Multi-Term Tourism Demand Forecasting with Online Data

Abstract

As an important service-oriented industry, the recovery and prosperity of the tourism industry not only benefits the income growth of tourism practitioners, but also plays an important role in the economic development of the country and region. However, due to the significant coexistence of randomness and periodicity in tourism demand, practitioners need to fully consider the long-term, short-term, and real-time tourism demand in decision-making. Therefore, it is important to conduct forecasting for tourism demand in different terms.

The current application of big data has led to an unprecedented enrichment of online tourism data resources. Online data can effectively reflect tourist behavior and tendencies, so online data has the potential in tourism demand forecasting. However, tourism managers mainly face several problems in processing online data. Firstly, the tourism demand data under different terms exhibit differentiated time series data characteristics (complex seasonality, holiday effects, real-time spatial-temporal effects, gravity effects, etc.), making it difficult to use a unified method for processing and forecasting. In addition, the influencing factors vary in different terms, and the changes reflected in tourism demand are also different. Therefore, selecting data with better feature capture capabilities and proposing more accurate and robust forecasting methods for different tourism demand terms has significant theoretical and practical significance.

This study characterizes different data characteristics and influencing factors of long-term, short-term, one-day-ahead, and real-time tourism demand. Then, four new approaches for forecasting tourism demand in different terms based on online data are proposed. The four proposed methods effectively capture the characteristics of tourism demand in different terms and improve the forecasting performance. Besides, the findings provide practical guidance for tourism practitioners, government, tourists, and others.

Specifically, this study has four innovations:

(1) Long-term tourism demand forecasting based on multi-source online data. Traditional long-term tourism demand forecasting methods often struggle to fully utilize low-predictive online data, limiting the extraction of influential factors. To address these limitations, this study proposes an integrated forecasting model based on tourist behavior and word-of-mouth effect theories. Online search data and tourist review data are employed as proxy variables for consumers' long-term tourism preferences. The proposed SAE-VMD-NN model is constructed using Variational Mode Decomposition (VMD), Stacked Autoencoder (SAE), Natural Language Processing (NLP), and Artificial Neural Network (NN). This study achieves the first-time information extraction and efficient application of low-correlation multi-source online data in long-term tourism demand forecasting research. The proposed model also addresses the cross-frequency issue and effectively improves the ability to capture long-term tourism demand characteristics and related prediction performance.

(2) Short-term tourism demand forecasting approach considering complex seasonality and holiday effects. Traditional methods often fail to consider complex seasonality and holiday effects. This study proposes an integrated FA-Prophet forecasting model that incorporates complex seasonality and holiday effects. The model utilizes short-term web search data and online weather forecast data as proxy variables for consumers' short-term tourism demand. Empirical results demonstrate that the proposed method effectively improves the accuracy of short-term tourism demand forecasting. Additionally, weather forecast data and online data are shown to significantly enhance prediction performance. The proposed model addresses the limitations of traditional algorithms that struggle to balance complex seasonality and holiday effects.

(3) One-day-ahead tourism demand forecasting based on different tourist search behavior. Previous studies have typically used search engine data as proxy variables for tourist behavior. However, due to diverse search motivations, search engine's predictive power for one-day-ahead tourist demand behavior is limited. This study systematically compares the performance of map search data and search engine data based on tourist behavior and characteristics of map search data in one-day-ahead tourism demand forecasting. We propose a tourism demand forecasting model based on Bayesian Model Averaging. For the first time, this study validates the capability of map search data in predicting one-day-ahead tourism demand and introduces a novel model based on statistical integration methods.

(4) Real-time tourism demand forecasting for multiple attractions based on gravity and spatial-temporal effects. Previous studies concerning real-time tourism demand forecasting for multiple attractions have preponderantly relied on data-driven machine learning, frequently neglecting the idiosyncratic traits of tourist behavior within the tourism industry. This study, which mainly based on the theory of gravity effect and spatial-temporal effect, utilizes IoT real-time data and exogenous variables to propose a real-time tourism demand forecasting model for multiple attractions, which is called G-STGCN, considering both gravity effect and spatial-temporal effect. Specifically, this model introduces an adjacency matrix model optimized based on gravity effect and the model integrates several exogenous variables. The approach effectively captures the spatial-temporal relationships of tourist behavior, and improving the forecasting performance of multiple attractions. Additionally, this study further demonstrates the impact of attraction attributes on tourist behavior and model forecasting performance.

Date of Award	2 Jan 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Gengzhong Feng (External Supervisor), Hoi Shou Alan CHAN (Supervisor), Kwok Leung TSUI (Co-supervisor) & Kwai Sang CHIN (Co-supervisor)