Are customer reviews just reviews? Hotel forecasting using sentiment analysis

Abstract

Despite the important role of the sentiment of customer reviews in the tourist decision-making process, it has received limited attention from the field of tourism demand forecasting. This study aims to examine the sentiment information of customer reviews and explore its potential in enhancing hotel demand forecast. Empirically, four Macau luxury hotels are selected and their customer reviews are crawled from two popular online platforms. A deep learning method, the Long Short-Term Memory model, is used to extract sentiment information from consumer reviews. Three sentiment indices, namely, bullish index, average index and variance index, are constructed and examined. The effectiveness of these sentiment indices is further evaluated by the autoregressive integrated moving average with exogenous variables model. Empirical results indicate that the inclusion of the sentiment indices helps with improving forecast accuracy. The findings of this study further emphasise the importance of textual content to hotel practitioners in terms of strategy formulation, revenue management and competitiveness enhancement.

Keywords

sentiment indices big data hotel demand forecasting online review long short-term memory model

Introduction

Demand forecasting is a core element of hotel revenue management. Accurate forecasts of hotel reservation, occupancy and guest arrivals are crucial for hoteliers to allocate hotel resources efficiently and develop appropriate price strategies (Dergiades et al., 2018; Frechtling, 2001; Song and Li, 2008; Weatherford and Kimes, 2003). A large body of studies has investigated hotel demand forecasts using various techniques, with a primary focus on low-frequency demand series such as weekly or monthly. In those studies, traditional economic variables, including income, tourism prices and exchange rates, are used as independent variables in the modelling process (Wu et al., 2017a). More recently, with the rapid development of information technology, the rich content from the internet has introduced an important opportunity to upgrade tourism-related data and generate timely demand forecasts. Big data from the internet can reflect tourists’ preference and their decision-making progress in real time, providing abundant and high-quality information for demand forecasting research (Yang et al., 2014). Tourism-related big data generated by tourists, such as web search data or social media data, have been applied to hotel and tourism demand forecasting research. Choi and Varian (2012) introduced Google Trends data as new explanatory variables to model tourism demand, revealing the considerable potential of search data in tourism demand forecasting. Pan and Yang (2017) adopted multi-source data, including web search data, weather data and local website traffic data, to construct an accurate forecasting model of hotel occupancy, emphasising the superiority of big data in tourism demand analysis. In addition to web search data, social media data are also proved to be effective in improving forecasting accuracy (Li et al., 2020; Önder et al., 2019a).

Among the applications of big data in tourism literature, the information of text content has not been widely used for hotel forecasting. According to Ady and Quadri-Felitti (2015), 95% of travellers search for online hotel reviews, and one-third of them consider the opinions expressed in reviews before making a booking decision (Hu et al., 2017). Therefore, online hotel reviews can reflect tourist behaviours to some extent and social media textual data could be considered when modelling and anticipating tourists’ needs. Text content analysis of social media textual data such as online reviews has been increasingly used to understand tourist perception such as experience, guest satisfaction and helpfulness, allowing hoteliers to meet guest needs in a timely and specific manner (Ghose and Ipeirotis, 2011; Xiang et al., 2015; Xu and Li, 2016).

Sentiment analysis is an effective way to capture tourist sentiment and emotional contagion from online textual reviews. Compared with traditional tourism demand determinants, sentiment analysis has three advantages. Firstly, it can capture tourist preference by extracting tourist sentiment. Secondly, it is capable of predicting tourist behaviour because sentiment in the review serves as an emotional contagion that delivers a positive or negative signal from previous tourists to prospective tourists (Pugh, 2001; Subramony and Pugh, 2015). Finally, it quantifies the unstructured big data using text mining techniques.

Sentiment analysis has been applied in various fields such as product sales forecasting (Fan et al., 2017; Yu et al., 2010; Zhang et al., 2020) and finance forecasting (Audrino et al., 2020; Nti et al., 2020). However, the use of sentiment analysis has not elicited considerable attention from tourism and hotel demand forecasting literature, with Önder et al. (2019b) and Colladon et al. (2019) as two exceptions. In Önder et al. (2019b), texts from online news media coverage were rated as either positive or negative via automated semantic routines. They made an initial attempt to use web sentiment of news media coverage to predict tourist arrivals of four European cities, demonstrating a superior forecast accuracy by incorporating web sentiment data into forecast models. Colladon et al. (2019) used review text from travel forums of TripAdvisor to make tourism demand predictions. The machine learning algorithm included in the software Condor was used to classify the sentiment polarity of community posts and calculate the sentiment value. The results have shown that sentiment-based text data can be used as an important explanatory variable in a tourism demand model to improve forecast accuracy. Önder et al. (2019b) and Colladon et al. (2019) demonstrate that online sentiment-based text data can be regarded as an effective data source and possess a great potential to improve tourism demand forecasts. With the rapid and diverse development of information technology, taking advantage of big data analysis and gaining useful insight from various user-generated content data is essential (Li et al., 2018a). However, no study has considered the use of sentiment analysis for hotel demand forecasting. The present study attempts to explore the predictive ability of the sentiment feature extracted from customer reviews for hotel demand forecasting.

Techniques of sentiment analysis have been further developed and improved. With developments in computer science, some deep learning models, such as the Long Short-Term Memory model (LSTM), have been used for sentiment polarity judgement, achieving promising results (Priyadarshini and Cotton, 2021; Sun et al., 2017). Compared with sentiment analysis techniques used in Önder et al. (2019b) and Colladon et al. (2019), LSTM has the advantages in handling sequences of any length and learning the temporal dependence from the data through the cell state (Hochreiter and Schmidhuber, 1997), which is well suited for processing sequential data. Thus, this study is designed to use LSTM for sentiment analysis.

Sentiment analysis has been used in the tourism context, but existing studies only extract the overall sentiment polarity of customer reviews. Further sentiment information such as diversity has not been considered. Therefore, this study constructs three sentiment indices to capture different aspects of the review sentiment, namely, the bullish sentiment index (BSI), the average sentiment index (ASI) and the variance sentiment index (VSI). The BSI borrows the idea from the stock market, which can be used to reflect the sentiment orientation of tourists on a particular day. The ASI and VSI are measured to evaluate the central tendency and dispersion of daily sentiment conveyed by tourists. The objective of the present study is threefold. Firstly, this study explores the potential of the sentiment feature extracted from customer review text in improving the accuracy of hotel demand forecasts. Secondly, this study examines the effectiveness of the LSTM algorithm as a tool for sentiment index construction. Lastly, three sentiment indices, namely, BSI, ASI and VSI, are constructed and examined for their performances in hotel demand forecasting.

