Abstract: As Internet search behavior reflects the needs and preferences of users, it can be used as a practical tool for demand forecasting. Therefore, this paper proposes a two-stage decomposition ensemble model with Internet search data (e.g. Baidu search index) for air passenger demand forecasting. The first stage of the proposed model is Internet search data preprocessing. By expanding the candidate set of the keywords, denoising the keywords time series, selecting appropriate keywords from that candidate set, and decomposing each series of the selected keywords into three components (i.e. seasonal factor, trend-cycle component and irregular component) by seasonal decomposition, three databases on keywords series are obtained accordingly. The second stage of the proposed model is prediction and evaluation. Firstly, the original air passenger demand time series is decomposed into three components by seasonal decomposition. Then, the three components are predicted independently and these prediction results of the components are combined as an aggregated output. The empirical results show that the proposed model achieves better forecasting performance than the benchmark models, and it can provide a valuable reference for making decisions in transportation management.

Key words: Internet search data, air passenger demand, data preprocessing, decomposition ensemble prediction