Tourism demand point-interval forecasting using global–local information extraction network

Abstract

The forward-looking estimation of tourism demand is pivotal for effective resource allocation and strategic planning, making robust predictive models valuable decision-making tools. Current research predominantly emphasizes isolated types of information, often neglecting comprehensive integration of spatiotemporal data and auxiliary factors, as well as the critical role of interval forecasting in decision support. To address these gaps, we propose the Global–Local Information Extraction Network (GLIEN), a novel deep learning model for point-interval forecasting. GLIEN dynamically integrates spatiotemporal information and auxiliary factors at a global level, followed by precise analysis of each city within the global information through the designed local blocks. During the information integration phase, enhanced by an error correction mechanism, the model generates both point and interval forecasts under global information adjustments. To mitigate challenges associated with small sample sizes in tourism datasets, our proposed K-fold training strategy enhances the model’s capacity to absorb data diversity to some extent. Empirical analysis of the Hainan and Hawaii datasets demonstrates that the GLIEN outperforms all benchmark models in both point and interval forecasting across different forecast horizons. Results also highlight the error correction strategy’s role in refining interval coverage and bandwidth, while the K-fold training significantly boosts forecasting accuracy. This research offers critical insights for tourism resource management and planning, marking the first attempt at point-interval forecasting for tourism demand.