Literature review

Hotel and tourism demand forecasting with big data

Accurate demand forecasts have recently received increasing attention from academic and industry practitioners in hotel and tourism demand forecasting (Li and Wu, 2019). Many empirical studies examine forecasting accuracy using more advanced forecasting techniques, such as combination forecasting (Li et al., 2019a), multivariate forecasting (Chen et al., 2019a) and pattern recognition forecasting (Hu et al., 2021). Moreover, AI-based techniques are also attracting increasing attention from the field of tourism demand forecasting since 2000 (Song et al., 2019). In recent years, tourism industry practitioners have focused more on quantitative evidence and used it as a reference for strategy and policy formulations (Li and Wu, 2019). With improved understanding and increasing availability of online big data, many researchers have attempted to forecast demand for hotels and tourism using online data. Four types of internet big data have been applied in hotel and tourism demand forecast literature, namely, web search and website traffic data, social media statistics, online textual data and online photo data. The most prevailing data type used in tourism demand forecasting is web search data, which reflects tourists’ preference and their decision-making progress by focussing on the records of web searching behaviour (Dergiades et al., 2018; Ghose et al., 2014). Pan et al. (2012) explore the usefulness of search query volume in hotel demand forecasting using five tourist-related Google search queries in the context of 110 hotels in Charleston, South Carolina. Bangwayo-Skeete and Skeete (2015) integrate a mixed-data sampling approach into the tourism demand modelling process to incorporate weekly Google search queries of ‘hotel and flights’ with monthly tourist arrivals to the Caribbean countries. Rivera (2016) uses a dynamic linear model to forecast tourism demand in Puerto Rico with the integration of Google Trends data and hotel registrations. Hu and Song (2020) further combine Google search data with traditional economic variables, such as tourist income, exchange rate and prices, to predict the volume for tourist arrivals from Hong Kong to Macau. Their findings suggest further enhancement on the effectiveness of big data in combination with causal variables in tourism demand forecasting practice.

Social media statistics are also frequently used in recent forecasting literature. Gunter et al. (2019) innovate by using the number of ‘Likes’ on Facebook posts for tourism demand forecasting. The study has shown that Likes data, as one important category of Facebook statistics, can enhance forecast accuracy. In a similar study on Austrian cities, Önder et al. (2019a) further confirm the predictive power of the Likes data on tourism demand. Some studies have sought to forecast hotel and tourism demand with online review data, but these attempts are mainly restricted to the use of numerical variables. Li et al. (2020) incorporate Baidu search engine data with online review data (volume and average rating) from Ctrip and Qunar to forecast weekly tourist arrivals to Mount Siguniang, finding an improved forecasting performance with data from multiple platforms. Khatibi et al. (2020) generate predictions for the counts of visitation to national parks in the United States and museums in the United Kingdom, with the application of review counts crawled from TripAdvisor. The findings suggest that the prediction based on the combination of review counts and environmental features outperforms other counterparts. Concerning the cancellation forecasting of hotel reservations, Antonio et al. (2019) adopt the award-winning ensemble tree-based machine learning algorithm, namely, the XGBoost model, to make hotel cancellation predictions using the volumes and ratings of online review data from multiple sources such as Booking.com and TripAdvisor. In addition to the counts of Likes and reviews, the metadata of online photos have been used in forecasting practices in hospitality research (Chen et al., 2019c; Colladon et al., 2019).

Despite the frequent appearance of online review counts and ratings in tourism demand forecast literature, only a few studies have investigated the usefulness of internal information of the review text in hotel and tourism demand analysis. Önder et al. (2019b) extract the sentiment information of online news media coverage via automated semantic routines to predict tourist arrivals to four European cities. The results indicate that the inclusion of sentiment-based textual data further improves the accuracy of tourism demand forecasts. Starosta et al. (2019) use a single-output neural network to extract the sentiment of news coverage in the German-speaking media and construct sentiment index as a proxy variable to supplement tourism demand information, revealing a significant correlation between the sentiment information of news coverage and tourism demand. More recently, studies that explore the predictive power of review textual data from travel social media are emerging. Colladon et al. (2019) use travel forum data that contains review text to predict international arrivals to seven European capital cities. They applied methods and tools of social network and semantic analysis to investigate user-generated content data retrieved from TripAdvisor. By considering three dimensions of online social interactions (Gloor et al., 2017), namely, degree of interactivity, degree of connectivity and language use, and using methods of social network and semantic analysis, Colladon et al. (2019) investigate the performance of user-generated contents in tourism demand forecasting process. Sentiment-based textual data from travel social media are further confirmed to have great potential to make predictions and can be regarded as an important indicator for the understanding of hotel and tourism demand.

Online textual reviews analysis in hotels

Comparing with review ratings and review statistics, online textual reviews of hotels include richer information, which reflects customer perceptions and experience (Xu and Li, 2016). Online textual reviews can serve as a feedback mechanism to help hoteliers anticipate the demands of guests and, therefore, increase the competitiveness of the hotel. Some studies have focused on the text content of the reviews to find the salient attributes on guests’ perceptions of their hotel stay experience (Berezina et al., 2016; Zhao et al., 2019). Specifically, text analysis methods, including latent semantic analysis (Xu and Li, 2016) and content analysis (Zhang and Mao, 2012), are used to explore and examine the effectiveness of these attributes. Empirical results of these studies show that textual review mining can effectively help practitioners to understand and anticipate the needs and preferences of guests, thereby providing significant business value in the big data era (Berezina et al., 2016; Zhao et al., 2019). Apart from identifying the attributes of online textual reviews of hotels, some studies also explore the potential of textual reviews on hotel room sales forecasting (Blal and Sturman, 2014), language style (Wu et al., 2017b), review helpfulness (Ma et al., 2018) and the relationship between textual reviews and online rating (Geetha et al., 2017; Li et al., 2019b). Among others, the identification of reviewers’ sentiment forms the basis of text analysis in most research. In particular, Geetha et al. (2017) investigate the relationship between online customer reviews and customer rating by analysing the sentiment polarity of reviews. Gao et al. (2018) consider online reviews as opinion-rich resources and propose a novel sentiment analysis model to extract comparative relations from online reviews and to construct competitiveness analysis that helps restaurants to identify top competitors and develop improvement strategies accordingly. Zhang et al. (2016) examine a sentimental interplay between reviews and ratings and reveal that satisfied or neutral customers are more inclined to show emotional signals. Accordingly, sentiment information contained in customer reviews can be regarded as a meaningful information resource to identify latent needs and behavioural preferences of tourists.

Given that online textual reviews are opinion-rich resources that can effectively reflect customer needs, investigating reviews as complementary information to demand time series in the analysis of hotel demand is logical. Several social media websites, such as Ctrip and TripAdvisor, have established online review systems to encourage tourists to upload their reviews, which could affect the behaviours of other tourists by influencing their decision-making progress. However, only a few studies have conducted hotel demand forecasting by mining the sentiment value of textual reviews. The present study aims to extend the current hotel demand forecasting literature by extracting sentiment variables from hotel review texts and integrating these extracted variables into hotel demand forecasting models.

