comparison. Findings Elastic Net delivered the best out-of-sample accuracy (R2 = 0.894; RMSE = 4.02; MAE = 2.80), while CatBoost produced a smoother, more conservative trajectory. Standardized Elastic Net coefficients and CatBoost SHAP values consistently identify visitor volume (VIS) and museum visitors (MV) as the dominant drivers. The equal-weight ensemble yields a balanced 2025–2027 nominal reference path with a modest post-pandemic increase. The benchmark ARIMA (2,1,2) performs substantially worse across error metrics, supporting the superiority of the proposed multivariate ML framework. Originality/value The study contributes a generalizable forecasting workflow that (1) benchmarks regularized linear and gradient-boosted models under a common rolling-origin design and (2) shows that a transparent simple-average ensemble yields more robust medium-term reference scenarios under shocks and structural breaks. Türkiye is used as a demonstration case.