Tourism demand forecasting with multi-source data: a hybrid framework integrating denoising, signal decomposition, and machine learning

Abstract Accurate short-term tourism-demand forecasting supports operational planning and sustainable destination management. Existing hybrid approaches frequently combine signal decomposition with machine-learning models, yet preprocessing steps are often applied asymmetrically, and their individual contributions remain unclear. This study proposes a two-sided operator framework that integrates predictor denoising and target decomposition within a unified denoise–decompose–learn structure. Both operators are implemented under a leakage-free expanding-window protocol, and performance is evaluated in a stage-wise manner so that incremental gains can be identified empirically. The framework is tested using weekly tourist arrivals to Mount Siguniang from 2016 to 2020 (195 observations), together with search-engine, online-review, and weather predictors. Variational Mode Decomposition (VMD) combined with Empirical Wavelet Transform (EWT) denoising and XGBoost delivers the strongest overall performance. The VMD–EWT specification achieves an MAE of 2,873.94, an RMSE of 3,511.18, and $$R^{2}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup> <mml:mi>R</mml:mi> <mml:mn>2</mml:mn> </mml:msup> </mml:math> of 0.9276, outperforming both classical time-series benchmarks and standalone machine-learning models. Compared with the raw XGBoost baseline, the combined model produces statistically significant reductions in large forecast errors across hard-week subsets, with persistent Diebold–Mariano significance under both absolute and squared loss. Component-wise analysis further shows that frequency-constrained decomposition redistributes predictive information across scales: high-frequency components exhibit strong autoregressive memory, mid-frequency components integrate behavioural and environmental signals, and lower-frequency components capture structural variation. The results indicate that frequency-aware decomposition is the principal source of tail-risk reduction, while predictor denoising provides additional refinement during volatile periods. The proposed framework offers a modular and reproducible structure for handling heterogeneous and noisy predictors in single-destination forecasting contexts.