Sentiment analysis used in forecasting

Sentiment analysis is an artificial intelligence technology to evaluate text emotion objectively (Lohr, 2012). The analysis mainly focuses on lexicon construction, feature extraction, polarity determination and semantic computing. Sentiment polarity determination is one of the fundamental tasks of sentiment analysis, which is widely used to obtain the emotional information that the reviewers express and analyse their emotional states. Many methods are available to conduct sentiment polarity classification, including lexicon-based, keyword-based and machine learning methods (Cambria et al., 2016; Giachanou and Crestani, 2016). To date, sentiment indicator extraction with sentiment polarity classification is among one of the most common methods in forecasting. Thus, indicators are mainly used in product sales forecasting (Fan et al., 2017; Lau et al., 2018), unemployment forecasting (Rambaccussing and Kwiatkowski, 2020) and stock market fluctuation forecasting (Bollen et al., 2011; Nti et al., 2020; Yu et al., 2013). Fan et al. (2017) propose a novel model that combines the Bass model and Naïve Bayesian (NB) sentiment analysis and incorporates historical sales data and online reviews in the process of sales forecasting. Their results demonstrate that the combination method improves the forecasting performance in comparison with the benchmark sales forecasting models. Rambaccussing and Kwiatkowski (2020) focus on the UK economic policies reported in newspapers and propose linear support vector machine as the technique of sentiment analysis in their forecast of unemployment. Considerable improvements in prediction are found when using popular printed media-based indices.

Although sentiment information of social media textual data has been widely used in the field of sales and financial forecasting, its applications in hotel and tourism forecasting literature remain rare. Nevertheless, Önder et al. (2019b) and Colladon et al. (2019) have attempted to use web sentiment indicators for tourism demand forecasting, noting that sentiment variables can properly complement the existing explanatory variables. With developments in computer science, sentiment analysis techniques have also been further developed and improved. Some deep learning models, such as Convolutional neural network (CNN), recurrent neural network (RNN) and Long Short-Term Memory model (LSTM), have been used for sentiment polarity judgement (Priyadarshini and Cotton, 2021; Sun et al., 2017).

Among these deep learning techniques, LSTM has the advantage of handling sequences of any length and learning the temporal dependence from the data through the cell state (Hochreiter and Schmidhuber, 1997). Moreover, LSTM is designed for better memory storage to avoid the problem of gradient exploding or vanishing compared with standard RNN. In addition to these general benefits, LSTM can also learn the lead or lag order that varied with circumstances, making it well suited for processing sequence data (Bi et al., 2020; Sutskever et al., 2014). Owing to these advantages, this model has been used for tourism forecasting (Bi et al., 2020; Law et al., 2019) and sentiment extraction (Abdi et al., 2019; Liu and Guo 2019). Therefore, this study aims to adopt this promising technique to construct sentiment indices from hotel online reviews and then include these innovative variables into the autoregressive integrated moving average with exogenous variables (ARIMAX) model to examine its hotel forecasting ability.

Methodology

A forecasting framework with sentiment-based review data is proposed and illustrated in Figure 1. The first stage is the sentiment analysis and index construction by training the LSTM model, whereas the second stage uses the constructed sentiment indices in autoregressive integrated moving average with an exogenous variables model (ARIMAX) for hotel demand forecasting. The LSTM and NB models are used for sentiment polarity determination of hotel reviews, which are crawled from Ctrip and Qunar. In this step, positive or negative sentiment value is identified for each review. Sentiment information is extracted by the means of average, variance and bullish sentiment statistics matching with counts of online reviews every day. Three composite indices, namely, ASI, VSI and BSI, are then calculated according to sentiment values of the reviews. These three indices capture different aspects of the review sentiment. ASI evaluates the average sentiment conveyed by tourists and VSI measures the dispersion of sentiments. BSI is used to reflect the sentiment orientation of tourists who stay in a hotel on a particular day. Finally, the effectiveness of constructed indices is tested in an ARIMAX model by comparing its hotel demand forecast performance with benchmark models. In the present study, the forecasting performance is measured by the mean absolute percentage error (MAPE) and the root mean square error (RMSE).

Data collection and preprocessing

Ctrip.com and Qunar.com are the two prevailing online travel social media platforms in China. They are selected as the review data sources in the present study not only for their abundant data volume but also for the dominant market shares they have occupied in China. Ctrip and Qunar have long-established online review systems to encourage tourists to upload reviews for hotel stays, and both travel social media platforms restrict review posting to only the actual guests of the host hotel. The review data in the present study are acquired using an author-generated python crawling system, which accesses the API of the two social media and crawls the content from the hotel webpage. The crawled review dataset includes hotel name, textual review, review rating, posting date and travel type.

Macau is chosen to be the empirical context of the present study in testing the construction of sentiment indices. Specifically, online reviews of four Macau luxury hotels on Ctrip and Qunar, namely, Hotel Lisboa (LM), Venetian Macau (VM), Sheraton Macau (SM) and Parisian Macau (PM), are used to construct the sentiment indices. The demand for these four hotels by mainland Chinese travellers is further used in a forecasting practice to test the effectiveness of constructed indices. Macau is chosen for its attractiveness in short-haul travels from surrounding regions (Fong, 2017). Short-haul tourists are easier to be influenced by online information to make timely decisions, in comparison with medium- and long-haul tourists (Hu and Song, 2020). Furthermore, after exploring data from some other hotels in Macau, the review data of the chosen four luxury hotels reveals higher quality (in terms of missing values and outliers), which is valuable for the empirical analysis in the present study. The review volumes each day are used as the proxy of daily hotel demand by mainland Chinese visitors, given that the two social media platforms restrict review posting by only actual hotel guests (Chen et al., 2019b).

The review data retrieved from Qunar are available after October 2016, whereas the review data from Ctrip are only available after November 2017. The time span of the data is further stringed to after March 2018 due to a large number of missing values and outliers before that date. To avoid potential errors that could be caused by the Hong Kong social unrest in the second half of 2019, we only use the reviews data of four Macau luxury hotels between 1 March 2018 and 31 March 2019. After preliminary data cleaning, our sample contains 12,118 reviews for SM, 15,560 reviews for VM, 9317 reviews for PM and 12,533 reviews for LM. The reviews used in this study use simplified Chinese characters.

Further data preprocessing is carried out to eliminate the symbol marks and stopwords. Moreover, word segmentation and tokenisation are used to prepare the phrases for sentiment analysis using the LSTM. To calculate the sentiment indices from hotel reviews, a labelled hotel review sentiment corpus is prepared for AI-based model training. The prepared corpus includes 4000 positive or negative labelled reviews from the annotated corpus of Chinese hotel reviews (Tan, 2020).

Sentiment analysis and index construction

In the present study, the LSTM model (Abdi et al., 2019) and the NB model (Yu et al., 2013) are used for sentiment analysis. The sentiment of reviews is categorised as either positive or negative. In the NB model, each sentence is represented using the bag-of-words approach, where each word is weighted using term-frequency and inverse document frequency. The multinomial NB function is then used to analyse the sentiment of labelled hotel reviews. However, due to the difficulty of the NB model in capturing information from sequential data and the neglect of the syntactic structure, distinguish the meaning of different sentences by itself is not feasible for the NB model. On this note, the LSTM network is proposed to deal with sequential data of the reviews for sentiment analysis by using memory cells that preserve memory states over long periods. The LSTM model has the advantage of identifying the syntactic structure and learning long-term dependencies among the input sequences on the basis of the means of the Word2vec approach (Li et al., 2018b). The LSTM memory cell contains three main gates to regulate and store sentiment information: input gate, output gate and forgetting gate, all of which jointly determine the information stored in a given cell.

