Time and feature varying tourism demand forecasting

Abstract

Choosing appropriate weights for individual models represents a major challenge in combination forecasting. Most research has used constant or time-varying weights during stable periods, ignoring dynamic weights that account for the latent features in multisource data during uncertain periods. We introduce an innovative approach that employs a time- and feature-varying ensemble learning–based meta-learner to consolidate individual model forecasts. The proposed model integrates statistical, machine learning, and deep learning models, along with economic and search engine data, to forecast visitor arrivals in Hong Kong and Sanya City, China. Results show that the proposed model surpasses most individual models and typical combination methods in stable and uncertain times. The findings highlight the proposed model's ability to yield consistent and reliable predictions across a variety of scenarios, particularly during volatile periods.