Leveraging online reviews for hotel demand forecasting: A deep learning approach

Abstract

The use of online reviews for forecasting hotel demand has gained increasing interest in recent years. However, prior studies have primarily focused on sentiment information and do not capture sufficient signals for accurate hotel demand forecasting. Furthermore, the complex feature interactions within multivariate time series complicate hotel demand forecasting. Guided by systematic functional linguistics (SFL) theory, this study proposes an analytic framework consisting of ideational, textual, and interpersonal functions to extract signals from online reviews. Besides, we propose a novel long short-term memory interaction-based convolutional neural network (LICNN) model for hotel demand forecasting. The results indicate that incorporating online review features reduces root mean squared error (RMSE) by at least 2.2 % and at most 46.6 %, mean absolute error (MAE) by at least 3.2 % and at most 44.8 %, and mean absolute percentage error (MAPE) by at least 3.5 % and at most 44.6 %. Moreover, compared with baseline models, our proposed LICNN model achieves the lowest RMSE (at least 15.8 % and at most 53.1 % improvements), MAE (at least 8.1 % and at most 56.1 % improvements), and MAPE (at least 12.9 % and at most 44.8 % improvements). The ablation study highlights the value of extracting feature interactions in demand forecasting.