The Chinese word embedding corpus of Zhihu trained by the Skip-Gram model of the Word2vec approach (Li et al., 2018b) is used to conduct word embedding representation for labelled hotel sentiment reviews and hotel reviews of the four chosen Macau luxury hotels. The LSTM model is trained to analyse the word vector to obtain the sentiment value. This section with the LSTM model focuses on sentiment value assignment, and the forecasting ability will be examined in the next section.

The annotated hotel reviews after vectorisation are divided into a training set (80%) and a test set (20%) for each hotel. The LSTM model is trained with the feature word vector from the training set. In particular, an Adam optimiser with loss function binary cross-entropy is adopted as the optimiser, the activation ‘sigmoid’ is used to classify the sentiment of reviews data into two categories, and the threshold of sentiment value is set to 0.5 to distinguish positive and negative sentiments (scores greater than 0.5 are perceived as positive and scores lower than 0.5 are set as negative).

Then, in the test dataset, four standard measures, namely, Accuracy, Precision, Recall and F1-score, are used to evaluate the classification performance of LSTM and NB (Schütze et al., 2008). Accuracy (Acc) refers to a ratio of correctly predicted reviews to the total number of reviews, which is the most intuitive measure. Precision (Prec) refers to the percentage of correctly predicted positive (negative) reviews in the total number of predicted positive (negative) reviews. Recall (Rec) refers to the percentage of correctly predicted positive (negative) reviews in the total number of actual positive (negative) reviews. F1-score (F1) is a composite score calculated by the corresponding Precision and Recall. The following equations illustrate the computation of these four standard measures:

Acc = \frac{TP + TN}{TP + FP + FN + TN}

Pre c_{P} = \frac{TP}{TP + FP}, Pre c_{N} = \frac{TN}{TN + FN}

Re c_{P} = \frac{TP}{TP + FN}, Re c_{N} = \frac{TN}{TN + FP}

F 1_{P} = \frac{2 \cdot Pre c_{P} \cdot Re c_{P}}{Pre c_{P} + Re c_{P}}, F 1_{N} = \frac{2 \cdot Pre c_{N} \cdot Re c_{N}}{Pre c_{N} + Re c_{N}}

where True Positive/Ture Negative means the number of reviews that are correctly categorised as positive/negative, whereas False Positive/False Negative refers to the number of reviews that are categorised in the wrong group.

The outcomes of sentiment analysis of the NB and LSTM models are compared according to the four standard measures, and the better model is used in hotel demand forecasting. Three indices are constructed based on the sentiment analysis results. The BSI borrows the idea from the stock market, where the share of buy signals reflects the sentiment expressed by the traders (Antweiler and Frank, 2004; Li et al., 2018b). In the present study, BSI is used to reflect the sentiment orientation of tourists who stay in a hotel on a particular day. The BSI of online reviews in Macau luxury hotel h on date i is calculated as

BS I_{h}^{i} = \ln (\frac{1 + N_{P}^{i}}{1 + N_{N}^{i}}) .

where

N_{P}^{i}

and

N_{N}^{i}

denote the number of positive and negative reviews labelled by the sentiment model on the i^th day, respectively. The value of BSI may be either negative or positive, and higher BSI indicates a more satisfactory experience for the tourists.

The ASI and VSI are measured to evaluate central and dispersion of the sentiment in the reviews, as shown below

AS I_{h}^{i} = \frac{\sum_{j = 1}^{n} S_{j}^{i}}{n}

VS I_{h}^{i} = \frac{\sum_{j = 1}^{n} {(S_{j}^{i} - AS I_{h}^{i})}^{2}}{n}

where

S_{j}^{i}

denotes the sentiment value of the j^th review on the i^th day and n denotes the total number of reviews. ASI evaluates the average sentiment conveyed by tourists, representing the central tendency of daily sentiment. Higher ASI indicates that tourists generally have a better travel experience. The VSI measures the dispersion of sentiments, whereas higher VSI indicates that tourists have a more varied emotional experience.

Forecasting models and performance measures

The ARIMAX model is frequently used in hotel and tourism forecasting (Li et al., 2017; Pan et al., 2012; Volchek et al., 2019; Yang et al., 2015) with superior forecasting performance. The present study uses the ARIMAX model to verify the predictive power of sentiment indices extracted from hotel reviews. Six dummy variables (representing Monday through Saturday, with Sunday omitted to avoid multicollinearity) are included to capture the weekly seasonality in the data.

A univariate ARIMA model and a Naïve model are also estimated as benchmark models. The forecasts are only based on hotel review series from March 2018 to March 2019. The analyses are carried out using software R.

Two measures of forecasting accuracy are adopted to evaluate the forecasting performance of the proposed models: the MAPE and RMSE. These measures are the two most widely used measurements to assess forecasting accuracy by examining the difference between forecasts and real demand (Song and Li, 2008; Song et al., 2019; Wu et al., 2017a).

They are calculated as follows

MAPE = \frac{1}{n} \sum_{i = 1}^{n} \frac{| y_{i} - {\hat{y}}_{i} |}{y_{i}}

RMSE = \sqrt{\frac{1}{n} \sum_{i = 1}^{n} {(y_{i} - {\hat{y}}_{i})}^{2}}

where

{\hat{y}}_{i}

represents the hotel room demand measured by review counts in day i, and

y_{i}

is the observed review counts in day i.

Empirical results

In the present study, daily review counts are used as the proxy of hotel demand for the empirical test of sentiment indices. We adopt review counts to measure hotel demand due to two reasons. Firstly, in our chosen platforms, only tourists with real experience in the hotel can post their reviews. Thus, the review counts can reflect the trend and fluctuation of the hotel guest booking. Secondly, according to Chen et al. (2019b), review counts have been used as the proxy of demand and it is highly related to real booking. Therefore, this study uses daily review counts as the proxy because real booking data are not accessible.

We use the daily reviews data series of four Macau luxury hotels (Hotel Lisboa, VM, SM and PM), covering the period from 1 March 2018 to 31 March 2019. The entire time series is further divided into a training set of 365 observations from 1 March 2018 to 28 February 2019 and a testing set of 31 observations from 1 March 2019 to 31 March 2019. Rolling forecasts are generated by the proposed model to evaluate the forecasting performance. The hotel demands of four luxury hotels are plotted in Figure 2.

Sentiment analysis

The performances of the LSTM model and the NB model in sentiment analysis are evaluated by the four aforementioned measures. Table 1 summarises the evaluation results. In general, the LSTM achieves a better performance in comparison with the NB model in all the measurement aspects. In particular, the LSTM model outperforms the NB model in positive sentiment in terms of Precision (by 6.2%) and in negative sentiment in terms of Recall (by 8.2%). The F1-scores of LSTM are 3.3% and 6.1% higher than those of the NB model in positive sentiment and negative sentiment, respectively. Therefore, the trained LSTM model is used to analyse the sentiment information of customer reviews crawled from Ctrip and Qunar for four Macau luxury hotels and three sentiment indices from different measured levels, namely, BSI, ASI and VSI, are constructed for each hotel. The daily trends of BSI, ASI and VSI of four Macau luxury hotels are shown in Figures 3–5.

