Stacked autoencoder with echo-state regression for tourism demand forecasting using search query data

摘要

Accurate tourism demand forecasting is fundamental in the tourism industry, while effective tourism demand forecasting using search query data (SQD) has become popular in the tourism management field. SQD is a type of statistical time series provided by search engines that can reflect netizens' attention on certain events. Scholars attempt to establish a reasonable relationship between tourism demand and SQD for its timeliness and comprehensiveness. The current study proposes an effective deep learning technique called stacked autoencoder with echo-state regression (SAEN) to accurately forecast tourist flow based on search query data. In the proposed SAEN approach, stacked autoencoder is adopted to hierarchically learn high-level predictive indicators from substantial SQD and connected by an echo-state regression layer to model the nonlinear time series relationship between tourism flow and the learned indicators. Four realistic applications (i.e., one comparative case and three extended cases in the US and China with different SQD sources) are used to verify the forecasting performance of SAEN. Numerical results indicate that SAEN is better than the current literature findings, including time series approach, econometric model, common machine learning algorithms, and state-of-the-art deep learning techniques. The structure parameters of SAEN are further analyzed empirically and theoretically. Moreover, this study determined a different impact of network depth and echo-state reservoir scale on the performance of SAEN. The proposed SAEN can be an appropriate alternative for tourism demand forecasting in complex data situations. (C) 2018 Elsevier B.V. All rights reserved.