A Data-Centric Framework for Forecasting Tourist Flows from Traffic Sensors

The paper develops a data-centric framework for forecasting tourist flows using high-frequency traffic-sensor data as a proxy for mobility and tourism dynamics. Two related datasets are considered: a benchmark based on hourly motorway-toll flows and a large regional monitoring system for the Aosta Valley. Several machine-learning and deep-learning models are compared, including Support Vector Regression, Random Forest, XGBoost, and LSTM networks. In highly stationary settings, moderate-complexity machine-learning models, especially tree-based ensembles, outperform deeper architectures in predictive accuracy, computational efficiency, and interpretability, while deep models tend to smooth sharp peaks. The analysis is extended to adaptive ensembles whose model weights change with recent forecasting performance, time-of-day and seasonal indicators, and data variability. Preliminary evidence suggests that these adaptive combinations reduce forecasting errors during abrupt changes, seasonal transitions, and exogenous shocks, providing a robust and interpretable framework for real-time tourism and mobility forecasting.