Table 1.

The comparisons between LSTM and NB for the sentiment classification.

Models	Predicted polarity	Precision	Recall	F1-score	Accuracy
LSTM	Positive	0.877	0.822	0.849	0.855
LSTM	Negative	0.835	0.887	0.861	0.855
NB	Positive	0.826	0.818	0.822	0.819
NB	Negative	0.812	0.820	0.816	0.819

Hotel demand forecasting

The constructed sentiment indices are tested in a hotel demand forecast model for their effectiveness as hotel demand indicators. The ARIMAX model is used to verify the predictive power of sentiment indices. An ARIMAX model (Box and Jenkins, 1970; Pankratz, 2009) takes the form of

(1 - \sum_{i = 1}^{p} Φ_{i} L^{i}) {(1 - L)}^{d} Y_{t} = \sum_{k = 1}^{u} \sum_{i = 1}^{v} η_{k i} L^{i} X_{k t} + (1 + \sum_{i = 1}^{q} θ_{i} L^{i}) ϵ_{t},

where

X_{k t}

represents the explanatory variable;

η_{k i}

is the coefficient for explanatory variables

X_{k t}

; and

L^{i}

denotes the i^th lags.

In this study, three variables are included: BSI, ASI and VSI. Thus, the ARIMAX model can be specified as

(1 - \sum_{h = 1}^{p} Φ_{h} L^{h}) {(1 - L)}^{d} H_{i, t} = \sum_{k = K_{S}}^{K_{E}} \sum_{m = M_{S}}^{M_{E}} \sum_{n = N_{S}}^{N_{E}} (α_{i, k} BS I_{i, t - k} + β_{i, m} AS I_{i, t - m} + γ_{i, n} VS I_{i, t - n}) + (1 + \sum_{j = 1}^{q} θ_{j} L^{j}) ϵ_{i, t}

where

H_{i, t}

represents the hotel demand of hotel i on date t and

ϵ_{i, t}

is the error term. These two terms and their lagged values form the standard ARMA model. The sentiment indices BSI, ASI and VSI are incorporated into the model as extra independent variables. The indices p and q govern the AR and MA order of the ARIMA model, whereas K_S, M_S and N_S (K_E, M_E and N_E) are the start (end) of lagged periods of BSI, ASI and VSI to be included, respectively. Lag periods from one to 10 of each sentiment index are initially incorporated into the ARIMAX model and the statistical significance of all coefficients are examined. For each sentiment index, the lagged variables are dropped from both ends until the occurrence of the first significant coefficient. Specifically, lags two and three are preserved for SM and lags four to six for LM, lags two to five for VM and lags two to seven for PM. Table 2 presents the estimation results of the ARIMAX model with the optimal lags of sentiment indices for all four luxury hotels.

Table 2.

Estimation results of ARIMAX models.

	Sheraton Macau	Hotel Lisboa	Venetian Macau	Parisian Macau
Bullish sentiment index (−2)	0.137* (0.0791)	—	0.122** (0.0584)	0.147*** (0.049)
Bullish sentiment index (−3)	0.108* (0.0589)	—	0.129** (0.0571)	0.117** (0.0578)
Bullish sentiment index (−4)	—	0.711* (0.369)	0.219*** (0.0553)	0.214*** (0.0582)
Bullish sentiment index (−5)	—	0.739 (0.519)	0.109** (0.0510)	0.467*** (0.0573)
Bullish sentiment index (−6)	—	0.963* (0.523)	—	0.403* (0.604)
Bullish sentiment index (−7)	—	—	—	0.118** (0.0581)
Average sentiment index (−2)	0.911 (0.573)	—	1.473** (0.605)	1.290*** (0.493)
Average sentiment index (−3)	0.964** (0.425)	—	0.481 (0.612)	0.664 (0.609)
Average sentiment index (−4)	—	0.282* (0.144)	0.439* (0.239)	0.702 (0.631)
Average sentiment index (−5)	—	0.416 (0.384)	0.493 (0.566)	0.656* (0.372)
Average sentiment index (−6)	—	0.518 (0.383)	—	0.678 (0.624)
Average sentiment index (−7)	—	—	—	0.164 (0.507)
Variance sentiment index (−2)	1.539* (0.842)	—	1.686*** (0.576)	0.635 (0.685)
Variance sentiment index (−3)	1.074* (0.620)	—	1.676*** (0.564)	2.191*** (0.817)
Variance sentiment index (−4)	—	—	0.744 (1.155)	1.004 (0.827)
Variance sentiment index (−5)	—	—	1.087* (0.612)	2.468*** (0.812)
Variance sentiment index (−6)	—	—	—	2.169*** (0.903)
Variance sentiment index (−7)	—	—	—	1.743** (0.845)
AR (1)	−0.946*** (0.0630)	—	0.136** (0.0651)	1.002*** (0.032)
AR (2)	—	—	—	−0.950*** (0.076)
MA (1)	0.268*** (0.0692)	−0.781*** (0.0501)	0.824*** (0.0382)	−1.817*** (0.0591)
MA (2)	0.670*** (0.0520)	−0.149** (0.0659)	—	1.767*** (0.139)
MA (3)	—	0.0932* (0.0486)	—	−0.814*** (0.0780)
Log likelihood	−50.7	−165.33	−106.73	−280.14
AIC	109.41	340.67	219.47	576.27
BIC	125	360.56	231.16	607.45

Notes: ***, ** and *significance at the 0.01, 0.05 and 0.10 levels, respectively. Standard errors are presented in parentheses.

The model is trained on the training set and rolling forecasts of 1–7-day-ahead are generated on the testing set (1 March 2019 to 31 March 2019). Forecast accuracy of the proposed model is calculated using the rolling forecasts and compared with those of Naïve and ARIMA models as benchmarks. These forecast accuracy measures are presented in Table 3

Table 3.

Forecasting accuracy.

Horizon	Model	Sheraton		Parisian		Lisboa		Venetian
Horizon	Model	MAPE	RMSE	MAPE	RMSE	MAPE	RMSE	MAPE	RMSE
1	Naïve	26.642	9.692	23.195	5.656	28.735	4.099	23.722	10.347
	ARIMA	19.192	7.334	18.529	5.064	15.357	2.838	15.023	6.427
	ARIMAX	18.525	7.312	18.753	5.218	15.239	2.926	14.516	5.945
2	Naïve	28.661	10.442	20.946	5.518	22.511	3.927	19.232	7.829
	ARIMA	18.052	6.802	20.445	5.415	16.369	3.143	15.148	6.566
	ARIMAX	17.483	6.699	19.401	5.249	16.062	3.058	15.261	6.532
3	Naïve	23.112	9.384	23.016	6.350	24.012	4.087	19.145	8.502
	ARIMA	17.413	6.938	21.684	5.551	15.989	2.896	15.698	6.491
	ARIMAX	16.741	6.803	19.008	5.391	16.851	3.006	14.785	6.197
4	Naïve	28.285	10.158	22.399	6.732	27.343	4.742	22.413	9.726
	ARIMA	18.556	7.116	20.510	5.458	15.813	2.862	14.653	6.411
	ARIMAX	17.428	6.834	20.982	5.853	15.503	2.823	14.097	6.262
5	Naïve	28.445	10.279	22.346	6.858	25.648	4.944	26.031	11.108
	ARIMA	17.352	6.756	20.055	5.271	16.424	2.861	16.837	6.851
	ARIMAX	16.442	6.303	19.542	5.227	16.415	2.856	16.968	7.058
6	Naïve	29.682	10.931	21.408	6.701	27.704	5.205	25.842	11.134
	ARIMA	19.897	7.615	20.494	5.343	17.164	2.995	20.318	5.314
	ARIMAX	18.027	7.023	20.291	5.894	16.823	2.987	17.457	5.034
7	Naïve	30.164	10.963	23.422	6.965	28.326	5.371	23.993	11.023
	ARIMA	18.457	7.368	20.260	5.298	16.658	2.896	17.199	6.824
	ARIMAX	17.495	6.925	21.124	5.927	16.224	2.891	17.791	6.944
Average	Naïve	27.856	10.264	22.390	6.397	26.326	4.625	22.911	9.953
	ARIMA	18.417	7.133	20.282	5.343	16.253	2.927	16.411	6.412
	ARIMAX	17.448	6.843	19.871	5.537	16.159	2.935	15.839	6.282

Note: ARIMAX: autoregressive integrated moving average with exogenous variables; MAPE: mean absolute percentage error; RMSE: root mean square error. The bold and italic numbers represent the best performance compared to the other models.

In general, the incorporation of sentiment indices improves the accuracy of short-term forecasts of 1–7-day-ahead forecasts. In particular, as for SM, the inclusion of sentiment indices improves the forecasting accuracy for all forecasts from one to 7 days ahead, demonstrating the predictive potential of sentiment-based review data. The forecast results of the other three hotels are not decisive. In the case of PM, the ARIMAX model incorporating sentiment indices outperforms the benchmarks for forecast horizons 2, 3, 5 and 6 according to MAPE, whereas the ARIMA model performs better for 1-day-ahead, 4-day-ahead and 7-day-ahead forecasts. For LM, the ARIMAX model with sentiment indices produces the lowest MAPE values for forecast horizons 1, 2, 4, 5, 6 and 7, but not for the 3-day-ahead forecast. For VM, the ARIMAX model with sentiment indices provides better forecast accuracy for forecast horizons 1, 2, 3, 4 and 6 according to RMSE and the best forecast performance except for h = 2, 5 and 7 according to MAPE. For all four hotels, the average MAPEs of the ARIMAX model with sentiment indices are the lowest. Sentiment information contained in textual reviews has considerable predictive potential to deliver accurate hotel demand forecasts. The present results extend the previous studies on tourism demand forecasting with big data, especially for review data.

Conclusion and future research directions

Conclusion

With the rapid and diverse development of the internet and smartphones, review platforms such as Ctrip and Qunar have become important platforms for users to express their feelings and search for information. A large volume of data is generated every day, potentially indicating the interests and preferences of users and enhance demand analysis research (Volchek et al., 2019). On the basis of an empirical investigation of four Macau luxury hotels, this study aims to investigate the predictive ability of online textual reviews on hotel demand. The LSTM is found to produce better review sentiment values in the sentiment analysis, and three construct sentiment indices are all proved to be effective in improving the forecast accuracy in the analysis of hotel demands. Three sentiment indices are proposed in the present study, namely, BSI, VSI and ASI. The empirical results reveal that, compared with the benchmark model without sentiment indices, hotel demand forecasting incorporating sentiment indices could improve the forecasting performance in most contexts. Nonetheless, the optimal lags of sentiment indices to be included in the forecasting practice are different across hotels, reflecting a heterogeneous choice behaviour of consumers in selecting different hotels.

This study contributes to the hospitality forecasting study from three perspectives. Firstly, the present study is among the first to use customer textual reviews for hotel demand forecasting by the mean of extracting information through deep learning sentiment analysis. While using online review data for predictions, the majority of previous studies focus merely on the numerical aspect, such as reviews volume or reviews rating (Antonio et al., 2019; Li, et al., 2020). Önder et al. (2019b) and Colladon et al. (2019) represent two exceptions that adopted textual information in their model of hotel and tourism demand. Secondly, compared with previous studies that only extracted the overall sentiment polarity of textual big data for demand analysis (Colladon et al., 2019; Önder et al., 2019), the present study quantifies sentiment information of customer reviews at different levels of measurement and constructs three sentiment indices (the BSI, VSI and ASI) for hotel demand analysis, shedding light on future hotel and tourism demand forecasting research. Thirdly, the present study represents one of the initial attempts of using LSTM to extract sentiment information from hotel reviews in the field of hotel and tourism demand forecast, following some advancements in sentiment analysis literature (Sun et al., 2017).

The findings of this study provide hotel managers managerial implications. Firstly, sentiment information contained in customer reviews can be regarded as a valuable information resource to identify the latent needs of customers (Geetha et al., 2017; Zhang et al., 2016). Constant monitoring of the sentiment indices proposed in the present study could help the managers to understand the attitudinal dynamics of the guests on hotel performance. Thus, information is crucial to the evaluation of performance, the assurance of service quality and the enhancement of the competitiveness of the hotel. Secondly, online customer reviews data are timely to reflect tourist behaviours or behaviour intentions. Thus they benefit to produce accurate demand forecasts especially under the COVID-19 outbreak.

Limitations and future research directions

A few limitations emerge from the findings of the present study. Firstly, bias may exist when the number of hotel reviews is used as the proxy of hotel demand. Further studies are encouraged to use actual hotel demand data whenever data availability allows. Secondly, the present study has a particular focus on short-haul travel from mainland China to Macau. Further investigations could explore the forecasting performance of sentiment information in other contexts, such as long-haul hotel demand forecasting. Moreover, the outbreak of the COVID-19 pandemic has caused a severe impact on the global tourism industry. The present study used pre-COVID-19 data to investigate the use of sentiment analysis in hotel demand forecast during a normal period. Future studies could consider the use of travel reviews for advanced hotel demand forecasting such as scenario forecasting and probabilistic forecasting (Qiu et al., 2021; Wu et al., 2021) in the context of COVID-19 to examine the predictive ability of sentiment analysis under crisis.

As one of the pioneering studies to use textual reviews and sentiment analysis in hotel and tourism demand forecasting literature, the present study also reveals some promising directions for future studies. In addition to sentiment information, textual reviews contain a wide variety of information that can be useful in hotel demand forecasting. The themes extracted from guest reviews could be used to infer the different needs of reviewers and therefore provide the hotel with directions for improvement. Therefore, a fine-grained sentiment analysis, such as joint sentiment-topic analysis (Li et al., 2019b), could be conducted to identify and extract multidimensional demand information, serving the purpose of precise marketing and enhancing the competitiveness of the hotel. Furthermore, economic determinants of demand are frequently used in hotel demand forecasting methods (Hu and Song, 2020). The predictive ability of the combination of these economic determinants and online customer reviews also needs further exploration.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by National Natural Science Foundation of China (No. 72071218 and 71573289), Guangdong Basic and Applied Basic Research Foundation (2020B1515020031), and Start-up Research Grant of University of Macau (SRG2019-00182-FBA).

ORCID iD

Doris Chenguang Wu

Shiteng Zhong

Richard T R Qiu

Author biographies

Doris Chenguang Wu, PhD, is professor of tourism forecasting in the Business School at Sun Yat-sen University, China. Her research interests include tourism demand forecasting and tourism big data analytics.

Shiteng Zhong, is a PhD student in the Business School at Sun Yat-sen University, China. His research interest is tourism demand forecasting.

Richard T R Qiu, PhD, is an assistant professor in the Department of Integrated Resort and Tourism Management, Faculty of Business Administration at University of Macau. His research interests are tourist choice behaviour, tourism economics and tourism demand forecasting.

Ji Wu, PhD, is an associate professor in the Business School at Sun Yat-sen University, China. His research interests include big data analytics, online community, and e-commerce.

References

Abdi

Shamsuddin

Hasan

, et al. (2019) Deep learning-based sentiment classification of evaluative text based on Multi-feature fusion. Information Processing & Management 56: 1245–1259.

Ady

Quadri-Felitti

(2015) Consumer Research Identifies How to Present Travel Review Content for More Bookings. Available at: http://webcache.googleusercontent.com/search

Antonio

De Almeida

Nunes

(2019) Big data in hotel revenue management: Exploring cancellation drivers to gain insights into booking cancellation behavior. Cornell Hospitality Quarterly 60(4): 298–319.

Antweiler

Frank

(2004) Is all that talk just noise? The information content of internet stock message boards. The Journal of Finance 59(3): 1259–1294.

Audrino

Sigrist

Ballinari

(2020) The impact of sentiment and attention measures on stock market volatility. International Journal of Forecasting 36: 334–357.

Bangwayo-Skeete

Skeete

(2015) Can Google data improve the forecasting performance of tourist arrivals? Mixed-data sampling approach. Tourism Management 46: 454–464.

Berezina

Bilgihan

Cobanoglu

, et al. (2016) Understanding satisfied and dissatisfied hotel customers: Text mining of online hotel reviews. Journal of Hospitality Marketing & Management 25(1): 1–24.

J-W

Liu

(2020) Daily tourism volume forecasting for tourist attractions. Annals of Tourism Research 83: 102923.

Blal

Sturman

(2014) The differential effects of the quality and quantity of online reviews on hotel room sales. Cornell Hospitality Quarterly 55(4): 365–375.

10.

Bollen

Mao

Zeng

(2011) Twitter mood predicts the stock market. Journal of Computational Science 2(1): 1–8.

11.

Box

GEP

Jenkins

(1970) Time Series Analysis, Forecasting and Control. San Francisco, CA: Holden Day.

12.

Cambria

Das

Bandyopadhyay

, et al. (2017) Affective computing and sentiment analysis. In: A Practical Guide to Sentiment Analysis. Cham: Springer, 1–10.

13.

Chen

, et al. (2019a) Forecasting seasonal tourism demand using a multiseries structural time series method. Journal of Travel Research 58(1): 92–103.

14.

Chen

, et al. (2019b) Measuring and managing the externality of managerial responses to online customer reviews. Information Systems Research 30: 81–96.

15.

Chen

Zheng

, et al. (2019c) Geo-tagged photo metadata processing method for Beijing inbound tourism flow. ISPRS International Journal of Geo-Information 8(12): 556.

16.

Choi

Varian

(2012) Predicting the present with Google Trends. Economic Record 88: 2–9.

17.

Colladon

Guardabascio

Innarella

(2019) Using social network and semantic analysis to analyze online travel forums and forecast tourism demand. Decision Support Systems 123: 113075.

18.

Dergiades

Mavragani

Pan

(2018) Google Trends and tourists’ arrivals: Emerging biases and proposed corrections. Tourism Management 66: 108–120.

19.

Fan

Z-P

Che

Y-J

Chen

Z-Y

(2017) Product sales forecasting using online reviews and historical sales data: A method combining the Bass model and sentiment analysis. Journal of Business Research 74: 90–100.

20.

Frechtling

(2001) Forecasting Tourism Demand: Methods and Strategies. Oxford: Butterworth-Heinemann.

21.

Fong

DKC

(2017) Macau Visitor Profile Study 2017. Macau: Institute for the Study of Commercial Gaming, University of Macau.

22.

Gao

Tang

Wang

, et al. (2018) Identifying competitors through comparative relation mining of online reviews in the restaurant industry. International Journal of Hospitality Management 71: 19–32.

23.

Geetha

Singha

Sinha

(2017) Relationship between customer sentiment and online customer ratings for hotels - an empirical analysis. Tourism Management 61: 43–54.

24.

Ghose

Ipeirotis

(2011) Estimating the helpfulness and economic impact of product reviews: Mining text and reviewer characteristics. IEEE Transactions on Knowledge and Data Engineering 23(10): 1498–1512.

25.

Ghose

Ipeirotis

(2014) Examining the impact of ranking on consumer behavior and search engine revenue. Management Science 60(7): 1632–1654.

26.

Giachanou

Crestani

(2016) Like it or not: A survey of twitter sentiment analysis methods. ACM Computing Surveys 49(2): 1–41.

27.

Gloor

Colladon

Giacomelli

, et al. (2017) The impact of virtual mirroring on customer satisfaction. Journal of Business Research 75: 67–76.

28.

Gunter

Önder

Gindl

(2019) Exploring the predictive ability of LIKES of posts on the Facebook pages of four major city DMOs in Austria. Tourism Economics 25(3): 375–401.

29.

Qiu

RTR

, et al. (2021) Hierarchical pattern recognition for tourism demand forecasting. Tourism Management 84: 104263.

30.

Y-H

Chen

Y-L

Chou

H-L

(2017) Opinion mining from online hotel reviews - a text summarization approach. Information Processing & Management 53(2): 436–449.

31.

Song

(2020) Data source combination for tourism demand forecasting. Tourism Economics 26(7): 1248–1265.

32.

Hochreiter

Schmidhuber

(1997) Long short-term memory. Neural computation 9(8): 1735–1780.

33.

Khatibi

Belém

Couto da Silva

, et al. (2020) Fine-grained tourism prediction: Impact of social and environmental features. Information Processing & Management 57(2): 102057.

34.

Lau

RYK

Zhang

(2018) Parallel aspect-oriented sentiment analysis for sales forecasting with big data. Production and Operations Management 27(10): 1775–1794.

35.

Law

Fong

DKC

, et al. (2019) Tourism demand forecasting: A deep learning approach. Annals of tourism research 75: 410–423.

36.

(2019) Introduction to the special issue: Tourism forecasting–New trends and issues. Tourism Economics 25(2): 135481661881680.

37.

Zhou

, et al. (2019a) The combination of interval forecasts in tourism. Annals of Tourism Research 75: 363–378.

38.

(2020) Forecasting tourism demand with multisource big data. Annals of Tourism Research 83: 102912.

39.

Tang

, et al. (2018a) Big data in tourism research: A literature review. Tourism Management 68: 301–323.

40.

van Dalen

van Rees

(2018b) More than just noise? Examining the information content of stock microblogs on financial markets. Journal of Information Technology 33(1): 50–69.

41.

Pan

Law

, et al. (2017) Forecasting tourism demand with composite search index. Tourism Management 59: 57–66.

42.

Mai

(2019b) The effect of online reviews on product sales: A joint sentiment-topic analysis. Information & Management 56(2): 172–184.

43.

Liu

Guo

(2019) Bidirectional LSTM with attention mechanism and convolutional layer for text classification. Neurocomputing 337: 325–338.

44.

Lohr

(2012) The Age of Big Data. New York Times, p. 11.

45.

Xiang

, et al. (2018) Effects of user-provided photos on hotel review helpfulness: An analytical approach with deep leaning. International Journal of Hospitality Management 71: 120–131.

46.

Nti

Adekoya

Weyori

(2020) Predicting stock market price movement using sentiment analysis: Evidence from ghana. Applied Computer Systems 25: 33–42.

47.

Önder

Gunter

Gindl

(2019a) Utilizing Facebook statistics in tourism demand modeling and destination marketing. Journal of Travel Research 59(2): 195–208.

48.

Önder

Gunter

Scharl

(2019b) Forecasting tourist arrivals with the help of web sentiment: A mixed-frequency modeling approach for big data. Tourism Analysis 24(4): 437–452.

49.

Pan

Song

(2012) Forecasting hotel room demand using search engine data. Journal of Hospitality and Tourism Technology 3(3): 196–210.

50.

Pan

Yang

(2017) Forecasting destination weekly hotel occupancy with big data. Journal of Travel Research 56(7): 957–970.

51.

Pankratz

(2009) Forecasting with univariate Box-Jenkins models: Concepts and cases. Hoboken: John Wiley & Sons, Vol. 224.

52.

Priyadarshini

Cotton

(2021) A novel LSTM-CNN-grid search-based deep neural network for sentiment analysis. The Journal of Supercomputing: 1–22.

53.

Pugh

(2001) Service with a smile: Emotional contagion in the service encounter. Academy of Management Journal 44(5): 1018–1027.

54.

Qiu

RTR

Dropsy

, et al. (2021) Visitor arrivals forecasts amid COVID-19: A perspective from the Asia and Pacific team. Annals of Tourism Research 88: 103155.

55.

Rambaccussing

Kwiatkowski

(2020) Forecasting with news sentiment: Evidence with UK newspapers. International Journal of Forecasting 36(4): 1501–1516.

56.

Rivera

(2016) A dynamic linear model to forecast hotel registrations in Puerto Rico using Google Trends data. Tourism Management 57: 12–20.

57.

Schütze

Manning

Raghavan

(2008) Introduction to information retrieval . Cambridge: Cambridge University Press.

58.

Song

(2008) Tourism demand modelling and forecasting-A review of recent research. Tourism Management 29(2): 203–220.

59.

Song

Qiu

RTR

Park

(2019) A review of research on tourism demand forecasting: Launching the Annals of Tourism Research Curated Collection on tourism demand forecasting. Annals of Tourism Research 75: 338–362.

60.

Starosta

Budz

Krutwig

(2019) The impact of German-speaking online media on tourist arrivals in popular tourist destinations for Europeans. Applied Economics 51(14): 1558–1573.

61.

Subramony

Pugh

(2015) Services management research: Review, integration, and future directions. Journal of Management 41(1): 349–373.

62.

Sun

Luo

Chen

(2017) A review of natural language processing techniques for opinion mining systems. Information Fusion 36: 10–25.

63.

Sutskever

Vinyals

(2014) Sequence to sequence learning with neural networks. In: Advances in Neural Information Processing Systems, pp. 3104–3112.

64.

Tan

(2020) ChnSentiCorp-Htl-ba-4000. Available at: http://github.com/Frances255/ZH-NLP/ (accessed 28 November 2020).

65.

Volchek

Liu

Song

, et al. (2019) Forecasting tourist arrivals at attractions: Search engine empowered methodologies. Tourism Economics 25(3): 425–447.

66.

Weatherford

Kimes

S E

(2003) A comparison of forecasting methods for hotel revenue management. International Journal of Forecasting 19(3): 401–415.

67.

Cao

Wen

, et al. (2021) Scenario forecasting for global tourism. Journal of Hospitality & Tourism Research 45(1): 28–51.

68.

Song

Shen

(2017a) New developments in tourism and hotel demand modeling and forecasting. International Journal of Contemporary Hospitality Management 29: 507–529.

69.

Shen

Fan

, et al. (2017b) The impact of language style on consumers^′ reactions to online reviews. Tourism Management 59: 590–596.

70.

Xiang

Schwartz

Gerdes

, et al. (2015) What can big data and text analytics tell us about hotel guest experience and satisfaction? International Journal of Hospitality Management 44: 120–130.

71.

(2016) The antecedents of customer satisfaction and dissatisfaction toward various types of hotels: A text mining approach. International Journal of Hospitality Management 55: 57–69.

72.

Yang

Pan

Evans

, et al. (2015) Forecasting Chinese tourist volume with search engine data. Tourism Management 46: 386–397.

73.

Yang

Pan

Song

(2014) Predicting hotel demand using destination marketing organization’s web traffic data. Journal of Travel Research 53(4): 433–447.

74.

Liu

Huang

, et al. (2010) Mining online reviews for predicting sales performance: A case study in the movie domain. IEEE Transactions on Knowledge and Data Engineering 24(4): 720–734.

75.

L-C

J-L

Chang

P-C

, et al. (2013) Using a contextual entropy model to expand emotion words and their intensity for the sentiment classification of stock market news. Knowledge-Based Systems 41: 89–97.

76.

Zhang

Tian

Y-X

Fan

L-W

(2020) Improving the Bass model’s predictive power through online reviews, search traffic and macroeconomic data. Annals of Operations Research 295: 881–922.

77.

Zhang

Mao

(2012) Image of all hotel scales on travel blogs: Its impact on customer loyalty. Journal of Hospitality Marketing & Management 21(2): 113–131.

78.

Zhang

, et al. (2016) Sentimental interplay between structured and unstructured user-generated contents. Online Information Review 40(1): 119–145.

79.

Zhao

Wang

(2019) Predicting overall customer satisfaction: Big data evidence from hotel online textual reviews. International Journal of Hospitality Management 76: 111–121.